metal : switch to execution barriers + fix one of the barriers

ref: https://github.com/ggerganov/ggml/pull/621/commits/1914017863d2f9ab8ecc0281cc2a56d683668b92

.editorconfig

@@ -15,14 +15,8 @@ indent_size = 4
 [Makefile]
 indent_style = tab
 
-[scripts/*.mk]
-indent_style = tab
-
 [prompts/*.txt]
 insert_final_newline = unset
 
 [examples/server/public/*]
 indent_size = 2
-
-[examples/llama.swiftui/llama.swiftui.xcodeproj/*]
-indent_style = tab

.github/workflows/docker.yml

@@ -52,36 +52,6 @@ jobs:
           username: ${{ github.repository_owner }}
           password: ${{ secrets.GITHUB_TOKEN }}
 
-      # https://github.com/jlumbroso/free-disk-space/tree/54081f138730dfa15788a46383842cd2f914a1be#example
-      - name: Free Disk Space (Ubuntu)
-        uses: jlumbroso/free-disk-space@main
-        with:
-          # this might remove tools that are actually needed,
-          # if set to "true" but frees about 6 GB
-          tool-cache: false
-
-          # all of these default to true, but feel free to set to
-          # "false" if necessary for your workflow
-          android: true
-          dotnet: true
-          haskell: true
-          large-packages: true
-          docker-images: true
-          swap-storage: true
-
-      - name: Determine tag name
-        id: tag
-        shell: bash
-        run: |
-          BUILD_NUMBER="$(git rev-list --count HEAD)"
-          SHORT_HASH="$(git rev-parse --short=7 HEAD)"
-          if [[ "${{ env.BRANCH_NAME }}" == "master" ]]; then
-            echo "name=b${BUILD_NUMBER}" >> $GITHUB_OUTPUT
-          else
-            SAFE_NAME=$(echo "${{ env.BRANCH_NAME }}" | tr '/' '-')
-            echo "name=${SAFE_NAME}-b${BUILD_NUMBER}-${SHORT_HASH}" >> $GITHUB_OUTPUT
-          fi
-
       - name: Build and push Docker image (versioned)
         if: github.event_name == 'push'
         uses: docker/build-push-action@v4
@@ -89,7 +59,7 @@ jobs:
           context: .
           push: true
           platforms: ${{ matrix.config.platforms }}
-          tags: "ghcr.io/${{ github.repository_owner }}/llama.cpp:${{ matrix.config.tag }}-${{ env.COMMIT_SHA }}"
+          tags: "ghcr.io/ggerganov/llama.cpp:${{ matrix.config.tag }}-${{ env.COMMIT_SHA }}"
           file: ${{ matrix.config.dockerfile }}
 
       - name: Build and push Docker image (tagged)
@@ -98,5 +68,5 @@ jobs:
           context: .
           push: ${{ github.event_name == 'push' }}
           platforms: ${{ matrix.config.platforms }}
-          tags: "ghcr.io/${{ github.repository_owner }}/llama.cpp:${{ matrix.config.tag }}" , "ghcr.io/${{ github.repository_owner }}/llama.cpp:${{ matrix.config.tag }}-${{ steps.tag.outputs.name }}"
+          tags: "ghcr.io/ggerganov/llama.cpp:${{ matrix.config.tag }}"
           file: ${{ matrix.config.dockerfile }}

CMakeLists.txt

@@ -91,7 +91,6 @@ set(LLAMA_CUDA_KQUANTS_ITER "2" CACHE STRING "llama: iters./thread per block for
 set(LLAMA_CUDA_PEER_MAX_BATCH_SIZE "128" CACHE STRING
                                              "llama: max. batch size for using peer access")
 option(LLAMA_HIPBLAS                         "llama: use hipBLAS"                               OFF)
-option(LLAMA_HIP_UMA                         "llama: use HIP unified memory architecture"       OFF)
 option(LLAMA_CLBLAST                         "llama: use CLBlast"                               OFF)
 option(LLAMA_METAL                           "llama: use Metal"                                 ${LLAMA_METAL_DEFAULT})
 option(LLAMA_METAL_NDEBUG                    "llama: disable Metal debugging"                   OFF)
@@ -292,12 +291,7 @@ if (LLAMA_CUBLAS)
         add_compile_definitions(GGML_CUDA_PEER_MAX_BATCH_SIZE=${LLAMA_CUDA_PEER_MAX_BATCH_SIZE})
 
