Upload folder using huggingface_hub

.gitignore

@@ -81,12 +81,12 @@ tests/test-tokenizer-0
 koboldcpp.so
 koboldcpp_failsafe.so
 koboldcpp_openblas.so
-koboldcpp_openblas_noavx2.so
+koboldcpp_noavx2.so
 koboldcpp_clblast.so
 koboldcpp.dll
 koboldcpp_failsafe.dll
 koboldcpp_openblas.dll
-koboldcpp_openblas_noavx2.dll
+koboldcpp_noavx2.dll
 koboldcpp_clblast.dll
 koboldcpp_cublas.dll
 cublas64_11.dll

CMakeLists.txt

@@ -3,9 +3,9 @@
 # IT WILL NOT BE UPDATED OR MAINTAINED !!!
 
 message(STATUS "============== ============== ==============")
-message(STATUS "WARNING! Do NOT use this file. It is UNSUPPORTED for normal users. Use MAKE instead.")
-message(STATUS "It is ONLY for CUBLAS build testing on windows visual studio. IT WILL NOT BE UPDATED OR MAINTAINED !!!")
-message(STATUS "IF YOU ARE SEEING THIS, you MUST ONLY be building AN EXPERIMENAL WINDOWS CUBLAS BUILD! NOTHING ELSE WILL BE SUPPORTED !!!")
+message(STATUS "WARNING! Recommend NOT to use this file. It is UNSUPPORTED for normal users. Use MAKE instead.")
+message(STATUS "It is ONLY for CUBLAS builds on windows visual studio. IT WILL OVERWRITE YOUR EXISTING MAKEFILE !!!")
+message(STATUS "IF YOU ARE SEEING THIS, you MUST ONLY be building CUBLAS BUILDS! NOTHING ELSE WILL BE SUPPORTED !!!")
 message(STATUS "============== ============== ==============")
 
 cmake_minimum_required(VERSION 3.12) # Don't bump this version for no reason
@@ -43,11 +43,12 @@ if (NOT MSVC)
 endif()
 
 # 3rd party libs
-option(LLAMA_CUBLAS                          "llama: use cuBLAS"                                ON)
+option(LLAMA_CUBLAS                          "llama: use CUDA"                                ON)
+set(LLAMA_CUDA_MMQ_Y       "64" CACHE STRING "llama: y tile size for mmq CUDA kernels")
 set(LLAMA_CUDA_DMMV_X      "32" CACHE STRING "llama: x stride for dmmv CUDA kernels")
 set(LLAMA_CUDA_DMMV_Y       "1" CACHE STRING "llama: y block size for dmmv CUDA kernels")
 set(LLAMA_CUDA_MMV_Y        "1" CACHE STRING "llama: y block size for mmv CUDA kernels")
-option(LLAMA_CUDA_DMMV_F16                   "llama: use 16 bit floats for dmmv CUDA kernels"   OFF)
+option(LLAMA_CUDA_F16                        "llama: use 16 bit floats for dmmv CUDA kernels"   OFF)
 set(LLAMA_CUDA_KQUANTS_ITER "2" CACHE STRING "llama: iters./thread per block for Q2_K/Q6_K")
 option(LLAMA_K_QUANTS                        "llama: use k-quants"                              ON)
 
@@ -79,13 +80,15 @@ if (LLAMA_CUBLAS)
         set(GGML_V2_LEGACY_CUDA_SOURCES otherarch/ggml_v2-cuda-legacy.cu otherarch/ggml_v2-cuda-legacy.h)
 
         add_compile_definitions(GGML_USE_CUBLAS)
+        #add_compile_definitions(GGML_CUDA_CUBLAS) #remove to not use cublas
+        add_compile_definitions(GGML_CUDA_MMQ_Y=${LLAMA_CUDA_MMQ_Y})
         #add_compile_definitions(GGML_CUDA_FORCE_DMMV) #non dmmv broken for me
 
         add_compile_definitions(GGML_CUDA_DMMV_X=${LLAMA_CUDA_DMMV_X})
         add_compile_definitions(GGML_CUDA_DMMV_Y=${LLAMA_CUDA_DMMV_Y})
         add_compile_definitions(GGML_CUDA_MMV_Y=${LLAMA_CUDA_MMV_Y})
-        if (LLAMA_CUDA_DMMV_F16)
-            add_compile_definitions(GGML_CUDA_DMMV_F16)
+        if (LLAMA_CUDA_F16 OR LLAMA_CUDA_DMMV_F16)
+            add_compile_definitions(GGML_CUDA_F16)
         endif()
         add_compile_definitions(K_QUANTS_PER_ITERATION=${LLAMA_CUDA_KQUANTS_ITER})
 
@@ -96,10 +99,19 @@ if (LLAMA_CUBLAS)
         endif()
 
     if (NOT DEFINED CMAKE_CUDA_ARCHITECTURES)
-        if (LLAMA_CUDA_DMMV_F16)
-            set(CMAKE_CUDA_ARCHITECTURES "60;61") # needed for f16 CUDA intrinsics
+        # 52 == lowest CUDA 12 standard
+        # 60 == f16 CUDA intrinsics
+        # 61 == integer CUDA intrinsics
+        # 70 == (assumed) compute capability at which unrolling a loop in mul_mat_q kernels is faster
+        if (LLAMA_CUDA_F16 OR LLAMA_CUDA_DMMV_F16)
+            set(CMAKE_CUDA_ARCHITECTURES "60;61;70") # needed for f16 CUDA intrinsics
         else()
-            set(CMAKE_CUDA_ARCHITECTURES "37;52;61") # lowest CUDA 12 standard + lowest for integer intrinsics
+            message("CUDA Toolkit Version: ${CUDAToolkit_VERSION}")
+            if(CUDAToolkit_VERSION VERSION_GREATER 12)
+                 set(CMAKE_CUDA_ARCHITECTURES "52;61;70") # lowest CUDA 12 standard + lowest for integer intrinsics
+            else()
+                 set(CMAKE_CUDA_ARCHITECTURES "37;52;61;70") # lowest CUDA 12 standard + lowest for integer intrinsics
+            endif()
         endif()
     endif()
     message(STATUS "Using CUDA architectures: ${CMAKE_CUDA_ARCHITECTURES}")
@@ -120,6 +132,7 @@ if (LLAMA_ALL_WARNINGS)
             -Wshadow
             -Wstrict-prototypes
             -Wpointer-arith
+            -Wmissing-prototypes
         )
         set(cxx_flags
             -Wall
@@ -259,6 +272,8 @@ endif()
 add_library(ggml OBJECT
             ggml.c
             ggml.h
+            ggml-alloc.c
+            ggml-alloc.h
             k_quants.h
             k_quants.c
             ${GGML_SOURCES_CUDA})

Makefile

@@ -1,4 +1,4 @@
-default: koboldcpp koboldcpp_failsafe koboldcpp_openblas koboldcpp_openblas_noavx2 koboldcpp_clblast koboldcpp_cublas
+default: koboldcpp koboldcpp_failsafe koboldcpp_openblas koboldcpp_noavx2 koboldcpp_clblast koboldcpp_cublas
 tools: quantize_gpt2 quantize_gptj quantize_llama quantize_neox quantize_mpt
 dev: koboldcpp_openblas
 dev2: koboldcpp_clblast
@@ -42,7 +42,7 @@ endif
 
 # keep standard at C11 and C++11
 CFLAGS   = -I.              -I./include -I./include/CL -I./otherarch -I./otherarch/tools -Ofast -DNDEBUG -std=c11   -fPIC -DGGML_USE_K_QUANTS
-CXXFLAGS = -I. -I./examples -I./include -I./include/CL -I./otherarch -I./otherarch/tools -O3 -DNDEBUG -std=c++11 -fPIC -DGGML_USE_K_QUANTS
+CXXFLAGS = -I. -I./examples -I./include -I./include/CL -I./otherarch -I./otherarch/tools -Ofast -DNDEBUG -std=c++11 -fPIC -DGGML_USE_K_QUANTS
 LDFLAGS  =
 
 # these are used on windows, to build some libraries with extra old device compatibility
@@ -165,20 +165,34 @@ else ifdef LLAMA_CUDA_DMMV_Y
 else
 	NVCCFLAGS += -DGGML_CUDA_MMV_Y=1
 endif # LLAMA_CUDA_MMV_Y
+ifdef LLAMA_CUDA_F16
+	NVCCFLAGS += -DGGML_CUDA_F16
+endif # LLAMA_CUDA_F16
 ifdef LLAMA_CUDA_DMMV_F16
-	NVCCFLAGS += -DGGML_CUDA_DMMV_F16
+	NVCCFLAGS += -DGGML_CUDA_F16
 endif # LLAMA_CUDA_DMMV_F16
 ifdef LLAMA_CUDA_KQUANTS_ITER
 	NVCCFLAGS += -DK_QUANTS_PER_ITERATION=$(LLAMA_CUDA_KQUANTS_ITER)
 else
 	NVCCFLAGS += -DK_QUANTS_PER_ITERATION=2
 endif
+ifdef LLAMA_CUDA_MMQ_Y
+	NVCCFLAGS += -DGGML_CUDA_MMQ_Y=$(LLAMA_CUDA_MMQ_Y)
+else
+	NVCCFLAGS += -DGGML_CUDA_MMQ_Y=64
+endif # LLAMA_CUDA_MMQ_Y
+#ifdef LLAMA_CUDA_CUBLAS
+#	NVCCFLAGS += -DGGML_CUDA_CUBLAS
+#endif # LLAMA_CUDA_CUBLAS
+ifdef LLAMA_CUDA_CCBIN
+	NVCCFLAGS += -ccbin $(LLAMA_CUDA_CCBIN)
+endif
 ggml-cuda.o: ggml-cuda.cu ggml-cuda.h
-	$(NVCC) $(NVCCFLAGS) $(CXXFLAGS) $(CUBLAS_FLAGS) $(CUBLAS_CXXFLAGS) -Wno-pedantic -c $< -o $@
+	$(NVCC) $(NVCCFLAGS) $(subst -Ofast,-O3,$(CXXFLAGS)) $(CUBLAS_FLAGS) $(CUBLAS_CXXFLAGS) -Wno-pedantic -c $< -o $@
 ggml_v2-cuda.o: otherarch/ggml_v2-cuda.cu otherarch/ggml_v2-cuda.h
-	$(NVCC) $(NVCCFLAGS) $(CXXFLAGS) $(CUBLAS_FLAGS) $(CUBLAS_CXXFLAGS) -Wno-pedantic -c $< -o $@
+	$(NVCC) $(NVCCFLAGS) $(subst -Ofast,-O3,$(CXXFLAGS)) $(CUBLAS_FLAGS) $(CUBLAS_CXXFLAGS) -Wno-pedantic -c $< -o $@
 ggml_v2-cuda-legacy.o: otherarch/ggml_v2-cuda-legacy.cu otherarch/ggml_v2-cuda-legacy.h
-	$(NVCC) $(NVCCFLAGS) $(CXXFLAGS) $(CUBLAS_FLAGS) $(CUBLAS_CXXFLAGS) -Wno-pedantic -c $< -o $@
+	$(NVCC) $(NVCCFLAGS) $(subst -Ofast,-O3,$(CXXFLAGS)) $(CUBLAS_FLAGS) $(CUBLAS_CXXFLAGS) -Wno-pedantic -c $< -o $@
 endif # LLAMA_CUBLAS
 
 ifdef LLAMA_METAL
@@ -213,7 +227,7 @@ endif
 DEFAULT_BUILD =
 FAILSAFE_BUILD =
 OPENBLAS_BUILD =
-OPENBLAS_NOAVX2_BUILD =
+NOAVX2_BUILD =
 CLBLAST_BUILD =
 CUBLAS_BUILD =
 
@@ -221,7 +235,7 @@ ifeq ($(OS),Windows_NT)
 	DEFAULT_BUILD = $(CXX) $(CXXFLAGS)  $^ -shared -o [email protected] $(LDFLAGS)
 	FAILSAFE_BUILD = $(CXX) $(CXXFLAGS) $^ -shared -o [email protected] $(LDFLAGS)
 	OPENBLAS_BUILD = $(CXX) $(CXXFLAGS) $^ lib/libopenblas.lib -shared -o [email protected] $(LDFLAGS)
-	OPENBLAS_NOAVX2_BUILD = $(CXX) $(CXXFLAGS) $^ lib/libopenblas.lib -shared -o [email protected] $(LDFLAGS)
+	NOAVX2_BUILD = $(CXX) $(CXXFLAGS) $^ -shared -o [email protected] $(LDFLAGS)
 	CLBLAST_BUILD = $(CXX) $(CXXFLAGS) $^ lib/OpenCL.lib lib/clblast.lib -shared -o [email protected] $(LDFLAGS)
 
 ifdef LLAMA_CUBLAS
@@ -233,7 +247,7 @@ else
 	FAILSAFE_BUILD = $(CXX) $(CXXFLAGS) $^ -shared -o [email protected] $(LDFLAGS)
 	ifdef LLAMA_OPENBLAS
 	OPENBLAS_BUILD = $(CXX) $(CXXFLAGS) $^ $(ARCH_ADD) -lopenblas -shared -o [email protected] $(LDFLAGS)
-	OPENBLAS_NOAVX2_BUILD = $(CXX) $(CXXFLAGS) $^ $(ARCH_ADD) -lopenblas -shared -o [email protected] $(LDFLAGS)
+	NOAVX2_BUILD = $(CXX) $(CXXFLAGS) $^ $(ARCH_ADD) -lopenblas -shared -o [email protected] $(LDFLAGS)
 	endif
 	ifdef LLAMA_CLBLAST
         ifeq ($(UNAME_S),Darwin)
@@ -283,8 +297,8 @@ ggml_openblas.o: ggml.c ggml.h
 	$(CC)  $(CFLAGS) $(FULLCFLAGS) $(OPENBLAS_FLAGS) -c $< -o $@
 ggml_failsafe.o: ggml.c ggml.h
 	$(CC)  $(CFLAGS) $(NONECFLAGS) -c $< -o $@
-ggml_openblas_noavx2.o: ggml.c ggml.h
-	$(CC)  $(CFLAGS) $(SIMPLECFLAGS) $(OPENBLAS_FLAGS) -c $< -o $@
+ggml_noavx2.o: ggml.c ggml.h
+	$(CC)  $(CFLAGS) $(SIMPLECFLAGS) -c $< -o $@
 ggml_clblast.o: ggml.c ggml.h
 	$(CC)  $(CFLAGS) $(FULLCFLAGS) $(CLBLAST_FLAGS) -c $< -o $@
 ggml_cublas.o: ggml.c ggml.h
@@ -298,6 +312,10 @@ k_quants_noavx2.o: k_quants.c k_quants.h ggml.h ggml-cuda.h
 k_quants_failsafe.o: k_quants.c k_quants.h ggml.h ggml-cuda.h
 	$(CC)  $(CFLAGS) $(NONECFLAGS) -c $< -o $@
 
+#there's no intrinsics or special gpu ops used here, so we can have a universal object
+ggml-alloc.o: ggml-alloc.c ggml.h ggml-alloc.h
+	$(CC)  $(CFLAGS) -c $< -o $@
+
 #version 2 libs
 ggml_v2.o: otherarch/ggml_v2.c otherarch/ggml_v2.h
 	$(CC)  $(CFLAGS) $(FULLCFLAGS) -c $< -o $@
@@ -305,8 +323,8 @@ ggml_v2_openblas.o: otherarch/ggml_v2.c otherarch/ggml_v2.h
 	$(CC)  $(CFLAGS) $(FULLCFLAGS) $(OPENBLAS_FLAGS) -c $< -o $@
 ggml_v2_failsafe.o: otherarch/ggml_v2.c otherarch/ggml_v2.h
 	$(CC)  $(CFLAGS) $(NONECFLAGS) -c $< -o $@
-ggml_v2_openblas_noavx2.o: otherarch/ggml_v2.c otherarch/ggml_v2.h
-	$(CC)  $(CFLAGS) $(SIMPLECFLAGS) $(OPENBLAS_FLAGS) -c $< -o $@
+ggml_v2_noavx2.o: otherarch/ggml_v2.c otherarch/ggml_v2.h
+	$(CC)  $(CFLAGS) $(SIMPLECFLAGS) -c $< -o $@
 ggml_v2_clblast.o: otherarch/ggml_v2.c otherarch/ggml_v2.h
 	$(CC)  $(CFLAGS) $(FULLCFLAGS) $(CLBLAST_FLAGS) -c $< -o $@
 ggml_v2_cublas.o: otherarch/ggml_v2.c otherarch/ggml_v2.h
@@ -327,10 +345,12 @@ ggml_v2-opencl-legacy.o: otherarch/ggml_v2-opencl-legacy.c otherarch/ggml_v2-ope
 	$(CC) $(CFLAGS) -c $< -o $@
 
 # intermediate objects
-llama.o: llama.cpp ggml.h ggml-cuda.h llama.h llama-util.h
+llama.o: llama.cpp ggml.h ggml-alloc.h ggml-cuda.h ggml-metal.h llama.h llama-util.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@
 common.o: examples/common.cpp examples/common.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@
+console.o: examples/console.cpp examples/console.h
+	$(CXX) $(CXXFLAGS) -c $< -o $@
 grammar-parser.o: examples/grammar-parser.cpp examples/grammar-parser.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@
 expose.o: expose.cpp expose.h
@@ -348,37 +368,37 @@ gpttype_adapter_cublas.o: $(GPTTYPE_ADAPTER)
 	$(CXX) $(CXXFLAGS) $(CUBLAS_FLAGS) -c $< -o $@
 
 clean:
-	rm -vf *.o main quantize_llama quantize_gpt2 quantize_gptj quantize_neox quantize_mpt quantize-stats perplexity embedding benchmark-matmult save-load-state main.exe quantize_llama.exe quantize_gptj.exe quantize_gpt2.exe quantize_neox.exe quantize_mpt.exe koboldcpp.dll koboldcpp_openblas.dll koboldcpp_failsafe.dll koboldcpp_openblas_noavx2.dll koboldcpp_clblast.dll koboldcpp_cublas.dll koboldcpp.so koboldcpp_openblas.so koboldcpp_failsafe.so koboldcpp_openblas_noavx2.so koboldcpp_clblast.so koboldcpp_cublas.so
+	rm -vf *.o main quantize_llama quantize_gpt2 quantize_gptj quantize_neox quantize_mpt quantize-stats perplexity embedding benchmark-matmult save-load-state main.exe quantize_llama.exe quantize_gptj.exe quantize_gpt2.exe quantize_neox.exe quantize_mpt.exe koboldcpp.dll koboldcpp_openblas.dll koboldcpp_failsafe.dll koboldcpp_noavx2.dll koboldcpp_clblast.dll koboldcpp_cublas.dll koboldcpp.so koboldcpp_openblas.so koboldcpp_failsafe.so koboldcpp_noavx2.so koboldcpp_clblast.so koboldcpp_cublas.so
 
-main: examples/main/main.cpp build-info.h ggml.o k_quants.o llama.o common.o grammar-parser.o $(OBJS)
+main: examples/main/main.cpp build-info.h ggml.o k_quants.o ggml-alloc.o llama.o common.o console.o grammar-parser.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 	@echo
 	@echo '====  Run ./main -h for help.  ===='
 	@echo
 
 #generated libraries
-koboldcpp: ggml.o ggml_v2.o ggml_v1.o expose.o common.o gpttype_adapter.o k_quants.o $(OBJS)
+koboldcpp: ggml.o ggml_v2.o ggml_v1.o expose.o common.o gpttype_adapter.o k_quants.o ggml-alloc.o $(OBJS)
 	$(DEFAULT_BUILD)
-koboldcpp_openblas: ggml_openblas.o ggml_v2_openblas.o ggml_v1.o expose.o common.o gpttype_adapter.o k_quants.o $(OBJS)
+koboldcpp_openblas: ggml_openblas.o ggml_v2_openblas.o ggml_v1.o expose.o common.o gpttype_adapter.o k_quants.o ggml-alloc.o $(OBJS)
 	$(OPENBLAS_BUILD)
-koboldcpp_failsafe: ggml_failsafe.o ggml_v2_failsafe.o ggml_v1_failsafe.o expose.o common.o gpttype_adapter_failsafe.o k_quants_failsafe.o $(OBJS)
+koboldcpp_failsafe: ggml_failsafe.o ggml_v2_failsafe.o ggml_v1_failsafe.o expose.o common.o gpttype_adapter_failsafe.o k_quants_failsafe.o ggml-alloc.o $(OBJS)
 	$(FAILSAFE_BUILD)
-koboldcpp_openblas_noavx2: ggml_openblas_noavx2.o ggml_v2_openblas_noavx2.o ggml_v1_failsafe.o expose.o common.o gpttype_adapter_failsafe.o k_quants_noavx2.o $(OBJS)
-	$(OPENBLAS_NOAVX2_BUILD)
-koboldcpp_clblast: ggml_clblast.o ggml_v2_clblast.o ggml_v1.o expose.o common.o gpttype_adapter_clblast.o ggml-opencl.o ggml_v2-opencl.o ggml_v2-opencl-legacy.o k_quants.o $(OBJS)
+koboldcpp_noavx2: ggml_noavx2.o ggml_v2_noavx2.o ggml_v1_failsafe.o expose.o common.o gpttype_adapter_failsafe.o k_quants_noavx2.o ggml-alloc.o $(OBJS)
+	$(NOAVX2_BUILD)
+koboldcpp_clblast: ggml_clblast.o ggml_v2_clblast.o ggml_v1.o expose.o common.o gpttype_adapter_clblast.o ggml-opencl.o ggml_v2-opencl.o ggml_v2-opencl-legacy.o k_quants.o ggml-alloc.o $(OBJS)
 	$(CLBLAST_BUILD)
-koboldcpp_cublas: ggml_cublas.o ggml_v2_cublas.o ggml_v1.o expose.o common.o gpttype_adapter_cublas.o k_quants.o $(CUBLAS_OBJS) $(OBJS)
+koboldcpp_cublas: ggml_cublas.o ggml_v2_cublas.o ggml_v1.o expose.o common.o gpttype_adapter_cublas.o k_quants.o ggml-alloc.o $(CUBLAS_OBJS) $(OBJS)
 	$(CUBLAS_BUILD)
 
-quantize_llama: examples/quantize/quantize.cpp ggml.o llama.o k_quants.o
+quantize_llama: examples/quantize/quantize.cpp ggml.o llama.o k_quants.o ggml-alloc.o
 	$(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS)
-quantize_gptj: ggml.o llama.o k_quants.o otherarch/tools/gptj_quantize.cpp otherarch/tools/common-ggml.cpp
+quantize_gptj: ggml.o llama.o k_quants.o ggml-alloc.o otherarch/tools/gptj_quantize.cpp otherarch/tools/common-ggml.cpp
 	$(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS)
-quantize_gpt2: ggml.o llama.o k_quants.o otherarch/tools/gpt2_quantize.cpp otherarch/tools/common-ggml.cpp
+quantize_gpt2: ggml.o llama.o k_quants.o ggml-alloc.o otherarch/tools/gpt2_quantize.cpp otherarch/tools/common-ggml.cpp
 	$(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS)
-quantize_neox: ggml.o llama.o k_quants.o otherarch/tools/neox_quantize.cpp otherarch/tools/common-ggml.cpp
+quantize_neox: ggml.o llama.o k_quants.o ggml-alloc.o otherarch/tools/neox_quantize.cpp otherarch/tools/common-ggml.cpp
 	$(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS)
-quantize_mpt: ggml.o llama.o k_quants.o otherarch/tools/mpt_quantize.cpp otherarch/tools/common-ggml.cpp
+quantize_mpt: ggml.o llama.o k_quants.o ggml-alloc.o otherarch/tools/mpt_quantize.cpp otherarch/tools/common-ggml.cpp
 	$(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS)
 
 

README.md

@@ -1,6 +1,80 @@
----
-title: Koboldcpp
-sdk: docker
-colorFrom: blue
-colorTo: green
----
+# koboldcpp
+
+KoboldCpp is an easy-to-use AI text-generation software for GGML models. It's a single self contained distributable from Concedo, that builds off llama.cpp, and adds a versatile Kobold API endpoint, additional format support, backward compatibility, as well as a fancy UI with persistent stories, editing tools, save formats, memory, world info, author's note, characters, scenarios and everything Kobold and Kobold Lite have to offer.
+
+![Preview](media/preview.png)
+
+## Usage
+- **[Download the latest .exe release here](https://github.com/LostRuins/koboldcpp/releases/latest)** or clone the git repo.
+- Windows binaries are provided in the form of **koboldcpp.exe**, which is a pyinstaller wrapper for a few **.dll** files and **koboldcpp.py**. If you feel concerned, you may prefer to rebuild it yourself with the provided makefiles and scripts.
+- Weights are not included, you can use the official llama.cpp `quantize.exe` to generate them from your official weight files (or download them from other places such as [TheBloke's Huggingface](https://huggingface.co/TheBloke).
+- To run, execute **koboldcpp.exe** or drag and drop your quantized `ggml_model.bin` file onto the .exe, and then connect with Kobold or Kobold Lite. If you're not on windows, then run the script **KoboldCpp.py** after compiling the libraries.
+- Launching with no command line arguments displays a GUI containing a subset of configurable settings. Generally you dont have to change much besides the `Presets` and `GPU Layers`. Read the `--help` for more info about each settings.
+- By default, you can connect to http://localhost:5001
+- You can also run it using the command line `koboldcpp.exe [ggml_model.bin] [port]`. For info, please check `koboldcpp.exe --help`
+- Default context size to small? Try `--contextsize 3072` to 1.5x your context size! without much perplexity gain. Note that you'll have to increase the max context in the Kobold Lite UI as well (click and edit the number text field).
+- Big context too slow? Try the `--smartcontext` flag to reduce prompt processing frequency. Also, you can try to run with your GPU using CLBlast, with `--useclblast` flag for a speedup
+- Want even more speedup? Combine `--useclblast` with `--gpulayers` to offload entire layers to the GPU! **Much faster, but uses more VRAM**. Experiment to determine number of layers to offload, and reduce by a few if you run out of memory.
+- If you are having crashes or issues, you can try turning off BLAS with the `--noblas` flag. You can also try running in a non-avx2 compatibility mode with `--noavx2`. Lastly, you can try turning off mmap with `--nommap`.
+
+For more information, be sure to run the program with the `--help` flag.
+
+## OSX and Linux
+- You will have to compile your binaries from source. A makefile is provided, simply run `make`
+- If you want you can also link your own install of OpenBLAS manually with `make LLAMA_OPENBLAS=1`
+- Alternatively, if you want you can also link your own install of CLBlast manually with `make LLAMA_CLBLAST=1`, for this you will need to obtain and link OpenCL and CLBlast libraries.
+  - For Arch Linux: Install `cblas` `openblas` and `clblast`.
+  - For Debian: Install `libclblast-dev` and `libopenblas-dev`.
+- For a full featured build, do `make LLAMA_OPENBLAS=1 LLAMA_CLBLAST=1 LLAMA_CUBLAS=1`
+- After all binaries are built, you can run the python script with the command `koboldcpp.py [ggml_model.bin] [port]`
+- Note: Many OSX users have found that the using Accelerate is actually faster than OpenBLAS. To try, you may wish to run with `--noblas` and compare speeds.
+
+## Compiling on Windows
+- You're encouraged to use the .exe released, but if you want to compile your binaries from source at Windows, the easiest way is:
+  - Use the latest release of w64devkit (https://github.com/skeeto/w64devkit). Be sure to use the "vanilla one", not i686 or other different stuff. If you try they will conflit with the precompiled libs!
+  - Make sure you are using the w64devkit integrated terminal, then run 'make' at the KoboldCpp source folder. This will create the .dll files.
+  - If you want to generate the .exe file, make sure you have the python module PyInstaller installed with pip ('pip install PyInstaller').
+  - Run the script make_pyinstaller.bat at a regular terminal (or Windows Explorer).
+  - The koboldcpp.exe file will be at your dist folder.
+- If you wish to use your own version of the additional Windows libraries (OpenCL, CLBlast and OpenBLAS), you can do it with:
+  - OpenCL - tested with https://github.com/KhronosGroup/OpenCL-SDK . If you wish to compile it, follow the repository instructions. You will need vcpkg.
+  - CLBlast - tested with https://github.com/CNugteren/CLBlast . If you wish to compile it you will need to reference the OpenCL files. It will only generate the ".lib" file if you compile using MSVC.
+  - OpenBLAS - tested with https://github.com/xianyi/OpenBLAS .
+  - Move the respectives .lib files to the /lib folder of your project, overwriting the older files.
+  - Also, replace the existing versions of the corresponding .dll files located in the project directory root (e.g. libopenblas.dll).
+  - Make the KoboldCPP project using the instructions above.
+
+## Android (Termux) Alternative method
+- See https://github.com/ggerganov/llama.cpp/pull/1828/files
+
+## Using CuBLAS
+- If you're on Windows with an Nvidia GPU you can get CUDA support out of the box using the `--usecublas` flag, make sure you select the correct .exe with CUDA support.
+- You can attempt a CuBLAS build with `LLAMA_CUBLAS=1` or using the provided CMake file (best for visual studio users). If you use the CMake file to build, copy the `koboldcpp_cublas.dll` generated into the same directory as the `koboldcpp.py` file. If you are bundling executables, you may need to include CUDA dynamic libraries (such as `cublasLt64_11.dll` and `cublas64_11.dll`) in order for the executable to work correctly on a different PC.
+
+## Questions and Help
+- **First, please check out [The KoboldCpp FAQ and Knowledgebase](https://github.com/LostRuins/koboldcpp/wiki) which may already have answers to your questions! Also please search through past issues and discussions.**
+- If you cannot find an answer, open an issue on this github, or find us on the [KoboldAI Discord](https://koboldai.org/discord).
+
+## Considerations
+- For Windows: No installation, single file executable, (It Just Works)
+- Since v1.0.6, requires libopenblas, the prebuilt windows binaries are included in this repo. If not found, it will fall back to a mode without BLAS.
+- Since v1.15, requires CLBlast if enabled, the prebuilt windows binaries are included in this repo. If not found, it will fall back to a mode without CLBlast.
+- Since v1.33, you can set the context size to be above what the model supports officially. It does increases perplexity but should still work well below 4096 even on untuned models. (For GPT-NeoX, GPT-J, and LLAMA models) Customize this with `--ropeconfig`.
+- **I plan to keep backwards compatibility with ALL past llama.cpp AND alpaca.cpp models**. But you are also encouraged to reconvert/update your models if possible for best results.
+
+## License
+- The original GGML library and llama.cpp by ggerganov are licensed under the MIT License
+- However, Kobold Lite is licensed under the AGPL v3.0 License
+- The other files are also under the AGPL v3.0 License unless otherwise stated
+
+## Notes
+- Generation delay scales linearly with original prompt length. If OpenBLAS is enabled then prompt ingestion becomes about 2-3x faster. This is automatic on windows, but will require linking on OSX and Linux. CLBlast speeds this up even further, and `--gpulayers` + `--useclblast` more so.
+- I have heard of someone claiming a false AV positive report. The exe is a simple pyinstaller bundle that includes the necessary python scripts and dlls to run. If this still concerns you, you might wish to rebuild everything from source code using the makefile, and you can rebuild the exe yourself with pyinstaller by using `make_pyinstaller.bat`
+- Supported GGML models (Includes backward compatibility for older versions/legacy GGML models, though some newer features might be unavailable):
+  - LLAMA and LLAMA2 (LLaMA / Alpaca / GPT4All / Vicuna / Koala / Pygmalion 7B / Metharme 7B / WizardLM and many more)
+  - GPT-2 / Cerebras
+  - GPT-J
+  - RWKV
+  - GPT-NeoX / Pythia / StableLM / Dolly / RedPajama
+  - MPT models
+

convert.py

@@ -133,7 +133,7 @@ TENSORS_SET = set(TENSORS_LIST)
 
 def find_n_mult(n_ff: int, n_embd: int) -> int:
     # hardcoded magic range
-    for n_mult in range(256, 1, -1):
+    for n_mult in range(8192, 1, -1):
         calc_ff = (((8*n_embd) // 3 + n_mult - 1) // n_mult)*n_mult
         if calc_ff == n_ff:
             return n_mult
@@ -141,11 +141,12 @@ def find_n_mult(n_ff: int, n_embd: int) -> int:
 
 @dataclass
 class Params:
-    n_vocab: int
-    n_embd:  int
-    n_mult:  int
-    n_head:  int
-    n_layer: int
+    n_vocab:   int
+    n_embd:    int
+    n_mult:    int
+    n_head:    int
+    n_layer:   int
+    n_kv_head: Optional[int]  # This parameter is only used for Llama 2
 
     @staticmethod
     def guessed(model: 'LazyModel') -> 'Params':
@@ -167,11 +168,12 @@ class Params:
         n_head=n_embd // 128 # guessed
 
         return Params(
-            n_vocab = n_vocab,
-            n_embd  = n_embd,
-            n_mult  = 256,
-            n_head  = n_head,
-            n_layer = n_layer,
+            n_vocab   = n_vocab,
+            n_embd    = n_embd,
+            n_mult    = 256,
+            n_head    = n_head,
+            n_layer   = n_layer,
+            n_kv_head = None,
         )
 
     @staticmethod
@@ -183,15 +185,17 @@ class Params:
         n_head  = config["num_attention_heads"];
         n_layer = config["num_hidden_layers"];
         n_ff    = config["intermediate_size"];
+        n_kv_head = config.get("num_key_value_heads")
 
         n_mult = find_n_mult(n_ff, n_embd);
 
         return Params(
-            n_vocab = n_vocab,
-            n_embd  = n_embd,
-            n_mult  = n_mult,
-            n_head  = n_head,
-            n_layer = n_layer,
+            n_vocab   = n_vocab,
+            n_embd    = n_embd,
+            n_mult    = n_mult,
+            n_head    = n_head,
+            n_layer   = n_layer,
+            n_kv_head = n_kv_head,
         )
 
     # LLaMA v2 70B params.json
@@ -200,21 +204,22 @@ class Params:
     def loadOriginalParamsJson(model: 'LazyModel', config_path: 'Path') -> 'Params':
         config = json.load(open(config_path))
 
-        n_vocab = config["vocab_size"];
-        n_embd  = config["dim"];
-        n_head  = config["n_heads"];
-        n_layer = config["n_layers"];
-        n_mult  = config["multiple_of"];
+        n_vocab   = config["vocab_size"];
+        n_embd    = config["dim"];
+        n_head    = config["n_heads"];
+        n_layer   = config["n_layers"];
+        n_mult    = config["multiple_of"];
 
         if n_vocab == -1:
             n_vocab = model["tok_embeddings.weight"].shape[0]
 
         return Params(
-            n_vocab = n_vocab,
-            n_embd  = n_embd,
-            n_mult  = n_mult,
-            n_head  = n_head,
-            n_layer = n_layer,
+            n_vocab   = n_vocab,
+            n_embd    = n_embd,
+            n_mult    = n_mult,
+            n_head    = n_head,
+            n_layer   = n_layer,
+            n_kv_head = None,
         )
 
     @staticmethod
@@ -317,10 +322,12 @@ class GGMLVocab:
 Vocab = Union[SentencePieceVocab, GGMLVocab]
 
 
-def permute(weights: NDArray, n_head: int) -> NDArray:
+def permute(weights: NDArray, n_head: int, n_kv_head: Optional[int] = None) -> NDArray:
+    if n_kv_head is not None and n_head != n_kv_head:
+        n_head //= n_kv_head
     return (weights.reshape(n_head, 2, weights.shape[0] // n_head // 2, *weights.shape[1:])
-                   .swapaxes(1, 2)
-                   .reshape(weights.shape))
+                .swapaxes(1, 2)
+                .reshape(weights.shape))
 
 
 def dequantize_q4(qvalues_pack32: NDArray, scales: NDArray, addends: Optional[NDArray], g_idx: Optional[NDArray]) -> NDArray:
@@ -368,7 +375,7 @@ class Tensor(metaclass=ABCMeta):
     @abstractmethod
     def astype(self, data_type: DataType) -> 'Tensor': ...
     @abstractmethod
-    def permute(self, n_head: int) -> 'Tensor': ...
+    def permute(self, n_head: int, n_kv_head: Optional[int] = None) -> 'Tensor': ...
     @abstractmethod
     def permute_part(self, n_part: int, n_head: int) -> 'UnquantizedTensor': ...
     @abstractmethod
@@ -406,8 +413,8 @@ class UnquantizedTensor(Tensor):
         r = self.ndarray.shape[0] // 3
         return UnquantizedTensor(self.ndarray[r * n_part : r * n_part + r, ...])
 
-    def permute(self, n_head: int) -> 'UnquantizedTensor':
-        return UnquantizedTensor(permute(self.ndarray, n_head))
+    def permute(self, n_head: int, n_kv_head: Optional[int] = None) -> 'UnquantizedTensor':
+        return UnquantizedTensor(permute(self.ndarray, n_head, n_kv_head))
 
 
 def load_unquantized(lazy_tensor: 'LazyTensor', expected_dtype: Any = None, convert: bool = False) -> NDArray:
@@ -455,26 +462,34 @@ class GGMLQuantizedTensor(Tensor):
     def to_ggml(self) -> 'GGMLQuantizedTensor':
         return self
 
-    def permute(self, n_head: int) -> 'GGMLQuantizedTensor':
-        return GGMLQuantizedTensor(permute(self.ndarray, n_head), self.shape, self.data_type)
+    def permute(self, n_head: int, n_kv_head: Optional[int] = None) -> 'GGMLQuantizedTensor':
+        return GGMLQuantizedTensor(permute(self.ndarray, n_head, n_kv_head), self.shape, self.data_type)
+
+    def permute_part(self, n_part: int, n_head: int) -> 'UnquantizedTensor':
+        r = self.ndarray.shape[0] // 3
+        return UnquantizedTensor(permute(self.ndarray[r * n_part : r * n_part + r, ...], n_head))
 
+    def part(self, n_part: int) -> 'UnquantizedTensor':
+        r = self.ndarray.shape[0] // 3
+        return UnquantizedTensor(self.ndarray[r * n_part : r * n_part + r, ...])
 
 GGMLCompatibleTensor = Union[UnquantizedTensor, GGMLQuantizedTensor]
 
 
 class DeferredPermutedTensor(Tensor):
-    def __init__(self, base: Tensor, n_head: int) -> None:
+    def __init__(self, base: Tensor, n_head: int, n_kv_head: Optional[int] = None) -> None:
         self.base = base
         self.n_head = n_head
+        self.n_kv_head = n_kv_head
         self.data_type = self.base.data_type
 
     def astype(self, data_type: DataType) -> Tensor:
-        return self.base.astype(data_type).permute(self.n_head)
+        return self.base.astype(data_type).permute(self.n_head, self.n_kv_head)
 
     def to_ggml(self) -> GGMLCompatibleTensor:
-        return self.base.to_ggml().permute(self.n_head)
+        return self.base.to_ggml().permute(self.n_head, self.n_kv_head)
 
-    def permute(self, n_head: int) -> Tensor:
+    def permute(self, n_head: int, n_kv_head: Optional[int] = None) -> Tensor:
         raise Exception("shouldn't permute twice")
 
 
@@ -566,8 +581,8 @@ class GPTQForLLaMaQuantizedTensor(Tensor):
         ret.data_type = QuantizedDataType(groupsize=new_groupsize, have_addends=True, have_g_idx=False)
         return ret
 
-    def permute(self, n_head: int) -> Tensor:
-        return DeferredPermutedTensor(self, n_head)
+    def permute(self, n_head: int, n_kv_head: Optional[int] = None) -> Tensor:
+        return DeferredPermutedTensor(self, n_head, n_kv_head)
 
     def to_ggml(self) -> GGMLQuantizedTensor:
         # The output format looks like this:
@@ -698,10 +713,10 @@ def merge_multifile_models(models_plus: List[ModelPlus]) -> ModelPlus:
     return ModelPlus(model, paths, format, vocab)
 
 
-def permute_lazy(lazy_tensor: LazyTensor, n_head: int) -> LazyTensor:
+def permute_lazy(lazy_tensor: LazyTensor, n_head: int, n_kv_head: Optional[int] = None) -> LazyTensor:
     def load() -> Tensor:
-        return lazy_tensor.load().permute(n_head)
-    return LazyTensor(load, lazy_tensor.shape, lazy_tensor.data_type, f'permute({n_head}) ' + lazy_tensor.description)
+        return lazy_tensor.load().permute(n_head, n_kv_head)
+    return LazyTensor(load, lazy_tensor.shape, lazy_tensor.data_type, f'permute({n_head}, {n_kv_head}) ' + lazy_tensor.description)
 
 def permute_part_lazy(lazy_tensor: LazyTensor, n_part: int, n_head: int) -> LazyTensor:
     def load() -> Tensor:
@@ -726,7 +741,7 @@ def convert_transformers_to_orig(model: LazyModel, params: Params) -> LazyModel:
     for i in itertools.count():
         if f"model.layers.{i}.self_attn.q_proj.weight" in model:
             out[f"layers.{i}.attention.wq.weight"] = permute_lazy(model[f"model.layers.{i}.self_attn.q_proj.weight"], params.n_head)
-            out[f"layers.{i}.attention.wk.weight"] = permute_lazy(model[f"model.layers.{i}.self_attn.k_proj.weight"], params.n_head)
+            out[f"layers.{i}.attention.wk.weight"] = permute_lazy(model[f"model.layers.{i}.self_attn.k_proj.weight"], params.n_head, params.n_kv_head)
             out[f"layers.{i}.attention.wv.weight"] = model[f"model.layers.{i}.self_attn.v_proj.weight"]
         elif f"model.layers.{i}.self_attn.W_pack.weight" in model:
             out[f"layers.{i}.attention.wq.weight"] = permute_part_lazy(model[f"model.layers.{i}.self_attn.W_pack.weight"], 0, params.n_head)

examples/CMakeLists.txt

@@ -13,6 +13,8 @@ set(TARGET common)
 add_library(${TARGET} OBJECT
     common.h
     common.cpp
+    console.h
+    console.cpp
     grammar-parser.h
     grammar-parser.cpp
     )

examples/common.cpp

@@ -25,7 +25,6 @@
 #else
 #include <sys/ioctl.h>
 #include <unistd.h>
-#include <wchar.h>
 #endif
 
 #if defined(_MSC_VER)
@@ -195,6 +194,12 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
                 break;
             }
             params.rope_freq_scale = std::stof(argv[i]);
+        } else if (arg == "--rope-scale") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.rope_freq_scale = 1.0f/std::stof(argv[i]);
         } else if (arg == "--memory-f32") {
             params.memory_f16 = false;
         } else if (arg == "--top-p") {
@@ -329,6 +334,8 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
             params.instruct = true;
         } else if (arg == "--multiline-input") {
             params.multiline_input = true;
+        } else if (arg == "--simple-io") {
+            params.simple_io = true;
         } else if (arg == "--color") {
             params.use_color = true;
         } else if (arg == "--mlock") {
@@ -352,7 +359,7 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
 #ifdef GGML_USE_CUBLAS
             params.main_gpu = std::stoi(argv[i]);
 #else
-      fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS. It is not possible to set a main GPU.\n");
+            fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS. It is not possible to set a main GPU.\n");
 #endif
         } else if (arg == "--tensor-split" || arg == "-ts") {
             if (++i >= argc) {
@@ -376,13 +383,19 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
                 }
             }
 #else
-      fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS. It is not possible to set a tensor split.\n");
+            fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS. It is not possible to set a tensor split.\n");
+#endif // GGML_USE_CUBLAS
+        } else if (arg == "--mul-mat-q" || arg == "-mmq") {
+#ifdef GGML_USE_CUBLAS
+            params.mul_mat_q = true;
+#else
+            fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS. It is not possible to use mul_mat_q kernels.\n");
 #endif // GGML_USE_CUBLAS
         } else if (arg == "--low-vram" || arg == "-lv") {
 #ifdef GGML_USE_CUBLAS
             params.low_vram = true;
 #else
-      fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS. It is not possible to set lower vram usage.\n");
+            fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS. It is not possible to set lower vram usage.\n");
 #endif // GGML_USE_CUBLAS
         } else if (arg == "--no-mmap") {
             params.use_mmap = false;
@@ -402,8 +415,14 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
             params.antiprompt.push_back(argv[i]);
         } else if (arg == "--perplexity") {
             params.perplexity = true;
-        } else if (arg == "--perplexity-lines") {
-            params.perplexity_lines = true;
+        } else if (arg == "--hellaswag") {
+            params.hellaswag = true;
+        } else if (arg == "--hellaswag-tasks") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.hellaswag_tasks = std::stoi(argv[i]);
         } else if (arg == "--ignore-eos") {
             params.logit_bias[llama_token_eos()] = -INFINITY;
         } else if (arg == "--no-penalize-nl") {
@@ -551,16 +570,18 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     fprintf(stdout, "  --cfg-negative-prompt PROMPT \n");
     fprintf(stdout, "                        negative prompt to use for guidance. (default: empty)\n");
     fprintf(stdout, "  --cfg-scale N         strength of guidance (default: %f, 1.0 = disable)\n", params.cfg_scale);
-    fprintf(stdout, "  --rope-freq-base N    RoPE base frequency (default: %.1f)\n", params.rope_freq_base);
-    fprintf(stdout, "  --rope-freq-scale N   RoPE frequency scaling factor (default: %g)\n", params.rope_freq_scale);
+    fprintf(stdout, "  --rope-scale N        RoPE context linear scaling factor, inverse of --rope-freq-scale (default: %g)\n", 1.0f/params.rope_freq_scale);
+    fprintf(stdout, "  --rope-freq-base N    RoPE base frequency, used by NTK-aware scaling (default: %.1f)\n", params.rope_freq_base);
+    fprintf(stdout, "  --rope-freq-scale N   RoPE frequency linear scaling factor, inverse of --rope-scale (default: %g)\n", params.rope_freq_scale);
     fprintf(stdout, "  --ignore-eos          ignore end of stream token and continue generating (implies --logit-bias 2-inf)\n");
     fprintf(stdout, "  --no-penalize-nl      do not penalize newline token\n");
     fprintf(stdout, "  --memory-f32          use f32 instead of f16 for memory key+value (default: disabled)\n");
     fprintf(stdout, "                        not recommended: doubles context memory required and no measurable increase in quality\n");
     fprintf(stdout, "  --temp N              temperature (default: %.1f)\n", (double)params.temp);
     fprintf(stdout, "  --perplexity          compute perplexity over each ctx window of the prompt\n");
-    fprintf(stdout, "  --perplexity-lines    compute perplexity over each line of the prompt\n");
-    fprintf(stdout, "  --keep                number of tokens to keep from the initial prompt (default: %d, -1 = all)\n", params.n_keep);
+    fprintf(stdout, "  --hellaswag           compute HellaSwag score over random tasks from datafile supplied with -f\n");
+    fprintf(stdout, "  --hellaswag-tasks N   number of tasks to use when computing the HellaSwag score (default: %zu)\n", params.hellaswag_tasks);
+    fprintf(stdout, "  --keep N              number of tokens to keep from the initial prompt (default: %d, -1 = all)\n", params.n_keep);
     fprintf(stdout, "  --chunks N            max number of chunks to process (default: %d, -1 = all)\n", params.n_chunks);
     if (llama_mlock_supported()) {
         fprintf(stdout, "  --mlock               force system to keep model in RAM rather than swapping or compressing\n");
@@ -578,10 +599,14 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     fprintf(stdout, "                        how to split tensors across multiple GPUs, comma-separated list of proportions, e.g. 3,1\n");
     fprintf(stdout, "  -mg i, --main-gpu i   the GPU to use for scratch and small tensors\n" );
     fprintf(stdout, "  -lv, --low-vram       don't allocate VRAM scratch buffer\n" );
+    fprintf(stdout, "  -mmq, --mul-mat-q     use experimental mul_mat_q CUDA kernels instead of cuBLAS. TEMP!!!\n" );
+    fprintf(stdout, "                        Reduces VRAM usage by 700/970/1430 MiB for 7b/13b/33b but prompt processing speed\n" );
+    fprintf(stdout, "                        is still suboptimal, especially q2_K, q3_K, q5_K, and q6_K.\n" );
 #endif
     fprintf(stdout, "  --mtest               compute maximum memory usage\n");
     fprintf(stdout, "  --export              export the computation graph to 'llama.ggml'\n");
     fprintf(stdout, "  --verbose-prompt      print prompt before generation\n");
+    fprintf(stderr, "  --simple-io           use basic IO for better compatibility in subprocesses and limited consoles\n");
     fprintf(stdout, "  --lora FNAME          apply LoRA adapter (implies --no-mmap)\n");
     fprintf(stdout, "  --lora-base FNAME     optional model to use as a base for the layers modified by the LoRA adapter\n");
     fprintf(stdout, "  -m FNAME, --model FNAME\n");
@@ -630,6 +655,7 @@ struct llama_context_params llama_context_params_from_gpt_params(const gpt_param
     lparams.main_gpu        = params.main_gpu;
     lparams.tensor_split    = params.tensor_split;
     lparams.low_vram        = params.low_vram;
+    lparams.mul_mat_q       = params.mul_mat_q;
     lparams.seed            = params.seed;
     lparams.f16_kv          = params.memory_f16;
     lparams.use_mmap        = params.use_mmap;
@@ -673,376 +699,3 @@ std::tuple<struct llama_model *, struct llama_context *> llama_init_from_gpt_par
 
