Move llava back to examples

.github/ISSUE_TEMPLATE/bug.md

@@ -1,8 +1,7 @@
 ---
-name: Issue and enhancement template
+name: Bug template
-about: Used to report issues and request enhancements for llama.cpp
+about: Used to report bugs in llama.cpp
-title: "[User] Insert summary of your issue or enhancement.."
+labels: ["bug-unconfirmed"]
-labels: ''
 assignees: ''
 
 ---
@@ -46,7 +45,7 @@ $ g++ --version
 
 # Failure Information (for bugs)
 
-Please help provide information about the failure if this is a bug. If it is not a bug, please remove the rest of this template.
+Please help provide information about the failure / bug.
 
 # Steps to Reproduce
 

.github/ISSUE_TEMPLATE/enhancement.md

@@ -0,0 +1,28 @@
+---
+name: Enhancement template
+about: Used to request enhancements for llama.cpp
+labels: ["enhancement"]
+assignees: ''
+
+---
+
+# Prerequisites
+
+Please answer the following questions for yourself before submitting an issue.
+
+- [ ] I am running the latest code. Development is very rapid so there are no tagged versions as of now.
+- [ ] I carefully followed the [README.md](https://github.com/ggerganov/llama.cpp/blob/master/README.md).
+- [ ] I [searched using keywords relevant to my issue](https://docs.github.com/en/issues/tracking-your-work-with-issues/filtering-and-searching-issues-and-pull-requests) to make sure that I am creating a new issue that is not already open (or closed).
+- [ ] I reviewed the [Discussions](https://github.com/ggerganov/llama.cpp/discussions), and have a new bug or useful enhancement to share.
+
+# Feature Description
+
+Please provide a detailed written description of what you were trying to do, and what you expected `llama.cpp` to do as an enhancement.
+
+# Motivation
+
+Please provide a detailed written description of reasons why this feature is necessary and how it is useful to `llama.cpp` users.
+
+# Possible Implementation
+
+If you have an idea as to how it can be implemented, please write a detailed description. Feel free to give links to external sources or share visuals that might be helpful to understand the details better.

.gitignore

@@ -10,10 +10,12 @@
 *.gcno
 *.gcda
 *.dot
+*.bat
 *.metallib
 .DS_Store
 .build/
 .cache/
+.ccls-cache/
 .direnv/
 .envrc
 .swiftpm
@@ -63,7 +65,7 @@ models-mnt
 /parallel
 /train-text-from-scratch
 /vdot
-build-info.h
+/common/build-info.cpp
 arm_neon.h
 compile_commands.json
 CMakeSettings.json

CMakeLists.txt

@@ -44,7 +44,7 @@ endif()
 
 # general
 option(LLAMA_STATIC                     "llama: static link libraries"                          OFF)
-option(LLAMA_NATIVE                     "llama: enable -march=native flag"                      ON)
+option(LLAMA_NATIVE                     "llama: enable -march=native flag"                      OFF)
 option(LLAMA_LTO                        "llama: enable link time optimization"                  OFF)
 
 # debug
@@ -82,6 +82,7 @@ set(LLAMA_BLAS_VENDOR "Generic" CACHE STRING "llama: BLAS library vendor")
 option(LLAMA_CUBLAS                          "llama: use CUDA"                                  OFF)
 #option(LLAMA_CUDA_CUBLAS                     "llama: use cuBLAS for prompt processing"          OFF)
 option(LLAMA_CUDA_FORCE_DMMV                 "llama: use dmmv instead of mmvq CUDA kernels"     OFF)
+option(LLAMA_CUDA_FORCE_MMQ                  "llama: use mmq kernels instead of cuBLAS"         OFF)
 set(LLAMA_CUDA_DMMV_X      "32" CACHE STRING "llama: x stride for dmmv CUDA kernels")
 set(LLAMA_CUDA_MMV_Y        "1" CACHE STRING "llama: y block size for mmv CUDA kernels")
 option(LLAMA_CUDA_F16                        "llama: use 16 bit floats for some calculations"   OFF)
@@ -93,46 +94,12 @@ option(LLAMA_CLBLAST                         "llama: use CLBlast"
 option(LLAMA_METAL                           "llama: use Metal"                                 ${LLAMA_METAL_DEFAULT})
 option(LLAMA_METAL_NDEBUG                    "llama: disable Metal debugging"                   OFF)
 option(LLAMA_MPI                             "llama: use MPI"                                   OFF)
-option(LLAMA_K_QUANTS                        "llama: use k-quants"                              ON)
 option(LLAMA_QKK_64                          "llama: use super-block size of 64 for k-quants"   OFF)
 
 option(LLAMA_BUILD_TESTS                "llama: build tests"    ${LLAMA_STANDALONE})
 option(LLAMA_BUILD_EXAMPLES             "llama: build examples" ${LLAMA_STANDALONE})
 option(LLAMA_BUILD_SERVER               "llama: build server example"                           ON)
 
-#
-# Build info header
-#
-
-# Generate initial build-info.h
-include(${CMAKE_CURRENT_SOURCE_DIR}/scripts/build-info.cmake)
-
-if(EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/.git")
-    set(GIT_DIR "${CMAKE_CURRENT_SOURCE_DIR}/.git")
-
-    # Is git submodule
-    if(NOT IS_DIRECTORY "${GIT_DIR}")
-        file(READ ${GIT_DIR} REAL_GIT_DIR_LINK)
-        string(REGEX REPLACE "gitdir: (.*)\n$" "\\1" REAL_GIT_DIR ${REAL_GIT_DIR_LINK})
-        set(GIT_DIR "${CMAKE_CURRENT_SOURCE_DIR}/${REAL_GIT_DIR}")
-    endif()
-
-    # Add a custom target for build-info.h
-    add_custom_target(BUILD_INFO ALL DEPENDS "${CMAKE_CURRENT_SOURCE_DIR}/build-info.h")
-
-    # Add a custom command to rebuild build-info.h when .git/index changes
-    add_custom_command(
-        OUTPUT "${CMAKE_CURRENT_SOURCE_DIR}/build-info.h"
-        COMMENT "Generating build details from Git"
-        COMMAND ${CMAKE_COMMAND} -DMSVC=${MSVC} -DCMAKE_C_COMPILER_VERSION=${CMAKE_C_COMPILER_VERSION} -DCMAKE_C_COMPILER_ID=${CMAKE_C_COMPILER_ID} -DCMAKE_VS_PLATFORM_NAME=${CMAKE_VS_PLATFORM_NAME} -DCMAKE_C_COMPILER=${CMAKE_C_COMPILER} -P "${CMAKE_CURRENT_SOURCE_DIR}/scripts/build-info.cmake"
-        WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}
-        DEPENDS "${GIT_DIR}/index"
-        VERBATIM
-    )
-else()
-    message(WARNING "Git repository not found; to enable automatic generation of build info, make sure Git is installed and the project is a Git repository.")
-endif()
-
 #
 # Compile flags
 #
@@ -277,13 +244,8 @@ if (LLAMA_BLAS)
     endif()
 endif()
 
-if (LLAMA_K_QUANTS)
+if (LLAMA_QKK_64)
-    set(GGML_HEADERS_EXTRA k_quants.h)
+    add_compile_definitions(GGML_QKK_64)
-    set(GGML_SOURCES_EXTRA k_quants.c)
-    add_compile_definitions(GGML_USE_K_QUANTS)
-    if (LLAMA_QKK_64)
-        add_compile_definitions(GGML_QKK_64)
-    endif()
 endif()
 
 if (LLAMA_CUBLAS)
@@ -305,6 +267,9 @@ if (LLAMA_CUBLAS)
         if (LLAMA_CUDA_FORCE_DMMV)
             add_compile_definitions(GGML_CUDA_FORCE_DMMV)
         endif()
+        if (LLAMA_CUDA_FORCE_MMQ)
+            add_compile_definitions(GGML_CUDA_FORCE_MMQ)
+        endif()
         add_compile_definitions(GGML_CUDA_DMMV_X=${LLAMA_CUDA_DMMV_X})
         add_compile_definitions(GGML_CUDA_MMV_Y=${LLAMA_CUDA_MMV_Y})
         if (DEFINED LLAMA_CUDA_DMMV_Y)
@@ -331,6 +296,7 @@ if (LLAMA_CUBLAS)
             set(CMAKE_CUDA_ARCHITECTURES "60;61;70") # needed for f16 CUDA intrinsics
         else()
             set(CMAKE_CUDA_ARCHITECTURES "52;61;70") # lowest CUDA 12 standard + lowest for integer intrinsics
+            #set(CMAKE_CUDA_ARCHITECTURES "") # use this to compile much faster, but only F16 models work
         endif()
     endif()
     message(STATUS "Using CUDA architectures: ${CMAKE_CUDA_ARCHITECTURES}")
@@ -404,6 +370,9 @@ if (LLAMA_HIPBLAS)
         if (LLAMA_CUDA_FORCE_DMMV)
             target_compile_definitions(ggml-rocm PRIVATE GGML_CUDA_FORCE_DMMV)
         endif()
+        if (LLAMA_CUDA_FORCE_MMQ)
+            target_compile_definitions(ggml-rocm PRIVATE GGML_CUDA_FORCE_MMQ)
+        endif()
         target_compile_definitions(ggml-rocm PRIVATE GGML_CUDA_DMMV_X=${LLAMA_CUDA_DMMV_X})
         target_compile_definitions(ggml-rocm PRIVATE GGML_CUDA_MMV_Y=${LLAMA_CUDA_MMV_Y})
         target_compile_definitions(ggml-rocm PRIVATE K_QUANTS_PER_ITERATION=${LLAMA_CUDA_KQUANTS_ITER})
@@ -665,6 +634,8 @@ add_library(ggml OBJECT
             ggml-alloc.h
             ggml-backend.c
             ggml-backend.h
+            ggml-quants.c
+            ggml-quants.h
             ${GGML_SOURCES_CUDA} ${GGML_HEADERS_CUDA}
             ${GGML_SOURCES_OPENCL} ${GGML_HEADERS_OPENCL}
             ${GGML_SOURCES_METAL} ${GGML_HEADERS_METAL}
@@ -810,7 +781,6 @@ endif()
 #
 
 add_subdirectory(common)
-add_subdirectory(llava)
 
 if (LLAMA_BUILD_TESTS AND NOT CMAKE_JS_VERSION)
     include(CTest)

Makefile

@@ -1,7 +1,7 @@
 # Define the default target now so that it is always the first target
 BUILD_TARGETS = \
 	main quantize quantize-stats perplexity embedding vdot q8dot train-text-from-scratch convert-llama2c-to-ggml \
-	simple batched batched-bench save-load-state server embd-input-test gguf llama-bench llava baby-llama beam-search  \
+	simple batched batched-bench save-load-state server gguf llama-bench llava baby-llama beam-search  \
 	speculative infill benchmark-matmult parallel finetune export-lora tests/test-c.o
 
 # Binaries only useful for tests
@@ -342,13 +342,9 @@ else
 	MK_CXXFLAGS += -march=rv64gcv -mabi=lp64d
 endif
 
-ifndef LLAMA_NO_K_QUANTS
-	MK_CPPFLAGS += -DGGML_USE_K_QUANTS
-	OBJS     += k_quants.o
 ifdef LLAMA_QKK_64
 	MK_CPPFLAGS += -DGGML_QKK_64
 endif
-endif
 
 ifndef LLAMA_NO_ACCELERATE
 	# Mac OS - include Accelerate framework.
@@ -365,7 +361,7 @@ ifdef LLAMA_MPI
 	MK_CPPFLAGS += -DGGML_USE_MPI
 	MK_CFLAGS   += -Wno-cast-qual
 	MK_CXXFLAGS += -Wno-cast-qual
-	OBJS     += ggml-mpi.o
+	OBJS        += ggml-mpi.o
 endif # LLAMA_MPI
 
 ifdef LLAMA_OPENBLAS
@@ -382,7 +378,7 @@ endif # LLAMA_BLIS
 ifdef LLAMA_CUBLAS
 	MK_CPPFLAGS  += -DGGML_USE_CUBLAS -I/usr/local/cuda/include -I/opt/cuda/include -I$(CUDA_PATH)/targets/x86_64-linux/include
 	MK_LDFLAGS   += -lcublas -lculibos -lcudart -lcublasLt -lpthread -ldl -lrt -L/usr/local/cuda/lib64 -L/opt/cuda/lib64 -L$(CUDA_PATH)/targets/x86_64-linux/lib
-	OBJS      += ggml-cuda.o
+	OBJS         += ggml-cuda.o
 	NVCCFLAGS = --forward-unknown-to-host-compiler -use_fast_math
 ifdef LLAMA_CUDA_NVCC
 	NVCC = $(LLAMA_CUDA_NVCC)
@@ -397,6 +393,9 @@ endif # CUDA_DOCKER_ARCH
 ifdef LLAMA_CUDA_FORCE_DMMV
 	NVCCFLAGS += -DGGML_CUDA_FORCE_DMMV
 endif # LLAMA_CUDA_FORCE_DMMV
+ifdef LLAMA_CUDA_FORCE_MMQ
+	NVCCFLAGS += -DGGML_CUDA_FORCE_MMQ
+endif # LLAMA_CUDA_FORCE_MMQ
 ifdef LLAMA_CUDA_DMMV_X
 	NVCCFLAGS += -DGGML_CUDA_DMMV_X=$(LLAMA_CUDA_DMMV_X)
 else
@@ -494,11 +493,6 @@ ggml-mpi.o: ggml-mpi.c ggml-mpi.h
 	$(CC) $(CFLAGS) -c $< -o $@
 endif # LLAMA_MPI
 
-ifndef LLAMA_NO_K_QUANTS
-k_quants.o: k_quants.c k_quants.h
-	$(CC) $(CFLAGS) -c $< -o $@
-endif # LLAMA_NO_K_QUANTS
-
 # combine build flags with cmdline overrides
 override CFLAGS        := $(MK_CPPFLAGS) $(CPPFLAGS) $(MK_CFLAGS) $(CFLAGS)
 override CXXFLAGS      := $(MK_CPPFLAGS) $(CPPFLAGS) $(MK_CXXFLAGS) $(CXXFLAGS)
@@ -539,13 +533,16 @@ ggml-alloc.o: ggml-alloc.c ggml.h ggml-alloc.h
 ggml-backend.o: ggml-backend.c ggml.h ggml-backend.h
 	$(CC)  $(CFLAGS)   -c $< -o $@
 
-OBJS += ggml-alloc.o ggml-backend.o
+ggml-quants.o: ggml-quants.c ggml.h ggml-quants.h
+	$(CC) $(CFLAGS)    -c $< -o $@
+
+OBJS += ggml-alloc.o ggml-backend.o ggml-quants.o
 
 llama.o: llama.cpp ggml.h ggml-alloc.h ggml-backend.h ggml-cuda.h ggml-metal.h llama.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@
 
-COMMON_H_DEPS = common/common.h common/sampling.h build-info.h common/log.h
+COMMON_H_DEPS = common/common.h common/sampling.h common/log.h
-COMMON_DEPS   = $(COMMON_H_DEPS) common.o sampling.o
+COMMON_DEPS   = common.o sampling.o grammar-parser.o build-info.o
 
 common.o: common/common.cpp $(COMMON_H_DEPS)
 	$(CXX) $(CXXFLAGS) -c $< -o $@
@@ -566,54 +563,47 @@ libllama.so: llama.o ggml.o $(OBJS)
 	$(CXX) $(CXXFLAGS) -shared -fPIC -o $@ $^ $(LDFLAGS)
 
 clean:
-	rm -vrf *.o tests/*.o *.so *.dll benchmark-matmult build-info.h *.dot $(COV_TARGETS) $(BUILD_TARGETS) $(TEST_TARGETS)
+	rm -vrf *.o tests/*.o *.so *.dll benchmark-matmult common/build-info.cpp *.dot $(COV_TARGETS) $(BUILD_TARGETS) $(TEST_TARGETS)
 
 #
 # Examples
 #
 
-main: examples/main/main.cpp                                  build-info.h ggml.o llama.o $(COMMON_DEPS) console.o grammar-parser.o $(OBJS)
+main: examples/main/main.cpp                                  ggml.o llama.o $(COMMON_DEPS) console.o grammar-parser.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 	@echo
 	@echo '====  Run ./main -h for help.  ===='
 	@echo
 
-infill: examples/infill/infill.cpp                            build-info.h ggml.o llama.o $(COMMON_DEPS) console.o grammar-parser.o $(OBJS)
+infill: examples/infill/infill.cpp                            ggml.o llama.o $(COMMON_DEPS) console.o grammar-parser.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-simple: examples/simple/simple.cpp                            build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+simple: examples/simple/simple.cpp                            ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-batched: examples/batched/batched.cpp                         build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+batched: examples/batched/batched.cpp                         ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-batched-bench: examples/batched-bench/batched-bench.cpp       build-info.h ggml.o llama.o common.o $(OBJS)
+batched-bench: examples/batched-bench/batched-bench.cpp       build-info.o ggml.o llama.o common.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-quantize: examples/quantize/quantize.cpp                      build-info.h ggml.o llama.o $(OBJS)
+quantize: examples/quantize/quantize.cpp                      build-info.o ggml.o llama.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-quantize-stats: examples/quantize-stats/quantize-stats.cpp    build-info.h ggml.o llama.o $(OBJS)
+quantize-stats: examples/quantize-stats/quantize-stats.cpp    build-info.o ggml.o llama.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-perplexity: examples/perplexity/perplexity.cpp                build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+perplexity: examples/perplexity/perplexity.cpp                ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-embedding: examples/embedding/embedding.cpp                   build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+embedding: examples/embedding/embedding.cpp                   ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-save-load-state: examples/save-load-state/save-load-state.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+save-load-state: examples/save-load-state/save-load-state.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-server: examples/server/server.cpp examples/server/httplib.h examples/server/json.hpp examples/server/index.html.hpp examples/server/index.js.hpp examples/server/completion.js.hpp build-info.h ggml.o llama.o $(COMMON_DEPS) grammar-parser.o $(OBJS)
+server: examples/server/server.cpp examples/server/httplib.h examples/server/json.hpp examples/server/index.html.hpp examples/server/index.js.hpp examples/server/completion.js.hpp examples/llava/clip.cpp examples/llava/clip.h common/stb_image.h ggml.o llama.o $(COMMON_DEPS) grammar-parser.o $(OBJS)
-	$(CXX) $(CXXFLAGS) -Iexamples/server $(filter-out %.h,$(filter-out %.hpp,$^)) -o $@ $(LDFLAGS) $(LWINSOCK2)
+	$(CXX) $(CXXFLAGS) -Iexamples/server $(filter-out %.h,$(filter-out %.hpp,$^)) -o $@ $(LDFLAGS) $(LWINSOCK2) -Wno-cast-qual
-
-$(LIB_PRE)embdinput$(DSO_EXT): examples/embd-input/embd-input.h examples/embd-input/embd-input-lib.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS)
-	$(CXX) --shared $(CXXFLAGS) $(filter-out %.h,$(filter-out %.hpp,$^)) -o $@ $(LDFLAGS)
-
-
-embd-input-test: $(LIB_PRE)embdinput$(DSO_EXT) examples/embd-input/embd-input-test.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS)
-	$(CXX) $(CXXFLAGS) $(filter-out %$(DSO_EXT),$(filter-out %.h,$(filter-out %.hpp,$^))) -o $@ $(LDFLAGS) -L. -lembdinput
 
 gguf: examples/gguf/gguf.cpp ggml.o llama.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
@@ -624,7 +614,7 @@ train-text-from-scratch: examples/train-text-from-scratch/train-text-from-scratc
 convert-llama2c-to-ggml: examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp ggml.o llama.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-llama-bench: examples/llama-bench/llama-bench.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+llama-bench: examples/llama-bench/llama-bench.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
 llava: examples/llava/llava.cpp examples/llava/llava-utils.h examples/llava/clip.cpp examples/llava/clip.h common/stb_image.h ggml.o llama.o $(COMMON_DEPS) $(OBJS)
@@ -633,19 +623,19 @@ llava: examples/llava/llava.cpp examples/llava/llava-utils.h examples/llava/clip
 baby-llama: examples/baby-llama/baby-llama.cpp ggml.o llama.o $(COMMON_DEPS) train.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-beam-search: examples/beam-search/beam-search.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+beam-search: examples/beam-search/beam-search.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-finetune: examples/finetune/finetune.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) train.o $(OBJS)
+finetune: examples/finetune/finetune.cpp ggml.o llama.o $(COMMON_DEPS) train.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-export-lora: examples/export-lora/export-lora.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+export-lora: examples/export-lora/export-lora.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-speculative: examples/speculative/speculative.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) grammar-parser.o $(OBJS)
+speculative: examples/speculative/speculative.cpp ggml.o llama.o $(COMMON_DEPS) grammar-parser.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-parallel: examples/parallel/parallel.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+parallel: examples/parallel/parallel.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
 ifdef LLAMA_METAL
@@ -658,7 +648,7 @@ swift: examples/batched.swift
 	(cd examples/batched.swift; make build)
 endif
 
-build-info.h: $(wildcard .git/index) scripts/build-info.sh
+common/build-info.cpp: $(wildcard .git/index) scripts/build-info.sh
 	@sh scripts/build-info.sh $(CC) > $@.tmp
 	@if ! cmp -s $@.tmp $@; then \
 		mv $@.tmp $@; \
@@ -666,13 +656,16 @@ build-info.h: $(wildcard .git/index) scripts/build-info.sh
 		rm $@.tmp; \
 	fi
 
+build-info.o: common/build-info.cpp
+	$(CXX) $(CXXFLAGS) -c $(filter-out %.h,$^) -o $@
+
 #
 # Tests
 #
 
 tests: $(TEST_TARGETS)
 
-benchmark-matmult: examples/benchmark/benchmark-matmult.cpp build-info.h ggml.o $(OBJS)
+benchmark-matmult: examples/benchmark/benchmark-matmult.cpp build-info.o ggml.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
 run-benchmark-matmult: benchmark-matmult
@@ -686,40 +679,40 @@ vdot: pocs/vdot/vdot.cpp ggml.o $(OBJS)
 q8dot: pocs/vdot/q8dot.cpp ggml.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS)
 
-tests/test-llama-grammar: tests/test-llama-grammar.cpp build-info.h ggml.o $(COMMON_DEPS) grammar-parser.o $(OBJS)
+tests/test-llama-grammar: tests/test-llama-grammar.cpp ggml.o $(COMMON_DEPS) grammar-parser.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-tests/test-grammar-parser: tests/test-grammar-parser.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) grammar-parser.o $(OBJS)
+tests/test-grammar-parser: tests/test-grammar-parser.cpp ggml.o llama.o $(COMMON_DEPS) grammar-parser.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-tests/test-double-float: tests/test-double-float.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+tests/test-double-float: tests/test-double-float.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-tests/test-grad0: tests/test-grad0.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+tests/test-grad0: tests/test-grad0.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-tests/test-opt: tests/test-opt.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+tests/test-opt: tests/test-opt.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-tests/test-quantize-fns: tests/test-quantize-fns.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+tests/test-quantize-fns: tests/test-quantize-fns.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-tests/test-quantize-perf: tests/test-quantize-perf.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+tests/test-quantize-perf: tests/test-quantize-perf.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-tests/test-sampling: tests/test-sampling.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+tests/test-sampling: tests/test-sampling.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-tests/test-tokenizer-0-falcon: tests/test-tokenizer-0-falcon.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+tests/test-tokenizer-0-falcon: tests/test-tokenizer-0-falcon.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-tests/test-tokenizer-0-llama: tests/test-tokenizer-0-llama.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+tests/test-tokenizer-0-llama: tests/test-tokenizer-0-llama.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-tests/test-tokenizer-1-bpe: tests/test-tokenizer-1-bpe.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+tests/test-tokenizer-1-bpe: tests/test-tokenizer-1-bpe.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-tests/test-tokenizer-1-llama: tests/test-tokenizer-1-llama.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+tests/test-tokenizer-1-llama: tests/test-tokenizer-1-llama.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
 tests/test-c.o: tests/test-c.c llama.h

Package.swift

@@ -42,13 +42,12 @@ let package = Package(
                 "llama.cpp",
                 "ggml-alloc.c",
                 "ggml-backend.c",
-                "k_quants.c",
+                "ggml-quants.c",
             ] + additionalSources,
             resources: resources,
             publicHeadersPath: "spm-headers",
             cSettings: [
                 .unsafeFlags(["-Wno-shorten-64-to-32", "-O3", "-DNDEBUG"]),
-                .define("GGML_USE_K_QUANTS"),
                 .define("GGML_USE_ACCELERATE")
                 // NOTE: NEW_LAPACK will required iOS version 16.4+
                 // We should consider add this in the future when we drop support for iOS 14

README.md

@@ -11,12 +11,8 @@ Inference of [LLaMA](https://arxiv.org/abs/2302.13971) model in pure C/C++
 
 ### Hot topics
 
-- ‼️ Breaking change: `rope_freq_base` and `rope_freq_scale` must be set to zero to use the model default values: [#3401](https://github.com/ggerganov/llama.cpp/pull/3401)
+- LLaVA support: https://github.com/ggerganov/llama.cpp/pull/3436
-- Parallel decoding + continuous batching support added: [#3228](https://github.com/ggerganov/llama.cpp/pull/3228) \
+- ‼️ BPE tokenizer update: existing Falcon and Starcoder `.gguf` models will need to be reconverted: [#3252](https://github.com/ggerganov/llama.cpp/pull/3252)
-  **Devs should become familiar with the new API**
-- Local Falcon 180B inference on Mac Studio
-
-  https://github.com/ggerganov/llama.cpp/assets/1991296/98abd4e8-7077-464c-ae89-aebabca7757e
 
 ----
 
@@ -89,21 +85,23 @@ as the main playground for developing new features for the [ggml](https://github
 - [X] [Vicuna](https://github.com/ggerganov/llama.cpp/discussions/643#discussioncomment-5533894)
 - [X] [Koala](https://bair.berkeley.edu/blog/2023/04/03/koala/)
 - [X] [OpenBuddy 🐶 (Multilingual)](https://github.com/OpenBuddy/OpenBuddy)
-- [X] [Pygmalion 7B / Metharme 7B](#using-pygmalion-7b--metharme-7b)
+- [X] [Pygmalion/Metharme](#using-pygmalion-7b--metharme-7b)
 - [X] [WizardLM](https://github.com/nlpxucan/WizardLM)
-- [X] [Baichuan-7B](https://huggingface.co/baichuan-inc/baichuan-7B) and its derivations (such as [baichuan-7b-sft](https://huggingface.co/hiyouga/baichuan-7b-sft))
+- [X] [Baichuan 1 & 2](https://huggingface.co/models?search=baichuan-inc/Baichuan) + [derivations](https://huggingface.co/hiyouga/baichuan-7b-sft)
-- [X] [Aquila-7B](https://huggingface.co/BAAI/Aquila-7B) / [AquilaChat-7B](https://huggingface.co/BAAI/AquilaChat-7B)
+- [X] [Aquila 1 & 2](https://huggingface.co/models?search=BAAI/Aquila)
 - [X] [Starcoder models](https://github.com/ggerganov/llama.cpp/pull/3187)
 - [X] [Mistral AI v0.1](https://huggingface.co/mistralai/Mistral-7B-v0.1)
 - [X] [Refact](https://huggingface.co/smallcloudai/Refact-1_6B-fim)
-- [X] [Bloom](https://github.com/ggerganov/llama.cpp/pull/3553)
+- [X] [Persimmon 8B](https://github.com/ggerganov/llama.cpp/pull/3410)
 - [X] [MPT](https://github.com/ggerganov/llama.cpp/pull/3417)
+- [X] [Bloom](https://github.com/ggerganov/llama.cpp/pull/3553)
+
 
 **Bindings:**
 
 - Python: [abetlen/llama-cpp-python](https://github.com/abetlen/llama-cpp-python)
 - Go: [go-skynet/go-llama.cpp](https://github.com/go-skynet/go-llama.cpp)
-- Node.js: [withcatai/node-llama-cpp](https://github.com/withcatai/node-llama-cpp), [hlhr202/llama-node](https://github.com/hlhr202/llama-node)
+- Node.js: [withcatai/node-llama-cpp](https://github.com/withcatai/node-llama-cpp)
 - Ruby: [yoshoku/llama_cpp.rb](https://github.com/yoshoku/llama_cpp.rb)
 - Rust: [mdrokz/rust-llama.cpp](https://github.com/mdrokz/rust-llama.cpp)
 - C#/.NET: [SciSharp/LLamaSharp](https://github.com/SciSharp/LLamaSharp)
@@ -206,7 +204,7 @@ https://user-images.githubusercontent.com/1991296/224442907-7693d4be-acaa-4e01-8
 
 ## Usage
 
-Here are the steps for the LLaMA-7B model.
+Here are the end-to-end binary build and model conversion steps for the LLaMA-7B model.
 
 ### Get the Code
 
@@ -573,6 +571,18 @@ python3 convert.py models/7B/
 
 When running the larger models, make sure you have enough disk space to store all the intermediate files.
 
+### Running on Windows with prebuilt binaries
+
+You will find prebuilt Windows binaries on the release page.
+
+Simply download and extract the latest zip package of choice: (e.g. `llama-b1380-bin-win-avx2-x64.zip`)
+
+From the unzipped folder, open a terminal/cmd window here and place a pre-converted `.gguf` model file. Test out the main example like so:
+
+```
+.\main -m llama-2-7b.Q4_0.gguf -n 128
+```
+
 ### Memory/Disk Requirements
 
 As the models are currently fully loaded into memory, you will need adequate disk space to save them and sufficient RAM to load them. At the moment, memory and disk requirements are the same.
@@ -952,7 +962,6 @@ docker run --gpus all -v /path/to/models:/models local/llama.cpp:light-cuda -m /
 
 - [main](./examples/main/README.md)
 - [server](./examples/server/README.md)
-- [embd-input](./examples/embd-input/README.md)
 - [jeopardy](./examples/jeopardy/README.md)
 - [BLIS](./docs/BLIS.md)
 - [Performance troubleshooting](./docs/token_generation_performance_tips.md)

build-info.h

@@ -0,0 +1,9 @@
+#ifndef BUILD_INFO_H
+#define BUILD_INFO_H
+
+#define BUILD_NUMBER 1418
+#define BUILD_COMMIT "9d02956"
+#define BUILD_COMPILER "clang version 16.0.0"
+#define BUILD_TARGET "x86_64-unknown-linux-gnu"
+
+#endif // BUILD_INFO_H

build.zig

@@ -10,7 +10,6 @@ const Maker = struct {
     builder: *std.build.Builder,
     target: CrossTarget,
     optimize: Mode,
-    config_header: *ConfigHeader,
     enable_lto: bool,
 
     include_dirs: ArrayList([]const u8),
@@ -41,26 +40,24 @@ const Maker = struct {
         const commit_hash = try std.ChildProcess.exec(
             .{ .allocator = builder.allocator, .argv = &.{ "git", "rev-parse", "HEAD" } },
         );
-        const config_header = builder.addConfigHeader(
+        try std.fs.cwd().writeFile("common/build-info.cpp", builder.fmt(
-            .{ .style = .blank, .include_path = "build-info.h" },
+            \\int LLAMA_BUILD_NUMBER = {};
-            .{
+            \\char const *LLAMA_COMMIT = "{s}";
-                .BUILD_NUMBER = 0,
+            \\char const *LLAMA_COMPILER = "Zig {s}";
-                .BUILD_COMMIT = commit_hash.stdout[0 .. commit_hash.stdout.len - 1], // omit newline
+            \\char const *LLAMA_BUILD_TARGET = "{s}";
-                .BUILD_COMPILER = builder.fmt("Zig {s}", .{zig_version}),
+            \\
-                .BUILD_TARGET = try target.allocDescription(builder.allocator),
+        , .{ 0, commit_hash.stdout[0 .. commit_hash.stdout.len - 1], zig_version, try target.allocDescription(builder.allocator) }));
-            },
-        );
         var m = Maker{
             .builder = builder,
             .target = target,
             .optimize = builder.standardOptimizeOption(.{}),
-            .config_header = config_header,
             .enable_lto = false,
             .include_dirs = ArrayList([]const u8).init(builder.allocator),
             .cflags = ArrayList([]const u8).init(builder.allocator),
             .cxxflags = ArrayList([]const u8).init(builder.allocator),
             .objs = ArrayList(*Compile).init(builder.allocator),
         };
+
         try m.addCFlag("-std=c11");
         try m.addCxxFlag("-std=c++11");
         try m.addProjectInclude(&.{});
@@ -72,7 +69,7 @@ const Maker = struct {
         const o = m.builder.addObject(.{ .name = name, .target = m.target, .optimize = m.optimize });
         if (o.target.getAbi() != .msvc)
             o.defineCMacro("_GNU_SOURCE", null);
-        o.addConfigHeader(m.config_header);
+
         if (std.mem.endsWith(u8, src, ".c")) {
             o.addCSourceFiles(&.{src}, m.cflags.items);
             o.linkLibC();
@@ -85,7 +82,6 @@ const Maker = struct {
                 o.linkLibCpp();
             }
         }
-        o.addConfigHeader(m.config_header);
         for (m.include_dirs.items) |i| o.addIncludePath(.{ .path = i });
         o.want_lto = m.enable_lto;
         return o;
@@ -105,7 +101,6 @@ const Maker = struct {
             // linkLibCpp already add (libc++ + libunwind + libc)
             e.linkLibCpp();
         }
-        e.addConfigHeader(m.config_header);
         m.builder.installArtifact(e);
         e.want_lto = m.enable_lto;
         return e;
@@ -116,30 +111,27 @@ pub fn build(b: *std.build.Builder) !void {
     var make = try Maker.init(b);
     make.enable_lto = b.option(bool, "lto", "Enable LTO optimization, (default: false)") orelse false;
 
-    if (b.option(bool, "k-quants", "Enable K-quants, (default: true)") orelse true) {
-        try make.addFlag("-DGGML_USE_K_QUANTS");
-        const k_quants = make.obj("k_quants", "k_quants.c");
-        try make.objs.append(k_quants);
-    }
-
     const ggml = make.obj("ggml", "ggml.c");
     const ggml_alloc = make.obj("ggml-alloc", "ggml-alloc.c");
     const ggml_backend = make.obj("ggml-backend", "ggml-backend.c");
+    const ggml_quants = make.obj("ggml-quants", "ggml-quants.c");
     const llama = make.obj("llama", "llama.cpp");
+    const buildinfo = make.obj("common", "common/build-info.cpp");
     const common = make.obj("common", "common/common.cpp");
     const console = make.obj("console", "common/console.cpp");
     const sampling = make.obj("sampling", "common/sampling.cpp");
     const grammar_parser = make.obj("grammar-parser", "common/grammar-parser.cpp");
     const train = make.obj("train", "common/train.cpp");
+    const clip = make.obj("clip", "examples/llava/clip.cpp");
 
-    _ = make.exe("main", "examples/main/main.cpp", &.{ ggml, ggml_alloc, ggml_backend, llama, common, sampling, console, grammar_parser });
+    _ = make.exe("main", "examples/main/main.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common, buildinfo, sampling, console, grammar_parser });
-    _ = make.exe("quantize", "examples/quantize/quantize.cpp", &.{ ggml, ggml_alloc, ggml_backend, llama, common });
+    _ = make.exe("quantize", "examples/quantize/quantize.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common, buildinfo });
-    _ = make.exe("perplexity", "examples/perplexity/perplexity.cpp", &.{ ggml, ggml_alloc, ggml_backend, llama, common });
+    _ = make.exe("perplexity", "examples/perplexity/perplexity.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common, buildinfo });
-    _ = make.exe("embedding", "examples/embedding/embedding.cpp", &.{ ggml, ggml_alloc, ggml_backend, llama, common });
+    _ = make.exe("embedding", "examples/embedding/embedding.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common, buildinfo });
-    _ = make.exe("finetune", "examples/finetune/finetune.cpp", &.{ ggml, ggml_alloc, ggml_backend, llama, common, train });
+    _ = make.exe("finetune", "examples/finetune/finetune.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common, buildinfo, train });
-    _ = make.exe("train-text-from-scratch", "examples/train-text-from-scratch/train-text-from-scratch.cpp", &.{ ggml, ggml_alloc, ggml_backend, llama, common, train });
+    _ = make.exe("train-text-from-scratch", "examples/train-text-from-scratch/train-text-from-scratch.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common, buildinfo, train });
 
-    const server = make.exe("server", "examples/server/server.cpp", &.{ ggml, ggml_alloc, ggml_backend, llama, common, sampling, grammar_parser });
+    const server = make.exe("server", "examples/server/server.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common, buildinfo, sampling, grammar_parser, clip });
     if (server.target.isWindows()) {
         server.linkSystemLibrary("ws2_32");
     }

ci/run.sh

@@ -208,6 +208,8 @@ function gg_run_open_llama_3b_v2 {
     (time ./bin/perplexity --model ${model_q5_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
     (time ./bin/perplexity --model ${model_q6_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
 
+    (time ./bin/save-load-state --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+
     function check_ppl {
         qnt="$1"
         ppl=$(echo "$2" | grep -oE "[0-9]+\.[0-9]+" | tail -n 1)
@@ -296,6 +298,7 @@ function gg_sum_open_llama_3b_v2 {
     gg_printf '- q4_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_k.log)"
     gg_printf '- q5_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q5_k.log)"
     gg_printf '- q6_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q6_k.log)"
+    gg_printf '- save-load-state: \n```\n%s\n```\n' "$(cat $OUT/${ci}-save-load-state.log)"
     gg_printf '- shakespeare (f16):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-f16.log)"
     gg_printf '- shakespeare (f16 lora):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-lora-f16.log)"
     gg_printf '- shakespeare (q8_0):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-q8_0.log)"
@@ -382,6 +385,8 @@ function gg_run_open_llama_7b_v2 {
     (time ./bin/perplexity --model ${model_q5_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
     (time ./bin/perplexity --model ${model_q6_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
 
+    (time ./bin/save-load-state --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+
     function check_ppl {
         qnt="$1"
         ppl=$(echo "$2" | grep -oE "[0-9]+\.[0-9]+" | tail -n 1)
@@ -470,6 +475,7 @@ function gg_sum_open_llama_7b_v2 {
     gg_printf '- q4_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_k.log)"
     gg_printf '- q5_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q5_k.log)"
     gg_printf '- q6_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q6_k.log)"
+    gg_printf '- save-load-state: \n```\n%s\n```\n' "$(cat $OUT/${ci}-save-load-state.log)"
     gg_printf '- shakespeare (f16):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-f16.log)"
     gg_printf '- shakespeare (f16 lora):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-lora-f16.log)"
     #gg_printf '- shakespeare (q8_0):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-q8_0.log)"

common/CMakeLists.txt

@@ -1,5 +1,43 @@
 # common
 
+
+# Build info header
+#
+
+if(EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/../.git")
+    set(GIT_DIR "${CMAKE_CURRENT_SOURCE_DIR}/../.git")
+
+    # Is git submodule
+    if(NOT IS_DIRECTORY "${GIT_DIR}")
+        file(READ ${GIT_DIR} REAL_GIT_DIR_LINK)
+        string(REGEX REPLACE "gitdir: (.*)\n$" "\\1" REAL_GIT_DIR ${REAL_GIT_DIR_LINK})
+        set(GIT_DIR "${CMAKE_CURRENT_SOURCE_DIR}/${REAL_GIT_DIR}")
+    endif()
+
+    set(GIT_INDEX "${GIT_DIR}/index")
+else()
+    message(WARNING "Git repository not found; to enable automatic generation of build info, make sure Git is installed and the project is a Git repository.")
+    set(GIT_INDEX "")
+endif()
+
+# Add a custom command to rebuild build-info.cpp when .git/index changes
+add_custom_command(
+    OUTPUT "${CMAKE_CURRENT_SOURCE_DIR}/build-info.cpp"
+    COMMENT "Generating build details from Git"
+    COMMAND ${CMAKE_COMMAND} -DMSVC=${MSVC} -DCMAKE_C_COMPILER_VERSION=${CMAKE_C_COMPILER_VERSION}
+            -DCMAKE_C_COMPILER_ID=${CMAKE_C_COMPILER_ID} -DCMAKE_VS_PLATFORM_NAME=${CMAKE_VS_PLATFORM_NAME}
+            -DCMAKE_C_COMPILER=${CMAKE_C_COMPILER} -P "${CMAKE_CURRENT_SOURCE_DIR}/../scripts/build-info.cmake"
+    WORKING_DIRECTORY "${CMAKE_CURRENT_SOURCE_DIR}/.."
+    DEPENDS "${CMAKE_CURRENT_SOURCE_DIR}/build-info.cpp.in" ${GIT_INDEX}
+    VERBATIM
+)
+set(TARGET build_info)
+add_library(${TARGET} OBJECT build-info.cpp)
+if (BUILD_SHARED_LIBS)
+    set_target_properties(${TARGET} PROPERTIES POSITION_INDEPENDENT_CODE ON)
+endif()
+
+
 set(TARGET common)
 
 add_library(${TARGET} OBJECT
@@ -22,4 +60,4 @@ endif()
 
 target_include_directories(${TARGET} PUBLIC .)
 target_compile_features(${TARGET} PUBLIC cxx_std_11)
-target_link_libraries(${TARGET} PRIVATE llama)
+target_link_libraries(${TARGET} PRIVATE llama build_info)

common/build-info.cpp.in

@@ -0,0 +1,4 @@
+int LLAMA_BUILD_NUMBER = @BUILD_NUMBER@;
+char const *LLAMA_COMMIT = "@BUILD_COMMIT@";
+char const *LLAMA_COMPILER = "@BUILD_COMPILER@";
+char const *LLAMA_BUILD_TARGET = "@BUILD_TARGET@";

common/common.cpp

@@ -1,5 +1,4 @@
 #include "common.h"
-#include "build-info.h"
 #include "llama.h"
 
 #include <algorithm>
@@ -103,11 +102,26 @@ void process_escapes(std::string& input) {
 }
 
 bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
+    bool result = true;
+    try {
+        if (!gpt_params_parse_ex(argc, argv, params)) {
+            gpt_print_usage(argc, argv, gpt_params());
+            exit(0);
+        }
+    }
+    catch (const std::invalid_argument & ex) {
+        fprintf(stderr, "%s\n", ex.what());
+        gpt_print_usage(argc, argv, gpt_params());
+        exit(1);
+    }
+    return result;
+}
+
+bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
     bool invalid_param = false;
     std::string arg;
-    gpt_params default_params;
     const std::string arg_prefix = "--";
-    llama_sampling_params & sparams = params.sampling_params;
+    llama_sampling_params & sparams = params.sparams;
 
     for (int i = 1; i < argc; i++) {
         arg = argv[i];
@@ -204,12 +218,52 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
                 break;
             }
             params.rope_freq_scale = std::stof(argv[i]);
+        } else if (arg == "--rope-scaling") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            std::string value(argv[i]);
+            /**/ if (value == "none")   { params.rope_scaling_type = LLAMA_ROPE_SCALING_NONE; }
+            else if (value == "linear") { params.rope_scaling_type = LLAMA_ROPE_SCALING_LINEAR; }
+            else if (value == "yarn")   { params.rope_scaling_type = LLAMA_ROPE_SCALING_YARN; }
+            else { invalid_param = true; break; }
         } 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 == "--yarn-orig-ctx") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.yarn_orig_ctx = std::stoi(argv[i]);
+        } else if (arg == "--yarn-ext-factor") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.yarn_ext_factor = std::stof(argv[i]);
+        } else if (arg == "--yarn-attn-factor") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.yarn_attn_factor = std::stof(argv[i]);
+        } else if (arg == "--yarn-beta-fast") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.yarn_beta_fast = std::stof(argv[i]);
+        } else if (arg == "--yarn-beta-slow") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.yarn_beta_slow = std::stof(argv[i]);
         } else if (arg == "--memory-f32") {
             params.memory_f16 = false;
         } else if (arg == "--top-p") {
@@ -218,12 +272,19 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
                 break;
             }
             sparams.top_p = std::stof(argv[i]);
+        } else if (arg == "--min-p") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            sparams.min_p = std::stof(argv[i]);
         } else if (arg == "--temp") {
             if (++i >= argc) {
                 invalid_param = true;
                 break;
             }
             sparams.temp = std::stof(argv[i]);
+            sparams.temp = std::max(sparams.temp, 0.0f);
         } else if (arg == "--tfs") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -241,25 +302,26 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
                 invalid_param = true;
                 break;
             }
-            sparams.repeat_last_n = std::stoi(argv[i]);
+            sparams.penalty_last_n = std::stoi(argv[i]);
+            sparams.n_prev = std::max(sparams.n_prev, sparams.penalty_last_n);
         } else if (arg == "--repeat-penalty") {
             if (++i >= argc) {
                 invalid_param = true;
                 break;
             }
-            sparams.repeat_penalty = std::stof(argv[i]);
+            sparams.penalty_repeat = std::stof(argv[i]);
         } else if (arg == "--frequency-penalty") {
             if (++i >= argc) {
                 invalid_param = true;
                 break;
             }
-            sparams.frequency_penalty = std::stof(argv[i]);
+            sparams.penalty_freq = std::stof(argv[i]);
         } else if (arg == "--presence-penalty") {
             if (++i >= argc) {
                 invalid_param = true;
                 break;
             }
-            sparams.presence_penalty = std::stof(argv[i]);
+            sparams.penalty_present = std::stof(argv[i]);
         } else if (arg == "--mirostat") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -546,11 +608,8 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
                 break;
             }
         } else if (arg == "-h" || arg == "--help") {
-            gpt_print_usage(argc, argv, default_params);
+            return false;
-#ifndef LOG_DISABLE_LOGS
+
-            log_print_usage();
-#endif // LOG_DISABLE_LOGS
-            exit(0);
         } else if (arg == "--random-prompt") {
             params.random_prompt = true;
         } else if (arg == "--in-prefix-bos") {
@@ -572,7 +631,7 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
                 invalid_param = true;
                 break;
             }
-            params.grammar = argv[i];
+            sparams.grammar = argv[i];
         } else if (arg == "--grammar-file") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -587,7 +646,7 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
             std::copy(
                 std::istreambuf_iterator<char>(file),
                 std::istreambuf_iterator<char>(),
-                std::back_inserter(params.grammar)
+                std::back_inserter(sparams.grammar)
             );
 #ifndef LOG_DISABLE_LOGS
         // Parse args for logging parameters
@@ -609,28 +668,24 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
         // End of Parse args for logging parameters
 #endif // LOG_DISABLE_LOGS
         } else {
-            fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
+            throw std::invalid_argument("error: unknown argument: " + arg);
-            gpt_print_usage(argc, argv, default_params);
-            exit(1);
         }
     }
     if (invalid_param) {
-        fprintf(stderr, "error: invalid parameter for argument: %s\n", arg.c_str());
+        throw std::invalid_argument("error: invalid parameter for argument: " + arg);
-        gpt_print_usage(argc, argv, default_params);
-        exit(1);
     }
     if (params.prompt_cache_all &&
             (params.interactive || params.interactive_first ||
              params.instruct)) {
-        fprintf(stderr, "error: --prompt-cache-all not supported in interactive mode yet\n");
+
-        gpt_print_usage(argc, argv, default_params);
+        throw std::invalid_argument("error: --prompt-cache-all not supported in interactive mode yet\n");
-        exit(1);
     }
 
     if (params.escape) {
         process_escapes(params.prompt);
         process_escapes(params.input_prefix);
         process_escapes(params.input_suffix);
+        process_escapes(sparams.cfg_negative_prompt);
         for (auto & antiprompt : params.antiprompt) {
             process_escapes(antiprompt);
         }
@@ -640,8 +695,9 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
 }
 
 void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
-    const llama_sampling_params & sparams = params.sampling_params;
+    const llama_sampling_params & sparams = params.sparams;
 
+    printf("\n");
     printf("usage: %s [options]\n", argv[0]);
     printf("\n");
     printf("options:\n");
@@ -676,12 +732,13 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     printf("  -b N, --batch-size N  batch size for prompt processing (default: %d)\n", params.n_batch);
     printf("  --top-k N             top-k sampling (default: %d, 0 = disabled)\n", sparams.top_k);
     printf("  --top-p N             top-p sampling (default: %.1f, 1.0 = disabled)\n", (double)sparams.top_p);
+    printf("  --min-p N             min-p sampling (default: %.1f, 0.0 = disabled)\n", (double)sparams.min_p);
     printf("  --tfs N               tail free sampling, parameter z (default: %.1f, 1.0 = disabled)\n", (double)sparams.tfs_z);
     printf("  --typical N           locally typical sampling, parameter p (default: %.1f, 1.0 = disabled)\n", (double)sparams.typical_p);
-    printf("  --repeat-last-n N     last n tokens to consider for penalize (default: %d, 0 = disabled, -1 = ctx_size)\n", sparams.repeat_last_n);
+    printf("  --repeat-last-n N     last n tokens to consider for penalize (default: %d, 0 = disabled, -1 = ctx_size)\n", sparams.penalty_last_n);
-    printf("  --repeat-penalty N    penalize repeat sequence of tokens (default: %.1f, 1.0 = disabled)\n", (double)sparams.repeat_penalty);
+    printf("  --repeat-penalty N    penalize repeat sequence of tokens (default: %.1f, 1.0 = disabled)\n", (double)sparams.penalty_repeat);
-    printf("  --presence-penalty N  repeat alpha presence penalty (default: %.1f, 0.0 = disabled)\n", (double)sparams.presence_penalty);
+    printf("  --presence-penalty N  repeat alpha presence penalty (default: %.1f, 0.0 = disabled)\n", (double)sparams.penalty_present);
-    printf("  --frequency-penalty N repeat alpha frequency penalty (default: %.1f, 0.0 = disabled)\n", (double)sparams.frequency_penalty);
+    printf("  --frequency-penalty N repeat alpha frequency penalty (default: %.1f, 0.0 = disabled)\n", (double)sparams.penalty_freq);
     printf("  --mirostat N          use Mirostat sampling.\n");
     printf("                        Top K, Nucleus, Tail Free and Locally Typical samplers are ignored if used.\n");
     printf("                        (default: %d, 0 = disabled, 1 = Mirostat, 2 = Mirostat 2.0)\n", sparams.mirostat);
@@ -698,9 +755,16 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     printf("  --cfg-negative-prompt-file FNAME\n");
     printf("                        negative prompt file to use for guidance. (default: empty)\n");
     printf("  --cfg-scale N         strength of guidance (default: %f, 1.0 = disable)\n", sparams.cfg_scale);
-    printf("  --rope-scale N        RoPE context linear scaling factor, inverse of --rope-freq-scale\n");
+    printf("  --rope-scaling {none,linear,yarn}\n");
+    printf("                        RoPE frequency scaling method, defaults to linear unless specified by the model\n");
+    printf("  --rope-scale N        RoPE context scaling factor, expands context by a factor of N\n");
     printf("  --rope-freq-base N    RoPE base frequency, used by NTK-aware scaling (default: loaded from model)\n");
-    printf("  --rope-freq-scale N   RoPE frequency linear scaling factor (default: loaded from model)\n");
+    printf("  --rope-freq-scale N   RoPE frequency scaling factor, expands context by a factor of 1/N\n");
+    printf("  --yarn-orig-ctx N     YaRN: original context size of model (default: 0 = model training context size)\n");
+    printf("  --yarn-ext-factor N   YaRN: extrapolation mix factor (default: 1.0, 0.0 = full interpolation)\n");
+    printf("  --yarn-attn-factor N  YaRN: scale sqrt(t) or attention magnitude (default: 1.0)\n");
+    printf("  --yarn-beta-slow N    YaRN: high correction dim or alpha (default: %.1f)\n", params.yarn_beta_slow);
+    printf("  --yarn-beta-fast N    YaRN: low correction dim or beta (default: %.1f)\n", params.yarn_beta_fast);
     printf("  --ignore-eos          ignore end of stream token and continue generating (implies --logit-bias 2-inf)\n");
     printf("  --no-penalize-nl      do not penalize newline token\n");
     printf("  --memory-f32          use f32 instead of f16 for memory key+value (default: disabled)\n");
@@ -741,7 +805,7 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
 #endif // GGML_USE_CUBLAS
 #endif
     printf("  --verbose-prompt      print prompt before generation\n");
-    fprintf(stderr, "  --simple-io           use basic IO for better compatibility in subprocesses and limited consoles\n");
+    printf("  --simple-io           use basic IO for better compatibility in subprocesses and limited consoles\n");
     printf("  --lora FNAME          apply LoRA adapter (implies --no-mmap)\n");
     printf("  --lora-scaled FNAME S apply LoRA adapter with user defined scaling S (implies --no-mmap)\n");
     printf("  --lora-base FNAME     optional model to use as a base for the layers modified by the LoRA adapter\n");
@@ -752,6 +816,9 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     printf("  -ld LOGDIR, --logdir LOGDIR\n");
     printf("                        path under which to save YAML logs (no logging if unset)\n");
     printf("\n");
+#ifndef LOG_DISABLE_LOGS
+    log_print_usage();
+#endif // LOG_DISABLE_LOGS
 }
 
 std::string get_system_info(const gpt_params & params) {
@@ -805,21 +872,48 @@ struct llama_model_params llama_model_params_from_gpt_params(const gpt_params &
 struct llama_context_params llama_context_params_from_gpt_params(const gpt_params & params) {
     auto cparams = llama_context_default_params();
 