         if (LLAMA_STATIC)
-            if (WIN32)
-                # As of 12.3.1 CUDA Tookit for Windows does not offer a static cublas library
-                set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} CUDA::cudart_static CUDA::cublas CUDA::cublasLt)
-            else ()
-                set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} CUDA::cudart_static CUDA::cublas_static CUDA::cublasLt_static)
-            endif()
+            set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} CUDA::cudart_static CUDA::cublas_static CUDA::cublasLt_static)
         else()
             set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} CUDA::cudart CUDA::cublas CUDA::cublasLt)
         endif()
@@ -378,9 +372,6 @@ if (LLAMA_HIPBLAS)
     if (${hipblas_FOUND} AND ${hip_FOUND})
         message(STATUS "HIP and hipBLAS found")
         add_compile_definitions(GGML_USE_HIPBLAS GGML_USE_CUBLAS)
-        if (LLAMA_HIP_UMA)
-            add_compile_definitions(GGML_HIP_UMA)
-        endif()
         add_library(ggml-rocm OBJECT ggml-cuda.cu ggml-cuda.h)
         if (BUILD_SHARED_LIBS)
             set_target_properties(ggml-rocm PROPERTIES POSITION_INDEPENDENT_CODE ON)
@@ -406,102 +397,57 @@ if (LLAMA_HIPBLAS)
     endif()
 endif()
 
-function(get_flags CCID CCVER)
-    set(C_FLAGS "")
-    set(CXX_FLAGS "")
-
-    if (CCID MATCHES "Clang")
-        set(C_FLAGS   -Wunreachable-code-break -Wunreachable-code-return)
-        set(CXX_FLAGS -Wunreachable-code-break -Wunreachable-code-return -Wmissing-prototypes -Wextra-semi)
-
-        if (
-            (CCID STREQUAL "Clang"      AND CCVER VERSION_GREATER_EQUAL 3.8.0) OR
-            (CCID STREQUAL "AppleClang" AND CCVER VERSION_GREATER_EQUAL 7.3.0)
-        )
-            set(C_FLAGS ${C_FLAGS} -Wdouble-promotion)
-        endif()
-    elseif (CCID STREQUAL "GNU")
-        set(C_FLAGS   -Wdouble-promotion)
-        set(CXX_FLAGS -Wno-array-bounds)
-
-        if (CCVER VERSION_GREATER_EQUAL 7.1.0)
-            set(CXX_FLAGS ${CXX_FLAGS} -Wno-format-truncation)
-        endif()
-        if (CCVER VERSION_GREATER_EQUAL 8.1.0)
-            set(CXX_FLAGS ${CXX_FLAGS} -Wextra-semi)
-        endif()
-    endif()
-
-    set(GF_C_FLAGS   ${C_FLAGS}   PARENT_SCOPE)
-    set(GF_CXX_FLAGS ${CXX_FLAGS} PARENT_SCOPE)
-endfunction()
-
 if (LLAMA_ALL_WARNINGS)
     if (NOT MSVC)
-        set(WARNING_FLAGS -Wall -Wextra -Wpedantic -Wcast-qual -Wno-unused-function)
-        set(C_FLAGS       -Wshadow -Wstrict-prototypes -Wpointer-arith -Wmissing-prototypes
-                          -Werror=implicit-int -Werror=implicit-function-declaration)
-        set(CXX_FLAGS     -Wmissing-declarations -Wmissing-noreturn)
+        set(warning_flags -Wall -Wextra -Wpedantic -Wcast-qual -Wno-unused-function)
+        set(c_flags -Wshadow -Wstrict-prototypes -Wpointer-arith -Wmissing-prototypes -Werror=implicit-int -Werror=implicit-function-declaration)
+        set(cxx_flags -Wmissing-declarations -Wmissing-noreturn)
+        set(host_cxx_flags "")
 