     return std::make_tuple(model, lctx);
 }
-
-void console_init(console_state & con_st) {
-#if defined(_WIN32)
-    // Windows-specific console initialization
-    DWORD dwMode = 0;
-    con_st.hConsole = GetStdHandle(STD_OUTPUT_HANDLE);
-    if (con_st.hConsole == INVALID_HANDLE_VALUE || !GetConsoleMode(con_st.hConsole, &dwMode)) {
-        con_st.hConsole = GetStdHandle(STD_ERROR_HANDLE);
-        if (con_st.hConsole != INVALID_HANDLE_VALUE && (!GetConsoleMode(con_st.hConsole, &dwMode))) {
-            con_st.hConsole = NULL;
-        }
-    }
-    if (con_st.hConsole) {
-        // Enable ANSI colors on Windows 10+
-        if (con_st.use_color && !(dwMode & ENABLE_VIRTUAL_TERMINAL_PROCESSING)) {
-            SetConsoleMode(con_st.hConsole, dwMode | ENABLE_VIRTUAL_TERMINAL_PROCESSING);
-        }
-        // Set console output codepage to UTF8
-        SetConsoleOutputCP(CP_UTF8);
-    }
-    HANDLE hConIn = GetStdHandle(STD_INPUT_HANDLE);
-    if (hConIn != INVALID_HANDLE_VALUE && GetConsoleMode(hConIn, &dwMode)) {
-        // Set console input codepage to UTF16
-        _setmode(_fileno(stdin), _O_WTEXT);
-
-        // Turn off ICANON (ENABLE_LINE_INPUT) and ECHO (ENABLE_ECHO_INPUT)
-        dwMode &= ~(ENABLE_LINE_INPUT | ENABLE_ECHO_INPUT);
-        SetConsoleMode(hConIn, dwMode);
-    }
-#else
-    // POSIX-specific console initialization
-    struct termios new_termios;
-    tcgetattr(STDIN_FILENO, &con_st.prev_state);
-    new_termios = con_st.prev_state;
-    new_termios.c_lflag &= ~(ICANON | ECHO);
-    new_termios.c_cc[VMIN] = 1;
-    new_termios.c_cc[VTIME] = 0;
-    tcsetattr(STDIN_FILENO, TCSANOW, &new_termios);
-
-    con_st.tty = fopen("/dev/tty", "w+");
-    if (con_st.tty != nullptr) {
-        con_st.out = con_st.tty;
-    }
-
-    setlocale(LC_ALL, "");
-#endif
-}
-
-void console_cleanup(console_state & con_st) {
-    // Reset console color
-    console_set_color(con_st, CONSOLE_COLOR_DEFAULT);
-
-#if !defined(_WIN32)
-    if (con_st.tty != nullptr) {
-        con_st.out = stdout;
-        fclose(con_st.tty);
-        con_st.tty = nullptr;
-    }
-    // Restore the terminal settings on POSIX systems
-    tcsetattr(STDIN_FILENO, TCSANOW, &con_st.prev_state);
-#endif
-}
-
-/* Keep track of current color of output, and emit ANSI code if it changes. */
-void console_set_color(console_state & con_st, console_color_t color) {
-    if (con_st.use_color && con_st.color != color) {
-        fflush(stdout);
-        switch(color) {
-            case CONSOLE_COLOR_DEFAULT:
-                fprintf(con_st.out, ANSI_COLOR_RESET);
-                break;
-            case CONSOLE_COLOR_PROMPT:
-                fprintf(con_st.out, ANSI_COLOR_YELLOW);
-                break;
-            case CONSOLE_COLOR_USER_INPUT:
-                fprintf(con_st.out, ANSI_BOLD ANSI_COLOR_GREEN);
-                break;
-            case CONSOLE_COLOR_ERROR:
-                fprintf(con_st.out, ANSI_BOLD ANSI_COLOR_RED);
-                break;
-        }
-        con_st.color = color;
-        fflush(con_st.out);
-    }
-}
-
-char32_t getchar32() {
-#if defined(_WIN32)
-    HANDLE hConsole = GetStdHandle(STD_INPUT_HANDLE);
-    wchar_t high_surrogate = 0;
-
-    while (true) {
-        INPUT_RECORD record;
-        DWORD count;
-        if (!ReadConsoleInputW(hConsole, &record, 1, &count) || count == 0) {
-            return WEOF;
-        }
-
-        if (record.EventType == KEY_EVENT && record.Event.KeyEvent.bKeyDown) {
-            wchar_t wc = record.Event.KeyEvent.uChar.UnicodeChar;
-            if (wc == 0) {
-                continue;
-            }
-
-            if ((wc >= 0xD800) && (wc <= 0xDBFF)) { // Check if wc is a high surrogate
-                high_surrogate = wc;
-                continue;
-            } else if ((wc >= 0xDC00) && (wc <= 0xDFFF)) { // Check if wc is a low surrogate
-                if (high_surrogate != 0) { // Check if we have a high surrogate
-                    return ((high_surrogate - 0xD800) << 10) + (wc - 0xDC00) + 0x10000;
-                }
-            }
-
-            high_surrogate = 0; // Reset the high surrogate
-            return static_cast<char32_t>(wc);
-        }
-    }
-#else
-    wchar_t wc = getwchar();
-    if (static_cast<wint_t>(wc) == WEOF) {
-        return WEOF;
-    }
-
-#if WCHAR_MAX == 0xFFFF
-    if ((wc >= 0xD800) && (wc <= 0xDBFF)) { // Check if wc is a high surrogate
-        wchar_t low_surrogate = getwchar();
-        if ((low_surrogate >= 0xDC00) && (low_surrogate <= 0xDFFF)) { // Check if the next wchar is a low surrogate
-            return (static_cast<char32_t>(wc & 0x03FF) << 10) + (low_surrogate & 0x03FF) + 0x10000;
-        }
-    }
-    if ((wc >= 0xD800) && (wc <= 0xDFFF)) { // Invalid surrogate pair
-        return 0xFFFD; // Return the replacement character U+FFFD
-    }
-#endif
-
-    return static_cast<char32_t>(wc);
-#endif
-}
-
-void pop_cursor(console_state & con_st) {
-#if defined(_WIN32)
-    if (con_st.hConsole != NULL) {
-        CONSOLE_SCREEN_BUFFER_INFO bufferInfo;
-        GetConsoleScreenBufferInfo(con_st.hConsole, &bufferInfo);
-
-        COORD newCursorPosition = bufferInfo.dwCursorPosition;
-        if (newCursorPosition.X == 0) {
-            newCursorPosition.X = bufferInfo.dwSize.X - 1;
-            newCursorPosition.Y -= 1;
-        } else {
-            newCursorPosition.X -= 1;
-        }
-
-        SetConsoleCursorPosition(con_st.hConsole, newCursorPosition);
-        return;
-    }
-#endif
-    putc('\b', con_st.out);
-}
-
-int estimateWidth(char32_t codepoint) {
-#if defined(_WIN32)
-    return 1;
-#else
-    return wcwidth(codepoint);
-#endif
-}
-
-int put_codepoint(console_state & con_st, const char* utf8_codepoint, size_t length, int expectedWidth) {
-#if defined(_WIN32)
-    CONSOLE_SCREEN_BUFFER_INFO bufferInfo;
-    if (!GetConsoleScreenBufferInfo(con_st.hConsole, &bufferInfo)) {
-        // go with the default
-        return expectedWidth;
-    }
-    COORD initialPosition = bufferInfo.dwCursorPosition;
-    DWORD nNumberOfChars = length;
-    WriteConsole(con_st.hConsole, utf8_codepoint, nNumberOfChars, &nNumberOfChars, NULL);
-
-    CONSOLE_SCREEN_BUFFER_INFO newBufferInfo;
-    GetConsoleScreenBufferInfo(con_st.hConsole, &newBufferInfo);
-
-    // Figure out our real position if we're in the last column
-    if (utf8_codepoint[0] != 0x09 && initialPosition.X == newBufferInfo.dwSize.X - 1) {
-        DWORD nNumberOfChars;
-        WriteConsole(con_st.hConsole, &" \b", 2, &nNumberOfChars, NULL);
-        GetConsoleScreenBufferInfo(con_st.hConsole, &newBufferInfo);
-    }
-
-    int width = newBufferInfo.dwCursorPosition.X - initialPosition.X;
-    if (width < 0) {
-        width += newBufferInfo.dwSize.X;
-    }
-    return width;
-#else
-    // we can trust expectedWidth if we've got one
-    if (expectedWidth >= 0 || con_st.tty == nullptr) {
-        fwrite(utf8_codepoint, length, 1, con_st.out);
-        return expectedWidth;
-    }
-
-    fputs("\033[6n", con_st.tty); // Query cursor position
-    int x1, x2, y1, y2;
-    int results = 0;
-    results = fscanf(con_st.tty, "\033[%d;%dR", &y1, &x1);
-
-    fwrite(utf8_codepoint, length, 1, con_st.tty);
-
-    fputs("\033[6n", con_st.tty); // Query cursor position
-    results += fscanf(con_st.tty, "\033[%d;%dR", &y2, &x2);
-
-    if (results != 4) {
-        return expectedWidth;
-    }
-
-    int width = x2 - x1;
-    if (width < 0) {
-        // Calculate the width considering text wrapping
-        struct winsize w;
-        ioctl(STDOUT_FILENO, TIOCGWINSZ, &w);
-        width += w.ws_col;
-    }
-    return width;
-#endif
-}
-
-void replace_last(console_state & con_st, char ch) {
-#if defined(_WIN32)
-    pop_cursor(con_st);
-    put_codepoint(con_st, &ch, 1, 1);
-#else
-    fprintf(con_st.out, "\b%c", ch);
-#endif
-}
-
-void append_utf8(char32_t ch, std::string & out) {
-    if (ch <= 0x7F) {
-        out.push_back(static_cast<unsigned char>(ch));
-    } else if (ch <= 0x7FF) {
-        out.push_back(static_cast<unsigned char>(0xC0 | ((ch >> 6) & 0x1F)));
-        out.push_back(static_cast<unsigned char>(0x80 | (ch & 0x3F)));
-    } else if (ch <= 0xFFFF) {
-        out.push_back(static_cast<unsigned char>(0xE0 | ((ch >> 12) & 0x0F)));
-        out.push_back(static_cast<unsigned char>(0x80 | ((ch >> 6) & 0x3F)));
-        out.push_back(static_cast<unsigned char>(0x80 | (ch & 0x3F)));
-    } else if (ch <= 0x10FFFF) {
-        out.push_back(static_cast<unsigned char>(0xF0 | ((ch >> 18) & 0x07)));
-        out.push_back(static_cast<unsigned char>(0x80 | ((ch >> 12) & 0x3F)));
-        out.push_back(static_cast<unsigned char>(0x80 | ((ch >> 6) & 0x3F)));
-        out.push_back(static_cast<unsigned char>(0x80 | (ch & 0x3F)));
-    } else {
-        // Invalid Unicode code point
-    }
-}
-
-// Helper function to remove the last UTF-8 character from a string
-void pop_back_utf8_char(std::string & line) {
-    if (line.empty()) {
-        return;
-    }
-
-    size_t pos = line.length() - 1;
-
-    // Find the start of the last UTF-8 character (checking up to 4 bytes back)
-    for (size_t i = 0; i < 3 && pos > 0; ++i, --pos) {
-        if ((line[pos] & 0xC0) != 0x80) break; // Found the start of the character
-    }
-    line.erase(pos);
-}
-
-bool console_readline(console_state & con_st, std::string & line) {
-    console_set_color(con_st, CONSOLE_COLOR_USER_INPUT);
-    if (con_st.out != stdout) {
-        fflush(stdout);
-    }
-
-    line.clear();
-    std::vector<int> widths;
-    bool is_special_char = false;
-    bool end_of_stream = false;
-
-    char32_t input_char;
-    while (true) {
-        fflush(con_st.out); // Ensure all output is displayed before waiting for input
-        input_char = getchar32();
-
-        if (input_char == '\r' || input_char == '\n') {
-            break;
-        }
-
-        if (input_char == (char32_t) WEOF || input_char == 0x04 /* Ctrl+D*/) {
-            end_of_stream = true;
-            break;
-        }
-
-        if (is_special_char) {
-            console_set_color(con_st, CONSOLE_COLOR_USER_INPUT);
-            replace_last(con_st, line.back());
-            is_special_char = false;
-        }
-
-        if (input_char == '\033') { // Escape sequence
-            char32_t code = getchar32();
-            if (code == '[' || code == 0x1B) {
-                // Discard the rest of the escape sequence
-                while ((code = getchar32()) != (char32_t) WEOF) {
-                    if ((code >= 'A' && code <= 'Z') || (code >= 'a' && code <= 'z') || code == '~') {
-                        break;
-                    }
-                }
-            }
-        } else if (input_char == 0x08 || input_char == 0x7F) { // Backspace
-            if (!widths.empty()) {
-                int count;
-                do {
-                    count = widths.back();
-                    widths.pop_back();
-                    // Move cursor back, print space, and move cursor back again
-                    for (int i = 0; i < count; i++) {
-                        replace_last(con_st, ' ');
-                        pop_cursor(con_st);
-                    }
-                    pop_back_utf8_char(line);
-                } while (count == 0 && !widths.empty());
-            }
-        } else {
-            int offset = line.length();
-            append_utf8(input_char, line);
-            int width = put_codepoint(con_st, line.c_str() + offset, line.length() - offset, estimateWidth(input_char));
-            if (width < 0) {
-                width = 0;
-            }
-            widths.push_back(width);
-        }
-
-        if (!line.empty() && (line.back() == '\\' || line.back() == '/')) {
-            console_set_color(con_st, CONSOLE_COLOR_PROMPT);
-            replace_last(con_st, line.back());
-            is_special_char = true;
-        }
-    }
-
-    bool has_more = con_st.multiline_input;
-    if (is_special_char) {
-        replace_last(con_st, ' ');
-        pop_cursor(con_st);
-
-        char last = line.back();
-        line.pop_back();
-        if (last == '\\') {
-            line += '\n';
-            fputc('\n', con_st.out);
-            has_more = !has_more;
-        } else {
-            // llama will just eat the single space, it won't act as a space
-            if (line.length() == 1 && line.back() == ' ') {
-                line.clear();
-                pop_cursor(con_st);
-            }
-            has_more = false;
-        }
-    } else {
-        if (end_of_stream) {
-            has_more = false;
-        } else {
-            line += '\n';
-            fputc('\n', con_st.out);
-        }
-    }
-
-    fflush(con_st.out);
-    return has_more;
-}

examples/common.h

@@ -11,11 +11,6 @@
 #include <unordered_map>
 #include <tuple>
 
-#if !defined (_WIN32)
-#include <stdio.h>
-#include <termios.h>
-#endif
-
 //
 // CLI argument parsing
 //
@@ -70,7 +65,11 @@ struct gpt_params {
     std::string lora_adapter = "";  // lora adapter path
     std::string lora_base    = "";  // base model path for the lora adapter
 
-    bool low_vram          = false;   // if true, reduce VRAM usage at the cost of performance
+    bool hellaswag         = false; // compute HellaSwag score over random tasks from datafile supplied in prompt
+    size_t hellaswag_tasks = 400;   // number of tasks to use when computing the HellaSwag score
+
+    bool low_vram          = false; // if true, reduce VRAM usage at the cost of performance
+    bool mul_mat_q         = false; // if true, use experimental mul_mat_q kernels
     bool memory_f16        = true;  // use f16 instead of f32 for memory kv
     bool random_prompt     = false; // do not randomize prompt if none provided
     bool use_color         = false; // use color to distinguish generations and inputs
@@ -81,12 +80,12 @@ struct gpt_params {
     bool embedding         = false; // get only sentence embedding
     bool interactive_first = false; // wait for user input immediately
     bool multiline_input   = false; // reverse the usage of `\`
+    bool simple_io         = false; // improves compatibility with subprocesses and limited consoles
 
     bool input_prefix_bos  = false; // prefix BOS to user inputs, preceding input_prefix
     bool instruct          = false; // instruction mode (used for Alpaca models)
     bool penalize_nl       = true;  // consider newlines as a repeatable token
     bool perplexity        = false; // compute perplexity over the prompt
-    bool perplexity_lines  = false; // compute perplexity over each line of the prompt
     bool use_mmap          = true;  // use mmap for faster loads
     bool use_mlock         = false; // use mlock to keep model in memory
     bool mem_test          = false; // compute maximum memory usage
@@ -113,42 +112,3 @@ std::vector<llama_token> llama_tokenize(struct llama_context * ctx, const std::s
 
 std::tuple<struct llama_model *, struct llama_context *> llama_init_from_gpt_params(const gpt_params & params);
 struct llama_context_params llama_context_params_from_gpt_params(const gpt_params & params);
-
-//
-// Console utils
-//
-
-#define ANSI_COLOR_RED     "\x1b[31m"
-#define ANSI_COLOR_GREEN   "\x1b[32m"
-#define ANSI_COLOR_YELLOW  "\x1b[33m"
-#define ANSI_COLOR_BLUE    "\x1b[34m"
-#define ANSI_COLOR_MAGENTA "\x1b[35m"
-#define ANSI_COLOR_CYAN    "\x1b[36m"
-#define ANSI_COLOR_RESET   "\x1b[0m"
-#define ANSI_BOLD          "\x1b[1m"
-
-enum console_color_t {
-    CONSOLE_COLOR_DEFAULT=0,
-    CONSOLE_COLOR_PROMPT,
-    CONSOLE_COLOR_USER_INPUT,
-    CONSOLE_COLOR_ERROR
-};
-
-struct console_state {
-    bool multiline_input = false;
-    bool use_color = false;
-    console_color_t color = CONSOLE_COLOR_DEFAULT;
-
-    FILE* out = stdout;
-#if defined (_WIN32)
-    void* hConsole;
-#else
-    FILE* tty = nullptr;
-    termios prev_state;
-#endif
-};
-
-void console_init(console_state & con_st);
-void console_cleanup(console_state & con_st);
-void console_set_color(console_state & con_st, console_color_t color);
-bool console_readline(console_state & con_st, std::string & line);

examples/console.cpp

@@ -0,0 +1,496 @@
+#include "console.h"
+#include <vector>
+#include <iostream>
+
+#if defined(_WIN32)
+#define WIN32_LEAN_AND_MEAN
+#ifndef NOMINMAX
+#define NOMINMAX
+#endif
+#include <windows.h>
+#include <fcntl.h>
+#include <io.h>
+#else
+#include <climits>
+#include <sys/ioctl.h>
+#include <unistd.h>
+#include <wchar.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <signal.h>
+#include <termios.h>
+#endif
+
+#define ANSI_COLOR_RED     "\x1b[31m"
+#define ANSI_COLOR_GREEN   "\x1b[32m"
+#define ANSI_COLOR_YELLOW  "\x1b[33m"
+#define ANSI_COLOR_BLUE    "\x1b[34m"
+#define ANSI_COLOR_MAGENTA "\x1b[35m"
+#define ANSI_COLOR_CYAN    "\x1b[36m"
+#define ANSI_COLOR_RESET   "\x1b[0m"
+#define ANSI_BOLD          "\x1b[1m"
+
+namespace console {
+
+    //
+    // Console state
+    //
+
+    static bool      advanced_display = false;
+    static bool      simple_io        = true;
+    static display_t current_display  = reset;
+
+    static FILE*     out              = stdout;
+
+#if defined (_WIN32)
+    static void*     hConsole;
+#else
+    static FILE*     tty              = nullptr;
+    static termios   initial_state;
+#endif
+
+    //
+    // Init and cleanup
+    //
+
+    void init(bool use_simple_io, bool use_advanced_display) {
+        advanced_display = use_advanced_display;
+        simple_io = use_simple_io;
+#if defined(_WIN32)
+        // Windows-specific console initialization
+        DWORD dwMode = 0;
+        hConsole = GetStdHandle(STD_OUTPUT_HANDLE);
+        if (hConsole == INVALID_HANDLE_VALUE || !GetConsoleMode(hConsole, &dwMode)) {
+            hConsole = GetStdHandle(STD_ERROR_HANDLE);
+            if (hConsole != INVALID_HANDLE_VALUE && (!GetConsoleMode(hConsole, &dwMode))) {
+                hConsole = nullptr;
+                simple_io = true;
+            }
+        }
+        if (hConsole) {
+            // Enable ANSI colors on Windows 10+
+            if (advanced_display && !(dwMode & ENABLE_VIRTUAL_TERMINAL_PROCESSING)) {
+                SetConsoleMode(hConsole, dwMode | ENABLE_VIRTUAL_TERMINAL_PROCESSING);
+            }
+            // Set console output codepage to UTF8
+            SetConsoleOutputCP(CP_UTF8);
+        }
+        HANDLE hConIn = GetStdHandle(STD_INPUT_HANDLE);
+        if (hConIn != INVALID_HANDLE_VALUE && GetConsoleMode(hConIn, &dwMode)) {
+            // Set console input codepage to UTF16
+            _setmode(_fileno(stdin), _O_WTEXT);
+
+            // Set ICANON (ENABLE_LINE_INPUT) and ECHO (ENABLE_ECHO_INPUT)
+            if (simple_io) {
+                dwMode |= ENABLE_LINE_INPUT | ENABLE_ECHO_INPUT;
+            } else {
+                dwMode &= ~(ENABLE_LINE_INPUT | ENABLE_ECHO_INPUT);
+            }
+            if (!SetConsoleMode(hConIn, dwMode)) {
+                simple_io = true;
+            }
+        }
+#else
+        // POSIX-specific console initialization
+        if (!simple_io) {
+            struct termios new_termios;
+            tcgetattr(STDIN_FILENO, &initial_state);
+            new_termios = initial_state;
+            new_termios.c_lflag &= ~(ICANON | ECHO);
+            new_termios.c_cc[VMIN] = 1;
+            new_termios.c_cc[VTIME] = 0;
+            tcsetattr(STDIN_FILENO, TCSANOW, &new_termios);
+
+            tty = fopen("/dev/tty", "w+");
+            if (tty != nullptr) {
+                out = tty;
+            }
+        }
+
+        setlocale(LC_ALL, "");
+#endif
+    }
+
+    void cleanup() {
+        // Reset console display
+        set_display(reset);
+
+#if !defined(_WIN32)
+        // Restore settings on POSIX systems
+        if (!simple_io) {
+            if (tty != nullptr) {
+                out = stdout;
+                fclose(tty);
+                tty = nullptr;
+            }
+            tcsetattr(STDIN_FILENO, TCSANOW, &initial_state);
+        }
+#endif
+    }
+
+    //
+    // Display and IO
+    //
+
+    // Keep track of current display and only emit ANSI code if it changes
+    void set_display(display_t display) {
+        if (advanced_display && current_display != display) {
+            fflush(stdout);
+            switch(display) {
+                case reset:
+                    fprintf(out, ANSI_COLOR_RESET);
+                    break;
+                case prompt:
+                    fprintf(out, ANSI_COLOR_YELLOW);
+                    break;
+                case user_input:
+                    fprintf(out, ANSI_BOLD ANSI_COLOR_GREEN);
+                    break;
+                case error:
+                    fprintf(out, ANSI_BOLD ANSI_COLOR_RED);
+            }
+            current_display = display;
+            fflush(out);
+        }
+    }
+
+    char32_t getchar32() {
+#if defined(_WIN32)
+        HANDLE hConsole = GetStdHandle(STD_INPUT_HANDLE);
+        wchar_t high_surrogate = 0;
+
+        while (true) {
+            INPUT_RECORD record;
+            DWORD count;
+            if (!ReadConsoleInputW(hConsole, &record, 1, &count) || count == 0) {
+                return WEOF;
+            }
+
+            if (record.EventType == KEY_EVENT && record.Event.KeyEvent.bKeyDown) {
+                wchar_t wc = record.Event.KeyEvent.uChar.UnicodeChar;
+                if (wc == 0) {
+                    continue;
+                }
+
+                if ((wc >= 0xD800) && (wc <= 0xDBFF)) { // Check if wc is a high surrogate
+                    high_surrogate = wc;
+                    continue;
+                }
+                if ((wc >= 0xDC00) && (wc <= 0xDFFF)) { // Check if wc is a low surrogate
+                    if (high_surrogate != 0) { // Check if we have a high surrogate
+                        return ((high_surrogate - 0xD800) << 10) + (wc - 0xDC00) + 0x10000;
+                    }
+                }
+
+                high_surrogate = 0; // Reset the high surrogate
+                return static_cast<char32_t>(wc);
+            }
+        }
+#else
+        wchar_t wc = getwchar();
+        if (static_cast<wint_t>(wc) == WEOF) {
+            return WEOF;
+        }
+
+#if WCHAR_MAX == 0xFFFF
+        if ((wc >= 0xD800) && (wc <= 0xDBFF)) { // Check if wc is a high surrogate
+            wchar_t low_surrogate = getwchar();
+            if ((low_surrogate >= 0xDC00) && (low_surrogate <= 0xDFFF)) { // Check if the next wchar is a low surrogate
+                return (static_cast<char32_t>(wc & 0x03FF) << 10) + (low_surrogate & 0x03FF) + 0x10000;
+            }
+        }
+        if ((wc >= 0xD800) && (wc <= 0xDFFF)) { // Invalid surrogate pair
+            return 0xFFFD; // Return the replacement character U+FFFD
+        }
+#endif
+
+        return static_cast<char32_t>(wc);
+#endif
+    }
+
+    void pop_cursor() {
+#if defined(_WIN32)
+        if (hConsole != NULL) {
+            CONSOLE_SCREEN_BUFFER_INFO bufferInfo;
+            GetConsoleScreenBufferInfo(hConsole, &bufferInfo);
+
+            COORD newCursorPosition = bufferInfo.dwCursorPosition;
+            if (newCursorPosition.X == 0) {
+                newCursorPosition.X = bufferInfo.dwSize.X - 1;
+                newCursorPosition.Y -= 1;
+            } else {
+                newCursorPosition.X -= 1;
+            }
+
+            SetConsoleCursorPosition(hConsole, newCursorPosition);
+            return;
+        }
+#endif
+        putc('\b', out);
+    }
+
+    int estimateWidth(char32_t codepoint) {
+#if defined(_WIN32)
+        return 1;
+#else
+        return wcwidth(codepoint);
+#endif
+    }
+
+    int put_codepoint(const char* utf8_codepoint, size_t length, int expectedWidth) {
+#if defined(_WIN32)
+        CONSOLE_SCREEN_BUFFER_INFO bufferInfo;
+        if (!GetConsoleScreenBufferInfo(hConsole, &bufferInfo)) {
+            // go with the default
+            return expectedWidth;
+        }
+        COORD initialPosition = bufferInfo.dwCursorPosition;
+        DWORD nNumberOfChars = length;
+        WriteConsole(hConsole, utf8_codepoint, nNumberOfChars, &nNumberOfChars, NULL);
+
+        CONSOLE_SCREEN_BUFFER_INFO newBufferInfo;
+        GetConsoleScreenBufferInfo(hConsole, &newBufferInfo);
+
+        // Figure out our real position if we're in the last column
+        if (utf8_codepoint[0] != 0x09 && initialPosition.X == newBufferInfo.dwSize.X - 1) {
+            DWORD nNumberOfChars;
+            WriteConsole(hConsole, &" \b", 2, &nNumberOfChars, NULL);
+            GetConsoleScreenBufferInfo(hConsole, &newBufferInfo);
+        }
+
+        int width = newBufferInfo.dwCursorPosition.X - initialPosition.X;
+        if (width < 0) {
+            width += newBufferInfo.dwSize.X;
+        }
+        return width;
+#else
+        // We can trust expectedWidth if we've got one
+        if (expectedWidth >= 0 || tty == nullptr) {
+            fwrite(utf8_codepoint, length, 1, out);
+            return expectedWidth;
+        }
+
+        fputs("\033[6n", tty); // Query cursor position
+        int x1;
+        int y1;
+        int x2;
+        int y2;
+        int results = 0;
+        results = fscanf(tty, "\033[%d;%dR", &y1, &x1);
+
+        fwrite(utf8_codepoint, length, 1, tty);
+
+        fputs("\033[6n", tty); // Query cursor position
+        results += fscanf(tty, "\033[%d;%dR", &y2, &x2);
+
+        if (results != 4) {
+            return expectedWidth;
+        }
+
+        int width = x2 - x1;
+        if (width < 0) {
+            // Calculate the width considering text wrapping
+            struct winsize w;
+            ioctl(STDOUT_FILENO, TIOCGWINSZ, &w);
+            width += w.ws_col;
+        }
+        return width;
+#endif
+    }
+
+    void replace_last(char ch) {
+#if defined(_WIN32)
+        pop_cursor();
+        put_codepoint(&ch, 1, 1);
+#else
+        fprintf(out, "\b%c", ch);
+#endif
+    }
+
+    void append_utf8(char32_t ch, std::string & out) {
+        if (ch <= 0x7F) {
+            out.push_back(static_cast<unsigned char>(ch));
+        } else if (ch <= 0x7FF) {
+            out.push_back(static_cast<unsigned char>(0xC0 | ((ch >> 6) & 0x1F)));
+            out.push_back(static_cast<unsigned char>(0x80 | (ch & 0x3F)));
+        } else if (ch <= 0xFFFF) {
+            out.push_back(static_cast<unsigned char>(0xE0 | ((ch >> 12) & 0x0F)));
+            out.push_back(static_cast<unsigned char>(0x80 | ((ch >> 6) & 0x3F)));
+            out.push_back(static_cast<unsigned char>(0x80 | (ch & 0x3F)));
+        } else if (ch <= 0x10FFFF) {
+            out.push_back(static_cast<unsigned char>(0xF0 | ((ch >> 18) & 0x07)));
+            out.push_back(static_cast<unsigned char>(0x80 | ((ch >> 12) & 0x3F)));
+            out.push_back(static_cast<unsigned char>(0x80 | ((ch >> 6) & 0x3F)));
+            out.push_back(static_cast<unsigned char>(0x80 | (ch & 0x3F)));
+        } else {
+            // Invalid Unicode code point
+        }
+    }
+
+    // Helper function to remove the last UTF-8 character from a string
+    void pop_back_utf8_char(std::string & line) {
+        if (line.empty()) {
+            return;
+        }
+
+        size_t pos = line.length() - 1;
+
+        // Find the start of the last UTF-8 character (checking up to 4 bytes back)
+        for (size_t i = 0; i < 3 && pos > 0; ++i, --pos) {
+            if ((line[pos] & 0xC0) != 0x80) {
+                break; // Found the start of the character
+            }
+        }
+        line.erase(pos);
+    }
+
+    bool readline_advanced(std::string & line, bool multiline_input) {
+        if (out != stdout) {
+            fflush(stdout);
+        }
+
+        line.clear();
+        std::vector<int> widths;
+        bool is_special_char = false;
+        bool end_of_stream = false;
+
+        char32_t input_char;
+        while (true) {
+            fflush(out); // Ensure all output is displayed before waiting for input
+            input_char = getchar32();
+
+            if (input_char == '\r' || input_char == '\n') {
+                break;
+            }
+
+            if (input_char == (char32_t) WEOF || input_char == 0x04 /* Ctrl+D*/) {
+                end_of_stream = true;
+                break;
+            }
+
+            if (is_special_char) {
+                set_display(user_input);
+                replace_last(line.back());
+                is_special_char = false;
+            }
+
+            if (input_char == '\033') { // Escape sequence
+                char32_t code = getchar32();
+                if (code == '[' || code == 0x1B) {
+                    // Discard the rest of the escape sequence
+                    while ((code = getchar32()) != (char32_t) WEOF) {
+                        if ((code >= 'A' && code <= 'Z') || (code >= 'a' && code <= 'z') || code == '~') {
+                            break;
+                        }
+                    }
+                }
+            } else if (input_char == 0x08 || input_char == 0x7F) { // Backspace
+                if (!widths.empty()) {
+                    int count;
+                    do {
+                        count = widths.back();
+                        widths.pop_back();
+                        // Move cursor back, print space, and move cursor back again
+                        for (int i = 0; i < count; i++) {
+                            replace_last(' ');
+                            pop_cursor();
+                        }
+                        pop_back_utf8_char(line);
+                    } while (count == 0 && !widths.empty());
+                }
+            } else {
+                int offset = line.length();
+                append_utf8(input_char, line);
+                int width = put_codepoint(line.c_str() + offset, line.length() - offset, estimateWidth(input_char));
+                if (width < 0) {
+                    width = 0;
+                }
+                widths.push_back(width);
+            }
+
+            if (!line.empty() && (line.back() == '\\' || line.back() == '/')) {
+                set_display(prompt);
+                replace_last(line.back());
+                is_special_char = true;
+            }
+        }
+
+        bool has_more = multiline_input;
+        if (is_special_char) {
+            replace_last(' ');
+            pop_cursor();
+
+            char last = line.back();
+            line.pop_back();
+            if (last == '\\') {
+                line += '\n';
+                fputc('\n', out);
+                has_more = !has_more;
+            } else {
+                // llama will just eat the single space, it won't act as a space
+                if (line.length() == 1 && line.back() == ' ') {
+                    line.clear();
+                    pop_cursor();
+                }
+                has_more = false;
+            }
+        } else {
+            if (end_of_stream) {
+                has_more = false;
+            } else {
+                line += '\n';
+                fputc('\n', out);
+            }
+        }
+
+        fflush(out);
+        return has_more;
+    }
+
+    bool readline_simple(std::string & line, bool multiline_input) {
+#if defined(_WIN32)
+        std::wstring wline;
+        if (!std::getline(std::wcin, wline)) {
+            // Input stream is bad or EOF received
+            line.clear();
+            GenerateConsoleCtrlEvent(CTRL_C_EVENT, 0);
+            return false;
+        }
+
+        int size_needed = WideCharToMultiByte(CP_UTF8, 0, &wline[0], (int)wline.size(), NULL, 0, NULL, NULL);
+        line.resize(size_needed);
+        WideCharToMultiByte(CP_UTF8, 0, &wline[0], (int)wline.size(), &line[0], size_needed, NULL, NULL);
+#else
+        if (!std::getline(std::cin, line)) {
+            // Input stream is bad or EOF received
+            line.clear();
+            return false;
+        }
+#endif
+        if (!line.empty()) {
+            char last = line.back();
+            if (last == '/') { // Always return control on '/' symbol
+                line.pop_back();
+                return false;
+            }
+            if (last == '\\') { // '\\' changes the default action
+                line.pop_back();
+                multiline_input = !multiline_input;
+            }
+        }
+        line += '\n';
+
+        // By default, continue input if multiline_input is set
+        return multiline_input;
+    }
+
+    bool readline(std::string & line, bool multiline_input) {
+        set_display(user_input);
+
+        if (simple_io) {
+            return readline_simple(line, multiline_input);
+        }
+        return readline_advanced(line, multiline_input);
+    }
+
+}

examples/console.h

@@ -0,0 +1,19 @@
+// Console functions
+
+#pragma once
+
+#include <string>
+
+namespace console {
+    enum display_t {
+        reset = 0,
+        prompt,
+        user_input,
+        error
+    };
+
+    void init(bool use_simple_io, bool use_advanced_display);
+    void cleanup();
+    void set_display(display_t display);
+    bool readline(std::string & line, bool multiline_input);
+}

examples/embd-input/embd-input-lib.cpp

@@ -30,7 +30,7 @@ struct MyModel* create_mymodel(int argc, char ** argv) {
     fprintf(stderr, "%s: build = %d (%s)\n", __func__, BUILD_NUMBER, BUILD_COMMIT);
 
     if (params.seed == LLAMA_DEFAULT_SEED) {
-        params.seed = time(NULL);
+        params.seed = uint32_t(time(NULL));
     }
     fprintf(stderr, "%s: seed  = %d\n", __func__, params.seed);
 

examples/grammar-parser.cpp

@@ -405,7 +405,7 @@ namespace grammar_parser {
             for (size_t i = 0, end = state.rules.size(); i < end; i++) {
                 // fprintf(file, "%zu: ", i);
                 // print_rule_binary(file, state.rules[i]);
-                print_rule(file, i, state.rules[i], symbol_id_names);
+                print_rule(file, uint32_t(i), state.rules[i], symbol_id_names);
                 // fprintf(file, "\n");
             }
         } catch (const std::exception & err) {

examples/json-schema-to-grammar.py

@@ -0,0 +1,132 @@
+import argparse
+import json
+import re
+import sys
+
+# whitespace is constrained to a single space char to prevent model "running away" in
+# whitespace. Also maybe improves generation quality?
+SPACE_RULE = '" "?'
+
+PRIMITIVE_RULES = {
+    'boolean': '("true" | "false") space',
+    'number': '("-"? ([0-9] | [1-9] [0-9]*)) ("." [0-9]+)? ([eE] [-+]? [0-9]+)? space',
+    'integer': '("-"? ([0-9] | [1-9] [0-9]*)) space',
+    'string': r''' "\"" (
+        [^"\\] |
+        "\\" (["\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F])
+      )* "\"" space ''',
+    'null': '"null" space',
+}
+
+INVALID_RULE_CHARS_RE = re.compile(r'[^a-zA-Z0-9-]+')
+GRAMMAR_LITERAL_ESCAPE_RE = re.compile(r'[\r\n"]')
+GRAMMAR_LITERAL_ESCAPES = {'\r': '\\r', '\n': '\\n', '"': '\\"'}
+
+
+class SchemaConverter:
+    def __init__(self, prop_order):
+        self._prop_order = prop_order
+        self._rules = {'space': SPACE_RULE}
+
+    def _format_literal(self, literal):
+        escaped = GRAMMAR_LITERAL_ESCAPE_RE.sub(
+            lambda m: GRAMMAR_LITERAL_ESCAPES.get(m.group(0)), json.dumps(literal)
+        )
+        return f'"{escaped}"'
+
+    def _add_rule(self, name, rule):
+        esc_name = INVALID_RULE_CHARS_RE.sub('-', name)
+        if esc_name not in self._rules or self._rules[esc_name] == rule:
+            key = esc_name
+        else:
+            i = 0
+            while f'{esc_name}{i}' in self._rules:
+                i += 1
+            key = f'{esc_name}{i}'
+        self._rules[key] = rule
+        return key
+
+    def visit(self, schema, name):
+        schema_type = schema.get('type')
+        rule_name = name or 'root'
+
+        if 'oneOf' in schema or 'anyOf' in schema:
+            rule = ' | '.join((
+                self.visit(alt_schema, f'{name}{"-" if name else ""}{i}')
+                for i, alt_schema in enumerate(schema.get('oneOf') or schema['anyOf'])
+            ))
+            return self._add_rule(rule_name, rule)
+
+        elif 'const' in schema:
+            return self._add_rule(rule_name, self._format_literal(schema['const']))
+
+        elif 'enum' in schema:
+            rule = ' | '.join((self._format_literal(v) for v in schema['enum']))
+            return self._add_rule(rule_name, rule)
+
+        elif schema_type == 'object' and 'properties' in schema:
+            # TODO: `required` keyword
+            prop_order = self._prop_order
+            prop_pairs = sorted(
+                schema['properties'].items(),
+                # sort by position in prop_order (if specified) then by key
+                key=lambda kv: (prop_order.get(kv[0], len(prop_order)), kv[0]),
+            )
+
+            rule = '"{" space'
+            for i, (prop_name, prop_schema) in enumerate(prop_pairs):
+                prop_rule_name = self.visit(prop_schema, f'{name}{"-" if name else ""}{prop_name}')
+                if i > 0:
+                    rule += ' "," space'
+                rule += fr' {self._format_literal(prop_name)} space ":" space {prop_rule_name}'
+            rule += ' "}" space'
+
+            return self._add_rule(rule_name, rule)
+
+        elif schema_type == 'array' and 'items' in schema:
+            # TODO `prefixItems` keyword
+            item_rule_name = self.visit(schema['items'], f'{name}{"-" if name else ""}item')
+            rule = f'"[" space ({item_rule_name} ("," space {item_rule_name})*)? "]" space'
+            return self._add_rule(rule_name, rule)
+
+        else:
+            assert schema_type in PRIMITIVE_RULES, f'Unrecognized schema: {schema}'
+            return self._add_rule(
+                'root' if rule_name == 'root' else schema_type,
+                PRIMITIVE_RULES[schema_type]
+            )
+
+    def format_grammar(self):
+        return '\n'.join((f'{name} ::= {rule}' for name, rule in self._rules.items()))
+
+
+def main(args_in = None):
+    parser = argparse.ArgumentParser(
+        description='''
+            Generates a grammar (suitable for use in ./main) that produces JSON conforming to a
+            given JSON schema. Only a subset of JSON schema features are supported; more may be
+            added in the future.
+        ''',
+    )
+    parser.add_argument(
+        '--prop-order',
+        default=[],
+        type=lambda s: s.split(','),
+        help='''
+            comma-separated property names defining the order of precedence for object properties;
+            properties not specified here are given lower precedence than those that are, and are
+            sorted alphabetically
+        '''
+    )
+    parser.add_argument('schema', help='file containing JSON schema ("-" for stdin)')
+    args = parser.parse_args(args_in)
+
+    schema = json.load(sys.stdin if args.schema == '-' else open(args.schema))
+    prop_order = {name: idx for idx, name in enumerate(args.prop_order)}
+    converter = SchemaConverter(prop_order)
+    converter.visit(schema, '')
+    print(converter.format_grammar())
+
+
+if __name__ == '__main__':
+    main()

examples/llama.vim

@@ -0,0 +1,132 @@
+" Requires an already running llama.cpp server
+" To install either copy or symlink to ~/.vim/autoload/llama.vim
+" Then start with either :call llama#doLlamaGen(),
+" or add a keybind to your vimrc such as
+" nnoremap Z :call llama#doLlamaGen()<CR>
+" Similarly, you could add an insert mode keybind with
+" inoremap <C-B> <Cmd>call llama#doLlamaGen()<CR>
+"
+" g:llama_api_url and g:llama_overrides can be configured in your .vimrc
+" let g:llama_api_url = "192.168.1.10:8080"
+" llama_overrides can also be set through buffer/window scopes. For instance
+" autocmd filetype python let b:llama_overrides = {"temp": 0.2}
+" Could be added to your .vimrc to automatically set a lower temperature when
+" editing a python script
+" Additionally, an override dict can be stored at the top of a file
+" !*{"stop": ["User:"]}
+" Could be added to the start of your chatlog.txt to set the stopping token
+" These parameter dicts are merged together from lowest to highest priority:
+" server default -> g:llama_overrides -> w:llama_overrides ->
+" b:llama_overrides -> in file (!*) overrides
+"
+" Sublists (like logit_bias and stop) are overridden, not merged
+" Example override:
+" !*{"logit_bias": [[13, -5], [2, false]], "temperature": 1, "top_k": 5, "top_p": 0.5, "n_predict": 256, "repeat_last_n": 256, "repeat_penalty": 1.17647}
+if !exists("g:llama_api_url")
+    let g:llama_api_url= "127.0.0.1:8080"
+endif
+if !exists("g:llama_overrides")
+   let g:llama_overrides = {}
+endif
+const s:querydata = {"n_predict": 256, "stop": [ "\n" ], "stream": v:true }
+const s:curlcommand = ['curl','--data-raw', "{\"prompt\":\"### System:\"}", '--silent', '--no-buffer', '--request', 'POST', '--url', g:llama_api_url .. '/completion', '--header', "Content-Type: application/json"]
+let s:linedict = {}
+
+func s:callbackHandler(bufn, channel, msg)
+   if len(a:msg) < 3
+      return
+   elseif a:msg[0] == "d"
+      let l:msg = a:msg[6:-1]
+   else
+      let l:msg = a:msg
+   endif
+   let l:decoded_msg = json_decode(l:msg)
+   let l:newtext = split(l:decoded_msg['content'], "\n", 1)
+   if len(l:newtext) > 0
+      call setbufline(a:bufn, s:linedict[a:bufn], getbufline(a:bufn, s:linedict[a:bufn])[0] .. newtext[0])
+   else
+      echo "nothing genned"
+   endif
+   if len(newtext) > 1
+      let l:failed = appendbufline(a:bufn, s:linedict[a:bufn], newtext[1:-1])
+      let s:linedict[a:bufn] = s:linedict[a:bufn] + len(newtext)-1
+   endif
+   if has_key(l:decoded_msg, "stop") && l:decoded_msg.stop
+       echo "Finished generation"
+   endif
+endfunction
+
+func llama#doLlamaGen()
+   if exists("b:job")
+      if job_status(b:job) == "run"
+         call job_stop(b:job)
+         return
+      endif
+   endif
+
+   let l:cbuffer = bufnr("%")
+   let s:linedict[l:cbuffer] = line('$')
+   let l:buflines = getbufline(l:cbuffer, 1, 1000)
+   let l:querydata = copy(s:querydata)
+   call extend(l:querydata, g:llama_overrides)
+   if exists("w:llama_overrides")
+      call extend(l:querydata, w:llama_overrides)
+   endif
+   if exists("b:llama_overrides")
+      call extend(l:querydata, b:llama_overrides)
+   endif
+   if l:buflines[0][0:1] == '!*'
+      let l:userdata = json_decode(l:buflines[0][2:-1])
+      call extend(l:querydata, l:userdata)
+      let l:buflines = l:buflines[1:-1]
+   endif
+   let l:querydata.prompt = join(l:buflines, "\n")
+   let l:curlcommand = copy(s:curlcommand)
+   let l:curlcommand[2] = json_encode(l:querydata)
+   let b:job = job_start(l:curlcommand, {"callback": function("s:callbackHandler", [l:cbuffer])})
+endfunction
+
+" Echos the tokkenization of the provided string , or cursor to end of word
+" Onus is placed on the user to include the preceding space
+func llama#tokenizeWord(...)
+    if (a:0 > 0)
+        let l:input = a:1
+    else
+        exe "normal \"*ye"
+        let l:input = @*
+    endif
+    let l:querydata = {"content": l:input}
+    let l:curlcommand = copy(s:curlcommand)
+    let l:curlcommand[2] = json_encode(l:querydata)
+    let l:curlcommand[8] = g:llama_api_url .. "/tokenize"
+   let s:token_job = job_start(l:curlcommand, {"callback": function("s:tokenizeWordCallback", [l:input])})
+endfunction
+
+func s:tokenizeWordCallback(plaintext, channel, msg)
+    echo '"' .. a:plaintext ..'" - ' .. string(json_decode(a:msg).tokens)
+endfunction
+
+
+" Echos the token count of the entire buffer (or provided string)
+" Example usage :echo llama#tokenCount()
+func llama#tokenCount(...)
+    if (a:0 > 0)
+        let l:buflines = a:1
+    else
+        let l:buflines = getline(1,1000)
+        if l:buflines[0][0:1] == '!*'
+            let l:buflines = l:buflines[1:-1]
+        endif
+        let l:buflines = join(l:buflines, "\n")
+    endif
+    let l:querydata = {"content": l:buflines}
+    let l:curlcommand = copy(s:curlcommand)
+    let l:curlcommand[2] = json_encode(l:querydata)
+    let l:curlcommand[8] = g:llama_api_url .. "/tokenize"
+   let s:token_job = job_start(l:curlcommand, {"callback": "s:tokenCountCallback"})
+endfunction
+
+func s:tokenCountCallback(channel, msg)
+    let resp = json_decode(a:msg)
+    echo len(resp.tokens)
+endfunction

examples/llm.vim

@@ -1,3 +1,5 @@
+" Basic plugin example
+
 function! Llm()
 
   let url = "http://127.0.0.1:8080/completion"
@@ -16,8 +18,10 @@ function! Llm()
   " Extract the content field from the response
   let content = json_decode(response).content
 
+  let split_newlines = split(content, '\n', 1)
+
   " Insert the content at the cursor position
-  call setline(line('.'), getline('.') . content)
+  call setline(line('.'), [ getline('.') . split_newlines[0] ] + split_newlines[1:])
 endfunction
 
 command! Llm call Llm()

examples/main/README.md

@@ -140,6 +140,12 @@ The `--ctx-size` option allows you to set the size of the prompt context used by
 
 -   `-c N, --ctx-size N`: Set the size of the prompt context (default: 512). The LLaMA models were built with a context of 2048, which will yield the best results on longer input/inference. However, increasing the context size beyond 2048 may lead to unpredictable results.
 
+### Extended Context Size
+
+Some fine-tuned models have extened the context length by scaling RoPE. For example, if the original pretrained model have a context length (max sequence length) of 4096 (4k) and the fine-tuned model have 32k. That is a scaling factor of 8, and should work by setting the above `--ctx-size` to 32768 (32k) and `--rope-scale` to 8.
+
+- `--rope-scale N`: Where N is the linear scaling factor used by the fine-tuned model.
+
 ### Keep Prompt
 
 The `--keep` option allows users to retain the original prompt when the model runs out of context, ensuring a connection to the initial instruction or conversation topic is maintained.
@@ -202,9 +208,9 @@ Example usage: `--top-p 0.95`
 
 -   `--tfs N`: Enable tail free sampling with parameter z (default: 1.0, 1.0 = disabled).
 
-Tail free sampling (TFS) is a text generation technique that aims to reduce the impact of less likely tokens, which may be less relevant, less coherent, or nonsensical, on the output. The method adjusts the logits (token probabilities) by raising them to the power of the parameter z. A higher value of z (e.g., 2.0) will further suppress less likely tokens from the tail of the distribution, while a value of 1.0 disables the effect of TFS. By setting the parameter z, you can control how much the probabilities of less likely tokens are reduced.
+Tail free sampling (TFS) is a text generation technique that aims to reduce the impact of less likely tokens, which may be less relevant, less coherent, or nonsensical, on the output. Similar to Top-P it tries to determine the bulk of the most likely tokens dynamically. But TFS filters out logits based on the second derivative of their probabilities. Adding tokens is stopped after the sum of the second derivatives reaches the parameter z. In short: TFS looks how quickly the probabilities of the tokens decrease and cuts off the tail of unlikely tokens using the parameter z. Typical values for z are in the range of 0.9 to 0.95. A value of 1.0 would include all tokens, and thus disables the effect of TFS.
 