-    cparams.n_ctx           = params.n_ctx;
+    cparams.n_ctx             = params.n_ctx;
-    cparams.n_batch         = params.n_batch;
+    cparams.n_batch           = params.n_batch;
-    cparams.n_threads       = params.n_threads;
+    cparams.n_threads         = params.n_threads;
-    cparams.n_threads_batch = params.n_threads_batch == -1 ? params.n_threads : params.n_threads_batch;
+    cparams.n_threads_batch   = params.n_threads_batch == -1 ? params.n_threads : params.n_threads_batch;
-    cparams.mul_mat_q       = params.mul_mat_q;
+    cparams.mul_mat_q         = params.mul_mat_q;
-    cparams.seed            = params.seed;
+    cparams.seed              = params.seed;
-    cparams.f16_kv          = params.memory_f16;
+    cparams.f16_kv            = params.memory_f16;
-    cparams.logits_all      = params.logits_all;
+    cparams.logits_all        = params.logits_all;
-    cparams.embedding       = params.embedding;
+    cparams.embedding         = params.embedding;
-    cparams.rope_freq_base  = params.rope_freq_base;
+    cparams.rope_scaling_type = params.rope_scaling_type;
-    cparams.rope_freq_scale = params.rope_freq_scale;
+    cparams.rope_freq_base    = params.rope_freq_base;
+    cparams.rope_freq_scale   = params.rope_freq_scale;
+    cparams.yarn_ext_factor   = params.yarn_ext_factor;
+    cparams.yarn_attn_factor  = params.yarn_attn_factor;
+    cparams.yarn_beta_fast    = params.yarn_beta_fast;
+    cparams.yarn_beta_slow    = params.yarn_beta_slow;
+    cparams.yarn_orig_ctx     = params.yarn_orig_ctx;
 
     return cparams;
 }
 
+void llama_batch_clear(struct llama_batch & batch) {
+    batch.n_tokens = 0;
+}
+
+void llama_batch_add(
+                 struct llama_batch & batch,
+                        llama_token   id,
+                          llama_pos   pos,
+    const std::vector<llama_seq_id> & seq_ids,
+                               bool   logits) {
+    batch.token   [batch.n_tokens] = id;
+    batch.pos     [batch.n_tokens] = pos,
+    batch.n_seq_id[batch.n_tokens] = seq_ids.size();
+    for (size_t i = 0; i < seq_ids.size(); ++i) {
+        batch.seq_id[batch.n_tokens][i] = seq_ids[i];
+    }
+    batch.logits  [batch.n_tokens] = logits;
+
+    batch.n_tokens++;
+}
+
 std::tuple<struct llama_model *, struct llama_context *> llama_init_from_gpt_params(gpt_params & params) {
     auto mparams = llama_model_params_from_gpt_params(params);
 
@@ -857,15 +951,15 @@ std::tuple<struct llama_model *, struct llama_context *> llama_init_from_gpt_par
     }
 
     if (params.ignore_eos) {
-        params.sampling_params.logit_bias[llama_token_eos(lctx)] = -INFINITY;
+        params.sparams.logit_bias[llama_token_eos(model)] = -INFINITY;
     }
 
     {
         LOG("warming up the model with an empty run\n");
 
-        std::vector<llama_token> tmp = { llama_token_bos(lctx), llama_token_eos(lctx), };
+        std::vector<llama_token> tmp = { llama_token_bos(model), llama_token_eos(model), };
         llama_decode(lctx, llama_batch_get_one(tmp.data(), std::min(tmp.size(), (size_t) params.n_batch), 0, 0));
-        llama_kv_cache_tokens_rm(lctx, -1, -1);
+        llama_kv_cache_clear(lctx);
         llama_reset_timings(lctx);
     }
 
@@ -879,21 +973,23 @@ std::tuple<struct llama_model *, struct llama_context *> llama_init_from_gpt_par
 std::vector<llama_token> llama_tokenize(
   const struct llama_context * ctx,
            const std::string & text,
-                        bool   add_bos) {
+                        bool   add_bos,
-    return llama_tokenize(llama_get_model(ctx), text, add_bos);
+                        bool   special) {
+    return llama_tokenize(llama_get_model(ctx), text, add_bos, special);
 }
 
 std::vector<llama_token> llama_tokenize(
     const struct llama_model * model,
            const std::string & text,
-                        bool   add_bos) {
+                        bool   add_bos,
+                        bool   special) {
     // upper limit for the number of tokens
     int n_tokens = text.length() + add_bos;
     std::vector<llama_token> result(n_tokens);
-    n_tokens = llama_tokenize(model, text.data(), text.length(), result.data(), result.size(), add_bos);
+    n_tokens = llama_tokenize(model, text.data(), text.length(), result.data(), result.size(), add_bos, special);
     if (n_tokens < 0) {
         result.resize(-n_tokens);
-        int check = llama_tokenize(model, text.data(), text.length(), result.data(), result.size(), add_bos);
+        int check = llama_tokenize(model, text.data(), text.length(), result.data(), result.size(), add_bos, special);
         GGML_ASSERT(check == -n_tokens);
     } else {
         result.resize(n_tokens);
@@ -916,7 +1012,7 @@ std::string llama_token_to_piece(const struct llama_context * ctx, llama_token t
 }
 
 std::string llama_detokenize_spm(llama_context * ctx, const std::vector<llama_token> & tokens) {
-    const llama_token bos_id = llama_token_bos(ctx);
+    const llama_token bos_id = llama_token_bos(llama_get_model(ctx));
 
     std::string piece;
     std::string result;
@@ -1100,28 +1196,28 @@ std::string get_sortable_timestamp() {
 
 void dump_non_result_info_yaml(FILE * stream, const gpt_params & params, const llama_context * lctx,
                                const std::string & timestamp, const std::vector<int> & prompt_tokens, const char * model_desc) {
-    const llama_sampling_params & sparams = params.sampling_params;
+    const llama_sampling_params & sparams = params.sparams;
 
-    fprintf(stream, "build_commit: %s\n", BUILD_COMMIT);
+    fprintf(stream, "build_commit: %s\n",        LLAMA_COMMIT);
-    fprintf(stream, "build_number: %d\n", BUILD_NUMBER);
+    fprintf(stream, "build_number: %d\n",        LLAMA_BUILD_NUMBER);
-    fprintf(stream, "cpu_has_arm_fma: %s\n", ggml_cpu_has_arm_fma() ? "true" : "false");
+    fprintf(stream, "cpu_has_arm_fma: %s\n",     ggml_cpu_has_arm_fma()     ? "true" : "false");
-    fprintf(stream, "cpu_has_avx: %s\n", ggml_cpu_has_avx() ? "true" : "false");
+    fprintf(stream, "cpu_has_avx: %s\n",         ggml_cpu_has_avx()         ? "true" : "false");
-    fprintf(stream, "cpu_has_avx2: %s\n", ggml_cpu_has_avx2() ? "true" : "false");
+    fprintf(stream, "cpu_has_avx2: %s\n",        ggml_cpu_has_avx2()        ? "true" : "false");
-    fprintf(stream, "cpu_has_avx512: %s\n", ggml_cpu_has_avx512() ? "true" : "false");
+    fprintf(stream, "cpu_has_avx512: %s\n",      ggml_cpu_has_avx512()      ? "true" : "false");
     fprintf(stream, "cpu_has_avx512_vbmi: %s\n", ggml_cpu_has_avx512_vbmi() ? "true" : "false");
     fprintf(stream, "cpu_has_avx512_vnni: %s\n", ggml_cpu_has_avx512_vnni() ? "true" : "false");
-    fprintf(stream, "cpu_has_blas: %s\n", ggml_cpu_has_blas() ? "true" : "false");
+    fprintf(stream, "cpu_has_blas: %s\n",        ggml_cpu_has_blas()        ? "true" : "false");
-    fprintf(stream, "cpu_has_cublas: %s\n", ggml_cpu_has_cublas() ? "true" : "false");
+    fprintf(stream, "cpu_has_cublas: %s\n",      ggml_cpu_has_cublas()      ? "true" : "false");
-    fprintf(stream, "cpu_has_clblast: %s\n", ggml_cpu_has_clblast() ? "true" : "false");
+    fprintf(stream, "cpu_has_clblast: %s\n",     ggml_cpu_has_clblast()     ? "true" : "false");
-    fprintf(stream, "cpu_has_fma: %s\n", ggml_cpu_has_fma() ? "true" : "false");
+    fprintf(stream, "cpu_has_fma: %s\n",         ggml_cpu_has_fma()         ? "true" : "false");
-    fprintf(stream, "cpu_has_gpublas: %s\n", ggml_cpu_has_gpublas() ? "true" : "false");
+    fprintf(stream, "cpu_has_gpublas: %s\n",     ggml_cpu_has_gpublas()     ? "true" : "false");
-    fprintf(stream, "cpu_has_neon: %s\n", ggml_cpu_has_neon() ? "true" : "false");
+    fprintf(stream, "cpu_has_neon: %s\n",        ggml_cpu_has_neon()        ? "true" : "false");
-    fprintf(stream, "cpu_has_f16c: %s\n", ggml_cpu_has_f16c() ? "true" : "false");
+    fprintf(stream, "cpu_has_f16c: %s\n",        ggml_cpu_has_f16c()        ? "true" : "false");
-    fprintf(stream, "cpu_has_fp16_va: %s\n", ggml_cpu_has_fp16_va() ? "true" : "false");
+    fprintf(stream, "cpu_has_fp16_va: %s\n",     ggml_cpu_has_fp16_va()     ? "true" : "false");
-    fprintf(stream, "cpu_has_wasm_simd: %s\n", ggml_cpu_has_wasm_simd() ? "true" : "false");
+    fprintf(stream, "cpu_has_wasm_simd: %s\n",   ggml_cpu_has_wasm_simd()   ? "true" : "false");
-    fprintf(stream, "cpu_has_blas: %s\n", ggml_cpu_has_blas() ? "true" : "false");
+    fprintf(stream, "cpu_has_blas: %s\n",        ggml_cpu_has_blas()        ? "true" : "false");
-    fprintf(stream, "cpu_has_sse3: %s\n", ggml_cpu_has_sse3() ? "true" : "false");
+    fprintf(stream, "cpu_has_sse3: %s\n",        ggml_cpu_has_sse3()        ? "true" : "false");
-    fprintf(stream, "cpu_has_vsx: %s\n", ggml_cpu_has_vsx() ? "true" : "false");
+    fprintf(stream, "cpu_has_vsx: %s\n",         ggml_cpu_has_vsx()         ? "true" : "false");
 
 #ifdef NDEBUG
     fprintf(stream, "debug: false\n");
@@ -1155,13 +1251,13 @@ void dump_non_result_info_yaml(FILE * stream, const gpt_params & params, const l
     fprintf(stream, "ctx_size: %d # default: 512\n", params.n_ctx);
     fprintf(stream, "escape: %s # default: false\n", params.escape ? "true" : "false");
     fprintf(stream, "file: # never logged, see prompt instead. Can still be specified for input.\n");
-    fprintf(stream, "frequency_penalty: %f # default: 0.0 \n", sparams.frequency_penalty);
+    fprintf(stream, "frequency_penalty: %f # default: 0.0 \n", sparams.penalty_freq);
-    dump_string_yaml_multiline(stream, "grammar", params.grammar.c_str());
+    dump_string_yaml_multiline(stream, "grammar", sparams.grammar.c_str());
     fprintf(stream, "grammar-file: # never logged, see grammar instead. Can still be specified for input.\n");
     fprintf(stream, "hellaswag: %s # default: false\n", params.hellaswag ? "true" : "false");
     fprintf(stream, "hellaswag_tasks: %zu # default: 400\n", params.hellaswag_tasks);
 
-    const auto logit_bias_eos = sparams.logit_bias.find(llama_token_eos(lctx));
+    const auto logit_bias_eos = sparams.logit_bias.find(llama_token_eos(llama_get_model(lctx)));
     const bool ignore_eos = logit_bias_eos != sparams.logit_bias.end() && logit_bias_eos->second == -INFINITY;
     fprintf(stream, "ignore_eos: %s # default: false\n", ignore_eos ? "true" : "false");
 
@@ -1215,14 +1311,14 @@ void dump_non_result_info_yaml(FILE * stream, const gpt_params & params, const l
     fprintf(stream, "numa: %s # default: false\n", params.numa ? "true" : "false");
     fprintf(stream, "ppl_output_type: %d # default: 0\n", params.ppl_output_type);
     fprintf(stream, "ppl_stride: %d # default: 0\n", params.ppl_stride);
-    fprintf(stream, "presence_penalty: %f # default: 0.0\n", sparams.presence_penalty);
+    fprintf(stream, "presence_penalty: %f # default: 0.0\n", sparams.penalty_present);
     dump_string_yaml_multiline(stream, "prompt", params.prompt.c_str());
     fprintf(stream, "prompt_cache: %s\n", params.path_prompt_cache.c_str());
     fprintf(stream, "prompt_cache_all: %s # default: false\n", params.prompt_cache_all ? "true" : "false");
     fprintf(stream, "prompt_cache_ro: %s # default: false\n", params.prompt_cache_ro ? "true" : "false");
     dump_vector_int_yaml(stream, "prompt_tokens", prompt_tokens);
     fprintf(stream, "random_prompt: %s # default: false\n", params.random_prompt ? "true" : "false");
-    fprintf(stream, "repeat_penalty: %f # default: 1.1\n", sparams.repeat_penalty);
+    fprintf(stream, "repeat_penalty: %f # default: 1.1\n", sparams.penalty_repeat);
 
     fprintf(stream, "reverse_prompt:\n");
     for (std::string ap : params.antiprompt) {
@@ -1249,6 +1345,7 @@ void dump_non_result_info_yaml(FILE * stream, const gpt_params & params, const l
     fprintf(stream, "threads: %d # default: %d\n", params.n_threads, std::thread::hardware_concurrency());
     fprintf(stream, "top_k: %d # default: 40\n", sparams.top_k);
     fprintf(stream, "top_p: %f # default: 0.95\n", sparams.top_p);
+    fprintf(stream, "min_p: %f # default: 0.0\n", sparams.min_p);
     fprintf(stream, "typical_p: %f # default: 1.0\n", sparams.typical_p);
     fprintf(stream, "verbose_prompt: %s # default: false\n", params.verbose_prompt ? "true" : "false");
 }

common/common.h

@@ -9,6 +9,7 @@
 #define LOG_NO_FILE_LINE_FUNCTION
 #include "log.h"
 
+#include <cmath>
 #include <string>
 #include <vector>
 #include <random>
@@ -25,11 +26,17 @@
 #define die(msg)          do { fputs("error: " msg "\n", stderr);                exit(1); } while (0)
 #define die_fmt(fmt, ...) do { fprintf(stderr, "error: " fmt "\n", __VA_ARGS__); exit(1); } while (0)
 
-#define print_build_info() do {                                                             \
+#define print_build_info() do {                                                                     \
-    fprintf(stderr, "%s: build = %d (%s)\n", __func__, BUILD_NUMBER, BUILD_COMMIT);         \
+    fprintf(stderr, "%s: build = %d (%s)\n", __func__, LLAMA_BUILD_NUMBER, LLAMA_COMMIT);           \
-    fprintf(stderr, "%s: built with %s for %s\n", __func__, BUILD_COMPILER, BUILD_TARGET);  \
+    fprintf(stderr, "%s: built with %s for %s\n", __func__, LLAMA_COMPILER, LLAMA_BUILD_TARGET);    \
 } while(0)
 
+// build info
+extern int LLAMA_BUILD_NUMBER;
+extern char const *LLAMA_COMMIT;
+extern char const *LLAMA_COMPILER;
+extern char const *LLAMA_BUILD_TARGET;
+
 //
 // CLI argument parsing
 //
@@ -54,9 +61,15 @@ struct gpt_params {
     int32_t n_beams                         = 0;    // if non-zero then use beam search of given width.
     float   rope_freq_base                  = 0.0f; // RoPE base frequency
     float   rope_freq_scale                 = 0.0f; // RoPE frequency scaling factor
+    float   yarn_ext_factor                 = NAN;  // YaRN extrapolation mix factor
+    float   yarn_attn_factor                = 1.0f; // YaRN magnitude scaling factor
+    float   yarn_beta_fast                  = 32.0f;// YaRN low correction dim
+    float   yarn_beta_slow                  = 1.0f; // YaRN high correction dim
+    int32_t yarn_orig_ctx                   = 0;    // YaRN original context length
+    int8_t  rope_scaling_type               = LLAMA_ROPE_SCALING_UNSPECIFIED;
 
     // // sampling parameters
-    struct llama_sampling_params sampling_params;
+    struct llama_sampling_params sparams;
 
     std::string model             = "models/7B/ggml-model-f16.gguf"; // model path
     std::string model_draft       = "";                              // draft model for speculative decoding
@@ -66,10 +79,10 @@ struct gpt_params {
     std::string path_prompt_cache = "";  // path to file for saving/loading prompt eval state
     std::string input_prefix      = "";  // string to prefix user inputs with
     std::string input_suffix      = "";  // string to suffix user inputs with
-    std::string grammar           = "";  // optional BNF-like grammar to constrain sampling
     std::vector<std::string> antiprompt; // string upon seeing which more user input is prompted
     std::string logdir            = "";  // directory in which to save YAML log files
 
+    // TODO: avoid tuple, use struct
     std::vector<std::tuple<std::string, float>> lora_adapter; // lora adapter path with user defined scale
     std::string lora_base  = "";                              // base model path for the lora adapter
 
@@ -110,6 +123,8 @@ struct gpt_params {
     std::string image = ""; // path to an image file
 };
 
+bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params);
+
 bool gpt_params_parse(int argc, char ** argv, gpt_params & params);
 
 void gpt_print_usage(int argc, char ** argv, const gpt_params & params);
@@ -124,10 +139,23 @@ void process_escapes(std::string& input);
 // Model utils
 //
 
+// TODO: avoid tuplue, use struct
 std::tuple<struct llama_model *, struct llama_context *> llama_init_from_gpt_params(gpt_params & params);
-struct llama_model_params   llama_model_params_from_gpt_params(const gpt_params & params);
+
+struct llama_model_params   llama_model_params_from_gpt_params  (const gpt_params & params);
 struct llama_context_params llama_context_params_from_gpt_params(const gpt_params & params);
 
+// Batch utils
+
+void llama_batch_clear(struct llama_batch & batch);
+
+void llama_batch_add(
+                 struct llama_batch & batch,
+                        llama_token   id,
+                          llama_pos   pos,
+    const std::vector<llama_seq_id> & seq_ids,
+                               bool   logits);
+
 //
 // Vocab utils
 //
@@ -137,12 +165,14 @@ struct llama_context_params llama_context_params_from_gpt_params(const gpt_param
 std::vector<llama_token> llama_tokenize(
   const struct llama_context * ctx,
            const std::string & text,
-                        bool   add_bos);
+                        bool   add_bos,
+                        bool   special = false);
 
 std::vector<llama_token> llama_tokenize(
     const struct llama_model * model,
            const std::string & text,
-                        bool   add_bos);
+                        bool   add_bos,
+                        bool   special = false);
 
 // tokenizes a token into a piece
 // should work similar to Python's `tokenizer.id_to_piece`

common/grammar-parser.cpp

@@ -399,7 +399,7 @@ namespace grammar_parser {
     void print_grammar(FILE * file, const parse_state & state) {
         try {
             std::map<uint32_t, std::string> symbol_id_names;
-            for (auto kv : state.symbol_ids) {
+            for (const auto & kv : state.symbol_ids) {
                 symbol_id_names[kv.second] = kv.first;
             }
             for (size_t i = 0, end = state.rules.size(); i < end; i++) {

common/log.h

@@ -97,37 +97,56 @@
     #define LOG_TEE_TARGET stderr
 #endif
 
+// Utility for synchronizing log configuration state
+//  since std::optional was introduced only in c++17
+enum LogTriState
+{
+    LogTriStateSame,
+    LogTriStateFalse,
+    LogTriStateTrue
+};
+
 // Utility to obtain "pid" like unique process id and use it when creating log files.
 inline std::string log_get_pid()
 {
-    static std::string pid;
+   static std::string pid;
-    if (pid.empty())
+   if (pid.empty())
-    {
+   {
-        // std::this_thread::get_id() is the most portable way of obtaining a "process id"
+       // std::this_thread::get_id() is the most portable way of obtaining a "process id"
-        //  it's not the same as "pid" but is unique enough to solve multiple instances
+       //  it's not the same as "pid" but is unique enough to solve multiple instances
-        //  trying to write to the same log.
+       //  trying to write to the same log.
-        std::stringstream ss;
+       std::stringstream ss;
-        ss << std::this_thread::get_id();
+       ss << std::this_thread::get_id();
-        pid = ss.str();
+       pid = ss.str();
-    }
+   }
 
-    return pid;
+   return pid;
 }
 
 // Utility function for generating log file names with unique id based on thread id.
 //  invocation with log_filename_generator( "llama", "log" ) creates a string "llama.<number>.log"
 //  where the number is a runtime id of the current thread.
 
-#define log_filename_generator(log_file_basename, log_file_extension) log_filename_generator_impl(log_file_basename, log_file_extension)
+#define log_filename_generator(log_file_basename, log_file_extension) log_filename_generator_impl(LogTriStateSame, log_file_basename, log_file_extension)
 
 // INTERNAL, DO NOT USE
-inline std::string log_filename_generator_impl(const std::string & log_file_basename, const std::string & log_file_extension)
+inline std::string log_filename_generator_impl(LogTriState multilog, const std::string & log_file_basename, const std::string & log_file_extension)
 {
+    static bool _multilog = false;
+
+    if (multilog != LogTriStateSame)
+    {
+        _multilog = multilog == LogTriStateTrue;
+    }
+
     std::stringstream buf;
 
     buf << log_file_basename;
-    buf << ".";
+    if (_multilog)
-    buf << log_get_pid();
+    {
+        buf << ".";
+        buf << log_get_pid();
+    }
     buf << ".";
     buf << log_file_extension;
 
@@ -212,15 +231,6 @@ inline std::string log_filename_generator_impl(const std::string & log_file_base
     #define LOG_TEE_FLF_VAL ,""
 #endif
 
-// Utility for synchronizing log configuration state
-//  since std::optional was introduced only in c++17
-enum LogTriState
-{
-    LogTriStateSame,
-    LogTriStateFalse,
-    LogTriStateTrue
-};
-
 // INTERNAL, DO NOT USE
 //  USE LOG() INSTEAD
 //
@@ -314,16 +324,23 @@ enum LogTriState
 #endif
 
 // INTERNAL, DO NOT USE
-inline FILE *log_handler1_impl(bool change = false, LogTriState disable = LogTriStateSame, const std::string & filename = LOG_DEFAULT_FILE_NAME, FILE *target = nullptr)
+inline FILE *log_handler1_impl(bool change = false, LogTriState append = LogTriStateSame, LogTriState disable = LogTriStateSame, const std::string & filename = LOG_DEFAULT_FILE_NAME, FILE *target = nullptr)
 {
-    static bool _initialized{false};
+    static bool _initialized = false;
-    static bool _disabled{(filename.empty() && target == nullptr)};
+    static bool _append = false;
+    static bool _disabled = filename.empty() && target == nullptr;
     static std::string log_current_filename{filename};
     static FILE *log_current_target{target};
     static FILE *logfile = nullptr;
 
     if (change)
     {
+        if (append != LogTriStateSame)
+        {
+            _append = append == LogTriStateTrue;
+            return logfile;
+        }
+
         if (disable == LogTriStateTrue)
         {
             // Disable primary target
@@ -376,7 +393,7 @@ inline FILE *log_handler1_impl(bool change = false, LogTriState disable = LogTri
             }
         }
 
-        logfile = fopen(filename.c_str(), "w");
+        logfile = fopen(filename.c_str(), _append ? "a" : "w");
     }
 
     if (!logfile)
@@ -397,9 +414,9 @@ inline FILE *log_handler1_impl(bool change = false, LogTriState disable = LogTri
 }
 
 // INTERNAL, DO NOT USE
-inline FILE *log_handler2_impl(bool change = false, LogTriState disable = LogTriStateSame, FILE *target = nullptr, const std::string & filename = LOG_DEFAULT_FILE_NAME)
+inline FILE *log_handler2_impl(bool change = false, LogTriState append = LogTriStateSame, LogTriState disable = LogTriStateSame, FILE *target = nullptr, const std::string & filename = LOG_DEFAULT_FILE_NAME)
 {
-    return log_handler1_impl(change, disable, filename, target);
+    return log_handler1_impl(change, append, disable, filename, target);
 }
 
 // Disables logs entirely at runtime.
@@ -410,7 +427,7 @@ inline FILE *log_handler2_impl(bool change = false, LogTriState disable = LogTri
 // INTERNAL, DO NOT USE
 inline FILE *log_disable_impl()
 {
-    return log_handler1_impl(true, LogTriStateTrue);
+    return log_handler1_impl(true, LogTriStateSame, LogTriStateTrue);
 }
 
 // Enables logs at runtime.
@@ -419,19 +436,31 @@ inline FILE *log_disable_impl()
 // INTERNAL, DO NOT USE
 inline FILE *log_enable_impl()
 {
-    return log_handler1_impl(true, LogTriStateFalse);
+    return log_handler1_impl(true, LogTriStateSame, LogTriStateFalse);
 }
 
 // Sets target fir logs, either by a file name or FILE* pointer (stdout, stderr, or any valid FILE*)
 #define log_set_target(target) log_set_target_impl(target)
 
 // INTERNAL, DO NOT USE
-inline FILE *log_set_target_impl(const std::string & filename) { return log_handler1_impl(true, LogTriStateSame, filename); }
+inline FILE *log_set_target_impl(const std::string & filename) { return log_handler1_impl(true, LogTriStateSame, LogTriStateSame, filename); }
-inline FILE *log_set_target_impl(FILE *target) { return log_handler2_impl(true, LogTriStateSame, target); }
+inline FILE *log_set_target_impl(FILE *target) { return log_handler2_impl(true, LogTriStateSame, LogTriStateSame, target); }
 
 // INTERNAL, DO NOT USE
 inline FILE *log_handler() { return log_handler1_impl(); }
 
+// Enable or disable creating separate log files for each run.
+//  can ONLY be invoked BEFORE first log use.
+#define log_multilog(enable) log_filename_generator_impl((enable) ? LogTriStateTrue : LogTriStateFalse, "", "")
+// Enable or disable append mode for log file.
+//  can ONLY be invoked BEFORE first log use.
+#define log_append(enable) log_append_impl(enable)
+// INTERNAL, DO NOT USE
+inline FILE *log_append_impl(bool enable)
+{
+    return log_handler1_impl(true, enable ? LogTriStateTrue : LogTriStateFalse, LogTriStateSame);
+}
+
 inline void log_test()
 {
     log_disable();
@@ -493,6 +522,18 @@ inline bool log_param_single_parse(const std::string & param)
         return true;
     }
 
+    if (param == "--log-new")
+    {
+        log_multilog(true);
+        return true;
+    }
+
+    if (param == "--log-append")
+    {
+        log_append(true);
+        return true;
+    }
+
     return false;
 }
 
@@ -522,7 +563,9 @@ inline void log_print_usage()
     printf("  --log-disable         Disable trace logs\n");
     printf("  --log-enable          Enable trace logs\n");
     printf("  --log-file            Specify a log filename (without extension)\n");
-    printf("                        Log file will be tagged with unique ID and written as \"<name>.<ID>.log\"\n"); /*  */
+    printf("  --log-new             Create a separate new log file on start. "
+                                   "Each log file will have unique name: \"<name>.<ID>.log\"\n");
+    printf("  --log-append          Don't truncate the old log file.\n");
 }
 
 #define log_dump_cmdline(argc, argv) log_dump_cmdline_impl(argc, argv)
@@ -579,38 +622,75 @@ inline std::string log_var_to_string_impl(const std::vector<int> & var)
     return buf.str();
 }
 
-#define LOG_TOKENS_TOSTR_PRETTY(ctx, tokens)                                 \
+template <typename C, typename T>
-    [&tokens, &ctx]()                                                        \
+inline std::string LOG_TOKENS_TOSTR_PRETTY(const C & ctx, const T & tokens)
-    {                                                                        \
+{
-        std::stringstream buf;                                               \
+    std::stringstream buf;
-        buf << "[ ";                                                         \
+    buf << "[ ";
-                                                                             \
+
-        bool first = true;                                                   \
+    bool first = true;
-        for (const auto &token : tokens)                                     \
+    for (const auto &token : tokens)
-        {                                                                    \
+    {
-            if (!first)                                                      \
+        if (!first) {
-                buf << ", ";                                                 \
+            buf << ", ";
-            else                                                             \
+        } else {
-                first = false;                                               \
+            first = false;
-                                                                             \
+        }
-            auto detokenized = llama_token_to_piece(ctx, token);             \
+
-                                                                             \
+        auto detokenized = llama_token_to_piece(ctx, token);
-            detokenized.erase(                                               \
+
-                std::remove_if(                                              \
+        detokenized.erase(
-                    detokenized.begin(),                                     \
+            std::remove_if(
-                    detokenized.end(),                                       \
+                detokenized.begin(),
-                    [](const unsigned char c) { return !std::isprint(c); }), \
+                detokenized.end(),
-                detokenized.end());                                          \
+                [](const unsigned char c) { return !std::isprint(c); }),
-                                                                             \
+            detokenized.end());
-            buf                                                              \
+
-                << "'" << detokenized << "'"                                 \
+        buf
-                << ":" << std::to_string(token);                             \
+            << "'" << detokenized << "'"
-        }                                                                    \
+            << ":" << std::to_string(token);
-        buf << " ]";                                                         \
+    }
-                                                                             \
+    buf << " ]";
-        return buf.str();                                                    \
+
-    }()                                                                      \
+    return buf.str();
-        .c_str()
+}
+
+template <typename C, typename B>
+inline std::string LOG_BATCH_TOSTR_PRETTY(const C & ctx, const B & batch)
+{
+    std::stringstream buf;
+    buf << "[ ";
+
+    bool first = true;
+    for (int i = 0; i < batch.n_tokens; ++i)
+    {
+        if (!first) {
+            buf << ", ";
+        } else {
+            first = false;
+        }
+
+        auto detokenized = llama_token_to_piece(ctx, batch.token[i]);
+
+        detokenized.erase(
+            std::remove_if(
+                detokenized.begin(),
+                detokenized.end(),
+                [](const unsigned char c) { return !std::isprint(c); }),
+            detokenized.end());
+
+        buf
+            << "\n" << std::to_string(i)
+            << ":token '" << detokenized << "'"
+            << ":pos " << std::to_string(batch.pos[i])
+            << ":n_seq_id  " << std::to_string(batch.n_seq_id[i])
+            << ":seq_id " << std::to_string(batch.seq_id[i][0])
+            << ":logits " << std::to_string(batch.logits[i]);
+    }
+    buf << " ]";
+
+    return buf.str();
+}
 
 #ifdef LOG_DISABLE_LOGS
 

common/sampling.cpp

@@ -1,113 +1,163 @@
 #include "sampling.h"
 
-llama_sampling_context::~llama_sampling_context() {
+struct llama_sampling_context * llama_sampling_init(const struct llama_sampling_params & params) {
-    for (auto & it : sequence_contexts) {
+    struct llama_sampling_context * result = new llama_sampling_context();
-        if (it.second.grammar != NULL) {
+
-            llama_grammar_free(it.second.grammar);
+    result->params  = params;
-            it.second.grammar = NULL;
+    result->grammar = nullptr;
+
+    // if there is a grammar, parse it
+    if (!params.grammar.empty()) {
+        result->parsed_grammar = grammar_parser::parse(params.grammar.c_str());
+
+        // will be empty (default) if there are parse errors
+        if (result->parsed_grammar.rules.empty()) {
+            fprintf(stderr, "%s: failed to parse grammar\n", __func__);
+            return nullptr;
         }
+
+        std::vector<const llama_grammar_element *> grammar_rules(result->parsed_grammar.c_rules());
+
+        result->grammar = llama_grammar_init(
+                grammar_rules.data(),
+                grammar_rules.size(), result->parsed_grammar.symbol_ids.at("root"));
     }
+
+    result->prev.resize(params.n_prev);
+
+    return result;
 }
 
-llama_sampling_context llama_sampling_context_init(
+void llama_sampling_free(struct llama_sampling_context * ctx) {
-        const struct gpt_params & params,
+    if (ctx->grammar != NULL) {
-                  llama_grammar * grammar) {
+        llama_grammar_free(ctx->grammar);
-  llama_sampling_context result;
+    }
 
-  result.params = params.sampling_params;
+    delete ctx;
-  result.grammar = grammar;
-  return result;
 }
 
-// Note: Creates the context if it doesn't exist, so this always return something.
+void llama_sampling_reset(llama_sampling_context * ctx) {
-llama_sampler_sequence_context & llama_sampling_get_sequence_context(
+    if (ctx->grammar != NULL) {
-              llama_sampling_context & ctx_sampling,
+        llama_grammar_free(ctx->grammar);
-        const llama_seq_id             seq) {
+        ctx->grammar = NULL;
-    const auto it = ctx_sampling.sequence_contexts.find(seq);
-    if (it != ctx_sampling.sequence_contexts.end()) {
-        return it->second;
     }
-    llama_sampler_sequence_context new_ctx = {
+
-        2.0f * ctx_sampling.params.mirostat_tau,
+    if (!ctx->parsed_grammar.rules.empty()) {
-        ctx_sampling.grammar != NULL ? llama_grammar_copy(ctx_sampling.grammar) : NULL,
+        std::vector<const llama_grammar_element *> grammar_rules(ctx->parsed_grammar.c_rules());
-    };
+
-    return ctx_sampling.sequence_contexts.insert({seq, new_ctx}).first->second;
+        ctx->grammar = llama_grammar_init(
+                grammar_rules.data(),
+                grammar_rules.size(), ctx->parsed_grammar.symbol_ids.at("root"));
+    }
+
+    std::fill(ctx->prev.begin(), ctx->prev.end(), 0);
+    ctx->cur.clear();
 }
 
-bool llama_sampling_context_reset(
+void llama_sampling_cp(llama_sampling_context * src, llama_sampling_context * dst) {
-              llama_sampling_context & ctx_sampling,
+    if (dst->grammar) {
-        const llama_seq_id             seq) {
+        llama_grammar_free(dst->grammar);
-    const auto it = ctx_sampling.sequence_contexts.find(seq);
+        dst->grammar = nullptr;
-    if (it == ctx_sampling.sequence_contexts.end()) return false;
-    if (it->second.grammar != NULL) {
-        llama_grammar_free(it->second.grammar);
-        it->second.grammar = NULL;
     }
-    ctx_sampling.sequence_contexts.erase(it);
+
-    return true;
+    if (src->grammar) {
+        dst->grammar = llama_grammar_copy(src->grammar);
+    }
+
+    dst->prev = src->prev;
+}
+
+llama_token llama_sampling_last(llama_sampling_context * ctx) {
+    return ctx->prev.back();
+}
+
+std::string llama_sampling_prev_str(llama_sampling_context * ctx_sampling, llama_context * ctx_main, int n) {
+    const int size = ctx_sampling->prev.size();
+
+    n = std::min(n, size);
+
+    std::string result;
+
+    for (int i = size - n; i < size; i++) {
+        result += llama_token_to_piece(ctx_main, ctx_sampling->prev[i]);
+    }
+
+    return result;
+}
+
+std::string llama_sampling_print(const llama_sampling_params & params) {
+    char result[1024];
+
+    snprintf(result, sizeof(result),
+            "\trepeat_last_n = %d, repeat_penalty = %.3f, frequency_penalty = %.3f, presence_penalty = %.3f\n"
+            "\ttop_k = %d, tfs_z = %.3f, top_p = %.3f, min_p = %.3f, typical_p = %.3f, temp = %.3f\n"
+            "\tmirostat = %d, mirostat_lr = %.3f, mirostat_ent = %.3f",
+            params.penalty_last_n, params.penalty_repeat, params.penalty_freq, params.penalty_present,
+            params.top_k, params.tfs_z, params.top_p, params.min_p, params.typical_p, params.temp,
+            params.mirostat, params.mirostat_eta, params.mirostat_tau);
+
+    return std::string(result);
 }
 
 llama_token llama_sampling_sample(
-                  struct llama_context * ctx,
+                  struct llama_sampling_context * ctx_sampling,
-                  struct llama_context * ctx_guidance,
+                  struct llama_context * ctx_main,
-                  struct llama_sampling_context & ctx_sampling,
+                  struct llama_context * ctx_cfg,
-        const std::vector<llama_token> & last_tokens,
+                  const int idx) {
-         std::vector<llama_token_data> & candidates,
+    const llama_sampling_params & params = ctx_sampling->params;
-        const                      int   idx,
+
-                          llama_seq_id   seq) {
+    const int n_vocab = llama_n_vocab(llama_get_model(ctx_main));
-    const int n_ctx   = llama_n_ctx(ctx);
-    const int n_vocab = llama_n_vocab(llama_get_model(ctx));
 
-    const llama_sampling_params & params = ctx_sampling.params;
     const float   temp            = params.temp;
     const int32_t top_k           = params.top_k <= 0 ? n_vocab : params.top_k;
     const float   top_p           = params.top_p;
+    const float   min_p           = params.min_p;
     const float   tfs_z           = params.tfs_z;
     const float   typical_p       = params.typical_p;
-    const int32_t repeat_last_n   = params.repeat_last_n < 0 ? n_ctx : params.repeat_last_n;
+    const int32_t penalty_last_n  = params.penalty_last_n < 0 ? params.n_prev : params.penalty_last_n;
-    const float   repeat_penalty  = params.repeat_penalty;
+    const float   penalty_repeat  = params.penalty_repeat;
-    const float   alpha_presence  = params.presence_penalty;
+    const float   penalty_freq    = params.penalty_freq;
-    const float   alpha_frequency = params.frequency_penalty;
+    const float   penalty_present = params.penalty_present;
     const int     mirostat        = params.mirostat;
     const float   mirostat_tau    = params.mirostat_tau;
     const float   mirostat_eta    = params.mirostat_eta;
     const bool    penalize_nl     = params.penalize_nl;
 
+    auto & prev = ctx_sampling->prev;
+    auto & cur  = ctx_sampling->cur;
+
     llama_token id = 0;
 
-    float * logits = llama_get_logits_ith(ctx, idx);
+    float * logits = llama_get_logits_ith(ctx_main, idx);
 
-    // Apply params.logit_bias map
+    // apply params.logit_bias map
     for (auto it = params.logit_bias.begin(); it != params.logit_bias.end(); it++) {
         logits[it->first] += it->second;
     }
 
-    candidates.clear();
+    cur.clear();
+
     for (llama_token token_id = 0; token_id < n_vocab; token_id++) {
-        candidates.emplace_back(llama_token_data{token_id, logits[token_id], 0.0f});
+        cur.emplace_back(llama_token_data{token_id, logits[token_id], 0.0f});
     }
 
-    llama_token_data_array cur_p = { candidates.data(), candidates.size(), false };
+    llama_token_data_array cur_p = { cur.data(), cur.size(), false };
 
-    if (ctx_guidance) {
+    if (ctx_cfg) {
-        llama_sample_classifier_free_guidance(ctx, &cur_p, ctx_guidance, params.cfg_scale);
+        llama_sample_classifier_free_guidance(ctx_main, &cur_p, ctx_cfg, params.cfg_scale);
     }
 
     // apply penalties
-    if (!last_tokens.empty()) {
+    if (!prev.empty()) {
-        const float nl_logit = logits[llama_token_nl(ctx)];
+        const float nl_logit = logits[llama_token_nl(llama_get_model(ctx_main))];
-        const int last_n_repeat = std::min(std::min((int)last_tokens.size(), repeat_last_n), n_ctx);
 
-        llama_sample_repetition_penalty(ctx, &cur_p,
+        llama_sample_repetition_penalties(ctx_main, &cur_p,
-                last_tokens.data() + last_tokens.size() - last_n_repeat,
+                prev.data() + prev.size() - penalty_last_n,
-                last_n_repeat, repeat_penalty);
+                penalty_last_n, penalty_repeat, penalty_freq, penalty_present);
-        llama_sample_frequency_and_presence_penalties(ctx, &cur_p,
-                last_tokens.data() + last_tokens.size() - last_n_repeat,
-                last_n_repeat, alpha_frequency, alpha_presence);
 
         if (!penalize_nl) {
             for (size_t idx = 0; idx < cur_p.size; idx++) {
-                if (cur_p.data[idx].id == llama_token_nl(ctx)) {
+                if (cur_p.data[idx].id == llama_token_nl(llama_get_model(ctx_main))) {
                     cur_p.data[idx].logit = nl_logit;
                     break;
                 }
@@ -115,52 +165,65 @@ llama_token llama_sampling_sample(
         }
     }
 
-    llama_sampler_sequence_context & ctx_seq = llama_sampling_get_sequence_context(ctx_sampling, seq);
+    if (ctx_sampling->grammar != NULL) {
-
+        llama_sample_grammar(ctx_main, &cur_p, ctx_sampling->grammar);
-    if (ctx_seq.grammar != NULL) {
-        llama_sample_grammar(ctx, &cur_p, ctx_seq.grammar);
     }
 
-    if (temp <= 0) {
+    if (temp < 0.0) {
-        // Greedy sampling
+        // greedy sampling, with probs
-        id = llama_sample_token_greedy(ctx, &cur_p);
+        llama_sample_softmax(ctx_main, &cur_p);
+        id = cur_p.data[0].id;
+    } else if (temp == 0.0) {
+        // greedy sampling, no probs
+        id = llama_sample_token_greedy(ctx_main, &cur_p);
     } else {
         if (mirostat == 1) {
             const int mirostat_m = 100;
-            llama_sample_temp(ctx, &cur_p, temp);
+            llama_sample_temp(ctx_main, &cur_p, temp);
-            id = llama_sample_token_mirostat(ctx, &cur_p, mirostat_tau, mirostat_eta, mirostat_m, &ctx_seq.mirostat_mu);
+            id = llama_sample_token_mirostat(ctx_main, &cur_p, mirostat_tau, mirostat_eta, mirostat_m, &ctx_sampling->mirostat_mu);
         } else if (mirostat == 2) {
-            llama_sample_temp(ctx, &cur_p, temp);
+            llama_sample_temp(ctx_main, &cur_p, temp);
-            id = llama_sample_token_mirostat_v2(ctx, &cur_p, mirostat_tau, mirostat_eta, &ctx_seq.mirostat_mu);
+            id = llama_sample_token_mirostat_v2(ctx_main, &cur_p, mirostat_tau, mirostat_eta, &ctx_sampling->mirostat_mu);
         } else {
-            // Temperature sampling
+            // temperature sampling
             size_t min_keep = std::max(1, params.n_probs);
-            llama_sample_top_k      (ctx, &cur_p, top_k, min_keep);
-            llama_sample_tail_free  (ctx, &cur_p, tfs_z, min_keep);
-            llama_sample_typical    (ctx, &cur_p, typical_p, min_keep);
-            llama_sample_top_p      (ctx, &cur_p, top_p, min_keep);
-            llama_sample_temp(ctx, &cur_p, temp);
 
-            {
+            llama_sample_top_k    (ctx_main, &cur_p, top_k,     min_keep);
-                const int n_top = 10;
+            llama_sample_tail_free(ctx_main, &cur_p, tfs_z,     min_keep);
-                LOG("top %d candidates:\n", n_top);
+            llama_sample_typical  (ctx_main, &cur_p, typical_p, min_keep);
+            llama_sample_top_p    (ctx_main, &cur_p, top_p,     min_keep);
+            llama_sample_min_p    (ctx_main, &cur_p, min_p,     min_keep);
+            llama_sample_temp     (ctx_main, &cur_p, temp);
 
-                for (int i = 0; i < n_top; i++) {
+            id = llama_sample_token(ctx_main, &cur_p);
-                    const llama_token id = cur_p.data[i].id;
-                    (void)id; // To avoid a warning that id is unused when logging is disabled.
-                    LOG(" - %5d: '%12s' (%.3f)\n", id, llama_token_to_piece(ctx, id).c_str(), cur_p.data[i].p);
-                }
-            }
 
-            id = llama_sample_token(ctx, &cur_p);
+            //{
+            //    const int n_top = 10;
+            //    LOG("top %d candidates:\n", n_top);
 
-            LOG("sampled token: %5d: '%s'\n", id, llama_token_to_piece(ctx, id).c_str());
+            //    for (int i = 0; i < n_top; i++) {
+            //        const llama_token id = cur_p.data[i].id;
+            //        (void)id; // To avoid a warning that id is unused when logging is disabled.
+            //        LOG(" - %5d: '%12s' (%.3f)\n", id, llama_token_to_piece(ctx_main, id).c_str(), cur_p.data[i].p);
+            //    }
+            //}
+
+            LOG("sampled token: %5d: '%s'\n", id, llama_token_to_piece(ctx_main, id).c_str());
         }
     }
 
-    if (ctx_seq.grammar != NULL) {
-        llama_grammar_accept_token(ctx, ctx_seq.grammar, id);
-    }
-
     return id;
 }
+
+void llama_sampling_accept(
+        struct llama_sampling_context * ctx_sampling,
+        struct llama_context * ctx_main,
+        llama_token id,
+        bool apply_grammar) {
+    ctx_sampling->prev.erase(ctx_sampling->prev.begin());
+    ctx_sampling->prev.push_back(id);
+
+    if (ctx_sampling->grammar != NULL && apply_grammar) {
+        llama_grammar_accept_token(ctx_main, ctx_sampling->grammar, id);
+    }
+}

common/sampling.h

@@ -2,107 +2,109 @@
 
 #include "llama.h"
 
+#include "grammar-parser.h"
+
 #include <string>
 #include <vector>
 #include <unordered_map>
 
 // sampling parameters
 typedef struct llama_sampling_params {
+    int32_t n_prev            = 64;    // number of previous tokens to remember
+    int32_t n_probs           = 0;     // if greater than 0, output the probabilities of top n_probs tokens.
     int32_t top_k             = 40;    // <= 0 to use vocab size
     float   top_p             = 0.95f; // 1.0 = disabled
+    float   min_p             = 0.05f; // 0.0 = disabled
     float   tfs_z             = 1.00f; // 1.0 = disabled
     float   typical_p         = 1.00f; // 1.0 = disabled
     float   temp              = 0.80f; // 1.0 = disabled
-    float   repeat_penalty    = 1.10f; // 1.0 = disabled
+    int32_t penalty_last_n    = 64;    // last n tokens to penalize (0 = disable penalty, -1 = context size)
-    int32_t repeat_last_n     = 64;    // last n tokens to penalize (0 = disable penalty, -1 = context size)
+    float   penalty_repeat    = 1.10f; // 1.0 = disabled
-    float   frequency_penalty = 0.00f; // 0.0 = disabled
+    float   penalty_freq      = 0.00f; // 0.0 = disabled
-    float   presence_penalty  = 0.00f; // 0.0 = disabled
+    float   penalty_present   = 0.00f; // 0.0 = disabled
     int32_t mirostat          = 0;     // 0 = disabled, 1 = mirostat, 2 = mirostat 2.0
     float   mirostat_tau      = 5.00f; // target entropy
     float   mirostat_eta      = 0.10f; // learning rate
-
     bool    penalize_nl       = true;  // consider newlines as a repeatable token
 
-    int32_t n_probs           = 0;     // if greater than 0, output the probabilities of top n_probs tokens.
+    std::string grammar;  // optional BNF-like grammar to constrain sampling
 
     // Classifier-Free Guidance
     // https://arxiv.org/abs/2306.17806
-    std::string cfg_negative_prompt;   // string to help guidance
+    std::string cfg_negative_prompt; // string to help guidance
-    float       cfg_scale     = 1.f;   // How strong is guidance
+    float       cfg_scale     = 1.f; // how strong is guidance
 
     std::unordered_map<llama_token, float> logit_bias; // logit bias for specific tokens
-
 } llama_sampling_params;
 