-        set(C_FLAGS   ${WARNING_FLAGS} ${C_FLAGS})
-        set(CXX_FLAGS ${WARNING_FLAGS} ${CXX_FLAGS})
+        if (CMAKE_C_COMPILER_ID MATCHES "Clang")
+            set(warning_flags ${warning_flags} -Wunreachable-code-break -Wunreachable-code-return)
+            set(host_cxx_flags ${host_cxx_flags} -Wmissing-prototypes -Wextra-semi)
 
-        get_flags(${CMAKE_CXX_COMPILER_ID} ${CMAKE_CXX_COMPILER_VERSION})
+            if (
+                (CMAKE_C_COMPILER_ID STREQUAL "Clang"      AND CMAKE_C_COMPILER_VERSION VERSION_GREATER_EQUAL 3.8.0) OR
+                (CMAKE_C_COMPILER_ID STREQUAL "AppleClang" AND CMAKE_C_COMPILER_VERSION VERSION_GREATER_EQUAL 7.3.0)
+            )
+                set(c_flags ${c_flags} -Wdouble-promotion)
+            endif()
+        elseif (CMAKE_C_COMPILER_ID STREQUAL "GNU")
+            set(c_flags ${c_flags} -Wdouble-promotion)
+            set(host_cxx_flags ${host_cxx_flags} -Wno-array-bounds)
 
-        add_compile_options("$<$<COMPILE_LANGUAGE:C>:${C_FLAGS};${GF_C_FLAGS}>"
-                            "$<$<COMPILE_LANGUAGE:CXX>:${CXX_FLAGS};${GF_CXX_FLAGS}>")
+            if (CMAKE_CXX_COMPILER_VERSION VERSION_GREATER_EQUAL 7.1.0)
+                set(host_cxx_flags ${host_cxx_flags} -Wno-format-truncation)
+            endif()
+            if (CMAKE_CXX_COMPILER_VERSION VERSION_GREATER_EQUAL 8.1.0)
+                set(host_cxx_flags ${host_cxx_flags} -Wextra-semi)
+            endif()
+        endif()
     else()
         # todo : msvc
-        set(C_FLAGS   "")
-        set(CXX_FLAGS "")
     endif()
+
+    set(c_flags   ${c_flags}   ${warning_flags})
+    set(cxx_flags ${cxx_flags} ${warning_flags})
+    add_compile_options("$<$<COMPILE_LANGUAGE:C>:${c_flags}>"
+                        "$<$<COMPILE_LANGUAGE:CXX>:${cxx_flags}>"
+                        "$<$<COMPILE_LANGUAGE:CXX>:${host_cxx_flags}>")
+
 endif()
 
-if (LLAMA_CUBLAS)
-    set(CUDA_FLAGS ${CXX_FLAGS} -use_fast_math)
-    if (NOT MSVC)
-        set(CUDA_FLAGS ${CUDA_FLAGS} -Wno-pedantic)
-    endif()
-
-    if (LLAMA_ALL_WARNINGS AND NOT MSVC)
-        set(NVCC_CMD ${CMAKE_CUDA_COMPILER} .c)
-        if (NOT CMAKE_CUDA_HOST_COMPILER STREQUAL "")
-            set(NVCC_CMD ${NVCC_CMD} -ccbin ${CMAKE_CUDA_HOST_COMPILER})
-        endif()
-
-        execute_process(
-            COMMAND ${NVCC_CMD} -Xcompiler --version
-            OUTPUT_VARIABLE CUDA_CCFULLVER
-            ERROR_QUIET
-        )
-
-        if (NOT CUDA_CCFULLVER MATCHES clang)
-            set(CUDA_CCID "GNU")
-            execute_process(
-                COMMAND ${NVCC_CMD} -Xcompiler "-dumpfullversion -dumpversion"
-                OUTPUT_VARIABLE CUDA_CCVER
-                ERROR_QUIET
-            )
-        else()
-            if (CUDA_CCFULLVER MATCHES Apple)
-                set(CUDA_CCID "AppleClang")
-            else()
-                set(CUDA_CCID "Clang")
-            endif()
-            string(REGEX REPLACE "^.* version ([0-9.]*).*$" "\\1" CUDA_CCVER ${CUDA_CCFULLVER})
-        endif()
-
-        message("-- CUDA host compiler is ${CUDA_CCID} ${CUDA_CCVER}")
-
-        get_flags(${CUDA_CCID} ${CUDA_CCVER})
-        list(JOIN GF_CXX_FLAGS " " CUDA_CXX_FLAGS)  # pass host compiler flags as a single argument
-        if (NOT CUDA_CXX_FLAGS STREQUAL "")
-            set(CUDA_FLAGS ${CUDA_FLAGS} -Xcompiler ${CUDA_CXX_FLAGS})
-        endif()
-    endif()
-
-    add_compile_options("$<$<COMPILE_LANGUAGE:CUDA>:${CUDA_FLAGS}>")
+if (NOT MSVC)
+    set(cuda_flags -Wno-pedantic)
 endif()
+set(cuda_flags ${cxx_flags} -use_fast_math ${cuda_flags})
+
+list(JOIN host_cxx_flags " " cuda_host_flags)  # pass host compiler flags as a single argument
+if (NOT cuda_host_flags STREQUAL "")
+    set(cuda_flags ${cuda_flags} -Xcompiler ${cuda_host_flags})
+endif()
+
+add_compile_options("$<$<COMPILE_LANGUAGE:CUDA>:${cuda_flags}>")
 