-Example usage: `--tfs 2.0`
+Example usage: `--tfs 0.95`
 
 ### Locally Typical Sampling
 

examples/main/main.cpp

@@ -4,6 +4,7 @@
 #endif
 
 #include "common.h"
+#include "console.h"
 #include "llama.h"
 #include "build-info.h"
 #include "grammar-parser.h"
@@ -35,9 +36,7 @@
 #pragma warning(disable: 4244 4267) // possible loss of data
 #endif
 
-static console_state con_st;
 static llama_context ** g_ctx;
-
 static bool is_interacting = false;
 
 #if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__)) || defined (_WIN32)
@@ -46,7 +45,7 @@ void sigint_handler(int signo) {
         if (!is_interacting) {
             is_interacting=true;
         } else {
-            console_cleanup(con_st);
+            console::cleanup();
             printf("\n");
             llama_print_timings(*g_ctx);
             _exit(130);
@@ -64,10 +63,8 @@ int main(int argc, char ** argv) {
 
     // save choice to use color for later
     // (note for later: this is a slightly awkward choice)
-    con_st.use_color = params.use_color;
-    con_st.multiline_input = params.multiline_input;
-    console_init(con_st);
-    atexit([]() { console_cleanup(con_st); });
+    console::init(params.simple_io, params.use_color);
+    atexit([]() { console::cleanup(); });
 
     if (params.perplexity) {
         printf("\n************\n");
@@ -373,7 +370,7 @@ int main(int argc, char ** argv) {
 
     if (params.interactive) {
         const char *control_message;
-        if (con_st.multiline_input) {
+        if (params.multiline_input) {
             control_message = " - To return control to LLaMa, end your input with '\\'.\n"
                               " - To return control without starting a new line, end your input with '/'.\n";
         } else {
@@ -401,7 +398,7 @@ int main(int argc, char ** argv) {
     int n_past_guidance    = 0;
 
     // the first thing we will do is to output the prompt, so set color accordingly
-    console_set_color(con_st, CONSOLE_COLOR_PROMPT);
+    console::set_display(console::prompt);
 
     std::vector<llama_token> embd;
     std::vector<llama_token> embd_guidance;
@@ -422,9 +419,9 @@ int main(int argc, char ** argv) {
             // Ensure the input doesn't exceed the context size by truncating embd if necessary.
             if ((int)embd.size() > max_embd_size) {
                 auto skipped_tokens = embd.size() - max_embd_size;
-                console_set_color(con_st, CONSOLE_COLOR_ERROR);
+                console::set_display(console::error);
                 printf("<<input too long: skipped %zu token%s>>", skipped_tokens, skipped_tokens != 1 ? "s" : "");
-                console_set_color(con_st, CONSOLE_COLOR_DEFAULT);
+                console::set_display(console::reset);
                 fflush(stdout);
                 embd.resize(max_embd_size);
             }
@@ -667,7 +664,7 @@ int main(int argc, char ** argv) {
         }
         // reset color to default if we there is no pending user input
         if (input_echo && (int)embd_inp.size() == n_consumed) {
-            console_set_color(con_st, CONSOLE_COLOR_DEFAULT);
+            console::set_display(console::reset);
         }
 
         // if not currently processing queued inputs;
@@ -693,7 +690,7 @@ int main(int argc, char ** argv) {
                     if (last_output.find(antiprompt.c_str(), search_start_pos) != std::string::npos) {
                         if (params.interactive) {
                             is_interacting = true;
-                            console_set_color(con_st, CONSOLE_COLOR_USER_INPUT);
+                            console::set_display(console::user_input);
                         }
                         is_antiprompt = true;
                         fflush(stdout);
@@ -714,7 +711,7 @@ int main(int argc, char ** argv) {
 
                     is_interacting = true;
                     printf("\n");
-                    console_set_color(con_st, CONSOLE_COLOR_USER_INPUT);
+                    console::set_display(console::user_input);
                     fflush(stdout);
                 } else if (params.instruct) {
                     is_interacting = true;
@@ -739,12 +736,12 @@ int main(int argc, char ** argv) {
                 std::string line;
                 bool another_line = true;
                 do {
-                    another_line = console_readline(con_st, line);
+                    another_line = console::readline(line, params.multiline_input);
                     buffer += line;
                 } while (another_line);
 
                 // done taking input, reset color
-                console_set_color(con_st, CONSOLE_COLOR_DEFAULT);
+                console::set_display(console::reset);
 
                 // Add tokens to embd only if the input buffer is non-empty
                 // Entering a empty line lets the user pass control back

examples/perplexity/perplexity.cpp

@@ -121,8 +121,23 @@ void perplexity(llama_context * ctx, const gpt_params & params) {
     printf("\n");
 }
 
-void perplexity_lines(llama_context * ctx, const gpt_params & params) {
-    // Calculates perplexity over each line of the prompt
+void hellaswag_score(llama_context * ctx, const gpt_params & params) {
+    // Calculates hellaswag score (acc_norm) from prompt
+    //
+    // Data extracted from the HellaSwag validation dataset (MIT license) https://github.com/rowanz/hellaswag/blob/master/data/hellaswag_val.jsonl
+    // All used data fields are preprocessed as in https://github.com/EleutherAI/lm-evaluation-harness/blob/df3da98c5405deafd519c2ddca52bb7c3fe36bef/lm_eval/tasks/hellaswag.py#L62-L68
+    //
+    // All 10042 tasks should be extracted to keep the results standardized like other implementations.
+    //
+    // Datafile layout:
+    // ['??'] denotes json fields
+    // 6 lines per task:
+    // ['activity_label'] + ": " +['ctx']  - The first part of the query, the context
+    // ['label'] - The index the best common sense ending aka gold ending
+    // ['endings'][0] - Endings added to the first part of the query
+    // ['endings'][1]
+    // ['endings'][2]
+    // ['endings'][3]
 
     std::vector<std::string> prompt_lines;
     std::istringstream strstream(params.prompt);
@@ -132,63 +147,149 @@ void perplexity_lines(llama_context * ctx, const gpt_params & params) {
         prompt_lines.push_back(line);
     }
 
-    const int n_vocab = llama_n_vocab(ctx);
+    if( prompt_lines.size() % 6 != 0) {
+        fprintf(stderr, "%s : number of lines in prompt not a multiple of 6.\n", __func__);
+        return;
+    }
 
-    int counttotal   = 0;
-    size_t n_lines = prompt_lines.size();
+    size_t hs_task_count = prompt_lines.size()/6;
+    fprintf(stderr, "%s : loaded %zu tasks from prompt.\n", __func__, hs_task_count);
 
-    double nll = 0.0;
+    // This is needed as usual for LLaMA models
+    bool prepend_bos = true;
+
+    // Number of tasks to use when computing the score
+    if ( params.hellaswag_tasks < hs_task_count  ) {
+        hs_task_count = params.hellaswag_tasks;
+    }
 
-    fprintf(stderr, "%s: calculating perplexity over %lu lines\n", __func__, n_lines);
+    // The tasks should be randomized so the score stabilizes quickly.
+    bool randomize_tasks = true;
 
-    printf("\nLine\tPPL line\tPPL cumulative\n");
+    // The random seed should not impact the final result if the computation is done over enough tasks, so kept hardcoded for now
+    std::mt19937 rng(1);
 
-    for (size_t i = 0; i < n_lines; ++i) {
+    // Dataholder for hellaswag tasks
+    struct hs_data_t {
+        std::string context;
+        size_t gold_ending_idx;
+        std::string ending[4];
+        size_t ending_logprob_count[4];
+        double ending_logprob[4];
+    };
 
-        // Tokenize and insert BOS at start
-        std::vector<int> batch_embd = ::llama_tokenize(ctx, prompt_lines[i], true);
+    fprintf(stderr, "%s : selecting %zu %s tasks.\n", __func__, hs_task_count, (randomize_tasks?"randomized":"the first")  );
 
-        size_t batch_size  = batch_embd.size();
+    // Select and read data from prompt lines
+    hs_data_t *hs_data = new hs_data_t[hs_task_count];
+    for (size_t i=0; i < hs_task_count; i++) {
+        size_t idx = i;
 
-        // Stop if line is too long
-        if( batch_size > (size_t)params.n_ctx ) {
-            fprintf(stderr, "%s : tokens in line %lu > n_ctxl\n", __func__, i);
-            return;
+        // Select a random example of those left in the prompt
+        if (randomize_tasks) {
+            std::uniform_int_distribution<size_t> dist(0, prompt_lines.size()/6-1 ) ;
+            idx = dist(rng);
         }
 
-        if (llama_eval(ctx, batch_embd.data(), batch_size, 0, params.n_threads)) {
-            fprintf(stderr, "%s : failed to eval\n", __func__);
-            return;
+        hs_data[i].context = prompt_lines[idx*6];
+        hs_data[i].gold_ending_idx = std::stoi( prompt_lines[idx*6+1] );
+        for (size_t j=0; j < 4; j++) {
+            hs_data[i].ending[j] = " " + prompt_lines[idx*6+2+j];
         }
 
-        const auto batch_logits = llama_get_logits(ctx);
-        std::vector<float> logits;
-        logits.insert(logits.end(), batch_logits, batch_logits + batch_size * n_vocab);
+        // Delete the selected random example from the prompt
+        if (randomize_tasks) {
+            prompt_lines.erase( std::next(prompt_lines.begin(),idx*6)  , std::next(prompt_lines.begin(),idx*6+6) );
+        }
+    }
 
-        double nllline = 0.0;
-        int countline = 0;
+    fprintf(stderr, "%s : calculating hellaswag score over selected tasks.\n", __func__);
+    printf("\ntask\tacc_norm\n");
 
-        // Perplexity over second half of the line
-        for (size_t j = batch_size/2; j < batch_size - 1; ++j) {
-            // Calculate probability of next token, given the previous ones.
-            const std::vector<float> tok_logits(
-                logits.begin() + (j + 0) * n_vocab,
-                logits.begin() + (j + 1) * n_vocab);
+    double acc = 0.0f;
+    const int n_vocab = llama_n_vocab(ctx);
+
+    for (size_t task_idx = 0; task_idx < hs_task_count; task_idx++) {
+
+        // Tokenize the context to count tokens
+        std::vector<int> context_embd = ::llama_tokenize(ctx, hs_data[task_idx].context, prepend_bos);
+        size_t context_size = context_embd.size();
+
+        for (size_t ending_idx=0;ending_idx<4;ending_idx++) {
+
+            // Tokenize the query
+            std::vector<int> query_embd = ::llama_tokenize(ctx, hs_data[task_idx].context + hs_data[task_idx].ending[ending_idx], prepend_bos);
+            size_t query_size = query_embd.size();
+
+            // Stop if query wont fit the ctx window
+            if (query_size > (size_t)params.n_ctx) {
+                fprintf(stderr, "%s : number of tokens in query %zu > n_ctxl\n", __func__, query_size);
+                return;
+            }
 
-            const float prob = softmax(tok_logits)[batch_embd[ j + 1]];
+            // Speedup small evaluations by evaluating atleast 32 tokens
+            if (query_size < 32) {
+                query_embd.resize(32);
+            }
+
+            // Evaluate the query
+            if (llama_eval(ctx, query_embd.data(), query_embd.size(), 0, params.n_threads)) {
+                fprintf(stderr, "%s : failed to eval\n", __func__);
+                return;
+            }
+
+            const auto query_logits = llama_get_logits(ctx);
+            std::vector<float> logits;
+            logits.insert(logits.end(), query_logits, query_logits + query_size * n_vocab);
+
+            hs_data[task_idx].ending_logprob_count[ending_idx] = 0;
+            hs_data[task_idx].ending_logprob[ending_idx] = 0.0f;
+
+            // Calculate the logprobs over the ending
+            for (size_t j = context_size-1; j < query_size - 1; j++) {
+                // Calculate probability of next token, given the previous ones.
+                const std::vector<float> tok_logits(
+                    logits.begin() + (j + 0) * n_vocab,
+                    logits.begin() + (j + 1) * n_vocab);
+
+                const float prob = softmax(tok_logits)[query_embd[ j + 1]];
+
+                hs_data[task_idx].ending_logprob[ending_idx] += std::log(prob);
+                hs_data[task_idx].ending_logprob_count[ending_idx]++;
+            }
+
+            // Calculate the mean token logprob for acc_norm
+            hs_data[task_idx].ending_logprob[ending_idx] /= hs_data[task_idx].ending_logprob_count[ending_idx];
+
+
+//            printf("task %lu, ending %lu, whole_len %lu, context_len %lu, ending_logprob_count %lu, ending_logprob %.4f\n",
+//                task_idx,ending_idx,whole_size,context_size, hs_data[task_idx].ending_logprob_count[ending_idx], hs_data[task_idx].ending_logprob[ending_idx] );
+        }
 
-            nllline += -std::log(prob);
-            ++countline;
+        // Find the ending with maximum logprob
+        size_t ending_logprob_max_idx = -1;
+        double ending_logprob_max_val = -INFINITY;
+        for (size_t j=0; j < 4; j++) {
+            if (hs_data[task_idx].ending_logprob[j] > ending_logprob_max_val) {
+                ending_logprob_max_idx = j;
+                ending_logprob_max_val =  hs_data[task_idx].ending_logprob[j];
+            }
         }
 
-        nll += nllline;
-        counttotal += countline;
+//        printf("max logprob ending idx %lu, gold ending idx %lu\n", ending_logprob_max_idx, hs_data[task_idx].gold_ending_idx);
 
-        // perplexity is e^(average negative log-likelihood)
-        printf("%lu\t%.8lf\t%.8lf\n", i + 1, std::exp(nllline/countline), std::exp(nll / counttotal) );
+        // If the gold ending got the maximum logprobe add one accuracy point
+        if (ending_logprob_max_idx == hs_data[task_idx].gold_ending_idx) {
+            acc += 1.0;
+        }
+
+        // Print the accumulated accuracy mean x 100
+        printf("%zu\t%.8lf\n",task_idx+1, acc/double(task_idx+1)*100.0);
         fflush(stdout);
     }
 
+    delete [] hs_data;
+
     printf("\n");
 }
 
@@ -240,8 +341,8 @@ int main(int argc, char ** argv) {
                 params.n_threads, std::thread::hardware_concurrency(), llama_print_system_info());
     }
 
-    if (params.perplexity_lines) {
-        perplexity_lines(ctx, params);
+    if (params.hellaswag) {
+        hellaswag_score(ctx, params);
     } else {
         perplexity(ctx, params);
     }

examples/save-load-state/save-load-state.cpp

@@ -26,6 +26,7 @@ int main(int argc, char ** argv) {
     auto lparams = llama_context_default_params();
 
     lparams.n_ctx     = params.n_ctx;
+    lparams.n_gqa     = params.n_gqa;
     lparams.seed      = params.seed;
     lparams.f16_kv    = params.memory_f16;
     lparams.use_mmap  = params.use_mmap;

examples/server-llama2-13B.sh

@@ -0,0 +1,26 @@
+#!/bin/bash
+
+set -e
+
+cd "$(dirname "$0")/.." || exit
+
+# Specify the model you want to use here:
+MODEL="${MODEL:-./models/llama-2-13b-chat.ggmlv3.q5_K_M.bin}"
+PROMPT_TEMPLATE=${PROMPT_TEMPLATE:-./prompts/chat-system.txt}
+
+# Adjust to the number of CPU cores you want to use.
+N_THREAD="${N_THREAD:-12}"
+
+# Note: you can also override the generation options by specifying them on the command line:
+GEN_OPTIONS="${GEN_OPTIONS:---ctx_size 4096 --batch-size 1024}"
+
+
+# shellcheck disable=SC2086 # Intended splitting of GEN_OPTIONS
+./server $GEN_OPTIONS \
+  --model "$MODEL" \
+  --threads "$N_THREAD" \
+  --rope-freq-scale 1.0 \
+  "$@"
+
+# I used this to test the model with mps, but omitted it from the general purpose. If you want to use it, just specify it on the command line.
+# -ngl 1 \

examples/server/README.md

@@ -151,6 +151,8 @@ node .
 
     `mirostat_eta`: Set the Mirostat learning rate, parameter eta (default: 0.1).
 
+    `grammar`: Set grammar for grammar-based sampling (default: no grammar)
+
     `seed`: Set the random number generator (RNG) seed (default: -1, -1 = random seed).
 
     `ignore_eos`: Ignore end of stream token and continue generating (default: false).
@@ -163,7 +165,7 @@ node .
 
     `content`: Set the text to tokenize.
 
-    Note that the special `BOS` token is not added in fron of the text and also a space character is not inserted automatically as it is for `/completion`.
+    Note that the special `BOS` token is not added in front of the text and also a space character is not inserted automatically as it is for `/completion`.
 
 -   **POST** `/embedding`: Generate embedding of a given text just as [the embedding example](../embedding) does.
 

examples/server/chat-llama2.sh

@@ -0,0 +1,109 @@
+#!/bin/bash
+
+API_URL="${API_URL:-http://127.0.0.1:8080}"
+
+CHAT=(
+    "Hello, Assistant."
+    "Hello. How may I help you today?"
+)
+
+INSTRUCTION="A chat between a curious human and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the human's questions."
+
+trim() {
+    shopt -s extglob
+    set -- "${1##+([[:space:]])}"
+    printf "%s" "${1%%+([[:space:]])}"
+}
+
+trim_trailing() {
+    shopt -s extglob
+    printf "%s" "${1%%+([[:space:]])}"
+}
+
+format_prompt() {
+    if [[ "${#CHAT[@]}" -eq 0 ]]; then
+        echo -n "[INST] <<SYS>>\n${INSTRUCTION}\n<</SYS>>"
+    else
+        LAST_INDEX=$(( ${#CHAT[@]} - 1 ))
+        echo -n "${CHAT[$LAST_INDEX]}\n[INST] $1 [/INST]"
+    fi
+}
+
+tokenize() {
+    curl \
+        --silent \
+        --request POST \
+        --url "${API_URL}/tokenize" \
+        --header "Content-Type: application/json" \
+        --data-raw "$(jq -ns --arg content "$1" '{content:$content}')" \
+    | jq '.tokens[]'
+}
+
+N_KEEP=$(tokenize "[INST] <<SYS>>\n${INSTRUCTION}\n<</SYS>>" | wc -l)
+
+chat_completion() {
+    PROMPT="$(trim_trailing "$(format_prompt "$1")")"
+    DATA="$(echo -n "$PROMPT" | jq -Rs --argjson n_keep $N_KEEP '{
+        prompt: .,
+        temperature: 0.2,
+        top_k: 40,
+        top_p: 0.9,
+        n_keep: $n_keep,
+        n_predict: 1024,
+        stop: ["[INST]"],
+        stream: true
+    }')"
+
+    # Create a temporary file to hold the Python output
+    TEMPFILE=$(mktemp)
+
+    exec 3< <(curl \
+        --silent \
+        --no-buffer \
+        --request POST \
+        --url "${API_URL}/completion" \
+        --header "Content-Type: application/json" \
+        --data-raw "${DATA}")
+
+    python -c "
+import json
+import sys
+
+answer = ''
+while True:
+    line = sys.stdin.readline()
+    if not line:
+        break
+    if line.startswith('data: '):
+        json_content = line[6:].strip()
+        content = json.loads(json_content)['content']
+        sys.stdout.write(content)
+        sys.stdout.flush()
+        answer += content
+
+answer = answer.rstrip('\n')
+
+# Write the answer to the temporary file
+with open('$TEMPFILE', 'w') as f:
+    f.write(answer)
+    " <&3
+
+    exec 3<&-
+
+    # Read the answer from the temporary file
+    ANSWER=$(cat $TEMPFILE)
+
+    # Clean up the temporary file
+    rm $TEMPFILE
+
+    printf "\n"
+
+    CHAT+=("$1" "$(trim "$ANSWER")")
+}
+
+while true; do
+    echo -en "\033[0;32m"  # Green color
+    read -r -e -p "> " QUESTION
+    echo -en "\033[0m"  # Reset color
+    chat_completion "${QUESTION}"
+done

examples/server/completion.js.hpp

@@ -87,289 +87,342 @@ unsigned char completion_js[] = {
   0x20, 0x54, 0x65, 0x78, 0x74, 0x44, 0x65, 0x63, 0x6f, 0x64, 0x65, 0x72,
   0x28, 0x29, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x6c, 0x65, 0x74, 0x20, 0x63,
   0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x20, 0x3d, 0x20, 0x22, 0x22, 0x3b,
-  0x0a, 0x0a, 0x20, 0x20, 0x74, 0x72, 0x79, 0x20, 0x7b, 0x0a, 0x20, 0x20,
-  0x20, 0x20, 0x6c, 0x65, 0x74, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x20, 0x3d,
-  0x20, 0x74, 0x72, 0x75, 0x65, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20,
-  0x77, 0x68, 0x69, 0x6c, 0x65, 0x20, 0x28, 0x63, 0x6f, 0x6e, 0x74, 0x29,
-  0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e,
-  0x73, 0x74, 0x20, 0x72, 0x65, 0x73, 0x75, 0x6c, 0x74, 0x20, 0x3d, 0x20,
-  0x61, 0x77, 0x61, 0x69, 0x74, 0x20, 0x72, 0x65, 0x61, 0x64, 0x65, 0x72,
-  0x2e, 0x72, 0x65, 0x61, 0x64, 0x28, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x72, 0x65, 0x73, 0x75, 0x6c,
-  0x74, 0x2e, 0x64, 0x6f, 0x6e, 0x65, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x62, 0x72, 0x65, 0x61, 0x6b, 0x3b,
-  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x73, 0x73, 0x65, 0x20, 0x61,
-  0x6e, 0x73, 0x77, 0x65, 0x72, 0x73, 0x20, 0x69, 0x6e, 0x20, 0x74, 0x68,
-  0x65, 0x20, 0x66, 0x6f, 0x72, 0x6d, 0x20, 0x6d, 0x75, 0x6c, 0x74, 0x69,
-  0x70, 0x6c, 0x65, 0x20, 0x6c, 0x69, 0x6e, 0x65, 0x73, 0x20, 0x6f, 0x66,
-  0x3a, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x5c, 0x6e, 0x20, 0x77, 0x69,
-  0x74, 0x68, 0x20, 0x64, 0x61, 0x74, 0x61, 0x20, 0x61, 0x6c, 0x77, 0x61,
-  0x79, 0x73, 0x20, 0x70, 0x72, 0x65, 0x73, 0x65, 0x6e, 0x74, 0x20, 0x61,
-  0x73, 0x20, 0x61, 0x20, 0x6b, 0x65, 0x79, 0x2e, 0x20, 0x69, 0x6e, 0x20,
-  0x6f, 0x75, 0x72, 0x20, 0x63, 0x61, 0x73, 0x65, 0x20, 0x77, 0x65, 0x0a,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x6d, 0x61, 0x69,
-  0x6e, 0x6c, 0x79, 0x20, 0x63, 0x61, 0x72, 0x65, 0x20, 0x61, 0x62, 0x6f,
-  0x75, 0x74, 0x20, 0x74, 0x68, 0x65, 0x20, 0x64, 0x61, 0x74, 0x61, 0x3a,
-  0x20, 0x6b, 0x65, 0x79, 0x20, 0x68, 0x65, 0x72, 0x65, 0x2c, 0x20, 0x77,
-  0x68, 0x69, 0x63, 0x68, 0x20, 0x77, 0x65, 0x20, 0x65, 0x78, 0x70, 0x65,
-  0x63, 0x74, 0x20, 0x61, 0x73, 0x20, 0x6a, 0x73, 0x6f, 0x6e, 0x0a, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x74,
-  0x65, 0x78, 0x74, 0x20, 0x3d, 0x20, 0x64, 0x65, 0x63, 0x6f, 0x64, 0x65,
-  0x72, 0x2e, 0x64, 0x65, 0x63, 0x6f, 0x64, 0x65, 0x28, 0x72, 0x65, 0x73,
-  0x75, 0x6c, 0x74, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x29, 0x3b, 0x0a,
-  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x70, 0x61,
-  0x72, 0x73, 0x65, 0x20, 0x61, 0x6c, 0x6c, 0x20, 0x73, 0x73, 0x65, 0x20,
-  0x65, 0x76, 0x65, 0x6e, 0x74, 0x73, 0x20, 0x61, 0x6e, 0x64, 0x20, 0x61,
-  0x64, 0x64, 0x20, 0x74, 0x68, 0x65, 0x6d, 0x20, 0x74, 0x6f, 0x20, 0x72,
-  0x65, 0x73, 0x75, 0x6c, 0x74, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x72, 0x65, 0x67, 0x65, 0x78, 0x20,
-  0x3d, 0x20, 0x2f, 0x5e, 0x28, 0x5c, 0x53, 0x2b, 0x29, 0x3a, 0x5c, 0x73,
-  0x28, 0x2e, 0x2a, 0x29, 0x24, 0x2f, 0x67, 0x6d, 0x3b, 0x0a, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x66, 0x6f, 0x72, 0x20, 0x28, 0x63, 0x6f, 0x6e,
-  0x73, 0x74, 0x20, 0x6d, 0x61, 0x74, 0x63, 0x68, 0x20, 0x6f, 0x66, 0x20,
-  0x74, 0x65, 0x78, 0x74, 0x2e, 0x6d, 0x61, 0x74, 0x63, 0x68, 0x41, 0x6c,
-  0x6c, 0x28, 0x72, 0x65, 0x67, 0x65, 0x78, 0x29, 0x29, 0x20, 0x7b, 0x0a,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x73, 0x75,
-  0x6c, 0x74, 0x5b, 0x6d, 0x61, 0x74, 0x63, 0x68, 0x5b, 0x31, 0x5d, 0x5d,
-  0x20, 0x3d, 0x20, 0x6d, 0x61, 0x74, 0x63, 0x68, 0x5b, 0x32, 0x5d, 0x0a,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x73, 0x69, 0x6e, 0x63, 0x65, 0x20,
-  0x77, 0x65, 0x20, 0x6b, 0x6e, 0x6f, 0x77, 0x20, 0x74, 0x68, 0x69, 0x73,
-  0x20, 0x69, 0x73, 0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x2e, 0x63, 0x70,
-  0x70, 0x2c, 0x20, 0x6c, 0x65, 0x74, 0x27, 0x73, 0x20, 0x6a, 0x75, 0x73,
-  0x74, 0x20, 0x64, 0x65, 0x63, 0x6f, 0x64, 0x65, 0x20, 0x74, 0x68, 0x65,
-  0x20, 0x6a, 0x73, 0x6f, 0x6e, 0x20, 0x69, 0x6e, 0x20, 0x64, 0x61, 0x74,
-  0x61, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x73, 0x75,
-  0x6c, 0x74, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x20, 0x3d, 0x20, 0x4a, 0x53,
-  0x4f, 0x4e, 0x2e, 0x70, 0x61, 0x72, 0x73, 0x65, 0x28, 0x72, 0x65, 0x73,
-  0x75, 0x6c, 0x74, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x29, 0x3b, 0x0a, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74,
-  0x20, 0x2b, 0x3d, 0x20, 0x72, 0x65, 0x73, 0x75, 0x6c, 0x74, 0x2e, 0x64,
-  0x61, 0x74, 0x61, 0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x3b,
-  0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x79,
-  0x69, 0x65, 0x6c, 0x64, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x79,
-  0x69, 0x65, 0x6c, 0x64, 0x20, 0x72, 0x65, 0x73, 0x75, 0x6c, 0x74, 0x3b,
-  0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x69,
-  0x66, 0x20, 0x77, 0x65, 0x20, 0x67, 0x6f, 0x74, 0x20, 0x61, 0x20, 0x73,
-  0x74, 0x6f, 0x70, 0x20, 0x74, 0x6f, 0x6b, 0x65, 0x6e, 0x20, 0x66, 0x72,
-  0x6f, 0x6d, 0x20, 0x73, 0x65, 0x72, 0x76, 0x65, 0x72, 0x2c, 0x20, 0x77,
-  0x65, 0x20, 0x77, 0x69, 0x6c, 0x6c, 0x20, 0x62, 0x72, 0x65, 0x61, 0x6b,
-  0x20, 0x68, 0x65, 0x72, 0x65, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x69, 0x66, 0x20, 0x28, 0x72, 0x65, 0x73, 0x75, 0x6c, 0x74, 0x2e, 0x64,
-  0x61, 0x74, 0x61, 0x2e, 0x73, 0x74, 0x6f, 0x70, 0x29, 0x20, 0x7b, 0x0a,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28,
-  0x72, 0x65, 0x73, 0x75, 0x6c, 0x74, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x2e,
-  0x67, 0x65, 0x6e, 0x65, 0x72, 0x61, 0x74, 0x69, 0x6f, 0x6e, 0x5f, 0x73,
-  0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x29, 0x20, 0x7b, 0x0a, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x67, 0x65, 0x6e,
-  0x65, 0x72, 0x61, 0x74, 0x69, 0x6f, 0x6e, 0x5f, 0x73, 0x65, 0x74, 0x74,
-  0x69, 0x6e, 0x67, 0x73, 0x20, 0x3d, 0x20, 0x72, 0x65, 0x73, 0x75, 0x6c,
-  0x74, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x2e, 0x67, 0x65, 0x6e, 0x65, 0x72,
-  0x61, 0x74, 0x69, 0x6f, 0x6e, 0x5f, 0x73, 0x65, 0x74, 0x74, 0x69, 0x6e,
-  0x67, 0x73, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x62, 0x72,
-  0x65, 0x61, 0x6b, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d,
-  0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x7d, 0x20, 0x63,
-  0x61, 0x74, 0x63, 0x68, 0x20, 0x28, 0x65, 0x29, 0x20, 0x7b, 0x0a, 0x20,
-  0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x65, 0x2e, 0x6e, 0x61, 0x6d,
-  0x65, 0x20, 0x21, 0x3d, 0x3d, 0x20, 0x27, 0x41, 0x62, 0x6f, 0x72, 0x74,
-  0x45, 0x72, 0x72, 0x6f, 0x72, 0x27, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x6f, 0x6c, 0x65, 0x2e,
-  0x65, 0x72, 0x72, 0x6f, 0x72, 0x28, 0x22, 0x6c, 0x6c, 0x61, 0x6d, 0x61,
-  0x20, 0x65, 0x72, 0x72, 0x6f, 0x72, 0x3a, 0x20, 0x22, 0x2c, 0x20, 0x65,
-  0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20,
-  0x20, 0x74, 0x68, 0x72, 0x6f, 0x77, 0x20, 0x65, 0x3b, 0x0a, 0x20, 0x20,
-  0x7d, 0x0a, 0x20, 0x20, 0x66, 0x69, 0x6e, 0x61, 0x6c, 0x6c, 0x79, 0x20,
-  0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x72, 0x6f,
-  0x6c, 0x6c, 0x65, 0x72, 0x2e, 0x61, 0x62, 0x6f, 0x72, 0x74, 0x28, 0x29,
-  0x3b, 0x0a, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x72, 0x65, 0x74,
-  0x75, 0x72, 0x6e, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x3b,
-  0x0a, 0x7d, 0x0a, 0x0a, 0x2f, 0x2f, 0x20, 0x43, 0x61, 0x6c, 0x6c, 0x20,
-  0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x2c, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72,
-  0x6e, 0x20, 0x61, 0x6e, 0x20, 0x65, 0x76, 0x65, 0x6e, 0x74, 0x20, 0x74,
-  0x61, 0x72, 0x67, 0x65, 0x74, 0x20, 0x74, 0x68, 0x61, 0x74, 0x20, 0x79,
-  0x6f, 0x75, 0x20, 0x63, 0x61, 0x6e, 0x20, 0x73, 0x75, 0x62, 0x63, 0x72,
-  0x69, 0x62, 0x65, 0x20, 0x74, 0x6f, 0x0a, 0x2f, 0x2f, 0x0a, 0x2f, 0x2f,
-  0x20, 0x45, 0x78, 0x61, 0x6d, 0x70, 0x6c, 0x65, 0x3a, 0x0a, 0x2f, 0x2f,
-  0x0a, 0x2f, 0x2f, 0x20, 0x20, 0x20, 0x20, 0x69, 0x6d, 0x70, 0x6f, 0x72,
-  0x74, 0x20, 0x7b, 0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x45, 0x76, 0x65,
-  0x6e, 0x74, 0x54, 0x61, 0x72, 0x67, 0x65, 0x74, 0x20, 0x7d, 0x20, 0x66,
-  0x72, 0x6f, 0x6d, 0x20, 0x27, 0x2f, 0x63, 0x6f, 0x6d, 0x70, 0x6c, 0x65,
-  0x74, 0x69, 0x6f, 0x6e, 0x2e, 0x6a, 0x73, 0x27, 0x0a, 0x2f, 0x2f, 0x0a,
-  0x2f, 0x2f, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20,
-  0x63, 0x6f, 0x6e, 0x6e, 0x20, 0x3d, 0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61,
-  0x45, 0x76, 0x65, 0x6e, 0x74, 0x54, 0x61, 0x72, 0x67, 0x65, 0x74, 0x28,
-  0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x29, 0x0a, 0x2f, 0x2f, 0x20, 0x20,
-  0x20, 0x20, 0x63, 0x6f, 0x6e, 0x6e, 0x2e, 0x61, 0x64, 0x64, 0x45, 0x76,
-  0x65, 0x6e, 0x74, 0x4c, 0x69, 0x73, 0x74, 0x65, 0x6e, 0x65, 0x72, 0x28,
-  0x22, 0x6d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x22, 0x2c, 0x20, 0x28,
-  0x63, 0x68, 0x75, 0x6e, 0x6b, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a,
-  0x2f, 0x2f, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x64, 0x6f, 0x63, 0x75,
-  0x6d, 0x65, 0x6e, 0x74, 0x2e, 0x77, 0x72, 0x69, 0x74, 0x65, 0x28, 0x63,
-  0x68, 0x75, 0x6e, 0x6b, 0x2e, 0x64, 0x65, 0x74, 0x61, 0x69, 0x6c, 0x2e,
-  0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x29, 0x0a, 0x2f, 0x2f, 0x20,
-  0x20, 0x20, 0x20, 0x7d, 0x29, 0x0a, 0x2f, 0x2f, 0x0a, 0x65, 0x78, 0x70,
-  0x6f, 0x72, 0x74, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x6c, 0x6c,
-  0x61, 0x6d, 0x61, 0x45, 0x76, 0x65, 0x6e, 0x74, 0x54, 0x61, 0x72, 0x67,
-  0x65, 0x74, 0x20, 0x3d, 0x20, 0x28, 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74,
-  0x2c, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x20, 0x3d, 0x20, 0x7b,
-  0x7d, 0x2c, 0x20, 0x63, 0x6f, 0x6e, 0x66, 0x69, 0x67, 0x20, 0x3d, 0x20,
-  0x7b, 0x7d, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x63,
-  0x6f, 0x6e, 0x73, 0x74, 0x20, 0x65, 0x76, 0x65, 0x6e, 0x74, 0x54, 0x61,
-  0x72, 0x67, 0x65, 0x74, 0x20, 0x3d, 0x20, 0x6e, 0x65, 0x77, 0x20, 0x45,
-  0x76, 0x65, 0x6e, 0x74, 0x54, 0x61, 0x72, 0x67, 0x65, 0x74, 0x28, 0x29,
-  0x3b, 0x0a, 0x20, 0x20, 0x28, 0x61, 0x73, 0x79, 0x6e, 0x63, 0x20, 0x28,
-  0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x6c,
-  0x65, 0x74, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x20, 0x3d,
-  0x20, 0x22, 0x22, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x66, 0x6f, 0x72,
-  0x20, 0x61, 0x77, 0x61, 0x69, 0x74, 0x20, 0x28, 0x63, 0x6f, 0x6e, 0x73,
-  0x74, 0x20, 0x63, 0x68, 0x75, 0x6e, 0x6b, 0x20, 0x6f, 0x66, 0x20, 0x6c,
-  0x6c, 0x61, 0x6d, 0x61, 0x28, 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x2c,
-  0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2c, 0x20, 0x63, 0x6f, 0x6e,
-  0x66, 0x69, 0x67, 0x29, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x63, 0x68, 0x75, 0x6e, 0x6b, 0x2e,
-  0x64, 0x61, 0x74, 0x61, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x20,
-  0x2b, 0x3d, 0x20, 0x63, 0x68, 0x75, 0x6e, 0x6b, 0x2e, 0x64, 0x61, 0x74,
-  0x61, 0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x3b, 0x0a, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x65, 0x76, 0x65, 0x6e, 0x74,
-  0x54, 0x61, 0x72, 0x67, 0x65, 0x74, 0x2e, 0x64, 0x69, 0x73, 0x70, 0x61,
-  0x74, 0x63, 0x68, 0x45, 0x76, 0x65, 0x6e, 0x74, 0x28, 0x6e, 0x65, 0x77,
-  0x20, 0x43, 0x75, 0x73, 0x74, 0x6f, 0x6d, 0x45, 0x76, 0x65, 0x6e, 0x74,
-  0x28, 0x22, 0x6d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x22, 0x2c, 0x20,
-  0x7b, 0x20, 0x64, 0x65, 0x74, 0x61, 0x69, 0x6c, 0x3a, 0x20, 0x63, 0x68,
-  0x75, 0x6e, 0x6b, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x20, 0x7d, 0x29, 0x29,
-  0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x63, 0x68, 0x75, 0x6e,
-  0x6b, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x2e, 0x67, 0x65, 0x6e, 0x65, 0x72,
-  0x61, 0x74, 0x69, 0x6f, 0x6e, 0x5f, 0x73, 0x65, 0x74, 0x74, 0x69, 0x6e,
-  0x67, 0x73, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x65, 0x76, 0x65, 0x6e, 0x74, 0x54, 0x61, 0x72, 0x67, 0x65,
-  0x74, 0x2e, 0x64, 0x69, 0x73, 0x70, 0x61, 0x74, 0x63, 0x68, 0x45, 0x76,
-  0x65, 0x6e, 0x74, 0x28, 0x6e, 0x65, 0x77, 0x20, 0x43, 0x75, 0x73, 0x74,
-  0x6f, 0x6d, 0x45, 0x76, 0x65, 0x6e, 0x74, 0x28, 0x22, 0x67, 0x65, 0x6e,
-  0x65, 0x72, 0x61, 0x74, 0x69, 0x6f, 0x6e, 0x5f, 0x73, 0x65, 0x74, 0x74,
-  0x69, 0x6e, 0x67, 0x73, 0x22, 0x2c, 0x20, 0x7b, 0x20, 0x64, 0x65, 0x74,
-  0x61, 0x69, 0x6c, 0x3a, 0x20, 0x63, 0x68, 0x75, 0x6e, 0x6b, 0x2e, 0x64,
+  0x0a, 0x20, 0x20, 0x6c, 0x65, 0x74, 0x20, 0x6c, 0x65, 0x66, 0x74, 0x6f,
+  0x76, 0x65, 0x72, 0x20, 0x3d, 0x20, 0x22, 0x22, 0x3b, 0x20, 0x2f, 0x2f,
+  0x20, 0x42, 0x75, 0x66, 0x66, 0x65, 0x72, 0x20, 0x66, 0x6f, 0x72, 0x20,
+  0x70, 0x61, 0x72, 0x74, 0x69, 0x61, 0x6c, 0x6c, 0x79, 0x20, 0x72, 0x65,
+  0x61, 0x64, 0x20, 0x6c, 0x69, 0x6e, 0x65, 0x73, 0x0a, 0x0a, 0x20, 0x20,
+  0x74, 0x72, 0x79, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x65,
+  0x74, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x20, 0x3d, 0x20, 0x74, 0x72, 0x75,
+  0x65, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x77, 0x68, 0x69, 0x6c,
+  0x65, 0x20, 0x28, 0x63, 0x6f, 0x6e, 0x74, 0x29, 0x20, 0x7b, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x72,
+  0x65, 0x73, 0x75, 0x6c, 0x74, 0x20, 0x3d, 0x20, 0x61, 0x77, 0x61, 0x69,
+  0x74, 0x20, 0x72, 0x65, 0x61, 0x64, 0x65, 0x72, 0x2e, 0x72, 0x65, 0x61,
+  0x64, 0x28, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69,
+  0x66, 0x20, 0x28, 0x72, 0x65, 0x73, 0x75, 0x6c, 0x74, 0x2e, 0x64, 0x6f,
+  0x6e, 0x65, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x62, 0x72, 0x65, 0x61, 0x6b, 0x3b, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x2f, 0x2f, 0x20, 0x41, 0x64, 0x64, 0x20, 0x61, 0x6e, 0x79, 0x20, 0x6c,
+  0x65, 0x66, 0x74, 0x6f, 0x76, 0x65, 0x72, 0x20, 0x64, 0x61, 0x74, 0x61,
+  0x20, 0x74, 0x6f, 0x20, 0x74, 0x68, 0x65, 0x20, 0x63, 0x75, 0x72, 0x72,
+  0x65, 0x6e, 0x74, 0x20, 0x63, 0x68, 0x75, 0x6e, 0x6b, 0x20, 0x6f, 0x66,
+  0x20, 0x64, 0x61, 0x74, 0x61, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x74, 0x65, 0x78, 0x74, 0x20, 0x3d,
+  0x20, 0x6c, 0x65, 0x66, 0x74, 0x6f, 0x76, 0x65, 0x72, 0x20, 0x2b, 0x20,
+  0x64, 0x65, 0x63, 0x6f, 0x64, 0x65, 0x72, 0x2e, 0x64, 0x65, 0x63, 0x6f,
+  0x64, 0x65, 0x28, 0x72, 0x65, 0x73, 0x75, 0x6c, 0x74, 0x2e, 0x76, 0x61,
+  0x6c, 0x75, 0x65, 0x29, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x2f, 0x2f, 0x20, 0x43, 0x68, 0x65, 0x63, 0x6b, 0x20, 0x69, 0x66,
+  0x20, 0x74, 0x68, 0x65, 0x20, 0x6c, 0x61, 0x73, 0x74, 0x20, 0x63, 0x68,
+  0x61, 0x72, 0x61, 0x63, 0x74, 0x65, 0x72, 0x20, 0x69, 0x73, 0x20, 0x61,
+  0x20, 0x6c, 0x69, 0x6e, 0x65, 0x20, 0x62, 0x72, 0x65, 0x61, 0x6b, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20,
+  0x65, 0x6e, 0x64, 0x73, 0x57, 0x69, 0x74, 0x68, 0x4c, 0x69, 0x6e, 0x65,
+  0x42, 0x72, 0x65, 0x61, 0x6b, 0x20, 0x3d, 0x20, 0x74, 0x65, 0x78, 0x74,
+  0x2e, 0x65, 0x6e, 0x64, 0x73, 0x57, 0x69, 0x74, 0x68, 0x28, 0x27, 0x5c,
+  0x6e, 0x27, 0x29, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x2f, 0x2f, 0x20, 0x53, 0x70, 0x6c, 0x69, 0x74, 0x20, 0x74, 0x68, 0x65,
+  0x20, 0x74, 0x65, 0x78, 0x74, 0x20, 0x69, 0x6e, 0x74, 0x6f, 0x20, 0x6c,
+  0x69, 0x6e, 0x65, 0x73, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6c,
+  0x65, 0x74, 0x20, 0x6c, 0x69, 0x6e, 0x65, 0x73, 0x20, 0x3d, 0x20, 0x74,
+  0x65, 0x78, 0x74, 0x2e, 0x73, 0x70, 0x6c, 0x69, 0x74, 0x28, 0x27, 0x5c,
+  0x6e, 0x27, 0x29, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x2f, 0x2f, 0x20, 0x49, 0x66, 0x20, 0x74, 0x68, 0x65, 0x20, 0x74, 0x65,
+  0x78, 0x74, 0x20, 0x64, 0x6f, 0x65, 0x73, 0x6e, 0x27, 0x74, 0x20, 0x65,
+  0x6e, 0x64, 0x20, 0x77, 0x69, 0x74, 0x68, 0x20, 0x61, 0x20, 0x6c, 0x69,
+  0x6e, 0x65, 0x20, 0x62, 0x72, 0x65, 0x61, 0x6b, 0x2c, 0x20, 0x74, 0x68,
+  0x65, 0x6e, 0x20, 0x74, 0x68, 0x65, 0x20, 0x6c, 0x61, 0x73, 0x74, 0x20,
+  0x6c, 0x69, 0x6e, 0x65, 0x20, 0x69, 0x73, 0x20, 0x69, 0x6e, 0x63, 0x6f,
+  0x6d, 0x70, 0x6c, 0x65, 0x74, 0x65, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x2f, 0x2f, 0x20, 0x53, 0x74, 0x6f, 0x72, 0x65, 0x20, 0x69, 0x74,
+  0x20, 0x69, 0x6e, 0x20, 0x6c, 0x65, 0x66, 0x74, 0x6f, 0x76, 0x65, 0x72,
+  0x20, 0x74, 0x6f, 0x20, 0x62, 0x65, 0x20, 0x61, 0x64, 0x64, 0x65, 0x64,
+  0x20, 0x74, 0x6f, 0x20, 0x74, 0x68, 0x65, 0x20, 0x6e, 0x65, 0x78, 0x74,
+  0x20, 0x63, 0x68, 0x75, 0x6e, 0x6b, 0x20, 0x6f, 0x66, 0x20, 0x64, 0x61,
+  0x74, 0x61, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20,
+  0x28, 0x21, 0x65, 0x6e, 0x64, 0x73, 0x57, 0x69, 0x74, 0x68, 0x4c, 0x69,
+  0x6e, 0x65, 0x42, 0x72, 0x65, 0x61, 0x6b, 0x29, 0x20, 0x7b, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x65, 0x66, 0x74, 0x6f,
+  0x76, 0x65, 0x72, 0x20, 0x3d, 0x20, 0x6c, 0x69, 0x6e, 0x65, 0x73, 0x2e,
+  0x70, 0x6f, 0x70, 0x28, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x7d, 0x20, 0x65, 0x6c, 0x73, 0x65, 0x20, 0x7b, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x65, 0x66, 0x74, 0x6f, 0x76,
+  0x65, 0x72, 0x20, 0x3d, 0x20, 0x22, 0x22, 0x3b, 0x20, 0x2f, 0x2f, 0x20,
+  0x52, 0x65, 0x73, 0x65, 0x74, 0x20, 0x6c, 0x65, 0x66, 0x74, 0x6f, 0x76,
+  0x65, 0x72, 0x20, 0x69, 0x66, 0x20, 0x77, 0x65, 0x20, 0x68, 0x61, 0x76,
+  0x65, 0x20, 0x61, 0x20, 0x6c, 0x69, 0x6e, 0x65, 0x20, 0x62, 0x72, 0x65,
+  0x61, 0x6b, 0x20, 0x61, 0x74, 0x20, 0x74, 0x68, 0x65, 0x20, 0x65, 0x6e,
+  0x64, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x50, 0x61, 0x72, 0x73,
+  0x65, 0x20, 0x61, 0x6c, 0x6c, 0x20, 0x73, 0x73, 0x65, 0x20, 0x65, 0x76,
+  0x65, 0x6e, 0x74, 0x73, 0x20, 0x61, 0x6e, 0x64, 0x20, 0x61, 0x64, 0x64,
+  0x20, 0x74, 0x68, 0x65, 0x6d, 0x20, 0x74, 0x6f, 0x20, 0x72, 0x65, 0x73,
+  0x75, 0x6c, 0x74, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f,
+  0x6e, 0x73, 0x74, 0x20, 0x72, 0x65, 0x67, 0x65, 0x78, 0x20, 0x3d, 0x20,
+  0x2f, 0x5e, 0x28, 0x5c, 0x53, 0x2b, 0x29, 0x3a, 0x5c, 0x73, 0x28, 0x2e,
+  0x2a, 0x29, 0x24, 0x2f, 0x67, 0x6d, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x66, 0x6f, 0x72, 0x20, 0x28, 0x63, 0x6f, 0x6e, 0x73, 0x74,
+  0x20, 0x6c, 0x69, 0x6e, 0x65, 0x20, 0x6f, 0x66, 0x20, 0x6c, 0x69, 0x6e,
+  0x65, 0x73, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x6d, 0x61, 0x74, 0x63,
+  0x68, 0x20, 0x3d, 0x20, 0x72, 0x65, 0x67, 0x65, 0x78, 0x2e, 0x65, 0x78,
+  0x65, 0x63, 0x28, 0x6c, 0x69, 0x6e, 0x65, 0x29, 0x3b, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x6d, 0x61,
+  0x74, 0x63, 0x68, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x73, 0x75, 0x6c, 0x74, 0x5b,
+  0x6d, 0x61, 0x74, 0x63, 0x68, 0x5b, 0x31, 0x5d, 0x5d, 0x20, 0x3d, 0x20,
+  0x6d, 0x61, 0x74, 0x63, 0x68, 0x5b, 0x32, 0x5d, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x73, 0x69,
+  0x6e, 0x63, 0x65, 0x20, 0x77, 0x65, 0x20, 0x6b, 0x6e, 0x6f, 0x77, 0x20,
+  0x74, 0x68, 0x69, 0x73, 0x20, 0x69, 0x73, 0x20, 0x6c, 0x6c, 0x61, 0x6d,
+  0x61, 0x2e, 0x63, 0x70, 0x70, 0x2c, 0x20, 0x6c, 0x65, 0x74, 0x27, 0x73,
+  0x20, 0x6a, 0x75, 0x73, 0x74, 0x20, 0x64, 0x65, 0x63, 0x6f, 0x64, 0x65,
+  0x20, 0x74, 0x68, 0x65, 0x20, 0x6a, 0x73, 0x6f, 0x6e, 0x20, 0x69, 0x6e,
+  0x20, 0x64, 0x61, 0x74, 0x61, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x72, 0x65, 0x73, 0x75,
+  0x6c, 0x74, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x29, 0x20, 0x7b, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72,
+  0x65, 0x73, 0x75, 0x6c, 0x74, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x20, 0x3d,
+  0x20, 0x4a, 0x53, 0x4f, 0x4e, 0x2e, 0x70, 0x61, 0x72, 0x73, 0x65, 0x28,
+  0x72, 0x65, 0x73, 0x75, 0x6c, 0x74, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x29,
+  0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x20, 0x2b, 0x3d,
+  0x20, 0x72, 0x65, 0x73, 0x75, 0x6c, 0x74, 0x2e, 0x64, 0x61, 0x74, 0x61,
+  0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x3b, 0x0a, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f,
+  0x2f, 0x20, 0x79, 0x69, 0x65, 0x6c, 0x64, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x79, 0x69, 0x65, 0x6c,
+  0x64, 0x20, 0x72, 0x65, 0x73, 0x75, 0x6c, 0x74, 0x3b, 0x0a, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f,
+  0x2f, 0x20, 0x69, 0x66, 0x20, 0x77, 0x65, 0x20, 0x67, 0x6f, 0x74, 0x20,
+  0x61, 0x20, 0x73, 0x74, 0x6f, 0x70, 0x20, 0x74, 0x6f, 0x6b, 0x65, 0x6e,
+  0x20, 0x66, 0x72, 0x6f, 0x6d, 0x20, 0x73, 0x65, 0x72, 0x76, 0x65, 0x72,
+  0x2c, 0x20, 0x77, 0x65, 0x20, 0x77, 0x69, 0x6c, 0x6c, 0x20, 0x62, 0x72,
+  0x65, 0x61, 0x6b, 0x20, 0x68, 0x65, 0x72, 0x65, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20,
+  0x28, 0x72, 0x65, 0x73, 0x75, 0x6c, 0x74, 0x2e, 0x64, 0x61, 0x74, 0x61,
+  0x2e, 0x73, 0x74, 0x6f, 0x70, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69,
+  0x66, 0x20, 0x28, 0x72, 0x65, 0x73, 0x75, 0x6c, 0x74, 0x2e, 0x64, 0x61,
+  0x74, 0x61, 0x2e, 0x67, 0x65, 0x6e, 0x65, 0x72, 0x61, 0x74, 0x69, 0x6f,
+  0x6e, 0x5f, 0x73, 0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x29, 0x20,
+  0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x67, 0x65, 0x6e, 0x65, 0x72, 0x61,
+  0x74, 0x69, 0x6f, 0x6e, 0x5f, 0x73, 0x65, 0x74, 0x74, 0x69, 0x6e, 0x67,
+  0x73, 0x20, 0x3d, 0x20, 0x72, 0x65, 0x73, 0x75, 0x6c, 0x74, 0x2e, 0x64,
   0x61, 0x74, 0x61, 0x2e, 0x67, 0x65, 0x6e, 0x65, 0x72, 0x61, 0x74, 0x69,
-  0x6f, 0x6e, 0x5f, 0x73, 0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x20,
-  0x7d, 0x29, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d,
-  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x63,
-  0x68, 0x75, 0x6e, 0x6b, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x2e, 0x74, 0x69,
-  0x6d, 0x69, 0x6e, 0x67, 0x73, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x65, 0x76, 0x65, 0x6e, 0x74, 0x54, 0x61,
-  0x72, 0x67, 0x65, 0x74, 0x2e, 0x64, 0x69, 0x73, 0x70, 0x61, 0x74, 0x63,
-  0x68, 0x45, 0x76, 0x65, 0x6e, 0x74, 0x28, 0x6e, 0x65, 0x77, 0x20, 0x43,
-  0x75, 0x73, 0x74, 0x6f, 0x6d, 0x45, 0x76, 0x65, 0x6e, 0x74, 0x28, 0x22,
-  0x74, 0x69, 0x6d, 0x69, 0x6e, 0x67, 0x73, 0x22, 0x2c, 0x20, 0x7b, 0x20,
-  0x64, 0x65, 0x74, 0x61, 0x69, 0x6c, 0x3a, 0x20, 0x63, 0x68, 0x75, 0x6e,
-  0x6b, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x2e, 0x74, 0x69, 0x6d, 0x69, 0x6e,
-  0x67, 0x73, 0x20, 0x7d, 0x29, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20,
-  0x20, 0x20, 0x65, 0x76, 0x65, 0x6e, 0x74, 0x54, 0x61, 0x72, 0x67, 0x65,
-  0x74, 0x2e, 0x64, 0x69, 0x73, 0x70, 0x61, 0x74, 0x63, 0x68, 0x45, 0x76,
-  0x65, 0x6e, 0x74, 0x28, 0x6e, 0x65, 0x77, 0x20, 0x43, 0x75, 0x73, 0x74,
-  0x6f, 0x6d, 0x45, 0x76, 0x65, 0x6e, 0x74, 0x28, 0x22, 0x64, 0x6f, 0x6e,
-  0x65, 0x22, 0x2c, 0x20, 0x7b, 0x20, 0x64, 0x65, 0x74, 0x61, 0x69, 0x6c,
-  0x3a, 0x20, 0x7b, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x20,
-  0x7d, 0x20, 0x7d, 0x29, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x7d, 0x29, 0x28,
-  0x29, 0x3b, 0x0a, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20,
-  0x65, 0x76, 0x65, 0x6e, 0x74, 0x54, 0x61, 0x72, 0x67, 0x65, 0x74, 0x3b,
-  0x0a, 0x7d, 0x0a, 0x0a, 0x2f, 0x2f, 0x20, 0x43, 0x61, 0x6c, 0x6c, 0x20,
-  0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x2c, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72,
-  0x6e, 0x20, 0x61, 0x20, 0x70, 0x72, 0x6f, 0x6d, 0x69, 0x73, 0x65, 0x20,
-  0x74, 0x68, 0x61, 0x74, 0x20, 0x72, 0x65, 0x73, 0x6f, 0x6c, 0x76, 0x65,
-  0x73, 0x20, 0x74, 0x6f, 0x20, 0x74, 0x68, 0x65, 0x20, 0x63, 0x6f, 0x6d,
-  0x70, 0x6c, 0x65, 0x74, 0x65, 0x64, 0x20, 0x74, 0x65, 0x78, 0x74, 0x2e,
-  0x20, 0x54, 0x68, 0x69, 0x73, 0x20, 0x64, 0x6f, 0x65, 0x73, 0x20, 0x6e,
-  0x6f, 0x74, 0x20, 0x73, 0x75, 0x70, 0x70, 0x6f, 0x72, 0x74, 0x20, 0x73,
-  0x74, 0x72, 0x65, 0x61, 0x6d, 0x69, 0x6e, 0x67, 0x0a, 0x2f, 0x2f, 0x0a,
-  0x2f, 0x2f, 0x20, 0x45, 0x78, 0x61, 0x6d, 0x70, 0x6c, 0x65, 0x3a, 0x0a,
-  0x2f, 0x2f, 0x0a, 0x2f, 0x2f, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x6c,
-  0x61, 0x6d, 0x61, 0x50, 0x72, 0x6f, 0x6d, 0x69, 0x73, 0x65, 0x28, 0x70,
-  0x72, 0x6f, 0x6d, 0x70, 0x74, 0x29, 0x2e, 0x74, 0x68, 0x65, 0x6e, 0x28,
-  0x28, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x29, 0x20, 0x3d, 0x3e,
-  0x20, 0x7b, 0x0a, 0x2f, 0x2f, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x64, 0x6f, 0x63, 0x75, 0x6d, 0x65, 0x6e, 0x74, 0x2e, 0x77, 0x72, 0x69,
-  0x74, 0x65, 0x28, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x29, 0x0a,
-  0x2f, 0x2f, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x29, 0x0a, 0x2f, 0x2f,
-  0x0a, 0x2f, 0x2f, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6f, 0x72, 0x0a, 0x2f,
-  0x2f, 0x0a, 0x2f, 0x2f, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e,
-  0x73, 0x74, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x20, 0x3d,
-  0x20, 0x61, 0x77, 0x61, 0x69, 0x74, 0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61,
-  0x50, 0x72, 0x6f, 0x6d, 0x69, 0x73, 0x65, 0x28, 0x70, 0x72, 0x6f, 0x6d,
-  0x70, 0x74, 0x29, 0x0a, 0x2f, 0x2f, 0x20, 0x20, 0x20, 0x20, 0x20, 0x64,
-  0x6f, 0x63, 0x75, 0x6d, 0x65, 0x6e, 0x74, 0x2e, 0x77, 0x72, 0x69, 0x74,
-  0x65, 0x28, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x29, 0x0a, 0x2f,
-  0x2f, 0x0a, 0x65, 0x78, 0x70, 0x6f, 0x72, 0x74, 0x20, 0x63, 0x6f, 0x6e,
-  0x73, 0x74, 0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x50, 0x72, 0x6f, 0x6d,
-  0x69, 0x73, 0x65, 0x20, 0x3d, 0x20, 0x28, 0x70, 0x72, 0x6f, 0x6d, 0x70,
+  0x6f, 0x6e, 0x5f, 0x73, 0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x3b,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x20,
+  0x3d, 0x20, 0x66, 0x61, 0x6c, 0x73, 0x65, 0x3b, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x62,
+  0x72, 0x65, 0x61, 0x6b, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x7d, 0x20,
+  0x63, 0x61, 0x74, 0x63, 0x68, 0x20, 0x28, 0x65, 0x29, 0x20, 0x7b, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x65, 0x2e, 0x6e, 0x61,
+  0x6d, 0x65, 0x20, 0x21, 0x3d, 0x3d, 0x20, 0x27, 0x41, 0x62, 0x6f, 0x72,
+  0x74, 0x45, 0x72, 0x72, 0x6f, 0x72, 0x27, 0x29, 0x20, 0x7b, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x6f, 0x6c, 0x65,
+  0x2e, 0x65, 0x72, 0x72, 0x6f, 0x72, 0x28, 0x22, 0x6c, 0x6c, 0x61, 0x6d,
+  0x61, 0x20, 0x65, 0x72, 0x72, 0x6f, 0x72, 0x3a, 0x20, 0x22, 0x2c, 0x20,
+  0x65, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x74, 0x68, 0x72, 0x6f, 0x77, 0x20, 0x65, 0x3b, 0x0a, 0x20,
+  0x20, 0x7d, 0x0a, 0x20, 0x20, 0x66, 0x69, 0x6e, 0x61, 0x6c, 0x6c, 0x79,
+  0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x72,
+  0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x2e, 0x61, 0x62, 0x6f, 0x72, 0x74, 0x28,
+  0x29, 0x3b, 0x0a, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x72, 0x65,
+  0x74, 0x75, 0x72, 0x6e, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74,
+  0x3b, 0x0a, 0x7d, 0x0a, 0x0a, 0x2f, 0x2f, 0x20, 0x43, 0x61, 0x6c, 0x6c,
+  0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x2c, 0x20, 0x72, 0x65, 0x74, 0x75,
+  0x72, 0x6e, 0x20, 0x61, 0x6e, 0x20, 0x65, 0x76, 0x65, 0x6e, 0x74, 0x20,
+  0x74, 0x61, 0x72, 0x67, 0x65, 0x74, 0x20, 0x74, 0x68, 0x61, 0x74, 0x20,
+  0x79, 0x6f, 0x75, 0x20, 0x63, 0x61, 0x6e, 0x20, 0x73, 0x75, 0x62, 0x63,
+  0x72, 0x69, 0x62, 0x65, 0x20, 0x74, 0x6f, 0x0a, 0x2f, 0x2f, 0x0a, 0x2f,
+  0x2f, 0x20, 0x45, 0x78, 0x61, 0x6d, 0x70, 0x6c, 0x65, 0x3a, 0x0a, 0x2f,
+  0x2f, 0x0a, 0x2f, 0x2f, 0x20, 0x20, 0x20, 0x20, 0x69, 0x6d, 0x70, 0x6f,
+  0x72, 0x74, 0x20, 0x7b, 0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x45, 0x76,
+  0x65, 0x6e, 0x74, 0x54, 0x61, 0x72, 0x67, 0x65, 0x74, 0x20, 0x7d, 0x20,
+  0x66, 0x72, 0x6f, 0x6d, 0x20, 0x27, 0x2f, 0x63, 0x6f, 0x6d, 0x70, 0x6c,
+  0x65, 0x74, 0x69, 0x6f, 0x6e, 0x2e, 0x6a, 0x73, 0x27, 0x0a, 0x2f, 0x2f,
+  0x0a, 0x2f, 0x2f, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74,
+  0x20, 0x63, 0x6f, 0x6e, 0x6e, 0x20, 0x3d, 0x20, 0x6c, 0x6c, 0x61, 0x6d,
+  0x61, 0x45, 0x76, 0x65, 0x6e, 0x74, 0x54, 0x61, 0x72, 0x67, 0x65, 0x74,
+  0x28, 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x29, 0x0a, 0x2f, 0x2f, 0x20,
+  0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x6e, 0x2e, 0x61, 0x64, 0x64, 0x45,
+  0x76, 0x65, 0x6e, 0x74, 0x4c, 0x69, 0x73, 0x74, 0x65, 0x6e, 0x65, 0x72,
+  0x28, 0x22, 0x6d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x22, 0x2c, 0x20,
+  0x28, 0x63, 0x68, 0x75, 0x6e, 0x6b, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b,
+  0x0a, 0x2f, 0x2f, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x64, 0x6f, 0x63,
+  0x75, 0x6d, 0x65, 0x6e, 0x74, 0x2e, 0x77, 0x72, 0x69, 0x74, 0x65, 0x28,
+  0x63, 0x68, 0x75, 0x6e, 0x6b, 0x2e, 0x64, 0x65, 0x74, 0x61, 0x69, 0x6c,
+  0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x29, 0x0a, 0x2f, 0x2f,
+  0x20, 0x20, 0x20, 0x20, 0x7d, 0x29, 0x0a, 0x2f, 0x2f, 0x0a, 0x65, 0x78,
+  0x70, 0x6f, 0x72, 0x74, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x6c,
+  0x6c, 0x61, 0x6d, 0x61, 0x45, 0x76, 0x65, 0x6e, 0x74, 0x54, 0x61, 0x72,
+  0x67, 0x65, 0x74, 0x20, 0x3d, 0x20, 0x28, 0x70, 0x72, 0x6f, 0x6d, 0x70,
   0x74, 0x2c, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x20, 0x3d, 0x20,
   0x7b, 0x7d, 0x2c, 0x20, 0x63, 0x6f, 0x6e, 0x66, 0x69, 0x67, 0x20, 0x3d,
   0x20, 0x7b, 0x7d, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20,
-  0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x6e, 0x65, 0x77, 0x20, 0x50,
-  0x72, 0x6f, 0x6d, 0x69, 0x73, 0x65, 0x28, 0x61, 0x73, 0x79, 0x6e, 0x63,
-  0x20, 0x28, 0x72, 0x65, 0x73, 0x6f, 0x6c, 0x76, 0x65, 0x2c, 0x20, 0x72,
-  0x65, 0x6a, 0x65, 0x63, 0x74, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a,
-  0x20, 0x20, 0x20, 0x20, 0x6c, 0x65, 0x74, 0x20, 0x63, 0x6f, 0x6e, 0x74,
-  0x65, 0x6e, 0x74, 0x20, 0x3d, 0x20, 0x22, 0x22, 0x3b, 0x0a, 0x20, 0x20,
-  0x20, 0x20, 0x74, 0x72, 0x79, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x66, 0x6f, 0x72, 0x20, 0x61, 0x77, 0x61, 0x69, 0x74, 0x20,
-  0x28, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x63, 0x68, 0x75, 0x6e, 0x6b,
-  0x20, 0x6f, 0x66, 0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x28, 0x70, 0x72,
-  0x6f, 0x6d, 0x70, 0x74, 0x2c, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73,
-  0x2c, 0x20, 0x63, 0x6f, 0x6e, 0x66, 0x69, 0x67, 0x29, 0x29, 0x20, 0x7b,
-  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e,
-  0x74, 0x65, 0x6e, 0x74, 0x20, 0x2b, 0x3d, 0x20, 0x63, 0x68, 0x75, 0x6e,
-  0x6b, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x65,
-  0x6e, 0x74, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x73, 0x6f, 0x6c, 0x76,
-  0x65, 0x28, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x29, 0x3b, 0x0a,
-  0x20, 0x20, 0x20, 0x20, 0x7d, 0x20, 0x63, 0x61, 0x74, 0x63, 0x68, 0x20,
-  0x28, 0x65, 0x72, 0x72, 0x6f, 0x72, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x6a, 0x65, 0x63, 0x74, 0x28, 0x65,
-  0x72, 0x72, 0x6f, 0x72, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d,
-  0x0a, 0x20, 0x20, 0x7d, 0x29, 0x3b, 0x0a, 0x7d, 0x3b, 0x0a, 0x0a, 0x2f,
-  0x2a, 0x2a, 0x0a, 0x20, 0x2a, 0x20, 0x28, 0x64, 0x65, 0x70, 0x72, 0x65,
-  0x63, 0x61, 0x74, 0x65, 0x64, 0x29, 0x0a, 0x20, 0x2a, 0x2f, 0x0a, 0x65,
-  0x78, 0x70, 0x6f, 0x72, 0x74, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20,
-  0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x43, 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74,
-  0x65, 0x20, 0x3d, 0x20, 0x61, 0x73, 0x79, 0x6e, 0x63, 0x20, 0x28, 0x70,
-  0x61, 0x72, 0x61, 0x6d, 0x73, 0x2c, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x72,
-  0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x2c, 0x20, 0x63, 0x61, 0x6c, 0x6c, 0x62,
-  0x61, 0x63, 0x6b, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20,
-  0x66, 0x6f, 0x72, 0x20, 0x61, 0x77, 0x61, 0x69, 0x74, 0x20, 0x28, 0x63,
-  0x6f, 0x6e, 0x73, 0x74, 0x20, 0x63, 0x68, 0x75, 0x6e, 0x6b, 0x20, 0x6f,
-  0x66, 0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x28, 0x70, 0x61, 0x72, 0x61,
-  0x6d, 0x73, 0x2e, 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x2c, 0x20, 0x70,
-  0x61, 0x72, 0x61, 0x6d, 0x73, 0x2c, 0x20, 0x7b, 0x20, 0x63, 0x6f, 0x6e,
-  0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x20, 0x7d, 0x29, 0x29, 0x20,
-  0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x61, 0x6c, 0x6c, 0x62, 0x61,
-  0x63, 0x6b, 0x28, 0x63, 0x68, 0x75, 0x6e, 0x6b, 0x29, 0x3b, 0x0a, 0x20,
-  0x20, 0x7d, 0x0a, 0x7d, 0x0a, 0x0a, 0x2f, 0x2f, 0x20, 0x47, 0x65, 0x74,
-  0x20, 0x74, 0x68, 0x65, 0x20, 0x6d, 0x6f, 0x64, 0x65, 0x6c, 0x20, 0x69,
-  0x6e, 0x66, 0x6f, 0x20, 0x66, 0x72, 0x6f, 0x6d, 0x20, 0x74, 0x68, 0x65,
-  0x20, 0x73, 0x65, 0x72, 0x76, 0x65, 0x72, 0x2e, 0x20, 0x54, 0x68, 0x69,
-  0x73, 0x20, 0x69, 0x73, 0x20, 0x75, 0x73, 0x65, 0x66, 0x75, 0x6c, 0x20,
-  0x66, 0x6f, 0x72, 0x20, 0x67, 0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, 0x20,
-  0x74, 0x68, 0x65, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x78, 0x74, 0x20,
-  0x77, 0x69, 0x6e, 0x64, 0x6f, 0x77, 0x20, 0x61, 0x6e, 0x64, 0x20, 0x73,
-  0x6f, 0x20, 0x6f, 0x6e, 0x2e, 0x0a, 0x65, 0x78, 0x70, 0x6f, 0x72, 0x74,
-  0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61,
-  0x4d, 0x6f, 0x64, 0x65, 0x6c, 0x49, 0x6e, 0x66, 0x6f, 0x20, 0x3d, 0x20,
-  0x61, 0x73, 0x79, 0x6e, 0x63, 0x20, 0x28, 0x29, 0x20, 0x3d, 0x3e, 0x20,
-  0x7b, 0x0a, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x21, 0x67, 0x65, 0x6e,
-  0x65, 0x72, 0x61, 0x74, 0x69, 0x6f, 0x6e, 0x5f, 0x73, 0x65, 0x74, 0x74,
-  0x69, 0x6e, 0x67, 0x73, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x65, 0x76, 0x65, 0x6e, 0x74, 0x54,
+  0x61, 0x72, 0x67, 0x65, 0x74, 0x20, 0x3d, 0x20, 0x6e, 0x65, 0x77, 0x20,
+  0x45, 0x76, 0x65, 0x6e, 0x74, 0x54, 0x61, 0x72, 0x67, 0x65, 0x74, 0x28,
+  0x29, 0x3b, 0x0a, 0x20, 0x20, 0x28, 0x61, 0x73, 0x79, 0x6e, 0x63, 0x20,
+  0x28, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x6c, 0x65, 0x74, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x20,
+  0x3d, 0x20, 0x22, 0x22, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x66, 0x6f,
+  0x72, 0x20, 0x61, 0x77, 0x61, 0x69, 0x74, 0x20, 0x28, 0x63, 0x6f, 0x6e,
+  0x73, 0x74, 0x20, 0x63, 0x68, 0x75, 0x6e, 0x6b, 0x20, 0x6f, 0x66, 0x20,
+  0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x28, 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74,
+  0x2c, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2c, 0x20, 0x63, 0x6f,
+  0x6e, 0x66, 0x69, 0x67, 0x29, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x63, 0x68, 0x75, 0x6e, 0x6b,
+  0x2e, 0x64, 0x61, 0x74, 0x61, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74,
+  0x20, 0x2b, 0x3d, 0x20, 0x63, 0x68, 0x75, 0x6e, 0x6b, 0x2e, 0x64, 0x61,
+  0x74, 0x61, 0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x3b, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x65, 0x76, 0x65, 0x6e,
+  0x74, 0x54, 0x61, 0x72, 0x67, 0x65, 0x74, 0x2e, 0x64, 0x69, 0x73, 0x70,
+  0x61, 0x74, 0x63, 0x68, 0x45, 0x76, 0x65, 0x6e, 0x74, 0x28, 0x6e, 0x65,
+  0x77, 0x20, 0x43, 0x75, 0x73, 0x74, 0x6f, 0x6d, 0x45, 0x76, 0x65, 0x6e,
+  0x74, 0x28, 0x22, 0x6d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x22, 0x2c,
+  0x20, 0x7b, 0x20, 0x64, 0x65, 0x74, 0x61, 0x69, 0x6c, 0x3a, 0x20, 0x63,
+  0x68, 0x75, 0x6e, 0x6b, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x20, 0x7d, 0x29,
+  0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x63, 0x68, 0x75,
+  0x6e, 0x6b, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x2e, 0x67, 0x65, 0x6e, 0x65,
+  0x72, 0x61, 0x74, 0x69, 0x6f, 0x6e, 0x5f, 0x73, 0x65, 0x74, 0x74, 0x69,
+  0x6e, 0x67, 0x73, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x65, 0x76, 0x65, 0x6e, 0x74, 0x54, 0x61, 0x72, 0x67,
+  0x65, 0x74, 0x2e, 0x64, 0x69, 0x73, 0x70, 0x61, 0x74, 0x63, 0x68, 0x45,
+  0x76, 0x65, 0x6e, 0x74, 0x28, 0x6e, 0x65, 0x77, 0x20, 0x43, 0x75, 0x73,
+  0x74, 0x6f, 0x6d, 0x45, 0x76, 0x65, 0x6e, 0x74, 0x28, 0x22, 0x67, 0x65,
+  0x6e, 0x65, 0x72, 0x61, 0x74, 0x69, 0x6f, 0x6e, 0x5f, 0x73, 0x65, 0x74,
+  0x74, 0x69, 0x6e, 0x67, 0x73, 0x22, 0x2c, 0x20, 0x7b, 0x20, 0x64, 0x65,
+  0x74, 0x61, 0x69, 0x6c, 0x3a, 0x20, 0x63, 0x68, 0x75, 0x6e, 0x6b, 0x2e,
+  0x64, 0x61, 0x74, 0x61, 0x2e, 0x67, 0x65, 0x6e, 0x65, 0x72, 0x61, 0x74,
+  0x69, 0x6f, 0x6e, 0x5f, 0x73, 0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, 0x73,
+  0x20, 0x7d, 0x29, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28,
+  0x63, 0x68, 0x75, 0x6e, 0x6b, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x2e, 0x74,
+  0x69, 0x6d, 0x69, 0x6e, 0x67, 0x73, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x65, 0x76, 0x65, 0x6e, 0x74, 0x54,
+  0x61, 0x72, 0x67, 0x65, 0x74, 0x2e, 0x64, 0x69, 0x73, 0x70, 0x61, 0x74,
+  0x63, 0x68, 0x45, 0x76, 0x65, 0x6e, 0x74, 0x28, 0x6e, 0x65, 0x77, 0x20,
+  0x43, 0x75, 0x73, 0x74, 0x6f, 0x6d, 0x45, 0x76, 0x65, 0x6e, 0x74, 0x28,
+  0x22, 0x74, 0x69, 0x6d, 0x69, 0x6e, 0x67, 0x73, 0x22, 0x2c, 0x20, 0x7b,
+  0x20, 0x64, 0x65, 0x74, 0x61, 0x69, 0x6c, 0x3a, 0x20, 0x63, 0x68, 0x75,
+  0x6e, 0x6b, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x2e, 0x74, 0x69, 0x6d, 0x69,
+  0x6e, 0x67, 0x73, 0x20, 0x7d, 0x29, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x65, 0x76, 0x65, 0x6e, 0x74, 0x54, 0x61, 0x72, 0x67,
+  0x65, 0x74, 0x2e, 0x64, 0x69, 0x73, 0x70, 0x61, 0x74, 0x63, 0x68, 0x45,
+  0x76, 0x65, 0x6e, 0x74, 0x28, 0x6e, 0x65, 0x77, 0x20, 0x43, 0x75, 0x73,
+  0x74, 0x6f, 0x6d, 0x45, 0x76, 0x65, 0x6e, 0x74, 0x28, 0x22, 0x64, 0x6f,
+  0x6e, 0x65, 0x22, 0x2c, 0x20, 0x7b, 0x20, 0x64, 0x65, 0x74, 0x61, 0x69,
+  0x6c, 0x3a, 0x20, 0x7b, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74,
+  0x20, 0x7d, 0x20, 0x7d, 0x29, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x7d, 0x29,
+  0x28, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e,
+  0x20, 0x65, 0x76, 0x65, 0x6e, 0x74, 0x54, 0x61, 0x72, 0x67, 0x65, 0x74,
+  0x3b, 0x0a, 0x7d, 0x0a, 0x0a, 0x2f, 0x2f, 0x20, 0x43, 0x61, 0x6c, 0x6c,
+  0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x2c, 0x20, 0x72, 0x65, 0x74, 0x75,
+  0x72, 0x6e, 0x20, 0x61, 0x20, 0x70, 0x72, 0x6f, 0x6d, 0x69, 0x73, 0x65,
+  0x20, 0x74, 0x68, 0x61, 0x74, 0x20, 0x72, 0x65, 0x73, 0x6f, 0x6c, 0x76,
+  0x65, 0x73, 0x20, 0x74, 0x6f, 0x20, 0x74, 0x68, 0x65, 0x20, 0x63, 0x6f,
+  0x6d, 0x70, 0x6c, 0x65, 0x74, 0x65, 0x64, 0x20, 0x74, 0x65, 0x78, 0x74,
+  0x2e, 0x20, 0x54, 0x68, 0x69, 0x73, 0x20, 0x64, 0x6f, 0x65, 0x73, 0x20,
+  0x6e, 0x6f, 0x74, 0x20, 0x73, 0x75, 0x70, 0x70, 0x6f, 0x72, 0x74, 0x20,
+  0x73, 0x74, 0x72, 0x65, 0x61, 0x6d, 0x69, 0x6e, 0x67, 0x0a, 0x2f, 0x2f,
+  0x0a, 0x2f, 0x2f, 0x20, 0x45, 0x78, 0x61, 0x6d, 0x70, 0x6c, 0x65, 0x3a,
+  0x0a, 0x2f, 0x2f, 0x0a, 0x2f, 0x2f, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6c,
+  0x6c, 0x61, 0x6d, 0x61, 0x50, 0x72, 0x6f, 0x6d, 0x69, 0x73, 0x65, 0x28,
+  0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x29, 0x2e, 0x74, 0x68, 0x65, 0x6e,
+  0x28, 0x28, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x29, 0x20, 0x3d,
+  0x3e, 0x20, 0x7b, 0x0a, 0x2f, 0x2f, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x64, 0x6f, 0x63, 0x75, 0x6d, 0x65, 0x6e, 0x74, 0x2e, 0x77, 0x72,
+  0x69, 0x74, 0x65, 0x28, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x29,
+  0x0a, 0x2f, 0x2f, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x29, 0x0a, 0x2f,
+  0x2f, 0x0a, 0x2f, 0x2f, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6f, 0x72, 0x0a,
+  0x2f, 0x2f, 0x0a, 0x2f, 0x2f, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f,
+  0x6e, 0x73, 0x74, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x20,
+  0x3d, 0x20, 0x61, 0x77, 0x61, 0x69, 0x74, 0x20, 0x6c, 0x6c, 0x61, 0x6d,
+  0x61, 0x50, 0x72, 0x6f, 0x6d, 0x69, 0x73, 0x65, 0x28, 0x70, 0x72, 0x6f,
+  0x6d, 0x70, 0x74, 0x29, 0x0a, 0x2f, 0x2f, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x64, 0x6f, 0x63, 0x75, 0x6d, 0x65, 0x6e, 0x74, 0x2e, 0x77, 0x72, 0x69,
+  0x74, 0x65, 0x28, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x29, 0x0a,
+  0x2f, 0x2f, 0x0a, 0x65, 0x78, 0x70, 0x6f, 0x72, 0x74, 0x20, 0x63, 0x6f,
+  0x6e, 0x73, 0x74, 0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x50, 0x72, 0x6f,
+  0x6d, 0x69, 0x73, 0x65, 0x20, 0x3d, 0x20, 0x28, 0x70, 0x72, 0x6f, 0x6d,
+  0x70, 0x74, 0x2c, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x20, 0x3d,
+  0x20, 0x7b, 0x7d, 0x2c, 0x20, 0x63, 0x6f, 0x6e, 0x66, 0x69, 0x67, 0x20,
+  0x3d, 0x20, 0x7b, 0x7d, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20,
+  0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x6e, 0x65, 0x77, 0x20,
+  0x50, 0x72, 0x6f, 0x6d, 0x69, 0x73, 0x65, 0x28, 0x61, 0x73, 0x79, 0x6e,
+  0x63, 0x20, 0x28, 0x72, 0x65, 0x73, 0x6f, 0x6c, 0x76, 0x65, 0x2c, 0x20,
+  0x72, 0x65, 0x6a, 0x65, 0x63, 0x74, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x65, 0x74, 0x20, 0x63, 0x6f, 0x6e,
+  0x74, 0x65, 0x6e, 0x74, 0x20, 0x3d, 0x20, 0x22, 0x22, 0x3b, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x74, 0x72, 0x79, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x66, 0x6f, 0x72, 0x20, 0x61, 0x77, 0x61, 0x69, 0x74,
+  0x20, 0x28, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x63, 0x68, 0x75, 0x6e,
+  0x6b, 0x20, 0x6f, 0x66, 0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x28, 0x70,
+  0x72, 0x6f, 0x6d, 0x70, 0x74, 0x2c, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d,
+  0x73, 0x2c, 0x20, 0x63, 0x6f, 0x6e, 0x66, 0x69, 0x67, 0x29, 0x29, 0x20,
+  0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f,
+  0x6e, 0x74, 0x65, 0x6e, 0x74, 0x20, 0x2b, 0x3d, 0x20, 0x63, 0x68, 0x75,
+  0x6e, 0x6b, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x2e, 0x63, 0x6f, 0x6e, 0x74,
+  0x65, 0x6e, 0x74, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x73, 0x6f, 0x6c,
+  0x76, 0x65, 0x28, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x29, 0x3b,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x20, 0x63, 0x61, 0x74, 0x63, 0x68,
+  0x20, 0x28, 0x65, 0x72, 0x72, 0x6f, 0x72, 0x29, 0x20, 0x7b, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x6a, 0x65, 0x63, 0x74, 0x28,
+  0x65, 0x72, 0x72, 0x6f, 0x72, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x7d, 0x0a, 0x20, 0x20, 0x7d, 0x29, 0x3b, 0x0a, 0x7d, 0x3b, 0x0a, 0x0a,
+  0x2f, 0x2a, 0x2a, 0x0a, 0x20, 0x2a, 0x20, 0x28, 0x64, 0x65, 0x70, 0x72,
+  0x65, 0x63, 0x61, 0x74, 0x65, 0x64, 0x29, 0x0a, 0x20, 0x2a, 0x2f, 0x0a,
+  0x65, 0x78, 0x70, 0x6f, 0x72, 0x74, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74,
+  0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x43, 0x6f, 0x6d, 0x70, 0x6c, 0x65,
+  0x74, 0x65, 0x20, 0x3d, 0x20, 0x61, 0x73, 0x79, 0x6e, 0x63, 0x20, 0x28,
+  0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2c, 0x20, 0x63, 0x6f, 0x6e, 0x74,
+  0x72, 0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x2c, 0x20, 0x63, 0x61, 0x6c, 0x6c,
+  0x62, 0x61, 0x63, 0x6b, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20,
+  0x20, 0x66, 0x6f, 0x72, 0x20, 0x61, 0x77, 0x61, 0x69, 0x74, 0x20, 0x28,
+  0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x63, 0x68, 0x75, 0x6e, 0x6b, 0x20,
+  0x6f, 0x66, 0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x28, 0x70, 0x61, 0x72,
+  0x61, 0x6d, 0x73, 0x2e, 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x2c, 0x20,
+  0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2c, 0x20, 0x7b, 0x20, 0x63, 0x6f,
+  0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x20, 0x7d, 0x29, 0x29,
+  0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x61, 0x6c, 0x6c, 0x62,
+  0x61, 0x63, 0x6b, 0x28, 0x63, 0x68, 0x75, 0x6e, 0x6b, 0x29, 0x3b, 0x0a,
+  0x20, 0x20, 0x7d, 0x0a, 0x7d, 0x0a, 0x0a, 0x2f, 0x2f, 0x20, 0x47, 0x65,
+  0x74, 0x20, 0x74, 0x68, 0x65, 0x20, 0x6d, 0x6f, 0x64, 0x65, 0x6c, 0x20,
+  0x69, 0x6e, 0x66, 0x6f, 0x20, 0x66, 0x72, 0x6f, 0x6d, 0x20, 0x74, 0x68,
+  0x65, 0x20, 0x73, 0x65, 0x72, 0x76, 0x65, 0x72, 0x2e, 0x20, 0x54, 0x68,
+  0x69, 0x73, 0x20, 0x69, 0x73, 0x20, 0x75, 0x73, 0x65, 0x66, 0x75, 0x6c,
+  0x20, 0x66, 0x6f, 0x72, 0x20, 0x67, 0x65, 0x74, 0x74, 0x69, 0x6e, 0x67,
+  0x20, 0x74, 0x68, 0x65, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x78, 0x74,
+  0x20, 0x77, 0x69, 0x6e, 0x64, 0x6f, 0x77, 0x20, 0x61, 0x6e, 0x64, 0x20,
+  0x73, 0x6f, 0x20, 0x6f, 0x6e, 0x2e, 0x0a, 0x65, 0x78, 0x70, 0x6f, 0x72,
+  0x74, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x6c, 0x6c, 0x61, 0x6d,
+  0x61, 0x4d, 0x6f, 0x64, 0x65, 0x6c, 0x49, 0x6e, 0x66, 0x6f, 0x20, 0x3d,
+  0x20, 0x61, 0x73, 0x79, 0x6e, 0x63, 0x20, 0x28, 0x29, 0x20, 0x3d, 0x3e,
+  0x20, 0x7b, 0x0a, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x21, 0x67, 0x65,
+  0x6e, 0x65, 0x72, 0x61, 0x74, 0x69, 0x6f, 0x6e, 0x5f, 0x73, 0x65, 0x74,
+  0x74, 0x69, 0x6e, 0x67, 0x73, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x67, 0x65, 0x6e, 0x65, 0x72, 0x61, 0x74, 0x69, 0x6f, 0x6e, 0x5f,
+  0x73, 0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x20, 0x3d, 0x20, 0x61,
+  0x77, 0x61, 0x69, 0x74, 0x20, 0x66, 0x65, 0x74, 0x63, 0x68, 0x28, 0x22,
+  0x2f, 0x6d, 0x6f, 0x64, 0x65, 0x6c, 0x2e, 0x6a, 0x73, 0x6f, 0x6e, 0x22,
+  0x29, 0x2e, 0x74, 0x68, 0x65, 0x6e, 0x28, 0x72, 0x20, 0x3d, 0x3e, 0x20,
+  0x72, 0x2e, 0x6a, 0x73, 0x6f, 0x6e, 0x28, 0x29, 0x29, 0x3b, 0x0a, 0x20,
+  0x20, 0x7d, 0x0a, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20,
   0x67, 0x65, 0x6e, 0x65, 0x72, 0x61, 0x74, 0x69, 0x6f, 0x6e, 0x5f, 0x73,
-  0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x20, 0x3d, 0x20, 0x61, 0x77,
-  0x61, 0x69, 0x74, 0x20, 0x66, 0x65, 0x74, 0x63, 0x68, 0x28, 0x22, 0x2f,
-  0x6d, 0x6f, 0x64, 0x65, 0x6c, 0x2e, 0x6a, 0x73, 0x6f, 0x6e, 0x22, 0x29,
-  0x2e, 0x74, 0x68, 0x65, 0x6e, 0x28, 0x72, 0x20, 0x3d, 0x3e, 0x20, 0x72,
-  0x2e, 0x6a, 0x73, 0x6f, 0x6e, 0x28, 0x29, 0x29, 0x3b, 0x0a, 0x20, 0x20,
-  0x7d, 0x0a, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x67,
-  0x65, 0x6e, 0x65, 0x72, 0x61, 0x74, 0x69, 0x6f, 0x6e, 0x5f, 0x73, 0x65,
-  0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x3b, 0x0a, 0x7d, 0x0a
+  0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x3b, 0x0a, 0x7d, 0x0a
 };
-unsigned int completion_js_len = 4462;
+unsigned int completion_js_len = 5099;