-// per-sequence sampler context
-typedef struct llama_sampler_sequence_context {
-    float mirostat_mu; // mirostat sampler state
-    llama_grammar * grammar;
-} llama_sampler_sequence_context;
-
 // general sampler context
-typedef struct llama_sampling_context {
+// TODO: move to llama.h
-    ~llama_sampling_context();
+struct llama_sampling_context {
-
+    // parameters that will be used for sampling
-    // parameters that will be used for sampling and when creating
-    // new llama_sampler_sequence_context instances
     llama_sampling_params params;
 
-    // map of sequence ids to sampler contexts
+    // mirostat sampler state
-    std::unordered_map<llama_seq_id, llama_sampler_sequence_context> sequence_contexts;
+    float mirostat_mu;
 
-    // when non-NULL, new instances of llama_sampler_sequence_context
-    // will get a copy of the grammar here
-    // note: only the pointer is stored here, it is not a copy of
-    //       the grammar and shouldn't be freed
     llama_grammar * grammar;
-} llama_sampling_context;
+
+    // internal
+    grammar_parser::parse_state parsed_grammar;
+
+    // TODO: replace with ring-buffer
+    std::vector<llama_token>      prev;
+    std::vector<llama_token_data> cur;
+};
 
 #include "common.h"
 
 // Create a new sampling context instance.
-llama_sampling_context llama_sampling_context_init(
+struct llama_sampling_context * llama_sampling_init(const struct llama_sampling_params & params);
-        const struct gpt_params & params,
-                  llama_grammar * grammar = NULL);
 
-// Fetches the sampler context for the specified sequence id (defaults to 0).
+void llama_sampling_free(struct llama_sampling_context * ctx);
-// If the context for that sequence id doesn't already exist, it will be created with
-// default values based on the parameters in the ctx_sampling argument.
-llama_sampler_sequence_context & llama_sampling_get_sequence_context(
-              llama_sampling_context & ctx_sampling,
-        const llama_seq_id             seq = 0);
 
-// Reset the sampler context for the supplied sequence id (defaults to 0).
+// Reset the sampler context
-// This is necessary to reuse a sequence id or free memory used by sequences
+// - clear prev tokens
-// that are no longer required.
+// - reset grammar
-bool llama_sampling_context_reset(
+void llama_sampling_reset(llama_sampling_context * ctx);
-              llama_sampling_context & ctx_sampling,
+
-        const llama_seq_id             seq = 0);
+// Copy the sampler context
+void llama_sampling_cp(llama_sampling_context * src, llama_sampling_context * dst);
+
+// Get the last sampled token
+llama_token llama_sampling_last(llama_sampling_context * ctx);
+
+// Get a string representation of the last sampled tokens
+std::string llama_sampling_prev_str(llama_sampling_context * ctx_sampling, llama_context * ctx_main, int n);
+
+// Print sampling parameters into a string
+std::string llama_sampling_print(const llama_sampling_params & params);
 
 // this is a common sampling function used across the examples for convenience
 // it can serve as a starting point for implementing your own sampling function
 // Note: When using multiple sequences, it is the caller's responsibility to call
-//       llama_sampling_context_reset when a sequence ends
+//       llama_sampling_reset when a sequence ends
 //
 // required:
-//  - ctx:          context to use for sampling
+//  - ctx_main:     context to use for sampling
 //  - ctx_sampling: sampling-specific context
 //
 // optional:
-//  - ctx_guidance:  context to use for classifier-free guidance, ignore if NULL
+//  - ctx_cfg:      context to use for classifier-free guidance
-//  - last_tokens:   needed for repetition penalty, ignore if empty
+//  - idx:          sample from llama_get_logits_ith(ctx, idx)
-//  - idx:           sample from llama_get_logits_ith(ctx, idx)
-//  - seq:           sequence id to associate sampler state with
 //
 // returns:
 //  - token:      sampled token
 //  - candidates: vector of candidate tokens
 //
 llama_token llama_sampling_sample(
-                  struct llama_context * ctx,
+        struct llama_sampling_context * ctx_sampling,
-                  struct llama_context * ctx_guidance,
+        struct llama_context * ctx_main,
-                  struct llama_sampling_context & ctx_sampling,
+        struct llama_context * ctx_cfg,
-        const std::vector<llama_token> & last_tokens,
+        int idx = 0);
-         std::vector<llama_token_data> & candidates,
+
-        const                      int   idx = 0,
+void llama_sampling_accept(
-                          llama_seq_id   seq = 0);
+        struct llama_sampling_context * ctx_sampling,
+        struct llama_context * ctx_main,
+        llama_token id,
+        bool apply_grammar);

common/train.cpp

@@ -236,8 +236,8 @@ int64_t get_example_targets_batch(
     int64_t used_samples = 0;
 
     ggml_set_f32(target_probs, 0.0f);
-    llama_token bos = llama_token_bos(lctx);
+    llama_token bos = llama_token_bos(llama_get_model(lctx));
-    llama_token eos = llama_token_eos(lctx);
+    llama_token eos = llama_token_eos(llama_get_model(lctx));
     // printf("%s: example_id=%d n_batch=%d n_train_samples=%zu\n", __func__, example_id, n_batch, n_train_samples);
     for (int k=0; k<n_batch; ++k) {
         // printf("%s: batch %d\n", __func__, k);
@@ -863,7 +863,7 @@ size_t tokenize_file(
             (int) buf.size(),
             out_tokens.data(),
             (int) out_tokens.size(),
-            false);
+            false, false);
         if (n_tokens < 0) {
             out_tokens.resize(-n_tokens);
             n_tokens = llama_tokenize(
@@ -872,7 +872,7 @@ size_t tokenize_file(
                 (int) buf.size(),
                 out_tokens.data(),
                 (int) out_tokens.size(),
-                false);
+                false, false);
         }
         if (n_tokens >= 0) {
             out_tokens.resize(n_tokens);
@@ -924,7 +924,7 @@ size_t tokenize_file(
         for (llama_token token=0; token < n_vocab; ++token) {
             max_token_text_size = std::max(
                 max_token_text_size,
-                strlen(llama_token_get_text(lctx, token)));
+                strlen(llama_token_get_text(llama_get_model(lctx), token)));
         }
 
         // upper bound of context byte length.
@@ -966,7 +966,7 @@ size_t tokenize_file(
                     (int) buf_sample.size(),
                     tok_sample.data(),
                     (int) tok_sample.size(),
-                    false);
+                    false, false);
                 if (n_tokens < 0) {
                     tok_sample.resize(-n_tokens);
                     n_tokens = llama_tokenize(llama_get_model(lctx),
@@ -974,7 +974,7 @@ size_t tokenize_file(
                         (int) buf_sample.size(),
                         tok_sample.data(),
                         (int) tok_sample.size(),
-                        false);
+                        false, false);
                     GGML_ASSERT(n_tokens >= 0);
                 }
                 GGML_ASSERT(n_tokens <= (int) tok_sample.size());
@@ -1045,6 +1045,7 @@ struct train_params_common get_default_train_params_common() {
     params.n_batch    =    8;
     params.n_gradient_accumulation = 1;
     params.n_epochs   = -1;
+    params.n_gpu_layers = 0;
 
     params.custom_n_ctx = false;
 
@@ -1080,6 +1081,7 @@ struct train_params_common get_default_train_params_common() {
     params.adam_beta2          = 0.999f;
     params.adam_gclip          = 1.0f;
     params.adam_eps_f          = 0.0f;
+
     return params;
 }
 
@@ -1425,7 +1427,7 @@ void train_opt_callback(void * vdata, int accum_step, float * sched, bool * canc
 
         int impr_plot = -(int)(1 + (opt->loss_before - opt->loss_after) * 10.0f + 0.5f);
         if (impr_plot > 0) impr_plot = 0;
-        if (std::isnan(opt->loss_before) || std::isnan(opt->loss_before)) impr_plot = 0;
+        if (std::isnan(opt->loss_before) || std::isnan(opt->loss_after)) impr_plot = 0;
         printf("%s: iter=%6d sample=%zu/%zu sched=%f loss=%f",
             __func__, opt->iter, std::min(1+train->shuffle_next_sample, train->shuffle_sample_count), train->shuffle_sample_count,
             *sched, opt->loss_after);

common/train.h

@@ -44,6 +44,7 @@ struct train_params_common {
     int n_batch;
     int n_gradient_accumulation;
     int n_epochs;
+    int n_gpu_layers;
 
     bool custom_n_ctx;
 

convert-baichuan-hf-to-gguf.py

@@ -76,6 +76,7 @@ def parse_args() -> argparse.Namespace:
         "ftype", type=int, choices=[0, 1], default=1, nargs='?',
         help="output format - use 0 for float32, 1 for float16",
     )
+    parser.add_argument("--bigendian",   action="store_true",    help="model is executed on big endian machine")
     return parser.parse_args()
 
 args = parse_args()
@@ -86,6 +87,11 @@ if not dir_model.is_dir():
     print(f'Error: {args.model} is not a directory', file = sys.stderr)
     sys.exit(1)
 
+endianess = gguf.GGUFEndian.LITTLE
+if args.bigendian:
+    endianess = gguf.GGUFEndian.BIG
+endianess_str = "Big Endian" if args.bigendian else "Little Endian"
+print(f"gguf: Conversion Endianess {endianess}")
 # possible tensor data types
 #   ftype == 0 -> float32
 #   ftype == 1 -> float16
@@ -104,7 +110,7 @@ print("gguf: loading model "+dir_model.name)
 with open(dir_model / "config.json", "r", encoding="utf-8") as f:
     hparams = json.load(f)
 print("hello print: ",hparams["architectures"][0])
-if hparams["architectures"][0] != "BaichuanForCausalLM":
+if hparams["architectures"][0] != "BaichuanForCausalLM" and hparams["architectures"][0] != "BaiChuanForCausalLM":
     print("Model architecture not supported: " + hparams["architectures"][0])
 
     sys.exit()
@@ -113,7 +119,7 @@ if hparams["architectures"][0] != "BaichuanForCausalLM":
 num_parts = count_model_parts(dir_model)
 print(f"num_parts:{num_parts}\n")
 ARCH=gguf.MODEL_ARCH.BAICHUAN
-gguf_writer = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[ARCH])
+gguf_writer = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[ARCH], endianess=endianess)
 
 print("gguf: get model metadata")
 
@@ -157,7 +163,8 @@ gguf_writer.add_layer_norm_rms_eps(hparams["rms_norm_eps"])
 if "rope_scaling" in hparams and hparams["rope_scaling"] != None and "factor" in hparams["rope_scaling"]:
     if "type" in hparams["rope_scaling"]:
         if hparams["rope_scaling"]["type"] == "linear":
-            gguf_writer.add_rope_scale_linear(hparams["rope_scaling"]["factor"])
+            gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR)
+            gguf_writer.add_rope_scaling_factor(hparams["rope_scaling"]["factor"])
 
 
 # TOKENIZATION
@@ -224,7 +231,7 @@ gguf_writer.add_token_list(tokens)
 gguf_writer.add_token_scores(scores)
 gguf_writer.add_token_types(toktypes)
 
-special_vocab = gguf.SpecialVocab(dir_model)
+special_vocab = gguf.SpecialVocab(dir_model, n_vocab = len(tokens))
 special_vocab.add_to_gguf(gguf_writer)
 
 # TENSORS

convert-bloom-hf-to-gguf.py

@@ -118,18 +118,27 @@ tokenizer = AutoTokenizer.from_pretrained(dir_model)
 vocab_size = hparams.get("vocab_size", len(tokenizer.vocab))
 assert max(tokenizer.vocab.values()) < vocab_size
 
+added_vocab = tokenizer.get_added_vocab()
 reverse_vocab = {id: encoded_tok for encoded_tok, id in tokenizer.vocab.items()}
 
 for i in range(vocab_size):
-    tokens.append(reverse_vocab[i] if i in reverse_vocab else f"[PAD{i}]")
+    if i not in reverse_vocab:
-    scores.append(0.0)  # dummy
+        tokens.append(f"[PAD{i}]")
-    toktypes.append(gguf.TokenType.NORMAL)
+        toktypes.append(gguf.TokenType.USER_DEFINED)
+    elif reverse_vocab[i] in added_vocab:
+        tokens.append(reverse_vocab[i])
+        if tokenizer.added_tokens_decoder[i].special:
+            toktypes.append(gguf.TokenType.CONTROL)
+        else:
+            toktypes.append(gguf.TokenType.USER_DEFINED)
+    else:
+        tokens.append(reverse_vocab[i])
+        toktypes.append(gguf.TokenType.NORMAL)
 
 gguf_writer.add_token_list(tokens)
-gguf_writer.add_token_scores(scores)
 gguf_writer.add_token_types(toktypes)
 
-special_vocab = gguf.SpecialVocab(dir_model, load_merges=True)
+special_vocab = gguf.SpecialVocab(dir_model, load_merges=True, n_vocab = len(tokens))
 special_vocab.add_to_gguf(gguf_writer)
 
 # TENSORS

convert-falcon-hf-to-gguf.py

@@ -78,7 +78,7 @@ print("gguf: loading model "+dir_model.name)
 with open(dir_model / "config.json", "r", encoding="utf-8") as f:
     hparams = json.load(f)
 
-if hparams["architectures"][0] != "FalconForCausalLM":
+if hparams["architectures"][0] not in ("RWForCausalLM", "FalconForCausalLM"):
     print("Model architecture not supported: " + hparams["architectures"][0])
 
     sys.exit(1)
@@ -97,7 +97,17 @@ gguf_writer = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[ARCH])
 
 print("gguf: get model metadata")
 
-block_count = hparams["num_hidden_layers"]
+block_count = hparams.get("num_hidden_layers")
+if block_count is None:
+    block_count = hparams["n_layer"]  # old name
+
+n_head = hparams.get("num_attention_heads")
+if n_head is None:
+    n_head = hparams["n_head"]  # old name
+
+n_head_kv = hparams.get("num_kv_heads")
+if n_head_kv is None:
+    n_head_kv = hparams.get("n_head_kv", 1)  # old name
 
 gguf_writer.add_name("Falcon")
 gguf_writer.add_context_length(2048) # not in config.json
@@ -105,11 +115,8 @@ gguf_writer.add_tensor_data_layout("jploski") # qkv tensor transform
 gguf_writer.add_embedding_length(hparams["hidden_size"])
 gguf_writer.add_feed_forward_length(4 * hparams["hidden_size"])
 gguf_writer.add_block_count(block_count)
-gguf_writer.add_head_count(hparams["num_attention_heads"])
+gguf_writer.add_head_count(n_head)
-if "num_kv_heads" in hparams:
+gguf_writer.add_head_count_kv(n_head_kv)
-    gguf_writer.add_head_count_kv(hparams["num_kv_heads"])
-else:
-    gguf_writer.add_head_count_kv(1)
 gguf_writer.add_layer_norm_eps(hparams["layer_norm_epsilon"])
 gguf_writer.add_file_type(ftype)
 
@@ -145,17 +152,13 @@ gguf_writer.add_token_list(tokens)
 gguf_writer.add_token_scores(scores)
 gguf_writer.add_token_types(toktypes)
 
-special_vocab = gguf.SpecialVocab(dir_model, load_merges = True)
+special_vocab = gguf.SpecialVocab(dir_model, load_merges = True, n_vocab = len(tokens))
 special_vocab.add_to_gguf(gguf_writer)
 
 # TENSORS
 
 tensor_map = gguf.get_tensor_name_map(ARCH,block_count)
 
-# params for qkv transform
-n_head    = hparams["num_attention_heads"]
-n_head_kv = hparams["num_kv_heads"] if "num_kv_heads" in hparams else 1
-
 head_dim = hparams["hidden_size"] // n_head
 
 # tensor info

convert-gptneox-hf-to-gguf.py

@@ -123,18 +123,27 @@ tokenizer = AutoTokenizer.from_pretrained(dir_model)
 vocab_size = hparams.get("vocab_size", len(tokenizer.vocab))
 assert max(tokenizer.vocab.values()) < vocab_size
 
+added_vocab = tokenizer.get_added_vocab()
 reverse_vocab = {id: encoded_tok for encoded_tok, id in tokenizer.vocab.items()}
 
 for i in range(vocab_size):
-    tokens.append(reverse_vocab[i] if i in reverse_vocab else f"[PAD{i}]")
+    if i not in reverse_vocab:
-    scores.append(0.0) # dummy
+        tokens.append(f"[PAD{i}]")
-    toktypes.append(gguf.TokenType.NORMAL)
+        toktypes.append(gguf.TokenType.USER_DEFINED)
+    elif reverse_vocab[i] in added_vocab:
+        tokens.append(reverse_vocab[i])
+        if tokenizer.added_tokens_decoder[i].special:
+            toktypes.append(gguf.TokenType.CONTROL)
+        else:
+            toktypes.append(gguf.TokenType.USER_DEFINED)
+    else:
+        tokens.append(reverse_vocab[i])
+        toktypes.append(gguf.TokenType.NORMAL)
 
 gguf_writer.add_token_list(tokens)
-gguf_writer.add_token_scores(scores)
 gguf_writer.add_token_types(toktypes)
 
-special_vocab = gguf.SpecialVocab(dir_model, load_merges = True)
+special_vocab = gguf.SpecialVocab(dir_model, load_merges = True, n_vocab = len(tokens))
 special_vocab.add_to_gguf(gguf_writer)
 
 # TENSORS

convert-llama-ggml-to-gguf.py

@@ -388,7 +388,9 @@ def handle_metadata(cfg, hp):
         cfg.vocab_dir if cfg.vocab_dir is not None else cfg.model_metadata_dir,
         cfg.vocabtype )
     # FIXME: Respect cfg.vocab_dir?
-    svocab = gguf.SpecialVocab(cfg.model_metadata_dir)
+    svocab = gguf.SpecialVocab(cfg.model_metadata_dir,
+        load_merges = cfg.vocabtype == 'bpe',
+        n_vocab = vocab.vocab_size)
     convert.check_vocab_size(params, vocab)
     return (params, vocab, svocab)
 

convert-mpt-hf-to-gguf.py

@@ -98,6 +98,8 @@ gguf_writer.add_embedding_length(hparams["d_model"])
 gguf_writer.add_block_count(block_count)
 gguf_writer.add_feed_forward_length(4 * hparams["d_model"])
 gguf_writer.add_head_count(hparams["n_heads"])
+if kv_n_heads := hparams["attn_config"].get("kv_n_heads"):
+    gguf_writer.add_head_count_kv(kv_n_heads)
 gguf_writer.add_layer_norm_eps(1e-05)
 if hparams["attn_config"]["clip_qkv"] is not None:
     gguf_writer.add_clamp_kqv(hparams["attn_config"]["clip_qkv"])
@@ -126,18 +128,27 @@ vocab_size = hparams["vocab_size"]
 # ref: https://github.com/cmp-nct/ggllm.cpp/blob/master/falcon_convert.py
 tokenizer = AutoTokenizer.from_pretrained(dir_model)
 
+added_vocab = tokenizer.get_added_vocab()
 reverse_vocab = {id: encoded_tok for encoded_tok, id in tokenizer.vocab.items()}
 
 for i in range(vocab_size):
-    tokens.append(reverse_vocab[i] if i in reverse_vocab else f"[PAD{i}]")
+    if i not in reverse_vocab:
-    scores.append(0.0) # dummy
+        tokens.append(f"[PAD{i}]")
-    toktypes.append(gguf.TokenType.NORMAL)
+        toktypes.append(gguf.TokenType.USER_DEFINED)
+    elif reverse_vocab[i] in added_vocab:
+        tokens.append(reverse_vocab[i])
+        if tokenizer.added_tokens_decoder[i].special:
+            toktypes.append(gguf.TokenType.CONTROL)
+        else:
+            toktypes.append(gguf.TokenType.USER_DEFINED)
+    else:
+        tokens.append(reverse_vocab[i])
+        toktypes.append(gguf.TokenType.NORMAL)
 
 gguf_writer.add_token_list(tokens)
-gguf_writer.add_token_scores(scores)
 gguf_writer.add_token_types(toktypes)
 
-special_vocab = gguf.SpecialVocab(dir_model, load_merges = True)
+special_vocab = gguf.SpecialVocab(dir_model, load_merges = True, n_vocab = len(tokens))
 special_vocab.add_to_gguf(gguf_writer)
 
 # TENSORS

convert-refact-hf-to-gguf.py

@@ -139,18 +139,27 @@ tokenizer = AutoTokenizer.from_pretrained(dir_model)
 vocab_size = hparams.get("vocab_size", len(tokenizer.vocab))
 assert max(tokenizer.vocab.values()) < vocab_size
 
+added_vocab = tokenizer.get_added_vocab()
 reverse_vocab = {id: encoded_tok for encoded_tok, id in tokenizer.vocab.items()}
 
 for i in range(vocab_size):
-    tokens.append(reverse_vocab[i] if i in reverse_vocab else f"[PAD{i}]")
+    if i not in reverse_vocab:
-    scores.append(0.0) # dummy
+        tokens.append(f"[PAD{i}]")
-    toktypes.append(gguf.TokenType.NORMAL)
+        toktypes.append(gguf.TokenType.USER_DEFINED)
+    elif reverse_vocab[i] in added_vocab:
+        tokens.append(reverse_vocab[i])
+        if tokenizer.added_tokens_decoder[i].special:
+            toktypes.append(gguf.TokenType.CONTROL)
+        else:
+            toktypes.append(gguf.TokenType.USER_DEFINED)
+    else:
+        tokens.append(reverse_vocab[i])
+        toktypes.append(gguf.TokenType.NORMAL)
 
 gguf_writer.add_token_list(tokens)
-gguf_writer.add_token_scores(scores)
 gguf_writer.add_token_types(toktypes)
 
-special_vocab = gguf.SpecialVocab(dir_model, load_merges=True)
+special_vocab = gguf.SpecialVocab(dir_model, load_merges=True, n_vocab = len(tokens))
 special_vocab.add_to_gguf(gguf_writer)
 
 # TENSORS

convert-starcoder-hf-to-gguf.py

@@ -111,18 +111,26 @@ tokenizer = AutoTokenizer.from_pretrained(dir_model)
 vocab_size = hparams.get("vocab_size", len(tokenizer.vocab))
 assert max(tokenizer.vocab.values()) < vocab_size
 
+added_vocab = tokenizer.get_added_vocab()
 reverse_vocab = {id: encoded_tok for encoded_tok, id in tokenizer.vocab.items()}
 
 for i in range(vocab_size):
-    tokens.append(reverse_vocab[i] if i in reverse_vocab else f"[PAD{i}]")
+    if i not in reverse_vocab:
-    scores.append(0.0) # dummy
+        tokens.append(f"[PAD{i}]")
-    toktypes.append(gguf.TokenType.NORMAL)
+        toktypes.append(gguf.TokenType.USER_DEFINED)
+    elif reverse_vocab[i] in added_vocab:
+        tokens.append(reverse_vocab[i])
+        if tokenizer.added_tokens_decoder[i].special:
+            toktypes.append(gguf.TokenType.CONTROL)
+        else:
+            toktypes.append(gguf.TokenType.USER_DEFINED)
+    else:
+        tokens.append(reverse_vocab[i])
+        toktypes.append(gguf.TokenType.NORMAL)
 
 gguf_writer.add_token_list(tokens)
-gguf_writer.add_token_scores(scores)
 gguf_writer.add_token_types(toktypes)
-
+special_vocab = gguf.SpecialVocab(dir_model, load_merges = True, n_vocab = len(tokens))
-special_vocab = gguf.SpecialVocab(dir_model, load_merges = True)
 special_vocab.add_to_gguf(gguf_writer)
 
 # TENSORS

convert.py

@@ -151,8 +151,11 @@ class Params:
     n_head_kv:  int
     f_norm_eps: float
 
+    rope_scaling_type: gguf.RopeScalingType | None = None
     f_rope_freq_base: float | None = None
     f_rope_scale: float | None = None
+    n_orig_ctx: int | None = None
+    rope_finetuned: bool | None = None
 
     ftype: GGMLFileType | None = None
 
@@ -198,20 +201,20 @@ class Params:
     def loadHFTransformerJson(model: LazyModel, config_path: Path) -> Params:
         config = json.load(open(config_path))
 
-        n_vocab          = config["vocab_size"]
+        rope_scaling_type = f_rope_scale = n_orig_ctx = rope_finetuned = None
-        n_embd           = config["hidden_size"]
-        n_layer          = config["num_hidden_layers"]
-        n_ff             = config["intermediate_size"]
-        n_head           = config["num_attention_heads"]
-        n_head_kv        = config["num_key_value_heads"] if "num_key_value_heads" in config else n_head
-        f_norm_eps       = config["rms_norm_eps"]
-        f_rope_freq_base = config["rope_theta"] if "rope_theta" in config else None
-
         rope_scaling = config.get("rope_scaling")
-        if isinstance(rope_scaling, dict) and rope_scaling.get("type") == "linear":
+
-            f_rope_scale = config["rope_scaling"].get("factor")
+        if rope_scaling is not None and (typ := rope_scaling.get("type")):
-        else:
+            rope_factor = rope_scaling.get("factor")
-            f_rope_scale = None
+            f_rope_scale = rope_factor
+            if typ == "linear":
+                rope_scaling_type = gguf.RopeScalingType.LINEAR
+            elif typ == "yarn":
+                rope_scaling_type = gguf.RopeScalingType.YARN
+                n_orig_ctx = rope_scaling['original_max_position_embeddings']
+                rope_finetuned = rope_scaling['finetuned']
+            else:
+                raise NotImplementedError(f'Unknown rope scaling type: {typ}')
 
         if "max_sequence_length" in config:
             n_ctx = config["max_sequence_length"]
@@ -222,16 +225,19 @@ class Params:
                             "Suggestion: provide 'config.json' of the model in the same directory containing model files.")
 
         return Params(
-            n_vocab          = n_vocab,
+            n_vocab           = config["vocab_size"],
-            n_embd           = n_embd,
+            n_embd            = config["hidden_size"],
-            n_layer          = n_layer,
+            n_layer           = config["num_hidden_layers"],
-            n_ctx            = n_ctx,
+            n_ctx             = n_ctx,
-            n_ff             = n_ff,
+            n_ff              = config["intermediate_size"],
-            n_head           = n_head,
+            n_head            = (n_head := config["num_attention_heads"]),
-            n_head_kv        = n_head_kv,
+            n_head_kv         = config.get("num_key_value_heads", n_head),
-            f_norm_eps       = f_norm_eps,
+            f_norm_eps        = config["rms_norm_eps"],
-            f_rope_freq_base = f_rope_freq_base,
+            f_rope_freq_base  = config.get("rope_theta"),
-            f_rope_scale     = f_rope_scale,
+            rope_scaling_type = rope_scaling_type,
+            f_rope_scale      = f_rope_scale,
+            n_orig_ctx        = n_orig_ctx,
+            rope_finetuned    = rope_finetuned,
         )
 
     # LLaMA v2 70B params.json
@@ -240,17 +246,8 @@ class Params:
     def loadOriginalParamsJson(model: LazyModel, config_path: Path) -> Params:
         config = json.load(open(config_path))
 
-        n_vocab          = config["vocab_size"] if "vocab_size" in config else -1
-        n_embd           = config["dim"]
-        n_layer          = config["n_layers"]
-        n_ff             = -1
-        n_head           = config["n_heads"]
-        n_head_kv        = config["n_kv_heads"] if "n_kv_heads" in config else n_head
-        f_norm_eps       = config["norm_eps"]
-        f_rope_freq_base = config["rope_theta"] if "rope_theta" in config else None
-
         # hack to determine LLaMA v1 vs v2 vs CodeLlama
-        if f_rope_freq_base == 1000000:
+        if config.get("rope_theta") == 1000000:
             # CodeLlama
             n_ctx = 16384
         elif config["norm_eps"] == 1e-05:
@@ -260,22 +257,16 @@ class Params:
             # LLaMA v1
             n_ctx = 2048
 
-        if n_vocab == -1:
-            n_vocab = model["tok_embeddings.weight"].shape[0]
-
-        if n_ff == -1:
-            n_ff = model["layers.0.feed_forward.w1.weight"].shape[0]
-
         return Params(
-            n_vocab          = n_vocab,
+            n_vocab          = config.get("vocab_size", model["tok_embeddings.weight"].shape[0]),
-            n_embd           = n_embd,
+            n_embd           = config["dim"],
-            n_layer          = n_layer,
+            n_layer          = config["n_layers"],
             n_ctx            = n_ctx,
-            n_ff             = n_ff,
+            n_ff             = model["layers.0.feed_forward.w1.weight"].shape[0],
-            n_head           = n_head,
+            n_head           = (n_head := config["n_heads"]),
-            n_head_kv        = n_head_kv,
+            n_head_kv        = config.get("n_kv_heads", n_head),
-            f_norm_eps       = f_norm_eps,
+            f_norm_eps       = config["norm_eps"],
-            f_rope_freq_base = f_rope_freq_base,
+            f_rope_freq_base = config.get("rope_theta"),
         )
 
     @staticmethod
@@ -366,16 +357,19 @@ class SentencePieceVocab:
             added_tokens = {}
 
         vocab_size: int = self.sentencepiece_tokenizer.vocab_size()
-        expected_ids = list(range(vocab_size, vocab_size + len(added_tokens)))
-        actual_ids   = sorted(added_tokens.values())
-        if expected_ids != actual_ids:
-            raise Exception(f"Expected added token IDs to be sequential and start at {len(added_tokens)}; got {actual_ids}")
 
-        items = sorted(added_tokens.items(), key=lambda text_idx: text_idx[1])
+        new_tokens       = {id: piece for piece, id in added_tokens.items() if id >= vocab_size}
-        self.added_tokens_list = [text for (text, idx) in items]
+        expected_new_ids = list(range(vocab_size, vocab_size + len(new_tokens)))
-        self.vocab_size_base: int = vocab_size
+        actual_new_ids   = sorted(new_tokens.keys())
-        self.vocab_size: int = self.vocab_size_base + len(self.added_tokens_list)
+
-        self.fname_tokenizer = fname_tokenizer
+        if expected_new_ids != actual_new_ids:
+            raise ValueError(f"Expected new token IDs {expected_new_ids} to be sequential; got {actual_new_ids}")
+
+        # Token pieces that were added to the base vocabulary.
+        self.added_tokens_list  = [new_tokens[id] for id in actual_new_ids]
+        self.vocab_size_base    = vocab_size
+        self.vocab_size         = self.vocab_size_base + len(self.added_tokens_list)
+        self.fname_tokenizer    = fname_tokenizer
         self.fname_added_tokens = fname_added_tokens
 
     def sentencepiece_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]:
@@ -803,8 +797,8 @@ def check_vocab_size(params: Params, vocab: Vocab) -> None:
 
 
 class OutputFile:
-    def __init__(self, fname_out: Path) -> None:
+    def __init__(self, fname_out: Path, endianess:gguf.GGUFEndian=gguf.GGUFEndian.LITTLE) -> None:
-        self.gguf = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[ARCH])
+        self.gguf = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[ARCH], endianess=endianess)
 
     def add_meta_arch(self, params: Params) -> None:
         name = "LLaMA"
@@ -828,8 +822,16 @@ class OutputFile:
         if params.f_rope_freq_base is not None:
             self.gguf.add_rope_freq_base(params.f_rope_freq_base)
 
-        if params.f_rope_scale is not None:
+        if params.rope_scaling_type:
-            self.gguf.add_rope_scale_linear(params.f_rope_scale)
+            assert params.f_rope_scale is not None
+            self.gguf.add_rope_scaling_type(params.rope_scaling_type)
+            self.gguf.add_rope_scaling_factor(params.f_rope_scale)
+
+        if params.n_orig_ctx is not None:
+            self.gguf.add_rope_scaling_orig_ctx_len(params.n_orig_ctx)
+
+        if params.rope_finetuned is not None:
+            self.gguf.add_rope_scaling_finetuned(params.rope_finetuned)
 
         if params.ftype is not None:
             self.gguf.add_file_type(params.ftype)
@@ -875,10 +877,10 @@ class OutputFile:
         self.gguf.close()
 
     @staticmethod
-    def write_vocab_only(fname_out: Path, params: Params, vocab: Vocab, svocab: gguf.SpecialVocab) -> None:
+    def write_vocab_only(fname_out: Path, params: Params, vocab: Vocab, svocab: gguf.SpecialVocab, endianess:gguf.GGUFEndian=gguf.GGUFEndian.LITTLE) -> None:
         check_vocab_size(params, vocab)
 
-        of = OutputFile(fname_out)
+        of = OutputFile(fname_out, endianess=endianess)
 
         # meta data
         of.add_meta_arch(params)
@@ -903,10 +905,10 @@ class OutputFile:
         return dt.quantize(arr)
 
     @staticmethod
-    def write_all(fname_out: Path, ftype: GGMLFileType, params: Params, model: LazyModel, vocab: Vocab, svocab: gguf.SpecialVocab, concurrency: int = DEFAULT_CONCURRENCY) -> None:
+    def write_all(fname_out: Path, ftype: GGMLFileType, params: Params, model: LazyModel, vocab: Vocab, svocab: gguf.SpecialVocab, concurrency: int = DEFAULT_CONCURRENCY, endianess=gguf.GGUFEndian.LITTLE) -> None:
         check_vocab_size(params, vocab)
 
-        of = OutputFile(fname_out)
+        of = OutputFile(fname_out, endianess=endianess)
 
         # meta data
         of.add_meta_arch(params)
@@ -1123,8 +1125,9 @@ def main(args_in: list[str] | None = None) -> None:
     parser.add_argument("--vocabtype",   choices=["spm", "bpe"], help="vocab format (default: spm)", default="spm")
     parser.add_argument("--ctx",         type=int,               help="model training context (default: based on input)")
     parser.add_argument("--concurrency", type=int,               help=f"concurrency used for conversion (default: {DEFAULT_CONCURRENCY})", default = DEFAULT_CONCURRENCY)
-    args = parser.parse_args(args_in)
+    parser.add_argument("--bigendian",   action="store_true",    help="model is executed on big endian machine")
 
+    args = parser.parse_args(args_in)
     if args.dump_single:
         model_plus = lazy_load_file(args.model)
         do_dump_model(model_plus)
@@ -1138,6 +1141,9 @@ def main(args_in: list[str] | None = None) -> None:
     if args.dump:
         do_dump_model(model_plus)
         return
+    endianess = gguf.GGUFEndian.LITTLE
+    if args.bigendian:
+        endianess = gguf.GGUFEndian.BIG
 
     params = Params.load(model_plus)
     if params.n_ctx == -1:
@@ -1159,10 +1165,13 @@ def main(args_in: list[str] | None = None) -> None:
 
     vocab: Vocab
     if args.vocab_only:
-        assert args.outfile, "need --outfile if using --vocab-only"
+        if not args.outfile:
+            raise ValueError("need --outfile if using --vocab-only")
         # FIXME: Try to respect vocab_dir somehow?
         vocab = load_vocab(args.vocab_dir or args.model, args.vocabtype)
-        special_vocab = gguf.SpecialVocab(model_plus.paths[0].parent, load_merges = args.vocabtype == 'bpe')
+        special_vocab = gguf.SpecialVocab(model_plus.paths[0].parent,
+            load_merges = args.vocabtype == 'bpe',
+            n_vocab = vocab.vocab_size)
         outfile = args.outfile
         OutputFile.write_vocab_only(outfile, params, vocab, special_vocab)
         print(f"Wrote {outfile}")
@@ -1174,7 +1183,9 @@ def main(args_in: list[str] | None = None) -> None:
         vocab_dir = args.vocab_dir if args.vocab_dir else model_plus.paths[0].parent
         vocab = load_vocab(vocab_dir, args.vocabtype)
     # FIXME: Try to respect vocab_dir somehow?
-    special_vocab = gguf.SpecialVocab(model_plus.paths[0].parent, load_merges = args.vocabtype == 'bpe')
+    special_vocab = gguf.SpecialVocab(model_plus.paths[0].parent,
+        load_merges = args.vocabtype == 'bpe',
+        n_vocab = vocab.vocab_size)
 
     model   = model_plus.model
     model   = convert_model_names(model, params)
@@ -1185,7 +1196,7 @@ def main(args_in: list[str] | None = None) -> None:
     params.ftype = ftype
     print(f"Writing {outfile}, format {ftype}")
 
-    OutputFile.write_all(outfile, ftype, params, model, vocab, special_vocab, concurrency = args.concurrency)
+    OutputFile.write_all(outfile, ftype, params, model, vocab, special_vocab, concurrency = args.concurrency, endianess=endianess)
     print(f"Wrote {outfile}")
 
 

examples/CMakeLists.txt

@@ -12,26 +12,26 @@ include_directories(${CMAKE_CURRENT_SOURCE_DIR})
 
 if (EMSCRIPTEN)
 else()
-    add_subdirectory(main)
-    add_subdirectory(quantize)
-    add_subdirectory(quantize-stats)
-    add_subdirectory(perplexity)
-    add_subdirectory(embedding)
-    add_subdirectory(save-load-state)
-    add_subdirectory(benchmark)
     add_subdirectory(baby-llama)
-    add_subdirectory(train-text-from-scratch)
-    add_subdirectory(finetune)
-    add_subdirectory(convert-llama2c-to-ggml)
-    add_subdirectory(simple)
     add_subdirectory(batched)
     add_subdirectory(batched-bench)
-    add_subdirectory(speculative)
-    add_subdirectory(parallel)
-    add_subdirectory(embd-input)
-    add_subdirectory(llava)
-    add_subdirectory(llama-bench)
     add_subdirectory(beam-search)
+    add_subdirectory(benchmark)
+    add_subdirectory(convert-llama2c-to-ggml)
+    add_subdirectory(embedding)
+    add_subdirectory(finetune)
+    add_subdirectory(infill)
+    add_subdirectory(llama-bench)
+    add_subdirectory(llava)
+    add_subdirectory(main)
+    add_subdirectory(parallel)
+    add_subdirectory(perplexity)
+    add_subdirectory(quantize)
+    add_subdirectory(quantize-stats)
+    add_subdirectory(save-load-state)
+    add_subdirectory(simple)
+    add_subdirectory(speculative)
+    add_subdirectory(train-text-from-scratch)
     if (LLAMA_METAL)
         add_subdirectory(metal)
     endif()

examples/batched-bench/batched-bench.cpp

@@ -114,7 +114,7 @@ int main(int argc, char ** argv) {
         return 1;
     }
 
-    llama_batch batch = llama_batch_init(n_kv_max, 0);
+    llama_batch batch = llama_batch_init(n_kv_max, 0, 1);
 
     // decode in batches of ctx_params.n_batch tokens
     auto decode_helper = [](llama_context * ctx, llama_batch & batch, int32_t n_batch) {
@@ -123,11 +123,12 @@ int main(int argc, char ** argv) {
 
             llama_batch batch_view = {
                 n_tokens,
-                batch.token  + i,
+                batch.token    + i,
                 nullptr,
-                batch.pos    + i,
+                batch.pos      + i,
-                batch.seq_id + i,
+                batch.n_seq_id + i,
-                batch.logits + i,
+                batch.seq_id   + i,
+                batch.logits   + i,
                 0, 0, 0, // unused
             };
 
@@ -143,13 +144,8 @@ int main(int argc, char ** argv) {
 
     // warm up
     {
-        batch.n_tokens = 16;
+        for (int i = 0; i < 16; ++i) {
-
+            llama_batch_add(batch, 0, i, { 0 }, false);
-        for (int i = 0; i < batch.n_tokens; ++i) {
-            batch.token[i]  = 0;
-            batch.pos[i]    = i;
-            batch.seq_id[i] = 0;
-            batch.logits[i] = false;
         }
 
         if (!decode_helper(ctx, batch, ctx_params.n_batch)) {
@@ -158,6 +154,10 @@ int main(int argc, char ** argv) {
         }
     }
 
+    LOG_TEE("\n");
+    LOG_TEE("%s: n_kv_max = %d, is_pp_shared = %d, n_gpu_layers = %d, mmq = %d\n", __func__, n_kv_max, is_pp_shared, n_gpu_layers, mmq);
+    LOG_TEE("\n");
+
     LOG_TEE("|%6s | %6s | %4s | %6s | %8s | %8s | %8s | %8s | %8s | %8s |\n", "PP",     "TG",     "B",    "N_KV",     "T_PP s",   "S_PP t/s", "T_TG s",   "S_TG t/s", "T s",      "S t/s");
     LOG_TEE("|%6s-|-%6s-|-%4s-|-%6s-|-%8s-|-%8s-|-%8s-|-%8s-|-%8s-|-%8s-|\n", "------", "------", "----", "------", "--------", "--------", "--------", "--------", "--------", "--------");
 
@@ -174,19 +174,18 @@ int main(int argc, char ** argv) {
                     continue;
                 }
 
-                batch.n_tokens = is_pp_shared ? pp : pl*pp;
+                llama_batch_clear(batch);
 
-                for (int i = 0; i < batch.n_tokens; ++i) {
+                const int n_tokens = is_pp_shared ? pp : pl*pp;
-                    batch.token[i]  = 0;
+
-                    batch.pos[i]    = i;
+                for (int i = 0; i < n_tokens; ++i) {
-                    batch.seq_id[i] = 0;
+                    llama_batch_add(batch, 0, i, { 0 }, false);
-                    batch.logits[i] = false;
                 }
                 batch.logits[batch.n_tokens - 1] = true;
 
                 const auto t_pp_start = ggml_time_us();
 
-                llama_kv_cache_tokens_rm(ctx, -1, -1);
+                llama_kv_cache_clear(ctx);
 
                 if (!decode_helper(ctx, batch, ctx_params.n_batch)) {
                     LOG_TEE("%s: llama_decode() failed\n", __func__);
@@ -204,13 +203,10 @@ int main(int argc, char ** argv) {
                 const auto t_tg_start = ggml_time_us();
 
                 for (int i = 0; i < tg; ++i) {
-                    batch.n_tokens = pl;
+                    llama_batch_clear(batch);
 
                     for (int j = 0; j < pl; ++j) {
-                        batch.token[j]  = 0;
+                        llama_batch_add(batch, 0, pp + i, { j }, true);
-                        batch.pos[j]    = pp + i;
-                        batch.seq_id[j] = j;
-                        batch.logits[j] = true;
                     }
 
                     if (!decode_helper(ctx, batch, ctx_params.n_batch)) {

examples/batched.swift/Sources/main.swift

@@ -69,7 +69,7 @@ for id: llama_token in tokens {
 
 print("\n")
 
-var batch = llama_batch_init(max(Int32(tokens.count), Int32(n_parallel)), 0)
+var batch = llama_batch_init(max(Int32(tokens.count), Int32(n_parallel)), 0, 1)
 defer {
     llama_batch_free(batch)
 }
@@ -80,7 +80,12 @@ batch.n_tokens = Int32(tokens.count)
 for (i, token) in tokens.enumerated() {
     batch.token[i] = token
     batch.pos[i] = Int32(i)
-    batch.seq_id[i] = 0
+    batch.n_seq_id[i] = 1
+    // batch.seq_id[i][0] = 0
+    // TODO: is this the proper way to do this?
+    if let seq_id = batch.seq_id[i] {
+        seq_id[0] = 0
+    }
     batch.logits[i] = 0
 }
 
@@ -169,7 +174,10 @@ while n_cur <= n_len {
         // push this new token for next evaluation
         batch.token[Int(batch.n_tokens)] = new_token_id
         batch.pos[Int(batch.n_tokens)] = n_cur
-        batch.seq_id[Int(batch.n_tokens)] = Int32(i)
+        batch.n_seq_id[Int(batch.n_tokens)] = 1
+        if let seq_id = batch.seq_id[Int(batch.n_tokens)] {
+            seq_id[0] = Int32(i)
+        }
         batch.logits[Int(batch.n_tokens)] = 1
 
         i_batch[i] = batch.n_tokens
@@ -209,7 +217,7 @@ llama_print_timings(context)
 private func tokenize(text: String, add_bos: Bool) -> [llama_token] {
     let n_tokens = text.count + (add_bos ? 1 : 0)
     let tokens = UnsafeMutablePointer<llama_token>.allocate(capacity: n_tokens)
-    let tokenCount = llama_tokenize(model, text, Int32(text.count), tokens, Int32(n_tokens), add_bos)
+    let tokenCount = llama_tokenize(model, text, Int32(text.count), tokens, Int32(n_tokens), add_bos, /*special tokens*/ false)
     var swiftTokens: [llama_token] = []
     for i in 0 ..< tokenCount {
         swiftTokens.append(tokens[Int(i)])

examples/batched/batched.cpp

@@ -11,12 +11,19 @@ int main(int argc, char ** argv) {
     gpt_params params;
 
     if (argc == 1 || argv[1][0] == '-') {
-        printf("usage: %s MODEL_PATH [PROMPT] [PARALLEL]\n" , argv[0]);
+        printf("usage: %s MODEL_PATH [PROMPT] [PARALLEL] [LEN] [NGL]\n" , argv[0]);
         return 1 ;
     }
 
+    // number of parallel batches
     int n_parallel = 1;
 
+    // total length of the sequences including the prompt
+    int n_len = 32;
+
+    // number of layers to offload to the GPU
+    int n_gpu_layers = 0;
+
     if (argc >= 2) {
         params.model = argv[1];
     }
@@ -29,13 +36,18 @@ int main(int argc, char ** argv) {
         n_parallel = std::atoi(argv[3]);
     }
 
+    if (argc >= 5) {
+        n_len = std::atoi(argv[4]);
+    }
+
+    if (argc >= 6) {
+        n_gpu_layers = std::atoi(argv[5]);
+    }
+
     if (params.prompt.empty()) {
         params.prompt = "Hello my name is";
     }
 
-    // total length of the sequences including the prompt
-    const int n_len = 32;
-
     // init LLM
 
     llama_backend_init(params.numa);
@@ -44,7 +56,7 @@ int main(int argc, char ** argv) {
 
     llama_model_params model_params = llama_model_default_params();
 
-    // model_params.n_gpu_layers = 99; // offload all layers to the GPU
+    model_params.n_gpu_layers = n_gpu_layers;
 
     llama_model * model = llama_load_model_from_file(params.model.c_str(), model_params);
 
@@ -97,20 +109,15 @@ int main(int argc, char ** argv) {
 
     fflush(stderr);
 
-    // create a llama_batch with size 512
+    // create a llama_batch
     // we use this object to submit token data for decoding
-
+    llama_batch batch = llama_batch_init(std::max(tokens_list.size(), (size_t)n_parallel), 0, 1);
-    llama_batch batch = llama_batch_init(std::max(tokens_list.size(), (size_t)n_parallel), 0);
 
     // evaluate the initial prompt
-    batch.n_tokens = tokens_list.size();
+    for (size_t i = 0; i < tokens_list.size(); ++i) {
-
+        llama_batch_add(batch, tokens_list[i], i, { 0 }, false);
-    for (int32_t i = 0; i < batch.n_tokens; i++) {
-        batch.token[i]  = tokens_list[i];
-        batch.pos[i]    = i;
-        batch.seq_id[i] = 0;
-        batch.logits[i] = false;
     }
+    GGML_ASSERT(batch.n_tokens == (int) tokens_list.size());
 
     // llama_decode will output logits only for the last token of the prompt
     batch.logits[batch.n_tokens - 1] = true;
@@ -146,7 +153,7 @@ int main(int argc, char ** argv) {
 
     while (n_cur <= n_len) {
         // prepare the next batch
-        batch.n_tokens = 0;
+        llama_batch_clear(batch);
 
         // sample the next token for each parallel sequence / stream
         for (int32_t i = 0; i < n_parallel; ++i) {
@@ -180,7 +187,7 @@ int main(int argc, char ** argv) {
             //const llama_token new_token_id = llama_sample_token_greedy(ctx, &candidates_p);
 
             // is it an end of stream? -> mark the stream as finished
-            if (new_token_id == llama_token_eos(ctx) || n_cur == n_len) {
+            if (new_token_id == llama_token_eos(model) || n_cur == n_len) {
                 i_batch[i] = -1;
                 LOG_TEE("\n");
                 if (n_parallel > 1) {
@@ -198,15 +205,10 @@ int main(int argc, char ** argv) {
 
             streams[i] += llama_token_to_piece(ctx, new_token_id);
 
-            // push this new token for next evaluation
-            batch.token [batch.n_tokens] = new_token_id;
-            batch.pos   [batch.n_tokens] = n_cur;
-            batch.seq_id[batch.n_tokens] = i;
-            batch.logits[batch.n_tokens] = true;
-
             i_batch[i] = batch.n_tokens;
 
-            batch.n_tokens += 1;
+            // push this new token for next evaluation
+            llama_batch_add(batch, new_token_id, n_cur, { i }, true);
 
             n_decode += 1;
         }

examples/beam-search/beam-search.cpp

@@ -47,7 +47,7 @@ struct beam_search_callback_data {
 // In this case, end-of-beam (eob) is equivalent to end-of-sentence (eos) but this need not always be the same.
 // For example, eob can be flagged due to maximum token length, stop words, etc.
 static bool is_at_eob(const beam_search_callback_data & callback_data, const llama_token * tokens, size_t n_tokens) {
-    return n_tokens && tokens[n_tokens-1] == llama_token_eos(callback_data.ctx);
+    return n_tokens && tokens[n_tokens-1] == llama_token_eos(llama_get_model(callback_data.ctx));
 }
 
 // Function matching type llama_beam_search_callback_fn_t.

examples/benchmark/CMakeLists.txt

@@ -1,9 +1,6 @@
 set(TARGET benchmark)
 add_executable(${TARGET} benchmark-matmult.cpp)
 install(TARGETS ${TARGET} RUNTIME)
-target_link_libraries(${TARGET} PRIVATE llama ${CMAKE_THREAD_LIBS_INIT})
+target_link_libraries(${TARGET} PRIVATE llama build_info ${CMAKE_THREAD_LIBS_INIT})
 target_include_directories(${TARGET} PRIVATE ../../common)
 target_compile_features(${TARGET} PRIVATE cxx_std_11)
-if(TARGET BUILD_INFO)
-  add_dependencies(${TARGET} BUILD_INFO)
-endif()

examples/benchmark/benchmark-matmult.cpp

@@ -1,4 +1,3 @@
-#include "build-info.h"
 #include "common.h"
 #include "ggml.h"
 

examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp

@@ -536,7 +536,7 @@ static bool is_ggml_file(const char * filename) {
     if (file.size < 4) {
         return false;
     }
-    uint32_t magic = file.read_u32();
+    std::string magic = file.read_string(4);
     return magic == GGUF_MAGIC;
 }
 

examples/embd-input/.gitignore

@@ -1,4 +0,0 @@
-PandaGPT
-MiniGPT-4
-*.pth
-

examples/embd-input/CMakeLists.txt

@@ -1,17 +0,0 @@
-set(TARGET embdinput)
-add_library(${TARGET} embd-input-lib.cpp embd-input.h)
-install(TARGETS ${TARGET} LIBRARY)
-target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
-target_compile_features(${TARGET} PRIVATE cxx_std_11)
-if(TARGET BUILD_INFO)
-  add_dependencies(${TARGET} BUILD_INFO)
-endif()
-
-set(TARGET embd-input-test)
-add_executable(${TARGET} embd-input-test.cpp)
-install(TARGETS ${TARGET} RUNTIME)
-target_link_libraries(${TARGET} PRIVATE common llama embdinput ${CMAKE_THREAD_LIBS_INIT})
-target_compile_features(${TARGET} PRIVATE cxx_std_11)
-if(TARGET BUILD_INFO)
-  add_dependencies(${TARGET} BUILD_INFO)
-endif()

examples/embd-input/README.md

@@ -1,63 +0,0 @@
-### Examples for input embedding directly
-
-## Requirement
-build  `libembdinput.so`
-run the following comman in main dir (../../).
-```
-make
-```
-
-## [LLaVA](https://github.com/haotian-liu/LLaVA/) example  (llava.py)
-
-1. Obtian LLaVA model (following https://github.com/haotian-liu/LLaVA/ , use https://huggingface.co/liuhaotian/LLaVA-13b-delta-v1-1/).
-2. Convert it to ggml format.
-3. `llava_projection.pth` is [pytorch_model-00003-of-00003.bin](https://huggingface.co/liuhaotian/LLaVA-13b-delta-v1-1/blob/main/pytorch_model-00003-of-00003.bin).
-
-```
-import torch
-
-bin_path = "../LLaVA-13b-delta-v1-1/pytorch_model-00003-of-00003.bin"
-pth_path = "./examples/embd-input/llava_projection.pth"
-
-dic = torch.load(bin_path)
-used_key = ["model.mm_projector.weight","model.mm_projector.bias"]
-torch.save({k: dic[k] for k in used_key}, pth_path)
-```
-4. Check the path of LLaVA model and `llava_projection.pth` in `llava.py`.
-
-
-## [PandaGPT](https://github.com/yxuansu/PandaGPT) example (panda_gpt.py)
-
-1. Obtian PandaGPT lora model from https://github.com/yxuansu/PandaGPT. Rename the file to `adapter_model.bin`. Use [convert-lora-to-ggml.py](../../convert-lora-to-ggml.py) to convert it to ggml format.
-The `adapter_config.json` is
-```
-{
-  "peft_type": "LORA",
-  "fan_in_fan_out": false,
-  "bias": null,
-  "modules_to_save": null,
-  "r": 32,
-  "lora_alpha": 32,
-  "lora_dropout": 0.1,
-  "target_modules": ["q_proj", "k_proj", "v_proj", "o_proj"]
-}
-```
-2. Papare the `vicuna` v0 model.
-3. Obtain the [ImageBind](https://dl.fbaipublicfiles.com/imagebind/imagebind_huge.pth) model.
-4. Clone the PandaGPT source.
-```
-git clone https://github.com/yxuansu/PandaGPT
-```
-5. Install the requirement of PandaGPT.
-6. Check the path of PandaGPT source, ImageBind model, lora model and vicuna model in panda_gpt.py.
-
-## [MiniGPT-4](https://github.com/Vision-CAIR/MiniGPT-4/) example (minigpt4.py)
-
-1. Obtain MiniGPT-4 model from https://github.com/Vision-CAIR/MiniGPT-4/ and put it in `embd-input`.
-2. Clone the MiniGPT-4 source.
-```
-git clone https://github.com/Vision-CAIR/MiniGPT-4/
-```
-3. Install the requirement of PandaGPT.
-4. Papare the `vicuna` v0 model.
-5. Check the path of MiniGPT-4 source, MiniGPT-4 model and vicuna model in `minigpt4.py`.