 if (WIN32)
     add_compile_definitions(_CRT_SECURE_NO_WARNINGS)
@@ -525,7 +471,6 @@ endif()
 execute_process(
     COMMAND ${CMAKE_C_COMPILER} ${CMAKE_EXE_LINKER_FLAGS} -Wl,-v
     ERROR_VARIABLE output
-    OUTPUT_QUIET
 )
 if (output MATCHES "dyld-1015\.7")
     add_compile_definitions(HAVE_BUGGY_APPLE_LINKER)
@@ -648,11 +593,6 @@ else()
     message(STATUS "Unknown architecture")
 endif()
 
-if (MINGW)
-    # Target Windows 8 for PrefetchVirtualMemory
-    add_compile_definitions(_WIN32_WINNT=0x602)
-endif()
-
 #
 # POSIX conformance
 #

Makefile

@@ -26,6 +26,20 @@ ifndef UNAME_M
 UNAME_M := $(shell uname -m)
 endif
 
+ifeq '' '$(findstring clang,$(shell $(CC) --version))'
+	CC_IS_GCC=1
+	CC_VER := $(shell $(CC) -dumpfullversion -dumpversion | awk -F. '{ printf("%02d%02d%02d", $$1, $$2, $$3) }')
+else
+	CC_IS_CLANG=1
+	ifeq '' '$(findstring Apple,$(shell $(CC) --version))'
+		CC_IS_LLVM_CLANG=1
+	else
+		CC_IS_APPLE_CLANG=1
+	endif
+	CC_VER := $(shell $(CC) --version | sed -n 's/^.* version \([0-9.]*\).*$$/\1/p' \
+				| awk -F. '{ printf("%02d%02d%02d", $$1, $$2, $$3) }')
+endif
+
 # Mac OS + Arm can report x86_64
 # ref: https://github.com/ggerganov/whisper.cpp/issues/66#issuecomment-1282546789
 ifeq ($(UNAME_S),Darwin)
@@ -65,7 +79,7 @@ test: $(TEST_TARGETS)
 			./$$test_target; \
 		fi; \
 		if [ $$? -ne 0 ]; then \
-			printf 'Test %s FAILED!\n\n' $$test_target; \
+			printf 'Test $$test_target FAILED!\n\n' $$test_target; \
 			failures=$$(( failures + 1 )); \
 		else \
 			printf 'Test %s passed.\n\n' $$test_target; \
@@ -107,12 +121,12 @@ MK_CXXFLAGS = -std=c++11 -fPIC
 