examples/server/public/completion.js

@@ -43,6 +43,7 @@ export async function* llama(prompt, params = {}, config = {}) {
   const decoder = new TextDecoder();
 
   let content = "";
+  let leftover = ""; // Buffer for partially read lines
 
   try {
     let cont = true;
@@ -53,29 +54,47 @@ export async function* llama(prompt, params = {}, config = {}) {
         break;
       }
 
-      // sse answers in the form multiple lines of: value\n with data always present as a key. in our case we
-      // mainly care about the data: key here, which we expect as json
-      const text = decoder.decode(result.value);
+      // Add any leftover data to the current chunk of data
+      const text = leftover + decoder.decode(result.value);
 
-      // parse all sse events and add them to result
-      const regex = /^(\S+):\s(.*)$/gm;
-      for (const match of text.matchAll(regex)) {
-        result[match[1]] = match[2]
-      }
+      // Check if the last character is a line break
+      const endsWithLineBreak = text.endsWith('\n');
 
-      // since we know this is llama.cpp, let's just decode the json in data
-      result.data = JSON.parse(result.data);
-      content += result.data.content;
+      // Split the text into lines
+      let lines = text.split('\n');
 
-      // yield
-      yield result;
+      // If the text doesn't end with a line break, then the last line is incomplete
+      // Store it in leftover to be added to the next chunk of data
+      if (!endsWithLineBreak) {
+        leftover = lines.pop();
+      } else {
+        leftover = ""; // Reset leftover if we have a line break at the end
+      }
 
-      // if we got a stop token from server, we will break here
-      if (result.data.stop) {
-        if (result.data.generation_settings) {
-          generation_settings = result.data.generation_settings;
+      // Parse all sse events and add them to result
+      const regex = /^(\S+):\s(.*)$/gm;
+      for (const line of lines) {
+        const match = regex.exec(line);
+        if (match) {
+          result[match[1]] = match[2]
+          // since we know this is llama.cpp, let's just decode the json in data
+          if (result.data) {
+            result.data = JSON.parse(result.data);
+            content += result.data.content;
+
+            // yield
+            yield result;
+
+            // if we got a stop token from server, we will break here
+            if (result.data.stop) {
+              if (result.data.generation_settings) {
+                generation_settings = result.data.generation_settings;
+              }
+              cont = false;
+              break;
+            }
+          }
         }
-        break;
       }
     }
   } catch (e) {

examples/server/public/index.html

@@ -3,12 +3,11 @@
 <head>
   <meta charset="UTF-8">
   <meta name="viewport" content="width=device-width, initial-scale=1, maximum-scale=1" />
+  <meta name="color-scheme" content="light dark">
   <title>llama.cpp - chat</title>
 
   <style>
     body {
-      background-color: #fff;
-      color: #000;
       font-family: system-ui;
       font-size: 90%;
     }
@@ -283,8 +282,9 @@
 
       useEffect(() => {
         // scroll to bottom (if needed)
-        if (container.current && container.current.scrollHeight <= container.current.scrollTop + container.current.offsetHeight + 300) {
-          container.current.scrollTo(0, container.current.scrollHeight)
+        const parent = container.current.parentElement;
+        if (parent && parent.scrollHeight <= parent.scrollTop + parent.offsetHeight + 300) {
+          parent.scrollTo(0, parent.scrollHeight)
         }
       }, [messages])
 

examples/server/server.cpp

@@ -1,6 +1,7 @@
 #include "common.h"
 #include "llama.h"
 #include "build-info.h"
+#include "grammar-parser.h"
 
 #ifndef NDEBUG
 // crash the server in debug mode, otherwise send an http 500 error
@@ -195,6 +196,8 @@ struct llama_server_context
     llama_context *ctx = nullptr;
     gpt_params params;
 
+    llama_grammar *grammar = nullptr;
+
     bool truncated = false;
     bool stopped_eos = false;
     bool stopped_word = false;
@@ -226,6 +229,7 @@ struct llama_server_context
     void rewind()
     {
         params.antiprompt.clear();
+        params.grammar.clear();
         num_prompt_tokens = 0;
         num_tokens_predicted = 0;
         generated_text = "";
@@ -237,6 +241,7 @@ struct llama_server_context
         stopped_limit = false;
         stopping_word = "";
         multibyte_pending = 0;
+        grammar = nullptr;
 
         n_remain = 0;
         n_past = 0;
@@ -257,6 +262,33 @@ struct llama_server_context
         return true;
     }
 
+    bool loadGrammar()
+    {
+        if (!params.grammar.empty()) {
+            grammar_parser::parse_state parsed_grammar;
+
+            parsed_grammar = grammar_parser::parse(params.grammar.c_str());
+            // will be empty (default) if there are parse errors
+            if (parsed_grammar.rules.empty()) {
+                LOG_ERROR("grammar parse error", {{"grammar", params.grammar}});
+                return false;
+            }
+            grammar_parser::print_grammar(stderr, parsed_grammar);
+
+            {
+                auto it = params.logit_bias.find(llama_token_eos());
+                if (it != params.logit_bias.end() && it->second == -INFINITY) {
+                    LOG_WARNING("EOS token is disabled, which will cause most grammars to fail", {});
+                }
+            }
+
+            std::vector<const llama_grammar_element *> grammar_rules(parsed_grammar.c_rules());
+            grammar = llama_grammar_init(
+                grammar_rules.data(), grammar_rules.size(), parsed_grammar.symbol_ids.at("root"));
+        }
+        return true;
+    }
+
     void loadPrompt()
     {
         params.prompt.insert(0, 1, ' '); // always add a first space
@@ -420,6 +452,10 @@ struct llama_server_context
                 logits[llama_token_nl()] = nl_logit;
             }
 
+            if (grammar != nullptr) {
+                llama_sample_grammar(ctx, &candidates_p, grammar);
+            }
+
             if (temp <= 0)
             {
                 // Greedy sampling
@@ -457,10 +493,15 @@ struct llama_server_context
                 }
             }
 
+            if (grammar != nullptr) {
+                llama_grammar_accept_token(ctx, grammar, result.tok);
+            }
+
             for (size_t i = 0; i < std::min(candidates_p.size, (size_t)n_probs); ++i)
             {
                 result.probs.push_back({candidates_p.data[i].id, candidates_p.data[i].p});
             }
+
             last_n_tokens.erase(last_n_tokens.begin());
             last_n_tokens.push_back(result.tok);
             num_tokens_predicted++;
@@ -631,6 +672,9 @@ static void server_print_usage(const char *argv0, const gpt_params &params,
     fprintf(stdout, "                        how to split tensors across multiple GPUs, comma-separated list of proportions, e.g. 3,1\n");
     fprintf(stdout, "  -mg i, --main-gpu i   the GPU to use for scratch and small tensors\n");
     fprintf(stdout, "  -lv, --low-vram don't allocate VRAM scratch buffer\n");
+    fprintf(stdout, "  -mmq, --mul-mat-q     use experimental mul_mat_q CUDA kernels instead of cuBLAS. TEMP!!!\n" );
+    fprintf(stdout, "                        Reduces VRAM usage by 700/970/1430 MiB for 7b/13b/33b but prompt processing speed\n" );
+    fprintf(stdout, "                        is still suboptimal, especially q2_K, q3_K, q5_K, and q6_K.\n" );
 #endif
     fprintf(stdout, "  -m FNAME, --model FNAME\n");
     fprintf(stdout, "                        model path (default: %s)\n", params.model.c_str());
@@ -827,7 +871,7 @@ static void server_params_parse(int argc, char **argv, server_params &sparams,
                 }
             }
 #else
-            LOG_WARNING("llama.cpp was compiled without cuBLAS. It is not possible to set a tensor split.", {});
+            LOG_WARNING("llama.cpp was compiled without cuBLAS. It is not possible to set a tensor split.\n", {});
 #endif // GGML_USE_CUBLAS
         }
         else if (arg == "--low-vram" || arg == "-lv")
@@ -835,7 +879,15 @@ static void server_params_parse(int argc, char **argv, server_params &sparams,
 #ifdef GGML_USE_CUBLAS
             params.low_vram = true;
 #else
-            fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS. It is not possible to set lower vram usage.\n");
+            LOG_WARNING("warning: llama.cpp was compiled without cuBLAS. It is not possible to set lower vram usage.\n", {});
+#endif // GGML_USE_CUBLAS
+        }
+        else if (arg == "--mul-mat-q" || arg == "-mmq")
+        {
+#ifdef GGML_USE_CUBLAS
+            params.mul_mat_q = true;
+#else
+            LOG_WARNING("warning: llama.cpp was compiled without cuBLAS. It is not possible to use mul_mat_q kernels.\n", {});
 #endif // GGML_USE_CUBLAS
         }
         else if (arg == "--main-gpu" || arg == "-mg")
@@ -936,6 +988,7 @@ static json format_generation_settings(llama_server_context &llama)
         {"stream", llama.stream},
         {"logit_bias", llama.params.logit_bias},
         {"n_probs", llama.params.n_probs},
+        {"grammar", llama.params.grammar},
     };
 }
 
@@ -1037,6 +1090,7 @@ static void parse_options_completion(const json &body, llama_server_context &lla
     llama.params.n_keep = body.value("n_keep", default_params.n_keep);
     llama.params.seed = body.value("seed", default_params.seed);
     llama.params.prompt = body.value("prompt", default_params.prompt);
+    llama.params.grammar = body.value("grammar", default_params.grammar);
     llama.params.n_probs = body.value("n_probs", default_params.n_probs);
 
     llama.params.logit_bias.clear();
@@ -1168,6 +1222,12 @@ int main(int argc, char **argv)
 
         parse_options_completion(json::parse(req.body), llama);
 
+        if (!llama.loadGrammar())
+        {
+            res.status = 400;
+            return;
+        }
+
         llama.loadPrompt();
         llama.beginCompletion();
 
@@ -1263,7 +1323,11 @@ int main(int argc, char **argv)
                 sink.done();
                 return true;
             };
-            res.set_chunked_content_provider("text/event-stream", chunked_content_provider);
+            const auto on_complete = [&](bool) {
+                llama.mutex.unlock();
+            };
+            lock.release();
+            res.set_chunked_content_provider("text/event-stream", chunked_content_provider, on_complete);
         } });
 
     svr.Get("/model.json", [&llama](const Request &, Response &res)
@@ -1319,8 +1383,12 @@ int main(int argc, char **argv)
 
     svr.set_error_handler([](const Request &, Response &res)
                           {
-        res.set_content("File Not Found", "text/plain");
-        res.status = 404; });
+        if (res.status == 400) {
+            res.set_content("Invalid request", "text/plain");
+        } else {
+            res.set_content("File Not Found", "text/plain");
+            res.status = 404;
+        } });
 
     // set timeouts and change hostname and port
     svr.set_read_timeout(sparams.read_timeout);
@@ -1348,6 +1416,9 @@ int main(int argc, char **argv)
         return 1;
     }
 
+    if (llama.grammar != nullptr) {
+        llama_grammar_free(llama.grammar);
+    }
     llama_backend_free();
 
     return 0;

examples/simple/simple.cpp

@@ -123,7 +123,7 @@ int main(int argc, char ** argv)
         // Evaluate the tokens :
         //---------------------------------
 