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

@@ -1,221 +0,0 @@
-#include "build-info.h"
-#include "common.h"
-#include "embd-input.h"
-
-#include <cassert>
-#include <cinttypes>
-#include <cmath>
-#include <cstdio>
-#include <cstring>
-#include <ctime>
-#include <fstream>
-#include <iostream>
-#include <string>
-#include <vector>
-
-static llama_context ** g_ctx;
-
-extern "C" {
-
-struct MyModel* create_mymodel(int argc, char ** argv) {
-    gpt_params params;
-
-    if (!gpt_params_parse(argc, argv, params)) {
-        return nullptr;
-    }
-
-    print_build_info();
-
-    if (params.seed == LLAMA_DEFAULT_SEED) {
-        params.seed = uint32_t(time(NULL));
-    }
-    fprintf(stderr, "%s: seed  = %d\n", __func__, params.seed);
-
-    llama_backend_init(params.numa);
-
-    llama_model * model;
-    llama_context * ctx;
-
-    g_ctx = &ctx;
-
-    // load the model and apply lora adapter, if any
-    std::tie(model, ctx) = llama_init_from_gpt_params(params);
-    if (model == NULL) {
-        fprintf(stderr, "%s: error: unable to load model\n", __func__);
-        return nullptr;
-    }
-
-    // print system information
-    {
-        fprintf(stderr, "\n");
-        fprintf(stderr, "%s\n", get_system_info(params).c_str());
-    }
-    struct MyModel * ret = new MyModel();
-    ret->ctx = ctx;
-    ret->params = params;
-    ret->n_past = 0;
-    // printf("ctx: %d\n", ret->ctx);
-    return ret;
-}
-
-void free_mymodel(struct MyModel * mymodel) {
-    llama_context * ctx = mymodel->ctx;
-    llama_print_timings(ctx);
-    llama_free(ctx);
-    delete mymodel;
-}
-
-
-bool eval_float(void * model, float * input, int N){
-    MyModel * mymodel = (MyModel*)model;
-    llama_context * ctx = mymodel->ctx;
-    gpt_params params = mymodel->params;
-    int n_emb = llama_n_embd(llama_get_model(ctx));
-    int n_past = mymodel->n_past;
-    int n_batch = N; // params.n_batch;
-
-    for (int i = 0; i < (int) N; i += n_batch) {
-        int n_eval = (int) N - i;
-        if (n_eval > n_batch) {
-            n_eval = n_batch;
-        }
-        llama_batch batch = {  int32_t(n_eval), nullptr, (input+i*n_emb), nullptr, nullptr, nullptr, n_past, 1, 0, };
-        if (llama_decode(ctx, batch)) {
-            fprintf(stderr, "%s : failed to eval\n", __func__);
-            return false;
-        }
-        n_past += n_eval;
-    }
-    mymodel->n_past = n_past;
-    return true;
-}
-
-bool eval_tokens(void * model, std::vector<llama_token> tokens) {
-    MyModel * mymodel = (MyModel* )model;
-    llama_context * ctx;
-    ctx = mymodel->ctx;
-    gpt_params params = mymodel->params;
-    int n_past = mymodel->n_past;
-    for (int i = 0; i < (int) tokens.size(); i += params.n_batch) {
-        int n_eval = (int) tokens.size() - i;
-        if (n_eval > params.n_batch) {
-            n_eval = params.n_batch;
-        }
-        if (llama_decode(ctx, llama_batch_get_one(&tokens[i], n_eval, n_past, 0))) {
-            fprintf(stderr, "%s : failed to eval\n", __func__);
-            return false;
-        }
-        n_past += n_eval;
-    }
-    mymodel->n_past = n_past;
-    return true;
-}
-
-bool eval_id(struct MyModel* mymodel, int id) {
-    std::vector<llama_token> tokens;
-    tokens.push_back(id);
-    return eval_tokens(mymodel, tokens);
-}
-
-bool eval_string(struct MyModel * mymodel,const char* str){
-    llama_context * ctx = mymodel->ctx;
-    std::string str2 = str;
-    std::vector<llama_token> embd_inp = ::llama_tokenize(ctx, str2, true);
-    eval_tokens(mymodel, embd_inp);
-    return true;
-}
-
-llama_token sampling_id(struct MyModel* mymodel) {
-    llama_context* ctx = mymodel->ctx;
-    gpt_params params = mymodel->params;
-    llama_sampling_params & sparams = params.sampling_params;
-    // int n_ctx = llama_n_ctx(ctx);
-
-    // out of user input, sample next token
-    const float   temp            = sparams.temp;
-    const int32_t top_k           = sparams.top_k <= 0 ? llama_n_vocab(llama_get_model(ctx)) : sparams.top_k;
-    const float   top_p           = sparams.top_p;
-    const float   tfs_z           = sparams.tfs_z;
-    const float   typical_p       = sparams.typical_p;
-    // const int32_t repeat_last_n   = params.repeat_last_n < 0 ? n_ctx : params.repeat_last_n;
-    // const float   repeat_penalty  = params.repeat_penalty;
-    // const float   alpha_presence  = params.presence_penalty;
-    // const float   alpha_frequency = params.frequency_penalty;
-    const int     mirostat        = sparams.mirostat;
-    const float   mirostat_tau    = sparams.mirostat_tau;
-    const float   mirostat_eta    = sparams.mirostat_eta;
-    // const bool    penalize_nl     = params.penalize_nl;
-
-    llama_token id = 0;
-    {
-        auto logits  = llama_get_logits(ctx);
-        auto n_vocab = llama_n_vocab(llama_get_model(ctx));
-
-        // Apply params.logit_bias map
-        for (auto it = sparams.logit_bias.begin(); it != sparams.logit_bias.end(); it++) {
-            logits[it->first] += it->second;
-        }
-
-        std::vector<llama_token_data> candidates;
-        candidates.reserve(n_vocab);
-        for (llama_token token_id = 0; token_id < n_vocab; token_id++) {
-            candidates.emplace_back(llama_token_data{token_id, logits[token_id], 0.0f});
-        }
-
-        llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false };
-
-        // TODO: Apply penalties
-        // float nl_logit = logits[llama_token_nl(ctx)];
-        // auto last_n_repeat = std::min(std::min((int)last_n_tokens.size(), repeat_last_n), n_ctx);
-        // llama_sample_repetition_penalty(ctx, &candidates_p,
-        //      last_n_tokens.data() + last_n_tokens.size() - last_n_repeat,
-        //      last_n_repeat, repeat_penalty);
-        // llama_sample_frequency_and_presence_penalties(ctx, &candidates_p,
-        // last_n_tokens.data() + last_n_tokens.size() - last_n_repeat,
-        // last_n_repeat, alpha_frequency, alpha_presence);
-        // if (!penalize_nl) {
-        //     logits[llama_token_nl(ctx)] = nl_logit;
-        // }
-
-        if (temp <= 0) {
-            // Greedy sampling
-            id = llama_sample_token_greedy(ctx, &candidates_p);
-        } else {
-            if (mirostat == 1) {
-                static float mirostat_mu = 2.0f * mirostat_tau;
-                const int mirostat_m = 100;
-                llama_sample_temp(ctx, &candidates_p, temp);
-                id = llama_sample_token_mirostat(ctx, &candidates_p, mirostat_tau, mirostat_eta, mirostat_m, &mirostat_mu);
-            } else if (mirostat == 2) {
-                static float mirostat_mu = 2.0f * mirostat_tau;
-                llama_sample_temp(ctx, &candidates_p, temp);
-                id = llama_sample_token_mirostat_v2(ctx, &candidates_p, mirostat_tau, mirostat_eta, &mirostat_mu);
-            } else {
-                // Temperature sampling
-                llama_sample_top_k(ctx, &candidates_p, top_k, 1);
-                llama_sample_tail_free(ctx, &candidates_p, tfs_z, 1);
-                llama_sample_typical(ctx, &candidates_p, typical_p, 1);
-                llama_sample_top_p(ctx, &candidates_p, top_p, 1);
-                llama_sample_temp(ctx, &candidates_p, temp);
-                id = llama_sample_token(ctx, &candidates_p);
-            }
-        }
-    }
-
-    return id;
-}
-
-const char * sampling(struct MyModel * mymodel) {
-    llama_context * ctx = mymodel->ctx;
-    int id = sampling_id(mymodel);
-    static std::string ret;
-    if (id == llama_token_eos(ctx)) {
-        ret = "</s>";
-    } else {
-        ret = llama_token_to_piece(ctx, id);
-    }
-    eval_id(mymodel, id);
-    return ret.c_str();
-}
-
-}

examples/embd-input/embd-input-test.cpp

@@ -1,35 +0,0 @@
-#include "embd-input.h"
-#include <stdlib.h>
-#include <random>
-#include <string.h>
-
-int main(int argc, char** argv) {
-
-    auto mymodel = create_mymodel(argc, argv);
-    int N = 10;
-    int max_tgt_len = 500;
-    int n_embd = llama_n_embd(llama_get_model(mymodel->ctx));
-
-    // add random float embd to test evaluation
-    float * data = new float[N*n_embd];
-    std::default_random_engine e;
-    std::uniform_real_distribution<float>  u(0,1);
-    for (int i=0;i<N*n_embd;i++) {
-        data[i] = u(e);
-    }
-
-    eval_string(mymodel, "user: what is the color of the flag of UN?");
-    eval_float(mymodel, data, N);
-    eval_string(mymodel, "assistant:");
-    eval_string(mymodel, mymodel->params.prompt.c_str());
-    const char* tmp;
-    for (int i=0; i<max_tgt_len; i++) {
-        tmp = sampling(mymodel);
-        if (strcmp(tmp, "</s>")==0) break;
-        printf("%s", tmp);
-        fflush(stdout);
-    }
-    printf("\n");
-    free_mymodel(mymodel);
-    return 0;
-}

examples/embd-input/embd-input.h

@@ -1,27 +0,0 @@
-#ifndef _EMBD_INPUT_H_
-#define _EMBD_INPUT_H_ 1
-
-#include "common.h"
-#include "llama.h"
-
-extern "C" {
-
-typedef struct MyModel {
-    llama_context* ctx;
-    gpt_params params;
-    int n_past = 0;
-} MyModel;
-
-struct MyModel* create_mymodel(int argc, char ** argv);
-
-bool eval_float(void* model, float* input, int N);
-bool eval_tokens(void* model, std::vector<llama_token> tokens);
-bool eval_id(struct MyModel* mymodel, int id);
-bool eval_string(struct MyModel* mymodel, const char* str);
-const char * sampling(struct MyModel* mymodel);
-llama_token sampling_id(struct MyModel* mymodel);
-void free_mymodel(struct MyModel* mymodel);
-
-}
-
-#endif

examples/embd-input/embd_input.py

@@ -1,72 +0,0 @@
-#!/usr/bin/env python3
-import ctypes
-from ctypes import cdll, c_char_p, c_void_p, POINTER, c_float, c_int
-import numpy as np
-import os
-
-libc = cdll.LoadLibrary("./libembdinput.so")
-libc.sampling.restype=c_char_p
-libc.create_mymodel.restype=c_void_p
-libc.eval_string.argtypes=[c_void_p, c_char_p]
-libc.sampling.argtypes=[c_void_p]
-libc.eval_float.argtypes=[c_void_p, POINTER(c_float), c_int]
-
-
-class MyModel:
-    def __init__(self, args):
-        argc = len(args)
-        c_str = [c_char_p(i.encode()) for i in args]
-        args_c = (c_char_p * argc)(*c_str)
-        self.model = c_void_p(libc.create_mymodel(argc, args_c))
-        self.max_tgt_len = 512
-        self.print_string_eval = True
-
-    def __del__(self):
-        libc.free_mymodel(self.model)
-
-    def eval_float(self, x):
-        libc.eval_float(self.model, x.astype(np.float32).ctypes.data_as(POINTER(c_float)), x.shape[1])
-
-    def eval_string(self, x):
-        libc.eval_string(self.model, x.encode()) # c_char_p(x.encode()))
-        if self.print_string_eval:
-            print(x)
-
-    def eval_token(self, x):
-        libc.eval_id(self.model, x)
-
-    def sampling(self):
-        s = libc.sampling(self.model)
-        return s
-
-    def stream_generate(self, end="</s>"):
-        ret = b""
-        end = end.encode()
-        for _ in range(self.max_tgt_len):
-            tmp = self.sampling()
-            ret += tmp
-            yield tmp
-            if ret.endswith(end):
-                break
-
-    def generate_with_print(self, end="</s>"):
-        ret = b""
-        for i in self.stream_generate(end=end):
-            ret += i
-            print(i.decode(errors="replace"), end="", flush=True)
-        print("")
-        return ret.decode(errors="replace")
-
-
-    def generate(self, end="</s>"):
-        text = b"".join(self.stream_generate(end=end))
-        return text.decode(errors="replace")
-
-if __name__ == "__main__":
-    model = MyModel(["main", "--model", "../llama.cpp/models/ggml-vic13b-q4_1.bin", "-c", "2048"])
-    model.eval_string("""user: what is the color of the flag of UN?""")
-    x = np.random.random((5120,10))# , dtype=np.float32)
-    model.eval_float(x)
-    model.eval_string("""assistant:""")
-    for i in model.generate():
-        print(i.decode(errors="replace"), end="", flush=True)

examples/embd-input/llava.py

@@ -1,71 +0,0 @@
-#!/usr/bin/env python3
-import sys
-import os
-sys.path.insert(0, os.path.dirname(__file__))
-from embd_input import MyModel
-import numpy as np
-from torch import nn
-import torch
-from transformers import CLIPVisionModel,  CLIPImageProcessor
-from PIL import Image
-
-# model parameters from 'liuhaotian/LLaVA-13b-delta-v1-1'
-vision_tower = "openai/clip-vit-large-patch14"
-select_hidden_state_layer = -2
-# (vision_config.image_size // vision_config.patch_size) ** 2
-image_token_len = (224//14)**2
-
-class Llava:
-    def __init__(self, args):
-        self.image_processor = CLIPImageProcessor.from_pretrained(vision_tower)
-        self.vision_tower = CLIPVisionModel.from_pretrained(vision_tower)
-        self.mm_projector = nn.Linear(1024, 5120)
-        self.model = MyModel(["main", *args])
-
-    def load_projection(self, path):
-        state = torch.load(path)
-        self.mm_projector.load_state_dict({
-            "weight": state["model.mm_projector.weight"],
-            "bias": state["model.mm_projector.bias"]})
-
-    def chat(self, question):
-        self.model.eval_string("user: ")
-        self.model.eval_string(question)
-        self.model.eval_string("\nassistant: ")
-        return self.model.generate_with_print()
-
-    def chat_with_image(self, image, question):
-        with torch.no_grad():
-            embd_image = self.image_processor.preprocess(image, return_tensors='pt')['pixel_values'][0]
-            image_forward_out = self.vision_tower(embd_image.unsqueeze(0), output_hidden_states=True)
-            select_hidden_state = image_forward_out.hidden_states[select_hidden_state_layer]
-            image_feature = select_hidden_state[:, 1:]
-            embd_image = self.mm_projector(image_feature)
-            embd_image = embd_image.cpu().numpy()[0]
-        self.model.eval_string("user: ")
-        self.model.eval_token(32003-2) # im_start
-        self.model.eval_float(embd_image.T)
-        for i in range(image_token_len-embd_image.shape[0]):
-            self.model.eval_token(32003-3) # im_patch
-        self.model.eval_token(32003-1) # im_end
-        self.model.eval_string(question)
-        self.model.eval_string("\nassistant: ")
-        return self.model.generate_with_print()
-
-
-if __name__=="__main__":
-    # model form liuhaotian/LLaVA-13b-delta-v1-1
-    a = Llava(["--model", "./models/ggml-llava-13b-v1.1.bin", "-c", "2048"])
-    # Extract from https://huggingface.co/liuhaotian/LLaVA-13b-delta-v1-1/blob/main/pytorch_model-00003-of-00003.bin.
-    # Also here can use pytorch_model-00003-of-00003.bin directly.
-    a.load_projection(os.path.join(
-        os.path.dirname(__file__) ,
-        "llava_projection.pth"))
-    respose = a.chat_with_image(
-        Image.open("./media/llama1-logo.png").convert('RGB'),
-        "what is the text in the picture?")
-    respose
-    a.chat("what is the color of it?")
-
-
-

examples/embd-input/minigpt4.py

@@ -1,129 +0,0 @@
-#!/usr/bin/env python3
-import sys
-import os
-sys.path.insert(0, os.path.dirname(__file__))
-from embd_input import MyModel
-import numpy as np
-from torch import nn
-import torch
-from PIL import Image
-
-minigpt4_path = os.path.join(os.path.dirname(__file__), "MiniGPT-4")
-sys.path.insert(0, minigpt4_path)
-from minigpt4.models.blip2 import Blip2Base
-from minigpt4.processors.blip_processors import Blip2ImageEvalProcessor
-
-
-class MiniGPT4(Blip2Base):
-    """
-    MiniGPT4 model from https://github.com/Vision-CAIR/MiniGPT-4
-    """
-    def __init__(self,
-        args,
-        vit_model="eva_clip_g",
-        q_former_model="https://storage.googleapis.com/sfr-vision-language-research/LAVIS/models/BLIP2/blip2_pretrained_flant5xxl.pth",
-        img_size=224,
-        drop_path_rate=0,
-        use_grad_checkpoint=False,
-        vit_precision="fp32",
-        freeze_vit=True,
-        freeze_qformer=True,
-        num_query_token=32,
-        llama_model="",
-        prompt_path="",
-        prompt_template="",
-        max_txt_len=32,
-        end_sym='\n',
-        low_resource=False,  # use 8 bit and put vit in cpu
-        device_8bit=0
-    ):
-        super().__init__()
-        self.img_size = img_size
-        self.low_resource = low_resource
-        self.preprocessor = Blip2ImageEvalProcessor(img_size)
-
-        print('Loading VIT')
-        self.visual_encoder, self.ln_vision = self.init_vision_encoder(
-            vit_model, img_size, drop_path_rate, use_grad_checkpoint, vit_precision
-        )
-        print('Loading VIT Done')
-        print('Loading Q-Former')
-        self.Qformer, self.query_tokens = self.init_Qformer(
-            num_query_token, self.visual_encoder.num_features
-        )
-        self.Qformer.cls = None
-        self.Qformer.bert.embeddings.word_embeddings = None
-        self.Qformer.bert.embeddings.position_embeddings = None
-        for layer in self.Qformer.bert.encoder.layer:
-            layer.output = None
-            layer.intermediate = None
-        self.load_from_pretrained(url_or_filename=q_former_model)
-        print('Loading Q-Former Done')
-        self.llama_proj = nn.Linear(
-            self.Qformer.config.hidden_size, 5120 # self.llama_model.config.hidden_size
-        )
-        self.max_txt_len = max_txt_len
-        self.end_sym = end_sym
-        self.model = MyModel(["main", *args])
-        # system prompt
-        self.model.eval_string("Give the following image: <Img>ImageContent</Img>. "
-           "You will be able to see the image once I provide it to you. Please answer my questions."
-           "###")
-
-    def encode_img(self, image):
-        image = self.preprocessor(image)
-        image = image.unsqueeze(0)
-        device = image.device
-        if self.low_resource:
-            self.vit_to_cpu()
-            image = image.to("cpu")
-
-        with self.maybe_autocast():
-            image_embeds = self.ln_vision(self.visual_encoder(image)).to(device)
-            image_atts = torch.ones(image_embeds.size()[:-1], dtype=torch.long).to(device)
-
-            query_tokens = self.query_tokens.expand(image_embeds.shape[0], -1, -1)
-            query_output = self.Qformer.bert(
-                query_embeds=query_tokens,
-                encoder_hidden_states=image_embeds,
-                encoder_attention_mask=image_atts,
-                return_dict=True,
-            )
-
-            inputs_llama = self.llama_proj(query_output.last_hidden_state)
-            # atts_llama = torch.ones(inputs_llama.size()[:-1], dtype=torch.long).to(image.device)
-        return inputs_llama
-
-    def load_projection(self, path):
-        state = torch.load(path)["model"]
-        self.llama_proj.load_state_dict({
-            "weight": state["llama_proj.weight"],
-            "bias": state["llama_proj.bias"]})
-
-    def chat(self, question):
-        self.model.eval_string("Human: ")
-        self.model.eval_string(question)
-        self.model.eval_string("\n### Assistant:")
-        return self.model.generate_with_print(end="###")
-
-    def chat_with_image(self, image, question):
-        with torch.no_grad():
-            embd_image = self.encode_img(image)
-        embd_image = embd_image.cpu().numpy()[0]
-        self.model.eval_string("Human: <Img>")
-        self.model.eval_float(embd_image.T)
-        self.model.eval_string("</Img> ")
-        self.model.eval_string(question)
-        self.model.eval_string("\n### Assistant:")
-        return self.model.generate_with_print(end="###")
-
-
-if __name__=="__main__":
-    a = MiniGPT4(["--model", "./models/ggml-vicuna-13b-v0-q4_1.bin", "-c", "2048"])
-    a.load_projection(os.path.join(
-        os.path.dirname(__file__) ,
-        "pretrained_minigpt4.pth"))
-    respose = a.chat_with_image(
-        Image.open("./media/llama1-logo.png").convert('RGB'),
-        "what is the text in the picture?")
-    a.chat("what is the color of it?")

examples/embd-input/panda_gpt.py

@@ -1,99 +0,0 @@
-#!/usr/bin/env python3
-import sys
-import os
-sys.path.insert(0, os.path.dirname(__file__))
-from embd_input import MyModel
-import numpy as np
-from torch import nn
-import torch
-
-# use PandaGPT path
-panda_gpt_path = os.path.join(os.path.dirname(__file__), "PandaGPT")
-imagebind_ckpt_path = "./models/panda_gpt/"
-
-sys.path.insert(0, os.path.join(panda_gpt_path,"code","model"))
-from ImageBind.models import imagebind_model
-from ImageBind import data
-
-ModalityType = imagebind_model.ModalityType
-max_tgt_len = 400
-
-class PandaGPT:
-    def __init__(self, args):
-        self.visual_encoder,_ = imagebind_model.imagebind_huge(pretrained=True, store_path=imagebind_ckpt_path)
-        self.visual_encoder.eval()
-        self.llama_proj = nn.Linear(1024, 5120) # self.visual_hidden_size, 5120)
-        self.max_tgt_len = max_tgt_len
-        self.model = MyModel(["main", *args])
-        self.generated_text = ""
-        self.device = "cpu"
-
-    def load_projection(self, path):
-        state = torch.load(path, map_location="cpu")
-        self.llama_proj.load_state_dict({
-            "weight": state["llama_proj.weight"],
-            "bias": state["llama_proj.bias"]})
-
-    def eval_inputs(self, inputs):
-        self.model.eval_string("<Img>")
-        embds = self.extract_multimoal_feature(inputs)
-        for i in embds:
-            self.model.eval_float(i.T)
-        self.model.eval_string("</Img> ")
-
-    def chat(self, question):
-        return self.chat_with_image(None, question)
-
-    def chat_with_image(self, inputs, question):
-        if self.generated_text == "":
-            self.model.eval_string("###")
-        self.model.eval_string(" Human: ")
-        if inputs:
-            self.eval_inputs(inputs)
-        self.model.eval_string(question)
-        self.model.eval_string("\n### Assistant:")
-        ret = self.model.generate_with_print(end="###")
-        self.generated_text += ret
-        return ret
-
-    def extract_multimoal_feature(self, inputs):
-        features = []
-        for key in ["image", "audio", "video", "thermal"]:
-            if key + "_paths" in inputs:
-                embeds = self.encode_data(key, inputs[key+"_paths"])
-                features.append(embeds)
-        return features
-
-    def encode_data(self, data_type, data_paths):
-
-        type_map = {
-            "image": ModalityType.VISION,
-            "audio": ModalityType.AUDIO,
-            "video": ModalityType.VISION,
-            "thermal": ModalityType.THERMAL,
-        }
-        load_map = {
-            "image": data.load_and_transform_vision_data,
-            "audio": data.load_and_transform_audio_data,
-            "video": data.load_and_transform_video_data,
-            "thermal": data.load_and_transform_thermal_data
-        }
-
-        load_function = load_map[data_type]
-        key = type_map[data_type]
-
-        inputs = {key: load_function(data_paths, self.device)}
-        with torch.no_grad():
-            embeddings = self.visual_encoder(inputs)
-            embeds = embeddings[key]
-            embeds = self.llama_proj(embeds).cpu().numpy()
-        return embeds
-
-
-if __name__=="__main__":
-    a = PandaGPT(["--model", "./models/ggml-vicuna-13b-v0-q4_1.bin", "-c", "2048", "--lora", "./models/panda_gpt/ggml-adapter-model.bin","--temp", "0"])
-    a.load_projection("./models/panda_gpt/adapter_model.bin")
-    a.chat_with_image(
-        {"image_paths": ["./media/llama1-logo.png"]},
-        "what is the text in the picture? 'llama' or 'lambda'?")
-    a.chat("what is the color of it?")

examples/embedding/CMakeLists.txt

@@ -3,6 +3,3 @@ add_executable(${TARGET} embedding.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_11)
-if(TARGET BUILD_INFO)
-  add_dependencies(${TARGET} BUILD_INFO)
-endif()

examples/embedding/embedding.cpp

@@ -1,4 +1,3 @@
-#include "build-info.h"
 #include "common.h"
 #include "llama.h"
 

examples/finetune/finetune.cpp

@@ -529,13 +529,14 @@ static void init_lora(const struct my_llama_model * model, struct my_llama_lora
     set_param_lora(lora);
 
     // measure data size
-    struct ggml_allocr * alloc = NULL;
+    size_t size = 0;
-    alloc = ggml_allocr_new_measure(tensor_alignment);
+    for (struct ggml_tensor * t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) {
-    alloc_lora(alloc, lora);
+        size += GGML_PAD(ggml_nbytes(t), tensor_alignment);
+    }
 
     // allocate data
-    lora->data.resize(ggml_allocr_max_size(alloc) + tensor_alignment);
+    struct ggml_allocr * alloc = NULL;
-    ggml_allocr_free(alloc);
+    lora->data.resize(size + tensor_alignment);
     alloc = ggml_allocr_new(lora->data.data(), lora->data.size(), tensor_alignment);
     alloc_lora(alloc, lora);
     ggml_allocr_free(alloc);
@@ -641,8 +642,9 @@ static struct ggml_tensor * llama_build_lora_finetune_graphs(
         const int rope_mode = 0;
 
         return ggml_rope_custom(ctx,
-            t, KQ_pos, n_rot, rope_mode, n_ctx,
+            t, KQ_pos, n_rot, rope_mode, n_ctx, 0,
-            rope_freq_base, rope_freq_scale);
+            rope_freq_base, rope_freq_scale, 0.0f, 0.0f, 0.0f, 0.0f
+        );
     };
 
     set_name(tokens_input, "tokens_input");
@@ -651,7 +653,7 @@ static struct ggml_tensor * llama_build_lora_finetune_graphs(
     GGML_ASSERT(tokens_input->type == GGML_TYPE_I32);
 
     auto add_to_f32 = [] (struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * b) {
-        if (ggml_is_quantized(a->type)) {
+        if (ggml_is_quantized(a->type) || a->type == GGML_TYPE_F16) {
             return ggml_add_cast(ctx, a, b, GGML_TYPE_F32);
         } else if (a->type == GGML_TYPE_F32) {
             return ggml_add(ctx, a, b);
@@ -1458,6 +1460,17 @@ static bool train_params_parse(int argc, char ** argv, struct train_params * par
             }
             params->n_rank_w3 = std::stoi(argv[i]);
             params->custom_n_rank_w3 = true;
+        } else if (arg == "--gpu-layers" || arg == "-ngl" || arg == "--n-gpu-layers") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+#ifdef LLAMA_SUPPORTS_GPU_OFFLOAD
+            params->common.n_gpu_layers = std::stoi(argv[i]);
+#else
+            fprintf(stderr, "warning: not compiled with GPU offload support, --n-gpu-layers option will be ignored\n");
+            fprintf(stderr, "warning: see main README.md for information on enabling GPU BLAS support\n");
+#endif
         } else {
             fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
             train_print_usage(argc, argv, &default_params);
@@ -1544,6 +1557,7 @@ int main(int argc, char ** argv) {
     srand(params.common.seed);
 
     struct llama_model_params llama_mparams = llama_model_default_params();
+    llama_mparams.n_gpu_layers = params.common.n_gpu_layers;
     llama_mparams.vocab_only = false;
 
     printf("%s: model base = '%s'\n", __func__, params.fn_model_base);
@@ -1714,11 +1728,9 @@ int main(int argc, char ** argv) {
     struct ggml_tensor * target_probs  = ggml_new_tensor_3d(ctx_input, GGML_TYPE_F32, n_vocab,  n_tokens, n_batch);
 
     // measure required memory for input tensors
-    alloc = ggml_allocr_new_measure(tensor_alignment);
+    size_t max_input_size = GGML_PAD(ggml_nbytes(tokens_input), tensor_alignment) +
-    ggml_allocr_alloc(alloc, tokens_input);
+                            GGML_PAD(ggml_nbytes(target_probs), tensor_alignment) +
-    ggml_allocr_alloc(alloc, target_probs);
+                            tensor_alignment;
-    size_t max_input_size = ggml_allocr_max_size(alloc) + tensor_alignment;
-    ggml_allocr_free(alloc);
     printf("%s: input_size = %zu bytes (%.1f MB)\n", __func__, max_input_size, (float) max_input_size / (1024.0f*1024.0f));
 
     // allocate input tensors

examples/finetune/finetune.sh

@@ -0,0 +1,34 @@
+#!/bin/bash
+cd `dirname $0`
+cd ../..
+
+EXE="./finetune"
+
+if [[ ! $LLAMA_MODEL_DIR ]]; then LLAMA_MODEL_DIR="./models"; fi
+if [[ ! $LLAMA_TRAINING_DIR ]]; then LLAMA_TRAINING_DIR="."; fi
+
+# MODEL="$LLAMA_MODEL_DIR/openllama-3b-v2-q8_0.gguf" # This is the model the readme uses.
+MODEL="$LLAMA_MODEL_DIR/openllama-3b-v2.gguf" # An f16 model. Note in this case with "-g", you get an f32-format .BIN file that isn't yet supported if you use it with "main --lora" with GPU inferencing.
+
+while getopts "dg" opt; do
+  case $opt in
+    d)
+      DEBUGGER="gdb --args"
+      ;;
+    g)
+      EXE="./build/bin/Release/finetune"
+      GPUARG="--gpu-layers 25"
+      ;;
+  esac
+done
+
+$DEBUGGER $EXE \
+        --model-base $MODEL \
+        $GPUARG \
+        --checkpoint-in  chk-ol3b-shakespeare-LATEST.gguf \
+        --checkpoint-out chk-ol3b-shakespeare-ITERATION.gguf \
+        --lora-out lora-ol3b-shakespeare-ITERATION.bin \
+        --train-data "$LLAMA_TRAINING_DIR\shakespeare.txt" \
+        --save-every 10 \
+        --threads 10 --adam-iter 30 --batch 4 --ctx 64 \
+        --use-checkpointing

examples/infill/CMakeLists.txt

@@ -3,6 +3,3 @@ add_executable(${TARGET} infill.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_11)
-if(TARGET BUILD_INFO)
-  add_dependencies(${TARGET} BUILD_INFO)
-endif()

examples/infill/infill.cpp

@@ -2,7 +2,6 @@
 
 #include "console.h"
 #include "llama.h"
-#include "build-info.h"
 #include "grammar-parser.h"
 
 #include <cassert>
@@ -39,8 +38,8 @@ static gpt_params               * g_params;
 static std::vector<llama_token> * g_input_tokens;
 static std::ostringstream       * g_output_ss;
 static std::vector<llama_token> * g_output_tokens;
-static bool is_interacting = false;
 
+static bool is_interacting = false;
 
 static void write_logfile(
     const llama_context * ctx, const gpt_params & params, const llama_model * model,
@@ -104,7 +103,7 @@ static void sigint_handler(int signo) {
 
 int main(int argc, char ** argv) {
     gpt_params params;
-    llama_sampling_params & sparams = params.sampling_params;
+    llama_sampling_params & sparams = params.sparams;
     g_params = &params;
 
     if (!gpt_params_parse(argc, argv, params)) {
@@ -184,8 +183,8 @@ int main(int argc, char ** argv) {
         LOG_TEE("%s: warning: scaling RoPE frequency by %g.\n", __func__, params.rope_freq_scale);
     }
 
-    LOG_TEE("%s: build = %d (%s)\n", __func__, BUILD_NUMBER, BUILD_COMMIT);
+    LOG_TEE("%s: build = %d (%s)\n",      __func__, LLAMA_BUILD_NUMBER, LLAMA_COMMIT);
-    LOG_TEE("%s: built with %s for %s\n", __func__, BUILD_COMPILER, BUILD_TARGET);
+    LOG_TEE("%s: built with %s for %s\n", __func__, LLAMA_COMPILER, LLAMA_BUILD_TARGET);
 
     if (params.seed == LLAMA_DEFAULT_SEED) {
         params.seed = time(NULL);
@@ -246,23 +245,23 @@ int main(int argc, char ** argv) {
     if (suff_rm_leading_spc && inp_sfx[0] == space_token) {
         inp_sfx.erase(inp_sfx.begin());
     }
-    inp_pfx.insert(inp_pfx.begin(), llama_token_prefix(ctx));
+    inp_pfx.insert(inp_pfx.begin(), llama_token_prefix(model));
     if (add_bos) {
-        inp_pfx.insert(inp_pfx.begin(), llama_token_bos(ctx));
+        inp_pfx.insert(inp_pfx.begin(), llama_token_bos(model));
     }
-    inp_sfx.insert(inp_sfx.begin(), llama_token_suffix(ctx));
+    inp_sfx.insert(inp_sfx.begin(), llama_token_suffix(model));
     embd_inp = inp_pfx;
     embd_inp.insert(embd_inp.end(), inp_sfx.begin(), inp_sfx.end());
-    embd_inp.push_back(llama_token_middle(ctx));
+    embd_inp.push_back(llama_token_middle(model));
 
     LOG("prefix: \"%s\"\n", log_tostr(params.input_prefix));
     LOG("suffix: \"%s\"\n", log_tostr(params.input_suffix));
-    LOG("tokens: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd_inp));
+    LOG("tokens: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd_inp).c_str());
 
     // Should not run without any tokens
     if (embd_inp.empty()) {
-        embd_inp.push_back(llama_token_bos(ctx));
+        embd_inp.push_back(llama_token_bos(model));
-        LOG("embd_inp was considered empty and bos was added: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd_inp));
+        LOG("embd_inp was considered empty and bos was added: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd_inp).c_str());
     }
 
     // Tokenize negative prompt
@@ -273,10 +272,10 @@ int main(int argc, char ** argv) {
         LOG("cfg_negative_prompt: \"%s\"\n", log_tostr(sparams.cfg_negative_prompt));
 
         guidance_inp = ::llama_tokenize(ctx_guidance, sparams.cfg_negative_prompt, add_bos);
-        LOG("guidance_inp tokenized: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx_guidance, guidance_inp));
+        LOG("guidance_inp tokenized: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx_guidance, guidance_inp).c_str());
 
         std::vector<llama_token> original_inp = ::llama_tokenize(ctx, params.prompt, add_bos);
-        LOG("original_inp tokenized: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, original_inp));
+        LOG("original_inp tokenized: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, original_inp).c_str());
 
         original_prompt_len = original_inp.size();
         guidance_offset = (int)guidance_inp.size() - original_prompt_len;
@@ -294,8 +293,8 @@ int main(int argc, char ** argv) {
         params.n_keep = (int)embd_inp.size();
     }
 
-    LOG("inp_pfx: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, inp_pfx));
+    LOG("inp_pfx: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, inp_pfx).c_str());
-    LOG("inp_sfx: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, inp_sfx));
+    LOG("inp_sfx: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, inp_sfx).c_str());
 
 
     // enable interactive mode if interactive start is specified
@@ -358,39 +357,10 @@ int main(int argc, char ** argv) {
             LOG_TEE("Input suffix: '%s'\n", params.input_suffix.c_str());
         }
     }
-    LOG_TEE("sampling: repeat_last_n = %d, repeat_penalty = %f, presence_penalty = %f, frequency_penalty = %f, top_k = %d, tfs_z = %f, top_p = %f, typical_p = %f, temp = %f, mirostat = %d, mirostat_lr = %f, mirostat_ent = %f\n",
+    LOG_TEE("sampling: \n%s\n", llama_sampling_print(sparams).c_str());
-            sparams.repeat_last_n, sparams.repeat_penalty, sparams.presence_penalty, sparams.frequency_penalty, sparams.top_k, sparams.tfs_z, sparams.top_p, sparams.typical_p, sparams.temp, sparams.mirostat, sparams.mirostat_eta, sparams.mirostat_tau);
     LOG_TEE("generate: n_ctx = %d, n_batch = %d, n_predict = %d, n_keep = %d\n", n_ctx, params.n_batch, params.n_predict, params.n_keep);
     LOG_TEE("\n\n");
 
-    struct llama_grammar * grammar = NULL;
-    grammar_parser::parse_state parsed_grammar;
-
-    if (!params.grammar.empty()) {
-        parsed_grammar = grammar_parser::parse(params.grammar.c_str());
-        // will be empty (default) if there are parse errors
-        if (parsed_grammar.rules.empty()) {
-            return 1;
-        }
-        LOG_TEE("%s: grammar:\n", __func__);
-        grammar_parser::print_grammar(stderr, parsed_grammar);
-        LOG_TEE("\n");
-
-        {
-            auto it = sparams.logit_bias.find(llama_token_eos(ctx));
-            if (it != sparams.logit_bias.end() && it->second == -INFINITY) {
-                LOG_TEE("%s: warning: EOS token is disabled, which will cause most grammars to fail\n", __func__);
-            }
-        }
-
-        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"));
-    }
-
-    // TODO: replace with ring-buffer
-    std::vector<llama_token> last_tokens(n_ctx);
-    std::fill(last_tokens.begin(), last_tokens.end(), 0);
     LOG_TEE("\n#####  Infill mode  #####\n\n");
     if (params.infill) {
         printf("\n************\n");
@@ -433,11 +403,7 @@ int main(int argc, char ** argv) {
     std::vector<llama_token> embd;
     std::vector<llama_token> embd_guidance;
 
-    const int n_vocab = llama_n_vocab(model);
+    struct llama_sampling_context * ctx_sampling = llama_sampling_init(sparams);
-
-    llama_sampling_context ctx_sampling = llama_sampling_context_init(params, grammar);
-    std::vector<llama_token_data> candidates;
-    candidates.reserve(n_vocab);
 
     while (n_remain != 0 || params.interactive) {
         // predict
@@ -484,7 +450,7 @@ int main(int argc, char ** argv) {
 
                 LOG("after swap: n_past = %d, n_past_guidance = %d\n", n_past, n_past_guidance);
 
-                LOG("embd: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd));
+                LOG("embd: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd).c_str());
 
             }
 
@@ -512,7 +478,7 @@ int main(int argc, char ** argv) {
                     input_buf  = embd_guidance.data();
                     input_size = embd_guidance.size();
 
-                    LOG("guidance context: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd_guidance));
+                    LOG("guidance context: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd_guidance).c_str());
                 } else {
                     input_buf  = embd.data();
                     input_size = embd.size();
@@ -535,7 +501,7 @@ int main(int argc, char ** argv) {
                     n_eval = params.n_batch;
                 }
 
-                LOG("eval: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd));
+                LOG("eval: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd).c_str());
 
                 if (llama_decode(ctx, llama_batch_get_one(&embd[i], n_eval, n_past, 0))) {
                     LOG_TEE("%s : failed to eval\n", __func__);
@@ -554,12 +520,11 @@ int main(int argc, char ** argv) {
 
         if ((int) embd_inp.size() <= n_consumed && !is_interacting) {
 
-            const llama_token id = llama_sampling_sample(ctx, ctx_guidance, ctx_sampling, last_tokens, candidates);
+            const llama_token id = llama_sampling_sample(ctx_sampling, ctx, ctx_guidance);
 
-            last_tokens.erase(last_tokens.begin());
+            llama_sampling_accept(ctx_sampling, ctx, id, true);
-            last_tokens.push_back(id);
 
-            LOG("last: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, last_tokens));
+            LOG("last: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, ctx_sampling->prev).c_str());
 
             embd.push_back(id);
 
@@ -575,8 +540,11 @@ int main(int argc, char ** argv) {
             LOG("embd_inp.size(): %d, n_consumed: %d\n", (int) embd_inp.size(), n_consumed);
             while ((int) embd_inp.size() > n_consumed) {
                 embd.push_back(embd_inp[n_consumed]);
-                last_tokens.erase(last_tokens.begin());
+
-                last_tokens.push_back(embd_inp[n_consumed]);
+                // push the prompt in the sampling context in order to apply repetition penalties later
+                // for the prompt, we don't apply grammar rules
+                llama_sampling_accept(ctx_sampling, ctx, embd_inp[n_consumed], false);
+
                 ++n_consumed;
                 if ((int) embd.size() >= params.n_batch) {
                     break;
@@ -608,10 +576,10 @@ int main(int argc, char ** argv) {
         if ((int) embd_inp.size() <= n_consumed) {
 
             // deal with eot token in infill mode
-            if ((last_tokens.back() == llama_token_eot(ctx) || is_interacting) && params.interactive){
+            if ((llama_sampling_last(ctx_sampling) == llama_token_eot(model) || is_interacting) && params.interactive){
                 if(is_interacting && !params.interactive_first) {
                     // print an eot token
-                    printf("%s", llama_token_to_piece(ctx, llama_token_eot(ctx)).c_str());
+                    printf("%s", llama_token_to_piece(ctx, llama_token_eot(model)).c_str());
                 }
                 fflush(stdout);
                 printf("\n");
@@ -625,7 +593,7 @@ int main(int argc, char ** argv) {
                     buffer += line;
                 } while (another_line);
                 // check if we got an empty line, if so we use the old input
-                if(!buffer.empty() && !(buffer.length() == 1 && buffer[0] == '\n')) {
+                if (!buffer.empty() && !(buffer.length() == 1 && buffer[0] == '\n')) {
                     params.input_prefix = buffer;
                 }
                 buffer.clear();
@@ -635,7 +603,7 @@ int main(int argc, char ** argv) {
                     buffer += line;
                 } while (another_line);
                 // check if we got an empty line
-                if(!buffer.empty() && !(buffer.length() == 1 && buffer[0] == '\n')) {
+                if (!buffer.empty() && !(buffer.length() == 1 && buffer[0] == '\n')) {
                     params.input_suffix = buffer;
                 }
                 buffer.clear();
@@ -648,7 +616,7 @@ int main(int argc, char ** argv) {
                     process_escapes(params.input_suffix);
                 }
                 suff_rm_leading_spc = params.escape;
-                if (suff_rm_leading_spc && params.input_suffix.find_first_of(" ") == 0 && params.input_suffix.size() > 1) {
+                if (suff_rm_leading_spc && params.input_suffix.find_first_of(' ') == 0 && params.input_suffix.size() > 1) {
                     params.input_suffix.erase(0, 1);
                     suff_rm_leading_spc = false;
                 }
@@ -658,14 +626,14 @@ int main(int argc, char ** argv) {
                 if (suff_rm_leading_spc && inp_sfx[0] == space_token) {
                     inp_sfx.erase(inp_sfx.begin());
                 }
-                inp_pfx.insert(inp_pfx.begin(), llama_token_prefix(ctx));
+                inp_pfx.insert(inp_pfx.begin(), llama_token_prefix(model));
                 if (add_bos) {
-                    inp_pfx.insert(inp_pfx.begin(), llama_token_bos(ctx));
+                    inp_pfx.insert(inp_pfx.begin(), llama_token_bos(model));
                 }
-                inp_sfx.insert(inp_sfx.begin(), llama_token_suffix(ctx));
+                inp_sfx.insert(inp_sfx.begin(), llama_token_suffix(model));
                 embd_inp = inp_pfx;
                 embd_inp.insert(embd_inp.end(), inp_sfx.begin(), inp_sfx.end());
-                embd_inp.push_back(llama_token_middle(ctx));
+                embd_inp.push_back(llama_token_middle(model));
                 embd.clear();
                 embd_guidance.clear();
                 n_remain = params.n_predict;
@@ -675,7 +643,7 @@ int main(int argc, char ** argv) {
                 is_interacting = false;
             }
             // deal with end of text token in interactive mode
-            else if (last_tokens.back() == llama_token_eos(ctx)) {
+            else if (llama_sampling_last(ctx_sampling) == llama_token_eos(model)) {
                 LOG("found EOS token\n");
 
                 if (params.interactive) {
@@ -692,7 +660,7 @@ int main(int argc, char ** argv) {
 
                 if (params.input_prefix_bos) {
                     LOG("adding input prefix BOS token\n");
-                    embd_inp.push_back(llama_token_bos(ctx));
+                    embd_inp.push_back(llama_token_bos(model));
                 }
 
                 std::string buffer;
@@ -727,7 +695,7 @@ int main(int argc, char ** argv) {
                     const size_t original_size = embd_inp.size();
 
                     const auto line_inp = ::llama_tokenize(ctx, buffer, false);
-                    LOG("input tokens: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, line_inp));
+                    LOG("input tokens: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, line_inp).c_str());
 
                     embd_inp.insert(embd_inp.end(), line_inp.begin(), line_inp.end());
 
@@ -748,22 +716,14 @@ int main(int argc, char ** argv) {
 
             if (n_past > 0) {
                 if (is_interacting) {
-                    // reset grammar state if we're restarting generation
+                    llama_sampling_reset(ctx_sampling);
-                    if (grammar != NULL) {
-                        llama_grammar_free(grammar);
-
-                        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"));
-                    }
                 }
                 is_interacting = false;
             }
         }
 
         // end of text token
-        if (!embd.empty() && embd.back() == llama_token_eos(ctx) && !params.interactive) {
+        if (!embd.empty() && embd.back() == llama_token_eos(model) && !params.interactive) {
             break;
         }
 
@@ -775,7 +735,7 @@ int main(int argc, char ** argv) {
         }
     }
     if (!params.interactive && n_remain <= 0) {
-        printf("%s", llama_token_to_piece(ctx, llama_token_eot(ctx)).c_str());
+        printf("%s", llama_token_to_piece(ctx, llama_token_eot(model)).c_str());
         fflush(stdout);
     }
 
@@ -786,9 +746,7 @@ int main(int argc, char ** argv) {
     llama_free(ctx);
     llama_free_model(model);
 
-    if (grammar != NULL) {
+    llama_sampling_free(ctx_sampling);
-        llama_grammar_free(grammar);
-    }
     llama_backend_free();
 