 # -Ofast tends to produce faster code, but may not be available for some compilers.
 ifdef LLAMA_FAST
-MK_CFLAGS     += -Ofast
-HOST_CXXFLAGS += -Ofast
-MK_NVCCFLAGS  += -O3
+MK_CFLAGS        += -Ofast
+MK_HOST_CXXFLAGS += -Ofast
+MK_CUDA_CXXFLAGS += -O3
 else
-MK_CFLAGS     += -O3
-MK_CXXFLAGS   += -O3
+MK_CFLAGS        += -O3
+MK_CXXFLAGS      += -O3
 endif
 
 # clock_gettime came in POSIX.1b (1993)
@@ -206,6 +220,30 @@ MK_CFLAGS    += $(WARN_FLAGS) -Wshadow -Wstrict-prototypes -Wpointer-arith -Wmis
 				-Werror=implicit-function-declaration
 MK_CXXFLAGS  += $(WARN_FLAGS) -Wmissing-declarations -Wmissing-noreturn
 
+ifeq ($(CC_IS_CLANG), 1)
+	# clang options
+	MK_CFLAGS        += -Wunreachable-code-break -Wunreachable-code-return
+	MK_HOST_CXXFLAGS += -Wunreachable-code-break -Wunreachable-code-return -Wmissing-prototypes -Wextra-semi
+
+	ifneq '' '$(and $(CC_IS_LLVM_CLANG),$(filter 1,$(shell expr $(CC_VER) \>= 030800)))'
+		MK_CFLAGS += -Wdouble-promotion
+	endif
+	ifneq '' '$(and $(CC_IS_APPLE_CLANG),$(filter 1,$(shell expr $(CC_VER) \>= 070300)))'
+		MK_CFLAGS += -Wdouble-promotion
+	endif
+else
+	# gcc options
+	MK_CFLAGS        += -Wdouble-promotion
+	MK_HOST_CXXFLAGS += -Wno-array-bounds
+
+	ifeq ($(shell expr $(CC_VER) \>= 070100), 1)
+		MK_HOST_CXXFLAGS += -Wno-format-truncation
+	endif
+	ifeq ($(shell expr $(CC_VER) \>= 080100), 1)
+		MK_HOST_CXXFLAGS += -Wextra-semi
+	endif
+endif
+
 # this version of Apple ld64 is buggy
 ifneq '' '$(findstring dyld-1015.7,$(shell $(CC) $(LDFLAGS) -Wl,-v 2>&1))'
 	MK_CPPFLAGS += -DHAVE_BUGGY_APPLE_LINKER
@@ -256,8 +294,8 @@ ifndef RISCV
 
 ifeq ($(UNAME_M),$(filter $(UNAME_M),x86_64 i686 amd64))
 	# Use all CPU extensions that are available:
-	MK_CFLAGS     += -march=native -mtune=native
-	HOST_CXXFLAGS += -march=native -mtune=native
+	MK_CFLAGS   += -march=native -mtune=native
+	MK_HOST_CXXFLAGS += -march=native -mtune=native
 
 	# Usage AVX-only
 	#MK_CFLAGS   += -mfma -mf16c -mavx
@@ -268,15 +306,12 @@ ifeq ($(UNAME_M),$(filter $(UNAME_M),x86_64 i686 amd64))
 	#MK_CXXFLAGS += -mssse3
 endif
 
+# The stack is only 16-byte aligned on Windows, so don't let gcc emit aligned moves.
+# https://gcc.gnu.org/bugzilla/show_bug.cgi?id=54412
+# https://github.com/ggerganov/llama.cpp/issues/2922
 ifneq '' '$(findstring mingw,$(shell $(CC) -dumpmachine))'
-	# The stack is only 16-byte aligned on Windows, so don't let gcc emit aligned moves.
-	# https://gcc.gnu.org/bugzilla/show_bug.cgi?id=54412
-	# https://github.com/ggerganov/llama.cpp/issues/2922
 	MK_CFLAGS   += -Xassembler -muse-unaligned-vector-move
 	MK_CXXFLAGS += -Xassembler -muse-unaligned-vector-move
-
-	# Target Windows 8 for PrefetchVirtualMemory
-	MK_CPPFLAGS += -D_WIN32_WINNT=0x602
 endif
 