-        if ( llama_eval( ctx , tokens_list.data() , tokens_list.size() , llama_get_kv_cache_token_count( ctx ) , params.n_threads ) )
+        if ( llama_eval( ctx , tokens_list.data() , int(tokens_list.size()) , llama_get_kv_cache_token_count( ctx ) , params.n_threads ) )
         {
             fprintf( stderr,  "%s : failed to eval\n" , __func__ );
             return 1;

expose.h

@@ -30,6 +30,7 @@ struct load_model_inputs
     const int batch_size;
     const bool f16_kv;
     const bool low_vram;
+    const bool use_mmq;
     const char * executable_path;
     const char * model_filename;
     const char * lora_filename;
@@ -74,7 +75,7 @@ struct generation_inputs
 struct generation_outputs
 {
     int status = -1;
-    char text[16384]; //16kb should be enough for any response
+    char text[24576]; //24kb should be enough for any response
 };
 
 extern std::string executable_path;

ggml-alloc.c

@@ -0,0 +1,541 @@
+#include "ggml-alloc.h"
+#include "ggml.h"
+#include <assert.h>
+#include <stdarg.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+
+#define UNUSED(x) (void)(x)
+#define MAX(a, b) ((a) > (b) ? (a) : (b))
+
+//#define GGML_ALLOCATOR_DEBUG
+
+//#define AT_PRINTF printf
+#define AT_PRINTF(...) ((void)0)
+
+struct hash_node {
+    struct ggml_tensor * t;
+    int n_children;
+    int n_views;
+};
+
+static size_t hash(void * p) {
+    return (size_t)p % GGML_GRAPH_HASHTABLE_SIZE;
+}
+
+static struct hash_node * hash_get(struct hash_node hash_table[], struct ggml_tensor * t) {
+    size_t h = hash(t);
+
+    // linear probing
+    size_t i = h;
+    while (hash_table[i].t != NULL) {
+        if (hash_table[i].t == t) {
+            return &hash_table[i];
+        }
+        i = (i + 1) % GGML_GRAPH_HASHTABLE_SIZE;
+        if (i == h) {
+            // hash table is full
+            GGML_ASSERT(false);
+        }
+    }
+
+    hash_table[i].t = t;
+    return &hash_table[i];
+}
+
+// TODO: GGML_PAD ?
+static size_t aligned_offset(const void * buffer, size_t offset, size_t alignment) {
+    assert(alignment && !(alignment & (alignment - 1))); // power of 2
+    size_t align = (alignment - (((uintptr_t)buffer + offset) % alignment)) % alignment;
+    return offset + align;
+}
+
+struct free_block {
+    void * addr;
+    size_t size;
+};
+
+#define MAX_FREE_BLOCKS 256
+
+struct ggml_allocr {
+    void * data;
+    size_t size;
+    size_t alignment;
+    int n_free_blocks;
+    struct free_block free_blocks[MAX_FREE_BLOCKS];
+    struct hash_node hash_table[GGML_GRAPH_HASHTABLE_SIZE];
+    size_t max_size;
+    bool measure;
+
+#ifdef GGML_ALLOCATOR_DEBUG
+    struct ggml_tensor * allocated_tensors[1024];
+#endif
+};
+
+#ifdef GGML_ALLOCATOR_DEBUG
+static void add_allocated_tensor(struct ggml_allocator * alloc, struct ggml_tensor * tensor) {
+    for (int i = 0; i < 1024; i++) {
+        if (alloc->allocated_tensors[i] == NULL) {
+            alloc->allocated_tensors[i] = tensor;
+            return;
+        }
+    }
+    GGML_ASSERT(!"out of allocated_tensors");
+}
+static void remove_allocated_tensor(struct ggml_allocator * alloc, struct ggml_tensor * tensor) {
+    for (int i = 0; i < 1024; i++) {
+        if (alloc->allocated_tensors[i] == tensor ||
+            (alloc->allocated_tensors[i] != NULL && alloc->allocated_tensors[i]->data == tensor->data)) {
+            alloc->allocated_tensors[i] = NULL;
+            return;
+        }
+    }
+    printf("tried to free tensor %s not found\n", tensor->name);
+    GGML_ASSERT(!"tensor not found");
+}
+#endif
+
+
+static size_t ggml_allocator_get_alloc_size(struct ggml_allocr * alloc, struct ggml_tensor * tensor) {
+    return ggml_nbytes(tensor);
+
+    UNUSED(alloc);
+}
+
+void ggml_allocr_alloc(struct ggml_allocr * alloc, struct ggml_tensor * tensor) {
+    size_t size = ggml_allocator_get_alloc_size(alloc, tensor);
+    size = aligned_offset(NULL, size, alloc->alignment);
+
+    AT_PRINTF("%s: allocating %s (%zu bytes) - ", __func__, tensor->name, size);
+
+    size_t max_avail = 0;
+
+    // find the best fitting free block
+    int best_fit_block = -1;
+    size_t best_fit_size = SIZE_MAX;
+    for (int i = 0; i < alloc->n_free_blocks; i++) {
+        struct free_block * block = &alloc->free_blocks[i];
+        max_avail = MAX(max_avail, block->size);
+        if (block->size >= size && block->size <= best_fit_size) {
+            best_fit_block = i;
+            best_fit_size = block->size;
+        }
+    }
+
+    AT_PRINTF("block %d\n", best_fit_block);
+
+    if (best_fit_block == -1) {
+        fprintf(stderr, "%s: not enough space in the buffer (needed %zu, largest block available %zu)\n",
+                __func__, size, max_avail);
+        GGML_ASSERT(!"not enough space in the buffer");
+        return;
+    }
+    struct free_block * block = &alloc->free_blocks[best_fit_block];
+    void * addr = block->addr;
+    block->addr = (char*)block->addr + size;
+    block->size -= size;
+    if (block->size == 0) {
+        // remove block if empty
+        alloc->n_free_blocks--;
+        for (int j = best_fit_block; j < alloc->n_free_blocks; j++) {
+            alloc->free_blocks[j] = alloc->free_blocks[j+1];
+        }
+    }
+
+    tensor->data = addr;
+
+#ifdef GGML_ALLOCATOR_DEBUG
+    add_allocated_tensor(alloc, tensor);
+    size_t cur_max = (char*)addr - (char*)alloc->data + size;
+    if (cur_max > alloc->max_size) {
+        printf("max_size = %.2f MB: tensors: ", cur_max / 1024.0 / 1024.0);
+        for (int i = 0; i < 1024; i++) {
+            if (alloc->allocated_tensors[i]) {
+                printf("%s (%.2f MB) ", alloc->allocated_tensors[i]->name, ggml_nbytes(alloc->allocated_tensors[i]) / 1024.0 / 1024.0);
+            }
+        }
+        printf("\n");
+    }
+#endif
+
+    alloc->max_size = MAX(alloc->max_size, (char*)addr - (char*)alloc->data + size);
+}
+
+// this is a very naive implementation, but for our case the number of free blocks should be very small
+static void ggml_allocator_free_tensor(struct ggml_allocr * alloc, struct ggml_tensor * tensor) {
+    void * ptr = tensor->data;
+
+    if (ptr < alloc->data || (char*)ptr >= (char*)alloc->data + alloc->max_size) {
+        // the tensor was not allocated in this buffer
+        // this can happen because the graph allocator will try to free weights and other tensors from different buffers
+        // the easiest way to deal with this is just to ignore it
+        return;
+    }
+
+    size_t size = ggml_allocator_get_alloc_size(alloc, tensor);
+    size = aligned_offset(NULL, size, alloc->alignment);
+    AT_PRINTF("%s: freeing %s (%zu bytes) - n_free_blocks = %d\n", __func__, tensor->name, size, alloc->n_free_blocks);
+
+#ifdef GGML_ALLOCATOR_DEBUG
+    remove_allocated_tensor(alloc, tensor);
+#endif
+
+    // see if we can merge with an existing block
+    for (int i = 0; i < alloc->n_free_blocks; i++) {
+        struct free_block * block = &alloc->free_blocks[i];
+        // check if ptr is at the end of the block
+        if ((char*)block->addr + block->size == ptr) {
+            block->size += size;
+            // check if we can merge with the next block
+            if (i < alloc->n_free_blocks - 1 && (char*)block->addr + block->size == alloc->free_blocks[i+1].addr) {
+                block->size += alloc->free_blocks[i+1].size;
+                alloc->n_free_blocks--;
+                for (int j = i+1; j < alloc->n_free_blocks; j++) {
+                    alloc->free_blocks[j] = alloc->free_blocks[j+1];
+                }
+            }
+            return;
+        }
+        // check if ptr is at the beginning of the block
+        if ((char*)ptr + size == block->addr) {
+            block->addr = ptr;
+            block->size += size;
+            // check if we can merge with the previous block
+            if (i > 0 && (char*)alloc->free_blocks[i-1].addr + alloc->free_blocks[i-1].size == block->addr) {
+                alloc->free_blocks[i-1].size += block->size;
+                alloc->n_free_blocks--;
+                for (int j = i; j < alloc->n_free_blocks; j++) {
+                    alloc->free_blocks[j] = alloc->free_blocks[j+1];
+                }
+            }
+            return;
+        }
+    }
+    // otherwise, add a new block
+    GGML_ASSERT(alloc->n_free_blocks < MAX_FREE_BLOCKS && "out of free blocks");
+    // insert the new block in the correct position to keep the array sorted by address (to make merging blocks faster)
+    int insert_pos = 0;
+    while (insert_pos < alloc->n_free_blocks && alloc->free_blocks[insert_pos].addr < ptr) {
+        insert_pos++;
+    }
+    // shift all blocks from insert_pos onward to make room for the new block
+    for (int i = alloc->n_free_blocks; i > insert_pos; i--) {
+        alloc->free_blocks[i] = alloc->free_blocks[i-1];
+    }
+    // insert the new block
+    alloc->free_blocks[insert_pos].addr = ptr;
+    alloc->free_blocks[insert_pos].size = size;
+    alloc->n_free_blocks++;
+}
+
+void ggml_allocr_reset(struct ggml_allocr * alloc) {
+    alloc->n_free_blocks = 1;
+    size_t align_offset = aligned_offset(alloc->data, 0, alloc->alignment);
+    alloc->free_blocks[0].addr = (char *)alloc->data + align_offset;
+    alloc->free_blocks[0].size = alloc->size - align_offset;
+}
+
+struct ggml_allocr * ggml_allocr_new(void * data, size_t size, size_t alignment) {
+    struct ggml_allocr * alloc = (struct ggml_allocr *)malloc(sizeof(struct ggml_allocr) /* + n_free_blocks * sizeof(struct free_block) */);
+
+    *alloc = (struct ggml_allocr){
+        /*.data          = */ data,
+        /*.size          = */ size,
+        /*.alignment     = */ alignment,
+        /*.n_free_blocks = */ 0,
+        /*.free_blocks   = */ {{0}},
+        /*.hash_table    = */ {{0}},
+        /*.max_size      = */ 0,
+        /*.measure       = */ false,
+#ifdef GGML_ALLOCATOR_DEBUG
+        /*.allocated_tensors = */ = {0},
+#endif
+    };
+
+    ggml_allocr_reset(alloc);
+
+    return alloc;
+}
+
+// address and size of the buffer when measuring
+// it needs to be large enough to fit all the tensors, but it cannot overlap with other existing buffers
+static void * const MEASURE_BASE_ADDR = (void *) 0x1000;
+static const size_t MEASURE_MAX_SIZE  = 1ULL<<40; // 1 TB
+
+struct ggml_allocr * ggml_allocr_new_measure(size_t alignment) {
+    struct ggml_allocr * alloc = (struct ggml_allocr *)malloc(sizeof(struct ggml_allocr) /* + n_free_blocks * sizeof(struct free_block) */);
+
+    *alloc = (struct ggml_allocr){
+        /*.data          = */ MEASURE_BASE_ADDR,
+        /*.size          = */ MEASURE_MAX_SIZE,
+        /*.alignment     = */ alignment,
+        /*.n_free_blocks = */ 0,
+        /*.free_blocks   = */ {{0}},
+        /*.hash_table    = */ {{0}},
+        /*.max_size      = */ 0,
+        /*.measure       = */ true,
+#ifdef GGML_ALLOCATOR_DEBUG
+        /*.allocated_tensors = */ = {0},
+#endif
+    };
+
+    ggml_allocr_reset(alloc);
+
+    return alloc;
+}
+
+void ggml_allocr_free(struct ggml_allocr * alloc) {
+    free(alloc);
+}
+
+bool ggml_allocr_is_measure(struct ggml_allocr * alloc) {
+    return alloc->measure;
+}
+
+//////////// compute graph allocator
+
+static bool ggml_is_view(struct ggml_tensor * t) {
+    return t->op == GGML_OP_RESHAPE || t->op == GGML_OP_VIEW || t->op == GGML_OP_TRANSPOSE ||
+           t->op == GGML_OP_PERMUTE || t->op == GGML_OP_CPY;
+}
+
+static bool ggml_are_same_layout(const struct ggml_tensor * a, const struct ggml_tensor * b) {
+    if (a->type != b->type) {
+        return false;
+    }
+    for (int i = 0; i < GGML_MAX_DIMS; i++) {
+        if (a->ne[i] != b->ne[i]) {
+            return false;
+        }
+        if (a->nb[i] != b->nb[i]) {
+            return false;
+        }
+    }
+    return true;
+}
+
+static struct ggml_tensor * get_view_parent(struct ggml_tensor * t) {
+    switch (t->op) {
+        case GGML_OP_PERMUTE:
+        case GGML_OP_RESHAPE:
+        case GGML_OP_TRANSPOSE:
+        case GGML_OP_VIEW:
+            return t->src[0];
+        case GGML_OP_CPY:
+            return t->src[1];
+        default:
+            return NULL;
+    }
+}
+
+static struct ggml_tensor * get_view_source(struct ggml_tensor * t) {
+    struct ggml_tensor * parent = t;
+    do {
+        parent = get_view_parent(parent);
+    } while (ggml_is_view(parent));
+    return parent;
+}
+
+static bool ggml_op_can_inplace(enum ggml_op op) {
+    switch (op) {
+        case GGML_OP_SCALE:
+        case GGML_OP_DIAG_MASK_ZERO:
+        case GGML_OP_DIAG_MASK_INF:
+        case GGML_OP_ADD:
+        case GGML_OP_ADD1:
+        case GGML_OP_ACC:
+        case GGML_OP_SUB:
+        case GGML_OP_MUL:
+        case GGML_OP_DIV:
+        case GGML_OP_SQR:
+        case GGML_OP_SQRT:
+        case GGML_OP_LOG:
+        case GGML_OP_UNARY:
+        case GGML_OP_ROPE:
+        case GGML_OP_RMS_NORM:
+        case GGML_OP_SET:
+        case GGML_OP_SOFT_MAX:
+        case GGML_OP_CONT:
+            return true;
+
+        default:
+            return false;
+    }
+}
+
+static void allocate_node(struct ggml_allocr * alloc, struct ggml_tensor * node) {
+    struct hash_node * ht = alloc->hash_table;
+    if (node->data == NULL) {
+        if (ggml_is_view(node)) {
+            size_t offset;
+            switch(node->op) {
+                case GGML_OP_VIEW:
+                    memcpy(&offset, node->op_params, sizeof(size_t));
+                    node->data = (char *) node->src[0]->data + offset;
+                    break;
+                case GGML_OP_PERMUTE:
+                case GGML_OP_RESHAPE:
+                case GGML_OP_TRANSPOSE:
+                    node->data = node->src[0]->data;
+                    break;
+                case GGML_OP_CPY:
+                    node->data = node->src[1]->data;
+                    break;
+                default:
+                    GGML_ASSERT(!"unknown view op");
+                    break;
+            }
+        } else {
+            // see if we can reuse a parent's buffer (inplace)
+            if (ggml_op_can_inplace(node->op)) {
+                for (int i = 0; i < GGML_MAX_SRC; i++) {
+                    struct ggml_tensor * parent = node->src[i];
+                    if (parent == NULL) {
+                        break;
+                    }
+                    struct hash_node * p_hn = hash_get(ht, parent);
+                    if (parent->data != NULL && p_hn->n_children == 1 && p_hn->n_views == 0 && ggml_are_same_layout(node, parent)) {
+                        if (ggml_is_view(parent)) {
+                            struct ggml_tensor * view_src = get_view_source(parent);
+                            struct hash_node * view_src_hn = hash_get(ht, view_src);
+                            if (view_src_hn->n_views == 1 && view_src_hn->n_children == 0 && view_src->data == parent->data) {
+                                // TODO: the offset of the view parent must be kept to ensure that the op doesn't overwrite
+                                // the parent's data that it will need later (same layout requirement). the problem is that then
+                                // we cannot free the tensor because the original address of the allocation is lost.
+                                // adding a view_src pointer to the tensor would solve this and simplify the code dealing with views
+                                // for now, we only reuse the parent's data if the offset is zero (view_src->data == parent->data)
+                                AT_PRINTF("reusing view parent %s (%s) for %s\n", parent->name, view_src->name, node->name);
+                                node->data = parent->data;
+                                return;
+                            }
+                        }
+                        else {
+                            AT_PRINTF("reusing parent %s for %s\n", parent->name, node->name);
+                            node->data = parent->data;
+                        }
+                        return;
+                    }
+                }
+            }
+            ggml_allocr_alloc(alloc, node);
+        }
+    }
+}
+
+static size_t ggml_allocator_alloc_graph_tensors_n(
+    struct ggml_allocr * alloc,
+    struct ggml_cgraph ** graphs, int n_graphs,
+    struct ggml_tensor *** inputs, struct ggml_tensor *** outputs) {
+
+    // reset hash table
+    struct hash_node * ht = alloc->hash_table;
+    memset(ht, 0, sizeof(struct hash_node) * GGML_GRAPH_HASHTABLE_SIZE);
+
+    // count number of children and views
+    for (int g = 0; g < n_graphs; g++) {
+        struct ggml_cgraph * gf = graphs[g];
+        for (int i = 0; i < gf->n_nodes; i++) {
+            struct ggml_tensor * node = gf->nodes[i];
+
+            if (ggml_is_view(node)) {
+                struct ggml_tensor * view_src = get_view_source(node);
+                hash_get(ht, view_src)->n_views += 1;
+            }
+
+            for (int j = 0; j < GGML_MAX_SRC; j++) {
+                struct ggml_tensor * parent = node->src[j];
+                if (parent == NULL) {
+                    break;
+                }
+                hash_get(ht, parent)->n_children += 1;
+            }
+        }
+    }
+
+    // allocate tensors
+    for (int g = 0; g < n_graphs; g++) {
+        struct ggml_cgraph * gf = graphs[g];
+        AT_PRINTF("####### graph %d/%d\n", g, n_graphs);
+        // graph inputs are allocated first to ensure that they are not overwritten by each other
+        if (inputs != NULL && inputs[g] != NULL) {
+            for (int i = 0; inputs[g][i] != NULL; i++) {
+                struct ggml_tensor * input = inputs[g][i];
+                AT_PRINTF("input: %s\n", input->name);
+                allocate_node(alloc, input);
+            }
+        }
+        for (int i = 0; i < gf->n_nodes; i++) {
+            struct ggml_tensor * node = gf->nodes[i];
+
+            // allocate parents (leafs)
+            for (int j = 0; j < GGML_MAX_SRC; j++) {
+                struct ggml_tensor * parent = node->src[j];
+                if (parent == NULL) {
+                    break;
+                }
+                allocate_node(alloc, parent);
+            }
+
+            // allocate node
+            allocate_node(alloc, node);
+
+            AT_PRINTF("exec: %s (%s) <= ", ggml_op_name(node->op), node->name);
+            for (int j = 0; j < GGML_MAX_SRC; j++) {
+                struct ggml_tensor * parent = node->src[j];
+                if (parent == NULL) {
+                    break;
+                }
+                AT_PRINTF("%s", parent->name);
+                if (j < GGML_MAX_SRC - 1 && node->src[j + 1] != NULL) {
+                    AT_PRINTF(", ");
+                }
+            }
+            AT_PRINTF("\n");
+
+            // update parents
+            for (int j = 0; j < GGML_MAX_SRC; j++) {
+                struct ggml_tensor * parent = node->src[j];
+                if (parent == NULL) {
+                    break;
+                }
+                struct hash_node * p_hn = hash_get(ht, parent);
+                p_hn->n_children -= 1;
+
+                //AT_PRINTF("parent %s: %d children, %d views\n", parent->name, parent->n_children, parent->n_views);
+
+                if (p_hn->n_children == 0 && p_hn->n_views == 0) {
+                    if (ggml_is_view(parent)) {
+                        struct ggml_tensor * view_src = get_view_source(parent);
+                        struct hash_node * view_src_hn = hash_get(ht, view_src);
+                        view_src_hn->n_views -= 1;
+                        AT_PRINTF("view_src %s: %d children, %d views\n", view_src->name, view_src->n_children, view_src->n_views);
+                        if (view_src_hn->n_views == 0 && view_src_hn->n_children == 0 && view_src->data != node->data) {
+                            ggml_allocator_free_tensor(alloc, view_src);
+                        }
+                    }
+                    else {
+                        if (parent->data != node->data) {
+                            ggml_allocator_free_tensor(alloc, parent);
+                        }
+                    }
+                }
+            }
+            AT_PRINTF("\n");
+        }
+        // free graph outputs here that wouldn't be freed otherwise because they have no children
+        if (outputs != NULL && outputs[g] != NULL) {
+            for (int i = 0; outputs[g][i] != NULL; i++) {
+                struct ggml_tensor * output = outputs[g][i];
+                AT_PRINTF("output: %s\n", output->name);
+                ggml_allocator_free_tensor(alloc, output);
+            }
+        }
+    }
+
+    return alloc->max_size;
+}
+
+size_t ggml_allocr_alloc_graph(struct ggml_allocr * alloc, struct ggml_cgraph * graph) {
+    return ggml_allocator_alloc_graph_tensors_n(alloc, &graph, 1, NULL, NULL);
+}

ggml-alloc.h

@@ -0,0 +1,22 @@
+#pragma once
+
+#include "ggml.h"
+
+#ifdef  __cplusplus
+extern "C" {
+#endif
+
+
+GGML_API struct ggml_allocr * ggml_allocr_new(void * data, size_t size, size_t alignment);
+GGML_API struct ggml_allocr * ggml_allocr_new_measure(size_t alignment);
+
+GGML_API void   ggml_allocr_free(struct ggml_allocr * alloc);
+GGML_API bool   ggml_allocr_is_measure(struct ggml_allocr * alloc);
+GGML_API void   ggml_allocr_reset(struct ggml_allocr * alloc);
+GGML_API void   ggml_allocr_alloc(struct ggml_allocr * alloc, struct ggml_tensor * tensor);
+GGML_API size_t ggml_allocr_alloc_graph(struct ggml_allocr * alloc, struct ggml_cgraph * graph);
+
+
+#ifdef  __cplusplus
+}
+#endif

ggml-cuda.h

@@ -27,6 +27,7 @@ void   ggml_cuda_assign_buffers(struct ggml_tensor * tensor);
 void   ggml_cuda_assign_buffers_no_scratch(struct ggml_tensor * tensor);
 void   ggml_cuda_assign_buffers_force_inplace(struct ggml_tensor * tensor);
 void   ggml_cuda_set_main_device(int main_device);
+void   ggml_cuda_set_mul_mat_q(bool mul_mat_q);
 void   ggml_cuda_set_scratch_size(size_t scratch_size);
 void   ggml_cuda_free_scratch(void);
 bool   ggml_cuda_compute_forward(struct ggml_compute_params * params, struct ggml_tensor * tensor);

ggml-metal.m

@@ -7,6 +7,11 @@
 #import <Metal/Metal.h>
 #import <MetalPerformanceShaders/MetalPerformanceShaders.h>
 
+#undef MIN
+#undef MAX
+#define MIN(a, b) ((a) < (b) ? (a) : (b))
+#define MAX(a, b) ((a) > (b) ? (a) : (b))
+
 #ifdef GGML_METAL_NDEBUG
 #define metal_printf(...)
 #else
@@ -15,6 +20,8 @@
 
 #define UNUSED(x) (void)(x)
 
+#define GGML_MAX_CONCUR (2*GGML_MAX_NODES)
+
 struct ggml_metal_buffer {
     const char * name;
 
@@ -36,7 +43,7 @@ struct ggml_metal_context {
     int n_buffers;
     struct ggml_metal_buffer buffers[GGML_METAL_MAX_BUFFERS];
 
-    int concur_list[GGML_MAX_NODES];
+    int concur_list[GGML_MAX_CONCUR];
     int concur_list_len;
 
     // custom kernels
@@ -370,15 +377,15 @@ void ggml_metal_graph_find_concurrency(
         struct ggml_metal_context * ctx,
         struct ggml_cgraph * gf) {
     int search_depth = gf->n_nodes; //we only find concurrency in this range to avoid wasting too much time
-    int nodes_unused[GGML_MAX_NODES];
+    int nodes_unused[GGML_MAX_CONCUR];
 
-    for (int i = 0; i < GGML_MAX_NODES; i++) {ctx->concur_list[i] = 0;}
-    for (int i = 0; i < gf->n_nodes; i++) {nodes_unused[i] = 1;}
+    for (int i = 0; i < GGML_MAX_CONCUR; i++) { ctx->concur_list[i] = 0; }
+    for (int i = 0; i < gf->n_nodes;     i++) { nodes_unused[i]     = 1; }
     ctx->concur_list_len = 0;
 
-    int n_left = gf->n_nodes;
-    int n_start = 0; // all nodes before n_start at nodes_unused array have been sorted and store back to ctx->concur_list
-    int level_pos = 0;  // at ctx->concur_list, the last layer (level) ends at level_pos
+    int n_left    = gf->n_nodes;
+    int n_start   = 0; // all nodes before n_start at nodes_unused array have been sorted and store back to ctx->concur_list
+    int level_pos = 0; // at ctx->concur_list, the last layer (level) ends at level_pos
 
     while (n_left > 0) {
         // number of nodes at a layer (that can be issued concurrently)
@@ -386,28 +393,40 @@ void ggml_metal_graph_find_concurrency(
         for (int i = n_start; i < ((n_start + search_depth > gf->n_nodes) ? gf->n_nodes : n_start + search_depth); i++) {
             if (nodes_unused[i]) {
                 // if the requirements for gf->nodes[i] are satisfied
-                int exe_flag=1;
+                int exe_flag = 1;
+
                 // scan all srcs
                 for (int src_ind = 0; src_ind < GGML_MAX_SRC; src_ind++) {
                     struct ggml_tensor * src_cur = gf->nodes[i]->src[src_ind];
                     if (src_cur) {
                         // if is leaf nodes it's satisfied.
-                        if (src_cur->op == GGML_OP_NONE && src_cur->grad == NULL) {continue;}
+                        // TODO: ggml_is_leaf()
+                        if (src_cur->op == GGML_OP_NONE && src_cur->grad == NULL) {
+                            continue;
+                        }
 
                         // otherwise this src should be the output from previous nodes.
                         int is_found = 0;
+
                         // scan 2*search_depth back because we inserted barrier.
-                        for (int j = ((level_pos - 2*search_depth) < 0 ? 0 : (level_pos - 2*search_depth)); j < level_pos; j++) {
-                            if (gf->nodes[ctx->concur_list[j]] == src_cur) {is_found = 1; break;}
+                        //for (int j = ((level_pos - 2*search_depth) < 0 ? 0 : (level_pos - 2*search_depth)); j < level_pos; j++) {
+                        for (int j = MAX(0, level_pos - 2*search_depth); j < level_pos; j++) {
+                            if (ctx->concur_list[j] >= 0 && gf->nodes[ctx->concur_list[j]] == src_cur) {
+                                is_found = 1;
+                                break;
+                            }
+                        }
+                        if (is_found == 0) {
+                            exe_flag = 0;
+                            break;
                         }
-                        if (is_found == 0) {exe_flag = 0; break;}
                     }
                 }
                 if (exe_flag) {
                     // check if nodes[i]'s data will be overwritten by a node before nodes[i].
                     // if node[5] and node[3] write to the same memory region, then we can't issue node[5] before node[3]
                     int64_t data_start = (int64_t) gf->nodes[i]->data;
-                    int64_t length = (int64_t) ggml_nbytes(gf->nodes[i]);
+                    int64_t length     = (int64_t) ggml_nbytes(gf->nodes[i]);
                     for (int j = n_start; j < i; j++) {
                         if (nodes_unused[j] && gf->nodes[j]->op != GGML_OP_RESHAPE \
                                             && gf->nodes[j]->op != GGML_OP_VIEW \
@@ -416,9 +435,9 @@ void ggml_metal_graph_find_concurrency(
                             if (((int64_t)gf->nodes[j]->data) >= data_start + length || \
                                 ((int64_t)gf->nodes[j]->data) + (int64_t) ggml_nbytes(gf->nodes[j]) <= data_start) {
                                 continue;
-                            } else {
-                                exe_flag = 0;
                             }
+
+                            exe_flag = 0;
                         }
                     }
                 }
@@ -435,11 +454,13 @@ void ggml_metal_graph_find_concurrency(
         ctx->concur_list[level_pos + concurrency] = -1;
         ctx->concur_list_len++;
         // jump all sorted nodes at nodes_bak
-        while (!nodes_unused[n_start]) {n_start++;}
+        while (!nodes_unused[n_start]) {
+            n_start++;
+        }
         level_pos += concurrency + 1;
     }
 
-    if (ctx->concur_list_len > GGML_MAX_NODES) {
+    if (ctx->concur_list_len > GGML_MAX_CONCUR) {
         fprintf(stderr, "%s: too many elements for metal ctx->concur_list!\n", __func__);
     }
 }
@@ -453,7 +474,7 @@ void ggml_metal_graph_compute(
     // else fallback to serial dispatch
     MTLComputePassDescriptor * edesc = MTLComputePassDescriptor.computePassDescriptor;
 
-    const bool has_concur = ctx->concur_list_len && ctx->concur_list_len <= GGML_MAX_NODES;
+    const bool has_concur = ctx->concur_list_len && ctx->concur_list_len <= GGML_MAX_CONCUR;
 
     const int n_nodes  = has_concur ? ctx->concur_list_len      : gf->n_nodes;
     edesc.dispatchType = has_concur ? MTLDispatchTypeConcurrent : MTLDispatchTypeSerial;
@@ -718,7 +739,8 @@ void ggml_metal_graph_compute(
                             // TODO: needs to be updated after PR: https://github.com/ggerganov/ggml/pull/224
 
                             GGML_ASSERT(ne00 == ne10);
-                            GGML_ASSERT(ne02 == ne12);
+                            // GGML_ASSERT(ne02 == ne12); // Should be checked on individual data types until broadcast is implemented everywhere
+                            GGML_ASSERT(ne03 == ne13);
 
                             if (ggml_is_contiguous(src0) &&
                                 ggml_is_contiguous(src1) &&
@@ -746,11 +768,11 @@ void ggml_metal_graph_compute(
                                     initWithDevice:ctx->device transposeLeft:false transposeRight:true
                                         resultRows:ne11 resultColumns:ne01 interiorColumns:ne00 alpha:1.0 beta:0.0];
 
-                                // we need to do ne02 multiplications
+                                // we need to do ne12 multiplications
                                 // TODO: is there a way to do this in parallel - currently very slow ..
                                 // TODO: might be possible to offload part of the computation to ANE using Accelerate's CBLAS
-                                for (int64_t i02 = 0; i02 < ne02; ++i02) {
-                                    size_t offs_src0_cur = offs_src0 + i02*nb02;
+                                for (int64_t i02 = 0; i02 < ne12; ++i02) {
+                                    size_t offs_src0_cur = offs_src0 + i02/(ne12/ne02)*nb02; // gqa not used for now
                                     size_t offs_src1_cur = offs_src1 + i02*nb12;
                                     size_t offs_dst_cur  = offs_dst  + i02*nb2;
 
@@ -772,8 +794,6 @@ void ggml_metal_graph_compute(
                                 switch (src0t) {
                                     case GGML_TYPE_F16:
                                         {
-                                            GGML_ASSERT(ne02 == ne12);
-
                                             nth0 = 64;
                                             nth1 = 1;
                                             [encoder setComputePipelineState:ctx->pipeline_mul_mat_f16_f32];
@@ -853,16 +873,18 @@ void ggml_metal_graph_compute(
                                 [encoder setBuffer:id_dst  offset:offs_dst  atIndex:2];
                                 [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:3];
                                 [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:4];
-                                [encoder setBytes:&nb00 length:sizeof(nb00) atIndex:5];
-                                [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:6];
-                                [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:7];
-                                [encoder setBytes:&ne10 length:sizeof(ne10) atIndex:8];
-                                [encoder setBytes:&ne11 length:sizeof(ne11) atIndex:9];
-                                [encoder setBytes:&nb10 length:sizeof(nb10) atIndex:10];
-                                [encoder setBytes:&nb11 length:sizeof(nb11) atIndex:11];
-                                [encoder setBytes:&nb12 length:sizeof(nb12) atIndex:12];
-                                [encoder setBytes:&ne0  length:sizeof(ne0)  atIndex:13];
-                                [encoder setBytes:&ne1  length:sizeof(ne1)  atIndex:14];
+                                [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:5];
+                                [encoder setBytes:&nb00 length:sizeof(nb00) atIndex:6];
+                                [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:7];
+                                [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:8];
+                                [encoder setBytes:&ne10 length:sizeof(ne10) atIndex:9];
+                                [encoder setBytes:&ne11 length:sizeof(ne11) atIndex:10];
+                                [encoder setBytes:&ne12 length:sizeof(ne12) atIndex:11];
+                                [encoder setBytes:&nb10 length:sizeof(nb10) atIndex:12];
+                                [encoder setBytes:&nb11 length:sizeof(nb11) atIndex:13];
+                                [encoder setBytes:&nb12 length:sizeof(nb12) atIndex:14];
+                                [encoder setBytes:&ne0  length:sizeof(ne0)  atIndex:15];
+                                [encoder setBytes:&ne1  length:sizeof(ne1)  atIndex:16];
 
                                 if (src0t == GGML_TYPE_Q4_0 || src0t == GGML_TYPE_Q4_1 ||
                                     src0t == GGML_TYPE_Q2_K || src0t == GGML_TYPE_Q4_K) {

ggml-metal.metal

@@ -509,11 +509,13 @@ kernel void kernel_mul_mat_f16_f32(
         device       float * dst,
         constant   int64_t & ne00,
         constant   int64_t & ne01,
+        constant   int64_t & ne02,
         constant  uint64_t & nb00,
         constant  uint64_t & nb01,
         constant  uint64_t & nb02,
         constant   int64_t & ne10,
         constant   int64_t & ne11,
+        constant   int64_t & ne12,
         constant  uint64_t & nb10,
         constant  uint64_t & nb11,
         constant  uint64_t & nb12,
@@ -529,7 +531,7 @@ kernel void kernel_mul_mat_f16_f32(
     const int64_t r1 = tgpig.y;
     const int64_t im = tgpig.z;
 
-    device const half  * x = (device const half  *) (src0 + r0*nb01 + im*nb02);
+    device const half  * x = (device const half  *) (src0 + r0*nb01 + im/(ne12/ne02)*nb02);
     device const float * y = (device const float *) (src1 + r1*nb11 + im*nb12);
 
     sum[tpitg.x] = 0.0f;
@@ -552,6 +554,7 @@ kernel void kernel_mul_mat_f16_f32(
     }
 }
 
+
 kernel void kernel_alibi_f32(
         device const float * src0,
         device       float * dst,

ggml.c

@@ -195,8 +195,8 @@ typedef void * thread_ret_t;
 #define GGML_ALIGNED_MALLOC(size)  _aligned_malloc(size, GGML_MEM_ALIGN)
 #define GGML_ALIGNED_FREE(ptr)     _aligned_free(ptr)
 #else
-inline static void* ggml_aligned_malloc(size_t size) {
-    void* aligned_memory = NULL;
+inline static void * ggml_aligned_malloc(size_t size) {
+    void * aligned_memory = NULL;
 #ifdef GGML_USE_METAL
     int result = posix_memalign(&aligned_memory, getpagesize(), size);
 #else
@@ -3812,7 +3812,7 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
     "CROSS_ENTROPY_LOSS_BACK",
 };
 
-static_assert(GGML_OP_COUNT == 59, "GGML_OP_COUNT != 59");
+static_assert(GGML_OP_COUNT == 62, "GGML_OP_COUNT != 62");
 
 static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
     "none",
@@ -3884,7 +3884,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
     "cross_entropy_loss_back(x,y)",
 };
 
-static_assert(GGML_OP_COUNT == 59, "GGML_OP_COUNT != 59");
+static_assert(GGML_OP_COUNT == 62, "GGML_OP_COUNT != 62");
 
 static_assert(GGML_OP_POOL_COUNT == 2, "GGML_OP_POOL_COUNT != 2");
 
@@ -4072,8 +4072,8 @@ bool ggml_is_numa(void) {
 ////////////////////////////////////////////////////////////////////////////////
 
 void ggml_print_object(const struct ggml_object * obj) {
-    GGML_PRINT(" - ggml_object: offset = %zu, size = %zu, next = %p\n",
-            obj->offs, obj->size, (const void *) obj->next);
+    GGML_PRINT(" - ggml_object: type = %d, offset = %zu, size = %zu, next = %p\n",
+            obj->type, obj->offs, obj->size, (const void *) obj->next);
 }
 
 void ggml_print_objects(const struct ggml_context * ctx) {
@@ -4111,7 +4111,7 @@ size_t ggml_nbytes(const struct ggml_tensor * tensor) {
     //
     // is enough, but just in case, adding the second part
 
-    return MAX(tensor->ne[3]*tensor->nb[3], (ggml_nelements(tensor)*GGML_TYPE_SIZE[tensor->type])/GGML_BLCK_SIZE[tensor->type]);
+    return GGML_PAD(MAX(tensor->ne[3]*tensor->nb[3], (ggml_nelements(tensor)*GGML_TYPE_SIZE[tensor->type])/GGML_BLCK_SIZE[tensor->type]), GGML_MEM_ALIGN);
 }
 
 size_t ggml_nbytes_split(const struct ggml_tensor * tensor, int nrows_split) {
@@ -4213,7 +4213,7 @@ enum ggml_type ggml_ftype_to_ggml_type(enum ggml_ftype ftype) {
 }
 
 size_t ggml_tensor_overhead(void) {
-    return GGML_OBJECT_SIZE + GGML_TENSOR_SIZE + 16;
+    return GGML_OBJECT_SIZE + GGML_TENSOR_SIZE;
 }
 
 bool ggml_is_transposed(const struct ggml_tensor * tensor) {
@@ -4254,7 +4254,7 @@ static inline bool ggml_is_padded_1d(const struct ggml_tensor * tensor) {
         tensor->nb[3] == tensor->nb[2]*tensor->ne[2];
 }
 
-static inline bool ggml_are_same_shape(const struct ggml_tensor * t0, const struct ggml_tensor * t1) {
+bool ggml_are_same_shape(const struct ggml_tensor * t0, const struct ggml_tensor * t1) {
     static_assert(GGML_MAX_DIMS == 4, "GGML_MAX_DIMS is not 4 - update this function");
 
     return
@@ -4384,7 +4384,7 @@ struct ggml_context * ggml_init(struct ggml_init_params params) {
         return NULL;
     }
 
-    const size_t mem_size = (params.mem_size + GGML_MEM_ALIGN - 1) & ~(GGML_MEM_ALIGN - 1);
+    const size_t mem_size = params.mem_buffer ? params.mem_size : GGML_PAD(params.mem_size, GGML_MEM_ALIGN);
 
     *ctx = (struct ggml_context) {
         /*.mem_size           =*/ mem_size,
@@ -4473,12 +4473,14 @@ size_t ggml_get_max_tensor_size(const struct ggml_context * ctx) {
     struct ggml_object * obj = ctx->objects_begin;
 
     while (obj != NULL) {
-        struct ggml_tensor * tensor = (struct ggml_tensor *) ((char *) ctx->mem_buffer + obj->offs);
+        if (obj->type == GGML_OBJECT_TENSOR) {
+            struct ggml_tensor * tensor = (struct ggml_tensor *) ((char *) ctx->mem_buffer + obj->offs);
 
-        const size_t size = ggml_nbytes(tensor);
+            const size_t size = ggml_nbytes(tensor);
 
-        if (max_size < size) {
-            max_size = size;
+            if (max_size < size) {
+                max_size = size;
+            }
         }
 
         obj = obj->next;
@@ -4510,12 +4512,7 @@ static void ggml_scratch_load(struct ggml_context * ctx) {
 
 ////////////////////////////////////////////////////////////////////////////////
 
-static struct ggml_tensor * ggml_new_tensor_impl(
-        struct ggml_context * ctx,
-        enum   ggml_type type,
-        int    n_dims,
-        const int64_t* ne,
-        void*  data) {
+static struct ggml_object * ggml_new_object(struct ggml_context * ctx, enum ggml_object_type type, size_t size) {
     // always insert objects at the end of the context's memory pool
     struct ggml_object * obj_cur = ctx->objects_end;
 
@@ -4523,77 +4520,81 @@ static struct ggml_tensor * ggml_new_tensor_impl(
     const size_t cur_size = obj_cur == NULL ? 0 : obj_cur->size;
     const size_t cur_end  = cur_offs + cur_size;
 
-    size_t size_needed = 0;
-
-    if (data == NULL && !ctx->no_alloc) {
-        size_needed += GGML_TYPE_SIZE[type]*(ne[0]/GGML_BLCK_SIZE[type]);
-        for (int i = 1; i < n_dims; i++) {
-            size_needed *= ne[i];
-        }
-        // align to GGML_MEM_ALIGN
-        size_needed = ((size_needed + GGML_MEM_ALIGN - 1)/GGML_MEM_ALIGN)*GGML_MEM_ALIGN;
-    }
+    // align to GGML_MEM_ALIGN
+    size_t size_needed = GGML_PAD(size, GGML_MEM_ALIGN);
 
     char * const mem_buffer = ctx->mem_buffer;
     struct ggml_object * const obj_new = (struct ggml_object *)(mem_buffer + cur_end);
 
-    if (ctx->scratch.data == NULL || data != NULL) {
-        size_needed += GGML_TENSOR_SIZE;
+    if (cur_end + size_needed + GGML_OBJECT_SIZE > ctx->mem_size) {
+        GGML_PRINT("%s: not enough space in the context's memory pool (needed %zu, available %zu)\n",
+                __func__, cur_end + size_needed, ctx->mem_size);
+        assert(false);
+        return NULL;
+    }
 
-        if (cur_end + size_needed + GGML_OBJECT_SIZE > ctx->mem_size) {
-            GGML_PRINT("%s: not enough space in the context's memory pool (needed %zu, available %zu)\n",
-                    __func__, cur_end + size_needed + GGML_OBJECT_SIZE, ctx->mem_size);
-            assert(false);
-            return NULL;
-        }
+    *obj_new = (struct ggml_object) {
+        .offs = cur_end + GGML_OBJECT_SIZE,
+        .size = size_needed,
+        .next = NULL,
+        .type = type,
+    };
 
-        *obj_new = (struct ggml_object) {
-            .offs = cur_end + GGML_OBJECT_SIZE,
-            .size = size_needed,
-            .next = NULL,
-        };
+    ggml_assert_aligned(mem_buffer + obj_new->offs);
+
+    if (obj_cur != NULL) {
+        obj_cur->next = obj_new;
     } else {
-        if (ctx->scratch.offs + size_needed > ctx->scratch.size) {
-            GGML_PRINT("%s: not enough space in the scratch memory pool (needed %zu, available %zu)\n",
-                    __func__, ctx->scratch.offs + size_needed, ctx->scratch.size);
-            assert(false);
-            return NULL;
+        // this is the first object in this context
+        ctx->objects_begin = obj_new;
+    }
+
+    ctx->objects_end = obj_new;
+
+    //printf("%s: inserted new object at %zu, size = %zu\n", __func__, cur_end, obj_new->size);
+
+    return obj_new;
+}
+
+static struct ggml_tensor * ggml_new_tensor_impl(
+        struct ggml_context * ctx,
+        enum   ggml_type      type,
+        int                   n_dims,
+        const int64_t       * ne,
+        void                * data) {
+
+    assert(n_dims >= 1 && n_dims <= GGML_MAX_DIMS);
+
+    size_t data_size = 0;
+
+    if (data == NULL && !ctx->no_alloc) {
+        data_size += GGML_TYPE_SIZE[type]*(ne[0]/GGML_BLCK_SIZE[type]);
+        for (int i = 1; i < n_dims; i++) {
+            data_size *= ne[i];
         }
+    }
 
-        if (cur_end + GGML_TENSOR_SIZE + GGML_OBJECT_SIZE > ctx->mem_size) {
-            GGML_PRINT("%s: not enough space in the context's memory pool (needed %zu, available %zu)\n",
-                    __func__, cur_end + GGML_TENSOR_SIZE + GGML_OBJECT_SIZE, ctx->mem_size);
+    if (ctx->scratch.data != NULL && data == NULL) {
+        // allocate tensor data in the scratch buffer
+        if (ctx->scratch.offs + data_size > ctx->scratch.size) {
+            GGML_PRINT("%s: not enough space in the scratch memory pool (needed %zu, available %zu)\n",
+                    __func__, ctx->scratch.offs + data_size, ctx->scratch.size);
             assert(false);
             return NULL;
         }
 
         data = (char * const) ctx->scratch.data + ctx->scratch.offs;
 
-        *obj_new = (struct ggml_object) {
-            .offs = cur_end + GGML_OBJECT_SIZE,
-            .size = GGML_TENSOR_SIZE,
-            .next = NULL,
-        };
-
-        //printf("scratch offs = %zu, size_needed = %zu\n", ctx->scratch.offs, size_needed);
+        ctx->scratch.offs += data_size;
 
-        ctx->scratch.offs += size_needed;
+        data_size = 0;
     }
 
-    if (obj_cur != NULL) {
-        obj_cur->next = obj_new;
-    } else {
-        // this is the first object in this context
-        ctx->objects_begin = obj_new;
-    }
-
-    ctx->objects_end = obj_new;
+    struct ggml_object * const obj_new = ggml_new_object(ctx, GGML_OBJECT_TENSOR, GGML_TENSOR_SIZE + data_size);
 
-    //printf("%s: inserted new object at %zu, size = %zu\n", __func__, cur_end, obj_new->size);
-
-    struct ggml_tensor * const result = (struct ggml_tensor *)(mem_buffer + obj_new->offs);
+    // TODO: for recoverable errors, we would need to free the data allocated from the scratch buffer here
 
-    ggml_assert_aligned(result);
+    struct ggml_tensor * const result = (struct ggml_tensor *)((char *)ctx->mem_buffer + obj_new->offs);
 
     *result = (struct ggml_tensor) {
         /*.type         =*/ type,
@@ -4602,7 +4603,7 @@ static struct ggml_tensor * ggml_new_tensor_impl(
         /*.ne           =*/ { 1, 1, 1, 1 },
         /*.nb           =*/ { 0, 0, 0, 0 },
         /*.op           =*/ GGML_OP_NONE,
-        /*.op_params    =*/ {0},
+        /*.op_params    =*/ { 0 },
         /*.is_param     =*/ false,
         /*.grad         =*/ NULL,
         /*.src          =*/ { NULL },
@@ -4634,6 +4635,7 @@ static struct ggml_tensor * ggml_new_tensor_impl(
 }
 
 static void ggml_set_op_params(struct ggml_tensor * tensor, const void * params, size_t params_size) {
+    GGML_ASSERT(tensor != NULL); // silence -Warray-bounds warnings
     assert(params_size <= GGML_MAX_OP_PARAMS);
     memcpy(tensor->op_params, params, params_size);
 }
@@ -4650,22 +4652,22 @@ static void ggml_set_op_params_i32(struct ggml_tensor * tensor, uint32_t i, int3
 
 struct ggml_tensor * ggml_new_tensor(
         struct ggml_context * ctx,
-        enum   ggml_type type,
-        int    n_dims,
-        const int64_t * ne) {
+        enum   ggml_type      type,
+        int                   n_dims,
+        const int64_t       * ne) {
     return ggml_new_tensor_impl(ctx, type, n_dims, ne, NULL);
 }
 
 struct ggml_tensor * ggml_new_tensor_1d(
         struct ggml_context * ctx,
-        enum   ggml_type type,
+        enum   ggml_type      type,
         int64_t ne0) {
     return ggml_new_tensor(ctx, type, 1, &ne0);
 }
 
 struct ggml_tensor * ggml_new_tensor_2d(
         struct ggml_context * ctx,
-        enum   ggml_type type,
+        enum   ggml_type      type,
         int64_t ne0,
         int64_t ne1) {
     const int64_t ne[2] = { ne0, ne1 };
@@ -4674,7 +4676,7 @@ struct ggml_tensor * ggml_new_tensor_2d(
 
 struct ggml_tensor * ggml_new_tensor_3d(
         struct ggml_context * ctx,
-        enum   ggml_type type,
+        enum   ggml_type      type,
         int64_t ne0,
         int64_t ne1,
         int64_t ne2) {
@@ -4984,11 +4986,6 @@ enum ggml_unary_op ggml_get_unary_op(const struct ggml_tensor * tensor) {
     return (enum ggml_unary_op) ggml_get_op_params_i32(tensor, 0);
 }
 
-static void ggml_set_unary_op(struct ggml_tensor * tensor, enum ggml_unary_op op) {
-    GGML_ASSERT(tensor->op = GGML_OP_UNARY);
-    ggml_set_op_params_i32(tensor, 0, (int32_t) op);
-}
-
 const char * ggml_get_name(const struct ggml_tensor * tensor) {
     return tensor->name;
 }
@@ -5027,9 +5024,11 @@ struct ggml_tensor * ggml_get_tensor(struct ggml_context * ctx, const char * nam
     char * const mem_buffer = ctx->mem_buffer;
 
     while (obj != NULL) {
-        struct ggml_tensor * cur = (struct ggml_tensor *)(mem_buffer + obj->offs);
-        if (strcmp(cur->name, name) == 0) {
-            return cur;
+        if (obj->type == GGML_OBJECT_TENSOR) {
+            struct ggml_tensor * cur = (struct ggml_tensor *)(mem_buffer + obj->offs);
+            if (strcmp(cur->name, name) == 0) {
+                return cur;
+            }
         }
 
         obj = obj->next;
@@ -6243,6 +6242,27 @@ struct ggml_tensor * ggml_reshape_4d(
 
 // ggml_view_1d
 
+static struct ggml_tensor * ggml_view_tensor_offset(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        int                   n_dims,
+        const int64_t       * ne,
+        size_t                offset) {
+    // don't calculate an offset from an unallocated tensor
+    void * data = NULL;
+    if (a->data != NULL) {
+        data = (char *) a->data + offset;
+    }
+
+    struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, n_dims, ne, data);
+
+    ggml_format_name(result, "%s (view)", a->name);
+
+    ggml_set_op_params(result, &offset, sizeof(offset));
+
+    return result;
+}
+
 struct ggml_tensor * ggml_view_1d(
         struct ggml_context * ctx,
         struct ggml_tensor  * a,
@@ -6255,10 +6275,7 @@ struct ggml_tensor * ggml_view_1d(
         is_node = true;
     }
 
-    struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, 1, &ne0, (char *) a->data + offset);
-    ggml_format_name(result, "%s (view)", a->name);
-
-    ggml_set_op_params(result, &offset, sizeof(offset));
+    struct ggml_tensor * result = ggml_view_tensor_offset(ctx, a, 1, &ne0, offset);
 
     result->op   = GGML_OP_VIEW;
     result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
@@ -6285,10 +6302,7 @@ struct ggml_tensor * ggml_view_2d(
 
     const int64_t ne[GGML_MAX_DIMS] = { ne0, ne1, 1, 1 };
 
-    struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, 2, ne, (char *) a->data + offset);
-    ggml_format_name(result, "%s (view)", a->name);
-
-    ggml_set_op_params(result, &offset, sizeof(offset));
+    struct ggml_tensor * result = ggml_view_tensor_offset(ctx, a, 2, ne, offset);
 
     result->nb[1] = nb1;
     result->nb[2] = result->nb[1]*ne1;
@@ -6321,10 +6335,7 @@ struct ggml_tensor * ggml_view_3d(
 
     const int64_t ne[GGML_MAX_DIMS] = { ne0, ne1, ne2, 1 };
 
-    struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, 3, ne, (char *) a->data + offset);
-    ggml_format_name(result, "%s (view)", a->name);
-
-    ggml_set_op_params(result, &offset, sizeof(offset));
+    struct ggml_tensor * result = ggml_view_tensor_offset(ctx, a, 3, ne, offset);
 
     result->nb[1] = nb1;
     result->nb[2] = nb2;
@@ -6359,10 +6370,7 @@ struct ggml_tensor * ggml_view_4d(
 
     const int64_t ne[GGML_MAX_DIMS] = { ne0, ne1, ne2, ne3 };
 
-    struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, 4, ne, (char *) a->data + offset);
-    ggml_format_name(result, "%s (view)", a->name);
-
-    ggml_set_op_params(result, &offset, sizeof(offset));
+    struct ggml_tensor * result = ggml_view_tensor_offset(ctx, a, 4, ne, offset);
 
     result->nb[1] = nb1;
     result->nb[2] = nb2;
@@ -6433,7 +6441,7 @@ struct ggml_tensor * ggml_permute(
     result->src[0] = a;
 
     int32_t params[] = { axis0, axis1, axis2, axis3 };
-    ggml_set_op_params(result, &params, sizeof(params));
+    ggml_set_op_params(result, params, sizeof(params));
 
     return result;
 }
@@ -6559,7 +6567,7 @@ static struct ggml_tensor * ggml_diag_mask_inf_impl(
     struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
 
     int32_t params[] = { n_past, inplace ? 1 : 0 };
-    ggml_set_op_params(result, &params, sizeof(params));
+    ggml_set_op_params(result, params, sizeof(params));
 
     result->op   = GGML_OP_DIAG_MASK_INF;
     result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
@@ -6599,7 +6607,7 @@ static struct ggml_tensor * ggml_diag_mask_zero_impl(
     struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
 
     int32_t params[] = { n_past, inplace ? 1 : 0 };
-    ggml_set_op_params(result, &params, sizeof(params));
+    ggml_set_op_params(result, params, sizeof(params));
 
     result->op   = GGML_OP_DIAG_MASK_ZERO;
     result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
@@ -6715,9 +6723,9 @@ static struct ggml_tensor * ggml_rope_impl(
     struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
 
     int32_t params[6] = { n_past, n_dims, mode, n_ctx };
-    memcpy(params + 4, &freq_base, sizeof(float));
+    memcpy(params + 4, &freq_base,  sizeof(float));
     memcpy(params + 5, &freq_scale, sizeof(float));
-    ggml_set_op_params(result, &params, sizeof(params));
+    ggml_set_op_params(result, params, sizeof(params));
 
     result->op   = GGML_OP_ROPE;
     result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
@@ -6746,6 +6754,18 @@ struct ggml_tensor * ggml_rope_inplace(
     return ggml_rope_impl(ctx, a, n_past, n_dims, mode, n_ctx, 10000.0f, 1.0f, true);
 }
 
+struct ggml_tensor * ggml_rope_custom(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        int                   n_past,
+        int                   n_dims,
+        int                   mode,
+        int                   n_ctx,
+        float                 freq_base,
+        float                 freq_scale) {
+    return ggml_rope_impl(ctx, a, n_past, n_dims, mode, n_ctx, freq_base, freq_scale, false);
+}
+
 struct ggml_tensor * ggml_rope_custom_inplace(
         struct ggml_context * ctx,
         struct ggml_tensor  * a,
@@ -6779,7 +6799,7 @@ struct ggml_tensor * ggml_rope_back(
     struct ggml_tensor * result = ggml_dup_tensor(ctx, a);
 
     int32_t params[] = { n_past, n_dims, mode, n_ctx };
-    ggml_set_op_params(result, &params, sizeof(params));
+    ggml_set_op_params(result, params, sizeof(params));
 
     result->op   = GGML_OP_ROPE_BACK;
     result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
@@ -6810,7 +6830,7 @@ struct ggml_tensor * ggml_alibi(
 
     int32_t op_params[3] = { n_past, n_head };
     memcpy(op_params + 2, &bias_max, sizeof(float));
-    ggml_set_op_params(result, &op_params, sizeof(op_params));
+    ggml_set_op_params(result, op_params, sizeof(op_params));
 
     result->op   = GGML_OP_ALIBI;
     result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
@@ -6837,7 +6857,7 @@ struct ggml_tensor * ggml_clamp(
     struct ggml_tensor * result = ggml_view_tensor(ctx, a);
 
     float params[] = { min, max };
-    ggml_set_op_params(result, &params, sizeof(params));
+    ggml_set_op_params(result, params, sizeof(params));
 
     result->op   = GGML_OP_CLAMP;
     result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
@@ -6872,10 +6892,10 @@ GGML_API struct ggml_tensor * ggml_conv_1d(
         ggml_calc_conv_output_size(b->ne[0], a->ne[0], s0, p0, d0),
         a->ne[2], 1, 1,
     };
-    struct ggml_tensor* result = ggml_new_tensor(ctx, GGML_TYPE_F32, 2, ne);
+    struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 2, ne);
 
     int32_t params[] = { s0, p0, d0 };
-    ggml_set_op_params(result, &params, sizeof(params));
+    ggml_set_op_params(result, params, sizeof(params));
 
     result->op = GGML_OP_CONV_1D;
     result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
@@ -6887,10 +6907,10 @@ GGML_API struct ggml_tensor * ggml_conv_1d(
 