 #ifndef LOG_DISABLE_LOGS

examples/llama-bench/CMakeLists.txt

@@ -3,6 +3,3 @@ add_executable(${TARGET} llama-bench.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_11)
-if(TARGET BUILD_INFO)
-  add_dependencies(${TARGET} BUILD_INFO)
-endif()

examples/llama-bench/llama-bench.cpp

@@ -19,7 +19,6 @@
 #include "ggml.h"
 #include "llama.h"
 #include "common.h"
-#include "build-info.h"
 #include "ggml-cuda.h"
 
 // utils
@@ -641,8 +640,8 @@ struct test {
     }
 };
 
-const std::string test::build_commit = BUILD_COMMIT;
+const std::string test::build_commit = LLAMA_COMMIT;
-const int         test::build_number = BUILD_NUMBER;
+const int         test::build_number = LLAMA_BUILD_NUMBER;
 const bool        test::cuda         = !!ggml_cpu_has_cublas();
 const bool        test::opencl       = !!ggml_cpu_has_clblast();
 const bool        test::metal        = !!ggml_cpu_has_metal();
@@ -933,7 +932,7 @@ struct sql_printer : public printer {
 };
 
 static void test_prompt(llama_context * ctx, int n_prompt, int n_past, int n_batch, int n_threads) {
-    std::vector<llama_token> tokens(n_batch, llama_token_bos(ctx));
+    std::vector<llama_token> tokens(n_batch, llama_token_bos(llama_get_model(ctx)));
     int n_processed = 0;
 
     llama_set_n_threads(ctx, n_threads, n_threads);
@@ -946,7 +945,7 @@ static void test_prompt(llama_context * ctx, int n_prompt, int n_past, int n_bat
 }
 
 static void test_gen(llama_context * ctx, int n_gen, int n_past, int n_threads) {
-    llama_token token = llama_token_bos(ctx);
+    llama_token token = llama_token_bos(llama_get_model(ctx));
 
     llama_set_n_threads(ctx, n_threads, n_threads);
 
@@ -1037,7 +1036,7 @@ int main(int argc, char ** argv) {
 
         test t(inst, lmodel, ctx);
 
-        llama_kv_cache_tokens_rm(ctx, -1, -1);
+        llama_kv_cache_clear(ctx);
 
         // warmup run
         if (t.n_prompt > 0) {
@@ -1048,7 +1047,7 @@ int main(int argc, char ** argv) {
         }
 
         for (int i = 0; i < params.reps; i++) {
-            llama_kv_cache_tokens_rm(ctx, -1, -1);
+            llama_kv_cache_clear(ctx);
 
             uint64_t t_start = get_time_ns();
             if (t.n_prompt > 0) {

examples/llava/CMakeLists.txt

@@ -1,8 +1,22 @@
+set(TARGET llava)
+
+add_library(${TARGET} STATIC llava.cpp llava.h clip.cpp clip.h)
+target_link_libraries(${TARGET} PRIVATE ggml llama ${CMAKE_THREAD_LIBS_INIT})
+
+target_include_directories(${TARGET} PUBLIC .)
+target_include_directories(${TARGET} PUBLIC ../..)
+target_include_directories(${TARGET} PUBLIC ../../common)
+
+target_compile_features(${TARGET} PRIVATE cxx_std_11)
+if (NOT MSVC)
+    target_compile_options(${TARGET} PRIVATE -Wno-cast-qual) # stb_image.h
+    endif()
+if(TARGET BUILD_INFO)
+    add_dependencies(${TARGET} BUILD_INFO)
+endif()
+
 set(TARGET llava-cli)
 add_executable(${TARGET} llava-cli.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llama llava ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_11)
-if(TARGET BUILD_INFO)
-    add_dependencies(${TARGET} BUILD_INFO)
-endif()

examples/llava/clip.cpp

@@ -112,8 +112,7 @@ static float get_f32(const gguf_context * ctx, const std::string & key) {
 static struct ggml_tensor * get_tensor(struct ggml_context * ctx, const std::string & name) {
     struct ggml_tensor * cur = ggml_get_tensor(ctx, name.c_str());
     if (!cur) {
-        printf("unable to find tensor %s\n", name.c_str());
+        throw std::runtime_error(format("%s: unable to find tensor %s\n", __func__, name.c_str()));
-        throw std::runtime_error(format("unable to find tensor %s\n", name.c_str()));
     }
 
     return cur;
@@ -136,7 +135,7 @@ static std::string get_ftype(int ftype) {
     case 8:
         return "q8_0";
     default:
-        throw std::runtime_error(format("Unrecognized file type: %d\n", ftype));
+        throw std::runtime_error(format("%s: Unrecognized file type: %d\n", __func__, ftype));
     }
 }
 
@@ -462,6 +461,9 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
     };
 
     struct gguf_context * ctx = gguf_init_from_file(fname, params);
+    if (!ctx) {
+        throw std::runtime_error(format("%s: failed to load CLIP model from %s. Does this file exist?\n", __func__, fname));
+    }
 
     if (verbosity >= 1) {
         const int n_tensors = gguf_get_n_tensors(ctx);
@@ -608,8 +610,8 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
         int idx_mean = get_key_idx(ctx, KEY_IMAGE_MEAN);
         int idx_std = get_key_idx(ctx, KEY_IMAGE_STD);
         for (int i = 0; i < 3; ++i) {
-            new_clip->image_mean[i] = *((float *)gguf_get_arr_data(ctx, idx_mean));
+            new_clip->image_mean[i] = *((const float *)gguf_get_arr_data(ctx, idx_mean));
-            new_clip->image_std[i] = *((float *)gguf_get_arr_data(ctx, idx_std));
+            new_clip->image_std[i] = *((const float *)gguf_get_arr_data(ctx, idx_std));
         }
 
         if (verbosity >= 2) {

examples/llava/clip.h

@@ -1,7 +1,6 @@
 #ifndef CLIP_H
 #define CLIP_H
 
-#include "ggml.h"
 #include "llama.h"
 
 struct clip_ctx;

examples/llava/llava-surgery.py

@@ -16,13 +16,29 @@ checkpoint = torch.load(path)
 mm_tensors = [k for k, v in checkpoint.items() if k.startswith("model.mm_projector")]
 
 # store these tensors in a new dictionary and torch.save them
-projector = {name: checkpoint[name] for name in mm_tensors}
+projector = {name: checkpoint[name].float() for name in mm_tensors}
 torch.save(projector, f"{args.model}/llava.projector")
 
 # remove these tensors from the checkpoint and save it again
 for name in mm_tensors:
     del checkpoint[name]
 
+# BakLLaVA models contain CLIP tensors in it
+clip_tensors = [k for k, v in checkpoint.items() if k.startswith("model.vision_tower")]
+if len(clip_tensors) > 0:
+    clip = {name.replace("vision_tower.vision_tower.", ""): checkpoint[name].float() for name in clip_tensors}
+    torch.save(clip, f"{args.model}/llava.clip")
+
+    # remove these tensors
+    for name in clip_tensors:
+        del checkpoint[name]
+
+    # added tokens should be removed to be able to convert Mistral models
+    if os.path.exists(f"{args.model}/added_tokens.json"):
+        with open(f"{args.model}/added_tokens.json", "w") as f:
+            f.write("{}\n")
+
+
 torch.save(checkpoint, path)
 
 print("Done!")

examples/llava/llava-utils.h

@@ -12,6 +12,25 @@
 #include <cstdlib>
 #include <vector>
 
+// todo: remove this
+inline bool eval_image_embd(llama_context * ctx_llama, float * embd, int N, int n_batch, int * n_past) {
+    int n_embd  = llama_n_embd(llama_get_model(ctx_llama));
+
+    for (int i = 0; i < N; i += n_batch) {
+        int n_eval = N - i;
+        if (n_eval > n_batch) {
+            n_eval = n_batch;
+        }
+        llama_batch batch = {int32_t(n_eval), nullptr, (embd+i*n_embd), nullptr, nullptr, nullptr, nullptr, *n_past, 1, 0, };
+        if (llama_decode(ctx_llama, batch)) {
+            fprintf(stderr, "%s : failed to eval\n", __func__);
+            return false;
+        }
+        *n_past += n_eval;
+    }
+    return true;
+}
+
 inline bool eval_tokens(struct llama_context * ctx_llama, std::vector<llama_token> tokens, int n_batch, int * n_past) {
     int N = (int) tokens.size();
     for (int i = 0; i < N; i += n_batch) {
@@ -34,37 +53,39 @@ inline bool eval_id(struct llama_context * ctx_llama, int id, int * n_past) {
     return eval_tokens(ctx_llama, tokens, 1, n_past);
 }
 
-inline bool eval_string(struct llama_context * ctx_llama, const char* str, int n_batch, int * n_past){
+inline bool eval_string(struct llama_context * ctx_llama, const char* str, int n_batch, int * n_past, bool add_bos){
     std::string              str2     = str;
-    std::vector<llama_token> embd_inp = ::llama_tokenize(ctx_llama, str2, true);
+    std::vector<llama_token> embd_inp = ::llama_tokenize(ctx_llama, str2, add_bos);
     eval_tokens(ctx_llama, embd_inp, n_batch, n_past);
     return true;
 }
 
 // TODO: use common/sampling.h
 inline llama_token sample_id(llama_context * ctx_llama, gpt_params & params) {
-      // out of user input, sample next token
+    auto & sparams = params.sparams;
-    const float   temp      = params.sampling_params.temp;
+
-    const int32_t top_k     = params.sampling_params.top_k <= 0 ? llama_n_vocab(llama_get_model(ctx_llama)) : params.sampling_params.top_k;
+    // out of user input, sample next token
-    const float   top_p     = params.sampling_params.top_p;
+    const float   temp      = sparams.temp;
-    const float   tfs_z     = params.sampling_params.tfs_z;
+    const int32_t top_k     = sparams.top_k <= 0 ? llama_n_vocab(llama_get_model(ctx_llama)) : sparams.top_k;
-    const float   typical_p = params.sampling_params.typical_p;
+    const float   top_p     = sparams.top_p;
-      // const int32_t repeat_last_n   = params.sampling_params.repeat_last_n < 0 ? n_ctx : params.sampling_params.repeat_last_n;
+    const float   tfs_z     = sparams.tfs_z;
-      // const float   repeat_penalty  = params.sampling_params.repeat_penalty;
+    const float   typical_p = sparams.typical_p;
-      // const float   alpha_presence  = params.sampling_params.presence_penalty;
+    // const int32_t repeat_last_n   = sparams.repeat_last_n < 0 ? n_ctx : sparams.repeat_last_n;
-      // const float   alpha_frequency = params.sampling_params.frequency_penalty;
+    // const float   repeat_penalty  = sparams.repeat_penalty;
-    const int     mirostat     = params.sampling_params.mirostat;
+    // const float   alpha_presence  = sparams.presence_penalty;
-    const float   mirostat_tau = params.sampling_params.mirostat_tau;
+    // const float   alpha_frequency = sparams.frequency_penalty;
-    const float   mirostat_eta = params.sampling_params.mirostat_eta;
+    const int     mirostat     = sparams.mirostat;
-      // const bool    penalize_nl     = params.sampling_params.penalize_nl;
+    const float   mirostat_tau = sparams.mirostat_tau;
+    const float   mirostat_eta = sparams.mirostat_eta;
+    // const bool    penalize_nl     = sparams.penalize_nl;
 
     llama_token id = 0;
     {
         auto logits  = llama_get_logits(ctx_llama);
         auto n_vocab = llama_n_vocab(llama_get_model(ctx_llama));
 
-          // Apply params.logit_bias map
+        // Apply params.logit_bias map
-        for (auto it = params.sampling_params.logit_bias.begin(); it != params.sampling_params.logit_bias.end(); it++) {
+        for (auto it = sparams.logit_bias.begin(); it != sparams.logit_bias.end(); it++) {
             logits[it->first] += it->second;
         }
 
@@ -76,18 +97,18 @@ inline llama_token sample_id(llama_context * ctx_llama, gpt_params & params) {
 
         llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false };
 
-          // TODO: Apply penalties
+        // TODO: Apply penalties
-          // float nl_logit = logits[llama_token_nl(ctx)];
+        // float nl_logit = logits[llama_token_nl(ctx)];
-          // auto last_n_repeat = std::min(std::min((int)last_n_tokens.size(), repeat_last_n), n_ctx);
+        // auto last_n_repeat = std::min(std::min((int)last_n_tokens.size(), repeat_last_n), n_ctx);
-          // llama_sample_repetition_penalty(ctx, &candidates_p,
+        // llama_sample_repetition_penalty(ctx, &candidates_p,
-          //      last_n_tokens.data() + last_n_tokens.size() - last_n_repeat,
+        //      last_n_tokens.data() + last_n_tokens.size() - last_n_repeat,
-          //      last_n_repeat, repeat_penalty);
+        //      last_n_repeat, repeat_penalty);
-          // llama_sample_frequency_and_presence_penalties(ctx, &candidates_p,
+        // llama_sample_frequency_and_presence_penalties(ctx, &candidates_p,
-          // last_n_tokens.data() + last_n_tokens.size() - last_n_repeat,
+        // last_n_tokens.data() + last_n_tokens.size() - last_n_repeat,
-          // last_n_repeat, alpha_frequency, alpha_presence);
+        // last_n_repeat, alpha_frequency, alpha_presence);
-          // if (!penalize_nl) {
+        // if (!penalize_nl) {
-          //     logits[llama_token_nl(ctx)] = nl_logit;
+        //     logits[llama_token_nl(ctx)] = nl_logit;
-          // }
+        // }
 
         if (temp <= 0) {
               // Greedy sampling
@@ -120,7 +141,7 @@ inline llama_token sample_id(llama_context * ctx_llama, gpt_params & params) {
 inline const char * sample(struct llama_context * ctx_llama, gpt_params & params, int * n_past) {
     int id = sample_id(ctx_llama, params);
     static std::string ret;
-    if (id == llama_token_eos(ctx_llama)) {
+    if (id == llama_token_eos(llama_get_model(ctx_llama))) {
         ret = "</s>";
     } else {
         ret = llama_token_to_piece(ctx_llama, id);

examples/llava/llava.cpp

@@ -80,7 +80,7 @@ bool llava_eval_image_embed(llama_context * ctx_llama, const struct llava_image_
         if (n_eval > n_batch) {
             n_eval = n_batch;
         }
-        llama_batch batch = {int32_t(n_eval), nullptr, (image_embed->embed+i*n_embd), nullptr, nullptr, nullptr, *n_past, 1, 0, };
+        llama_batch batch = {int32_t(n_eval), nullptr, (image_embed->embed+i*n_embd), nullptr, nullptr, nullptr, nullptr, *n_past, 1, 0, };
         if (llama_decode(ctx_llama, batch)) {
             fprintf(stderr, "%s : failed to eval\n", __func__);
             return false;

examples/llava/test-llava.cpp

@@ -1,6 +0,0 @@
-#include <cstdio>
-
-int main(int argc, char ** argv) {
-    printf("dummy llava test\n");
-    return 0;
-}

examples/main-cmake-pkg/CMakeLists.txt

@@ -16,6 +16,8 @@ add_library(common OBJECT
     ${_common_path}/console.cpp
     ${_common_path}/grammar-parser.h
     ${_common_path}/grammar-parser.cpp
+    ${_common_path}/sampling.h
+    ${_common_path}/sampling.cpp
     )
 
 # WARNING: because build-info.h is auto-generated, it will only

examples/main/CMakeLists.txt

@@ -3,6 +3,3 @@ add_executable(${TARGET} main.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_11)
-if(TARGET BUILD_INFO)
-  add_dependencies(${TARGET} BUILD_INFO)
-endif()

examples/main/README.md

@@ -208,6 +208,14 @@ Top-p sampling, also known as nucleus sampling, is another text generation metho
 
 Example usage: `--top-p 0.95`
 
+### Min P Sampling
+
+-   `--min-p N`: Sets a minimum base probability threshold for token selection (default: 0.05).
+
+The Min-P sampling method was designed as an alternative to Top-P, and aims to ensure a balance of quality and variety. The parameter *p* represents the minimum probability for a token to be considered, relative to the probability of the most likely token. For example, with *p*=0.05 and the most likely token having a probability of 0.9, logits with a value less than 0.045 are filtered out.
+
+Example usage: `--min-p 0.05`
+
 ### Tail Free Sampling (TFS)
 
 -   `--tfs N`: Enable tail free sampling with parameter z (default: 1.0, 1.0 = disabled).

examples/main/main.cpp

@@ -2,8 +2,6 @@
 
 #include "console.h"
 #include "llama.h"
-#include "build-info.h"
-#include "grammar-parser.h"
 
 #include <cassert>
 #include <cinttypes>
@@ -109,7 +107,7 @@ int main(int argc, char ** argv) {
     if (!gpt_params_parse(argc, argv, params)) {
         return 1;
     }
-    llama_sampling_params & sparams = params.sampling_params;
+    llama_sampling_params & sparams = params.sparams;
 
 #ifndef LOG_DISABLE_LOGS
     log_set_target(log_filename_generator("main", "log"));
@@ -154,8 +152,8 @@ int main(int argc, char ** argv) {
         LOG_TEE("%s: warning: scaling RoPE frequency by %g.\n", __func__, params.rope_freq_scale);
     }
 
-    LOG_TEE("%s: build = %d (%s)\n", __func__, BUILD_NUMBER, BUILD_COMMIT);
+    LOG_TEE("%s: build = %d (%s)\n",      __func__, LLAMA_BUILD_NUMBER, LLAMA_COMMIT);
-    LOG_TEE("%s: built with %s for %s\n", __func__, BUILD_COMPILER, BUILD_TARGET);
+    LOG_TEE("%s: built with %s for %s\n", __func__, LLAMA_COMPILER, LLAMA_BUILD_TARGET);
 
     if (params.seed == LLAMA_DEFAULT_SEED) {
         params.seed = time(NULL);
@@ -238,19 +236,19 @@ int main(int argc, char ** argv) {
 
     if (params.interactive_first || params.instruct || !params.prompt.empty() || session_tokens.empty()) {
         LOG("tokenize the prompt\n");
-        embd_inp = ::llama_tokenize(ctx, params.prompt, add_bos);
+        embd_inp = ::llama_tokenize(ctx, params.prompt, add_bos, true);
     } else {
         LOG("use session tokens\n");
         embd_inp = session_tokens;
     }
 
     LOG("prompt: \"%s\"\n", log_tostr(params.prompt));
-    LOG("tokens: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd_inp));
+    LOG("tokens: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd_inp).c_str());
 
     // Should not run without any tokens
     if (embd_inp.empty()) {
-        embd_inp.push_back(llama_token_bos(ctx));
+        embd_inp.push_back(llama_token_bos(model));
-        LOG("embd_inp was considered empty and bos was added: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd_inp));
+        LOG("embd_inp was considered empty and bos was added: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd_inp).c_str());
     }
 
     // Tokenize negative prompt
@@ -260,11 +258,11 @@ int main(int argc, char ** argv) {
     if (ctx_guidance) {
         LOG("cfg_negative_prompt: \"%s\"\n", log_tostr(sparams.cfg_negative_prompt));
 
-        guidance_inp = ::llama_tokenize(ctx_guidance, sparams.cfg_negative_prompt, add_bos);
+        guidance_inp = ::llama_tokenize(ctx_guidance, sparams.cfg_negative_prompt, add_bos, true);
-        LOG("guidance_inp tokenized: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx_guidance, guidance_inp));
+        LOG("guidance_inp tokenized: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx_guidance, guidance_inp).c_str());
 
-        std::vector<llama_token> original_inp = ::llama_tokenize(ctx, params.prompt, add_bos);
+        std::vector<llama_token> original_inp = ::llama_tokenize(ctx, params.prompt, add_bos, true);
-        LOG("original_inp tokenized: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, original_inp));
+        LOG("original_inp tokenized: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, original_inp).c_str());
 
         original_prompt_len = original_inp.size();
         guidance_offset = (int)guidance_inp.size() - original_prompt_len;
@@ -299,7 +297,7 @@ int main(int argc, char ** argv) {
         }
 
         // remove any "future" tokens that we might have inherited from the previous session
-        llama_kv_cache_tokens_rm(ctx, n_matching_session_tokens, -1);
+        llama_kv_cache_seq_rm(ctx, -1, n_matching_session_tokens, -1);
     }
 
     LOGLN(
@@ -320,11 +318,11 @@ int main(int argc, char ** argv) {
     }
 
     // prefix & suffix for instruct mode
-    const auto inp_pfx = ::llama_tokenize(ctx, "\n\n### Instruction:\n\n", add_bos);
+    const auto inp_pfx = ::llama_tokenize(ctx, "\n\n### Instruction:\n\n", add_bos, true);
-    const auto inp_sfx = ::llama_tokenize(ctx, "\n\n### Response:\n\n",    false);
+    const auto inp_sfx = ::llama_tokenize(ctx, "\n\n### Response:\n\n",    false,   true);
 
-    LOG("inp_pfx: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, inp_pfx));
+    LOG("inp_pfx: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, inp_pfx).c_str());
-    LOG("inp_sfx: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, inp_sfx));
+    LOG("inp_sfx: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, inp_sfx).c_str());
 
     // in instruct mode, we inject a prefix and a suffix to each input by the user
     if (params.instruct) {
@@ -383,6 +381,12 @@ int main(int argc, char ** argv) {
         if (!params.antiprompt.empty()) {
             for (const auto & antiprompt : params.antiprompt) {
                 LOG_TEE("Reverse prompt: '%s'\n", antiprompt.c_str());
+                if (params.verbose_prompt) {
+                    auto tmp = ::llama_tokenize(ctx, antiprompt, false, true);
+                    for (int i = 0; i < (int) tmp.size(); i++) {
+                        LOG_TEE("%6d -> '%s'\n", tmp[i], llama_token_to_piece(ctx, tmp[i]).c_str());
+                    }
+                }
             }
         }
 
@@ -392,45 +396,27 @@ int main(int argc, char ** argv) {
 
         if (!params.input_prefix.empty()) {
             LOG_TEE("Input prefix: '%s'\n", params.input_prefix.c_str());
+            if (params.verbose_prompt) {
+                auto tmp = ::llama_tokenize(ctx, params.input_prefix, true, true);
+                for (int i = 0; i < (int) tmp.size(); i++) {
+                    LOG_TEE("%6d -> '%s'\n", tmp[i], llama_token_to_piece(ctx, tmp[i]).c_str());
+                }
+            }
         }
 
         if (!params.input_suffix.empty()) {
             LOG_TEE("Input suffix: '%s'\n", params.input_suffix.c_str());
-        }
+            if (params.verbose_prompt) {
-    }
+                auto tmp = ::llama_tokenize(ctx, params.input_suffix, false, true);
-    LOG_TEE("sampling: repeat_last_n = %d, repeat_penalty = %f, presence_penalty = %f, frequency_penalty = %f, top_k = %d, tfs_z = %f, top_p = %f, typical_p = %f, temp = %f, mirostat = %d, mirostat_lr = %f, mirostat_ent = %f\n",
+                for (int i = 0; i < (int) tmp.size(); i++) {
-            sparams.repeat_last_n, sparams.repeat_penalty, sparams.presence_penalty, sparams.frequency_penalty, sparams.top_k, sparams.tfs_z, sparams.top_p, sparams.typical_p, sparams.temp, sparams.mirostat, sparams.mirostat_eta, sparams.mirostat_tau);
+                    LOG_TEE("%6d -> '%s'\n", tmp[i], llama_token_to_piece(ctx, tmp[i]).c_str());
-    LOG_TEE("generate: n_ctx = %d, n_batch = %d, n_predict = %d, n_keep = %d\n", n_ctx, params.n_batch, params.n_predict, params.n_keep);
+                }
-    LOG_TEE("\n\n");
-
-    struct llama_grammar * grammar = NULL;
-    grammar_parser::parse_state parsed_grammar;
-
-    if (!params.grammar.empty()) {
-        parsed_grammar = grammar_parser::parse(params.grammar.c_str());
-        // will be empty (default) if there are parse errors
-        if (parsed_grammar.rules.empty()) {
-            return 1;
-        }
-        LOG_TEE("%s: grammar:\n", __func__);
-        grammar_parser::print_grammar(stderr, parsed_grammar);
-        LOG_TEE("\n");
-
-        {
-            auto it = sparams.logit_bias.find(llama_token_eos(ctx));
-            if (it != sparams.logit_bias.end() && it->second == -INFINITY) {
-                LOG_TEE("%s: warning: EOS token is disabled, which will cause most grammars to fail\n", __func__);
             }
         }
-
-        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"));
     }
-
+    LOG_TEE("sampling: \n%s\n", llama_sampling_print(sparams).c_str());
-    // TODO: replace with ring-buffer
+    LOG_TEE("generate: n_ctx = %d, n_batch = %d, n_predict = %d, n_keep = %d\n", n_ctx, params.n_batch, params.n_predict, params.n_keep);
-    std::vector<llama_token> last_tokens(n_ctx);
+    LOG_TEE("\n\n");
-    std::fill(last_tokens.begin(), last_tokens.end(), 0);
 
     if (params.interactive) {
         const char *control_message;
@@ -471,11 +457,7 @@ int main(int argc, char ** argv) {
     std::vector<llama_token> embd;
     std::vector<llama_token> embd_guidance;
 
-    const int n_vocab = llama_n_vocab(model);
+    struct llama_sampling_context * ctx_sampling = llama_sampling_init(sparams);
-
-    llama_sampling_context ctx_sampling = llama_sampling_context_init(params, grammar);
-    std::vector<llama_token_data> candidates;
-    candidates.reserve(n_vocab);
 
     while ((n_remain != 0 && !is_antiprompt) || params.interactive) {
         // predict
@@ -522,7 +504,7 @@ int main(int argc, char ** argv) {
 
                 LOG("after swap: n_past = %d, n_past_guidance = %d\n", n_past, n_past_guidance);
 
-                LOG("embd: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd));
+                LOG("embd: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd).c_str());
 
                 LOG("clear session path\n");
                 path_session.clear();
@@ -552,7 +534,6 @@ int main(int argc, char ** argv) {
 
             // evaluate tokens in batches
             // embd is typically prepared beforehand to fit within a batch, but not always
-
             if (ctx_guidance) {
                 int input_size = 0;
                 llama_token * input_buf = NULL;
@@ -574,7 +555,7 @@ int main(int argc, char ** argv) {
                     input_buf  = embd_guidance.data();
                     input_size = embd_guidance.size();
 
-                    LOG("guidance context: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd_guidance));
+                    LOG("guidance context: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd_guidance).c_str());
                 } else {
                     input_buf  = embd.data();
                     input_size = embd.size();
@@ -597,7 +578,7 @@ int main(int argc, char ** argv) {
                     n_eval = params.n_batch;
                 }
 
-                LOG("eval: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd));
+                LOG("eval: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd).c_str());
 
                 if (llama_decode(ctx, llama_batch_get_one(&embd[i], n_eval, n_past, 0))) {
                     LOG_TEE("%s : failed to eval\n", __func__);
@@ -627,12 +608,11 @@ int main(int argc, char ** argv) {
                 LOG("saved session to %s\n", path_session.c_str());
             }
 
-            const llama_token id = llama_sampling_sample(ctx, ctx_guidance, ctx_sampling, last_tokens, candidates);
+            const llama_token id = llama_sampling_sample(ctx_sampling, ctx, ctx_guidance);
 
-            last_tokens.erase(last_tokens.begin());
+            llama_sampling_accept(ctx_sampling, ctx, id, true);
-            last_tokens.push_back(id);
 
-            LOG("last: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, last_tokens));
+            LOG("last: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, ctx_sampling->prev).c_str());
 
             embd.push_back(id);
 
@@ -648,8 +628,11 @@ int main(int argc, char ** argv) {
             LOG("embd_inp.size(): %d, n_consumed: %d\n", (int) embd_inp.size(), n_consumed);
             while ((int) embd_inp.size() > n_consumed) {
                 embd.push_back(embd_inp[n_consumed]);
-                last_tokens.erase(last_tokens.begin());
+
-                last_tokens.push_back(embd_inp[n_consumed]);
+                // push the prompt in the sampling context in order to apply repetition penalties later
+                // for the prompt, we don't apply grammar rules
+                llama_sampling_accept(ctx_sampling, ctx, embd_inp[n_consumed], false);
+
                 ++n_consumed;
                 if ((int) embd.size() >= params.n_batch) {
                     break;
@@ -679,12 +662,10 @@ int main(int argc, char ** argv) {
 
         // if not currently processing queued inputs;
         if ((int) embd_inp.size() <= n_consumed) {
-            // check for reverse prompt
+            // check for reverse prompt in the last n_prev tokens
             if (!params.antiprompt.empty()) {
-                std::string last_output;
+                const int n_prev = 32;
-                for (auto id : last_tokens) {
+                const std::string last_output = llama_sampling_prev_str(ctx_sampling, ctx, n_prev);
-                    last_output += llama_token_to_piece(ctx, id);
-                }
 
                 is_antiprompt = false;
                 // Check if each of the reverse prompts appears at the end of the output.
@@ -711,13 +692,13 @@ int main(int argc, char ** argv) {
             }
 
             // deal with end of text token in interactive mode
-            if (last_tokens.back() == llama_token_eos(ctx)) {
+            if (llama_sampling_last(ctx_sampling) == llama_token_eos(model)) {
                 LOG("found EOS token\n");
 
                 if (params.interactive) {
                     if (!params.antiprompt.empty()) {
                         // tokenize and inject first reverse prompt
-                        const auto first_antiprompt = ::llama_tokenize(ctx, params.antiprompt.front(), false);
+                        const auto first_antiprompt = ::llama_tokenize(ctx, params.antiprompt.front(), false, true);
                         embd_inp.insert(embd_inp.end(), first_antiprompt.begin(), first_antiprompt.end());
                         is_antiprompt = true;
                     }
@@ -738,14 +719,13 @@ int main(int argc, char ** argv) {
 
                 if (params.input_prefix_bos) {
                     LOG("adding input prefix BOS token\n");
-                    embd_inp.push_back(llama_token_bos(ctx));
+                    embd_inp.push_back(llama_token_bos(model));
                 }
 
                 std::string buffer;
                 if (!params.input_prefix.empty()) {
                     LOG("appending input prefix: '%s'\n", params.input_prefix.c_str());
-                    buffer += params.input_prefix;
+                    printf("%s", params.input_prefix.c_str());
-                    printf("%s", buffer.c_str());
                 }
 
                 // color user input only
@@ -767,7 +747,6 @@ int main(int argc, char ** argv) {
                     // append input suffix if any
                     if (!params.input_suffix.empty()) {
                         LOG("appending input suffix: '%s'\n", params.input_suffix.c_str());
-                        buffer += params.input_suffix;
                         printf("%s", params.input_suffix.c_str());
                     }
 
@@ -781,11 +760,18 @@ int main(int argc, char ** argv) {
                         n_consumed = embd_inp.size();
                         embd_inp.insert(embd_inp.end(), inp_pfx.begin(), inp_pfx.end());
                     }
+                    if (params.escape) {
+                        process_escapes(buffer);
+                    }
 
-                    const auto line_inp = ::llama_tokenize(ctx, buffer, false);
+                    const auto line_pfx = ::llama_tokenize(ctx, params.input_prefix, false, true);
-                    LOG("input tokens: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, line_inp));
+                    const auto line_inp = ::llama_tokenize(ctx, buffer,              false, false);
+                    const auto line_sfx = ::llama_tokenize(ctx, params.input_suffix, false, true);
+                    LOG("input tokens: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, line_inp).c_str());
 
+                    embd_inp.insert(embd_inp.end(), line_pfx.begin(), line_pfx.end());
                     embd_inp.insert(embd_inp.end(), line_inp.begin(), line_inp.end());
+                    embd_inp.insert(embd_inp.end(), line_sfx.begin(), line_sfx.end());
 
                     // instruct mode: insert response suffix
                     if (params.instruct) {
@@ -810,22 +796,14 @@ int main(int argc, char ** argv) {
 
             if (n_past > 0) {
                 if (is_interacting) {
-                    // reset grammar state if we're restarting generation
+                    llama_sampling_reset(ctx_sampling);
-                    if (grammar != NULL) {
-                        llama_grammar_free(grammar);
-
-                        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"));
-                    }
                 }
                 is_interacting = false;
             }
         }
 
         // end of text token
-        if (!embd.empty() && embd.back() == llama_token_eos(ctx) && !(params.instruct || params.interactive)) {
+        if (!embd.empty() && embd.back() == llama_token_eos(model) && !(params.instruct || params.interactive)) {
             LOG_TEE(" [end of text]\n");
             break;
         }
@@ -850,9 +828,7 @@ int main(int argc, char ** argv) {
     llama_free(ctx);
     llama_free_model(model);
 
-    if (grammar != NULL) {
+    llama_sampling_free(ctx_sampling);
-        llama_grammar_free(grammar);
-    }
     llama_backend_free();
 
 #ifndef LOG_DISABLE_LOGS

examples/parallel/CMakeLists.txt

@@ -3,6 +3,3 @@ add_executable(${TARGET} parallel.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_11)
-if(TARGET BUILD_INFO)
-  add_dependencies(${TARGET} BUILD_INFO)
-endif()

examples/parallel/parallel.cpp

@@ -1,8 +1,6 @@
 // A basic application simulating a server with multiple clients.
 // The clients submite requests to the server and they are processed in parallel.
 
-#include "build-info.h"
-
 #include "common.h"
 #include "llama.h"
 
@@ -51,6 +49,12 @@ static std::vector<std::string> k_prompts = {
 };
 
 struct client {
+    ~client() {
+        if (ctx_sampling) {
+            llama_sampling_free(ctx_sampling);
+        }
+    }
+
     int32_t id = 0;
 
     llama_seq_id seq_id = -1;
@@ -68,7 +72,7 @@ struct client {
     std::string prompt;
     std::string response;
 
-    std::vector<llama_token> tokens_prev;
+    struct llama_sampling_context * ctx_sampling = nullptr;
 };
 
 static void print_date_time() {
@@ -125,8 +129,6 @@ int main(int argc, char ** argv) {
     params.logits_all = true;
     std::tie(model, ctx) = llama_init_from_gpt_params(params);
 
-    llama_sampling_context ctx_sampling = llama_sampling_context_init(params, NULL);
-
     // load the prompts from an external file if there are any
     if (params.prompt.empty()) {
         printf("\n\033[32mNo new questions so proceed with build-in defaults.\033[0m\n");
@@ -147,20 +149,15 @@ int main(int argc, char ** argv) {
     fprintf(stderr, "\n\n");
     fflush(stderr);
 
-    const int n_ctx   = llama_n_ctx(ctx);
+    const int n_ctx = llama_n_ctx(ctx);
-    const int n_vocab = llama_n_vocab(model);
 
     std::vector<client> clients(n_clients);
     for (size_t i = 0; i < clients.size(); ++i) {
         auto & client = clients[i];
         client.id = i;
-        client.tokens_prev.resize(std::max(256, params.n_predict));
+        client.ctx_sampling = llama_sampling_init(params.sparams);
-        std::fill(client.tokens_prev.begin(), client.tokens_prev.end(), 0);
     }
 
-    std::vector<llama_token_data> candidates;
-    candidates.reserve(n_vocab);
-
     std::vector<llama_token> tokens_system;
     tokens_system = ::llama_tokenize(ctx, k_system, true);
     const int32_t n_tokens_system = tokens_system.size();
@@ -169,7 +166,7 @@ int main(int argc, char ** argv) {
 
     // the max batch size is as large as the context to handle cases where we get very long input prompt from multiple
     // users. regardless of the size, the main loop will chunk the batch into a maximum of params.n_batch tokens at a time
-    llama_batch batch = llama_batch_init(n_ctx, 0);
+    llama_batch batch = llama_batch_init(n_ctx, 0, 1);
 
     int32_t n_total_prompt = 0;
     int32_t n_total_gen    = 0;
@@ -184,13 +181,8 @@ int main(int argc, char ** argv) {
     {
         LOG_TEE("%s: Evaluating the system prompt ...\n", __func__);
 
-        batch.n_tokens = n_tokens_system;
+        for (int32_t i = 0; i < n_tokens_system; ++i) {
-
+            llama_batch_add(batch, tokens_system[i], i, { 0 }, false);
-        for (int32_t i = 0; i < batch.n_tokens; ++i) {
-            batch.token[i]  = tokens_system[i];
-            batch.pos[i]    = i;
-            batch.seq_id[i] = 0;
-            batch.logits[i] = false;
         }
 
         if (llama_decode(ctx, batch) != 0) {
@@ -209,7 +201,7 @@ int main(int argc, char ** argv) {
     LOG_TEE("Processing requests ...\n\n");
 
     while (true) {
-        batch.n_tokens = 0;
+        llama_batch_clear(batch);
 
         // decode any currently ongoing sequences
         for (auto & client : clients) {
@@ -217,15 +209,11 @@ int main(int argc, char ** argv) {
                 continue;
             }
 
-            batch.token [batch.n_tokens] = client.sampled;
-            batch.pos   [batch.n_tokens] = n_tokens_system + client.n_prompt + client.n_decoded;
-            batch.seq_id[batch.n_tokens] = client.id;
-            batch.logits[batch.n_tokens] = true;
-
-            client.n_decoded += 1;
             client.i_batch = batch.n_tokens;
 
-            batch.n_tokens += 1;
+            llama_batch_add(batch, client.sampled, n_tokens_system + client.n_prompt + client.n_decoded, { client.id }, true);
+
+            client.n_decoded += 1;
         }
 
         if (batch.n_tokens == 0) {
@@ -250,18 +238,14 @@ int main(int argc, char ** argv) {
                     client.prompt   = client.input + "\nAssistant:";
                     client.response = "";
 
-                    std::fill(client.tokens_prev.begin(), client.tokens_prev.end(), 0);
+                    llama_sampling_reset(client.ctx_sampling);
 
                     // do not prepend BOS because we have a system prompt!
                     std::vector<llama_token> tokens_prompt;
                     tokens_prompt = ::llama_tokenize(ctx, client.prompt, false);
 
                     for (size_t i = 0; i < tokens_prompt.size(); ++i) {
-                        batch.token [batch.n_tokens] = tokens_prompt[i];
+                        llama_batch_add(batch, tokens_prompt[i], i + n_tokens_system, { client.id }, false);
-                        batch.pos   [batch.n_tokens] = i + n_tokens_system;
-                        batch.seq_id[batch.n_tokens] = client.id;
-                        batch.logits[batch.n_tokens] = false;
-                        batch.n_tokens += 1;
                     }
 
                     // extract the logits only for the last token
@@ -304,11 +288,12 @@ int main(int argc, char ** argv) {
 
             llama_batch batch_view = {
                 n_tokens,
-                batch.token  + i,
+                batch.token    + i,
                 nullptr,
-                batch.pos    + i,
+                batch.pos      + i,
-                batch.seq_id + i,
+                batch.n_seq_id + i,
-                batch.logits + i,
+                batch.seq_id   + i,
+                batch.logits   + i,
                 0, 0, 0, // unused
             };
 
@@ -341,7 +326,9 @@ int main(int argc, char ** argv) {
                 //printf("client %d, seq %d, token %d, pos %d, batch %d\n",
                 //        client.id, client.seq_id, client.sampled, client.n_decoded, client.i_batch);
 
-                const llama_token id = llama_sampling_sample(ctx, NULL, ctx_sampling, client.tokens_prev, candidates, client.i_batch - i, client.seq_id);
+                const llama_token id = llama_sampling_sample(client.ctx_sampling, ctx, NULL, client.i_batch - i);
+
+                llama_sampling_accept(client.ctx_sampling, ctx, id, true);
 
                 if (client.n_decoded == 1) {
                     // start measuring generation time after the first token to make sure all concurrent clients
@@ -349,11 +336,8 @@ int main(int argc, char ** argv) {
                     client.t_start_gen = ggml_time_us();
                 }
 
-                // remember which tokens were sampled - used for repetition penalties during sampling
-                client.tokens_prev.erase(client.tokens_prev.begin());
-                client.tokens_prev.push_back(id);
-
                 const std::string token_str = llama_token_to_piece(ctx, id);
+
                 client.response += token_str;
                 client.sampled = id;
 
@@ -361,7 +345,7 @@ int main(int argc, char ** argv) {
                 //        client.id, client.seq_id, id, client.n_decoded, client.i_batch, token_str.c_str());
 
                 if (client.n_decoded > 2 &&
-                        (id == llama_token_eos(ctx) ||
+                        (id == llama_token_eos(model) ||
                          (params.n_predict > 0 && client.n_decoded + client.n_prompt >= params.n_predict) ||
                          client.response.find("User:") != std::string::npos ||
                          client.response.find('\n') != std::string::npos)) {
@@ -386,7 +370,7 @@ int main(int argc, char ** argv) {
 
                     n_total_prompt += client.n_prompt;
                     n_total_gen    += client.n_decoded;
-                    llama_sampling_context_reset(ctx_sampling, client.seq_id);
+
                     client.seq_id = -1;
                 }
 

examples/perplexity/CMakeLists.txt

@@ -3,6 +3,3 @@ add_executable(${TARGET} perplexity.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_11)
-if(TARGET BUILD_INFO)
-  add_dependencies(${TARGET} BUILD_INFO)
-endif()

examples/perplexity/perplexity.cpp

@@ -1,4 +1,3 @@
-#include "build-info.h"
 #include "common.h"
 #include "llama.h"
 
@@ -210,7 +209,7 @@ static results_perplexity perplexity_v2(llama_context * ctx, const gpt_params &
         const auto t_start = std::chrono::high_resolution_clock::now();
 
         // clear the KV cache
-        llama_kv_cache_tokens_rm(ctx, -1, -1);
+        llama_kv_cache_clear(ctx);
 
         for (int j = 0; j < num_batches; ++j) {
             const int batch_start = start + j * n_batch;
@@ -227,7 +226,7 @@ static results_perplexity perplexity_v2(llama_context * ctx, const gpt_params &
 
             // add BOS token for the first batch of each chunk
             if (add_bos && j == 0) {
-                tokens[batch_start] = llama_token_bos(ctx);
+                tokens[batch_start] = llama_token_bos(llama_get_model(ctx));
             }
 
             const auto batch_logits = llama_get_logits(ctx);
@@ -339,7 +338,7 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par
         const auto t_start = std::chrono::high_resolution_clock::now();
 
         // clear the KV cache
-        llama_kv_cache_tokens_rm(ctx, -1, -1);
+        llama_kv_cache_clear(ctx);
 
         for (int j = 0; j < num_batches; ++j) {
             const int batch_start = start + j * n_batch;
@@ -350,7 +349,7 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par
 
             // add BOS token for the first batch of each chunk
             if (add_bos && j == 0) {
-                tokens[batch_start] = llama_token_bos(ctx);
+                tokens[batch_start] = llama_token_bos(llama_get_model(ctx));
             }
 
             if (llama_decode(ctx, llama_batch_get_one(tokens.data() + batch_start, batch_size, j * n_batch, 0))) {
@@ -573,7 +572,7 @@ static void hellaswag_score(llama_context * ctx, const gpt_params & params) {
         }
 
         // clear the KV cache
-        llama_kv_cache_tokens_rm(ctx, -1, -1);
+        llama_kv_cache_clear(ctx);
 
         auto logits = hellaswag_evaluate_tokens(ctx, query_embd, 0, params.n_batch, n_vocab);
         if (logits.empty()) {

examples/quantize-stats/CMakeLists.txt

@@ -1,6 +1,6 @@
 set(TARGET quantize-stats)
 add_executable(${TARGET} quantize-stats.cpp)
 install(TARGETS ${TARGET} RUNTIME)
-target_link_libraries(${TARGET} PRIVATE llama ${CMAKE_THREAD_LIBS_INIT})
+target_link_libraries(${TARGET} PRIVATE llama build_info ${CMAKE_THREAD_LIBS_INIT})
 target_include_directories(${TARGET} PRIVATE ../../common)
 target_compile_features(${TARGET} PRIVATE cxx_std_11)

examples/quantize-stats/quantize-stats.cpp

@@ -1,5 +1,4 @@
 #define LLAMA_API_INTERNAL
-#include "build-info.h"
 #include "common.h"
 #include "ggml.h"
 #include "llama.h"

examples/quantize/CMakeLists.txt

@@ -1,9 +1,6 @@
 set(TARGET quantize)
 add_executable(${TARGET} quantize.cpp)
 install(TARGETS ${TARGET} RUNTIME)
-target_link_libraries(${TARGET} PRIVATE llama ${CMAKE_THREAD_LIBS_INIT})
+target_link_libraries(${TARGET} PRIVATE llama build_info ${CMAKE_THREAD_LIBS_INIT})
 target_include_directories(${TARGET} PRIVATE ../../common)
 target_compile_features(${TARGET} PRIVATE cxx_std_11)
-if(TARGET BUILD_INFO)
-  add_dependencies(${TARGET} BUILD_INFO)
-endif()

examples/quantize/quantize.cpp

@@ -1,4 +1,3 @@
-#include "build-info.h"
 #include "common.h"
 #include "llama.h"
 
@@ -18,7 +17,6 @@ static const std::vector<struct quant_option> QUANT_OPTIONS = {
     { "Q4_1",   LLAMA_FTYPE_MOSTLY_Q4_1,   " 3.90G, +0.1585 ppl @ LLaMA-v1-7B", },
     { "Q5_0",   LLAMA_FTYPE_MOSTLY_Q5_0,   " 4.33G, +0.0683 ppl @ LLaMA-v1-7B", },
     { "Q5_1",   LLAMA_FTYPE_MOSTLY_Q5_1,   " 4.70G, +0.0349 ppl @ LLaMA-v1-7B", },
-#ifdef GGML_USE_K_QUANTS
     { "Q2_K",   LLAMA_FTYPE_MOSTLY_Q2_K,   " 2.63G, +0.6717 ppl @ LLaMA-v1-7B", },
     { "Q3_K",   LLAMA_FTYPE_MOSTLY_Q3_K_M, "alias for Q3_K_M" },
     { "Q3_K_S", LLAMA_FTYPE_MOSTLY_Q3_K_S, " 2.75G, +0.5551 ppl @ LLaMA-v1-7B", },
@@ -31,7 +29,6 @@ static const std::vector<struct quant_option> QUANT_OPTIONS = {
     { "Q5_K_S", LLAMA_FTYPE_MOSTLY_Q5_K_S, " 4.33G, +0.0400 ppl @ LLaMA-v1-7B", },
     { "Q5_K_M", LLAMA_FTYPE_MOSTLY_Q5_K_M, " 4.45G, +0.0122 ppl @ LLaMA-v1-7B", },
     { "Q6_K",   LLAMA_FTYPE_MOSTLY_Q6_K,   " 5.15G, -0.0008 ppl @ LLaMA-v1-7B", },
-#endif
     { "Q8_0",   LLAMA_FTYPE_MOSTLY_Q8_0,   " 6.70G, +0.0004 ppl @ LLaMA-v1-7B", },
     { "F16",    LLAMA_FTYPE_MOSTLY_F16,    "13.00G              @ 7B", },
     { "F32",    LLAMA_FTYPE_ALL_F32,       "26.00G              @ 7B", },
@@ -70,13 +67,14 @@ static bool try_parse_ftype(const std::string & ftype_str_in, llama_ftype & ftyp
 }
 
 // usage:
-//  ./quantize [--allow-requantize] [--leave-output-tensor] models/llama/ggml-model.gguf [models/llama/ggml-model-quant.gguf] type [nthreads]
+//  ./quantize [--allow-requantize] [--leave-output-tensor] [--pure] models/llama/ggml-model.gguf [models/llama/ggml-model-quant.gguf] type [nthreads]
 //
 [[noreturn]]
 static void usage(const char * executable) {
-    printf("usage: %s [--help] [--allow-requantize] [--leave-output-tensor] model-f32.gguf [model-quant.gguf] type [nthreads]\n\n", executable);
+    printf("usage: %s [--help] [--allow-requantize] [--leave-output-tensor] [--pure] model-f32.gguf [model-quant.gguf] type [nthreads]\n\n", executable);
     printf("  --allow-requantize: Allows requantizing tensors that have already been quantized. Warning: This can severely reduce quality compared to quantizing from 16bit or 32bit\n");
     printf("  --leave-output-tensor: Will leave output.weight un(re)quantized. Increases model size but may also increase quality, especially when requantizing\n");
+    printf("  --pure: Disable k-quant mixtures and quantize all tensors to the same type\n");
     printf("\nAllowed quantization types:\n");
     for (auto & it : QUANT_OPTIONS) {
         if (it.name != "COPY") {
@@ -103,6 +101,8 @@ int main(int argc, char ** argv) {
             params.quantize_output_tensor = false;
         } else if (strcmp(argv[arg_idx], "--allow-requantize") == 0) {
             params.allow_requantize = true;
+        } else if (strcmp(argv[arg_idx], "--pure") == 0) {
+            params.pure = true;
         } else {
             usage(argv[0]);
         }

examples/save-load-state/CMakeLists.txt

@@ -3,6 +3,3 @@ add_executable(${TARGET} save-load-state.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_11)
-if(TARGET BUILD_INFO)
-  add_dependencies(${TARGET} BUILD_INFO)
-endif()

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

@@ -1,4 +1,3 @@
-#include "build-info.h"
 #include "common.h"
 #include "llama.h"
 
@@ -8,10 +7,7 @@
 
 int main(int argc, char ** argv) {
     gpt_params params;
-    llama_sampling_params & sparams = params.sampling_params;
+
-    params.seed = 42;
-    params.n_threads = 4;
-    sparams.repeat_last_n = 64;
     params.prompt = "The quick brown fox";
 
     if (!gpt_params_parse(argc, argv, params)) {
@@ -25,56 +21,49 @@ int main(int argc, char ** argv) {
     }
 
     auto n_past = 0;
-    auto last_n_tokens_data = std::vector<llama_token>(sparams.repeat_last_n, 0);
+
+    std::string result0;
+    std::string result1;
 
     // init
     llama_model * model;
     llama_context * ctx;
 
-    std::tie(model, ctx) = llama_init_from_gpt_params( params );
+    std::tie(model, ctx) = llama_init_from_gpt_params(params);
-    if (model == nullptr) {
+    if (model == nullptr || ctx == nullptr) {
-        return 1;
+        fprintf(stderr, "%s : failed to init\n", __func__);
-    }
-    if (ctx == nullptr) {
-        llama_free_model(model);
         return 1;
     }
+
+    // tokenize prompt
     auto tokens = llama_tokenize(ctx, params.prompt, true);
-    auto n_prompt_tokens = tokens.size();
-    if (n_prompt_tokens < 1) {
-        fprintf(stderr, "%s : failed to tokenize prompt\n", __func__);
-        llama_free(ctx);
-        llama_free_model(model);
-        return 1;
-    }
 
     // evaluate prompt
-    llama_decode(ctx, llama_batch_get_one(tokens.data(), n_prompt_tokens, n_past, 0));
+    llama_decode(ctx, llama_batch_get_one(tokens.data(), tokens.size(), n_past, 0));
+    n_past += tokens.size();
 
-    last_n_tokens_data.insert(last_n_tokens_data.end(), tokens.data(), tokens.data() + n_prompt_tokens);
+    // save state (rng, logits, embedding and kv_cache) to file
-    n_past += n_prompt_tokens;
-
-    const size_t state_size = llama_get_state_size(ctx);
-    uint8_t * state_mem = new uint8_t[state_size];
-
-    // Save state (rng, logits, embedding and kv_cache) to file
     {
-        FILE *fp_write = fopen("dump_state.bin", "wb");
+        std::vector<uint8_t> state_mem(llama_get_state_size(ctx));
-        llama_copy_state_data(ctx, state_mem); // could also copy directly to memory mapped file
+
-        fwrite(state_mem, 1, state_size, fp_write);
+        {
-        fclose(fp_write);
+            FILE *fp_write = fopen("dump_state.bin", "wb");
+            llama_copy_state_data(ctx, state_mem.data()); // could also copy directly to memory mapped file
+            fwrite(state_mem.data(), 1, state_mem.size(), fp_write);
+            fclose(fp_write);
+        }
     }
 
     // save state (last tokens)
-    const auto last_n_tokens_data_saved = std::vector<llama_token>(last_n_tokens_data);
     const auto n_past_saved = n_past;
 
     // first run
-    printf("\n%s", params.prompt.c_str());
+    printf("\nfirst run: %s", params.prompt.c_str());
 
     for (auto i = 0; i < params.n_predict; i++) {
         auto * logits = llama_get_logits(ctx);
         auto n_vocab = llama_n_vocab(model);
+
         std::vector<llama_token_data> candidates;
         candidates.reserve(n_vocab);
         for (llama_token token_id = 0; token_id < n_vocab; token_id++) {
@@ -83,9 +72,10 @@ int main(int argc, char ** argv) {
         llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false };
         auto next_token = llama_sample_token(ctx, &candidates_p);
         auto next_token_str = llama_token_to_piece(ctx, next_token);
-        last_n_tokens_data.push_back(next_token);
 
         printf("%s", next_token_str.c_str());
+        result0 += next_token_str;
+
         if (llama_decode(ctx, llama_batch_get_one(&next_token, 1, n_past, 0))) {
             fprintf(stderr, "\n%s : failed to evaluate\n", __func__);
             llama_free(ctx);
@@ -103,32 +93,28 @@ int main(int argc, char ** argv) {
     // make new context
     auto * ctx2 = llama_new_context_with_model(model, llama_context_params_from_gpt_params(params));
 
-    // Load state (rng, logits, embedding and kv_cache) from file
+    printf("\nsecond run: %s", params.prompt.c_str());
-    {
-        FILE *fp_read = fopen("dump_state.bin", "rb");
-        if (state_size != llama_get_state_size(ctx2)) {
-            fprintf(stderr, "\n%s : failed to validate state size\n", __func__);
-            llama_free(ctx2);
-            llama_free_model(model);
-            return 1;
-        }
 
-        const size_t ret = fread(state_mem, 1, state_size, fp_read);
+    // load state (rng, logits, embedding and kv_cache) from file
-        if (ret != state_size) {
+    {
+        std::vector<uint8_t> state_mem(llama_get_state_size(ctx2));
+
+        FILE * fp_read = fopen("dump_state.bin", "rb");
+
+        const size_t ret = fread(state_mem.data(), 1, state_mem.size(), fp_read);
+        if (ret != state_mem.size()) {
             fprintf(stderr, "\n%s : failed to read state\n", __func__);
             llama_free(ctx2);
             llama_free_model(model);
             return 1;
         }
 
-        llama_set_state_data(ctx2, state_mem);  // could also read directly from memory mapped file
+        llama_set_state_data(ctx2, state_mem.data());
+
         fclose(fp_read);
     }
 
-    delete[] state_mem;
-
     // restore state (last tokens)
-    last_n_tokens_data = last_n_tokens_data_saved;
     n_past = n_past_saved;
 
     // second run
@@ -143,10 +129,11 @@ int main(int argc, char ** argv) {
         llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false };
         auto next_token = llama_sample_token(ctx2, &candidates_p);
         auto next_token_str = llama_token_to_piece(ctx2, next_token);
-        last_n_tokens_data.push_back(next_token);
 
         printf("%s", next_token_str.c_str());
-        if (llama_decode(ctx, llama_batch_get_one(&next_token, 1, n_past, 0))) {
+        result1 += next_token_str;
+
+        if (llama_decode(ctx2, llama_batch_get_one(&next_token, 1, n_past, 0))) {
             fprintf(stderr, "\n%s : failed to evaluate\n", __func__);
             llama_free(ctx2);
             llama_free_model(model);
@@ -155,10 +142,17 @@ int main(int argc, char ** argv) {
         n_past += 1;
     }
 
-    printf("\n\n");
+    printf("\n");
 
     llama_free(ctx2);
     llama_free_model(model);
 
+    if (result0 != result1) {
+        fprintf(stderr, "\n%s : error : the 2 generations are different\n", __func__);
+        return 1;
+    }
+
+    fprintf(stderr, "\n%s : success\n", __func__);
+
     return 0;
 }

examples/server/CMakeLists.txt

@@ -6,11 +6,8 @@ install(TARGETS ${TARGET} RUNTIME)
 target_compile_definitions(${TARGET} PRIVATE
     SERVER_VERBOSE=$<BOOL:${LLAMA_SERVER_VERBOSE}>
 )
-target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_link_libraries(${TARGET} PRIVATE common llama llava ${CMAKE_THREAD_LIBS_INIT})
 if (WIN32)
     TARGET_LINK_LIBRARIES(${TARGET} PRIVATE ws2_32)
 endif()
 target_compile_features(${TARGET} PRIVATE cxx_std_11)
-if(TARGET BUILD_INFO)
-  add_dependencies(${TARGET} BUILD_INFO)
-endif()

examples/server/README.md

@@ -24,6 +24,10 @@ Command line options:
 -   `--port`: Set the port to listen. Default: `8080`.
 -   `--path`: path from which to serve static files (default examples/server/public)
 -   `--embedding`: Enable embedding extraction, Default: disabled.
+-   `-np N`, `--parallel N`: Set the number of slots for process requests (default: 1)
+-   `-cb`, `--cont-batching`: enable continuous batching (a.k.a dynamic batching) (default: disabled)
+-   `-spf FNAME`, `--system-prompt-file FNAME` Set a file to load "a system prompt (initial prompt of all slots), this is useful for chat applications. [See more](#change-system-prompt-on-runtime)
+-   `--mmproj MMPROJ_FILE`: Path to a multimodal projector file for LLaVA.
 