 ifneq ($(filter aarch64%,$(UNAME_M)),)
@@ -360,10 +395,10 @@ 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
-	MK_NVCCFLAGS  = --forward-unknown-to-host-compiler -use_fast_math
+	NVCCFLAGS = --forward-unknown-to-host-compiler -use_fast_math
 
 ifdef LLAMA_DEBUG
-	MK_NVCCFLAGS += -lineinfo
+	NVCCFLAGS += -lineinfo
 endif
 
 ifdef LLAMA_CUDA_NVCC
@@ -372,52 +407,54 @@ else
 	NVCC = nvcc
 endif #LLAMA_CUDA_NVCC
 ifdef CUDA_DOCKER_ARCH
-	MK_NVCCFLAGS += -Wno-deprecated-gpu-targets -arch=$(CUDA_DOCKER_ARCH)
-else ifndef CUDA_POWER_ARCH
-	MK_NVCCFLAGS += -arch=native
+	NVCCFLAGS += -Wno-deprecated-gpu-targets -arch=$(CUDA_DOCKER_ARCH)
+else ifdef CUDA_POWER_ARCH
+	NVCCFLAGS +=
+else
+	NVCCFLAGS += -arch=native
 endif # CUDA_DOCKER_ARCH
 ifdef LLAMA_CUDA_FORCE_DMMV
-	MK_NVCCFLAGS += -DGGML_CUDA_FORCE_DMMV
+	NVCCFLAGS += -DGGML_CUDA_FORCE_DMMV
 endif # LLAMA_CUDA_FORCE_DMMV
 ifdef LLAMA_CUDA_FORCE_MMQ
-	MK_NVCCFLAGS += -DGGML_CUDA_FORCE_MMQ
+	NVCCFLAGS += -DGGML_CUDA_FORCE_MMQ
 endif # LLAMA_CUDA_FORCE_MMQ
 ifdef LLAMA_CUDA_DMMV_X
-	MK_NVCCFLAGS += -DGGML_CUDA_DMMV_X=$(LLAMA_CUDA_DMMV_X)
+	NVCCFLAGS += -DGGML_CUDA_DMMV_X=$(LLAMA_CUDA_DMMV_X)
 else
-	MK_NVCCFLAGS += -DGGML_CUDA_DMMV_X=32
+	NVCCFLAGS += -DGGML_CUDA_DMMV_X=32
 endif # LLAMA_CUDA_DMMV_X
 ifdef LLAMA_CUDA_MMV_Y
-	MK_NVCCFLAGS += -DGGML_CUDA_MMV_Y=$(LLAMA_CUDA_MMV_Y)
+	NVCCFLAGS += -DGGML_CUDA_MMV_Y=$(LLAMA_CUDA_MMV_Y)
 else ifdef LLAMA_CUDA_DMMV_Y
-	MK_NVCCFLAGS += -DGGML_CUDA_MMV_Y=$(LLAMA_CUDA_DMMV_Y) # for backwards compatibility
+	NVCCFLAGS += -DGGML_CUDA_MMV_Y=$(LLAMA_CUDA_DMMV_Y) # for backwards compatibility
 else
-	MK_NVCCFLAGS += -DGGML_CUDA_MMV_Y=1
+	NVCCFLAGS += -DGGML_CUDA_MMV_Y=1
 endif # LLAMA_CUDA_MMV_Y
 ifdef LLAMA_CUDA_F16
-	MK_NVCCFLAGS += -DGGML_CUDA_F16
+	NVCCFLAGS += -DGGML_CUDA_F16
 endif # LLAMA_CUDA_F16
 ifdef LLAMA_CUDA_DMMV_F16
-	MK_NVCCFLAGS += -DGGML_CUDA_F16
+	NVCCFLAGS += -DGGML_CUDA_F16
 endif # LLAMA_CUDA_DMMV_F16
 ifdef LLAMA_CUDA_KQUANTS_ITER
-	MK_NVCCFLAGS += -DK_QUANTS_PER_ITERATION=$(LLAMA_CUDA_KQUANTS_ITER)
+	NVCCFLAGS += -DK_QUANTS_PER_ITERATION=$(LLAMA_CUDA_KQUANTS_ITER)
 else
-	MK_NVCCFLAGS += -DK_QUANTS_PER_ITERATION=2
+	NVCCFLAGS += -DK_QUANTS_PER_ITERATION=2
 endif
 ifdef LLAMA_CUDA_PEER_MAX_BATCH_SIZE
-	MK_NVCCFLAGS += -DGGML_CUDA_PEER_MAX_BATCH_SIZE=$(LLAMA_CUDA_PEER_MAX_BATCH_SIZE)
+	NVCCFLAGS += -DGGML_CUDA_PEER_MAX_BATCH_SIZE=$(LLAMA_CUDA_PEER_MAX_BATCH_SIZE)
 else
-	MK_NVCCFLAGS += -DGGML_CUDA_PEER_MAX_BATCH_SIZE=128
+	NVCCFLAGS += -DGGML_CUDA_PEER_MAX_BATCH_SIZE=128
 endif # LLAMA_CUDA_PEER_MAX_BATCH_SIZE
 #ifdef LLAMA_CUDA_CUBLAS
-#	MK_NVCCFLAGS += -DGGML_CUDA_CUBLAS
+#	NVCCFLAGS += -DGGML_CUDA_CUBLAS
 #endif # LLAMA_CUDA_CUBLAS
 ifdef LLAMA_CUDA_CCBIN
-	MK_NVCCFLAGS += -ccbin $(LLAMA_CUDA_CCBIN)
+	NVCCFLAGS += -ccbin $(LLAMA_CUDA_CCBIN)
 endif
 ggml-cuda.o: ggml-cuda.cu ggml-cuda.h
-	$(NVCC) $(BASE_CXXFLAGS) $(NVCCFLAGS) -Wno-pedantic -Xcompiler "$(CUDA_CXXFLAGS)" -c $< -o $@
+	$(NVCC) $(NVCCFLAGS) -c $< -o $@
 endif # LLAMA_CUBLAS
 