 // ggml_conv_2d
 
-struct ggml_tensor* ggml_conv_2d(
-    struct ggml_context* ctx,
-    struct ggml_tensor * a,
-    struct ggml_tensor * b,
+struct ggml_tensor * ggml_conv_2d(
+    struct ggml_context * ctx,
+    struct ggml_tensor  * a,
+    struct ggml_tensor  * b,
     int                  s0,
     int                  s1,
     int                  p0,
@@ -6911,10 +6931,10 @@ struct ggml_tensor* ggml_conv_2d(
         ggml_calc_conv_output_size(b->ne[1], a->ne[1], s1, p1, d1),
         a->ne[3], b->ne[3],
     };
-    struct ggml_tensor* result = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne);
+    struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne);
 
     int32_t params[] = { s0, s1, p0, p1, d0, d1 };
-    ggml_set_op_params(result, &params, sizeof(params));
+    ggml_set_op_params(result, params, sizeof(params));
 
     result->op = GGML_OP_CONV_2D;
     result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
@@ -6927,7 +6947,7 @@ struct ggml_tensor* ggml_conv_2d(
 
 // ggml_conv_1d_ph
 
-struct ggml_tensor* ggml_conv_1d_ph(
+struct ggml_tensor * ggml_conv_1d_ph(
         struct ggml_context * ctx,
         struct ggml_tensor  * a,
         struct ggml_tensor  * b,
@@ -6945,7 +6965,7 @@ static int64_t ggml_calc_pool_output_size(int64_t ins, int ks, int s, int p) {
 
 // ggml_pool_1d
 
-struct ggml_tensor* ggml_pool_1d(
+struct ggml_tensor * ggml_pool_1d(
         struct ggml_context * ctx,
         struct ggml_tensor  * a,
         enum ggml_op_pool     op,
@@ -6964,10 +6984,10 @@ struct ggml_tensor* ggml_pool_1d(
         ggml_calc_pool_output_size(a->ne[0], k0, s0, p0),
         a->ne[1],
     };
-    struct ggml_tensor* result = ggml_new_tensor(ctx, GGML_TYPE_F32, 2, ne);
+    struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 2, ne);
 
     int32_t params[] = { op, k0, s0, p0 };
-    ggml_set_op_params(result, &params, sizeof(params));
+    ggml_set_op_params(result, params, sizeof(params));
 
     result->op = GGML_OP_POOL_1D;
     result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
@@ -6978,7 +6998,7 @@ struct ggml_tensor* ggml_pool_1d(
 
 // ggml_pool_2d
 
-struct ggml_tensor* ggml_pool_2d(
+struct ggml_tensor * ggml_pool_2d(
         struct ggml_context * ctx,
         struct ggml_tensor  * a,
         enum ggml_op_pool     op,
@@ -7001,10 +7021,10 @@ struct ggml_tensor* ggml_pool_2d(
         ggml_calc_pool_output_size(a->ne[1], k1, s1, p1),
         a->ne[2],
     };
-    struct ggml_tensor* result = ggml_new_tensor(ctx, GGML_TYPE_F32, 3, ne);
+    struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 3, ne);
 
     int32_t params[] = { op, k0, k1, s0, s1, p0, p1 };
-    ggml_set_op_params(result, &params, sizeof(params));
+    ggml_set_op_params(result, params, sizeof(params));
 
     result->op = GGML_OP_POOL_2D;
     result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
@@ -7172,7 +7192,7 @@ struct ggml_tensor * ggml_win_part(
     struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne);
 
     int32_t params[] = { npx, npy, w };
-    ggml_set_op_params(result, &params, sizeof(params));
+    ggml_set_op_params(result, params, sizeof(params));
 
     result->op   = GGML_OP_WIN_PART;
     result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
@@ -7202,7 +7222,7 @@ struct ggml_tensor * ggml_win_unpart(
     struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 3, ne);
 
     int32_t params[] = { w };
-    ggml_set_op_params(result, &params, sizeof(params));
+    ggml_set_op_params(result, params, sizeof(params));
 
     result->op   = GGML_OP_WIN_UNPART;
     result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
@@ -7226,7 +7246,7 @@ static struct ggml_tensor * ggml_unary_impl(
 
     struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
 
-    ggml_set_unary_op(result, op);
+    ggml_set_op_params_i32(result, 0, (int32_t) op);
 
     result->op   = GGML_OP_UNARY;
     result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
@@ -7331,7 +7351,7 @@ struct ggml_tensor * ggml_map_binary_inplace_f32(
     return ggml_map_binary_impl_f32(ctx, a, b, fun, true);
 }
 
-// ggml_map_custom1
+// ggml_map_custom1_f32
 
 static struct ggml_tensor * ggml_map_custom1_impl_f32(
         struct ggml_context          * ctx,
@@ -7348,7 +7368,7 @@ static struct ggml_tensor * ggml_map_custom1_impl_f32(
 
     ggml_set_op_params(result, (const void *) &fun, sizeof(fun));
 
-    result->op = GGML_OP_MAP_CUSTOM1;
+    result->op = GGML_OP_MAP_CUSTOM1_F32;
     result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
     result->src[0] = a;
 
@@ -7369,7 +7389,7 @@ struct ggml_tensor * ggml_map_custom1_inplace_f32(
     return ggml_map_custom1_impl_f32(ctx, a, fun, true);
 }
 
-// ggml_map_custom2
+// ggml_map_custom2_f32
 
 static struct ggml_tensor * ggml_map_custom2_impl_f32(
         struct ggml_context          * ctx,
@@ -7387,7 +7407,7 @@ static struct ggml_tensor * ggml_map_custom2_impl_f32(
 
     ggml_set_op_params(result, (const void *) &fun, sizeof(fun));
 
-    result->op = GGML_OP_MAP_CUSTOM2;
+    result->op = GGML_OP_MAP_CUSTOM2_F32;
     result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
     result->src[0] = a;
     result->src[1] = b;
@@ -7411,7 +7431,7 @@ struct ggml_tensor * ggml_map_custom2_inplace_f32(
     return ggml_map_custom2_impl_f32(ctx, a, b, fun, true);
 }
 
-// ggml_map_custom3
+// ggml_map_custom3_f32
 
 static struct ggml_tensor * ggml_map_custom3_impl_f32(
         struct ggml_context          * ctx,
@@ -7430,7 +7450,7 @@ static struct ggml_tensor * ggml_map_custom3_impl_f32(
 
     ggml_set_op_params(result, (const void *) &fun, sizeof(fun));
 
-    result->op = GGML_OP_MAP_CUSTOM3;
+    result->op = GGML_OP_MAP_CUSTOM3_F32;
     result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
     result->src[0] = a;
     result->src[1] = b;
@@ -7457,6 +7477,190 @@ struct ggml_tensor * ggml_map_custom3_inplace_f32(
     return ggml_map_custom3_impl_f32(ctx, a, b, c, fun, true);
 }
 
+// ggml_map_custom1
+struct ggml_map_custom1_op_params {
+    ggml_custom1_op_t fun;
+    int n_tasks;
+    void * userdata;
+};
+
+static struct ggml_tensor * ggml_map_custom1_impl(
+        struct ggml_context          * ctx,
+        struct ggml_tensor           * a,
+        const  ggml_custom1_op_t       fun,
+        int                            n_tasks,
+        void                         * userdata,
+        bool                           inplace) {
+    GGML_ASSERT(n_tasks == GGML_N_TASKS_MAX || n_tasks > 0);
+
+    bool is_node = false;
+
+    if (!inplace && a->grad) {
+        is_node = true;
+    }
+
+    struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
+
+    struct ggml_map_custom1_op_params params = {
+        /*.fun      =*/ fun,
+        /*.n_tasks  =*/ n_tasks,
+        /*.userdata =*/ userdata
+    };
+    ggml_set_op_params(result, (const void *) &params, sizeof(params));
+
+    result->op = GGML_OP_MAP_CUSTOM1;
+    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
+    result->src[0] = a;
+
+    return result;
+}
+
+struct ggml_tensor * ggml_map_custom1(
+        struct ggml_context          * ctx,
+        struct ggml_tensor           * a,
+        const  ggml_custom1_op_t       fun,
+        int                            n_tasks,
+        void                         * userdata) {
+    return ggml_map_custom1_impl(ctx, a, fun, n_tasks, userdata, false);
+}
+
+struct ggml_tensor * ggml_map_custom1_inplace(
+        struct ggml_context          * ctx,
+        struct ggml_tensor           * a,
+        const  ggml_custom1_op_t       fun,
+        int                            n_tasks,
+        void                         * userdata) {
+    return ggml_map_custom1_impl(ctx, a, fun, n_tasks, userdata, true);
+}
+
+// ggml_map_custom2
+
+struct ggml_map_custom2_op_params {
+    ggml_custom2_op_t fun;
+    int n_tasks;
+    void * userdata;
+};
+
+static struct ggml_tensor * ggml_map_custom2_impl(
+        struct ggml_context          * ctx,
+        struct ggml_tensor           * a,
+        struct ggml_tensor           * b,
+        const  ggml_custom2_op_t       fun,
+        int                            n_tasks,
+        void                         * userdata,
+        bool                           inplace) {
+    GGML_ASSERT(n_tasks == GGML_N_TASKS_MAX || n_tasks > 0);
+
+    bool is_node = false;
+
+    if (!inplace && (a->grad || b->grad)) {
+        is_node = true;
+    }
+
+    struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
+
+    struct ggml_map_custom2_op_params params = {
+        /*.fun      =*/ fun,
+        /*.n_tasks  =*/ n_tasks,
+        /*.userdata =*/ userdata
+    };
+    ggml_set_op_params(result, (const void *) &params, sizeof(params));
+
+    result->op = GGML_OP_MAP_CUSTOM2;
+    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
+    result->src[0] = a;
+    result->src[1] = b;
+
+    return result;
+}
+
+struct ggml_tensor * ggml_map_custom2(
+        struct ggml_context          * ctx,
+        struct ggml_tensor           * a,
+        struct ggml_tensor           * b,
+        const  ggml_custom2_op_t       fun,
+        int                            n_tasks,
+        void                         * userdata) {
+    return ggml_map_custom2_impl(ctx, a, b, fun, n_tasks, userdata, false);
+}
+
+struct ggml_tensor * ggml_map_custom2_inplace(
+        struct ggml_context          * ctx,
+        struct ggml_tensor           * a,
+        struct ggml_tensor           * b,
+        const  ggml_custom2_op_t       fun,
+        int                            n_tasks,
+        void                         * userdata) {
+    return ggml_map_custom2_impl(ctx, a, b, fun, n_tasks, userdata, true);
+}
+
+// ggml_map_custom3
+
+struct ggml_map_custom3_op_params {
+    ggml_custom3_op_t fun;
+    int n_tasks;
+    void * userdata;
+};
+
+static struct ggml_tensor * ggml_map_custom3_impl(
+        struct ggml_context          * ctx,
+        struct ggml_tensor           * a,
+        struct ggml_tensor           * b,
+        struct ggml_tensor           * c,
+        const  ggml_custom3_op_t       fun,
+        int                            n_tasks,
+        void                         * userdata,
+        bool                           inplace) {
+    GGML_ASSERT(n_tasks == GGML_N_TASKS_MAX || n_tasks > 0);
+
+    bool is_node = false;
+
+    if (!inplace && (a->grad || b->grad || c->grad)) {
+        is_node = true;
+    }
+
+    struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
+
+    struct ggml_map_custom3_op_params params = {
+        /*.fun      =*/ fun,
+        /*.n_tasks  =*/ n_tasks,
+        /*.userdata =*/ userdata
+    };
+    ggml_set_op_params(result, (const void *) &params, sizeof(params));
+
+    result->op = GGML_OP_MAP_CUSTOM3;
+    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
+    result->src[0] = a;
+    result->src[1] = b;
+    result->src[2] = c;
+
+    return result;
+}
+
+struct ggml_tensor * ggml_map_custom3(
+        struct ggml_context          * ctx,
+        struct ggml_tensor           * a,
+        struct ggml_tensor           * b,
+        struct ggml_tensor           * c,
+        const  ggml_custom3_op_t       fun,
+        int                            n_tasks,
+        void                         * userdata) {
+    return ggml_map_custom3_impl(ctx, a, b, c, fun, n_tasks, userdata, false);
+}
+
+struct ggml_tensor * ggml_map_custom3_inplace(
+        struct ggml_context          * ctx,
+        struct ggml_tensor           * a,
+        struct ggml_tensor           * b,
+        struct ggml_tensor           * c,
+        const  ggml_custom3_op_t       fun,
+        int                            n_tasks,
+        void                         * userdata) {
+    return ggml_map_custom3_impl(ctx, a, b, c, fun, n_tasks, userdata, true);
+}
+
+
+
 // ggml_cross_entropy_loss
 
 struct ggml_tensor * ggml_cross_entropy_loss(
@@ -9265,8 +9469,8 @@ static void ggml_compute_forward_sum_rows_f32(
     for (int64_t i3 = 0; i3 < ne03; i3++) {
         for (int64_t i2 = 0; i2 < ne02; i2++) {
             for (int64_t i1 = 0; i1 < ne01; i1++) {
-                float* src_row = (float *) ((char *) src0->data + i1*nb01 + i2*nb02 + i3*nb03);
-                float* dst_row = (float *) ((char *) dst->data  + i1*nb1  + i2*nb2  + i3*nb3);
+                float * src_row = (float *) ((char *) src0->data + i1*nb01 + i2*nb02 + i3*nb03);
+                float * dst_row = (float *) ((char *) dst->data  + i1*nb1  + i2*nb2  + i3*nb3);
                 float row_sum = 0;
                 ggml_vec_sum_f32(ne00, &row_sum, src_row);
                 dst_row[0] = row_sum;
@@ -10523,71 +10727,95 @@ static void ggml_compute_forward_mul_mat(
         return;
     }
 
-    // parallelize by src0 rows
-    const int64_t dr = (ne01 + nth - 1)/nth;
+    const void * wdata    = (src1->type == vec_dot_type) ? src1->data : params->wdata;
+    const size_t row_size = ne10*GGML_TYPE_SIZE[vec_dot_type]/GGML_BLCK_SIZE[vec_dot_type];
 
-    const int64_t ir10 = dr*ith;
-    const int64_t ir11 = MIN(ir10 + dr, ne01);
+    const int64_t nr0 = ne01;           // src0 rows
+    const int64_t nr1 = ne11*ne12*ne13; // src1 rows
 
-    // src1 rows
-    const int64_t nr1 = ne11*ne12*ne13;
+    //printf("nr0 = %lld, nr1 = %lld\n", nr0, nr1);
 
-    const void * wdata    = (src1->type == vec_dot_type) ? src1->data : params->wdata;
-    const size_t row_size = ne10*GGML_TYPE_SIZE[vec_dot_type]/GGML_BLCK_SIZE[vec_dot_type];
+    // distribute the thread work across the inner or outer loop based on which one is larger
 
-    for (int64_t ir1 = 0; ir1 < nr1; ++ir1) {
-        const int64_t i13 = (ir1/(ne12*ne11));
-        const int64_t i12 = (ir1 - i13*ne12*ne11)/ne11;
-        const int64_t i11 = (ir1 - i13*ne12*ne11 - i12*ne11);
-
-        const int64_t ir0 = (ir1/ne11)%(ne02*ne03);
-        const int64_t i03 = (ir0/(ne02));
-        // Hack for "Falcon multi-query-attention key stutter" / alternative to ggml_repeat2.
-        // See https://github.com/ggerganov/llama.cpp/issues/1602#issuecomment-1606087470:
-        // GG: this is likely the correct way to broadcast, though need some more thought
-        //     therefore leaving the comments to remind us for now
-        const int64_t i02 = (i12 / (ne12 / ne02));
-        // Original from PR/224 (and also essential/correct for non-broadcast matmuls in Falcon)
-        // const int64_t i02 = (ir0 - i03*ne02);
-
-        const int64_t i1 = i11;
-        const int64_t i2 = i12;
-        const int64_t i3 = i13;
-
-        const char * src0_row = (const char *) src0->data + (  0 + i02*nb02 + i03*nb03     );
-
-        // desc: when src1 is not a contiguous memory block we have to calculate the offset using the strides
-        //       if it is, then we have either copied the data to params->wdata and made it contiguous or we are using
-        //       the original src1 data pointer, so we should index using the indices directly
-        // TODO: this is a bit of a hack, we should probably have a better way to handle this
-        const char * src1_col = (const char *) wdata +
-            (src1_cont || src1->type != vec_dot_type
-             ? (i11      + i12*ne11 + i13*ne12*ne11)*row_size
-             : (i11*nb11 + i12*nb12 + i13*nb13));
-
-        float * dst_col = (float *) ((char *) dst->data + (i1*nb1 + i2*nb2 + i3*nb3));
-
-        for (int64_t ir = ir10; ir < ir11; ++ir) {
-            vec_dot(ne00, &dst_col[ir], src0_row + ir*nb01, src1_col);
-        }
+    const int64_t nth0 = nr0 > nr1 ? nth : 1; // parallelize by src0 rows
+    const int64_t nth1 = nr0 > nr1 ? 1 : nth; // parallelize by src1 rows
+
+    const int64_t ith0 = ith % nth0;
+    const int64_t ith1 = ith / nth0;
+
+    const int64_t dr0 = (nr0 + nth0 - 1)/nth0;
+    const int64_t dr1 = (nr1 + nth1 - 1)/nth1;
+
+    const int64_t ir010 = dr0*ith0;
+    const int64_t ir011 = MIN(ir010 + dr0, nr0);
+
+    const int64_t ir110 = dr1*ith1;
+    const int64_t ir111 = MIN(ir110 + dr1, nr1);
+
+    //printf("ir010 = %6lld, ir011 = %6lld, ir110 = %6lld, ir111 = %6lld\n", ir010, ir011, ir110, ir111);
+
+    // threads with no work simply yield (not sure if it helps)
+    if (ir010 >= ir011 || ir110 >= ir111) {
+        sched_yield();
+        return;
     }
 
-    //int64_t t1 = ggml_time_us();
-    //static int64_t acc = 0;
-    //acc += t1 - t0;
-    //if (t1 - t0 > 10) {
-    //    printf("\n");
-    //    printf("ne00 = %5d, ne01 = %5d, ne02 = %5d, ne03 = %5d\n", ne00, ne01, ne02, ne03);
-    //    printf("nb00 = %5d, nb01 = %5d, nb02 = %5d, nb03 = %5d\n", nb00, nb01, nb02, nb03);
-    //    printf("ne10 = %5d, ne11 = %5d, ne12 = %5d, ne13 = %5d\n", ne10, ne11, ne12, ne13);
+    assert(ne12 % ne02 == 0);
+    assert(ne13 % ne03 == 0);
 
-    //    printf("XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX task %d/%d: %d us, acc = %d\n", ith, nth, (int) (t1 - t0), (int) acc);
-    //}
-}
+    // broadcast factors
+    const int64_t r2 = ne12/ne02;
+    const int64_t r3 = ne13/ne03;
 
+    // block-tiling attempt
+    const int64_t blck_0 = 16;
+    const int64_t blck_1 = 16;
 
-// ggml_compute_forward_out_prod
+    // attempt to reduce false-sharing (does not seem to make a difference)
+    float tmp[16];
+
+    for (int64_t iir1 = ir110; iir1 < ir111; iir1 += blck_1) {
+        for (int64_t iir0 = ir010; iir0 < ir011; iir0 += blck_0) {
+            for (int64_t ir1 = iir1; ir1 < iir1 + blck_1 && ir1 < ir111; ++ir1) {
+                const int64_t i13 = (ir1/(ne12*ne11));
+                const int64_t i12 = (ir1 - i13*ne12*ne11)/ne11;
+                const int64_t i11 = (ir1 - i13*ne12*ne11 - i12*ne11);
+
+                // broadcast src0 into src1
+                const int64_t i03 = i13/r3;
+                const int64_t i02 = i12/r2;
+
+                const int64_t i1 = i11;
+                const int64_t i2 = i12;
+                const int64_t i3 = i13;
+
+                const char * src0_row = (const char *) src0->data + (0 + i02*nb02 + i03*nb03);
 
+                // desc: when src1 is not a contiguous memory block we have to calculate the offset using the strides
+                //       if it is, then we have either copied the data to params->wdata and made it contiguous or we are using
+                //       the original src1 data pointer, so we should index using the indices directly
+                // TODO: this is a bit of a hack, we should probably have a better way to handle this
+                const char * src1_col = (const char *) wdata +
+                    (src1_cont || src1->type != vec_dot_type
+                     ? (i11      + i12*ne11 + i13*ne12*ne11)*row_size
+                     : (i11*nb11 + i12*nb12 + i13*nb13));
+
+                float * dst_col = (float *) ((char *) dst->data + (i1*nb1 + i2*nb2 + i3*nb3));
+
+                //for (int64_t ir0 = iir0; ir0 < iir0 + blck_0 && ir0 < ir011; ++ir0) {
+                //    vec_dot(ne00, &dst_col[ir0], src0_row + ir0*nb01, src1_col);
+                //}
+
+                for (int64_t ir0 = iir0; ir0 < iir0 + blck_0 && ir0 < ir011; ++ir0) {
+                    vec_dot(ne00, &tmp[ir0 - iir0], src0_row + ir0*nb01, src1_col);
+                }
+                memcpy(&dst_col[iir0], tmp, (MIN(iir0 + blck_0, ir011) - iir0)*sizeof(float));
+            }
+        }
+    }
+}
+
+// ggml_compute_forward_out_prod
 
 static void ggml_compute_forward_out_prod_f32(
         const struct ggml_compute_params * params,
@@ -12871,7 +13099,7 @@ static void ggml_compute_forward_pool_1d(
         const struct ggml_tensor * src0,
               struct ggml_tensor * dst) {
 
-    const int32_t* opts = (const int32_t*)dst->op_params;
+    const int32_t * opts = (const int32_t *)dst->op_params;
     enum ggml_op_pool op = opts[0];
     const int k0 = opts[1];
     const int s0 = opts[2];
@@ -14204,24 +14432,6 @@ static void ggml_compute_forward_map_custom1_f32(
     fun(dst, a);
 }
 
-
-static void ggml_compute_forward_map_custom1(
-        const struct ggml_compute_params * params,
-        const struct ggml_tensor * a,
-        struct ggml_tensor * dst,
-        const ggml_custom1_op_f32_t fun) {
-    switch (a->type) {
-        case GGML_TYPE_F32:
-            {
-                ggml_compute_forward_map_custom1_f32(params, a, dst, fun);
-            } break;
-        default:
-            {
-                GGML_ASSERT(false);
-            } break;
-    }
-}
-
 // ggml_compute_forward_map_custom2
 
 static void ggml_compute_forward_map_custom2_f32(
@@ -14240,24 +14450,6 @@ static void ggml_compute_forward_map_custom2_f32(
 }
 
 
-static void ggml_compute_forward_map_custom2(
-        const struct ggml_compute_params * params,
-        const struct ggml_tensor * a,
-        const struct ggml_tensor * b,
-        struct ggml_tensor * dst,
-        const ggml_custom2_op_f32_t fun) {
-    switch (a->type) {
-        case GGML_TYPE_F32:
-            {
-                ggml_compute_forward_map_custom2_f32(params, a, b, dst, fun);
-            } break;
-        default:
-            {
-                GGML_ASSERT(false);
-            } break;
-    }
-}
-
 // ggml_compute_forward_map_custom3
 
 static void ggml_compute_forward_map_custom3_f32(
@@ -14276,24 +14468,52 @@ static void ggml_compute_forward_map_custom3_f32(
     fun(dst, a, b, c);
 }
 
+// ggml_compute_forward_map_custom1
+
+static void ggml_compute_forward_map_custom1(
+        const struct ggml_compute_params * params,
+        const struct ggml_tensor * a,
+              struct ggml_tensor * dst) {
+    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
+        return;
+    }
+
+    struct ggml_map_custom1_op_params * p = (struct ggml_map_custom1_op_params *) dst->op_params;
+
+    p->fun(dst, a, params->ith, params->nth, p->userdata);
+}
+
+// ggml_compute_forward_map_custom2
+
+static void ggml_compute_forward_map_custom2(
+        const struct ggml_compute_params * params,
+        const struct ggml_tensor * a,
+        const struct ggml_tensor * b,
+              struct ggml_tensor * dst) {
+    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
+        return;
+    }
+
+    struct ggml_map_custom2_op_params * p = (struct ggml_map_custom2_op_params *) dst->op_params;
+
+    p->fun(dst, a, b, params->ith, params->nth, p->userdata);
+}
+
+// ggml_compute_forward_map_custom3
 
 static void ggml_compute_forward_map_custom3(
         const struct ggml_compute_params * params,
         const struct ggml_tensor * a,
         const struct ggml_tensor * b,
         const struct ggml_tensor * c,
-        struct ggml_tensor * dst,
-        const ggml_custom3_op_f32_t fun) {
-    switch (a->type) {
-        case GGML_TYPE_F32:
-            {
-                ggml_compute_forward_map_custom3_f32(params, a, b, c, dst, fun);
-            } break;
-        default:
-            {
-                GGML_ASSERT(false);
-            } break;
+              struct ggml_tensor * dst) {
+    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
+        return;
     }
+
+    struct ggml_map_custom3_op_params * p = (struct ggml_map_custom3_op_params *) dst->op_params;
+
+    p->fun(dst, a, b, c, params->ith, params->nth, p->userdata);
 }
 
 // ggml_compute_forward_cross_entropy_loss
@@ -14815,25 +15035,40 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
                 ggml_compute_forward_map_binary(params, tensor->src[0], tensor->src[1], tensor, fun);
             }
             break;
-        case GGML_OP_MAP_CUSTOM1:
+        case GGML_OP_MAP_CUSTOM1_F32:
             {
                 ggml_custom1_op_f32_t fun;
                 memcpy(&fun, tensor->op_params, sizeof(fun));
-                ggml_compute_forward_map_custom1(params, tensor->src[0], tensor, fun);
+                ggml_compute_forward_map_custom1_f32(params, tensor->src[0], tensor, fun);
             }
             break;
-        case GGML_OP_MAP_CUSTOM2:
+        case GGML_OP_MAP_CUSTOM2_F32:
             {
                 ggml_custom2_op_f32_t fun;
                 memcpy(&fun, tensor->op_params, sizeof(fun));
-                ggml_compute_forward_map_custom2(params, tensor->src[0], tensor->src[1], tensor, fun);
+                ggml_compute_forward_map_custom2_f32(params, tensor->src[0], tensor->src[1], tensor, fun);
             }
             break;
-        case GGML_OP_MAP_CUSTOM3:
+        case GGML_OP_MAP_CUSTOM3_F32:
             {
                 ggml_custom3_op_f32_t fun;
                 memcpy(&fun, tensor->op_params, sizeof(fun));
-                ggml_compute_forward_map_custom3(params, tensor->src[0], tensor->src[1], tensor->src[2], tensor, fun);
+                ggml_compute_forward_map_custom3_f32(params, tensor->src[0], tensor->src[1], tensor->src[2], tensor, fun);
+            }
+            break;
+        case GGML_OP_MAP_CUSTOM1:
+            {
+                ggml_compute_forward_map_custom1(params, tensor->src[0], tensor);
+            }
+            break;
+        case GGML_OP_MAP_CUSTOM2:
+            {
+                ggml_compute_forward_map_custom2(params, tensor->src[0], tensor->src[1], tensor);
+            }
+            break;
+        case GGML_OP_MAP_CUSTOM3:
+            {
+                ggml_compute_forward_map_custom3(params, tensor->src[0], tensor->src[1], tensor->src[2], tensor);
             }
             break;
         case GGML_OP_CROSS_ENTROPY_LOSS:
@@ -15641,6 +15876,9 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
             } break;
         case GGML_OP_MAP_UNARY:
         case GGML_OP_MAP_BINARY:
+        case GGML_OP_MAP_CUSTOM1_F32:
+        case GGML_OP_MAP_CUSTOM2_F32:
+        case GGML_OP_MAP_CUSTOM3_F32:
         case GGML_OP_MAP_CUSTOM1:
         case GGML_OP_MAP_CUSTOM2:
         case GGML_OP_MAP_CUSTOM3:
@@ -15825,6 +16063,35 @@ struct ggml_cgraph ggml_build_backward(struct ggml_context * ctx, struct ggml_cg
     return result;
 }
 
+struct ggml_cgraph * ggml_new_graph(struct ggml_context * ctx) {
+    struct ggml_object * obj = ggml_new_object(ctx, GGML_OBJECT_GRAPH, GGML_GRAPH_SIZE);
+    struct ggml_cgraph * cgraph = (struct ggml_cgraph *) ((char *) ctx->mem_buffer + obj->offs);
+
+    *cgraph = (struct ggml_cgraph) {
+        /*.n_nodes      =*/ 0,
+        /*.n_leafs      =*/ 0,
+        /*.nodes        =*/ { NULL },
+        /*.grads        =*/ { NULL },
+        /*.leafs        =*/ { NULL },
+        /*.hash_table   =*/ { NULL },
+        /*.perf_runs    =*/ 0,
+        /*.perf_cycles  =*/ 0,
+        /*.perf_time_us =*/ 0,
+    };
+
+    return cgraph;
+}
+
+struct ggml_cgraph * ggml_build_forward_ctx(struct ggml_context * ctx, struct ggml_tensor * tensor) {
+    struct ggml_cgraph * cgraph = ggml_new_graph(ctx);
+    ggml_build_forward_impl(cgraph, tensor, false);
+    return cgraph;
+}
+
+size_t ggml_graph_overhead(void) {
+    return GGML_OBJECT_SIZE + GGML_PAD(GGML_GRAPH_SIZE, GGML_MEM_ALIGN);
+}
+
 //
 // thread data
 //
@@ -16401,11 +16668,38 @@ struct ggml_cplan ggml_graph_plan(struct ggml_cgraph * cgraph, int n_threads) {
             case GGML_OP_WIN_UNPART:
             case GGML_OP_MAP_UNARY:
             case GGML_OP_MAP_BINARY:
+            case GGML_OP_MAP_CUSTOM1_F32:
+            case GGML_OP_MAP_CUSTOM2_F32:
+            case GGML_OP_MAP_CUSTOM3_F32:
+                {
+                    n_tasks = 1;
+                } break;
             case GGML_OP_MAP_CUSTOM1:
+                {
+                    struct ggml_map_custom1_op_params * p = (struct ggml_map_custom1_op_params *) node->op_params;
+                    if (p->n_tasks == GGML_N_TASKS_MAX) {
+                        n_tasks = n_threads;
+                    } else {
+                        n_tasks = MIN(p->n_tasks, n_threads);
+                    }
+                } break;
             case GGML_OP_MAP_CUSTOM2:
+                {
+                    struct ggml_map_custom2_op_params * p = (struct ggml_map_custom2_op_params *) node->op_params;
+                    if (p->n_tasks == GGML_N_TASKS_MAX) {
+                        n_tasks = n_threads;
+                    } else {
+                        n_tasks = MIN(p->n_tasks, n_threads);
+                    }
+                } break;
             case GGML_OP_MAP_CUSTOM3:
                 {
-                    n_tasks = 1;
+                    struct ggml_map_custom3_op_params * p = (struct ggml_map_custom3_op_params *) node->op_params;
+                    if (p->n_tasks == GGML_N_TASKS_MAX) {
+                        n_tasks = n_threads;
+                    } else {
+                        n_tasks = MIN(p->n_tasks, n_threads);
+                    }
                 } break;
             case GGML_OP_CROSS_ENTROPY_LOSS:
                 {
@@ -16544,10 +16838,9 @@ void ggml_graph_reset(struct ggml_cgraph * cgraph) {
 void ggml_graph_compute_with_ctx(struct ggml_context * ctx, struct ggml_cgraph * cgraph, int n_threads) {
     struct ggml_cplan cplan = ggml_graph_plan(cgraph, n_threads);
 
-    struct ggml_tensor * buf = ggml_new_tensor_1d(ctx, GGML_TYPE_I8, cplan.work_size);
-    GGML_ASSERT(buf);
+    struct ggml_object * obj = ggml_new_object(ctx, GGML_OBJECT_WORK_BUFFER, cplan.work_size);
 
-    cplan.work_data = buf->data;
+    cplan.work_data = (uint8_t *)ctx->mem_buffer + obj->offs;
 
     ggml_graph_compute(cgraph, &cplan);
 }

ggml.h

@@ -183,6 +183,15 @@
 #    define GGML_API
 #endif
 
+// TODO: support for clang
+#ifdef __GNUC__
+#    define GGML_DEPRECATED(func, hint) func __attribute__((deprecated(hint)))
+#elif defined(_MSC_VER)
+#    define GGML_DEPRECATED(func, hint) __declspec(deprecated(hint)) func
+#else
+#    define GGML_DEPRECATED(func, hint) func
+#endif
+
 #include <stdint.h>
 #include <stddef.h>
 #include <stdbool.h>
@@ -208,6 +217,8 @@
 
 #define GGML_UNUSED(x) (void)(x)
 
+#define GGML_PAD(x, n) (((x) + (n) - 1) & ~((n) - 1))
+
 #define GGML_ASSERT(x) \
     do { \
         if (!(x)) { \
@@ -372,6 +383,10 @@ extern "C" {
         GGML_OP_MAP_UNARY,
         GGML_OP_MAP_BINARY,
 
+        GGML_OP_MAP_CUSTOM1_F32,
+        GGML_OP_MAP_CUSTOM2_F32,
+        GGML_OP_MAP_CUSTOM3_F32,
+
         GGML_OP_MAP_CUSTOM1,
         GGML_OP_MAP_CUSTOM2,
         GGML_OP_MAP_CUSTOM3,
@@ -395,6 +410,12 @@ extern "C" {
         GGML_UNARY_OP_SILU,
     };
 
+    enum ggml_object_type {
+        GGML_OBJECT_TENSOR,
+        GGML_OBJECT_GRAPH,
+        GGML_OBJECT_WORK_BUFFER
+    };
+
     // ggml object
     struct ggml_object {
         size_t offs;
@@ -402,7 +423,9 @@ extern "C" {
 
         struct ggml_object * next;
 
-        char padding[8];
+        enum ggml_object_type type;
+
+        char padding[4];
     };
 
     static const size_t GGML_OBJECT_SIZE = sizeof(struct ggml_object);
@@ -423,7 +446,7 @@ extern "C" {
         enum ggml_op op;
 
         // op params - allocated as int32_t for alignment
-        int32_t op_params[GGML_MAX_OP_PARAMS / sizeof(uint32_t)];
+        int32_t op_params[GGML_MAX_OP_PARAMS / sizeof(int32_t)];
 
         bool is_param;
 
@@ -484,6 +507,8 @@ extern "C" {
         int64_t perf_time_us;
     };
 
+    static const size_t GGML_GRAPH_SIZE = sizeof(struct ggml_cgraph);
+
     // scratch buffer
     struct ggml_scratch {
         size_t offs;
@@ -558,6 +583,8 @@ extern "C" {
     GGML_API bool ggml_is_contiguous(const struct ggml_tensor * tensor);
     GGML_API bool ggml_is_permuted  (const struct ggml_tensor * tensor);
 
+    GGML_API bool ggml_are_same_shape(const struct ggml_tensor * t0, const struct ggml_tensor * t1);
+
     // use this to compute the memory overhead of a tensor
     GGML_API size_t ggml_tensor_overhead(void);
 
@@ -1158,7 +1185,18 @@ extern "C" {
             int                   mode,
             int                   n_ctx);
 
-    // custom RoPE, in-place, returns view(a)
+    // custom RoPE
+    GGML_API struct ggml_tensor * ggml_rope_custom(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            int                   n_past,
+            int                   n_dims,
+            int                   mode,
+            int                   n_ctx,
+            float                 freq_base,
+            float                 freq_scale);
+
+    // in-place, returns view(a)
     GGML_API struct ggml_tensor * ggml_rope_custom_inplace(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
@@ -1217,7 +1255,7 @@ extern "C" {
 
     // conv_1d with padding = half
     // alias for ggml_conv_1d(a, b, s, a->ne[0]/2, d)
-    GGML_API struct ggml_tensor* ggml_conv_1d_ph(
+    GGML_API struct ggml_tensor * ggml_conv_1d_ph(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
             struct ggml_tensor  * b,
@@ -1230,7 +1268,7 @@ extern "C" {
         GGML_OP_POOL_COUNT,
     };
 
-    GGML_API struct ggml_tensor* ggml_pool_1d(
+    GGML_API struct ggml_tensor * ggml_pool_1d(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
             enum ggml_op_pool     op,
@@ -1238,7 +1276,7 @@ extern "C" {
             int                   s0, // stride
             int                   p0); // padding
 
-    GGML_API struct ggml_tensor* ggml_pool_2d(
+    GGML_API struct ggml_tensor * ggml_pool_2d(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
             enum ggml_op_pool     op,
@@ -1292,15 +1330,6 @@ extern "C" {
             int                   h0,
             int                   w);
 
-    // custom operators
-
-    typedef void (*ggml_unary_op_f32_t) (const int, float *, const float *);
-    typedef void (*ggml_binary_op_f32_t)(const int, float *, const float *, const float *);
-
-    typedef void (*ggml_custom1_op_f32_t)(struct ggml_tensor *, const struct ggml_tensor *);
-    typedef void (*ggml_custom2_op_f32_t)(struct ggml_tensor *, const struct ggml_tensor *, const struct ggml_tensor *);
-    typedef void (*ggml_custom3_op_f32_t)(struct ggml_tensor *, const struct ggml_tensor *, const struct ggml_tensor *, const struct ggml_tensor *);
-
     GGML_API struct ggml_tensor * ggml_unary(
             struct ggml_context * ctx,
              struct ggml_tensor * a,
@@ -1311,63 +1340,138 @@ extern "C" {
         struct ggml_tensor  * a,
         enum ggml_unary_op op);
 
-    GGML_API struct ggml_tensor * ggml_map_unary_f32(
+    // custom operators
+
+    typedef void (*ggml_unary_op_f32_t) (const int, float *, const float *);
+    typedef void (*ggml_binary_op_f32_t)(const int, float *, const float *, const float *);
+
+    typedef void (*ggml_custom1_op_f32_t)(struct ggml_tensor *, const struct ggml_tensor *);
+    typedef void (*ggml_custom2_op_f32_t)(struct ggml_tensor *, const struct ggml_tensor *, const struct ggml_tensor *);
+    typedef void (*ggml_custom3_op_f32_t)(struct ggml_tensor *, const struct ggml_tensor *, const struct ggml_tensor *, const struct ggml_tensor *);
+
+    GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_unary_f32(
             struct ggml_context        * ctx,
             struct ggml_tensor         * a,
-                   ggml_unary_op_f32_t   fun);
+                   ggml_unary_op_f32_t   fun),
+        "use ggml_map_custom1 instead");
 
-    GGML_API struct ggml_tensor * ggml_map_unary_inplace_f32(
+    GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_unary_inplace_f32(
             struct ggml_context        * ctx,
             struct ggml_tensor         * a,
-                   ggml_unary_op_f32_t   fun);
+                   ggml_unary_op_f32_t   fun),
+        "use ggml_map_custom1_inplace instead");
 
-    GGML_API struct ggml_tensor * ggml_map_binary_f32(
+    GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_binary_f32(
             struct ggml_context         * ctx,
             struct ggml_tensor          * a,
             struct ggml_tensor          * b,
-                   ggml_binary_op_f32_t   fun);
+                   ggml_binary_op_f32_t   fun),
+        "use ggml_map_custom2 instead");
 
-    GGML_API struct ggml_tensor * ggml_map_binary_inplace_f32(
+    GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_binary_inplace_f32(
             struct ggml_context         * ctx,
             struct ggml_tensor          * a,
             struct ggml_tensor          * b,
-                   ggml_binary_op_f32_t   fun);
+                   ggml_binary_op_f32_t   fun),
+        "use ggml_map_custom2_inplace instead");
 
-    GGML_API struct ggml_tensor * ggml_map_custom1_f32(
+    GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_custom1_f32(
             struct ggml_context          * ctx,
             struct ggml_tensor           * a,
-                   ggml_custom1_op_f32_t   fun);
+                   ggml_custom1_op_f32_t   fun),
+        "use ggml_map_custom1 instead");
 
-    GGML_API struct ggml_tensor * ggml_map_custom1_inplace_f32(
+    GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_custom1_inplace_f32(
             struct ggml_context          * ctx,
             struct ggml_tensor           * a,
-                   ggml_custom1_op_f32_t   fun);
+                   ggml_custom1_op_f32_t   fun),
+        "use ggml_map_custom1_inplace instead");
 
-    GGML_API struct ggml_tensor * ggml_map_custom2_f32(
+    GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_custom2_f32(
             struct ggml_context          * ctx,
             struct ggml_tensor           * a,
             struct ggml_tensor           * b,
-                   ggml_custom2_op_f32_t   fun);
+                   ggml_custom2_op_f32_t   fun),
+        "use ggml_map_custom2 instead");
 
-    GGML_API struct ggml_tensor * ggml_map_custom2_inplace_f32(
+    GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_custom2_inplace_f32(
             struct ggml_context          * ctx,
             struct ggml_tensor           * a,
             struct ggml_tensor           * b,
-                   ggml_custom2_op_f32_t   fun);
+                   ggml_custom2_op_f32_t   fun),
+        "use ggml_map_custom2_inplace instead");
 
-    GGML_API struct ggml_tensor * ggml_map_custom3_f32(
+    GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_custom3_f32(
             struct ggml_context          * ctx,
             struct ggml_tensor           * a,
             struct ggml_tensor           * b,
             struct ggml_tensor           * c,
-                   ggml_custom3_op_f32_t   fun);
+                   ggml_custom3_op_f32_t   fun),
+        "use ggml_map_custom3 instead");
 
-    GGML_API struct ggml_tensor * ggml_map_custom3_inplace_f32(
+    GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_custom3_inplace_f32(
             struct ggml_context          * ctx,
             struct ggml_tensor           * a,
             struct ggml_tensor           * b,
             struct ggml_tensor           * c,
-                   ggml_custom3_op_f32_t   fun);
+                   ggml_custom3_op_f32_t   fun),
+        "use ggml_map_custom3_inplace instead");
+
+    // custom operators v2
+
+    typedef void (*ggml_custom1_op_t)(struct ggml_tensor * dst , const struct ggml_tensor * a, int ith, int nth, void * userdata);
+    typedef void (*ggml_custom2_op_t)(struct ggml_tensor * dst , const struct ggml_tensor * a, const struct ggml_tensor * b, int ith, int nth, void * userdata);
+    typedef void (*ggml_custom3_op_t)(struct ggml_tensor * dst , const struct ggml_tensor * a, const struct ggml_tensor * b, const struct ggml_tensor * c, int ith, int nth, void * userdata);
+
+    #define GGML_N_TASKS_MAX -1
+
+    GGML_API struct ggml_tensor * ggml_map_custom1(
+            struct ggml_context   * ctx,
+            struct ggml_tensor    * a,
+            ggml_custom1_op_t       fun,
+            int                     n_tasks,
+            void                  * userdata);
+
+    GGML_API struct ggml_tensor * ggml_map_custom1_inplace(
+            struct ggml_context   * ctx,
+            struct ggml_tensor    * a,
+            ggml_custom1_op_t       fun,
+            int                     n_tasks,
+            void                  * userdata);
+
+    GGML_API struct ggml_tensor * ggml_map_custom2(
+            struct ggml_context   * ctx,
+            struct ggml_tensor    * a,
+            struct ggml_tensor    * b,
+            ggml_custom2_op_t       fun,
+            int                     n_tasks,
+            void                  * userdata);
+
+    GGML_API struct ggml_tensor * ggml_map_custom2_inplace(
+            struct ggml_context   * ctx,
+            struct ggml_tensor    * a,
+            struct ggml_tensor    * b,
+            ggml_custom2_op_t       fun,
+            int                     n_tasks,
+            void                  * userdata);
+
+    GGML_API struct ggml_tensor * ggml_map_custom3(
+            struct ggml_context   * ctx,
+            struct ggml_tensor    * a,
+            struct ggml_tensor    * b,
+            struct ggml_tensor    * c,
+            ggml_custom3_op_t       fun,
+            int                     n_tasks,
+            void                  * userdata);
+
+    GGML_API struct ggml_tensor * ggml_map_custom3_inplace(
+            struct ggml_context   * ctx,
+            struct ggml_tensor    * a,
+            struct ggml_tensor    * b,
+            struct ggml_tensor    * c,
+            ggml_custom3_op_t       fun,
+            int                     n_tasks,
+            void                  * userdata);
 
     // loss function
 
@@ -1390,11 +1494,17 @@ extern "C" {
             struct ggml_context * ctx,
             struct ggml_tensor  * tensor);
 
+
     GGML_API void ggml_build_forward_expand(struct ggml_cgraph * cgraph, struct ggml_tensor * tensor);
 
     GGML_API struct ggml_cgraph ggml_build_forward (struct ggml_tensor * tensor);
     GGML_API struct ggml_cgraph ggml_build_backward(struct ggml_context * ctx, struct ggml_cgraph * gf, bool keep);
 
+    // graph allocation in a context
+    GGML_API struct ggml_cgraph * ggml_new_graph        (struct ggml_context * ctx);
+    GGML_API struct ggml_cgraph * ggml_build_forward_ctx(struct ggml_context * ctx, struct ggml_tensor * tensor);
+    GGML_API size_t ggml_graph_overhead(void);
+
     // ggml_graph_plan() has to be called before ggml_graph_compute()
     // when plan.work_size > 0, caller must allocate memory for plan.work_data
     GGML_API struct ggml_cplan ggml_graph_plan   (struct ggml_cgraph * cgraph, int n_threads /*= GGML_DEFAULT_N_THREADS*/);

gpttype_adapter.cpp

@@ -243,7 +243,7 @@ void sample_temperature(llama_token_data_array * candidates_p, float temp)
     if (temp <= 0)
     {
         // Imitate greedy sampling
-        temp = 0.01f; //cannot be zero else div0
+        temp = 0.00390625f; //cannot be zero else div0, this is 1/256
         llama_sample_temperature(nullptr, candidates_p, temp);
         llama_sample_top_k(nullptr, candidates_p, 1, 1); //only want first candidate
     }
@@ -347,6 +347,10 @@ ModelLoadResult gpttype_load_model(const load_model_inputs inputs, FileFormat in
     debugmode = inputs.debugmode;
     unbanTokens = inputs.unban_tokens;
     blasbatchsize = inputs.blasbatchsize;
+    if(blasbatchsize<=0)
+    {
+        blasbatchsize = 8;
+    }
     params.memory_f16 = inputs.f16_kv;
     params.n_ctx = inputs.max_context_length;
 
@@ -374,7 +378,8 @@ ModelLoadResult gpttype_load_model(const load_model_inputs inputs, FileFormat in
         else
         {
             //approximate NTK aware ctx
-            rope_freq_base = (params.n_ctx <= 3072 ? 26000.0f : (params.n_ctx <= 4096 ? 32000.0f : (params.n_ctx <= 6144 ? 54000.0f : 82684.0f)));
+            rope_freq_base = (params.n_ctx <= 3072 ? 26000.0f : (params.n_ctx <= 4096 ? 32000.0f : (params.n_ctx <= 6144 ? 54000.0f : (params.n_ctx <= 8192 ? 82684.0f : (params.n_ctx <= 12288 ? 140000.0f : 200000.0f)))));
+
         }
 
         printf("Using automatic RoPE scaling (scale:%.3f, base:%.1f)\n",rope_freq_scale,rope_freq_base);
@@ -466,6 +471,7 @@ ModelLoadResult gpttype_load_model(const load_model_inputs inputs, FileFormat in
         llama_ctx_params.seed = -1;
         llama_ctx_params.f16_kv = inputs.f16_kv;
         llama_ctx_params.low_vram = inputs.low_vram;
+        llama_ctx_params.mul_mat_q = inputs.use_mmq;
         llama_ctx_params.logits_all = false;
         llama_ctx_params.use_mmap = inputs.use_mmap;
         llama_ctx_params.use_mlock = inputs.use_mlock;

grammars/json.gbnf

@@ -1,29 +1,25 @@
-# Grammar for subset of JSON - doesn't support full string or number syntax
-
-root  ::= object
-value ::= object | array | string | number | boolean | "null"
+root   ::= object
+value  ::= object | array | string | number | ("true" | "false" | "null") ws
 
 object ::=
   "{" ws (
             string ":" ws value
     ("," ws string ":" ws value)*
-  )? "}"
+  )? "}" ws
 
 array  ::=
   "[" ws (
             value
     ("," ws value)*
-  )? "]"
+  )? "]" ws
 
-string  ::=
+string ::=
   "\"" (
     [^"\\] |
     "\\" (["\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F]) # escapes
   )* "\"" ws
 
-# Only plain integers currently
-number  ::= "-"? [0-9]+ ws
-boolean ::= ("true" | "false") ws
+number ::= ("-"? ([0-9] | [1-9] [0-9]*)) ("." [0-9]+)? ([eE] [-+]? [0-9]+)? ws
 
 # Optional space: by convention, applied in this grammar after literal chars when allowed
 ws ::= ([ \t\n] ws)?

k_quants.c

@@ -39,6 +39,8 @@
 #define MIN(a, b) ((a) < (b) ? (a) : (b))
 #define MAX(a, b) ((a) > (b) ? (a) : (b))
 
+#define MM256_SET_M128I(a, b) _mm256_insertf128_si256(_mm256_castsi128_si256(b), (a), 1)
+
 //
 // 2-6 bit quantization in super-blocks
 //
@@ -1353,7 +1355,7 @@ void ggml_vec_dot_q2_K_q8_K(const int n, float * restrict s, const void * restri
         const __m256i all_scales = _mm256_cvtepi8_epi16(scales8);
         const __m128i l_scales = _mm256_extracti128_si256(all_scales, 0);
         const __m128i h_scales = _mm256_extracti128_si256(all_scales, 1);
-        const __m256i scales[2] = {_mm256_set_m128i(l_scales, l_scales), _mm256_set_m128i(h_scales, h_scales)};
+        const __m256i scales[2] = {MM256_SET_M128I(l_scales, l_scales), MM256_SET_M128I(h_scales, h_scales)};
 
         __m256i sumi = _mm256_setzero_si256();
 
@@ -1421,7 +1423,7 @@ void ggml_vec_dot_q2_K_q8_K(const int n, float * restrict s, const void * restri
         const __m128i summs_1 = _mm_madd_epi16(mins_1, _mm_loadu_si128((const __m128i*)&y[i].bsums[8]));
 