 ## Build
 
@@ -106,25 +110,25 @@ node index.js
 
 ## API Endpoints
 
--   **POST** `/completion`: Given a prompt, it returns the predicted completion.
+-   **POST** `/completion`: Given a `prompt`, it returns the predicted completion.
 
     *Options:*
 
+    `prompt`: Provide the prompt for this completion as a string or as an array of strings or numbers representing tokens. Internally, the prompt is compared to the previous completion and only the "unseen" suffix is evaluated. If the prompt is a string or an array with the first element given as a string, a `bos` token is inserted in the front like `main` does.
+
     `temperature`: Adjust the randomness of the generated text (default: 0.8).
 
     `top_k`: Limit the next token selection to the K most probable tokens (default: 40).
 
     `top_p`: Limit the next token selection to a subset of tokens with a cumulative probability above a threshold P (default: 0.95).
 
-    `n_predict`: Set the number of tokens to predict when generating text. **Note:** May exceed the set limit slightly if the last token is a partial multibyte character. When 0, no tokens will be generated but the prompt is evaluated into the cache. (default: -1, -1 = infinity).
+    `n_predict`: Set the maximum number of tokens to predict when generating text. **Note:** May exceed the set limit slightly if the last token is a partial multibyte character. When 0, no tokens will be generated but the prompt is evaluated into the cache. (default: -1, -1 = infinity).
 
-    `n_keep`: Specify the number of tokens from the initial prompt to retain when the model resets its internal context.
+    `n_keep`: Specify the number of tokens from the prompt to retain when the context size is exceeded and tokens need to be discarded.
-    By default, this value is set to 0 (meaning no tokens are kept). Use `-1` to retain all tokens from the initial prompt.
+    By default, this value is set to 0 (meaning no tokens are kept). Use `-1` to retain all tokens from the prompt.
 
     `stream`: It allows receiving each predicted token in real-time instead of waiting for the completion to finish. To enable this, set to `true`.
 
-    `prompt`: Provide a prompt as a string, or as an array of strings and numbers representing tokens. Internally, the prompt is compared, and it detects if a part has already been evaluated, and the remaining part will be evaluate. If the prompt is a string, or an array with the first element given as a string, a space is inserted in the front like main.cpp does.
-
     `stop`: Specify a JSON array of stopping strings.
     These words will not be included in the completion, so make sure to add them to the prompt for the next iteration (default: []).
 
@@ -158,6 +162,44 @@ node index.js
 
     `n_probs`: If greater than 0, the response also contains the probabilities of top N tokens for each generated token (default: 0)
 
+    `image_data`: An array of objects to hold base64-encoded image `data` and its `id`s to be reference in `prompt`. You can determine the place of the image in the prompt as in the following: `USER:[img-12]Describe the image in detail.\nASSISTANT:` In this case, `[img-12]` will be replaced by the embeddings of the image id 12 in the following `image_data` array: `{..., "image_data": [{"data": "<BASE64_STRING>", "id": 12}]}`. Use `image_data` only with multimodal models, e.g., LLaVA.
+
+    *Result JSON:*
+
+    Note: When using streaming mode (`stream`) only `content` and `stop` will be returned until end of completion.
+
+    `content`: Completion result as a string (excluding `stopping_word` if any). In case of streaming mode, will contain the next token as a string.
+
+    `stop`: Boolean for use with `stream` to check whether the generation has stopped (Note: This is not related to stopping words array `stop` from input options)
+
+    `generation_settings`: The provided options above excluding `prompt` but including `n_ctx`, `model`
+
+    `model`: The path to the model loaded with `-m`
+
+    `prompt`: The provided `prompt`
+
+    `stopped_eos`: Indicating whether the completion has stopped because it encountered the EOS token
+
+    `stopped_limit`: Indicating whether the completion stopped because `n_predict` tokens were generated before stop words or EOS was encountered
+
+    `stopped_word`: Indicating whether the completion stopped due to encountering a stopping word from `stop` JSON array provided
+
+    `stopping_word`: The stopping word encountered which stopped the generation (or "" if not stopped due to a stopping word)
+
+    `timings`: Hash of timing information about the completion such as the number of tokens `predicted_per_second`
+
+    `tokens_cached`: Number of tokens from the prompt which could be re-used from previous completion (`n_past`)
+
+    `tokens_evaluated`: Number of tokens evaluated in total from the prompt
+
+    `truncated`: Boolean indicating if the context size was exceeded during generation, i.e. the number of tokens provided in the prompt (`tokens_evaluated`) plus tokens generated (`tokens predicted`) exceeded the context size (`n_ctx`)
+
+    `slot_id`: Assign the completion task to an specific slot. If is -1 the task will be assigned to a Idle slot (default: -1)
+
+    `cache_prompt`: Save the prompt and generation for avoid reprocess entire prompt if a part of this isn't change (default: false)
+
+    `system_prompt`: Change the system prompt (initial prompt of all slots), this is useful for chat applications. [See more](#change-system-prompt-on-runtime)
+
 -   **POST** `/tokenize`: Tokenize a given text.
 
     *Options:*
@@ -188,8 +230,32 @@ node index.js
 
     It also accepts all the options of `/completion` except `stream` and `prompt`.
 
+-   **GET** `/props`: Return the required assistant name and anti-prompt to generate the prompt in case you have specified a system prompt for all slots.
+
 ## More examples
 
+### Change system prompt on runtime
+
+To use the server example to serve multiple chat-type clients while keeping the same system prompt, you can utilize the option `system_prompt` to achieve that. This only needs to be done once to establish it.
+
+`prompt`: Specify a context that you want all connecting clients to respect.
+
+`anti_prompt`: Specify the word you want to use to instruct the model to stop. This must be sent to each client through the `/props` endpoint.
+
+`assistant_name`: The bot's name is necessary for each customer to generate the prompt. This must be sent to each client through the `/props` endpoint.
+
+```json
+{
+    "system_prompt": {
+        "prompt": "Transcript of a never ending dialog, where the User interacts with an Assistant.\nThe Assistant is helpful, kind, honest, good at writing, and never fails to answer the User's requests immediately and with precision.\nUser: Recommend a nice restaurant in the area.\nAssistant: I recommend the restaurant \"The Golden Duck\". It is a 5 star restaurant with a great view of the city. The food is delicious and the service is excellent. The prices are reasonable and the portions are generous. The restaurant is located at 123 Main Street, New York, NY 10001. The phone number is (212) 555-1234. The hours are Monday through Friday from 11:00 am to 10:00 pm. The restaurant is closed on Saturdays and Sundays.\nUser: Who is Richard Feynman?\nAssistant: Richard Feynman was an American physicist who is best known for his work in quantum mechanics and particle physics. He was awarded the Nobel Prize in Physics in 1965 for his contributions to the development of quantum electrodynamics. He was a popular lecturer and author, and he wrote several books, including \"Surely You're Joking, Mr. Feynman!\" and \"What Do You Care What Other People Think?\".\nUser:",
+        "anti_prompt": "User:",
+        "assistant_name": "Assistant:"
+    }
+}
+```
+
+**NOTE**: You can do this automatically when starting the server by simply creating a .json file with these options and using the CLI option `-spf FNAME` or `--system-prompt-file FNAME`.
+
 ### Interactive mode
 
 Check the sample in [chat.mjs](chat.mjs).

examples/server/api_like_OAI.py

@@ -8,6 +8,7 @@ import json
 
 
 app = Flask(__name__)
+slot_id = -1
 
 parser = argparse.ArgumentParser(description="An example of using server.cpp with a similar API to OAI. It must be used together with server.cpp.")
 parser.add_argument("--chat-prompt", type=str, help="the top prompt in chat completions(default: 'A chat between a curious user and an artificial intelligence assistant. The assistant follows the given rules no matter what.\\n')", default='A chat between a curious user and an artificial intelligence assistant. The assistant follows the given rules no matter what.\\n')
@@ -77,7 +78,8 @@ def make_postData(body, chat=False, stream=False):
     if(is_present(body, "stop")): postData["stop"] += body["stop"]
     postData["n_keep"] = -1
     postData["stream"] = stream
-
+    postData["cache_prompt"] = True
+    postData["slot_id"] = slot_id
     return postData
 
 def make_resData(data, chat=False, promptToken=[]):
@@ -128,6 +130,7 @@ def make_resData_stream(data, chat=False, time_now = 0, start=False):
             }
         ]
     }
+    slot_id = data["slot_id"]
     if (chat):
         if (start):
             resData["choices"][0]["delta"] =  {

examples/server/chat.mjs

@@ -7,6 +7,11 @@ const args = process.argv.slice(2);
 const grammarJsonSchemaFile = args.find(
     (_, index) => args[index - 1] === "--grammar-json-schema"
 );
+
+const no_cached_prompt = args.find(
+    (_, index) => args[index - 1] === "--no-cache-prompt"
+) ?? "false";
+
 const grammarFile = args.find((_, index) => args[index - 1] === "--grammar");
 
 // Example usage: function,arguments
@@ -30,6 +35,9 @@ if (grammarFile) {
     grammar = readFileSync(grammarFile, 'utf-8')
 }
 
+// for cached prompt
+let slot_id = -1;
+
 const API_URL = 'http://127.0.0.1:8080'
 
 const chat = [
@@ -76,6 +84,8 @@ async function chat_completion(question) {
             top_p: 0.9,
             n_keep: n_keep,
             n_predict: 256,
+            cache_prompt: no_cached_prompt === "false",
+            slot_id: slot_id,
             stop: ["\n### Human:"], // stop completion after generating this
             grammar,
             stream: true,
@@ -92,6 +102,7 @@ async function chat_completion(question) {
         const t = Buffer.from(chunk).toString('utf8')
         if (t.startsWith('data: ')) {
             const message = JSON.parse(t.substring(6))
+            slot_id = message.slot_id
             answer += message.content
             process.stdout.write(message.content)
             if (message.stop) {

examples/server/public/index.html

@@ -125,6 +125,7 @@
       background-color: #222;
       color: #ddd;
     }
+
     code {
       font-family: monospace;
       padding: 0.1em 0.3em;
@@ -141,7 +142,8 @@
       display: inline;
     }
 
-    header, footer {
+    header,
+    footer {
       text-align: center;
     }
 
@@ -163,6 +165,7 @@
       0% {
         background-position: 0%;
       }
+
       100% {
         background-position: 100%;
       }
@@ -181,6 +184,7 @@
         --loading-color-1: #22222200;
         --loading-color-2: #222222ff;
       }
+
       .popover-content {
         background-color: black;
       }
@@ -194,6 +198,8 @@
 
     import { llama } from '/completion.js';
     import { SchemaConverter } from '/json-schema-to-grammar.mjs';
+    let selected_image = false;
+    var slot_id = -1;
 
     const session = signal({
       prompt: "This is a conversation between User and Llama, a friendly chatbot. Llama is helpful, kind, honest, good at writing, and never fails to answer any requests immediately and with precision.",
@@ -203,6 +209,7 @@
       type: "chat",  // "chat" | "completion"
       char: "Llama",
       user: "User",
+      image_selected: ''
     })
 
     const params = signal({
@@ -220,7 +227,9 @@
       mirostat_tau: 5, // target entropy
       mirostat_eta: 0.1, // learning rate
       grammar: '',
-      n_probs: 0, // no completion_probabilities
+      n_probs: 0, // no completion_probabilities,
+      image_data: [],
+      cache_prompt: true
     })
 
     /* START: Support for storing prompt templates and parameters in borwser LocalStorage */
@@ -270,6 +279,7 @@
       // saved templates were successfuly imported.
 
       console.log('Processing saved templates and updating default template')
+      params.value = { ...params.value, image_data: [] };
 
       //console.log(importedTemplates);
       savedUserTemplates.value = importedTemplates;
@@ -294,7 +304,9 @@
 
     function userTemplateApply(t) {
       session.value = t.data.session;
+      session.value = { ...session.value, image_selected: '' };
       params.value = t.data.params;
+      params.value = { ...params.value, image_data: [] };
     }
 
     function userTemplateResetToDefaultAndApply() {
@@ -385,20 +397,25 @@
         throw new Error("already running");
       }
       controller.value = new AbortController();
-      for await (const chunk of llama(prompt, llamaParams, {controller: controller.value})) {
+      for await (const chunk of llama(prompt, llamaParams, { controller: controller.value })) {
         const data = chunk.data;
 
         if (data.stop) {
           while (
             currentMessages.length > 0 &&
             currentMessages[currentMessages.length - 1].content.match(/\n$/) != null
-            ) {
+          ) {
             currentMessages.pop();
           }
           transcriptUpdate([...history, [char, currentMessages]])
           console.log("Completion finished: '", currentMessages.map(msg => msg.content).join(''), "', summary: ", data);
         } else {
           currentMessages.push(data);
+          slot_id = data.slot_id;
+          if (selected_image && !data.multimodal) {
+            alert("The server was not compiled for multimodal or the model projector can't be loaded.");
+            return;
+          }
           transcriptUpdate([...history, [char, currentMessages]])
         }
 
@@ -419,7 +436,7 @@
 
       transcriptUpdate([...session.value.transcript, ["{{user}}", msg]])
 
-      const prompt = template(session.value.template, {
+      let prompt = template(session.value.template, {
         message: msg,
         history: session.value.transcript.flatMap(
           ([name, data]) =>
@@ -434,9 +451,12 @@
             )
         ).join("\n"),
       });
-
+      if (selected_image) {
+        prompt = `A chat between a curious human and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the human's questions.\nUSER:[img-10]${msg}\nASSISTANT:`;
+      }
       await runLlama(prompt, {
         ...params.value,
+        slot_id: slot_id,
         stop: ["</s>", template("{{char}}:"), template("{{user}}:")],
       }, "{{char}}");
     }
@@ -446,10 +466,11 @@
         console.log('already running...');
         return;
       }
-      const {prompt} = session.value;
+      const { prompt } = session.value;
       transcriptUpdate([...session.value.transcript, ["", prompt]]);
       await runLlama(prompt, {
         ...params.value,
+        slot_id: slot_id,
         stop: [],
       }, "");
     }
@@ -467,6 +488,27 @@
       transcriptUpdate([]);
     }
 
+    const uploadImage = (e) => {
+      e.preventDefault();
+      document.getElementById("fileInput").click();
+      document.getElementById("fileInput").addEventListener("change", function (event) {
+        const selectedFile = event.target.files[0];
+        if (selectedFile) {
+          const reader = new FileReader();
+          reader.onload = function () {
+            const image_data = reader.result;
+            session.value = { ...session.value, image_selected: image_data };
+            params.value = {
+              ...params.value, image_data: [
+                { data: image_data.replace(/data:image\/[^;]+;base64,/, ''), id: 10 }]
+            }
+          };
+          selected_image = true;
+          reader.readAsDataURL(selectedFile);
+        }
+      });
+    }
+
     function MessageInput() {
       const message = useSignal("")
 
@@ -497,6 +539,7 @@
           </div>
           <div class="right">
             <button type="submit" disabled=${generating.value}>Send</button>
+            <button onclick=${uploadImage}>Upload Image</button>
             <button onclick=${stop} disabled=${!generating.value}>Stop</button>
             <button onclick=${reset}>Reset</button>
           </div>
@@ -540,7 +583,7 @@
             data;
           message = html`<${Markdownish} text=${template(text)} />`
         }
-        if(user) {
+        if (user) {
           return html`<p key=${index}><strong>${template(user)}:</strong> ${message}</p>`
         } else {
           return html`<p key=${index}>${message}</p>`
@@ -549,6 +592,7 @@
 
       return html`
         <section id="chat" ref=${container}>
+          <img style="width: 60%;${!session.value.image_selected ? `display: none;` : ``}" src="${session.value.image_selected}"/>
           ${messages.flatMap(chatLine)}
         </section>`;
     };
@@ -567,7 +611,7 @@
           const converter = new SchemaConverter(
             grammarJsonSchemaPropOrder.value
               .split(',')
-              .reduce((acc, cur, i) => ({...acc, [cur.trim()]: i}), {})
+              .reduce((acc, cur, i) => ({ ...acc, [cur.trim()]: i }), {})
           )
           converter.visit(schema, '')
           params.value = {
@@ -579,7 +623,7 @@
         }
       }
 
-      const FloatField = ({label, max, min, name, step, value}) => {
+      const FloatField = ({ label, max, min, name, step, value }) => {
         return html`
           <div>
             <label for="${name}">${label}</label>
@@ -589,7 +633,7 @@
         `
       };
 
-      const IntField = ({label, max, min, name, value}) => {
+      const IntField = ({ label, max, min, name, value }) => {
         return html`
           <div>
             <label for="${name}">${label}</label>
@@ -672,7 +716,7 @@
             ${GrammarControl()}
           </fieldset>
       `
-    );
+      );
 
       const CompletionConfigForm = () => (
         html`
@@ -694,20 +738,20 @@
           ${session.value.type === 'chat' ? ChatConfigForm() : CompletionConfigForm()}
 
           <fieldset class="two">
-            ${IntField({label: "Predictions", max: 2048, min: -1, name: "n_predict", value: params.value.n_predict})}
+            ${IntField({ label: "Predictions", max: 2048, min: -1, name: "n_predict", value: params.value.n_predict })}
-            ${FloatField({label: "Temperature", max: 1.5, min: 0.0, name: "temperature", step: 0.01, value: params.value.temperature})}
+            ${FloatField({ label: "Temperature", max: 1.5, min: 0.0, name: "temperature", step: 0.01, value: params.value.temperature })}
-            ${FloatField({label: "Penalize repeat sequence", max: 2.0, min: 0.0, name: "repeat_penalty", step: 0.01, value: params.value.repeat_penalty})}
+            ${FloatField({ label: "Penalize repeat sequence", max: 2.0, min: 0.0, name: "repeat_penalty", step: 0.01, value: params.value.repeat_penalty })}
-            ${IntField({label: "Consider N tokens for penalize", max: 2048, min: 0, name: "repeat_last_n", value: params.value.repeat_last_n})}
+            ${IntField({ label: "Consider N tokens for penalize", max: 2048, min: 0, name: "repeat_last_n", value: params.value.repeat_last_n })}
-            ${IntField({label: "Top-K sampling", max: 100, min: -1, name: "top_k", value: params.value.top_k})}
+            ${IntField({ label: "Top-K sampling", max: 100, min: -1, name: "top_k", value: params.value.top_k })}
-            ${FloatField({label: "Top-P sampling", max: 1.0, min: 0.0, name: "top_p", step: 0.01, value: params.value.top_p})}
+            ${FloatField({ label: "Top-P sampling", max: 1.0, min: 0.0, name: "top_p", step: 0.01, value: params.value.top_p })}
           </fieldset>
           <details>
             <summary>More options</summary>
             <fieldset class="two">
-              ${FloatField({label: "TFS-Z", max: 1.0, min: 0.0, name: "tfs_z", step: 0.01, value: params.value.tfs_z})}
+              ${FloatField({ label: "TFS-Z", max: 1.0, min: 0.0, name: "tfs_z", step: 0.01, value: params.value.tfs_z })}
-              ${FloatField({label: "Typical P", max: 1.0, min: 0.0, name: "typical_p", step: 0.01, value: params.value.typical_p})}
+              ${FloatField({ label: "Typical P", max: 1.0, min: 0.0, name: "typical_p", step: 0.01, value: params.value.typical_p })}
-              ${FloatField({label: "Presence penalty", max: 1.0, min: 0.0, name: "presence_penalty", step: 0.01, value: params.value.presence_penalty})}
+              ${FloatField({ label: "Presence penalty", max: 1.0, min: 0.0, name: "presence_penalty", step: 0.01, value: params.value.presence_penalty })}
-              ${FloatField({label: "Frequency penalty", max: 1.0, min: 0.0, name: "frequency_penalty", step: 0.01, value: params.value.frequency_penalty})}
+              ${FloatField({ label: "Frequency penalty", max: 1.0, min: 0.0, name: "frequency_penalty", step: 0.01, value: params.value.frequency_penalty })}
             </fieldset>
             <hr />
             <fieldset class="three">
@@ -716,11 +760,11 @@
                 <label><input type="radio" name="mirostat" value="1" checked=${params.value.mirostat == 1} oninput=${updateParamsInt} /> Mirostat v1</label>
                 <label><input type="radio" name="mirostat" value="2" checked=${params.value.mirostat == 2} oninput=${updateParamsInt} /> Mirostat v2</label>
               </div>
-              ${FloatField({label: "Mirostat tau", max: 10.0, min: 0.0, name: "mirostat_tau", step: 0.01, value: params.value.mirostat_tau})}
+              ${FloatField({ label: "Mirostat tau", max: 10.0, min: 0.0, name: "mirostat_tau", step: 0.01, value: params.value.mirostat_tau })}
-              ${FloatField({label: "Mirostat eta", max: 1.0, min: 0.0, name: "mirostat_eta", step: 0.01, value: params.value.mirostat_eta})}
+              ${FloatField({ label: "Mirostat eta", max: 1.0, min: 0.0, name: "mirostat_eta", step: 0.01, value: params.value.mirostat_eta })}
             </fieldset>
             <fieldset>
-              ${IntField({label: "Show Probabilities", max: 10, min: 0, name: "n_probs", value: params.value.n_probs})}
+              ${IntField({ label: "Show Probabilities", max: 10, min: 0, name: "n_probs", value: params.value.n_probs })}
             </fieldset>
           </details>
         </form>
@@ -759,20 +803,20 @@
         const popoverChildren = html`
           <div class="prob-set">
             ${probs.map((p, index) => {
-              return html`
+          return html`
                 <div
                   key=${index}
                   title=${`prob: ${p.prob}`}
                   style=${{
-                    padding: '0.3em',
+              padding: '0.3em',
-                    backgroundColor: p.tok_str === content ? probColor(p.prob) : 'transparent'
+              backgroundColor: p.tok_str === content ? probColor(p.prob) : 'transparent'
-                  }}
+            }}
                 >
                   <span>${p.tok_str}: </span>
                   <span>${Math.floor(p.prob * 100)}%</span>
                 </div>
               `
-            })}
+        })}
           </div>
         `
 
@@ -851,9 +895,9 @@
               ref=${popoverRef}
               class="popover-content"
               style=${{
-                top: position.value.top,
+            top: position.value.top,
-                left: position.value.left,
+            left: position.value.left,
-              }}
+          }}
             >
               ${props.popoverChildren}
             </div>
@@ -952,8 +996,11 @@
 </head>
 
 <body>
-  <div id="container"></div>
+  <div id="container">
+    <input type="file" id="fileInput" accept="image/*" style="display: none;">
+  </div>
   <div id="portal"></div>
 </body>
 
 </html>
+

examples/simple/simple.cpp

@@ -92,16 +92,11 @@ int main(int argc, char ** argv) {
     // create a llama_batch with size 512
     // we use this object to submit token data for decoding
 
-    llama_batch batch = llama_batch_init(512, 0);
+    llama_batch batch = llama_batch_init(512, 0, 1);
 
     // evaluate the initial prompt
-    batch.n_tokens = tokens_list.size();
+    for (size_t i = 0; i < tokens_list.size(); i++) {
-
+        llama_batch_add(batch, tokens_list[i], i, { 0 }, false);
-    for (int32_t i = 0; i < batch.n_tokens; i++) {
-        batch.token[i]  = tokens_list[i];
-        batch.pos[i]    = i;
-        batch.seq_id[i] = 0;
-        batch.logits[i] = false;
     }
 
     // llama_decode will output logits only for the last token of the prompt
@@ -138,7 +133,7 @@ int main(int argc, char ** argv) {
             const llama_token new_token_id = llama_sample_token_greedy(ctx, &candidates_p);
 
             // is it an end of stream?
-            if (new_token_id == llama_token_eos(ctx) || n_cur == n_len) {
+            if (new_token_id == llama_token_eos(model) || n_cur == n_len) {
                 LOG_TEE("\n");
 
                 break;
@@ -148,15 +143,10 @@ int main(int argc, char ** argv) {
             fflush(stdout);
 
             // prepare the next batch
-            batch.n_tokens = 0;
+            llama_batch_clear(batch);
 
             // push this new token for next evaluation
-            batch.token [batch.n_tokens] = new_token_id;
+            llama_batch_add(batch, new_token_id, n_cur, { 0 }, true);
-            batch.pos   [batch.n_tokens] = n_cur;
-            batch.seq_id[batch.n_tokens] = 0;
-            batch.logits[batch.n_tokens] = true;
-
-            batch.n_tokens += 1;
 
             n_decode += 1;
         }

examples/speculative/CMakeLists.txt

@@ -3,6 +3,3 @@ add_executable(${TARGET} speculative.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_11)
-if(TARGET BUILD_INFO)
-  add_dependencies(${TARGET} BUILD_INFO)
-endif()

examples/speculative/speculative.cpp

@@ -1,14 +1,27 @@
-#include "build-info.h"
-
 #include "common.h"
 #include "llama.h"
-#include "grammar-parser.h"
 
 #include <cmath>
 #include <cstdio>
 #include <string>
 #include <vector>
 
+#define SPEC_VOCAB_MAX_SIZE_DIFFERENCE  100
+#define SPEC_VOCAB_CHECK_START_TOKEN_ID 5
+
+struct seq_draft {
+    bool active   = false;
+    bool drafting = false;
+    bool skip     = false;
+
+    int i_batch_dft = 0;
+    std::vector<int> i_batch_tgt;
+
+    std::vector<llama_token> tokens;
+
+    struct llama_sampling_context * ctx_sampling;
+};
+
 int main(int argc, char ** argv) {
     gpt_params params;
 
@@ -21,6 +34,13 @@ int main(int argc, char ** argv) {
         return 1;
     }
 
+    // max number of parallel drafting sequences (i.e. tree branches)
+    const int n_seq_dft = params.n_parallel;
+
+    // TODO: make this configurable
+    const float p_accept = 0.80f;
+    const float p_split  = 0.10f;
+
 #ifndef LOG_DISABLE_LOGS
     log_set_target(log_filename_generator("speculative", "log"));
     LOG_TEE("Log start\n");
@@ -45,6 +65,33 @@ int main(int argc, char ** argv) {
     params.n_gpu_layers = params.n_gpu_layers_draft;
     std::tie(model_dft, ctx_dft) = llama_init_from_gpt_params(params);
 
+    {
+        const int n_vocab_tgt = llama_n_vocab(model_tgt);
+        const int n_vocab_dft = llama_n_vocab(model_dft);
+        const int vocab_diff  = n_vocab_tgt > n_vocab_dft
+            ? n_vocab_tgt - n_vocab_dft
+            : n_vocab_dft - n_vocab_tgt;
+
+        if (vocab_diff > SPEC_VOCAB_MAX_SIZE_DIFFERENCE) {
+            fprintf(stderr, "%s: error: draft model vocab must closely match target model to use speculation but ", __func__);
+            fprintf(stderr, "target vocab size %d does not match draft vocab size %d - difference %d, max allowed %d\n",
+                    n_vocab_tgt, llama_n_vocab(model_dft), vocab_diff, SPEC_VOCAB_MAX_SIZE_DIFFERENCE);
+            return 1;
+        }
+
+        for (int i = SPEC_VOCAB_CHECK_START_TOKEN_ID; i < std::min(n_vocab_tgt, n_vocab_dft); ++i) {
+            const char * token_text_tgt = llama_token_get_text(model_tgt, i);
+            const char * token_text_dft = llama_token_get_text(model_dft, i);
+            if (std::strcmp(token_text_tgt, token_text_dft) != 0) {
+                fprintf(stderr, "%s: error: draft model vocab must match target model to use speculation but ", __func__);
+                fprintf(stderr, "token %d content differs - target '%s', draft '%s'\n", i,
+                        llama_token_to_piece(ctx_tgt, i).c_str(),
+                        llama_token_to_piece(ctx_dft, i).c_str());
+                return 1;
+            }
+        }
+    }
+
     // tokenize the prompt
     std::vector<llama_token> inp;
     inp = ::llama_tokenize(ctx_tgt, params.prompt, true);
@@ -77,8 +124,6 @@ int main(int argc, char ** argv) {
     const auto t_enc_end = ggml_time_us();
 
     // the 2 models should have the same vocab
-    const int n_ctx   = llama_n_ctx(ctx_tgt);
-    const int n_vocab = llama_n_vocab(model_tgt);
     //GGML_ASSERT(n_vocab == llama_n_vocab(model_dft));
 
     // how many tokens to draft each time
@@ -91,116 +136,129 @@ int main(int argc, char ** argv) {
     int n_past_tgt = inp.size();
     int n_past_dft = inp.size();
 
-    std::vector<llama_token> drafted;
-
-    std::vector<llama_token> last_tokens(n_ctx);
-    std::fill(last_tokens.begin(), last_tokens.end(), 0);
-
-    for (auto & id : inp) {
-        last_tokens.erase(last_tokens.begin());
-        last_tokens.push_back(id);
-    }
-
-    std::vector<llama_token_data> candidates;
-    candidates.reserve(n_vocab);
-
     // used to determine end of generation
     bool has_eos = false;
 
-    // grammar stuff
+    // target model sampling context
-    struct llama_grammar * grammar_dft = NULL;
+    struct llama_sampling_context * ctx_sampling = llama_sampling_init(params.sparams);
-    struct llama_grammar * grammar_tgt = NULL;
 
-    grammar_parser::parse_state parsed_grammar;
+    // draft sequence data
+    std::vector<seq_draft> drafts(n_seq_dft);
 
-    // if requested - load the grammar, error checking is omitted for brevity
+    params.sparams.grammar.clear(); // the draft samplers will copy the target sampler's grammar
-    if (!params.grammar.empty()) {
+    params.sparams.temp = -1.0f;    // force greedy sampling with probs for the draft model
-        parsed_grammar = grammar_parser::parse(params.grammar.c_str());
-        // will be empty (default) if there are parse errors
-        if (parsed_grammar.rules.empty()) {
-            return 1;
-        }
 
-        std::vector<const llama_grammar_element *> grammar_rules(parsed_grammar.c_rules());
+    for (int s = 0; s < n_seq_dft; ++s) {
-        grammar_tgt = llama_grammar_init(grammar_rules.data(), grammar_rules.size(), parsed_grammar.symbol_ids.at("root"));
+        drafts[s].ctx_sampling = llama_sampling_init(params.sparams);
     }
 
-    llama_sampling_context ctx_sampling = llama_sampling_context_init(params, grammar_tgt);
+    llama_batch batch_dft = llama_batch_init(params.n_ctx, 0, 1);
+    llama_batch batch_tgt = llama_batch_init(params.n_ctx, 0, n_seq_dft);
 
     const auto t_dec_start = ggml_time_us();
 
-    while (true) {
+    // sample from the last token of the prompt
-        LOG("drafted: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx_dft, drafted));
+    drafts[0].i_batch_tgt.resize(1);
+    drafts[0].i_batch_tgt[0] = 0;
 
-        int i_dft = 0;
+    while (true) {
+        // print current draft sequences
+        for (int s = 0; s < n_seq_dft; ++s) {
+            if (!drafts[s].active) {
+                continue;
+            }
+
+            const auto & tokens = drafts[s].tokens;
+
+            LOG("draft %d: %s\n", s, LOG_TOKENS_TOSTR_PRETTY(ctx_dft, tokens).c_str());
+        }
+
+        int i_dft  = 0;
+        int s_keep = 0;
 
         while (true) {
+            LOG("sampling target: s_keep = %3d, i_dft = %3d, i_batch_tgt = %3d\n", s_keep, i_dft, drafts[s_keep].i_batch_tgt[i_dft]);
+
             // sample from the target model
-            llama_token id = llama_sampling_sample(ctx_tgt, NULL, ctx_sampling, last_tokens, candidates, i_dft);
+            llama_token id = llama_sampling_sample(ctx_sampling, ctx_tgt, NULL, drafts[s_keep].i_batch_tgt[i_dft]);
 
-            // remember which tokens were sampled - used for repetition penalties during sampling
+            llama_sampling_accept(ctx_sampling, ctx_tgt, id, true);
-            last_tokens.erase(last_tokens.begin());
-            last_tokens.push_back(id);
 
-            //LOG("last: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx_tgt, last_tokens));
+            //LOG("last: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx_tgt, ctx_sampling->prev).c_str());
 
             const std::string token_str = llama_token_to_piece(ctx_tgt, id);
+
             printf("%s", token_str.c_str());
             fflush(stdout);
 
-            if (id == llama_token_eos(ctx_tgt)) {
+            if (id == llama_token_eos(model_tgt)) {
                 has_eos = true;
             }
 
             ++n_predict;
 
-            // check if the draft matches the target
+            // check if the target token matches any of the drafts
-            if (i_dft < (int) drafted.size() && id == drafted[i_dft]) {
-                LOG("the sampled target token matches the %dth drafted token (%d, '%s') - accepted\n", i_dft, id, token_str.c_str());
-                ++n_accept;
-                ++n_past_tgt;
-                ++n_past_dft;
-                ++i_dft;
-
-                continue;
-            }
-
-            // the drafted token was rejected or we are out of drafted tokens
-
-            if (i_dft < (int) drafted.size()) {
-                LOG("the %dth drafted token (%d, '%s') does not match the sampled target token (%d, '%s') - rejected\n",
-                        i_dft, drafted[i_dft], llama_token_to_piece(ctx_dft, drafted[i_dft]).c_str(), id, token_str.c_str());
-            } else {
-                LOG("out of drafted tokens\n");
-            }
-
-            llama_kv_cache_seq_rm(ctx_dft, 0, n_past_dft, -1);
-            llama_decode(ctx_dft, llama_batch_get_one(&id, 1, n_past_dft, 0));
-            ++n_past_dft;
-
-            // heuristic for n_draft
             {
-                const int  n_draft_cur  = (int) drafted.size();
+                bool matches = false;
-                const bool all_accepted = i_dft == n_draft_cur;
 
-                LOG("n_draft      = %d\n", n_draft);
+                for (int s = 0; s < n_seq_dft; ++s) {
-                LOG("n_draft_cur  = %d\n", n_draft_cur);
+                    if (!drafts[s].active) {
-                LOG("i_dft        = %d\n", i_dft);
+                        continue;
-                LOG("all_accepted = %d\n", all_accepted);
+                    }
 
-                if (all_accepted && n_draft == n_draft_cur) {
+                    if (i_dft < (int) drafts[s].tokens.size() && id == drafts[s].tokens[i_dft]) {
-                    LOG(" - max drafted tokens accepted - n_draft += 8\n");
+                        LOG("the sampled target token matches the %dth drafted token of sequence %d (%d, '%s') - accepted\n", i_dft, s, id, token_str.c_str());
-                    n_draft = std::min(30, n_draft + 8);
+
-                } else if (all_accepted) {
+                        s_keep = s;
-                    LOG(" - partially drafted tokens accepted - no change\n");
+                        matches = true;
-                } else {
+                    } else {
-                    LOG(" - drafted token rejected - n_draft -= 1\n");
+                        drafts[s].active = false;
-                    n_draft = std::max(2, n_draft - 1);
+                    }
+                }
+
+                if (matches) {
+                    ++n_accept;
+                    ++n_past_tgt;
+                    ++n_past_dft;
+                    ++i_dft;
+
+                    continue;
                 }
             }
 
-            drafted.clear();
+            LOG("the sampled target token (%d, '%s') did not match, or we ran out of drafted tokens\n", id, token_str.c_str());
-            drafted.push_back(id);
+
+            // TODO: simplify
+            {
+                LOG("keeping sequence %d, n_past_tgt = %d, n_past_dft = %d\n", s_keep, n_past_tgt, n_past_dft);
+
+                llama_kv_cache_seq_keep(ctx_dft, s_keep);
+                llama_kv_cache_seq_cp  (ctx_dft, s_keep, 0, -1, -1);
+                llama_kv_cache_seq_keep(ctx_dft, 0);
+
+                llama_kv_cache_seq_rm  (ctx_tgt, s_keep, n_past_tgt, -1);
+                llama_kv_cache_seq_keep(ctx_tgt, s_keep);
+                llama_kv_cache_seq_cp  (ctx_tgt, s_keep, 0, -1, -1);
+                llama_kv_cache_seq_keep(ctx_tgt, 0);
+            }
+
+            for (int s = 0; s < n_seq_dft; ++s) {
+                drafts[s].active = false;
+                drafts[s].tokens.clear();
+                drafts[s].i_batch_tgt.clear();
+            }
+            // note: will be erased after the speculation phase
+            drafts[0].tokens.push_back(id);
+            drafts[0].i_batch_tgt.push_back(0);
+
+            llama_batch_clear(batch_dft);
+            llama_batch_add  (batch_dft, id, n_past_dft, { 0 }, true);
+
+            llama_kv_cache_seq_rm(ctx_dft, 0, n_past_dft, -1);
+            // LOG("dft batch: %s\n", LOG_BATCH_TOSTR_PRETTY(ctx_dft, batch_dft).c_str());
+            llama_decode         (ctx_dft, batch_dft);
+
+            ++n_past_dft;
 
             break;
         }
@@ -209,78 +267,151 @@ int main(int argc, char ** argv) {
             break;
         }
 
-        if (grammar_tgt) {
+        llama_sampling_cp(ctx_sampling, drafts[0].ctx_sampling);
-            if (grammar_dft) {
-                llama_grammar_free(grammar_dft);
-            }
-            // Note: Hardcoded to sequence id 0, if this ever supports parallel generation
-            //       that will need to change.
-            auto it = ctx_sampling.sequence_contexts.find(0);
-            GGML_ASSERT(it != ctx_sampling.sequence_contexts.end());
-            // This is necessary because each sequence id in sequence_contexts
-            // uses a copy of the original grammar.
-            grammar_dft = llama_grammar_copy(it->second.grammar);
 
-            LOG("copied target grammar to draft grammar\n");
+        int n_seq_cur  = 1;
-        }
-
-        // sample n_draft tokens from the draft model using greedy decoding
         int n_past_cur = n_past_dft;
+
+        for (int s = 0; s < n_seq_dft; ++s) {
+            drafts[s].active   = false;
+            drafts[s].drafting = false;
+        }
+        drafts[0].active      = true;
+        drafts[0].drafting    = true;
+        drafts[0].i_batch_dft = 0;
+
+        llama_batch_clear(batch_tgt);
+        llama_batch_add  (batch_tgt, drafts[0].tokens[0], n_past_tgt, { 0 }, true);
+
+        // sample n_draft tokens from the draft model using tree-based sampling
         for (int i = 0; i < n_draft; ++i) {
-            float * logits = llama_get_logits(ctx_dft);
+            batch_dft.n_tokens = 0;
 
-            candidates.clear();
+            for (int s = 0; s < n_seq_dft; ++s) {
-            for (llama_token token_id = 0; token_id < n_vocab; token_id++) {
+                drafts[s].skip = false;
-                candidates.emplace_back(llama_token_data{token_id, logits[token_id], 0.0f});
             }
 
-            llama_token_data_array cur_p = { candidates.data(), candidates.size(), false };
+            for (int s = 0; s < n_seq_dft; ++s) {
+                if (!drafts[s].drafting || drafts[s].skip) {
+                    continue;
+                }
 
-            if (grammar_dft != NULL) {
+                llama_sampling_sample(drafts[s].ctx_sampling, ctx_dft, NULL, drafts[s].i_batch_dft);
-                llama_sample_grammar(ctx_dft, &cur_p, grammar_dft);
+
+                const auto & cur_p = drafts[s].ctx_sampling->cur;
+
+                for (int k = 0; k < std::min(n_seq_dft + 3, (int) cur_p.size()); ++k) {
+                    LOG(" - draft candidate %3d for seq %3d, pos %3d: %6d (%8.3f) '%s'\n",
+                            k, s, i, cur_p[k].id, cur_p[k].p, llama_token_to_piece(ctx_dft, cur_p[k].id).c_str());
+                }
+
+                if (cur_p[0].p < p_accept) {
+                    LOG("stopping drafting for seq %3d, probability too low: %.3f < %.3f\n", s, cur_p[0].p, p_accept);
+                    drafts[s].drafting = false;
+                    continue;
+                }
+
+                std::vector<int> sa(1, s);
+
+                // attempt to split the branch if the probability is high enough
+                for (int f = 1; f < 8; ++f) {
+                    if (n_seq_cur < n_seq_dft && cur_p[f].p > p_split) {
+                        LOG("splitting seq %3d into %3d\n", s, n_seq_cur);
+
+                        llama_kv_cache_seq_rm(ctx_dft,    n_seq_cur, -1, -1);
+                        llama_kv_cache_seq_cp(ctx_dft, s, n_seq_cur, -1, -1);
+
+                        // all previous tokens from this branch are now also part of the new branch
+                        for (int t = 0; t < batch_tgt.n_tokens; ++t) {
+                            for (int p = 0; p < batch_tgt.n_seq_id[t]; ++p) {
+                                if (batch_tgt.seq_id[t][p] == s) {
+                                    batch_tgt.seq_id[t][batch_tgt.n_seq_id[t]] = n_seq_cur;
+                                    batch_tgt.n_seq_id[t]++;
+                                    break;
+                                }
+                            }
+                        }
+
+                        // copy the draft state
+                        drafts[n_seq_cur].active   = true;
+                        drafts[n_seq_cur].drafting = true;
+                        drafts[n_seq_cur].skip     = true;
+
+                        drafts[n_seq_cur].tokens      = drafts[s].tokens;
+                        drafts[n_seq_cur].i_batch_dft = drafts[s].i_batch_dft;
+                        drafts[n_seq_cur].i_batch_tgt = drafts[s].i_batch_tgt;
+
+                        llama_sampling_cp(drafts[s].ctx_sampling, drafts[n_seq_cur].ctx_sampling);
+
+                        sa.push_back(n_seq_cur);
+
+                        n_seq_cur++;
+                    } else {
+                        break;
+                    }
+                }
+
+                // add drafted token for each sequence
+                for (int is = 0; is < (int) sa.size(); ++is) {
+                    const llama_token id = cur_p[is].id;
+
+                    const int s = sa[is];
+
+                    llama_sampling_accept(drafts[s].ctx_sampling, ctx_dft, id, true);
+
+                    drafts[s].tokens.push_back(id);
+
+                    // add unique drafted tokens to the target batch
+                    drafts[s].i_batch_tgt.push_back(batch_tgt.n_tokens);
+
+                    llama_batch_add(batch_tgt, id, n_past_tgt + i + 1, { s }, true);
+
+                    // add the token to the batch for batched decoding with the draft model
+                    drafts[s].i_batch_dft = batch_dft.n_tokens;
+
+                    llama_batch_add(batch_dft, id, n_past_cur, { s }, true);
+
+                    if (batch_tgt.n_tokens > n_draft) {
+                        drafts[s].drafting = false;
+                    }
+                }
             }
 
-            // computes softmax and sorts the candidates
+            // no sequence is drafting anymore
-            llama_sample_softmax(ctx_dft, &cur_p);
+            if (batch_dft.n_tokens == 0) {
-
-            for (int i = 0; i < 3; ++i) {
-                LOG(" - draft candidate %3d: %6d (%8.3f) '%s'\n", i, cur_p.data[i].id, cur_p.data[i].p, llama_token_to_piece(ctx_dft, cur_p.data[i].id).c_str());
-            }
-
-            // TODO: better logic?
-            if (cur_p.data[0].p < 2*cur_p.data[1].p) {
-                LOG("stopping drafting, probability too low: %.3f < 2*%.3f\n", cur_p.data[0].p, cur_p.data[1].p);
                 break;
             }
 
-            // drafted token
+            // evaluate the drafted tokens on the draft model
-            const llama_token id = cur_p.data[0].id;
+            llama_decode(ctx_dft, batch_dft);
-
+            ++n_past_cur;
-            drafted.push_back(id);
             ++n_drafted;
 