 ifdef LLAMA_CLBLAST
@@ -439,22 +476,13 @@ ggml-opencl.o: ggml-opencl.cpp ggml-opencl.h
 endif # LLAMA_CLBLAST
 
 ifdef LLAMA_HIPBLAS
-
-	ifeq ($(wildcard /opt/rocm),)
-		ROCM_PATH	?= /usr
-		GPU_TARGETS ?= $(shell $(shell which amdgpu-arch))
-	else
-		ROCM_PATH	?= /opt/rocm
-		GPU_TARGETS ?= $(shell $(ROCM_PATH)/llvm/bin/amdgpu-arch)
-	endif
-	HIPCC                   ?= $(ROCM_PATH)/bin/hipcc
+	ROCM_PATH	?= /opt/rocm
+	HIPCC	    ?= $(ROCM_PATH)/bin/hipcc
+	GPU_TARGETS ?= $(shell $(ROCM_PATH)/llvm/bin/amdgpu-arch)
 	LLAMA_CUDA_DMMV_X       ?= 32
 	LLAMA_CUDA_MMV_Y        ?= 1
 	LLAMA_CUDA_KQUANTS_ITER ?= 2
 	MK_CPPFLAGS += -DGGML_USE_HIPBLAS -DGGML_USE_CUBLAS
-ifdef LLAMA_HIP_UMA
-	MK_CPPFLAGS += -DGGML_HIP_UMA
-endif # LLAMA_HIP_UMA
 	MK_LDFLAGS  += -L$(ROCM_PATH)/lib -Wl,-rpath=$(ROCM_PATH)/lib
 	MK_LDFLAGS	+= -lhipblas -lamdhip64 -lrocblas
 	HIPFLAGS    += $(addprefix --offload-arch=,$(GPU_TARGETS))
@@ -488,22 +516,16 @@ ggml-mpi.o: ggml-mpi.c ggml-mpi.h
 	$(CC) $(CFLAGS) -c $< -o $@
 endif # LLAMA_MPI
 