         // sumf += -dmin * summs in 32bits*8
-        acc = _mm256_add_ps(_mm256_mul_ps(_mm256_broadcast_ss(&dmin), _mm256_cvtepi32_ps(_mm256_set_m128i(summs_1, summs_0))), acc);
+        acc = _mm256_add_ps(_mm256_mul_ps(_mm256_broadcast_ss(&dmin), _mm256_cvtepi32_ps(MM256_SET_M128I(summs_1, summs_0))), acc);
 
         const __m128i scales_0 = _mm_cvtepi8_epi16(scales16);
         const __m128i scales_1 = _mm_cvtepi8_epi16(_mm_unpackhi_epi64(scales16, scales16));
@@ -1493,7 +1495,7 @@ void ggml_vec_dot_q2_K_q8_K(const int n, float * restrict s, const void * restri
         }
 
         // sumf += dall * isum - dmin * summs in 32bits
-        __m256i sumi = _mm256_set_m128i(sumi_1, sumi_0);
+        __m256i sumi = MM256_SET_M128I(sumi_1, sumi_0);
         acc = _mm256_add_ps(_mm256_mul_ps(_mm256_broadcast_ss(&dall), _mm256_cvtepi32_ps(sumi)), acc);
     }
 
@@ -1644,8 +1646,8 @@ void ggml_vec_dot_q2_K_q8_K(const int n, float * restrict s, const void * restri
         summs += dmin * smin;
 
         const __m128i q2bits = _mm_loadu_si128((const __m128i*)q2);
-        const __m256i q2_0 = _mm256_and_si256(_mm256_set_m128i(_mm_srli_epi16(q2bits, 2), q2bits), m3);
-        const __m256i q2_1 = _mm256_and_si256(_mm256_set_m128i(_mm_srli_epi16(q2bits, 6), _mm_srli_epi16(q2bits, 4)), m3);
+        const __m256i q2_0 = _mm256_and_si256(MM256_SET_M128I(_mm_srli_epi16(q2bits, 2), q2bits), m3);
+        const __m256i q2_1 = _mm256_and_si256(MM256_SET_M128I(_mm_srli_epi16(q2bits, 6), _mm_srli_epi16(q2bits, 4)), m3);
 
         const __m256i q8_0 = _mm256_loadu_si256((const __m256i*)(q8+ 0));
         const __m256i q8_1 = _mm256_loadu_si256((const __m256i*)(q8+32));
@@ -1709,10 +1711,10 @@ void ggml_vec_dot_q2_K_q8_K(const int n, float * restrict s, const void * restri
         const __m128i p2 = _mm_maddubs_epi16(q2_2, _mm256_extractf128_si256(q8_1, 0));
         const __m128i p3 = _mm_maddubs_epi16(q2_3, _mm256_extractf128_si256(q8_1, 1));
 
-        const __m256i p_0 = _mm256_set_m128i(_mm_cvtepi16_epi32(_mm_unpackhi_epi64(p0, p0)), _mm_cvtepi16_epi32(p0));
-        const __m256i p_1 = _mm256_set_m128i(_mm_cvtepi16_epi32(_mm_unpackhi_epi64(p1, p1)), _mm_cvtepi16_epi32(p1));
-        const __m256i p_2 = _mm256_set_m128i(_mm_cvtepi16_epi32(_mm_unpackhi_epi64(p2, p2)), _mm_cvtepi16_epi32(p2));
-        const __m256i p_3 = _mm256_set_m128i(_mm_cvtepi16_epi32(_mm_unpackhi_epi64(p3, p3)), _mm_cvtepi16_epi32(p3));
+        const __m256i p_0 = MM256_SET_M128I(_mm_cvtepi16_epi32(_mm_unpackhi_epi64(p0, p0)), _mm_cvtepi16_epi32(p0));
+        const __m256i p_1 = MM256_SET_M128I(_mm_cvtepi16_epi32(_mm_unpackhi_epi64(p1, p1)), _mm_cvtepi16_epi32(p1));
+        const __m256i p_2 = MM256_SET_M128I(_mm_cvtepi16_epi32(_mm_unpackhi_epi64(p2, p2)), _mm_cvtepi16_epi32(p2));
+        const __m256i p_3 = MM256_SET_M128I(_mm_cvtepi16_epi32(_mm_unpackhi_epi64(p3, p3)), _mm_cvtepi16_epi32(p3));
 
         acc = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(d * db[0]), _mm256_cvtepi32_ps(p_0)), acc);
         acc = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(d * db[1]), _mm256_cvtepi32_ps(p_1)), acc);
@@ -1917,7 +1919,7 @@ void ggml_vec_dot_q3_K_q8_K(const int n, float * restrict s, const void * restri
         const __m256i all_scales = _mm256_cvtepi8_epi16(scales128);
         const __m128i l_scales = _mm256_extracti128_si256(all_scales, 0);
         const __m128i h_scales = _mm256_extracti128_si256(all_scales, 1);
-        const __m256i scales[2] = {_mm256_set_m128i(l_scales, l_scales), _mm256_set_m128i(h_scales, h_scales)};
+        const __m256i scales[2] = {MM256_SET_M128I(l_scales, l_scales), MM256_SET_M128I(h_scales, h_scales)};
 
         // high bit
         const __m256i hbits = _mm256_loadu_si256((const __m256i*)x[i].hmask);
@@ -2128,7 +2130,7 @@ void ggml_vec_dot_q3_K_q8_K(const int n, float * restrict s, const void * restri
         }
 
         // multiply with block scale and accumulate
-        __m256i sumi = _mm256_set_m128i(sumi_1, sumi_0);
+        __m256i sumi = MM256_SET_M128I(sumi_1, sumi_0);
         acc = _mm256_add_ps(_mm256_mul_ps(_mm256_broadcast_ss(&d), _mm256_cvtepi32_ps(sumi)), acc);
 
     }
@@ -2303,13 +2305,13 @@ void ggml_vec_dot_q3_K_q8_K(const int n, float * restrict s, const void * restri
         aux16[0] = a & 0x0f0f;
         aux16[1] = (a >> 4) & 0x0f0f;
 
-        const __m256i scale_0 = _mm256_set_m128i(_mm_set1_epi16(aux8[2] - 8), _mm_set1_epi16(aux8[0] - 8));
-        const __m256i scale_1 = _mm256_set_m128i(_mm_set1_epi16(aux8[3] - 8), _mm_set1_epi16(aux8[1] - 8));
+        const __m256i scale_0 = MM256_SET_M128I(_mm_set1_epi16(aux8[2] - 8), _mm_set1_epi16(aux8[0] - 8));
+        const __m256i scale_1 = MM256_SET_M128I(_mm_set1_epi16(aux8[3] - 8), _mm_set1_epi16(aux8[1] - 8));
 
         memcpy(&aux64, x[i].hmask, 8);
 
         const __m128i haux = _mm_set_epi64x(aux64 >> 1, aux64 >> 0);
-        __m256i q3h_0 = _mm256_set_m128i(_mm_srli_epi16(haux, 2), haux);
+        __m256i q3h_0 = MM256_SET_M128I(_mm_srli_epi16(haux, 2), haux);
         __m256i q3h_1 = _mm256_srli_epi16(q3h_0, 4);
         q3h_0 = _mm256_slli_epi16(_mm256_andnot_si256(q3h_0, m1), 2);
         q3h_1 = _mm256_slli_epi16(_mm256_andnot_si256(q3h_1, m1), 2);
@@ -2318,7 +2320,7 @@ void ggml_vec_dot_q3_K_q8_K(const int n, float * restrict s, const void * restri
         const __m128i q3bits = _mm_loadu_si128((const __m128i*)q3);
 
         // prepare low and high bits
-        const __m256i q3aux  = _mm256_set_m128i(_mm_srli_epi16(q3bits, 2), q3bits);
+        const __m256i q3aux  = MM256_SET_M128I(_mm_srli_epi16(q3bits, 2), q3bits);
         const __m256i q3l_0 = _mm256_and_si256(q3aux, m3);
         const __m256i q3l_1 = _mm256_and_si256(_mm256_srli_epi16(q3aux, 4), m3);
 
@@ -2429,7 +2431,7 @@ void ggml_vec_dot_q3_K_q8_K(const int n, float * restrict s, const void * restri
 
         p16_0 = _mm_add_epi32(p16_0, p16_2);
         p16_1 = _mm_add_epi32(p16_1, p16_3);
-        __m256i p16 = _mm256_set_m128i(p16_1, p16_0);
+        __m256i p16 = MM256_SET_M128I(p16_1, p16_0);
 
         // multiply with block scale and accumulate
         acc = _mm256_add_ps(_mm256_mul_ps(_mm256_broadcast_ss(&d), _mm256_cvtepi32_ps(p16)), acc);
@@ -2620,7 +2622,7 @@ void ggml_vec_dot_q4_K_q8_K(const int n, float * restrict s, const void * restri
         acc_m = _mm_fmadd_ps(_mm_set1_ps(dmin), _mm_cvtepi32_ps(prod), acc_m);
 
         const __m128i sc128  = _mm256_extracti128_si256(mins_and_scales, 0);
-        const __m256i scales = _mm256_set_m128i(sc128, sc128);
+        const __m256i scales = MM256_SET_M128I(sc128, sc128);
 
         __m256i sumi = _mm256_setzero_si256();
 
@@ -2727,7 +2729,7 @@ void ggml_vec_dot_q4_K_q8_K(const int n, float * restrict s, const void * restri
         }
 
         __m256 vd = _mm256_set1_ps(d);
-        __m256i sumi = _mm256_set_m128i(sumi_1, sumi_0);
+        __m256i sumi = MM256_SET_M128I(sumi_1, sumi_0);
         acc = _mm256_add_ps(_mm256_mul_ps(vd, _mm256_cvtepi32_ps(sumi)), acc);
 
     }
@@ -2968,11 +2970,11 @@ void ggml_vec_dot_q4_K_q8_K(const int n, float * restrict s, const void * restri
 
         const __m128i p32_0 = _mm_madd_epi16(_mm_set1_epi16(scales[0]), p16_0);
         const __m128i p32_1 = _mm_madd_epi16(_mm_set1_epi16(scales[0]), p16_1);
-        acc = _mm256_add_ps(_mm256_mul_ps(vd, _mm256_cvtepi32_ps(_mm256_set_m128i(p32_1, p32_0))), acc);
+        acc = _mm256_add_ps(_mm256_mul_ps(vd, _mm256_cvtepi32_ps(MM256_SET_M128I(p32_1, p32_0))), acc);
 
         const __m128i p32_2 = _mm_madd_epi16(_mm_set1_epi16(scales[1]), p16_2);
         const __m128i p32_3 = _mm_madd_epi16(_mm_set1_epi16(scales[1]), p16_3);
-        acc = _mm256_add_ps(_mm256_mul_ps(vd, _mm256_cvtepi32_ps(_mm256_set_m128i(p32_3, p32_2))), acc);
+        acc = _mm256_add_ps(_mm256_mul_ps(vd, _mm256_cvtepi32_ps(MM256_SET_M128I(p32_3, p32_2))), acc);
 
     }
 
@@ -3160,7 +3162,7 @@ void ggml_vec_dot_q5_K_q8_K(const int n, float * restrict s, const void * restri
         summs += dmin * _mm_extract_epi32(hsum, 0);
 
         const __m128i sc128  = _mm256_extracti128_si256(mins_and_scales, 0);
-        const __m256i scales = _mm256_set_m128i(sc128, sc128);
+        const __m256i scales = MM256_SET_M128I(sc128, sc128);
 
         const __m256i hbits = _mm256_loadu_si256((const __m256i*)x[i].qh);
         __m256i hmask = mone;
@@ -3299,7 +3301,7 @@ void ggml_vec_dot_q5_K_q8_K(const int n, float * restrict s, const void * restri
         }
 
         __m256 vd = _mm256_set1_ps(d);
-        __m256i sumi = _mm256_set_m128i(sumi_1, sumi_0);
+        __m256i sumi = MM256_SET_M128I(sumi_1, sumi_0);
         acc = _mm256_add_ps(_mm256_mul_ps(vd, _mm256_cvtepi32_ps(sumi)), acc);
 
     }
@@ -3462,13 +3464,13 @@ void ggml_vec_dot_q5_K_q8_K(const int n, float * restrict s, const void * restri
 
         const __m256i q5bits = _mm256_loadu_si256((const __m256i*)q5);
 
-        const __m256i scale_l = _mm256_set_m128i(_mm_set1_epi16(x[i].scales[1]), _mm_set1_epi16(x[i].scales[0]));
-        const __m256i scale_h = _mm256_set_m128i(_mm_set1_epi16(x[i].scales[3]), _mm_set1_epi16(x[i].scales[2]));
+        const __m256i scale_l = MM256_SET_M128I(_mm_set1_epi16(x[i].scales[1]), _mm_set1_epi16(x[i].scales[0]));
+        const __m256i scale_h = MM256_SET_M128I(_mm_set1_epi16(x[i].scales[3]), _mm_set1_epi16(x[i].scales[2]));
 
         int64_t aux64;
         memcpy(&aux64, x[i].qh, 8);
         const __m128i haux128 = _mm_set_epi64x(aux64 >> 1, aux64);
-        const __m256i haux256 = _mm256_set_m128i(_mm_srli_epi16(haux128, 2), haux128);
+        const __m256i haux256 = MM256_SET_M128I(_mm_srli_epi16(haux128, 2), haux128);
 
         const __m256i q5h_0 = _mm256_slli_epi16(_mm256_andnot_si256(haux256, mone), 4);
         const __m256i q5h_1 = _mm256_slli_epi16(_mm256_andnot_si256(_mm256_srli_epi16(haux256, 4), mone), 4);
@@ -3543,7 +3545,7 @@ void ggml_vec_dot_q5_K_q8_K(const int n, float * restrict s, const void * restri
         const __m128i dot_0 = _mm_sub_epi32(_mm_add_epi32(p16_0, p16_2), _mm_add_epi32(s16_0, s16_2));
         const __m128i dot_1 = _mm_sub_epi32(_mm_add_epi32(p16_1, p16_3), _mm_add_epi32(s16_1, s16_3));
 
-        acc = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(_mm256_set_m128i(dot_1, dot_0))), acc);
+        acc = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(MM256_SET_M128I(dot_1, dot_0))), acc);
 
     }
 
@@ -3925,7 +3927,7 @@ void ggml_vec_dot_q6_K_q8_K(const int n, float * restrict s, const void * restri
 
         }
 
-        __m256i sumi = _mm256_set_m128i(sumi_1, sumi_0);
+        __m256i sumi = MM256_SET_M128I(sumi_1, sumi_0);
         acc = _mm256_add_ps(_mm256_mul_ps(_mm256_broadcast_ss(&d), _mm256_cvtepi32_ps(sumi)), acc);
     }
 
@@ -4083,8 +4085,8 @@ void ggml_vec_dot_q6_K_q8_K(const int n, float * restrict s, const void * restri
         const __m256i q4bits1 = _mm256_loadu_si256((const __m256i*)q4);
         const __m128i q4bitsH = _mm_loadu_si128((const __m128i*)qh);
 
-        const __m256i q4h_0 = _mm256_slli_epi16(_mm256_and_si256(_mm256_set_m128i(_mm_srli_epi16(q4bitsH, 2), q4bitsH), m2), 4);
-        const __m256i q4h_1 = _mm256_slli_epi16(_mm256_and_si256(_mm256_set_m128i(_mm_srli_epi16(q4bitsH, 6), _mm_srli_epi16(q4bitsH, 4)), m2), 4);
+        const __m256i q4h_0 = _mm256_slli_epi16(_mm256_and_si256(MM256_SET_M128I(_mm_srli_epi16(q4bitsH, 2), q4bitsH), m2), 4);
+        const __m256i q4h_1 = _mm256_slli_epi16(_mm256_and_si256(MM256_SET_M128I(_mm_srli_epi16(q4bitsH, 6), _mm_srli_epi16(q4bitsH, 4)), m2), 4);
 
         const __m256i q4_0 = _mm256_or_si256(_mm256_and_si256(q4bits1, m4), q4h_0);
         const __m256i q4_1 = _mm256_or_si256(_mm256_and_si256(_mm256_srli_epi16(q4bits1, 4), m4), q4h_1);
@@ -4177,7 +4179,7 @@ void ggml_vec_dot_q6_K_q8_K(const int n, float * restrict s, const void * restri
         sumi_0 = _mm_add_epi32(sumi_0, _mm_add_epi32(p16_0, p16_2));
         sumi_1 = _mm_add_epi32(sumi_1, _mm_add_epi32(p16_1, p16_3));
 
-        acc = _mm256_add_ps(_mm256_mul_ps(_mm256_broadcast_ss(&d), _mm256_cvtepi32_ps(_mm256_set_m128i(sumi_1, sumi_0))), acc);
+        acc = _mm256_add_ps(_mm256_mul_ps(_mm256_broadcast_ss(&d), _mm256_cvtepi32_ps(MM256_SET_M128I(sumi_1, sumi_0))), acc);
     }
 
     *s = hsum_float_8(acc);

koboldcpp.py

@@ -23,6 +23,7 @@ class load_model_inputs(ctypes.Structure):
                 ("batch_size", ctypes.c_int),
                 ("f16_kv", ctypes.c_bool),
                 ("low_vram", ctypes.c_bool),
+                ("use_mmq", ctypes.c_bool),
                 ("executable_path", ctypes.c_char_p),
                 ("model_filename", ctypes.c_char_p),
                 ("lora_filename", ctypes.c_char_p),
@@ -65,7 +66,7 @@ class generation_inputs(ctypes.Structure):
 
 class generation_outputs(ctypes.Structure):
     _fields_ = [("status", ctypes.c_int),
-                ("text", ctypes.c_char * 16384)]
+                ("text", ctypes.c_char * 24576)]
 
 handle = None
 
@@ -89,29 +90,30 @@ def pick_existant_file(ntoption,nonntoption):
 lib_default = pick_existant_file("koboldcpp.dll","koboldcpp.so")
 lib_failsafe = pick_existant_file("koboldcpp_failsafe.dll","koboldcpp_failsafe.so")
 lib_openblas = pick_existant_file("koboldcpp_openblas.dll","koboldcpp_openblas.so")
-lib_openblas_noavx2 = pick_existant_file("koboldcpp_openblas_noavx2.dll","koboldcpp_openblas_noavx2.so")
+lib_noavx2 = pick_existant_file("koboldcpp_noavx2.dll","koboldcpp_noavx2.so")
 lib_clblast = pick_existant_file("koboldcpp_clblast.dll","koboldcpp_clblast.so")
 lib_cublas = pick_existant_file("koboldcpp_cublas.dll","koboldcpp_cublas.so")
 
 
 def init_library():
     global handle
-    global lib_default,lib_failsafe,lib_openblas,lib_openblas_noavx2,lib_clblast,lib_cublas
+    global lib_default,lib_failsafe,lib_openblas,lib_noavx2,lib_clblast,lib_cublas
 
     libname = ""
-    use_blas = False # if true, uses OpenBLAS for acceleration. libopenblas.dll must exist in the same dir.
+    use_openblas = False # if true, uses OpenBLAS for acceleration. libopenblas.dll must exist in the same dir.
     use_clblast = False #uses CLBlast instead
     use_cublas = False #uses cublas instead
-    use_noavx2 = False #uses openblas with no avx2 instructions
+    use_noavx2 = False #uses no avx2 instructions
+    use_failsafe = False #uses no intrinsics, failsafe mode
     if args.noavx2:
         use_noavx2 = True
-        if not file_exists(lib_openblas_noavx2) or (os.name=='nt' and not file_exists("libopenblas.dll")):
-            print("Warning: OpenBLAS library file not found. Non-BLAS library will be used.")
-        elif args.noblas:
+        if not file_exists(lib_noavx2):
+            print("Warning: NoAVX2 library file not found. Failsafe library will be used.")
+        elif (args.noblas and args.nommap):
+            use_failsafe = True
             print("!!! Attempting to use FAILSAFE MODE !!!")
         else:
-            use_blas = True
-            print("Attempting to use non-avx2 compatibility library with OpenBLAS. A compatible libopenblas will be required.")
+            print("Attempting to use non-avx2 compatibility library.")
     elif args.useclblast:
         if not file_exists(lib_clblast) or (os.name=='nt' and not file_exists("clblast.dll")):
             print("Warning: CLBlast library file not found. Non-BLAS library will be used.")
@@ -130,22 +132,22 @@ def init_library():
         elif args.noblas:
             print("Attempting to library without OpenBLAS.")
         else:
-            use_blas = True
+            use_openblas = True
             print("Attempting to use OpenBLAS library for faster prompt ingestion. A compatible libopenblas will be required.")
             if sys.platform=="darwin":
                 print("Mac OSX note: Some people have found Accelerate actually faster than OpenBLAS. To compare, run Koboldcpp with --noblas instead.")
 
     if use_noavx2:
-        if use_blas:
-            libname = lib_openblas_noavx2
-        else:
+        if use_failsafe:
             libname = lib_failsafe
+        else:
+            libname = lib_noavx2
     else:
         if use_clblast:
             libname = lib_clblast
         elif use_cublas:
             libname = lib_cublas
-        elif use_blas:
+        elif use_openblas:
             libname = lib_openblas
         else:
             libname = lib_default
@@ -178,6 +180,7 @@ def load_model(model_filename):
     inputs.max_context_length = maxctx #initial value to use for ctx, can be overwritten
     inputs.threads = args.threads
     inputs.low_vram = (True if (args.usecublas and "lowvram" in args.usecublas) else False)
+    inputs.use_mmq = (True if (args.usecublas and "mmq" in args.usecublas) else False)
     inputs.blasthreads = args.blasthreads
     inputs.f16_kv = True
     inputs.use_mmap = (not args.nommap)
@@ -229,12 +232,17 @@ def load_model(model_filename):
     return ret
 
 def generate(prompt,max_length=20, max_context_length=512, temperature=0.8, top_k=120, top_a=0.0, top_p=0.85, typical_p=1.0, tfs=1.0, rep_pen=1.1, rep_pen_range=128, mirostat=0, mirostat_tau=5.0, mirostat_eta=0.1, sampler_order=[6,0,1,3,4,2,5], seed=-1, stop_sequence=[], stream_sse=False):
+    global maxctx
     inputs = generation_inputs()
     outputs = ctypes.create_unicode_buffer(ctypes.sizeof(generation_outputs))
     inputs.prompt = prompt.encode("UTF-8")
     if max_length >= max_context_length:
         max_length = max_context_length-1
     inputs.max_context_length = max_context_length   # this will resize the context buffer if changed
+    global showmaxctxwarning
+    if showmaxctxwarning and max_context_length > maxctx:
+        print(f"\n(Warning! Request max_context_length={max_context_length} exceeds allocated context size of {maxctx}. Consider launching with increased --contextsize to avoid errors. This message will only show once per session.)")
+        showmaxctxwarning = False
     inputs.max_length = max_length
     inputs.temperature = temperature
     inputs.top_k = top_k
@@ -295,9 +303,10 @@ maxhordectx = 1024
 maxhordelen = 256
 modelbusy = threading.Lock()
 defaultport = 5001
-KcppVersion = "1.37.1"
+KcppVersion = "1.40.1"
 showdebug = True
 showsamplerwarning = True
+showmaxctxwarning = True
 exitcounter = 0
 
 class ServerRequestHandler(http.server.SimpleHTTPRequestHandler):
@@ -393,6 +402,7 @@ class ServerRequestHandler(http.server.SimpleHTTPRequestHandler):
 
         current_token = 0
 
+        incomplete_token_buffer = bytearray()
         while not handle.has_finished():
             if current_token < handle.get_stream_count():
                 token = handle.new_token(current_token)
@@ -402,10 +412,14 @@ class ServerRequestHandler(http.server.SimpleHTTPRequestHandler):
 
                 current_token += 1
 
-                tokenStr = ctypes.string_at(token).decode("UTF-8","ignore")
-                event_data = {"token": tokenStr}
-                event_str = json.dumps(event_data)
-                await self.send_sse_event("message", event_str)
+                newbyte = ctypes.string_at(token)
+                incomplete_token_buffer += bytearray(newbyte)
+                tokenStr = incomplete_token_buffer.decode("UTF-8","ignore")
+                if tokenStr!="":
+                    incomplete_token_buffer.clear()
+                    event_data = {"token": tokenStr}
+                    event_str = json.dumps(event_data)
+                    await self.send_sse_event("message", event_str)
 
             await asyncio.sleep(0)
 
@@ -481,7 +495,7 @@ class ServerRequestHandler(http.server.SimpleHTTPRequestHandler):
             laste = handle.get_last_eval_time()
             lastc = handle.get_last_token_count()
             stopreason = handle.get_last_stop_reason()
-            response_body = (json.dumps({"last_process":lastp,"last_eval":laste,"last_token_count":lastc, "stop_reason":stopreason}).encode())
+            response_body = (json.dumps({"last_process":lastp,"last_eval":laste,"last_token_count":lastc, "stop_reason":stopreason, "idle":(0 if modelbusy.locked() else 1)}).encode())
 
         if response_body is None:
             self.send_response(404)
@@ -558,9 +572,12 @@ class ServerRequestHandler(http.server.SimpleHTTPRequestHandler):
                 newprompt = fullprompt
 
                 gen = asyncio.run(self.handle_request(genparams, newprompt, basic_api_flag, kai_sse_stream_flag))
+
                 try:
-                    self.send_response(200)
-                    self.end_headers()
+                    # Headers are already sent when streaming
+                    if not kai_sse_stream_flag:
+                        self.send_response(200)
+                        self.end_headers()
                     self.wfile.write(json.dumps(gen).encode())
                 except:
                     print("Generate: The response could not be sent, maybe connection was terminated?")
@@ -629,7 +646,7 @@ def RunServerMultiThreaded(addr, port, embedded_kailite = None):
             exitcounter = 999
             self.httpd.server_close()
 
-    numThreads = 6
+    numThreads = 8
     threadArr = []
     for i in range(numThreads):
         threadArr.append(Thread(i))
@@ -656,14 +673,13 @@ def show_new_gui():
         root.destroy()
         if not args.model_param:
             print("\nNo ggml model file was selected. Exiting.")
-            time.sleep(2)
+            time.sleep(3)
             sys.exit(2)
         return
 
     import customtkinter as ctk
-
     nextstate = 0 #0=exit, 1=launch, 2=oldgui
-    windowwidth = 520
+    windowwidth = 530
     windowheight = 500
     ctk.set_appearance_mode("dark")
     root = ctk.CTk()
@@ -684,13 +700,22 @@ def show_new_gui():
     tabcontentframe.grid_propagate(False)
 
     tabcontent = {}
-
+    lib_option_pairs = [
+        (lib_openblas, "Use OpenBLAS"),
+        (lib_clblast, "Use CLBlast"),
+        (lib_cublas, "Use CuBLAS"),
+        (lib_default, "Use No BLAS"),
+        (lib_noavx2, "NoAVX2 Mode (Old CPU)"),
+        (lib_failsafe, "Failsafe Mode (Old CPU)")]
+    openblas_option, clblast_option, cublas_option, default_option, noavx2_option, failsafe_option = (opt if file_exists(lib) or (os.name == 'nt' and file_exists(opt + ".dll")) else None for lib, opt in lib_option_pairs)
     # slider data
-    blasbatchsize_values = ["-1", "32", "64", "128", "256", "512", "1024"]
-    blasbatchsize_text = ["Don't Batch BLAS","32","64","128","256","512","1024"]
-    contextsize_text = ["512", "1024", "2048", "3072", "4096", "6144", "8192"]
-    runopts = ["Use OpenBLAS","Use CLBlast", "Use CuBLAS", "Use No BLAS","Use OpenBLAS (Old CPU, noavx2)","Failsafe Mode (Old CPU, noavx)"]
-
+    blasbatchsize_values = ["-1", "32", "64", "128", "256", "512", "1024", "2048"]
+    blasbatchsize_text = ["Don't Batch BLAS","32","64","128","256","512","1024","2048"]
+    contextsize_text = ["512", "1024", "2048", "3072", "4096", "6144", "8192", "12288", "16384"]
+    runopts = [opt for lib, opt in lib_option_pairs if file_exists(lib) or os.name == 'nt' and file_exists(opt + ".dll")]
+    antirunopts = [opt.replace("Use ", "") for lib, opt in lib_option_pairs if not file_exists(lib) or os.name == 'nt' and not file_exists(opt + ".dll")]
+    if not any(runopts):
+        show_gui_warning("No Backend Available")
     def tabbuttonaction(name):
         for t in tabcontent:
             if name == t:
@@ -756,6 +781,32 @@ def show_new_gui():
         button.grid(row=row+1, column=1, stick="nw")
         return
 
+    def show_tooltip(event, tooltip_text=None):
+        if hasattr(show_tooltip, "_tooltip"):
+            tooltip = show_tooltip._tooltip
+        else:
+            tooltip = ctk.CTkToplevel(root)
+            tooltip.configure(fg_color="#ffffe0")
+            tooltip.withdraw()
+            tooltip.overrideredirect(True)
+            tooltip_label = ctk.CTkLabel(tooltip, text=tooltip_text, text_color="#000000", fg_color="#ffffe0")
+            tooltip_label.pack(expand=True, padx=2, pady=1)
+            show_tooltip._tooltip = tooltip
+        x, y = root.winfo_pointerxy()
+        tooltip.wm_geometry(f"+{x + 10}+{y + 10}")
+        tooltip.deiconify()
+    def hide_tooltip(event):
+        if hasattr(show_tooltip, "_tooltip"):
+            tooltip = show_tooltip._tooltip
+            tooltip.withdraw()
+    def setup_backend_tooltip(parent):
+        num_backends_built = makelabel(parent, str(len(runopts)) + "/6", 5, 2)
+        num_backends_built.grid(row=1, column=2, padx=0, pady=0)
+        num_backends_built.configure(text_color="#00ff00")
+        # Bind the backend count label with the tooltip function
+        num_backends_built.bind("<Enter>", lambda event: show_tooltip(event, f"This is the number of backends you have built and available." + (f"\nMissing: {', '.join(antirunopts)}" if len(runopts) != 6 else "")))
+        num_backends_built.bind("<Leave>", hide_tooltip)
+
     # Vars - should be in scope to be used by multiple widgets
     gpulayers_var = ctk.StringVar(value="0")
     threads_var = ctk.StringVar(value=str(default_threads))
@@ -770,6 +821,7 @@ def show_new_gui():
     debugmode = ctk.IntVar()
 
     lowvram_var = ctk.IntVar()
+    mmq_var = ctk.IntVar()
 
     blas_threads_var = ctk.StringVar()
     blas_size_var = ctk.IntVar()
@@ -806,11 +858,13 @@ def show_new_gui():
     quick_tab = tabcontent["Quick Launch"]
 
     # gpu options
-    quick_gpu_layers_entry,quick_gpu_layers_label = makelabelentry(quick_tab,"GPU Layers:", gpulayers_var, 4, 50)
+    quick_gpu_layers_entry,quick_gpu_layers_label = makelabelentry(quick_tab,"GPU Layers:", gpulayers_var, 5, 50)
     quick_gpu_selector_label = makelabel(quick_tab, "GPU ID:", 3)
     quick_gpu_selector_box = ctk.CTkComboBox(quick_tab, values=["1","2","3"], width=60, variable=gpu_choice_var, state="readonly")
     CUDA_quick_gpu_selector_box = ctk.CTkComboBox(quick_tab, values=["1","2","3","All"], width=60, variable=gpu_choice_var, state="readonly")
-    quick_lowvram_box = makecheckbox(quick_tab,  "Low VRAM", lowvram_var, 5)
+    quick_lowvram_box = makecheckbox(quick_tab,  "Low VRAM", lowvram_var, 4,0)
+    quick_mmq_box = makecheckbox(quick_tab,  "Use QuantMatMul (mmq)", mmq_var, 4,1)
+
 
     def changerunmode(a,b,c):
         index = runopts_var.get()
@@ -836,9 +890,13 @@ def show_new_gui():
         if index == "Use CuBLAS":
             lowvram_box.grid(row=4, column=0, padx=8, pady=1,  stick="nw")
             quick_lowvram_box.grid(row=4, column=0, padx=8, pady=1,  stick="nw")
+            mmq_box.grid(row=4, column=1, padx=8, pady=1,  stick="nw")
+            quick_mmq_box.grid(row=4, column=1, padx=8, pady=1,  stick="nw")
         else:
             lowvram_box.grid_forget()
             quick_lowvram_box.grid_forget()
+            mmq_box.grid_forget()
+            quick_mmq_box.grid_forget()
 
         if index == "Use CLBlast" or index == "Use CuBLAS":
             gpu_layers_label.grid(row=5, column=0, padx = 8, pady=1, stick="nw")
@@ -856,19 +914,21 @@ def show_new_gui():
 
     runoptbox = ctk.CTkComboBox(quick_tab, values=runopts, width=180,variable=runopts_var, state="readonly")
     runoptbox.grid(row=1, column=1,padx=8, stick="nw")
-    runoptbox.set("Use OpenBLAS")
+    runoptbox.set(runopts[0]) # Set to first available option
+
+    # Tell user how many backends are available
+    setup_backend_tooltip(quick_tab)
 
     # threads
     makelabelentry(quick_tab, "Threads:" , threads_var, 8, 50)
 
     # blas batch size
-    makeslider(quick_tab, "BLAS Batch Size:", blasbatchsize_text, blas_size_var, 0, 6, 12, set=5)
+    makeslider(quick_tab, "BLAS Batch Size:", blasbatchsize_text, blas_size_var, 0, 7, 12, set=5)
 
     # quick boxes
     quick_boxes = {"Launch Browser": launchbrowser , "High Priority" : highpriority, "Streaming Mode":stream, "Use SmartContext":smartcontext, "Unban Tokens":unbantokens, "Disable MMAP":disablemmap,}
     for idx, name, in enumerate(quick_boxes):
         makecheckbox(quick_tab, name, quick_boxes[name], int(idx/2) +20, idx%2)
-
     # context size
     makeslider(quick_tab, "Context Size:", contextsize_text, context_var, 0, len(contextsize_text)-1, 30, set=2)
 
@@ -879,19 +939,24 @@ def show_new_gui():
     hardware_tab = tabcontent["Hardware"]
 
     # gpu options
-    gpu_layers_entry,gpu_layers_label = makelabelentry(hardware_tab,"GPU Layers:", gpulayers_var, 4, 50)
+    gpu_layers_entry,gpu_layers_label = makelabelentry(hardware_tab,"GPU Layers:", gpulayers_var, 5, 50)
     gpu_selector_label = makelabel(hardware_tab, "GPU ID:", 3)
     gpu_selector_box = ctk.CTkComboBox(hardware_tab, values=["1","2","3"], width=60, variable=gpu_choice_var, state="readonly")
     CUDA_gpu_selector_box = ctk.CTkComboBox(hardware_tab, values=["1","2","3", "All"], width=60, variable=gpu_choice_var, state="readonly")
-    lowvram_box = makecheckbox(hardware_tab,  "Low VRAM", lowvram_var, 5)
+    lowvram_box = makecheckbox(hardware_tab,  "Low VRAM", lowvram_var, 4,0)
+    mmq_box = makecheckbox(hardware_tab,  "Use QuantMatMul (mmq)", mmq_var, 4,1)
 
     # presets selector
     makelabel(hardware_tab, "Presets:", 1)
     runoptbox = ctk.CTkComboBox(hardware_tab, values=runopts,  width=180,variable=runopts_var, state="readonly")
     runoptbox.grid(row=1, column=1,padx=8, stick="nw")
-    runoptbox.set("Use OpenBLAS")
+    runoptbox.set(runopts[0]) # Set to first available option
     runopts_var.trace('w', changerunmode)
     changerunmode(1,1,1)
+
+    # Tell user how many backends are available
+    setup_backend_tooltip(hardware_tab)
+
     # threads
     makelabelentry(hardware_tab, "Threads:" , threads_var, 8, 50)
 
@@ -904,7 +969,7 @@ def show_new_gui():
     # blas thread specifier
     makelabelentry(hardware_tab, "BLAS threads:" , blas_threads_var, 11, 50)
     # blas batch size
-    makeslider(hardware_tab, "BLAS Batch Size:", blasbatchsize_text, blas_size_var, 0, 6, 12, set=5)
+    makeslider(hardware_tab, "BLAS Batch Size:", blasbatchsize_text, blas_size_var, 0, 7, 12, set=5)
     # force version
     makelabelentry(hardware_tab, "Force Version:" , version_var, 100, 50)
 
@@ -931,7 +996,7 @@ def show_new_gui():
     togglemiro(1,1,1)
 
     # context size
-    makeslider(tokens_tab, "Context Size:",contextsize_text, context_var, 0, 4, 20, set=2)
+    makeslider(tokens_tab, "Context Size:",contextsize_text, context_var, 0, len(contextsize_text)-1, 20, set=2)
 
 
     customrope_scale_entry, customrope_scale_label = makelabelentry(tokens_tab, "RoPE Scale:", customrope_scale)
@@ -1017,20 +1082,22 @@ def show_new_gui():
         gpuchoiceidx = 0
         if gpu_choice_var.get()!="All":
             gpuchoiceidx = int(gpu_choice_var.get())-1
-        if runopts_var.get() == runopts[1]:
+        if runopts_var.get() == "Use CLBlast":
             args.useclblast = [[0,0], [1,0], [0,1]][gpuchoiceidx]
-        if runopts_var.get() == runopts[2]:
+        if runopts_var.get() == "Use CuBLAS":
             if gpu_choice_var.get()=="All":
                 args.usecublas = ["lowvram"] if lowvram_var.get() == 1 else ["normal"]
             else:
                 args.usecublas = ["lowvram",str(gpuchoiceidx)] if lowvram_var.get() == 1 else ["normal",str(gpuchoiceidx)]
+            if mmq_var.get()==1:
+                args.usecublas.append("mmq")
         if gpulayers_var.get():
             args.gpulayers = int(gpulayers_var.get())
-        if runopts_var.get()==runopts[3]:
+        if runopts_var.get()=="Use No BLAS":
             args.noblas = True
-        if runopts_var.get()==runopts[4]:
+        if runopts_var.get()=="NoAVX2 Mode (Old CPU)":
             args.noavx2 = True
-        if runopts_var.get()==runopts[5]:
+        if runopts_var.get()=="Failsafe Mode (Old CPU)":
             args.noavx2 = True
             args.noblas = True
             args.nommap = True
@@ -1040,7 +1107,7 @@ def show_new_gui():
         args.blasbatchsize = int(blasbatchsize_values[int(blas_size_var.get())])
         args.forceversion = 0 if version_var.get()=="" else int(version_var.get())
 
-        args.mirostat = [int(mirostat_var.get()), float(mirostat_tau.get()), float(mirostat_eta.get())] if usemirostat.get()==1 else None
+        args.usemirostat = [int(mirostat_var.get()), float(mirostat_tau.get()), float(mirostat_eta.get())] if usemirostat.get()==1 else None
         args.contextsize = int(contextsize_text[context_var.get()])
 
         if customrope_var.get()==1:
@@ -1069,38 +1136,39 @@ def show_new_gui():
         stream.set(1 if "stream" in dict and dict["stream"] else 0)
         smartcontext.set(1 if "smartcontext" in dict and dict["smartcontext"] else 0)
         unbantokens.set(1 if "unbantokens" in dict and dict["unbantokens"] else 0)
-        runopts_var.set(runopts[0])
         if "useclblast" in dict and dict["useclblast"]:
-            runopts_var.set(runopts[1])
-            gpu_choice_var.set(str(["0 0", "1 0", "0 1"].index(str(dict["useclblast"][0]) + " " + str(dict["useclblast"][1])) + 1))
+            if clblast_option is not None:
+                runopts_var.set(clblast_option)
+                gpu_choice_var.set(str(["0 0", "1 0", "0 1"].index(str(dict["useclblast"][0]) + " " + str(dict["useclblast"][1])) + 1))
         elif "usecublas" in dict and dict["usecublas"]:
-            runopts_var.set(runopts[2])
-            if len(dict["usecublas"])==1:
-                lowvram_var.set(1 if dict["usecublas"][0]=="lowvram" else 0)
-            else:
+            if cublas_option is not None:
+                runopts_var.set(cublas_option)
                 lowvram_var.set(1 if "lowvram" in dict["usecublas"] else 0)
-                gpu_choice_var.set("1")
+                mmq_var.set(1 if "mmq" in dict["usecublas"] else 0)
+                gpu_choice_var.set("All")
                 for g in range(3):
                     if str(g) in dict["usecublas"]:
                         gpu_choice_var.set(str(g+1))
                         break
-        if "gpulayers" in dict and dict["gpulayers"]:
-            gpulayers_var.set(dict["gpulayers"])
-
-        if  "noavx2" in dict and "noblas" in dict and dict["noblas"] and dict["noavx2"]:
-            runopts_var.set(runopts[5])
+        elif  "noavx2" in dict and "noblas" in dict and dict["noblas"] and dict["noavx2"]:
+            if failsafe_option is not None:
+                runopts_var.set(failsafe_option)
         elif "noavx2" in dict and dict["noavx2"]:
-            runopts_var.set(runopts[4])
+            if noavx2_option is not None:
+                runopts_var.set(noavx2_option)
         elif "noblas" in dict and dict["noblas"]:
-            runopts_var.set(runopts[3])
+            if default_option is not None:
+                runopts_var.set(default_option)
+        elif openblas_option is not None:
+                runopts_var.set(openblas_option)
+        if "gpulayers" in dict and dict["gpulayers"]:
+            gpulayers_var.set(dict["gpulayers"])
         if "blasthreads" in dict and dict["blasthreads"]:
             blas_threads_var.set(str(dict["blasthreads"]))
         else:
             blas_threads_var.set("")
-
         if "contextsize" in dict and dict["contextsize"]:
             context_var.set(contextsize_text.index(str(dict["contextsize"])))
-
         if "ropeconfig" in dict and dict["ropeconfig"] and len(dict["ropeconfig"])>1:
             if dict["ropeconfig"][0]>0:
                 customrope_var.set(1)
@@ -1114,11 +1182,11 @@ def show_new_gui():
         if "forceversion" in dict and dict["forceversion"]:
             version_var.set(str(dict["forceversion"]))
 
-        if "mirostat" in dict and dict["mirostat"] and len(dict["mirostat"])>1:
-            usemirostat.set(0 if str(dict["mirostat"][0])=="0" else 1)
-            mirostat_var.set(str(dict["mirostat"][0]))
-            mirostat_tau.set(str(dict["mirostat"][1]))
-            mirostat_eta.set(str(dict["mirostat"][2]))
+        if "usemirostat" in dict and dict["usemirostat"] and len(dict["usemirostat"])>1:
+            usemirostat.set(0 if str(dict["usemirostat"][0])=="0" else 1)
+            mirostat_var.set(str(dict["usemirostat"][0]))
+            mirostat_tau.set(str(dict["usemirostat"][1]))
+            mirostat_eta.set(str(dict["usemirostat"][2]))
 
         if "model_param" in dict and dict["model_param"]:
             model_var.set(dict["model_param"])
@@ -1165,18 +1233,26 @@ def show_new_gui():
             import_vars(dict)
         pass
 
-    ctk.CTkButton(tabs , text = "Launch", fg_color="#2f8d3c", command = guilaunch, width=80, height = 35 ).grid(row=1,column=1, stick="se", padx= 25, pady=5)
+    def display_help():
+        try:
+            import webbrowser as wb
+            wb.open("https://github.com/LostRuins/koboldcpp/wiki")
+        except:
+            print("Cannot launch help browser.")
+
+    ctk.CTkButton(tabs , text = "Launch", fg_color="#2f8d3c", hover_color="#2faa3c", command = guilaunch, width=80, height = 35 ).grid(row=1,column=1, stick="se", padx= 25, pady=5)
 
-    ctk.CTkButton(tabs , text = "Save", fg_color="#084a66", command = save_config, width=60, height = 35 ).grid(row=1,column=1, stick="sw", padx= 5, pady=5)
-    ctk.CTkButton(tabs , text = "Load", fg_color="#084a66", command = load_config, width=60, height = 35 ).grid(row=1,column=1, stick="sw", padx= 70, pady=5)
+    ctk.CTkButton(tabs , text = "Save", fg_color="#084a66", hover_color="#085a88", command = save_config, width=60, height = 35 ).grid(row=1,column=1, stick="sw", padx= 5, pady=5)
+    ctk.CTkButton(tabs , text = "Load", fg_color="#084a66", hover_color="#085a88", command = load_config, width=60, height = 35 ).grid(row=1,column=1, stick="sw", padx= 70, pady=5)
+    ctk.CTkButton(tabs , text = "Help", fg_color="#992222", hover_color="#bb3333", command = display_help, width=60, height = 35 ).grid(row=1,column=1, stick="sw", padx= 135, pady=5)
 
-    ctk.CTkButton(tabs , text = "Old GUI", fg_color="#084a66", command = switch_old_gui, width=100, height = 35 ).grid(row=1,column=0, stick="sw", padx= 5, pady=5)
+    ctk.CTkButton(tabs , text = "Old GUI", fg_color="#084a66", hover_color="#085a88", command = switch_old_gui, width=100, height = 35 ).grid(row=1,column=0, stick="sw", padx= 5, pady=5)
     # runs main loop until closed or launch clicked
     root.mainloop()
 
     if nextstate==0:
         print("Exiting by user request.")
-        time.sleep(2)
+        time.sleep(3)
         sys.exit()
     elif nextstate==2:
         time.sleep(0.1)
@@ -1187,16 +1263,23 @@ def show_new_gui():
 
         if not args.model_param:
             print("\nNo ggml model file was selected. Exiting.")
-            time.sleep(2)
+            time.sleep(3)
             sys.exit(2)
 
-def show_gui_warning():
+def show_gui_warning(issue=None):
     from tkinter import messagebox
     import tkinter as tk
     root = tk.Tk()
     root.attributes("-alpha", 0)
-    messagebox.showerror(title="New GUI failed, using Old GUI", message="The new GUI failed to load.\n\nTo use new GUI, please install the customtkinter python module.")
-    root.destroy()
+    if issue == "No Backend Available":
+        messagebox.showerror(title="No Backends Available!", message="KoboldCPP couldn't locate any backends to use.\n\nTo use the program, please run the 'make' command from the directory.")
+        root.destroy()
+        print("No Backend Available (i.e Default, OpenBLAS, CLBlast, CuBLAS). To use the program, please run the 'make' command from the directory.")
+        time.sleep(3)
+        sys.exit(2)
+    else:
+        messagebox.showerror(title="New GUI failed, using Old GUI", message="The new GUI failed to load.\n\nTo use new GUI, please install the customtkinter python module.")
+        root.destroy()
 
 def show_old_gui():
     import tkinter as tk
@@ -1223,11 +1306,11 @@ def show_old_gui():
         tk.Label(root, text = "(Note: KoboldCpp only works with GGML model formats!)",
                 font = ("Arial", 9)).grid(row=1,column=0)
 
-        blasbatchopts = ["Don't Batch BLAS","BLAS = 32","BLAS = 64","BLAS = 128","BLAS = 256","BLAS = 512","BLAS = 1024"]
+        blasbatchopts = ["Don't Batch BLAS","BLAS = 32","BLAS = 64","BLAS = 128","BLAS = 256","BLAS = 512","BLAS = 1024","BLAS = 2048"]
         blaschoice = tk.StringVar()
         blaschoice.set("BLAS = 512")
 
-        runopts = ["Use OpenBLAS","Use CLBLast GPU #1","Use CLBLast GPU #2","Use CLBLast GPU #3","Use CuBLAS GPU","Use No BLAS","Use OpenBLAS (Old CPU, noavx2)","Failsafe Mode (Old CPU, noavx)"]
+        runopts = ["Use OpenBLAS","Use CLBLast GPU #1","Use CLBLast GPU #2","Use CLBLast GPU #3","Use CuBLAS GPU","Use No BLAS","NoAVX2 Mode (Old CPU)","Failsafe Mode (Old CPU)"]
         runchoice = tk.StringVar()
         runchoice.set("Use OpenBLAS")
 
@@ -1286,7 +1369,7 @@ def show_old_gui():
 
         if launchclicked==False:
             print("Exiting by user request.")
-            time.sleep(2)
+            time.sleep(3)
             sys.exit()
 
         #load all the vars
@@ -1318,7 +1401,6 @@ def show_old_gui():
             args.noavx2 = True
             args.noblas = True
             args.nommap = True
-            print("[Failsafe Mode : mmap is disabled.]")
 
         if selblaschoice==blasbatchopts[0]:
             args.blasbatchsize = -1
@@ -1334,6 +1416,8 @@ def show_old_gui():
             args.blasbatchsize = 512
         if selblaschoice==blasbatchopts[6]:
             args.blasbatchsize = 1024
+        if selblaschoice==blasbatchopts[7]:
+            args.blasbatchsize = 2048
 
         root = tk.Tk()
         root.attributes("-alpha", 0)
@@ -1341,7 +1425,7 @@ def show_old_gui():
         root.destroy()
         if not args.model_param:
             print("\nNo ggml model file was selected. Exiting.")
-            time.sleep(2)
+            time.sleep(3)
             sys.exit(2)
 
     else:
@@ -1351,7 +1435,7 @@ def show_old_gui():
         root.destroy()
         if not args.model_param:
             print("\nNo ggml model file was selected. Exiting.")
-            time.sleep(2)
+            time.sleep(3)
             sys.exit(2)
 