-            // no need to evaluate the last drafted token, since we won't use the result
+            if (batch_tgt.n_tokens > n_draft) {
-            if (i == n_draft - 1) {
                 break;
             }
-
-            // evaluate the drafted token on the draft model
-            llama_kv_cache_seq_rm(ctx_dft, 0, n_past_cur, -1);
-            llama_decode(ctx_dft, llama_batch_get_one(&drafted.back(), 1, n_past_cur, 0));
-            ++n_past_cur;
-
-            if (grammar_dft != NULL) {
-                llama_grammar_accept_token(ctx_dft, grammar_dft, id);
-            }
         }
 
         // evaluate the target model on the drafted tokens
-        llama_kv_cache_seq_rm(ctx_tgt, 0, n_past_tgt, -1);
+        {
-        llama_decode(ctx_tgt, llama_batch_get_one(drafted.data(), drafted.size(), n_past_tgt, 0));
+            llama_kv_cache_seq_keep(ctx_tgt, 0);
-        ++n_past_tgt;
+            for (int s = 1; s < n_seq_dft; ++s) {
+                llama_kv_cache_seq_cp(ctx_tgt, 0, s, -1, -1);
+            }
 
-        // the first token is always proposed by the traget model before the speculation loop
+            // LOG("target batch: %s\n", LOG_BATCH_TOSTR_PRETTY(ctx_tgt, batch_tgt).c_str());
-        drafted.erase(drafted.begin());
+            llama_decode(ctx_tgt, batch_tgt);
+            ++n_past_tgt;
+        }
+
+        // the first token is always proposed by the traget model before the speculation loop so we erase it here
+        for (int s = 0; s < n_seq_dft; ++s) {
+            if (!drafts[s].active) {
+                continue;
+            }
+
+            drafts[s].tokens.erase(drafts[s].tokens.begin());
+        }
     }
 
     auto t_dec_end = ggml_time_us();
@@ -288,9 +419,8 @@ int main(int argc, char ** argv) {
     LOG_TEE("\n\n");
 
     LOG_TEE("encoded %4d tokens in %8.3f seconds, speed: %8.3f t/s\n", n_input,   (t_enc_end - t_enc_start) / 1e6f, inp.size() / ((t_enc_end - t_enc_start) / 1e6f));
-    LOG_TEE("decoded %4d tokens in %8.3f seconds, speed: %8.3f t/s\n", n_predict, (t_dec_end - t_dec_start) / 1e6f, n_predict / ((t_dec_end - t_dec_start) / 1e6f));
+    LOG_TEE("decoded %4d tokens in %8.3f seconds, speed: %8.3f t/s\n", n_predict, (t_dec_end - t_dec_start) / 1e6f, n_predict  / ((t_dec_end - t_dec_start) / 1e6f));
 
-    // TODO: make sure these numbers are computed correctly
     LOG_TEE("\n");
     LOG_TEE("n_draft   = %d\n", n_draft);
     LOG_TEE("n_predict = %d\n", n_predict);
@@ -304,16 +434,19 @@ int main(int argc, char ** argv) {
     LOG_TEE("\ntarget:\n");
     llama_print_timings(ctx_tgt);
 
+    llama_sampling_free(ctx_sampling);
+    for (int s = 0; s < n_seq_dft; ++s) {
+        llama_sampling_free(drafts[s].ctx_sampling);
+    }
+
+    llama_batch_free(batch_dft);
+
     llama_free(ctx_tgt);
     llama_free_model(model_tgt);
 
     llama_free(ctx_dft);
     llama_free_model(model_dft);
 
-    if (grammar_dft != NULL) {
-        llama_grammar_free(grammar_dft);
-        llama_grammar_free(grammar_tgt);
-    }
     llama_backend_free();
 
     fprintf(stderr, "\n\n");

examples/train-text-from-scratch/train-text-from-scratch.cpp

@@ -253,13 +253,14 @@ static void init_model(struct my_llama_model * model) {
     set_param_model(model);
 
     // measure data size
-    struct ggml_allocr * alloc = NULL;
+    size_t size = 0;
-    alloc = ggml_allocr_new_measure(tensor_alignment);
+    for (struct ggml_tensor * t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) {
-    alloc_model(alloc, model);
+        size += GGML_PAD(ggml_nbytes(t), tensor_alignment);
+    }
 
     // allocate data
-    model->data.resize(ggml_allocr_max_size(alloc) + tensor_alignment);
+    struct ggml_allocr * alloc = NULL;
-    ggml_allocr_free(alloc);
+    model->data.resize(size + tensor_alignment);
     alloc = ggml_allocr_new(model->data.data(), model->data.size(), tensor_alignment);
     alloc_model(alloc, model);
     ggml_allocr_free(alloc);
@@ -348,9 +349,9 @@ static struct ggml_tensor * llama_build_train_graphs(
         // not capturing these, to silcence warnings
         const int rope_mode = 0;
 
-        return ggml_rope_custom(ctx,
+        return ggml_rope_custom(
-            t, KQ_pos, n_rot, rope_mode, n_ctx,
+            ctx, t, KQ_pos, n_rot, rope_mode, n_ctx, 0, rope_freq_base, rope_freq_scale, 0.0f, 1.0f, 0.0f, 0.0f
-            rope_freq_base, rope_freq_scale);
+        );
     };
 
     set_name(tokens_input, "tokens_input");
@@ -1094,11 +1095,9 @@ int main(int argc, char ** argv) {
     struct ggml_tensor * target_probs  = ggml_new_tensor_3d(ctx_input, GGML_TYPE_F32, n_vocab,  n_tokens, n_batch);
 
     // measure required memory for input tensors
-    alloc = ggml_allocr_new_measure(tensor_alignment);
+    size_t max_input_size = GGML_PAD(ggml_nbytes(tokens_input), tensor_alignment) +
-    ggml_allocr_alloc(alloc, tokens_input);
+                            GGML_PAD(ggml_nbytes(target_probs), tensor_alignment) +
-    ggml_allocr_alloc(alloc, target_probs);
+                            tensor_alignment;
-    size_t max_input_size = ggml_allocr_max_size(alloc) + tensor_alignment;
-    ggml_allocr_free(alloc);
     printf("%s: input_size = %zu bytes (%.1f MB)\n", __func__, max_input_size, (float) max_input_size / (1024.0f*1024.0f));
 
     // allocate input tensors

flake.lock

@@ -5,11 +5,11 @@
         "systems": "systems"
       },
       "locked": {
-        "lastModified": 1692799911,
+        "lastModified": 1694529238,
-        "narHash": "sha256-3eihraek4qL744EvQXsK1Ha6C3CR7nnT8X2qWap4RNk=",
+        "narHash": "sha256-zsNZZGTGnMOf9YpHKJqMSsa0dXbfmxeoJ7xHlrt+xmY=",
         "owner": "numtide",
         "repo": "flake-utils",
-        "rev": "f9e7cf818399d17d347f847525c5a5a8032e4e44",
+        "rev": "ff7b65b44d01cf9ba6a71320833626af21126384",
         "type": "github"
       },
       "original": {
@@ -20,11 +20,11 @@
     },
     "nixpkgs": {
       "locked": {
-        "lastModified": 1692913444,
+        "lastModified": 1698318101,
-        "narHash": "sha256-1SvMQm2DwofNxXVtNWWtIcTh7GctEVrS/Xel/mdc6iY=",
+        "narHash": "sha256-gUihHt3yPD7bVqg+k/UVHgngyaJ3DMEBchbymBMvK1E=",
         "owner": "NixOS",
         "repo": "nixpkgs",
-        "rev": "18324978d632ffc55ef1d928e81630c620f4f447",
+        "rev": "63678e9f3d3afecfeafa0acead6239cdb447574c",
         "type": "github"
       },
       "original": {

flake.nix

@@ -11,8 +11,7 @@
         meta.mainProgram = "llama";
         inherit (pkgs.stdenv) isAarch32 isAarch64 isDarwin;
         buildInputs = with pkgs; [ openmpi ];
-        osSpecific = with pkgs; buildInputs ++
+        osSpecific = with pkgs; buildInputs ++ (
-        (
           if isAarch64 && isDarwin then
             with pkgs.darwin.apple_sdk_11_0.frameworks; [
               Accelerate
@@ -51,6 +50,9 @@
         };
         llama-python =
           pkgs.python3.withPackages (ps: with ps; [ numpy sentencepiece ]);
+        # TODO(Green-Sky): find a better way to opt-into the heavy ml python runtime
+        llama-python-extra =
+          pkgs.python3.withPackages (ps: with ps; [ numpy sentencepiece torchWithoutCuda transformers ]);
         postPatch = ''
           substituteInPlace ./ggml-metal.m \
             --replace '[bundle pathForResource:@"ggml-metal" ofType:@"metal"];' "@\"$out/bin/ggml-metal.metal\";"
@@ -93,12 +95,15 @@
         };
         packages.rocm = pkgs.stdenv.mkDerivation {
           inherit name src meta postPatch nativeBuildInputs postInstall;
-          buildInputs = with pkgs; buildInputs ++ [ hip hipblas rocblas ];
+          buildInputs = with pkgs.rocmPackages; buildInputs ++ [ clr hipblas rocblas ];
           cmakeFlags = cmakeFlags ++ [
             "-DLLAMA_HIPBLAS=1"
             "-DCMAKE_C_COMPILER=hipcc"
             "-DCMAKE_CXX_COMPILER=hipcc"
-            "-DCMAKE_POSITION_INDEPENDENT_CODE=ON"
+            # Build all targets supported by rocBLAS. When updating search for TARGET_LIST_ROCM
+            # in github.com/ROCmSoftwarePlatform/rocBLAS/blob/develop/CMakeLists.txt
+            # and select the line that matches the current nixpkgs version of rocBLAS.
+            "-DAMDGPU_TARGETS=gfx803;gfx900;gfx906:xnack-;gfx908:xnack-;gfx90a:xnack+;gfx90a:xnack-;gfx940;gfx941;gfx942;gfx1010;gfx1012;gfx1030;gfx1100;gfx1101;gfx1102"
           ];
         };
         apps.llama-server = {
@@ -126,5 +131,9 @@
           buildInputs = [ llama-python ];
           packages = nativeBuildInputs ++ osSpecific;
         };
+        devShells.extra = pkgs.mkShell {
+          buildInputs = [ llama-python-extra ];
+          packages = nativeBuildInputs ++ osSpecific;
+        };
       });
 }

ggml-impl.h

@@ -0,0 +1,237 @@
+#pragma once
+
+#include "ggml.h"
+
+// GGML internal header
+
+#include <assert.h>
+#include <stddef.h>
+#include <stdbool.h>
+#include <string.h> // memcpy
+#include <math.h>   // fabsf
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+// static_assert should be a #define, but if it's not,
+// fall back to the _Static_assert C11 keyword.
+// if C99 - static_assert is noop
+// ref: https://stackoverflow.com/a/53923785/4039976
+#ifndef static_assert
+#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201100L)
+#define static_assert(cond, msg) _Static_assert(cond, msg)
+#else
+#define static_assert(cond, msg) struct global_scope_noop_trick
+#endif
+#endif
+
+// __FMA__ and __F16C__ are not defined in MSVC, however they are implied with AVX2/AVX512
+#if defined(_MSC_VER) && (defined(__AVX2__) || defined(__AVX512F__))
+#ifndef __FMA__
+#define __FMA__
+#endif
+#ifndef __F16C__
+#define __F16C__
+#endif
+#ifndef __SSE3__
+#define __SSE3__
+#endif
+#endif
+
+#undef MIN
+#undef MAX
+
+#define MIN(a, b) ((a) < (b) ? (a) : (b))
+#define MAX(a, b) ((a) > (b) ? (a) : (b))
+
+// 16-bit float
+// on Arm, we use __fp16
+// on x86, we use uint16_t
+#if defined(__ARM_NEON) && !defined(_MSC_VER)
+
+// if YCM cannot find <arm_neon.h>, make a symbolic link to it, for example:
+//
+//   $ ln -sfn /Library/Developer/CommandLineTools/usr/lib/clang/13.1.6/include/arm_neon.h ./src/
+//
+#include <arm_neon.h>
+
+#define GGML_COMPUTE_FP16_TO_FP32(x) ((float) (x))
+#define GGML_COMPUTE_FP32_TO_FP16(x) (x)
+
+#define GGML_FP16_TO_FP32(x) ((float) (x))
+#define GGML_FP32_TO_FP16(x) (x)
+
+#else
+
+#ifdef __wasm_simd128__
+#include <wasm_simd128.h>
+#else
+#ifdef __POWER9_VECTOR__
+#include <altivec.h>
+#undef bool
+#define bool _Bool
+#else
+#if defined(_MSC_VER) || defined(__MINGW32__)
+#include <intrin.h>
+#else
+#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__) || defined(__SSSE3__) || defined(__SSE3__)
+#if !defined(__riscv)
+#include <immintrin.h>
+#endif
+#endif
+#endif
+#endif
+#endif
+
+#ifdef __riscv_v_intrinsic
+#include <riscv_vector.h>
+#endif
+
+#ifdef __F16C__
+
+#ifdef _MSC_VER
+#define GGML_COMPUTE_FP16_TO_FP32(x) _mm_cvtss_f32(_mm_cvtph_ps(_mm_cvtsi32_si128(x)))
+#define GGML_COMPUTE_FP32_TO_FP16(x) _mm_extract_epi16(_mm_cvtps_ph(_mm_set_ss(x), 0), 0)
+#else
+#define GGML_COMPUTE_FP16_TO_FP32(x) _cvtsh_ss(x)
+#define GGML_COMPUTE_FP32_TO_FP16(x) _cvtss_sh(x, 0)
+#endif
+
+#elif defined(__POWER9_VECTOR__)
+
+#define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
+#define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x)
+/* the inline asm below is about 12% faster than the lookup method */
+#define GGML_FP16_TO_FP32(x) GGML_COMPUTE_FP16_TO_FP32(x)
+#define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x)
+
+static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
+    register float f;
+    register double d;
+    __asm__(
+        "mtfprd %0,%2\n"
+        "xscvhpdp %0,%0\n"
+        "frsp %1,%0\n" :
+        /* temp */ "=d"(d),
+        /* out */  "=f"(f):
+        /* in */   "r"(h));
+    return f;
+}
+
+static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
+    register double d;
+    register ggml_fp16_t r;
+    __asm__( /* xscvdphp can work on double or single precision */
+        "xscvdphp %0,%2\n"
+        "mffprd %1,%0\n" :
+        /* temp */ "=d"(d),
+        /* out */  "=r"(r):
+        /* in */   "f"(f));
+    return r;
+}
+
+#else
+
+// FP16 <-> FP32
+// ref: https://github.com/Maratyszcza/FP16
+
+static inline float fp32_from_bits(uint32_t w) {
+    union {
+        uint32_t as_bits;
+        float as_value;
+    } fp32;
+    fp32.as_bits = w;
+    return fp32.as_value;
+}
+
+static inline uint32_t fp32_to_bits(float f) {
+    union {
+        float as_value;
+        uint32_t as_bits;
+    } fp32;
+    fp32.as_value = f;
+    return fp32.as_bits;
+}
+
+static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
+    const uint32_t w = (uint32_t) h << 16;
+    const uint32_t sign = w & UINT32_C(0x80000000);
+    const uint32_t two_w = w + w;
+
+    const uint32_t exp_offset = UINT32_C(0xE0) << 23;
+#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) || defined(__GNUC__) && !defined(__STRICT_ANSI__)
+    const float exp_scale = 0x1.0p-112f;
+#else
+    const float exp_scale = fp32_from_bits(UINT32_C(0x7800000));
+#endif
+    const float normalized_value = fp32_from_bits((two_w >> 4) + exp_offset) * exp_scale;
+
+    const uint32_t magic_mask = UINT32_C(126) << 23;
+    const float magic_bias = 0.5f;
+    const float denormalized_value = fp32_from_bits((two_w >> 17) | magic_mask) - magic_bias;
+
+    const uint32_t denormalized_cutoff = UINT32_C(1) << 27;
+    const uint32_t result = sign |
+        (two_w < denormalized_cutoff ? fp32_to_bits(denormalized_value) : fp32_to_bits(normalized_value));
+    return fp32_from_bits(result);
+}
+
+static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
+#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) || defined(__GNUC__) && !defined(__STRICT_ANSI__)
+    const float scale_to_inf = 0x1.0p+112f;
+    const float scale_to_zero = 0x1.0p-110f;
+#else
+    const float scale_to_inf = fp32_from_bits(UINT32_C(0x77800000));
+    const float scale_to_zero = fp32_from_bits(UINT32_C(0x08800000));
+#endif
+    float base = (fabsf(f) * scale_to_inf) * scale_to_zero;
+
+    const uint32_t w = fp32_to_bits(f);
+    const uint32_t shl1_w = w + w;
+    const uint32_t sign = w & UINT32_C(0x80000000);
+    uint32_t bias = shl1_w & UINT32_C(0xFF000000);
+    if (bias < UINT32_C(0x71000000)) {
+        bias = UINT32_C(0x71000000);
+    }
+
+    base = fp32_from_bits((bias >> 1) + UINT32_C(0x07800000)) + base;
+    const uint32_t bits = fp32_to_bits(base);
+    const uint32_t exp_bits = (bits >> 13) & UINT32_C(0x00007C00);
+    const uint32_t mantissa_bits = bits & UINT32_C(0x00000FFF);
+    const uint32_t nonsign = exp_bits + mantissa_bits;
+    return (sign >> 16) | (shl1_w > UINT32_C(0xFF000000) ? UINT16_C(0x7E00) : nonsign);
+}
+
+#define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
+#define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x)
+
+#endif // __F16C__
+
+#endif // __ARM_NEON
+
+// precomputed f32 table for f16 (256 KB)
+// defined in ggml.c, initialized in ggml_init()
+extern float ggml_table_f32_f16[1 << 16];
+
+// On ARM NEON, it's quicker to directly convert x -> x instead of calling into ggml_lookup_fp16_to_fp32,
+// so we define GGML_FP16_TO_FP32 and GGML_FP32_TO_FP16 elsewhere for NEON.
+// This is also true for POWER9.
+#if !defined(GGML_FP16_TO_FP32) || !defined(GGML_FP32_TO_FP16)
+
+inline static float ggml_lookup_fp16_to_fp32(ggml_fp16_t f) {
+    uint16_t s;
+    memcpy(&s, &f, sizeof(uint16_t));
+    return ggml_table_f32_f16[s];
+}
+
+#define GGML_FP16_TO_FP32(x) ggml_lookup_fp16_to_fp32(x)
+#define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x)
+
+#endif
+
+    // TODO: backend v2 PR
+
+#ifdef __cplusplus
+}
+#endif

ggml-metal.m

@@ -62,6 +62,7 @@ struct ggml_metal_context {
     GGML_METAL_DECL_KERNEL(mul);
     GGML_METAL_DECL_KERNEL(mul_row); // TODO: avoid this extra kernel, instead extend the "mul" kernel to support broadcast
     GGML_METAL_DECL_KERNEL(scale);
+    GGML_METAL_DECL_KERNEL(scale_4);
     GGML_METAL_DECL_KERNEL(silu);
     GGML_METAL_DECL_KERNEL(relu);
     GGML_METAL_DECL_KERNEL(gelu);
@@ -73,6 +74,8 @@ struct ggml_metal_context {
     GGML_METAL_DECL_KERNEL(get_rows_f16);
     GGML_METAL_DECL_KERNEL(get_rows_q4_0);
     GGML_METAL_DECL_KERNEL(get_rows_q4_1);
+    GGML_METAL_DECL_KERNEL(get_rows_q5_0);
+    GGML_METAL_DECL_KERNEL(get_rows_q5_1);
     GGML_METAL_DECL_KERNEL(get_rows_q8_0);
     GGML_METAL_DECL_KERNEL(get_rows_q2_K);
     GGML_METAL_DECL_KERNEL(get_rows_q3_K);
@@ -87,6 +90,8 @@ struct ggml_metal_context {
     GGML_METAL_DECL_KERNEL(mul_mv_f16_f32_l4);
     GGML_METAL_DECL_KERNEL(mul_mv_q4_0_f32);
     GGML_METAL_DECL_KERNEL(mul_mv_q4_1_f32);
+    GGML_METAL_DECL_KERNEL(mul_mv_q5_0_f32);
+    GGML_METAL_DECL_KERNEL(mul_mv_q5_1_f32);
     GGML_METAL_DECL_KERNEL(mul_mv_q8_0_f32);
     GGML_METAL_DECL_KERNEL(mul_mv_q2_K_f32);
     GGML_METAL_DECL_KERNEL(mul_mv_q3_K_f32);
@@ -97,6 +102,8 @@ struct ggml_metal_context {
     GGML_METAL_DECL_KERNEL(mul_mm_f16_f32);
     GGML_METAL_DECL_KERNEL(mul_mm_q4_0_f32);
     GGML_METAL_DECL_KERNEL(mul_mm_q4_1_f32);
+    GGML_METAL_DECL_KERNEL(mul_mm_q5_0_f32);
+    GGML_METAL_DECL_KERNEL(mul_mm_q5_1_f32);
     GGML_METAL_DECL_KERNEL(mul_mm_q8_0_f32);
     GGML_METAL_DECL_KERNEL(mul_mm_q2_K_f32);
     GGML_METAL_DECL_KERNEL(mul_mm_q3_K_f32);
@@ -203,6 +210,10 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
             GGML_METAL_LOG_INFO("%s: default.metallib not found, loading from source\n", __func__);
 
             NSString * sourcePath = [bundle pathForResource:@"ggml-metal" ofType:@"metal"];
+            if (sourcePath == nil) {
+                GGML_METAL_LOG_WARN("%s: error: could not use bundle path to find ggml-metal.metal, falling back to trying cwd\n", __func__);
+                sourcePath = @"ggml-metal.metal";
+            }
             GGML_METAL_LOG_INFO("%s: loading '%s'\n", __func__, [sourcePath UTF8String]);
             NSString * src = [NSString stringWithContentsOfFile:sourcePath encoding:NSUTF8StringEncoding error:&error];
             if (error) {
@@ -227,14 +238,17 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
     // load kernels
     {
         NSError * error = nil;
-#define GGML_METAL_ADD_KERNEL(name) \
+
-        ctx->function_##name = [ctx->library newFunctionWithName:@"kernel_"#name]; \
+        /*
-        ctx->pipeline_##name = [ctx->device newComputePipelineStateWithFunction:ctx->function_##name error:&error]; \
         GGML_METAL_LOG_INFO("%s: loaded %-32s %16p | th_max = %4d | th_width = %4d\n", __func__, "kernel_"#name, (void *) ctx->pipeline_##name, \
                 (int) ctx->pipeline_##name.maxTotalThreadsPerThreadgroup, \
                 (int) ctx->pipeline_##name.threadExecutionWidth); \
+        */
+#define GGML_METAL_ADD_KERNEL(name) \
+        ctx->function_##name = [ctx->library newFunctionWithName:@"kernel_"#name]; \
+        ctx->pipeline_##name = [ctx->device newComputePipelineStateWithFunction:ctx->function_##name error:&error]; \
         if (error) { \
-          GGML_METAL_LOG_ERROR("%s: error: load pipeline error: %s\n", __func__, [[error description] UTF8String]); \
+            GGML_METAL_LOG_ERROR("%s: error: load pipeline error: %s\n", __func__, [[error description] UTF8String]); \
             return NULL; \
         }
 
@@ -243,6 +257,7 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
         GGML_METAL_ADD_KERNEL(mul);
         GGML_METAL_ADD_KERNEL(mul_row);
         GGML_METAL_ADD_KERNEL(scale);
+        GGML_METAL_ADD_KERNEL(scale_4);
         GGML_METAL_ADD_KERNEL(silu);
         GGML_METAL_ADD_KERNEL(relu);
         GGML_METAL_ADD_KERNEL(gelu);
@@ -254,6 +269,8 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
         GGML_METAL_ADD_KERNEL(get_rows_f16);
         GGML_METAL_ADD_KERNEL(get_rows_q4_0);
         GGML_METAL_ADD_KERNEL(get_rows_q4_1);
+        GGML_METAL_ADD_KERNEL(get_rows_q5_0);
+        GGML_METAL_ADD_KERNEL(get_rows_q5_1);
         GGML_METAL_ADD_KERNEL(get_rows_q8_0);
         GGML_METAL_ADD_KERNEL(get_rows_q2_K);
         GGML_METAL_ADD_KERNEL(get_rows_q3_K);
@@ -268,6 +285,8 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
         GGML_METAL_ADD_KERNEL(mul_mv_f16_f32_l4);
         GGML_METAL_ADD_KERNEL(mul_mv_q4_0_f32);
         GGML_METAL_ADD_KERNEL(mul_mv_q4_1_f32);
+        GGML_METAL_ADD_KERNEL(mul_mv_q5_0_f32);
+        GGML_METAL_ADD_KERNEL(mul_mv_q5_1_f32);
         GGML_METAL_ADD_KERNEL(mul_mv_q8_0_f32);
         GGML_METAL_ADD_KERNEL(mul_mv_q2_K_f32);
         GGML_METAL_ADD_KERNEL(mul_mv_q3_K_f32);
@@ -278,8 +297,10 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
             GGML_METAL_ADD_KERNEL(mul_mm_f32_f32);
             GGML_METAL_ADD_KERNEL(mul_mm_f16_f32);
             GGML_METAL_ADD_KERNEL(mul_mm_q4_0_f32);
-            GGML_METAL_ADD_KERNEL(mul_mm_q8_0_f32);
             GGML_METAL_ADD_KERNEL(mul_mm_q4_1_f32);
+            GGML_METAL_ADD_KERNEL(mul_mm_q5_0_f32);
+            GGML_METAL_ADD_KERNEL(mul_mm_q5_1_f32);
+            GGML_METAL_ADD_KERNEL(mul_mm_q8_0_f32);
             GGML_METAL_ADD_KERNEL(mul_mm_q2_K_f32);
             GGML_METAL_ADD_KERNEL(mul_mm_q3_K_f32);
             GGML_METAL_ADD_KERNEL(mul_mm_q4_K_f32);
@@ -335,6 +356,7 @@ void ggml_metal_free(struct ggml_metal_context * ctx) {
     GGML_METAL_DEL_KERNEL(mul);
     GGML_METAL_DEL_KERNEL(mul_row);
     GGML_METAL_DEL_KERNEL(scale);
+    GGML_METAL_DEL_KERNEL(scale_4);
     GGML_METAL_DEL_KERNEL(silu);
     GGML_METAL_DEL_KERNEL(relu);
     GGML_METAL_DEL_KERNEL(gelu);
@@ -346,6 +368,8 @@ void ggml_metal_free(struct ggml_metal_context * ctx) {
     GGML_METAL_DEL_KERNEL(get_rows_f16);
     GGML_METAL_DEL_KERNEL(get_rows_q4_0);
     GGML_METAL_DEL_KERNEL(get_rows_q4_1);
+    GGML_METAL_DEL_KERNEL(get_rows_q5_0);
+    GGML_METAL_DEL_KERNEL(get_rows_q5_1);
     GGML_METAL_DEL_KERNEL(get_rows_q8_0);
     GGML_METAL_DEL_KERNEL(get_rows_q2_K);
     GGML_METAL_DEL_KERNEL(get_rows_q3_K);
@@ -360,6 +384,8 @@ void ggml_metal_free(struct ggml_metal_context * ctx) {
     GGML_METAL_DEL_KERNEL(mul_mv_f16_f32_l4);
     GGML_METAL_DEL_KERNEL(mul_mv_q4_0_f32);
     GGML_METAL_DEL_KERNEL(mul_mv_q4_1_f32);
+    GGML_METAL_DEL_KERNEL(mul_mv_q5_0_f32);
+    GGML_METAL_DEL_KERNEL(mul_mv_q5_1_f32);
     GGML_METAL_DEL_KERNEL(mul_mv_q8_0_f32);
     GGML_METAL_DEL_KERNEL(mul_mv_q2_K_f32);
     GGML_METAL_DEL_KERNEL(mul_mv_q3_K_f32);
@@ -370,8 +396,10 @@ void ggml_metal_free(struct ggml_metal_context * ctx) {
         GGML_METAL_DEL_KERNEL(mul_mm_f32_f32);
         GGML_METAL_DEL_KERNEL(mul_mm_f16_f32);
         GGML_METAL_DEL_KERNEL(mul_mm_q4_0_f32);
-        GGML_METAL_DEL_KERNEL(mul_mm_q8_0_f32);
         GGML_METAL_DEL_KERNEL(mul_mm_q4_1_f32);
+        GGML_METAL_DEL_KERNEL(mul_mm_q5_0_f32);
+        GGML_METAL_DEL_KERNEL(mul_mm_q5_1_f32);
+        GGML_METAL_DEL_KERNEL(mul_mm_q8_0_f32);
         GGML_METAL_DEL_KERNEL(mul_mm_q2_K_f32);
         GGML_METAL_DEL_KERNEL(mul_mm_q3_K_f32);
         GGML_METAL_DEL_KERNEL(mul_mm_q4_K_f32);
@@ -905,15 +933,20 @@ void ggml_metal_graph_compute(
 
                             const float scale = *(const float *) src1->data;
 
-                            [encoder setComputePipelineState:ctx->pipeline_scale];
+                            int64_t n = ggml_nelements(dst);
+
+                            if (n % 4 == 0) {
+                                n /= 4;
+                                [encoder setComputePipelineState:ctx->pipeline_scale_4];
+                            } else {
+                                [encoder setComputePipelineState:ctx->pipeline_scale];
+                            }
+
                             [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
                             [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
                             [encoder setBytes:&scale length:sizeof(scale) atIndex:2];
 
-                            const int64_t n = ggml_nelements(dst);
+                            [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
-                            GGML_ASSERT(n % 4 == 0);
-
-                            [encoder dispatchThreadgroups:MTLSizeMake(n/4, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
                         } break;
                     case GGML_OP_UNARY:
                         switch (ggml_get_unary_op(gf->nodes[i])) {
@@ -968,11 +1001,15 @@ void ggml_metal_graph_compute(
                         } break;
                     case GGML_OP_SOFT_MAX:
                         {
-                            const int nth = MIN(32, ne00);
+                            int nth = 32; // SIMD width
 
                             if (ne00%4 == 0) {
                                 [encoder setComputePipelineState:ctx->pipeline_soft_max_4];
                             } else {
+                                do {
+                                    nth *= 2;
+                                } while (nth <= ne00 && nth <= 1024);
+                                nth /= 2;
                                 [encoder setComputePipelineState:ctx->pipeline_soft_max];
                             }
                             [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
@@ -980,8 +1017,9 @@ void ggml_metal_graph_compute(
                             [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:2];
                             [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:3];
                             [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:4];
+                            [encoder setThreadgroupMemoryLength:nth/32*sizeof(float) atIndex:0];
 
-                            [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
+                            [encoder dispatchThreadgroups:MTLSizeMake(ne01*ne02*ne03, 1, 1) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
                         } break;
                     case GGML_OP_DIAG_MASK_INF:
                         {
@@ -1052,6 +1090,8 @@ void ggml_metal_graph_compute(
                                     case GGML_TYPE_F16:  [encoder setComputePipelineState:ctx->pipeline_mul_mm_f16_f32];  break;
                                     case GGML_TYPE_Q4_0: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q4_0_f32]; break;
                                     case GGML_TYPE_Q4_1: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q4_1_f32]; break;
+                                    case GGML_TYPE_Q5_0: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q5_0_f32]; break;
+                                    case GGML_TYPE_Q5_1: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q5_1_f32]; break;
                                     case GGML_TYPE_Q8_0: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q8_0_f32]; break;
                                     case GGML_TYPE_Q2_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q2_K_f32]; break;
                                     case GGML_TYPE_Q3_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q3_K_f32]; break;
@@ -1121,6 +1161,24 @@ void ggml_metal_graph_compute(
                                             nth1 = 8;
                                             [encoder setComputePipelineState:ctx->pipeline_mul_mv_q4_1_f32];
                                         } break;
+                                    case GGML_TYPE_Q5_0:
+                                        {
+                                            GGML_ASSERT(ne02 == 1);
+                                            GGML_ASSERT(ne12 == 1);
+
+                                            nth0 = 8;
+                                            nth1 = 8;
+                                            [encoder setComputePipelineState:ctx->pipeline_mul_mv_q5_0_f32];
+                                        } break;
+                                    case GGML_TYPE_Q5_1:
+                                        {
+                                            GGML_ASSERT(ne02 == 1);
+                                            GGML_ASSERT(ne12 == 1);
+
+                                            nth0 = 8;
+                                            nth1 = 8;
+                                            [encoder setComputePipelineState:ctx->pipeline_mul_mv_q5_1_f32];
+                                        } break;
                                     case GGML_TYPE_Q8_0:
                                         {
                                             GGML_ASSERT(ne02 == 1);
@@ -1201,7 +1259,8 @@ void ggml_metal_graph_compute(
                                 [encoder setBytes:&ne1  length:sizeof(ne1)  atIndex:16];
                                 [encoder setBytes:&gqa  length:sizeof(gqa)  atIndex:17];
 
-                                if (src0t == GGML_TYPE_Q4_0 || src0t == GGML_TYPE_Q4_1 || src0t == GGML_TYPE_Q8_0 ||
+                                if (src0t == GGML_TYPE_Q4_0 || src0t == GGML_TYPE_Q4_1 ||
+                                    src0t == GGML_TYPE_Q5_0 || src0t == GGML_TYPE_Q5_1 || src0t == GGML_TYPE_Q8_0 ||
                                     src0t == GGML_TYPE_Q2_K) { // || src0t == GGML_TYPE_Q4_K) {
                                     [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
                                 }
@@ -1233,6 +1292,8 @@ void ggml_metal_graph_compute(
                                 case GGML_TYPE_F16:  [encoder setComputePipelineState:ctx->pipeline_get_rows_f16];  break;
                                 case GGML_TYPE_Q4_0: [encoder setComputePipelineState:ctx->pipeline_get_rows_q4_0]; break;
                                 case GGML_TYPE_Q4_1: [encoder setComputePipelineState:ctx->pipeline_get_rows_q4_1]; break;
+                                case GGML_TYPE_Q5_0: [encoder setComputePipelineState:ctx->pipeline_get_rows_q5_0]; break;
+                                case GGML_TYPE_Q5_1: [encoder setComputePipelineState:ctx->pipeline_get_rows_q5_1]; break;
                                 case GGML_TYPE_Q8_0: [encoder setComputePipelineState:ctx->pipeline_get_rows_q8_0]; break;
                                 case GGML_TYPE_Q2_K: [encoder setComputePipelineState:ctx->pipeline_get_rows_q2_K]; break;
                                 case GGML_TYPE_Q3_K: [encoder setComputePipelineState:ctx->pipeline_get_rows_q3_K]; break;
@@ -1339,14 +1400,18 @@ void ggml_metal_graph_compute(
 
                             const int nth = MIN(1024, ne00);
 
-                            const int n_past = ((int32_t *) dst->op_params)[0];
+                            const int n_past     = ((int32_t *) dst->op_params)[0];
-                            const int n_dims = ((int32_t *) dst->op_params)[1];
+                            const int n_dims     = ((int32_t *) dst->op_params)[1];
-                            const int mode   = ((int32_t *) dst->op_params)[2];
+                            const int mode       = ((int32_t *) dst->op_params)[2];
+                            const int n_orig_ctx = ((int32_t *) dst->op_params)[3];
 
-                            float freq_base;
+                            float freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow;
-                            float freq_scale;
+                            memcpy(&freq_base,   (int32_t *) dst->op_params +  5, sizeof(float));
-                            memcpy(&freq_base,  (int32_t *) dst->op_params + 4, sizeof(float));
+                            memcpy(&freq_scale,  (int32_t *) dst->op_params +  6, sizeof(float));
-                            memcpy(&freq_scale, (int32_t *) dst->op_params + 5, sizeof(float));
+                            memcpy(&ext_factor,  (int32_t *) dst->op_params +  7, sizeof(float));
+                            memcpy(&attn_factor, (int32_t *) dst->op_params +  8, sizeof(float));
+                            memcpy(&beta_fast,   (int32_t *) dst->op_params +  9, sizeof(float));
+                            memcpy(&beta_slow,   (int32_t *) dst->op_params + 10, sizeof(float));
 
                             switch (src0->type) {
                                 case GGML_TYPE_F32: [encoder setComputePipelineState:ctx->pipeline_rope_f32]; break;
@@ -1354,30 +1419,35 @@ void ggml_metal_graph_compute(
                                 default: GGML_ASSERT(false);
                             };
 
-                            [encoder setBuffer:id_src0 offset:offs_src0        atIndex:0];
+                            [encoder setBuffer:id_src0     offset:offs_src0        atIndex:0];
-                            [encoder setBuffer:id_src1 offset:offs_src1        atIndex:1];
+                            [encoder setBuffer:id_src1     offset:offs_src1        atIndex:1];
-                            [encoder setBuffer:id_dst  offset:offs_dst         atIndex:2];
+                            [encoder setBuffer:id_dst      offset:offs_dst         atIndex:2];
-                            [encoder setBytes:&ne00    length:sizeof( int64_t) atIndex:3];
+                            [encoder setBytes:&ne00        length:sizeof( int64_t) atIndex:3];
-                            [encoder setBytes:&ne01    length:sizeof( int64_t) atIndex:4];
+                            [encoder setBytes:&ne01        length:sizeof( int64_t) atIndex:4];
-                            [encoder setBytes:&ne02    length:sizeof( int64_t) atIndex:5];
+                            [encoder setBytes:&ne02        length:sizeof( int64_t) atIndex:5];
-                            [encoder setBytes:&ne03    length:sizeof( int64_t) atIndex:6];
+                            [encoder setBytes:&ne03        length:sizeof( int64_t) atIndex:6];
-                            [encoder setBytes:&nb00    length:sizeof(uint64_t) atIndex:7];
+                            [encoder setBytes:&nb00        length:sizeof(uint64_t) atIndex:7];
-                            [encoder setBytes:&nb01    length:sizeof(uint64_t) atIndex:8];
+                            [encoder setBytes:&nb01        length:sizeof(uint64_t) atIndex:8];
-                            [encoder setBytes:&nb02    length:sizeof(uint64_t) atIndex:9];
+                            [encoder setBytes:&nb02        length:sizeof(uint64_t) atIndex:9];
-                            [encoder setBytes:&nb03    length:sizeof(uint64_t) atIndex:10];
+                            [encoder setBytes:&nb03        length:sizeof(uint64_t) atIndex:10];
-                            [encoder setBytes:&ne0     length:sizeof( int64_t) atIndex:11];
+                            [encoder setBytes:&ne0         length:sizeof( int64_t) atIndex:11];
-                            [encoder setBytes:&ne1     length:sizeof( int64_t) atIndex:12];
+                            [encoder setBytes:&ne1         length:sizeof( int64_t) atIndex:12];
-                            [encoder setBytes:&ne2     length:sizeof( int64_t) atIndex:13];
+                            [encoder setBytes:&ne2         length:sizeof( int64_t) atIndex:13];
-                            [encoder setBytes:&ne3     length:sizeof( int64_t) atIndex:14];
+                            [encoder setBytes:&ne3         length:sizeof( int64_t) atIndex:14];
-                            [encoder setBytes:&nb0     length:sizeof(uint64_t) atIndex:15];
+                            [encoder setBytes:&nb0         length:sizeof(uint64_t) atIndex:15];
-                            [encoder setBytes:&nb1     length:sizeof(uint64_t) atIndex:16];
+                            [encoder setBytes:&nb1         length:sizeof(uint64_t) atIndex:16];
-                            [encoder setBytes:&nb2     length:sizeof(uint64_t) atIndex:17];
+                            [encoder setBytes:&nb2         length:sizeof(uint64_t) atIndex:17];
-                            [encoder setBytes:&nb3     length:sizeof(uint64_t) atIndex:18];
+                            [encoder setBytes:&nb3         length:sizeof(uint64_t) atIndex:18];
-                            [encoder setBytes:&n_past  length:sizeof(     int) atIndex:19];
+                            [encoder setBytes:&n_past      length:sizeof(     int) atIndex:19];
-                            [encoder setBytes:&n_dims  length:sizeof(     int) atIndex:20];
+                            [encoder setBytes:&n_dims      length:sizeof(     int) atIndex:20];
-                            [encoder setBytes:&mode    length:sizeof(     int) atIndex:21];
+                            [encoder setBytes:&mode        length:sizeof(     int) atIndex:21];
-                            [encoder setBytes:&freq_base  length:sizeof(float) atIndex:22];
+                            [encoder setBytes:&n_orig_ctx  length:sizeof(     int) atIndex:22];
-                            [encoder setBytes:&freq_scale length:sizeof(float) atIndex:23];
+                            [encoder setBytes:&freq_base   length:sizeof(   float) atIndex:23];
+                            [encoder setBytes:&freq_scale  length:sizeof(   float) atIndex:24];
+                            [encoder setBytes:&ext_factor  length:sizeof(   float) atIndex:25];
+                            [encoder setBytes:&attn_factor length:sizeof(   float) atIndex:26];
+                            [encoder setBytes:&beta_fast   length:sizeof(   float) atIndex:27];
+                            [encoder setBytes:&beta_slow   length:sizeof(   float) atIndex:28];
 
                             [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
                         } break;

ggml-metal.metal

@@ -18,6 +18,21 @@ typedef struct {
     uint8_t qs[QK4_1 / 2];  // nibbles / quants
 } block_q4_1;
 
+#define QK5_0 32
+typedef struct {
+    half d;                // delta
+    uint8_t qh[4];         // 5-th bit of quants
+    uint8_t qs[QK5_0 / 2]; // nibbles / quants
+} block_q5_0;
+
+#define QK5_1 32
+typedef struct {
+    half d;                 // delta
+    half m;                 // min
+    uint8_t qh[4];          // 5-th bit of quants
+    uint8_t qs[QK5_1 / 2];  // nibbles / quants
+} block_q5_1;
+
 #define QK8_0 32
 typedef struct {
     half    d;         // delta
@@ -110,9 +125,17 @@ kernel void kernel_mul_row(
 }
 
 kernel void kernel_scale(
+        device const float * src0,
+        device       float * dst,
+        constant     float & scale,
+        uint tpig[[thread_position_in_grid]]) {
+    dst[tpig] = src0[tpig] * scale;
+}
+
+kernel void kernel_scale_4(
         device const float4 * src0,
         device       float4 * dst,
-        constant     float & scale,
+        constant     float  & scale,
         uint tpig[[thread_position_in_grid]]) {
     dst[tpig] = src0[tpig] * scale;
 }
@@ -161,36 +184,73 @@ kernel void kernel_soft_max(
         constant   int64_t & ne00,
         constant   int64_t & ne01,
         constant   int64_t & ne02,
-        uint3 tgpig[[threadgroup_position_in_grid]],
+        threadgroup float  * buf [[threadgroup(0)]],
-        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint  tgpig[[threadgroup_position_in_grid]],
-        uint3   ntg[[threads_per_threadgroup]]) {
+        uint  tpitg[[thread_position_in_threadgroup]],
-    const int64_t i03 = tgpig[2];
+        uint  sgitg[[simdgroup_index_in_threadgroup]],
-    const int64_t i02 = tgpig[1];
+        uint  tiisg[[thread_index_in_simdgroup]],
-    const int64_t i01 = tgpig[0];
+        uint    ntg[[threads_per_threadgroup]]) {
+    const int64_t i03 = (tgpig) / (ne02*ne01);
+    const int64_t i02 = (tgpig - i03*ne02*ne01) / ne01;
+    const int64_t i01 = (tgpig - i03*ne02*ne01 - i02*ne01);
 
     device const float * psrc0 = src0 + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
     device       float * pdst  = dst  + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
 
     // parallel max
-    float lmax = tpitg[0] < ne00 ? psrc0[tpitg[0]] : -INFINITY;
+    float lmax = tpitg < ne00 ? psrc0[tpitg] : -INFINITY;
-    for (int i00 = tpitg[0] + ntg[0]; i00 < ne00; i00 += ntg[0]) {
+
+    for (int i00 = tpitg + ntg; i00 < ne00; i00 += ntg) {
         lmax = MAX(lmax, psrc0[i00]);
     }
-    const float max = simd_max(lmax);
+
+    float max = simd_max(lmax);
+    if (tiisg == 0) {
+        buf[sgitg] = max;
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    // broadcast, simd group number is ntg / 32
+    for (uint i = ntg / 32 / 2; i > 0; i /= 2) {
+       if (tpitg < i) {
+           buf[tpitg] = MAX(buf[tpitg], buf[tpitg + i]);
+       }
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    max = buf[0];
 
     // parallel sum
     float lsum = 0.0f;
-    for (int i00 = tpitg[0]; i00 < ne00; i00 += ntg[0]) {
+    for (int i00 = tpitg; i00 < ne00; i00 += ntg) {
         const float exp_psrc0 = exp(psrc0[i00] - max);
         lsum += exp_psrc0;
         // Remember the result of exp here. exp is expensive, so we really do not
-        // whish to compute it twice.
+        // wish to compute it twice.
         pdst[i00] = exp_psrc0;
     }
 
-    const float sum = simd_sum(lsum);
+    float sum = simd_sum(lsum);
+    if (tiisg == 0) {
+        buf[sgitg] = sum;
+    }
 
-    for (int i00 = tpitg[0]; i00 < ne00; i00 += ntg[0]) {
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    // broadcast, simd group number is ntg / 32
+    for (uint i = ntg / 32 / 2; i > 0; i /= 2) {
+       if (tpitg < i) {
+           buf[tpitg] += buf[tpitg + i];
+       }
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    sum = buf[0];
+
+    for (int i00 = tpitg; i00 < ne00; i00 += ntg) {
         pdst[i00] /= sum;
     }
 }
@@ -201,37 +261,73 @@ kernel void kernel_soft_max_4(
         constant   int64_t & ne00,
         constant   int64_t & ne01,
         constant   int64_t & ne02,
-        uint3 tgpig[[threadgroup_position_in_grid]],
+        threadgroup float  * buf [[threadgroup(0)]],
-        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint  tgpig[[threadgroup_position_in_grid]],
-        uint3   ntg[[threads_per_threadgroup]]) {
+        uint  tpitg[[thread_position_in_threadgroup]],
-    const int64_t i03 = tgpig[2];
+        uint  sgitg[[simdgroup_index_in_threadgroup]],
-    const int64_t i02 = tgpig[1];
+        uint  tiisg[[thread_index_in_simdgroup]],
-    const int64_t i01 = tgpig[0];
+        uint    ntg[[threads_per_threadgroup]]) {
+    const int64_t i03 = (tgpig) / (ne02*ne01);
+    const int64_t i02 = (tgpig - i03*ne02*ne01) / ne01;
+    const int64_t i01 = (tgpig - i03*ne02*ne01 - i02*ne01);
 
     device const float4 * psrc4 = (device const float4 *)(src0 + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00);
     device       float4 * pdst4 = (device       float4 *)(dst  + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00);
 
     // parallel max
-    float4 lmax4 = tpitg[0] < ne00/4 ? psrc4[tpitg[0]] : -INFINITY;
+    float4 lmax4 = tpitg < ne00/4 ? psrc4[tpitg] : -INFINITY;
-    for (int i00 = tpitg[0] + ntg[0]; i00 < ne00/4; i00 += ntg[0]) {
+
+    for (int i00 = tpitg + ntg; i00 < ne00/4; i00 += ntg) {
         lmax4 = fmax(lmax4, psrc4[i00]);
     }
-    float lmax = MAX(MAX(lmax4[0], lmax4[1]), MAX(lmax4[2], lmax4[3]));
 
-    const float max = simd_max(lmax);
+    const float lmax = MAX(MAX(lmax4[0], lmax4[1]), MAX(lmax4[2], lmax4[3]));
+    float max = simd_max(lmax);
+    if (tiisg == 0) {
+        buf[sgitg] = max;
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    // broadcast, simd group number is ntg / 32
+    for (uint i = ntg / 32 / 2; i > 0; i /= 2) {
+       if (tpitg < i) {
+           buf[tpitg] = MAX(buf[tpitg], buf[tpitg + i]);
+       }
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    max = buf[0];
 
     // parallel sum
     float4 lsum4 = 0.0f;
-    for (int i00 = tpitg[0]; i00 < ne00/4; i00 += ntg[0]) {
+    for (int i00 = tpitg; i00 < ne00/4; i00 += ntg) {
         const float4 exp_psrc4 = exp(psrc4[i00] - max);
         lsum4 += exp_psrc4;
         pdst4[i00] = exp_psrc4;
     }
-    float lsum = lsum4[0] + lsum4[1] + lsum4[2] + lsum4[3];
 
-    const float sum = simd_sum(lsum);
+    const float lsum = lsum4[0] + lsum4[1] + lsum4[2] + lsum4[3];
+    float sum = simd_sum(lsum);
+    if (tiisg == 0) {
+        buf[sgitg] = sum;
+    }
 
-    for (int i00 = tpitg[0]; i00 < ne00/4; i00 += ntg[0]) {
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    // broadcast, simd group number is ntg / 32
+    for (uint i = ntg / 32 / 2; i > 0; i /= 2) {
+       if (tpitg < i) {
+           buf[tpitg] += buf[tpitg + i];
+       }
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    sum = buf[0];
+
+    for (int i00 = tpitg; i00 < ne00/4; i00 += ntg) {
         pdst4[i00] /= sum;
     }
 }
@@ -251,7 +347,7 @@ kernel void kernel_diag_mask_inf(
         dst[i02*ne01*ne00 + i01*ne00 + i00] = -INFINITY;
     } else {
         dst[i02*ne01*ne00 + i01*ne00 + i00] = src0[i02*ne01*ne00 + i01*ne00 + i00];
-     }
+    }
 }
 
 kernel void kernel_diag_mask_inf_8(
@@ -399,8 +495,11 @@ kernel void kernel_rms_norm(
 // that corresponds to the missing bit shifts (1, 1/16, 1/256, 1/4096)
 inline float block_q_n_dot_y(device const block_q4_0 * qb_curr, float sumy, thread float * yl, int il) {
     float d = qb_curr->d;
+
     float2 acc = 0.f;
+
     device const uint16_t * qs = ((device const uint16_t *)qb_curr + 1 + il/2);
+
     for (int i = 0; i < 8; i+=2) {
         acc[0] += yl[i + 0] * (qs[i / 2] & 0x000F)
                 + yl[i + 1] * (qs[i / 2] & 0x0F00);
@@ -417,8 +516,11 @@ inline float block_q_n_dot_y(device const block_q4_0 * qb_curr, float sumy, thre
 inline float block_q_n_dot_y(device const block_q4_1 * qb_curr, float sumy, thread float * yl, int il) {
     float d = qb_curr->d;
     float m = qb_curr->m;
-    device const uint16_t * qs = ((device const uint16_t *)qb_curr + 2 + il/2);
+
     float2 acc = 0.f;
+
+    device const uint16_t * qs = ((device const uint16_t *)qb_curr + 2 + il/2);
+
     for (int i = 0; i < 8; i+=2) {
         acc[0] += yl[i + 0] * (qs[i / 2] & 0x000F)
                 + yl[i + 1] * (qs[i / 2] & 0x0F00);
@@ -428,6 +530,49 @@ inline float block_q_n_dot_y(device const block_q4_1 * qb_curr, float sumy, thre
     return d * (acc[0] + acc[1]) + sumy * m;
 }
 
+// function for calculate inner product between half a q5_0 block and 16 floats (yl), sumy is SUM(yl[i])
+// il indicates where the q5 quants begin (0 or QK5_0/4)
+// we assume that the yl's have been multiplied with the appropriate scale factor
+// that corresponds to the missing bit shifts (1, 1/16, 1/256, 1/4096)
+inline float block_q_n_dot_y(device const block_q5_0 * qb_curr, float sumy, thread float * yl, int il) {
+    float d = qb_curr->d;
+
+    float2 acc = 0.f;
+
+    device const uint16_t * qs =  ((device const uint16_t *)qb_curr + 3 + il/2);
+           const uint32_t   qh = *((device const uint32_t *)qb_curr->qh);
+
+    for (int i = 0; i < 8; i+=2) {
+        acc[0] += yl[i + 0] * ((qs[i / 2] & 0x000F) | ((qh >> (i+0+il        ) << 4 ) & 0x00010))
+                + yl[i + 1] * ((qs[i / 2] & 0x0F00) | ((qh >> (i+1+il        ) << 12) & 0x01000));
+        acc[1] += yl[i + 8] * ((qs[i / 2] & 0x00F0) | ((qh >> (i+0+il+QK5_0/2) << 8 ) & 0x00100))
+                + yl[i + 9] * ((qs[i / 2] & 0xF000) | ((qh >> (i+1+il+QK5_0/2) << 16) & 0x10000));
+    }
+    return d * (sumy * -16.f + acc[0] + acc[1]);
+}
+
+// function for calculate inner product between half a q5_1 block and 16 floats (yl), sumy is SUM(yl[i])
+// il indicates where the q5 quants begin (0 or QK5_1/4)
+// we assume that the yl's have been multiplied with the appropriate scale factor
+// that corresponds to the missing bit shifts (1, 1/16, 1/256, 1/4096)
+inline float block_q_n_dot_y(device const block_q5_1 * qb_curr, float sumy, thread float * yl, int il) {
+    float d = qb_curr->d;
+    float m = qb_curr->m;
+
+    float2 acc = 0.f;
+
+    device const uint16_t * qs =  ((device const uint16_t *)qb_curr + 4 + il/2);
+           const uint32_t   qh = *((device const uint32_t *)qb_curr->qh);
+
+    for (int i = 0; i < 8; i+=2) {
+        acc[0] += yl[i + 0] * ((qs[i / 2] & 0x000F) | ((qh >> (i+0+il        ) << 4 ) & 0x00010))
+                + yl[i + 1] * ((qs[i / 2] & 0x0F00) | ((qh >> (i+1+il        ) << 12) & 0x01000));
+        acc[1] += yl[i + 8] * ((qs[i / 2] & 0x00F0) | ((qh >> (i+0+il+QK5_0/2) << 8 ) & 0x00100))
+                + yl[i + 9] * ((qs[i / 2] & 0xF000) | ((qh >> (i+1+il+QK5_0/2) << 16) & 0x10000));
+    }
+    return d * (acc[0] + acc[1]) + sumy * m;
+}
+
 // putting them in the kernel cause a significant performance penalty
 #define N_DST 4        // each SIMD group works on 4 rows
 #define N_SIMDGROUP 2  // number of SIMD groups in a thread group
@@ -525,6 +670,43 @@ kernel void kernel_mul_mv_q4_1_f32(
      mul_vec_q_n_f32<block_q4_1, N_DST, N_SIMDGROUP, N_SIMDWIDTH>(src0,src1,dst,ne00,ne01,ne02,ne10,ne12,ne0,ne1,gqa,tgpig,tiisg,sgitg);
 }
 
+kernel void kernel_mul_mv_q5_0_f32(
+        device const  void * src0,
+        device const float * src1,
+        device       float * dst,
+        constant   int64_t & ne00,
+        constant   int64_t & ne01[[buffer(4)]],
+        constant   int64_t & ne02[[buffer(5)]],
+        constant   int64_t & ne10[[buffer(9)]],
+        constant   int64_t & ne12[[buffer(11)]],
+        constant   int64_t & ne0[[buffer(15)]],
+        constant   int64_t & ne1[[buffer(16)]],
+        constant   uint    & gqa[[buffer(17)]],
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint  tiisg[[thread_index_in_simdgroup]],
+        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
+    mul_vec_q_n_f32<block_q5_0, N_DST, N_SIMDGROUP, N_SIMDWIDTH>(src0,src1,dst,ne00,ne01,ne02,ne10,ne12,ne0,ne1,gqa,tgpig,tiisg,sgitg);
+}
+
+kernel void kernel_mul_mv_q5_1_f32(
+        device const  void * src0,
+        device const float * src1,
+        device       float * dst,
+        constant   int64_t & ne00,
+        constant   int64_t & ne01[[buffer(4)]],
+        constant   int64_t & ne02[[buffer(5)]],
+        constant   int64_t & ne10[[buffer(9)]],
+        constant   int64_t & ne12[[buffer(11)]],
+        constant   int64_t & ne0[[buffer(15)]],
+        constant   int64_t & ne1[[buffer(16)]],
+        constant   uint    & gqa[[buffer(17)]],
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint  tiisg[[thread_index_in_simdgroup]],
+        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
+    mul_vec_q_n_f32<block_q5_1, N_DST, N_SIMDGROUP, N_SIMDWIDTH>(src0,src1,dst,ne00,ne01,ne02,ne10,ne12,ne0,ne1,gqa,tgpig,tiisg,sgitg);
+}
+
+
 #define NB_Q8_0 8
 
 kernel void kernel_mul_mv_q8_0_f32(
@@ -879,6 +1061,45 @@ kernel void kernel_alibi_f32(
     }
 }
 
+static float rope_yarn_ramp(const float low, const float high, const int i0) {
+    const float y = (i0 / 2 - low) / max(0.001f, high - low);
+    return 1.0f - min(1.0f, max(0.0f, y));
+}
+
+// YaRN algorithm based on LlamaYaRNScaledRotaryEmbedding.py from https://github.com/jquesnelle/yarn
+// MIT licensed. Copyright (c) 2023 Jeffrey Quesnelle and Bowen Peng.
+static void rope_yarn(
+    float theta_extrap, float freq_scale, float corr_dims[2], int64_t i0, float ext_factor, float mscale,
+    thread float * cos_theta, thread float * sin_theta
+) {
+    // Get n-d rotational scaling corrected for extrapolation
+    float theta_interp = freq_scale * theta_extrap;
+    float theta = theta_interp;
+    if (ext_factor != 0.0f) {
+        float ramp_mix = rope_yarn_ramp(corr_dims[0], corr_dims[1], i0) * ext_factor;
+        theta = theta_interp * (1 - ramp_mix) + theta_extrap * ramp_mix;
+
+        // Get n-d magnitude scaling corrected for interpolation
+        mscale *= 1.0f + 0.1f * log(1.0f / freq_scale);
+    }
+    *cos_theta = cos(theta) * mscale;
+    *sin_theta = sin(theta) * mscale;
+}
+
+// Apparently solving `n_rot = 2pi * x * base^((2 * max_pos_emb) / n_dims)` for x, we get
+// `corr_fac(n_rot) = n_dims * log(max_pos_emb / (n_rot * 2pi)) / (2 * log(base))`
+static float rope_yarn_corr_factor(int n_dims, int n_orig_ctx, float n_rot, float base) {
+    return n_dims * log(n_orig_ctx / (n_rot * 2 * M_PI_F)) / (2 * log(base));
+}
+
+static void rope_yarn_corr_dims(
+    int n_dims, int n_orig_ctx, float freq_base, float beta_fast, float beta_slow, float dims[2]
+) {
+    // start and end correction dims
+    dims[0] = max(0.0f,         floor(rope_yarn_corr_factor(n_dims, n_orig_ctx, beta_fast, freq_base)));
+    dims[1] = min(n_dims - 1.0f, ceil(rope_yarn_corr_factor(n_dims, n_orig_ctx, beta_slow, freq_base)));
+}
+
 typedef void (rope_t)(
         device const    void * src0,
         device const int32_t * src1,
@@ -902,8 +1123,13 @@ typedef void (rope_t)(
         constant         int & n_past,
         constant         int & n_dims,
         constant         int & mode,
+        constant         int & n_orig_ctx,
         constant       float & freq_base,
         constant       float & freq_scale,
+        constant       float & ext_factor,
+        constant       float & attn_factor,
+        constant       float & beta_fast,
+        constant       float & beta_slow,
         uint  tiitg[[thread_index_in_threadgroup]],
         uint3 tptg[[threads_per_threadgroup]],
         uint3 tgpig[[threadgroup_position_in_grid]]);
@@ -932,8 +1158,13 @@ kernel void kernel_rope(
         constant         int & n_past,
         constant         int & n_dims,
         constant         int & mode,
+        constant         int & n_orig_ctx,
         constant       float & freq_base,
         constant       float & freq_scale,
+        constant       float & ext_factor,
+        constant       float & attn_factor,
+        constant       float & beta_fast,
+        constant       float & beta_slow,
         uint  tiitg[[thread_index_in_threadgroup]],
         uint3 tptg[[threads_per_threadgroup]],
         uint3 tgpig[[threadgroup_position_in_grid]]) {
@@ -943,19 +1174,22 @@ kernel void kernel_rope(
 
     const bool is_neox = mode & 2;
 
+    float corr_dims[2];
+    rope_yarn_corr_dims(n_dims, n_orig_ctx, freq_base, beta_fast, beta_slow, corr_dims);
+
     device const int32_t * pos = src1;
 
     const int64_t p = pos[i2];
 
-    const float theta_0 = freq_scale * (float)p;
+    const float theta_0 = (float)p;
     const float inv_ndims = -1.f/n_dims;
 
     if (!is_neox) {
         for (int64_t i0 = 2*tiitg; i0 < ne0; i0 += 2*tptg.x) {
 
             const float theta = theta_0 * pow(freq_base, inv_ndims*i0);
-            const float cos_theta = cos(theta);
+            float cos_theta, sin_theta;
-            const float sin_theta = sin(theta);
+            rope_yarn(theta, freq_scale, corr_dims, i0, ext_factor, attn_factor, &cos_theta, &sin_theta);
 
             device const T * const src = (device T *)((device char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
             device       T * dst_data  = (device T *)((device char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
@@ -970,9 +1204,12 @@ kernel void kernel_rope(
         for (int64_t ib = 0; ib < ne0/n_dims; ++ib) {
             for (int64_t ic = 2*tiitg; ic < n_dims; ic += 2*tptg.x) {
 
-                const float theta = theta_0 * pow(freq_base, inv_ndims*ic - ib);
+                // simplified from `(ib * n_dims + ic) * inv_ndims`
-                const float cos_theta = cos(theta);
+                const float cur_rot = inv_ndims*ic - ib;
-                const float sin_theta = sin(theta);
+
+                const float theta = theta_0 * pow(freq_base, cur_rot);
+                float cos_theta, sin_theta;
+                rope_yarn(theta, freq_scale, corr_dims, cur_rot, ext_factor, attn_factor, &cos_theta, &sin_theta);
 
                 const int64_t i0 = ib*n_dims + ic/2;
 
@@ -2149,6 +2386,62 @@ void dequantize_q4_1(device const block_q4_1 *xb, short il, thread type4x4 & reg
     }
 }
 
+template <typename type4x4>
+void dequantize_q5_0(device const block_q5_0 *xb, short il, thread type4x4 & reg) {
+    device const uint16_t * qs = ((device const uint16_t *)xb + 3);
+    const float d = xb->d;
+    const float md = -16.h * xb->d;
+    const ushort mask = il ? 0x00F0 : 0x000F;
+
+    const uint32_t qh = *((device const uint32_t *)xb->qh);
+
+    const int x_mv = il ? 4 : 0;
+
+    const int gh_mv = il ? 12 : 0;
+    const int gh_bk = il ?  0 : 4;
+
+    for (int i = 0; i < 8; i++) {
+        // extract the 5-th bits for x0 and x1
+        const uint8_t xh_0 = ((qh >> (gh_mv + 2*i  )) << gh_bk) & 0x10;
+        const uint8_t xh_1 = ((qh >> (gh_mv + 2*i+1)) << gh_bk) & 0x10;
+
+        // combine the 4-bits from qs with the 5th bit
+        const int32_t x0 = ((((qs[i]     ) & mask) >> x_mv) | xh_0);
+        const int32_t x1 = ((((qs[i] >> 8) & mask) >> x_mv) | xh_1);
+
+        reg[i/2][2*(i%2)+0] = d * x0 + md;
+        reg[i/2][2*(i%2)+1] = d * x1 + md;
+    }
+}
+
+template <typename type4x4>
+void dequantize_q5_1(device const block_q5_1 *xb, short il, thread type4x4 & reg) {
+    device const uint16_t * qs = ((device const uint16_t *)xb + 4);
+    const float d = xb->d;
+    const float m = xb->m;
+    const ushort mask = il ? 0x00F0 : 0x000F;
+
+    const uint32_t qh = *((device const uint32_t *)xb->qh);
+
+    const int x_mv = il ? 4 : 0;
+
+    const int gh_mv = il ? 12 : 0;
+    const int gh_bk = il ?  0 : 4;
+
+    for (int i = 0; i < 8; i++) {
+        // extract the 5-th bits for x0 and x1
+        const uint8_t xh_0 = ((qh >> (gh_mv + 2*i  )) << gh_bk) & 0x10;
+        const uint8_t xh_1 = ((qh >> (gh_mv + 2*i+1)) << gh_bk) & 0x10;
+
+        // combine the 4-bits from qs with the 5th bit
+        const int32_t x0 = ((((qs[i]     ) & mask) >> x_mv) | xh_0);
+        const int32_t x1 = ((((qs[i] >> 8) & mask) >> x_mv) | xh_1);
+
+        reg[i/2][2*(i%2)+0] = d * x0 + m;
+        reg[i/2][2*(i%2)+1] = d * x1 + m;
+    }
+}
+
 template <typename type4x4>
 void dequantize_q8_0(device const block_q8_0 *xb, short il, thread type4x4 & reg) {
     device const int8_t * qs = ((device const int8_t *)xb->qs);
@@ -2490,6 +2783,8 @@ template [[host_name("kernel_get_rows_f32")]]  kernel get_rows_t kernel_get_rows
 template [[host_name("kernel_get_rows_f16")]]  kernel get_rows_t kernel_get_rows<half4x4,    1, dequantize_f16>;
 template [[host_name("kernel_get_rows_q4_0")]] kernel get_rows_t kernel_get_rows<block_q4_0, 2, dequantize_q4_0>;
 template [[host_name("kernel_get_rows_q4_1")]] kernel get_rows_t kernel_get_rows<block_q4_1, 2, dequantize_q4_1>;
+template [[host_name("kernel_get_rows_q5_0")]] kernel get_rows_t kernel_get_rows<block_q5_0, 2, dequantize_q5_0>;
+template [[host_name("kernel_get_rows_q5_1")]] kernel get_rows_t kernel_get_rows<block_q5_1, 2, dequantize_q5_1>;
 template [[host_name("kernel_get_rows_q8_0")]] kernel get_rows_t kernel_get_rows<block_q8_0, 2, dequantize_q8_0>;
 template [[host_name("kernel_get_rows_q2_K")]] kernel get_rows_t kernel_get_rows<block_q2_K, QK_NL, dequantize_q2_K>;
 template [[host_name("kernel_get_rows_q3_K")]] kernel get_rows_t kernel_get_rows<block_q3_K, QK_NL, dequantize_q3_K>;
@@ -2518,6 +2813,8 @@ template [[host_name("kernel_mul_mm_f32_f32")]]  kernel mat_mm_t kernel_mul_mm<f
 template [[host_name("kernel_mul_mm_f16_f32")]]  kernel mat_mm_t kernel_mul_mm<half4x4,    1,     dequantize_f16>;
 template [[host_name("kernel_mul_mm_q4_0_f32")]] kernel mat_mm_t kernel_mul_mm<block_q4_0, 2,     dequantize_q4_0>;
 template [[host_name("kernel_mul_mm_q4_1_f32")]] kernel mat_mm_t kernel_mul_mm<block_q4_1, 2,     dequantize_q4_1>;
+template [[host_name("kernel_mul_mm_q5_0_f32")]] kernel mat_mm_t kernel_mul_mm<block_q5_0, 2,     dequantize_q5_0>;
+template [[host_name("kernel_mul_mm_q5_1_f32")]] kernel mat_mm_t kernel_mul_mm<block_q5_1, 2,     dequantize_q5_1>;
 template [[host_name("kernel_mul_mm_q8_0_f32")]] kernel mat_mm_t kernel_mul_mm<block_q8_0, 2,     dequantize_q8_0>;
 template [[host_name("kernel_mul_mm_q2_K_f32")]] kernel mat_mm_t kernel_mul_mm<block_q2_K, QK_NL, dequantize_q2_K>;
 template [[host_name("kernel_mul_mm_q3_K_f32")]] kernel mat_mm_t kernel_mul_mm<block_q3_K, QK_NL, dequantize_q3_K>;

ggml-opencl.cpp

@@ -1395,75 +1395,46 @@ static void ggml_cl_mul_f32(const ggml_tensor * src0, const ggml_tensor * src1,
     const int64_t ne01 = src0->ne[1];
     const int64_t ne02 = src0->ne[2];
     const int64_t ne03 = src0->ne[3];
-    const int64_t ne0 = ne00 * ne01 * ne02 * ne03;
     const int64_t ne10 = src1->ne[0];
     const int64_t ne11 = src1->ne[1];
     const int64_t ne12 = src1->ne[2];
     const int64_t ne13 = src1->ne[3];
-    const int64_t nb10 = src1->nb[0];
     const int nb2  = dst->nb[2];
     const int nb3  = dst->nb[3];
     size_t x_size;
     size_t d_size;
 
-    cl_mem d_X = ggml_cl_pool_malloc(ne0 * sizeof(float), &x_size); // src0
+    cl_mem d_X = ggml_cl_pool_malloc(ne00 * ne01 * sizeof(float), &x_size); // src0
     cl_mem d_Y = (cl_mem) src1->extra; // src1 is already on device, broadcasted.
-    cl_mem d_D = ggml_cl_pool_malloc(ne0 * sizeof(float), &d_size); // dst
+    cl_mem d_D = ggml_cl_pool_malloc(ne00 * ne01 * sizeof(float), &d_size); // dst
 
 
     for (int64_t i03 = 0; i03 < ne03; i03++) {
         for (int64_t i02 = 0; i02 < ne02; i02++) {
-            const int i0 = i03*ne02 + i02;
-
             cl_event ev;
 
             // copy src0 to device
-            CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_X, i0, src0, i03, i02, &ev));
+            CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_X, 0, src0, i03, i02, &ev));
 
-            if (nb10 == sizeof(float)) {
+            const int64_t i13 = i03%ne13;
-                // Contiguous, avoid overhead from queueing many kernel runs
+            const int64_t i12 = i02%ne12;
-                const int64_t i13 = i03%ne13;
+            const int i1 = i13*ne12*ne11 + i12*ne11;
-                const int64_t i12 = i02%ne12;
-                const int i1 = i13*ne12*ne11 + i12*ne11;
 
-                cl_int x_offset = 0;
+            cl_int x_offset = 0;
-                cl_int y_offset = i1*ne10;
+            cl_int y_offset = i1*ne10;
-                cl_int d_offset = 0;
+            cl_int d_offset = 0;
 
-                size_t global = ne00 * ne01;
+            size_t global = ne00 * ne01;
-                cl_int ky = ne10;
+            cl_int ky = ne10 * ne11;
-                CL_CHECK(clSetKernelArg(mul_f32_cl, 0, sizeof(cl_mem), &d_X));
-                CL_CHECK(clSetKernelArg(mul_f32_cl, 1, sizeof(cl_int), &x_offset));
-                CL_CHECK(clSetKernelArg(mul_f32_cl, 2, sizeof(cl_mem), &d_Y));
-                CL_CHECK(clSetKernelArg(mul_f32_cl, 3, sizeof(cl_int), &y_offset));
-                CL_CHECK(clSetKernelArg(mul_f32_cl, 4, sizeof(cl_mem), &d_D));
-                CL_CHECK(clSetKernelArg(mul_f32_cl, 5, sizeof(cl_int), &d_offset));
-                CL_CHECK(clSetKernelArg(mul_f32_cl, 6, sizeof(cl_int), &ky));
-                CL_CHECK(clEnqueueNDRangeKernel(queue, mul_f32_cl, 1, NULL, &global, NULL, 1, &ev, NULL));
-            } else {
-                for (int64_t i01 = 0; i01 < ne01; i01++) {
-                    const int64_t i13 = i03%ne13;
-                    const int64_t i12 = i02%ne12;
-                    const int64_t i11 = i01%ne11;
-                    const int i1 = i13*ne12*ne11 + i12*ne11 + i11;
 
-                    cl_int x_offset = i01*ne00;
+            CL_CHECK(clSetKernelArg(mul_f32_cl, 0, sizeof(cl_mem), &d_X));
-                    cl_int y_offset = i1*ne10;
+            CL_CHECK(clSetKernelArg(mul_f32_cl, 1, sizeof(cl_int), &x_offset));
-                    cl_int d_offset = i01*ne00;
+            CL_CHECK(clSetKernelArg(mul_f32_cl, 2, sizeof(cl_mem), &d_Y));
-
+            CL_CHECK(clSetKernelArg(mul_f32_cl, 3, sizeof(cl_int), &y_offset));
-                    // compute
+            CL_CHECK(clSetKernelArg(mul_f32_cl, 4, sizeof(cl_mem), &d_D));
-                    size_t global = ne00;
+            CL_CHECK(clSetKernelArg(mul_f32_cl, 5, sizeof(cl_int), &d_offset));
-                    cl_int ky = ne10;
+            CL_CHECK(clSetKernelArg(mul_f32_cl, 6, sizeof(cl_int), &ky));
-                    CL_CHECK(clSetKernelArg(mul_f32_cl, 0, sizeof(cl_mem), &d_X));
+            CL_CHECK(clEnqueueNDRangeKernel(queue, mul_f32_cl, 1, NULL, &global, NULL, 1, &ev, NULL));
-                    CL_CHECK(clSetKernelArg(mul_f32_cl, 1, sizeof(cl_int), &x_offset));
-                    CL_CHECK(clSetKernelArg(mul_f32_cl, 2, sizeof(cl_mem), &d_Y));
-                    CL_CHECK(clSetKernelArg(mul_f32_cl, 3, sizeof(cl_int), &y_offset));
-                    CL_CHECK(clSetKernelArg(mul_f32_cl, 4, sizeof(cl_mem), &d_D));
-                    CL_CHECK(clSetKernelArg(mul_f32_cl, 5, sizeof(cl_int), &d_offset));
-                    CL_CHECK(clSetKernelArg(mul_f32_cl, 6, sizeof(cl_int), &ky));
-                    CL_CHECK(clEnqueueNDRangeKernel(queue, mul_f32_cl, 1, NULL, &global, NULL, 1, &ev, NULL));
-                }
-            }
 
             CL_CHECK(clReleaseEvent(ev));
             CL_CHECK(clFinish(queue));
@@ -1518,46 +1489,45 @@ static void ggml_cl_mul_mat_f32(const ggml_tensor * src0, const ggml_tensor * sr
     cl_mem d_D = ggml_cl_pool_malloc(sizeof(float) * d_ne, &d_size);
 
     size_t x_offset = 0;
-    int64_t pi02 = -1;
-    int64_t pi03 = -1;
 
-    for (int64_t i13 = 0; i13 < ne13; i13++) {
+    for (int64_t i03 = 0; i03 < ne03; i03++) {
-        int64_t i03 = i13 / r3;
+        // TODO: copy src0 here when r3>1
+        for (int64_t i13 = i03 * r3, e13 = i13 + r3; i13 < e13; i13++) {
+            for (int64_t i02 = 0; i02 < ne02; i02++) {
+                if (src0->backend == GGML_BACKEND_GPU) {
+                    x_offset = (i03 * ne02 + i02) * x_ne;
+                } else {
+                    // copy src0 to device
+                    CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_X, 0, src0, i03, i02, NULL));
+                }
 
-        for (int64_t i12 = 0; i12 < ne12; i12++) {
+                for (int64_t i12 = i02 * r2, e12 = i12 + r2; i12 < e12; i12++) {
-            int64_t i02 = i12 / r2;
+                    // copy src1 to device
+                    CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_Y, 0, src1, i13, i12, NULL));
 
-            // copy data to device
+                    CL_CHECK(clFinish(queue));
-            if (src0->backend == GGML_BACKEND_GPU) {
+
-                x_offset = (i03 * ne02 + i02) * x_ne;
+                    // compute
-            } else if (i02 != pi02 || i03 != pi03) {
+                    cl_event ev_sgemm;
-                CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_X, 0, src0, i03, i02, NULL));
+                    clblast::StatusCode status = clblast::Gemm<cl_float>(clblast::Layout::kColMajor,
-                pi02 = i02;
+                                                               clblast::Transpose::kYes, clblast::Transpose::kNo,
-                pi03 = i03;
+                                                               ne01, ne11, ne10,
+                                                               alpha,
+                                                               d_X, x_offset, ne00,
+                                                               d_Y, 0, ne10,
+                                                               beta,
+                                                               d_D, 0, ne01,
+                                                               &queue, &ev_sgemm);
+
+                    if (status != clblast::StatusCode::kSuccess) {
+                        GGML_ASSERT(false);
+                    }
+
+                    // copy dst to host
+                    float * d = (float *) ((char *) dst->data + i12*nb2 + i13*nb3);
+                    CL_CHECK(clEnqueueReadBuffer(queue, d_D, true, 0, sizeof(float) * d_ne, d, 1, &ev_sgemm, NULL));
+                }
             }
-            CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_Y, 0, src1, i13, i12, NULL));
-
-            CL_CHECK(clFinish(queue));
-
-            // compute
-            cl_event ev_sgemm;
-            clblast::StatusCode status = clblast::Gemm<cl_float>(clblast::Layout::kColMajor,
-                                                       clblast::Transpose::kYes, clblast::Transpose::kNo,
-                                                       ne01, ne11, ne10,
-                                                       alpha,
-                                                       d_X, x_offset, ne00,
-                                                       d_Y, 0, ne10,
-                                                       beta,
-                                                       d_D, 0, ne01,
-                                                       &queue, &ev_sgemm);
-
-            if (status != clblast::StatusCode::kSuccess) {
-                GGML_ASSERT(false);
-            }
-
-            // copy dst to host
-            float * d = (float *) ((char *) dst->data + i12*nb2 + i13*nb3);
-            CL_CHECK(clEnqueueReadBuffer(queue, d_D, true, 0, sizeof(float) * d_ne, d, 1, &ev_sgemm, NULL));
         }
     }
 
@@ -1568,7 +1538,7 @@ static void ggml_cl_mul_mat_f32(const ggml_tensor * src0, const ggml_tensor * sr
     ggml_cl_pool_free(d_D, d_size);
 }
 
-static void ggml_cl_mul_mat_f16(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, void * wdata, size_t /* wsize */) {
+static void ggml_cl_mul_mat_f16(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, void * wdata, size_t wsize) {
     GGML_ASSERT(fp16_support);
 
     const int64_t ne00 = src0->ne[0];
@@ -1598,6 +1568,10 @@ static void ggml_cl_mul_mat_f16(const ggml_tensor * src0, const ggml_tensor * sr
     const int y_ne = ne11 * ne10;
     const int d_ne = ne11 * ne01;
 
+    GGML_ASSERT(wsize >= sizeof(ggml_fp16_t) * y_ne);
+    GGML_ASSERT(wsize >= sizeof(ggml_fp16_t) * d_ne);
+    ggml_fp16_t * const tmp = (ggml_fp16_t *) wdata;
+
     size_t x_size;
     size_t y_size;
     size_t d_size;
@@ -1614,74 +1588,70 @@ static void ggml_cl_mul_mat_f16(const ggml_tensor * src0, const ggml_tensor * sr
     bool src1_cont_cols = (size_t)nb11 == ne11*sizeof(float);
 
     size_t x_offset = 0;
-    int64_t pi02 = -1;
-    int64_t pi03 = -1;
 
-    for (int64_t i13 = 0; i13 < ne13; i13++) {
+    for (int64_t i03 = 0; i03 < ne03; i03++) {
-        int64_t i03 = i13 / r3;
+        // TODO: copy src0 here when r3>1
-
+        for (int64_t i13 = i03 * r3, e13 = i13 + r3; i13 < e13; i13++) {
-        for (int64_t i12 = 0; i12 < ne12; i12++) {
+            for (int64_t i02 = 0; i02 < ne02; i02++) {
-            int64_t i02 = i12 / r2;
+                if (src0->backend == GGML_BACKEND_GPU) {
-
+                    x_offset = (i03 * ne02 + i02) * x_ne;
-            // copy src0 to device
+                } else {
-            if (src0->backend == GGML_BACKEND_GPU) {
+                    // copy src0 to device
-                x_offset = (i03 * ne02 + i02) * x_ne;
+                    CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_X, 0, src0, i03, i02, NULL));
-            } else if (i02 != pi02 || i03 != pi03) {
-                CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_X, 0, src0, i03, i02, NULL));
-                pi02 = i02;
-                pi03 = i03;
-            }
-
-            // convert src1 to fp16
-            // TODO: use multiple threads
-            ggml_fp16_t * const tmp = (ggml_fp16_t *) wdata + (ne11 * ne10) * (i13 * ne12 + i12);
-            char * src1i = (char *) src1->data + i13*nb13 + i12*nb12;
-            if (src1_cont_rows) {
-                if (src1_cont_cols) {
-                    ggml_fp32_to_fp16_row((float *) src1i, tmp, ne10*ne11);
                 }
-                else {
+
-                    for (int64_t i11 = 0; i11 < ne11; i11++) {
+                for (int64_t i12 = i02 * r2, e12 = i12 + r2; i12 < e12; i12++) {
-                        ggml_fp32_to_fp16_row((float *) (src1i + i11*nb11), tmp + i11*ne10, ne10);
+                    // convert src1 to fp16
+                    // TODO: use multiple threads
+                    char * src1i = (char *) src1->data + i13*nb13 + i12*nb12;
+                    if (src1_cont_rows) {
+                        if (src1_cont_cols) {
+                            ggml_fp32_to_fp16_row((float *) src1i, tmp, ne10*ne11);
+                        }
+                        else {
+                            for (int64_t i11 = 0; i11 < ne11; i11++) {
+                                ggml_fp32_to_fp16_row((float *) (src1i + i11*nb11), tmp + i11*ne10, ne10);
+                            }
+                        }
                     }
-                }
+                    else {
-            }
+                        for (int64_t i11 = 0; i11 < ne11; i11++) {
-            else {
+                            for (int64_t i10 = 0; i10 < ne10; i10++) {
-                for (int64_t i11 = 0; i11 < ne11; i11++) {
+                                // very slow due to no inlining
-                    for (int64_t i10 = 0; i10 < ne10; i10++) {
+                                tmp[i11*ne10 + i10] = ggml_fp32_to_fp16(*(float *) (src1i + i11*nb11 + i10*nb10));
-                        // very slow due to no inlining
+                            }
-                        tmp[i11*ne10 + i10] = ggml_fp32_to_fp16(*(float *) (src1i + i11*nb11 + i10*nb10));
+                        }
                     }
+
+                    // copy src1 to device
+                    CL_CHECK(clEnqueueWriteBuffer(queue, d_Y, false, 0, sizeof(ggml_fp16_t) * y_ne, tmp, 0, NULL, NULL));
+
+                    CL_CHECK(clFinish(queue));
+
+                    // compute
+                    cl_event ev_sgemm;
+                    clblast::StatusCode status = clblast::Gemm<cl_half>(clblast::Layout::kColMajor,
+                                                               clblast::Transpose::kYes, clblast::Transpose::kNo,
+                                                               ne01, ne11, ne10,
+                                                               alpha,
+                                                               d_X, x_offset, ne00,
+                                                               d_Y, 0, ne10,
+                                                               beta,
+                                                               d_D, 0, ne01,
+                                                               &queue, &ev_sgemm);
+
+                    if (status != clblast::StatusCode::kSuccess) {
+                        GGML_ASSERT(false);
+                    }
+
+                    // copy dst to host, then convert to float
+                    CL_CHECK(clEnqueueReadBuffer(queue, d_D, true, 0, sizeof(ggml_fp16_t) * d_ne, tmp, 1, &ev_sgemm, NULL));
+
+                    float * d = (float *) ((char *) dst->data + i12*nb2 + i13*nb3);
+
+                    ggml_fp16_to_fp32_row(tmp, d, d_ne);
                 }
             }
-
-            // copy src1 to device
-            CL_CHECK(clEnqueueWriteBuffer(queue, d_Y, false, 0, sizeof(ggml_fp16_t) * y_ne, tmp, 0, NULL, NULL));
-
-            CL_CHECK(clFinish(queue));
-
-            // compute
-            cl_event ev_sgemm;
-            clblast::StatusCode status = clblast::Gemm<cl_half>(clblast::Layout::kColMajor,
-                                                       clblast::Transpose::kYes, clblast::Transpose::kNo,
-                                                       ne01, ne11, ne10,
-                                                       alpha,
-                                                       d_X, x_offset, ne00,
-                                                       d_Y, 0, ne10,
-                                                       beta,
-                                                       d_D, 0, ne01,
-                                                       &queue, &ev_sgemm);
-
-            if (status != clblast::StatusCode::kSuccess) {
-                GGML_ASSERT(false);
-            }
-
-            // copy dst to host, then convert to float
-            CL_CHECK(clEnqueueReadBuffer(queue, d_D, true, 0, sizeof(ggml_fp16_t) * d_ne, tmp, 1, &ev_sgemm, NULL));
-
-            float * d = (float *) ((char *) dst->data + i12*nb2 + i13*nb3);
-
-            ggml_fp16_to_fp32_row(tmp, d, d_ne);
         }
     }
 
@@ -1744,85 +1714,81 @@ static void ggml_cl_mul_mat_q_f32(const ggml_tensor * src0, const ggml_tensor *
     size_t ev_idx = 0;
     std::vector<cl_event> events;
 
-    int64_t pi02 = -1;
+    for (int64_t i03 = 0; i03 < ne03; i03++) {
-    int64_t pi03 = -1;
+        // TODO: copy and dequantize src0 here when r3>1
-
+        for (int64_t i13 = i03 * r3, e13 = i13 + r3; i13 < e13; i13++) {
-    for (int64_t i13 = 0; i13 < ne13; i13++) {
+            for (int64_t i02 = 0; i02 < ne02; i02++) {
-        int64_t i03 = i13 / r3;
+                // copy src0 to device if necessary
-
+                if (src0->backend == GGML_BACKEND_CPU) {
-        for (int64_t i12 = 0; i12 < ne12; i12++) {
-            int64_t i02 = i12 / r2;
-
-            // copy src0 to device if necessary
-            if (src0->backend == GGML_BACKEND_CPU) {
-                if (i02 != pi02 || i03 != pi03) {
                     events.emplace_back();
                     CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_Q, 0, src0, i03, i02, events.data() + ev_idx++));
-                    pi02 = i02;
+                } else if (src0->backend == GGML_BACKEND_GPU) {
-                    pi03 = i03;
+                    d_Q = (cl_mem) src0->extra;
-                }
+                } else {
-            } else if (src0->backend == GGML_BACKEND_GPU) {
-                d_Q = (cl_mem) src0->extra;
-            } else {
-                GGML_ASSERT(false);
-            }
-            if (mul_mat_vec) { // specialized dequantize_mul_mat_vec kernel
-                // copy src1 to device
-                events.emplace_back();
-                CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_Y, 0, src1, i13, i12, events.data() + ev_idx++));
-
-                // compute
-                const size_t global = ne01 * local;
-                const size_t offset = src0->backend == GGML_BACKEND_GPU ? (i03 * ne02 + i02) * x_bps : 0;
-                const cl_int ncols = ne00;
-                events.emplace_back();
-                CL_CHECK(clSetKernelArg(*dmmv, 0, sizeof(cl_mem), &d_Q));
-                CL_CHECK(clSetKernelArg(*dmmv, 1, sizeof(float) * local, NULL));
-                CL_CHECK(clSetKernelArg(*dmmv, 2, sizeof(cl_mem), &d_Y));
-                CL_CHECK(clSetKernelArg(*dmmv, 3, sizeof(cl_mem), &d_D));
-                CL_CHECK(clSetKernelArg(*dmmv, 4, sizeof(cl_int), &ncols));
-                CL_CHECK(clEnqueueNDRangeKernel(queue, *dmmv, 1, &offset, &global, &local, events.size() - 1, events.data(), events.data() + ev_idx++));
-            } else { // general dequantization kernel + CLBlast matrix matrix multiplication
-                // convert src0 to fp32 on device
-                const size_t global = x_ne / global_denom;
-                const size_t offset = src0->backend == GGML_BACKEND_GPU ? (i03 * ne02 + i02) * x_bps : 0;
-                CL_CHECK(clSetKernelArg(*to_fp32_cl, 0, sizeof(cl_mem), &d_Q));
-                CL_CHECK(clSetKernelArg(*to_fp32_cl, 1, sizeof(cl_mem), &d_X));
-                CL_CHECK(clEnqueueNDRangeKernel(queue, *to_fp32_cl, 1, offset > 0 ? &offset : NULL, &global, local > 0 ? &local : NULL, events.size(), !events.empty() ? events.data() : NULL, NULL));
-
-                // copy src1 to device
-                CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_Y, 0, src1, i13, i12, NULL));
-
-                events.emplace_back();
-
-                // wait for conversion
-                CL_CHECK(clFinish(queue));
-
-                // compute
-                clblast::StatusCode status = clblast::Gemm<cl_float>(clblast::Layout::kColMajor,
-                                                           clblast::Transpose::kYes, clblast::Transpose::kNo,
-                                                           ne01, ne11, ne10,
-                                                           alpha,
-                                                           d_X, 0, ne00,
-                                                           d_Y, 0, ne10,
-                                                           beta,
-                                                           d_D, 0, ne01,
-                                                           &queue, events.data() + ev_idx++);
-
-                if (status != clblast::StatusCode::kSuccess) {
                     GGML_ASSERT(false);
                 }
-            }
 
-            // copy dst to host
+                if (!mul_mat_vec) {
-            float * d = (float *) ((char *) dst->data + i12*nb2 + i13*nb3);
+                    // convert src0 to fp32 on device
-            CL_CHECK(clEnqueueReadBuffer(queue, d_D, true, 0, sizeof(float) * d_ne, d, 1, &events[events.size() - 1], NULL));
+                    const size_t global = x_ne / global_denom;
-            for (auto *event : events) {
+                    const size_t offset = src0->backend == GGML_BACKEND_GPU ? (i03 * ne02 + i02) * x_bps : 0;
-                clReleaseEvent(event);
+                    CL_CHECK(clSetKernelArg(*to_fp32_cl, 0, sizeof(cl_mem), &d_Q));
-            }
+                    CL_CHECK(clSetKernelArg(*to_fp32_cl, 1, sizeof(cl_mem), &d_X));
+                    CL_CHECK(clEnqueueNDRangeKernel(queue, *to_fp32_cl, 1, &offset, &global, local > 0 ? &local : NULL, events.size(), !events.empty() ? events.data() : NULL, NULL));
+                }
 
-            ev_idx = 0;
+                for (int64_t i12 = i02 * r2, e12 = i12 + r2; i12 < e12; i12++) {
-            events.clear();
+                    if (mul_mat_vec) { // specialized dequantize_mul_mat_vec kernel
+                        // copy src1 to device
+                        events.emplace_back();
+                        CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_Y, 0, src1, i13, i12, events.data() + ev_idx++));
+
+                        // compute
+                        const size_t global = ne01 * local;
+                        const size_t offset = src0->backend == GGML_BACKEND_GPU ? (i03 * ne02 + i02) * x_bps : 0;
+                        const cl_int ncols = ne00;
+                        events.emplace_back();
+                        CL_CHECK(clSetKernelArg(*dmmv, 0, sizeof(cl_mem), &d_Q));
+                        CL_CHECK(clSetKernelArg(*dmmv, 1, sizeof(float) * local, NULL));
+                        CL_CHECK(clSetKernelArg(*dmmv, 2, sizeof(cl_mem), &d_Y));
+                        CL_CHECK(clSetKernelArg(*dmmv, 3, sizeof(cl_mem), &d_D));
+                        CL_CHECK(clSetKernelArg(*dmmv, 4, sizeof(cl_int), &ncols));
+                        CL_CHECK(clEnqueueNDRangeKernel(queue, *dmmv, 1, &offset, &global, &local, events.size() - 1, events.data(), events.data() + ev_idx++));
+                    } else { // CLBlast matrix matrix multiplication
+                        // copy src1 to device
+                        CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_Y, 0, src1, i13, i12, NULL));
+
+                        // wait for conversion
+                        CL_CHECK(clFinish(queue));
+
+                        // compute
+                        events.emplace_back();
+                        clblast::StatusCode status = clblast::Gemm<cl_float>(clblast::Layout::kColMajor,
+                                                                   clblast::Transpose::kYes, clblast::Transpose::kNo,
+                                                                   ne01, ne11, ne10,
+                                                                   alpha,
+                                                                   d_X, 0, ne00,
+                                                                   d_Y, 0, ne10,
+                                                                   beta,
+                                                                   d_D, 0, ne01,
+                                                                   &queue, events.data() + ev_idx++);
+
+                        if (status != clblast::StatusCode::kSuccess) {
+                            GGML_ASSERT(false);
+                        }
+                    }
+
+                    // copy dst to host
+                    float * d = (float *) ((char *) dst->data + i12*nb2 + i13*nb3);
+                    CL_CHECK(clEnqueueReadBuffer(queue, d_D, true, 0, sizeof(float) * d_ne, d, 1, &events[events.size() - 1], NULL));
+                    for (auto *event : events) {
+                        clReleaseEvent(event);
+                    }
+
+                    ev_idx = 0;
+                    events.clear();
+                }
+            }
         }
     }
 
@@ -1897,8 +1863,8 @@ void ggml_cl_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor *
 }
 
 size_t ggml_cl_mul_mat_get_wsize(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst) {
-    if (ggml_cl_mul_mat_use_f16(src0, src1, dst)) {
+    if (src0->type == GGML_TYPE_F16 && ggml_cl_mul_mat_use_f16(src0, src1, dst)) {
-        return ggml_nelements(src1) * sizeof(ggml_fp16_t);
+        return sizeof(ggml_fp16_t) * std::max(src1->ne[0] * src1->ne[1], dst->ne[0] * dst->ne[1]);
     }
     return 0;
 }

ggml-quants.h

@@ -1,11 +1,63 @@
 #pragma once
 
-#include "ggml.h"
+#include "ggml-impl.h"
+
+// GGML internal header
 
 #include <stdint.h>
-#include <assert.h>
 #include <stddef.h>
 
+#define QK4_0 32
+typedef struct {
+    ggml_fp16_t d;          // delta
+    uint8_t qs[QK4_0 / 2];  // nibbles / quants
+} block_q4_0;
+static_assert(sizeof(block_q4_0) == sizeof(ggml_fp16_t) + QK4_0 / 2, "wrong q4_0 block size/padding");
+
+#define QK4_1 32
+typedef struct {
+    ggml_fp16_t d;          // delta
+    ggml_fp16_t m;          // min
+    uint8_t qs[QK4_1 / 2];  // nibbles / quants
+} block_q4_1;
+static_assert(sizeof(block_q4_1) == 2 * sizeof(ggml_fp16_t) + QK4_1 / 2, "wrong q4_1 block size/padding");
+
+#define QK5_0 32
+typedef struct {
+    ggml_fp16_t d;         // delta
+    uint8_t qh[4];         // 5-th bit of quants
+    uint8_t qs[QK5_0 / 2]; // nibbles / quants
+} block_q5_0;
+static_assert(sizeof(block_q5_0) == sizeof(ggml_fp16_t) + sizeof(uint32_t) + QK5_0 / 2, "wrong q5_0 block size/padding");
+
+#define QK5_1 32
+typedef struct {
+    ggml_fp16_t d;         // delta
+    ggml_fp16_t m;         // min
+    uint8_t qh[4];         // 5-th bit of quants
+    uint8_t qs[QK5_1 / 2]; // nibbles / quants
+} block_q5_1;
+static_assert(sizeof(block_q5_1) == 2 * sizeof(ggml_fp16_t) + sizeof(uint32_t) + QK5_1 / 2, "wrong q5_1 block size/padding");
+
+#define QK8_0 32
+typedef struct {
+    ggml_fp16_t d;         // delta
+    int8_t  qs[QK8_0];     // quants
+} block_q8_0;
+static_assert(sizeof(block_q8_0) == sizeof(ggml_fp16_t) + QK8_0, "wrong q8_0 block size/padding");
+
+#define QK8_1 32
+typedef struct {
+    float d;               // delta
+    float s;               // d * sum(qs[i])
+    int8_t  qs[QK8_1];     // quants
+} block_q8_1;
+static_assert(sizeof(block_q8_1) == 2*sizeof(float) + QK8_1, "wrong q8_1 block size/padding");
+
+//
+// Super-block quantization structures
+//
+
 // Super-block size
 #ifdef GGML_QKK_64
 #define QK_K 64
@@ -15,18 +67,6 @@
 #define K_SCALE_SIZE 12
 #endif
 
-#ifndef static_assert
-#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201100L)
-#define static_assert(cond, msg) _Static_assert(cond, msg)
-#else
-#define static_assert(cond, msg) struct global_scope_noop_trick
-#endif
-#endif
-
-//
-// Super-block quantization structures
-//
-
 // 2-bit quantization
 // weight is represented as x = a * q + b
 // 16 blocks of 16 elements each
@@ -127,6 +167,13 @@ static_assert(sizeof(block_q8_K) == sizeof(float) + QK_K + QK_K/16*sizeof(int16_
 
 
 // Quantization
+void quantize_row_q4_0_reference(const float * restrict x, block_q4_0 * restrict y, int k);
+void quantize_row_q4_1_reference(const float * restrict x, block_q4_1 * restrict y, int k);
+void quantize_row_q5_0_reference(const float * restrict x, block_q5_0 * restrict y, int k);
+void quantize_row_q5_1_reference(const float * restrict x, block_q5_1 * restrict y, int k);
+void quantize_row_q8_0_reference(const float * restrict x, block_q8_0 * restrict y, int k);
+void quantize_row_q8_1_reference(const float * restrict x, block_q8_1 * restrict y, int k);
+
 void quantize_row_q2_K_reference(const float * restrict x, block_q2_K * restrict y, int k);
 void quantize_row_q3_K_reference(const float * restrict x, block_q3_K * restrict y, int k);
 void quantize_row_q4_K_reference(const float * restrict x, block_q4_K * restrict y, int k);
@@ -134,6 +181,13 @@ void quantize_row_q5_K_reference(const float * restrict x, block_q5_K * restrict
 void quantize_row_q6_K_reference(const float * restrict x, block_q6_K * restrict y, int k);
 void quantize_row_q8_K_reference(const float * restrict x, block_q8_K * restrict y, int k);
 
+void quantize_row_q4_0(const float * restrict x, void * restrict y, int k);
+void quantize_row_q4_1(const float * restrict x, void * restrict y, int k);
+void quantize_row_q5_0(const float * restrict x, void * restrict y, int k);
+void quantize_row_q5_1(const float * restrict x, void * restrict y, int k);
+void quantize_row_q8_0(const float * restrict x, void * restrict y, int k);
+void quantize_row_q8_1(const float * restrict x, void * restrict y, int k);
+
 void quantize_row_q2_K(const float * restrict x, void * restrict y, int k);
 void quantize_row_q3_K(const float * restrict x, void * restrict y, int k);
 void quantize_row_q4_K(const float * restrict x, void * restrict y, int k);
@@ -142,6 +196,13 @@ void quantize_row_q6_K(const float * restrict x, void * restrict y, int k);
 void quantize_row_q8_K(const float * restrict x, void * restrict y, int k);
 
 // Dequantization
+void dequantize_row_q4_0(const block_q4_0 * restrict x, float * restrict y, int k);
+void dequantize_row_q4_1(const block_q4_1 * restrict x, float * restrict y, int k);
+void dequantize_row_q5_0(const block_q5_0 * restrict x, float * restrict y, int k);
+void dequantize_row_q5_1(const block_q5_1 * restrict x, float * restrict y, int k);
+void dequantize_row_q8_0(const block_q8_0 * restrict x, float * restrict y, int k);
+//void dequantize_row_q8_1(const block_q8_1 * restrict x, float * restrict y, int k);
+
 void dequantize_row_q2_K(const block_q2_K * restrict x, float * restrict y, int k);
 void dequantize_row_q3_K(const block_q3_K * restrict x, float * restrict y, int k);
 void dequantize_row_q4_K(const block_q4_K * restrict x, float * restrict y, int k);
@@ -150,16 +211,14 @@ void dequantize_row_q6_K(const block_q6_K * restrict x, float * restrict y, int
 void dequantize_row_q8_K(const block_q8_K * restrict x, float * restrict y, int k);
 
 // Dot product
+void ggml_vec_dot_q4_0_q8_0(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
+void ggml_vec_dot_q4_1_q8_1(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
+void ggml_vec_dot_q5_0_q8_0(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
+void ggml_vec_dot_q5_1_q8_1(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
+void ggml_vec_dot_q8_0_q8_0(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
+
 void ggml_vec_dot_q2_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
 void ggml_vec_dot_q3_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
 void ggml_vec_dot_q4_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
 void ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
 void ggml_vec_dot_q6_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
-
-// Quantization with histogram collection
-size_t ggml_quantize_q2_K(const float * src, void * dst, int n, int k, int64_t * hist);
-size_t ggml_quantize_q3_K(const float * src, void * dst, int n, int k, int64_t * hist);
-size_t ggml_quantize_q4_K(const float * src, void * dst, int n, int k, int64_t * hist);
-size_t ggml_quantize_q5_K(const float * src, void * dst, int n, int k, int64_t * hist);
-size_t ggml_quantize_q6_K(const float * src, void * dst, int n, int k, int64_t * hist);
-