 #A very simple and stripped down embedded horde worker with no dependencies
@@ -1396,7 +1480,7 @@ def run_horde_worker(args, api_key, worker_name):
     BRIDGE_AGENT = f"KoboldCppEmbedWorker:1:https://github.com/LostRuins/koboldcpp"
     cluster = "https://horde.koboldai.net"
     while exitcounter < 10:
-        time.sleep(2)
+        time.sleep(3)
         readygo = make_url_request(f'{epurl}/api/v1/info/version', None,'GET')
         if readygo:
             print("Embedded Horde Worker is started.")
@@ -1472,10 +1556,10 @@ def run_horde_worker(args, api_key, worker_name):
         time.sleep(1)
     if exitcounter<100:
         print("Horde Worker Shutdown - Too many errors.")
-        time.sleep(2)
+        time.sleep(3)
     else:
         print("Horde Worker Shutdown - Server Closing.")
-        time.sleep(1)
+        time.sleep(2)
     sys.exit(2)
 
 def main(args):
@@ -1499,7 +1583,7 @@ def main(args):
                 except Exception as ex2:
                     print("File selection GUI unsupported. Please check command line: script.py --help")
                     print("Reason for no GUI: " + str(ex2))
-                    time.sleep(2)
+                    time.sleep(3)
                     sys.exit(2)
 
     if args.hordeconfig and args.hordeconfig[0]!="":
@@ -1543,20 +1627,20 @@ def main(args):
     time.sleep(1)
     if not os.path.exists(args.model_param):
         print(f"Cannot find model file: {args.model_param}")
-        time.sleep(2)
+        time.sleep(3)
         sys.exit(2)
 
     if args.lora and args.lora[0]!="":
         if not os.path.exists(args.lora[0]):
             print(f"Cannot find lora file: {args.lora[0]}")
-            time.sleep(2)
+            time.sleep(3)
             sys.exit(2)
         else:
             args.lora[0] = os.path.abspath(args.lora[0])
             if len(args.lora) > 1:
                 if not os.path.exists(args.lora[1]):
                     print(f"Cannot find lora base: {args.lora[1]}")
-                    time.sleep(2)
+                    time.sleep(3)
                     sys.exit(2)
                 else:
                     args.lora[1] = os.path.abspath(args.lora[1])
@@ -1577,7 +1661,7 @@ def main(args):
 
     if not loadok:
         print("Could not load model: " + modelname)
-        time.sleep(2)
+        time.sleep(3)
         sys.exit(3)
     try:
         basepath = os.path.abspath(os.path.dirname(__file__))
@@ -1605,6 +1689,7 @@ def main(args):
 
     if args.hordeconfig and len(args.hordeconfig)>4:
         horde_thread = threading.Thread(target=run_horde_worker,args=(args,args.hordeconfig[3],args.hordeconfig[4]))
+        horde_thread.daemon = True
         horde_thread.start()
 
     print(f"Please connect to custom endpoint at {epurl}")
@@ -1631,8 +1716,8 @@ if __name__ == '__main__':
     parser.add_argument("--blasthreads", help="Use a different number of threads during BLAS if specified. Otherwise, has the same value as --threads",metavar=('[threads]'), type=int, default=0)
     parser.add_argument("--psutil_set_threads", help="Experimental flag. If set, uses psutils to determine thread count based on physical cores.", action='store_true')
     parser.add_argument("--highpriority", help="Experimental flag. If set, increases the process CPU priority, potentially speeding up generation. Use caution.", action='store_true')
-    parser.add_argument("--contextsize", help="Controls the memory allocated for maximum context size, only change if you need more RAM for big contexts. (default 2048)", type=int,choices=[512,1024,2048,3072,4096,6144,8192], default=2048)
-    parser.add_argument("--blasbatchsize", help="Sets the batch size used in BLAS processing (default 512). Setting it to -1 disables BLAS mode, but keeps other benefits like GPU offload.", type=int,choices=[-1,32,64,128,256,512,1024], default=512)
+    parser.add_argument("--contextsize", help="Controls the memory allocated for maximum context size, only change if you need more RAM for big contexts. (default 2048)", type=int,choices=[512,1024,2048,3072,4096,6144,8192,12288,16384], default=2048)
+    parser.add_argument("--blasbatchsize", help="Sets the batch size used in BLAS processing (default 512). Setting it to -1 disables BLAS mode, but keeps other benefits like GPU offload.", type=int,choices=[-1,32,64,128,256,512,1024,2048], default=512)
     parser.add_argument("--ropeconfig", help="If set, uses customized RoPE scaling from configured frequency scale and frequency base (e.g. --ropeconfig 0.25 10000). Otherwise, uses NTK-Aware scaling set automatically based on context size. For linear rope, simply set the freq-scale and ignore the freq-base",metavar=('[rope-freq-scale]', '[rope-freq-base]'), default=[0.0, 10000.0], type=float, nargs='+')
     parser.add_argument("--stream", help="Uses streaming when generating tokens. Only for the Kobold Lite UI.", action='store_true')
     parser.add_argument("--smartcontext", help="Reserving a portion of context to try processing less frequently.", action='store_true')
@@ -1649,7 +1734,7 @@ if __name__ == '__main__':
     compatgroup = parser.add_mutually_exclusive_group()
     compatgroup.add_argument("--noblas", help="Do not use OpenBLAS for accelerated prompt ingestion", action='store_true')
     compatgroup.add_argument("--useclblast", help="Use CLBlast for GPU Acceleration. Must specify exactly 2 arguments, platform ID and device ID (e.g. --useclblast 1 0).", type=int, choices=range(0,9), nargs=2)
-    compatgroup.add_argument("--usecublas", help="Use CuBLAS for GPU Acceleration. Requires CUDA. Select lowvram to not allocate VRAM scratch buffer. Enter a number afterwards to select and use 1 GPU. Leaving no number will use all GPUs.", nargs='*',metavar=('[lowvram|normal] [main GPU ID]'), choices=['normal', 'lowvram', '0', '1', '2'])
+    compatgroup.add_argument("--usecublas", help="Use CuBLAS for GPU Acceleration. Requires CUDA. Select lowvram to not allocate VRAM scratch buffer. Enter a number afterwards to select and use 1 GPU. Leaving no number will use all GPUs.", nargs='*',metavar=('[lowvram|normal] [main GPU ID] [mmq]'), choices=['normal', 'lowvram', '0', '1', '2', 'mmq'])
     parser.add_argument("--gpulayers", help="Set number of layers to offload to GPU when using GPU. Requires GPU.",metavar=('[GPU layers]'), type=int, default=0)
     parser.add_argument("--tensor_split", help="For CUDA with ALL GPU set only, ratio to split tensors across multiple GPUs, space-separated list of proportions, e.g. 7 3", metavar=('[Ratios]'), type=float, nargs='+')
 

llama-util.h

@@ -149,6 +149,46 @@ struct llama_file {
     }
 };
 
+// llama_context_data
+struct llama_data_context {
+    virtual void write(const void * src, size_t size) = 0;
+    virtual size_t get_size_written() = 0;
+    virtual ~llama_data_context() = default;
+};
+
+struct llama_data_buffer_context : llama_data_context {
+    uint8_t* ptr;
+    size_t size_written = 0;
+
+    llama_data_buffer_context(uint8_t * p) : ptr(p) {}
+
+    void write(const void * src, size_t size) override {
+        memcpy(ptr, src, size);
+        ptr += size;
+        size_written += size;
+    }
+
+    size_t get_size_written() override {
+        return size_written;
+    }
+};
+
+struct llama_data_file_context : llama_data_context {
+    llama_file* file;
+    size_t size_written = 0;
+
+    llama_data_file_context(llama_file * f) : file(f) {}
+
+    void write(const void * src, size_t size) override {
+        file->write_raw(src, size);
+        size_written += size;
+    }
+
+    size_t get_size_written() override {
+        return size_written;
+    }
+};
+
 #if defined(_WIN32)
 static std::string llama_format_win_err(DWORD err) {
     LPSTR buf;
@@ -179,7 +219,7 @@ struct llama_mmap {
         // prefetch/readahead impairs performance on NUMA systems
         if (numa) { prefetch = 0; }
 #ifdef __linux__
-        if (prefetch) { flags |= MAP_POPULATE; }
+        if (prefetch >= file->size) { flags |= MAP_POPULATE; }
 #endif
         addr = mmap(NULL, file->size, PROT_READ, flags, fd, 0);
         if (addr == MAP_FAILED) {
@@ -247,7 +287,7 @@ struct llama_mmap {
         #pragma message("warning: You are building for pre-Windows 8; prefetch not supported")
         #endif // _WIN32_WINNT >= _WIN32_WINNT_WIN8
         #else
-        printf("\nPrefetchVirtualMemory skipped in failsafe mode.");
+        printf("\nPrefetchVirtualMemory skipped in compatibility mode.\n");
         #endif
     }
 

llama.cpp

@@ -57,8 +57,14 @@
 #pragma warning(disable: 4244 4267) // possible loss of data
 #endif
 
+#if !defined(GGML_USE_CUBLAS) && !defined(GGML_USE_METAL)
+#include "ggml-alloc.h"
+#define LLAMA_USE_ALLOCATOR
+#else
 #define LLAMA_USE_SCRATCH
 #define LLAMA_MAX_SCRATCH_BUFFERS 16
+#endif
+
 
 // available llama models
 enum e_model {
@@ -144,7 +150,7 @@ static const std::map<e_model, size_t> & MEM_REQ_EVAL()
 }
 
 // amount of VRAM needed per batch size to hold temporary results
-// the values for 3b and 65b are not derived from testing but instead chosen conservatively
+// the values for 3b are not derived from testing but instead chosen conservatively
 static const std::map<e_model, size_t> & VRAM_REQ_SCRATCH_BASE()
 {
     static std::map<e_model, size_t> k_sizes = {
@@ -152,14 +158,14 @@ static const std::map<e_model, size_t> & VRAM_REQ_SCRATCH_BASE()
         { MODEL_7B,   512ull * kB },
         { MODEL_13B,  640ull * kB },
         { MODEL_30B,  768ull * kB },
-        { MODEL_65B, 1536ull * kB },
-        { MODEL_70B, 1536ull * kB }, // TODO (likely can be reduced)
+        { MODEL_65B, 1360ull * kB },
+        { MODEL_70B, 1360ull * kB },
     };
     return k_sizes;
 }
 
 // amount of VRAM needed per batch size and context to hold temporary results
-// the values for 3b and 65b are not derived from testing but instead chosen conservatively
+// the values for 3b are not derived from testing but instead chosen conservatively
 static const std::map<e_model, size_t> & VRAM_REQ_SCRATCH_PER_CONTEXT()
 {
     static std::map<e_model, size_t> k_sizes = {
@@ -167,8 +173,8 @@ static const std::map<e_model, size_t> & VRAM_REQ_SCRATCH_PER_CONTEXT()
         { MODEL_7B,  128ull },
         { MODEL_13B, 160ull },
         { MODEL_30B, 208ull },
-        { MODEL_65B, 416ull },
-        { MODEL_70B, 416ull }, // TODO (likely can be reduced)
+        { MODEL_65B, 320ull },
+        { MODEL_70B, 320ull },
     };
     return k_sizes;
 }
@@ -328,13 +334,22 @@ struct llama_model {
 
 struct llama_context {
     llama_context(const llama_model & model) : model(model), t_load_us(model.t_load_us), t_start_us(model.t_start_us) {}
-#ifdef GGML_USE_METAL
     ~llama_context() {
+        if (model_owner) {
+            delete &model;
+        }
+#ifdef GGML_USE_METAL
         if (ctx_metal) {
             ggml_metal_free(ctx_metal);
         }
-    }
 #endif
+#ifdef LLAMA_USE_ALLOCATOR
+        if (alloc) {
+            ggml_allocr_free(alloc);
+        }
+#endif
+    }
+
     std::mt19937 rng;
 
     bool has_evaluated_once = false;
@@ -372,7 +387,17 @@ struct llama_context {
     // memory buffers used to evaluate the model
     // TODO: move in llama_state
     llama_ctx_buffer buf_compute;
+
+#ifdef LLAMA_USE_ALLOCATOR
+    llama_ctx_buffer buf_alloc;
+    ggml_allocr * alloc = NULL;
+#endif
+
+#ifdef LLAMA_USE_SCRATCH
     llama_ctx_buffer buf_scratch[LLAMA_MAX_SCRATCH_BUFFERS];
+    int    buf_last = 0;
+    size_t buf_max_size[LLAMA_MAX_SCRATCH_BUFFERS] = { 0 };
+#endif
 
 #ifdef GGML_USE_METAL
     ggml_metal_context * ctx_metal = NULL;
@@ -382,9 +407,6 @@ struct llama_context {
     ggml_mpi_context * ctx_mpi = NULL;
 #endif
 
-    int    buf_last = 0;
-    size_t buf_max_size[LLAMA_MAX_SCRATCH_BUFFERS] = { 0 };
-
     void use_buf(struct ggml_context * ctx, int i) {
 #if defined(LLAMA_USE_SCRATCH)
         size_t last_size = 0;
@@ -726,12 +748,12 @@ struct llama_model_loader {
 
     void load_all_data(llama_progress_callback progress_callback, void *  progress_callback_user_data, llama_mlock * lmlock) {
         size_t data_size = 0;
-        size_t prefetch_size = 0;
+        size_t prefetch_size = file_loader->file.size;
         size_t lock_size = 0;
         for (const llama_load_tensor & lt : tensors_map.tensors) {
             data_size += lt.size;
-            if (lt.ggml_tensor->backend == GGML_BACKEND_CPU) {
-                prefetch_size += lt.size;
+            if (lt.ggml_tensor->backend != GGML_BACKEND_CPU) {
+                prefetch_size -= lt.size;
             }
         }
 
@@ -880,6 +902,7 @@ struct llama_context_params llama_context_default_params() {
         /*.progress_callback           =*/ nullptr,
         /*.progress_callback_user_data =*/ nullptr,
         /*.low_vram                    =*/ false,
+        /*.mul_mat_q                   =*/ false,
         /*.f16_kv                      =*/ true,
         /*.logits_all                  =*/ false,
         /*.vocab_only                  =*/ false,
@@ -914,6 +937,11 @@ bool llama_mlock_supported() {
     return llama_mlock::SUPPORTED;
 }
 
+int get_blas_batch_mul(int batch)
+{
+    return (batch>512?(batch>1024?4:2):1);
+}
+
 void llama_backend_init(bool numa) {
     ggml_time_init();
 
@@ -1007,6 +1035,7 @@ static void llama_model_load_internal(
         int n_gpu_layers,
         int main_gpu,
         const float * tensor_split,
+        const bool mul_mat_q,
         float rope_freq_base,
         float rope_freq_scale,
         bool low_vram,
@@ -1018,7 +1047,7 @@ static void llama_model_load_internal(
         void * progress_callback_user_data) {
 
     model.t_start_us = ggml_time_us();
-    size_t blasbatchmul = (n_batch>512?2:1);
+    size_t blasbatchmul = get_blas_batch_mul(n_batch);
 
     std::unique_ptr<llama_model_loader> ml(new llama_model_loader(fname, use_mmap));
 
@@ -1052,7 +1081,7 @@ static void llama_model_load_internal(
         // LLaMAv2
         // TODO: temporary until GGUF
         //patch for llama2 gqa
-        if (model.type == e_model::MODEL_65B && hparams.n_mult >= 4096) {
+        if (model.type == e_model::MODEL_65B && (hparams.n_mult >= 4096 && hparams.n_mult != 5504)) {
             fprintf(stderr, "%s: Applying KCPP Patch for 70B model, setting GQA to 8\n", __func__);
             n_gqa = 8;
         }
@@ -1141,9 +1170,11 @@ static void llama_model_load_internal(
     }
 
     (void) main_gpu;
+    (void) mul_mat_q;
 #if defined(GGML_USE_CUBLAS)
     fprintf(stderr, "%s: using CUDA for GPU acceleration\n", __func__);
     ggml_cuda_set_main_device(main_gpu);
+    ggml_cuda_set_mul_mat_q(mul_mat_q);
 #define LLAMA_BACKEND_OFFLOAD       GGML_BACKEND_GPU
 #define LLAMA_BACKEND_OFFLOAD_SPLIT GGML_BACKEND_GPU_SPLIT
 #elif defined(GGML_USE_CLBLAST)
@@ -1237,12 +1268,16 @@ static void llama_model_load_internal(
         const size_t scale = memory_type == GGML_TYPE_F32 ? 2 : 1;
 
         // this is the total memory required to run the inference
-        const size_t mem_required =
+        size_t mem_required =
             ctx_size +
-            mmapped_size - vram_weights + // weights in VRAM not in memory
+            mmapped_size - vram_weights; // weights in VRAM not in memory
+
+#ifndef LLAMA_USE_ALLOCATOR
+        mem_required +=
             blasbatchmul*MEM_REQ_SCRATCH0(hparams.n_ctx).at(model.type) +
             blasbatchmul*MEM_REQ_SCRATCH1().at(model.type) +
             blasbatchmul*MEM_REQ_EVAL().at(model.type);
+#endif
 
         // this is the memory required by one llama_state
         const size_t mem_required_state =
@@ -1348,6 +1383,7 @@ static bool llama_model_load(
         int n_gpu_layers,
         int main_gpu,
         const float * tensor_split,
+        const bool mul_mat_q,
         float rope_freq_base,
         float rope_freq_scale,
         bool low_vram,
@@ -1358,7 +1394,8 @@ static bool llama_model_load(
         llama_progress_callback progress_callback,
         void *progress_callback_user_data) {
     try {
-        llama_model_load_internal(fname, model, vocab, n_ctx, n_batch, n_gqa, rms_norm_eps, n_gpu_layers, main_gpu, tensor_split, rope_freq_base, rope_freq_scale, low_vram, memory_type,
+        llama_model_load_internal(fname, model, vocab, n_ctx, n_batch, n_gqa, rms_norm_eps, n_gpu_layers,
+                                  main_gpu, tensor_split, mul_mat_q, rope_freq_base, rope_freq_scale, low_vram, memory_type,
                                   use_mmap, use_mlock, vocab_only, progress_callback, progress_callback_user_data);
         return true;
     } catch (const std::exception & err) {
@@ -1367,32 +1404,15 @@ static bool llama_model_load(
     }
 }
 
-// evaluate the transformer
-//
-//   - lctx:      llama context
-//   - tokens:    new batch of tokens to process
-//   - embd       embeddings input
-//   - n_tokens   number of tokens
-//   - n_past:    the context size so far
-//   - n_threads: number of threads to use
-//
-static bool llama_eval_internal(
+static struct ggml_cgraph * llama_build_graph(
          llama_context & lctx,
      const llama_token * tokens,
            const float * embd,
                    int   n_tokens,
-                   int   n_past,
-                   int   n_threads,
-            const char * cgraph_fname) {
+                   int   n_past) {
 
     LLAMA_ASSERT((!tokens && embd) || (tokens && !embd));
 
-#ifdef GGML_USE_MPI
-    ggml_mpi_eval_init(lctx.ctx_mpi, &n_tokens, &n_past, &n_threads);
-#endif
-
-    const int64_t t_start_us = ggml_time_us();
-
     const int N = n_tokens;
 
     const auto & model   = lctx.model;
@@ -1408,10 +1428,8 @@ static bool llama_eval_internal(
     const int64_t n_head      = hparams.n_head;
     const int64_t n_head_kv   = hparams.n_head_kv;
     const int64_t n_embd_head = hparams.n_embd_head();
-    const int64_t n_vocab     = hparams.n_vocab;
     const int64_t n_embd_gqa  = hparams.n_embd_gqa();
 
-
     LLAMA_ASSERT(n_embd_head == hparams.n_rot);
 
     const float freq_base  = hparams.rope_freq_base;
@@ -1423,26 +1441,35 @@ static bool llama_eval_internal(
     auto & mem_per_token = lctx.mem_per_token;
     auto & buf_compute   = lctx.buf_compute;
 
+
     struct ggml_init_params params = {
         /*.mem_size   =*/ buf_compute.size,
         /*.mem_buffer =*/ buf_compute.addr,
         /*.no_alloc   =*/ false,
     };
 
-    struct ggml_context * ctx0 = ggml_init(params);
+#ifdef LLAMA_USE_ALLOCATOR
+    params.no_alloc = true;
+#endif
 
-    ggml_cgraph gf = {};
+    struct ggml_context * ctx0 = ggml_init(params);
 
-    // for big prompts, if BLAS is enabled, it is better to use only one thread
-    // otherwise, the threads are spin-lock waiting for the BLAS calls and are degrading the performance
-    n_threads = N >= 32 && ggml_cpu_has_blas() && !ggml_cpu_has_gpublas() ? 1 : n_threads;
+    ggml_cgraph * gf = ggml_new_graph(ctx0);
 
     struct ggml_tensor * cur;
     struct ggml_tensor * inpL;
 
     if (tokens) {
         struct ggml_tensor * inp_tokens = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N);
+
+#ifdef LLAMA_USE_ALLOCATOR
+        ggml_allocr_alloc(lctx.alloc, inp_tokens);
+        if (!ggml_allocr_is_measure(lctx.alloc)) {
+            memcpy(inp_tokens->data, tokens, N*ggml_element_size(inp_tokens));
+        }
+#else
         memcpy(inp_tokens->data, tokens, N*ggml_element_size(inp_tokens));
+#endif
         ggml_set_name(inp_tokens, "inp_tokens");
 
         inpL = ggml_get_rows(ctx0, model.tok_embeddings, inp_tokens);
@@ -1452,7 +1479,15 @@ static bool llama_eval_internal(
 #endif
 
         inpL = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, N);
+
+#ifdef LLAMA_USE_ALLOCATOR
+        ggml_allocr_alloc(lctx.alloc, inpL);
+        if (!ggml_allocr_is_measure(lctx.alloc)) {
+            memcpy(inpL->data, embd, N * n_embd * ggml_element_size(inpL));
+        }
+#else
         memcpy(inpL->data, embd, N * n_embd * ggml_element_size(inpL));
+#endif
     }
 
     const int i_gpu_start = n_layer - n_gpu_layers;
@@ -1479,6 +1514,17 @@ static bool llama_eval_internal(
     }
 #endif // GGML_USE_CUBLAS
 
+    struct ggml_tensor * KQ_scale = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, 1);
+#ifdef LLAMA_USE_ALLOCATOR
+    ggml_allocr_alloc(lctx.alloc, KQ_scale);
+    if (!ggml_allocr_is_measure(lctx.alloc)) {
+        ggml_set_f32(KQ_scale, 1.0f/sqrtf(float(n_embd)/n_head));
+    }
+#else
+    ggml_set_f32(KQ_scale, 1.0f/sqrtf(float(n_embd)/n_head));
+#endif
+    ggml_set_name(KQ_scale, "1/sqrt(n_embd_head)");
+
     for (int il = 0; il < n_layer; ++il) {
         ggml_format_name(inpL, "layer_inp_%d", il);
 
@@ -1548,8 +1594,8 @@ static bool llama_eval_internal(
                 ggml_set_name(v, "v");
 
                 // important: storing RoPE-ed version of K in the KV cache!
-                ggml_build_forward_expand(&gf, ggml_cpy(ctx0, Kcur, k));
-                ggml_build_forward_expand(&gf, ggml_cpy(ctx0, Vcur, v));
+                ggml_build_forward_expand(gf, ggml_cpy(ctx0, Kcur, k));
+                ggml_build_forward_expand(gf, ggml_cpy(ctx0, Vcur, v));
             }
 
             struct ggml_tensor * Q =
@@ -1574,9 +1620,6 @@ static bool llama_eval_internal(
             ggml_set_name(KQ, "KQ");
 
             // KQ_scaled = KQ / sqrt(n_embd_head)
-            struct ggml_tensor * KQ_scale = ggml_new_f32(ctx0, 1.0f/sqrtf(float(n_embd)/n_head));
-            ggml_set_name(KQ_scale, "1/sqrt(n_embd_head)");
-
             // KQ_scaled shape [n_past + N, N, n_head, 1]
             struct ggml_tensor * KQ_scaled = ggml_scale_inplace(ctx0, KQ, KQ_scale);
             offload_func_kq(KQ_scaled);
@@ -1692,9 +1735,6 @@ static bool llama_eval_internal(
 
     lctx.use_buf(ctx0, 0);
 
-    // used at the end to optionally extract the embeddings
-    struct ggml_tensor * embeddings = NULL;
-
     // norm
     {
         cur = ggml_rms_norm(ctx0, inpL, rms_norm_eps);
@@ -1705,8 +1745,6 @@ static bool llama_eval_internal(
         cur = ggml_mul(ctx0, cur, model.norm);
         // offload_func_nr(cur); // TODO CPU + GPU mirrored backend
         ggml_set_name(cur, "result_norm");
-
-        embeddings = cur;
     }
 
     // lm_head
@@ -1718,23 +1756,103 @@ static bool llama_eval_internal(
     // logits -> probs
     //cur = ggml_soft_max_inplace(ctx0, cur);
 
-    // run the computation
-    ggml_build_forward_expand(&gf, cur);
+    ggml_build_forward_expand(gf, cur);
 
-    // fprintf(stderr, "graph build time: %.3f ms (%d nodes, %d leafs)\n", (ggml_time_us() - t_start_us)/1000.0, gf.n_nodes, gf.n_leafs);
+    if (mem_per_token == 0) {
+        mem_per_token = ggml_used_mem(ctx0)/N;
+    }
+
+#if 0
+    printf("\n%s: used_mem: eval ctx %.3f MB, scratch %.3f MB %.3f MB, work buf %.3f MB, n_past = %d, N = %d\n", __func__,
+            ggml_used_mem(ctx0)/1024.0/1024.0,
+            lctx.get_buf_max_mem(0)/1024.0/1024.0,
+            lctx.get_buf_max_mem(1)/1024.0/1024.0,
+            lctx.work_buffer.size()/1024.0/1024.0,
+            n_past, N);
+#endif
+
+    ggml_free(ctx0);
+
+    return gf;
+}
+
+// evaluate the transformer
+//
+//   - lctx:      llama context
+//   - tokens:    new batch of tokens to process
+//   - embd       embeddings input
+//   - n_tokens   number of tokens
+//   - n_past:    the context size so far
+//   - n_threads: number of threads to use
+//
+static bool llama_eval_internal(
+         llama_context & lctx,
+     const llama_token * tokens,
+           const float * embd,
+                   int   n_tokens,
+                   int   n_past,
+                   int   n_threads,
+            const char * cgraph_fname) {
+
+    LLAMA_ASSERT((!tokens && embd) || (tokens && !embd));
+
+    const int64_t t_start_us = ggml_time_us();
+
+#ifdef GGML_USE_MPI
+    ggml_mpi_eval_init(lctx.ctx_mpi, &n_tokens, &n_past, &n_threads);
+#endif
+
+    const int N = n_tokens;
+
+    const auto & model   = lctx.model;
+    const auto & hparams = model.hparams;
+
+    const auto & kv_self = lctx.kv_self;
+
+    LLAMA_ASSERT(!!kv_self.ctx);
+
+    const int64_t n_embd      = hparams.n_embd;
+    const int64_t n_vocab     = hparams.n_vocab;
+
+#ifdef LLAMA_USE_ALLOCATOR
+    ggml_allocr_reset(lctx.alloc);
+#endif
+
+    ggml_cgraph * gf = llama_build_graph(lctx, tokens, embd, n_tokens, n_past);
+
+#ifdef LLAMA_USE_ALLOCATOR
+    ggml_allocr_alloc_graph(lctx.alloc, gf);
+#endif
+
+    // fprintf(stderr, "graph build time: %.3f ms (%d nodes, %d leafs)\n", (ggml_time_us() - t_start_us)/1000.0, gf->n_nodes, gf->n_leafs);
+
+    // for big prompts, if BLAS is enabled, it is better to use only one thread
+    // otherwise, the threads are spin-lock waiting for the BLAS calls and are degrading the performance
+    n_threads = N >= 32 && ggml_cpu_has_blas() && !ggml_cpu_has_gpublas() ? 1 : n_threads;
+
+    struct ggml_tensor * res = gf->nodes[gf->n_nodes - 1];
+    struct ggml_tensor * embeddings = gf->nodes[gf->n_nodes - 2];
+
+    LLAMA_ASSERT(strcmp(res->name, "result_output") == 0);
+    LLAMA_ASSERT(strcmp(embeddings->name, "result_norm") == 0);
 
 #if GGML_USE_MPI
-    ggml_mpi_graph_compute_pre(lctx.ctx_mpi, &gf, n_layer);
+    const int64_t n_layer = hparams.n_layer;
+    ggml_mpi_graph_compute_pre(lctx.ctx_mpi, gf, n_layer);
 #endif
 
 #ifdef GGML_USE_METAL
     if (lctx.ctx_metal && N == 1) {
-        if (!ggml_metal_if_optimized(lctx.ctx_metal)) {
-            ggml_metal_graph_find_concurrency(lctx.ctx_metal,&gf);
-        }
+        // TODO: disabled until #2413 is resolved
+        //if (!ggml_metal_if_optimized(lctx.ctx_metal)) {
+        //    ggml_metal_graph_find_concurrency(lctx.ctx_metal, gf);
+        //}
         ggml_metal_set_n_cb     (lctx.ctx_metal, n_threads);
-        ggml_metal_graph_compute(lctx.ctx_metal, &gf);
-        ggml_metal_get_tensor   (lctx.ctx_metal, cur);
+        ggml_metal_graph_compute(lctx.ctx_metal, gf);
+        ggml_metal_get_tensor   (lctx.ctx_metal, res);
+        if (!lctx.embedding.empty()) {
+            ggml_metal_get_tensor(lctx.ctx_metal, embeddings);
+        }
     } else {
         // IMPORTANT:
         // Since we don't have efficient Matrix x Matrix Metal multiplication yet, we fallback to vanilla
@@ -1752,34 +1870,32 @@ static bool llama_eval_internal(
             ggml_metal_get_tensor(lctx.ctx_metal, kv_self.v);
         }
 
-        ggml_graph_compute_helper(lctx.work_buffer, &gf, n_threads);
+        ggml_graph_compute_helper(lctx.work_buffer, gf, n_threads);
     }
 #else
-    ggml_graph_compute_helper(lctx.work_buffer, &gf, n_threads);
+    ggml_graph_compute_helper(lctx.work_buffer, gf, n_threads);
 #endif
 
 #if GGML_USE_MPI
-    ggml_mpi_graph_compute_post(lctx.ctx_mpi, &gf, n_layer);
+    ggml_mpi_graph_compute_post(lctx.ctx_mpi, gf, n_layer);
 #endif
 
     // update kv token count
     lctx.kv_self.n = n_past + N;
 
-    struct ggml_tensor * res = gf.nodes[gf.n_nodes - 1];
-
     if (cgraph_fname) {
-        ggml_graph_export(&gf, cgraph_fname);
+        ggml_graph_export(gf, cgraph_fname);
     }
 
 #ifdef GGML_PERF
     // print timing information per ggml operation (for debugging purposes)
     // requires GGML_PERF to be defined
-    ggml_graph_print(&gf);
+    ggml_graph_print(gf);
 #endif
 
     // plot the computation graph in dot format (for debugging purposes)
     //if (n_past%100 == 0) {
-    //    ggml_graph_dump_dot(&gf, NULL, "llama.dot");
+    //    ggml_graph_dump_dot(gf, NULL, "llama.dot");
     //}
 
     // extract logits
@@ -1804,21 +1920,6 @@ static bool llama_eval_internal(
         memcpy(embedding_out.data(), (float *) ggml_get_data(embeddings) + (n_embd*(N - 1)), sizeof(float)*n_embd);
     }
 
-    if (mem_per_token == 0) {
-        mem_per_token = ggml_used_mem(ctx0)/N;
-    }
-
-#if 0
-    printf("\n%s: used_mem: eval ctx %.3f MB, scratch %.3f MB %.3f MB, work buf %.3f MB, n_past = %d, N = %d\n", __func__,
-            ggml_used_mem(ctx0)/1024.0/1024.0,
-            lctx.get_buf_max_mem(0)/1024.0/1024.0,
-            lctx.get_buf_max_mem(1)/1024.0/1024.0,
-            lctx.work_buffer.size()/1024.0/1024.0,
-            n_past, N);
-#endif
-
-    ggml_free(ctx0);
-
     // measure the performance only for the single-token evals
     if (N == 1) {
         lctx.t_eval_us += ggml_time_us() - t_start_us;
@@ -1930,7 +2031,9 @@ struct llama_tokenizer {
             if (token == vocab_.token_to_id.end()) {
                 // output any symbols that did not form tokens as bytes.
                 for (int j = 0; j < (int) symbol.n; ++j) {
-                    llama_vocab::id token_id = static_cast<uint8_t>(symbol.text[j]) + 3;
+                    // NOTE: old version, before #2420 - not sure what are the implications of this
+                    //llama_vocab::id token_id = static_cast<uint8_t>(symbol.text[j]) + 3;
+                    llama_vocab::id token_id = vocab_.token_to_id.at(std::string(1, symbol.text[j]));
                     output.push_back(token_id);
                 }
             } else {
@@ -3107,7 +3210,7 @@ struct llama_model * llama_load_model_from_file(
     ggml_type memory_type = params.f16_kv ? GGML_TYPE_F16 : GGML_TYPE_F32;
 
     if (!llama_model_load(path_model, *model, model->vocab, params.n_ctx, params.n_batch, params.n_gqa, params.rms_norm_eps, params.n_gpu_layers,
-                params.main_gpu, params.tensor_split, params.rope_freq_base, params.rope_freq_scale,params.low_vram,
+                params.main_gpu, params.tensor_split, params.mul_mat_q, params.rope_freq_base, params.rope_freq_scale,params.low_vram,
                 memory_type, params.use_mmap, params.use_mlock, params.vocab_only, params.progress_callback,
                 params.progress_callback_user_data)) {
         delete model;
@@ -3136,7 +3239,7 @@ struct llama_context * llama_new_context_with_model(
         params.seed = time(NULL);
     }
 
-    size_t blasbatchmul = (params.n_batch>512?2:1);
+    size_t blasbatchmul = get_blas_batch_mul(params.n_batch);
 
     unsigned cur_percentage = 0;
     if (params.progress_callback == NULL) {
@@ -3186,10 +3289,47 @@ struct llama_context * llama_new_context_with_model(
             ctx->embedding.resize(hparams.n_embd);
         }
 
-        ctx->buf_compute.resize(blasbatchmul*MEM_REQ_EVAL().at(ctx->model.type));
+#ifdef LLAMA_USE_ALLOCATOR
+        {
+            static const size_t tensor_alignment = 32;
+            // the compute buffer is used to store the tensor and graph structs, while the allocator buffer is used for the tensor data
+            ctx->buf_compute.resize(ggml_tensor_overhead()*GGML_MAX_NODES + ggml_graph_overhead());
+
+            // create measure allocator
+            ctx->alloc = ggml_allocr_new_measure(tensor_alignment);
+
+            // build worst-case graph
+            int n_tokens = std::min((int)hparams.n_ctx, params.n_batch);
+            int n_past = hparams.n_ctx - n_tokens;
+            llama_token token = llama_token_bos(); // not actually used by llama_build_graph, but required to choose between token and embedding inputs graph
+            ggml_cgraph * gf = llama_build_graph(*ctx, &token, NULL, n_tokens, n_past);
+
+            // measure memory requirements for the graph
+            size_t alloc_size = ggml_allocr_alloc_graph(ctx->alloc, gf) + tensor_alignment;
+
+            fprintf(stderr, "%s: compute buffer total size = %7.2f MB\n", __func__, (ctx->buf_compute.size + alloc_size) / 1024.0 / 1024.0);
 
+            // debug - for comparison with scratch buffer
+            //size_t prev_req =
+            //    MEM_REQ_SCRATCH0(hparams.n_ctx).at(ctx->model.type) +
+            //    MEM_REQ_SCRATCH1().at(ctx->model.type) +
+            //    MEM_REQ_EVAL().at(ctx->model.type);
+            //fprintf(stderr, "%s: (debug) equivalent with scratch buffer = %7.2f MB\n", __func__, prev_req / 1024.0 / 1024.0);
+
+            // recreate allocator with exact memory requirements
+            ggml_allocr_free(ctx->alloc);
+
+            ctx->buf_alloc.resize(alloc_size);
+            ctx->alloc = ggml_allocr_new(ctx->buf_alloc.addr, ctx->buf_alloc.size, tensor_alignment);
+        }
+#else
+        ctx->buf_compute.resize(blasbatchmul*MEM_REQ_EVAL().at(ctx->model.type) + ggml_graph_overhead());
+#endif
+
+#ifdef LLAMA_USE_SCRATCH
         ctx->buf_scratch[0].resize(blasbatchmul*MEM_REQ_SCRATCH0(hparams.n_ctx).at(ctx->model.type));
         ctx->buf_scratch[1].resize(blasbatchmul*MEM_REQ_SCRATCH1().at(ctx->model.type));
+#endif
     }
 
 #ifdef GGML_USE_METAL
@@ -3259,9 +3399,6 @@ struct llama_context * llama_init_from_file(
 }
 
 void llama_free(struct llama_context * ctx) {
-    if (ctx->model_owner) {
-        delete &ctx->model;
-    }
     delete ctx;
 }
 
@@ -3620,10 +3757,20 @@ size_t llama_get_state_size(const struct llama_context * ctx) {
     return s_total;
 }
 
-// Copies the state to the specified destination address
-size_t llama_copy_state_data(struct llama_context * ctx, uint8_t * dst) {
-    uint8_t * out = dst;
-
+/** copy state data into either a buffer or file depending on the passed in context
+ *
+ * file context:
+ * llama_file file("/path", "wb");
+ * llama_data_file_context data_ctx(&file);
+ * llama_copy_state_data(ctx, &data_ctx);
+ *
+ * buffer context:
+ * std::vector<uint8_t> buf(max_size, 0);
+ * llama_data_buffer_context data_ctx(&buf.data());
+ * llama_copy_state_data(ctx, &data_ctx);
+ *
+*/
+void llama_copy_state_data_internal(struct llama_context * ctx, llama_data_context * data_ctx) {
     // copy rng
     {
         std::stringstream rng_ss;
@@ -3635,8 +3782,8 @@ size_t llama_copy_state_data(struct llama_context * ctx, uint8_t * dst) {
         memset(&rng_buf[0], 0, LLAMA_MAX_RNG_STATE);
         memcpy(&rng_buf[0], rng_ss.str().data(), rng_ss.str().size());
 
-        memcpy(out, &rng_size,   sizeof(rng_size));    out += sizeof(rng_size);
-        memcpy(out, &rng_buf[0], LLAMA_MAX_RNG_STATE); out += LLAMA_MAX_RNG_STATE;
+        data_ctx->write(&rng_size,   sizeof(rng_size));
+        data_ctx->write(&rng_buf[0], LLAMA_MAX_RNG_STATE);
     }
 
     // copy logits
@@ -3644,25 +3791,29 @@ size_t llama_copy_state_data(struct llama_context * ctx, uint8_t * dst) {
         const size_t logits_cap  = ctx->logits.capacity();
         const size_t logits_size = ctx->logits.size();
 
-        memcpy(out, &logits_cap,  sizeof(logits_cap));  out += sizeof(logits_cap);
-        memcpy(out, &logits_size, sizeof(logits_size)); out += sizeof(logits_size);
+        data_ctx->write(&logits_cap,  sizeof(logits_cap));
+        data_ctx->write(&logits_size, sizeof(logits_size));
 
         if (logits_size) {
-            memcpy(out, ctx->logits.data(), logits_size * sizeof(float));
+            data_ctx->write(ctx->logits.data(), logits_size * sizeof(float));
         }
 
-        out += logits_cap * sizeof(float);
+        // If there is a gap between the size and the capacity, write padding
+        size_t padding_size = (logits_cap - logits_size) * sizeof(float);
+        if (padding_size > 0) {
+            std::vector<uint8_t> padding(padding_size, 0); // Create a buffer filled with zeros
+            data_ctx->write(padding.data(), padding_size);
+        }
     }
 
     // copy embeddings
     {
         const size_t embedding_size = ctx->embedding.size();
 
-        memcpy(out, &embedding_size, sizeof(embedding_size)); out += sizeof(embedding_size);
+        data_ctx->write(&embedding_size, sizeof(embedding_size));
 
         if (embedding_size) {
-            memcpy(out, ctx->embedding.data(), embedding_size * sizeof(float));
-            out += embedding_size * sizeof(float);
+            data_ctx->write(ctx->embedding.data(), embedding_size * sizeof(float));
         }
     }
 
@@ -3671,14 +3822,14 @@ size_t llama_copy_state_data(struct llama_context * ctx, uint8_t * dst) {
         const auto & kv_self = ctx->kv_self;
         const auto & hparams = ctx->model.hparams;
         const int    n_layer = hparams.n_layer;
-        const int    n_embd  = hparams.n_embd;
+        const int    n_embd  = hparams.n_embd_gqa();
         const int    n_ctx   = hparams.n_ctx;
 
         const size_t kv_size = kv_self.buf.size;
         const int    kv_ntok = llama_get_kv_cache_token_count(ctx);
 
-        memcpy(out, &kv_size, sizeof(kv_size)); out += sizeof(kv_size);
-        memcpy(out, &kv_ntok, sizeof(kv_ntok)); out += sizeof(kv_ntok);
+        data_ctx->write(&kv_size, sizeof(kv_size));
+        data_ctx->write(&kv_ntok, sizeof(kv_ntok));
 
         if (kv_size) {
             const size_t elt_size = ggml_element_size(kv_self.k);
@@ -3687,12 +3838,12 @@ size_t llama_copy_state_data(struct llama_context * ctx, uint8_t * dst) {
             ggml_cgraph gf{};
 
             ggml_tensor * kout3d = ggml_new_tensor_3d(cpy_ctx, kv_self.k->type, n_embd, kv_ntok, n_layer);
-            kout3d->data = out;
-            out += ggml_nbytes(kout3d);
+            std::vector<uint8_t> kout3d_data(ggml_nbytes(kout3d), 0);
+            kout3d->data = kout3d_data.data();
 
             ggml_tensor * vout3d = ggml_new_tensor_3d(cpy_ctx, kv_self.v->type, kv_ntok, n_embd, n_layer);
-            vout3d->data = out;
-            out += ggml_nbytes(vout3d);
+            std::vector<uint8_t> vout3d_data(ggml_nbytes(vout3d), 0);
+            vout3d->data = vout3d_data.data();
 
             ggml_tensor * k3d = ggml_view_3d(cpy_ctx, kv_self.k,
                 n_embd, kv_ntok, n_layer,
@@ -3707,15 +3858,20 @@ size_t llama_copy_state_data(struct llama_context * ctx, uint8_t * dst) {
             ggml_graph_compute_helper(ctx->work_buffer, &gf, /*n_threads*/ 1);
 
             ggml_free(cpy_ctx);
+
+            // our data is now in the kout3d_data and vout3d_data buffers
+            // write them to file
+            data_ctx->write(kout3d_data.data(), kout3d_data.size());
+            data_ctx->write(vout3d_data.data(), vout3d_data.size());
         }
     }
+}
 
-    const size_t written  = out - dst;
-    const size_t max_size = llama_get_state_size(ctx);
-
-    LLAMA_ASSERT(written <= max_size);
+size_t llama_copy_state_data(struct llama_context * ctx, uint8_t * dst) {
+    llama_data_buffer_context data_ctx(dst);
+    llama_copy_state_data_internal(ctx, &data_ctx);
 
-    return written;
+    return data_ctx.get_size_written();
 }
 
 // Sets the state reading from the specified source address
@@ -3774,7 +3930,7 @@ size_t llama_set_state_data(struct llama_context * ctx, uint8_t * src) {
         const auto & kv_self = ctx->kv_self;
         const auto & hparams = ctx->model.hparams;
         const int    n_layer = hparams.n_layer;
-        const int    n_embd  = hparams.n_embd;
+        const int    n_embd  = hparams.n_embd_gqa();
         const int    n_ctx   = hparams.n_ctx;
 
         size_t kv_size;
@@ -3900,15 +4056,9 @@ bool llama_save_session_file(struct llama_context * ctx, const char * path_sessi
     file.write_u32((uint32_t) n_token_count);
     file.write_raw(tokens, sizeof(llama_token) * n_token_count);
 
-    // save the context state
-    {
-        const size_t n_state_size_max = llama_get_state_size(ctx);
-
-        std::vector<uint8_t> state_data(n_state_size_max);
-        const size_t n_state_size_cur = llama_copy_state_data(ctx, state_data.data());
-
-        file.write_raw(state_data.data(), n_state_size_cur);
-    }
+    // save the context state using stream saving
+    llama_data_file_context data_ctx(&file);
+    llama_copy_state_data_internal(ctx, &data_ctx);
 
     return true;
 }

llama.h

@@ -108,6 +108,7 @@ extern "C" {
 
         // Keep the booleans together to avoid misalignment during copy-by-value.
         bool low_vram;   // if true, reduce VRAM usage at the cost of performance
+        bool mul_mat_q;  // if true, use experimental mul_mat_q kernels
         bool f16_kv;     // use fp16 for KV cache
         bool logits_all; // the llama_eval() call computes all logits, not just the last one
         bool vocab_only; // only load the vocabulary, no weights

make_old_pyinstaller.bat

@@ -1,4 +1,4 @@
 echo This file is only for my own usage, please do not use it. I am lazy.
 
 set PATH=d:\\MainApplications\\KoboldAIGPT\\KoboldAI-Horde-Bridge\\python;d:\\MainApplications\\KoboldAIGPT\\KoboldAI-Horde-Bridge\\python\\Scripts;%PATH%
-PyInstaller --noconfirm --onefile --clean --console --collect-all customtkinter --icon "./niko.ico" --add-data "./klite.embd;." --add-data "./koboldcpp.dll;." --add-data "./koboldcpp_openblas.dll;." --add-data "./koboldcpp_failsafe.dll;." --add-data "./koboldcpp_openblas_noavx2.dll;." --add-data "./libopenblas.dll;." --add-data "./koboldcpp_clblast.dll;." --add-data "./clblast.dll;." --add-data "./rwkv_vocab.embd;." --add-data "./rwkv_world_vocab.embd;." "./koboldcpp.py" -n "koboldcpp_nocuda.exe"
+PyInstaller --noconfirm --onefile --clean --console --collect-all customtkinter --icon "./niko.ico" --add-data "./klite.embd;." --add-data "./koboldcpp.dll;." --add-data "./koboldcpp_openblas.dll;." --add-data "./koboldcpp_failsafe.dll;." --add-data "./koboldcpp_noavx2.dll;." --add-data "./libopenblas.dll;." --add-data "./koboldcpp_clblast.dll;." --add-data "./clblast.dll;." --add-data "./rwkv_vocab.embd;." --add-data "./rwkv_world_vocab.embd;." "./koboldcpp.py" -n "koboldcpp_nocuda.exe"

make_old_pyinstaller_cuda.bat

@@ -1,4 +1,4 @@
 echo This file is only for my own usage, please do not use it. I am lazy.
 
 set PATH=d:\\MainApplications\\KoboldAIGPT\\KoboldAI-Horde-Bridge\\python;d:\\MainApplications\\KoboldAIGPT\\KoboldAI-Horde-Bridge\\python\\Scripts;%PATH%
-PyInstaller --noconfirm --onefile --clean --console --collect-all customtkinter --icon "./nikogreen.ico" --add-data "./klite.embd;." --add-data "./koboldcpp.dll;." --add-data "./koboldcpp_openblas.dll;." --add-data "./koboldcpp_failsafe.dll;." --add-data "./koboldcpp_openblas_noavx2.dll;." --add-data "./libopenblas.dll;." --add-data "./koboldcpp_clblast.dll;." --add-data "./clblast.dll;." --add-data "./koboldcpp_cublas.dll;." --add-data "./cublas64_11.dll;." --add-data "./cublasLt64_11.dll;." --add-data "./cudart64_110.dll;." --add-data "./msvcp140.dll;." --add-data "./vcruntime140.dll;." --add-data "./vcruntime140_1.dll;." --add-data "./rwkv_vocab.embd;." --add-data "./rwkv_world_vocab.embd;." "./koboldcpp.py" -n "koboldcpp.exe"
+PyInstaller --noconfirm --onefile --clean --console --collect-all customtkinter --icon "./nikogreen.ico" --add-data "./klite.embd;." --add-data "./koboldcpp.dll;." --add-data "./koboldcpp_openblas.dll;." --add-data "./koboldcpp_failsafe.dll;." --add-data "./koboldcpp_noavx2.dll;." --add-data "./libopenblas.dll;." --add-data "./koboldcpp_clblast.dll;." --add-data "./clblast.dll;." --add-data "./koboldcpp_cublas.dll;." --add-data "./cublas64_11.dll;." --add-data "./cublasLt64_11.dll;." --add-data "./cudart64_110.dll;." --add-data "./msvcp140.dll;." --add-data "./vcruntime140.dll;." --add-data "./vcruntime140_1.dll;." --add-data "./rwkv_vocab.embd;." --add-data "./rwkv_world_vocab.embd;." "./koboldcpp.py" -n "koboldcpp.exe"

make_pyinstaller.bat

@@ -1 +1 @@
-PyInstaller --noconfirm --onefile --clean --console --collect-all customtkinter --icon "./niko.ico" --add-data "./klite.embd;." --add-data "./koboldcpp.dll;." --add-data "./koboldcpp_openblas.dll;." --add-data "./koboldcpp_failsafe.dll;." --add-data "./koboldcpp_openblas_noavx2.dll;." --add-data "./libopenblas.dll;." --add-data "./koboldcpp_clblast.dll;." --add-data "./clblast.dll;." --add-data "./rwkv_vocab.embd;." --add-data "./rwkv_world_vocab.embd;." "./koboldcpp.py" -n "koboldcpp.exe"
+PyInstaller --noconfirm --onefile --clean --console --collect-all customtkinter --icon "./niko.ico" --add-data "./klite.embd;." --add-data "./koboldcpp.dll;." --add-data "./koboldcpp_openblas.dll;." --add-data "./koboldcpp_failsafe.dll;." --add-data "./koboldcpp_noavx2.dll;." --add-data "./libopenblas.dll;." --add-data "./koboldcpp_clblast.dll;." --add-data "./clblast.dll;." --add-data "./rwkv_vocab.embd;." --add-data "./rwkv_world_vocab.embd;." "./koboldcpp.py" -n "koboldcpp.exe"

make_pyinstaller.sh

@@ -5,7 +5,7 @@ pyinstaller --noconfirm --onefile --clean --console --collect-all customtkinter
 --add-data "./koboldcpp.so:." \
 --add-data "./koboldcpp_openblas.so:." \
 --add-data "./koboldcpp_failsafe.so:." \
---add-data "./koboldcpp_openblas_noavx2.so:." \
+--add-data "./koboldcpp_noavx2.so:." \
 --add-data "./koboldcpp_clblast.so:." \
 --add-data "./rwkv_vocab.embd:." \
 --add-data "./rwkv_world_vocab.embd:." \

make_pyinstaller_hybrid_henk.bat

@@ -2,4 +2,4 @@ cd /d "%~dp0"
 copy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.4\bin\cudart64_110.dll" .\ /Y
 copy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.4\bin\cublasLt64_11.dll" .\ /Y
 copy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.4\bin\cublas64_11.dll" .\ /Y
-PyInstaller --noconfirm --onefile --collect-all customtkinter --clean --console --icon ".\niko.ico" --add-data "./klite.embd;." --add-data "./koboldcpp.dll;." --add-data "./koboldcpp_openblas.dll;." --add-data "./koboldcpp_failsafe.dll;." --add-data "./koboldcpp_openblas_noavx2.dll;." --add-data "./libopenblas.dll;." --add-data "./koboldcpp_clblast.dll;." --add-data "./clblast.dll;." --add-data "./koboldcpp_cublas.dll;." --add-data "./cudart64_110.dll;." --add-data "./cublasLt64_11.dll;." --add-data "./cublas64_11.dll;." --add-data "./rwkv_vocab.embd;." --add-data "C:/Windows/System32/msvcp140.dll;." --add-data "C:/Windows/System32/vcruntime140_1.dll;." "./koboldcpp.py" -n "koboldcpp.exe"
+PyInstaller --noconfirm --onefile --collect-all customtkinter --clean --console --icon ".\niko.ico" --add-data "./klite.embd;." --add-data "./koboldcpp.dll;." --add-data "./koboldcpp_openblas.dll;." --add-data "./koboldcpp_failsafe.dll;." --add-data "./koboldcpp_noavx2.dll;." --add-data "./libopenblas.dll;." --add-data "./koboldcpp_clblast.dll;." --add-data "./clblast.dll;." --add-data "./koboldcpp_cublas.dll;." --add-data "./cudart64_110.dll;." --add-data "./cublasLt64_11.dll;." --add-data "./cublas64_11.dll;." --add-data "./rwkv_vocab.embd;." --add-data "C:/Windows/System32/msvcp140.dll;." --add-data "C:/Windows/System32/vcruntime140_1.dll;." "./koboldcpp.py" -n "koboldcpp.exe"