-GF_CC := $(CC)
-include scripts/get-flags.mk
-
 # combine build flags with cmdline overrides
-override CFLAGS    := $(MK_CPPFLAGS) $(CPPFLAGS) $(MK_CFLAGS) $(GF_CFLAGS) $(CFLAGS)
-BASE_CXXFLAGS      := $(MK_CPPFLAGS) $(CPPFLAGS) $(MK_CXXFLAGS) $(CXXFLAGS)
-override CXXFLAGS  := $(BASE_CXXFLAGS) $(HOST_CXXFLAGS) $(GF_CXXFLAGS)
-override NVCCFLAGS := $(MK_NVCCFLAGS) $(NVCCFLAGS)
-override LDFLAGS   := $(MK_LDFLAGS) $(LDFLAGS)
+override CFLAGS        := $(MK_CPPFLAGS) $(CPPFLAGS) $(MK_CFLAGS) $(CFLAGS)
+override CXXFLAGS      := $(MK_CPPFLAGS) $(CPPFLAGS) $(MK_CXXFLAGS) $(CXXFLAGS)
+override CUDA_CXXFLAGS := $(MK_CUDA_CXXFLAGS) $(CUDA_CXXFLAGS)
+override HOST_CXXFLAGS := $(MK_HOST_CXXFLAGS) $(HOST_CXXFLAGS)
+override LDFLAGS       := $(MK_LDFLAGS) $(LDFLAGS)
 
-# identify CUDA host compiler
-ifdef LLAMA_CUBLAS
-GF_CC := $(NVCC) $(NVCCFLAGS) 2>/dev/null .c -Xcompiler
-include scripts/get-flags.mk
-CUDA_CXXFLAGS := $(GF_CXXFLAGS)
-endif
+# save CXXFLAGS before we add host-only options
+NVCCFLAGS := $(NVCCFLAGS) $(CXXFLAGS) $(CUDA_CXXFLAGS) -Wno-pedantic -Xcompiler "$(HOST_CXXFLAGS)"
+override CXXFLAGS += $(HOST_CXXFLAGS)
 
 #
 # Print build information
@@ -609,7 +631,7 @@ save-load-state: examples/save-load-state/save-load-state.cpp ggml.o llama.o $(C
 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) -Wno-cast-qual
 
-gguf: examples/gguf/gguf.cpp ggml.o $(OBJS)
+gguf: examples/gguf/gguf.cpp ggml.o llama.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
 train-text-from-scratch: examples/train-text-from-scratch/train-text-from-scratch.cpp ggml.o llama.o $(COMMON_DEPS) train.o $(OBJS)
@@ -713,16 +735,16 @@ tests/test-quantize-perf: tests/test-quantize-perf.cpp ggml.o $(OBJS)
 tests/test-sampling: tests/test-sampling.cpp ggml.o llama.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-tests/test-tokenizer-0-falcon: tests/test-tokenizer-0-falcon.cpp ggml.o llama.o $(COMMON_DEPS) console.o $(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 ggml.o llama.o $(COMMON_DEPS) console.o $(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 ggml.o llama.o $(COMMON_DEPS) console.o $(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 ggml.o llama.o $(COMMON_DEPS) console.o $(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-rope: tests/test-rope.cpp ggml.o $(OBJS)

README.md

@@ -10,11 +10,10 @@ Inference of [LLaMA](https://arxiv.org/abs/2302.13971) model in pure C/C++
 
 ### Hot topics
 
-- Collecting Apple Silicon performance stats:
-  - M-series: https://github.com/ggerganov/llama.cpp/discussions/4167
-  - A-series: https://github.com/ggerganov/llama.cpp/discussions/4508
-- Added Mixtral support: https://github.com/ggerganov/llama.cpp/pull/4406
+- **llama.h API change for handling KV cache offloading and data type: https://github.com/ggerganov/llama.cpp/pull/4309**
+- Using `llama.cpp` with AWS instances: https://github.com/ggerganov/llama.cpp/discussions/4225
 - Looking for contributions to improve and maintain the `server` example: https://github.com/ggerganov/llama.cpp/issues/4216
+- Collecting Apple Silicon performance stats: https://github.com/ggerganov/llama.cpp/discussions/4167
 
 ----
 
@@ -97,18 +96,7 @@ as the main playground for developing new features for the [ggml](https://github
 - [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)
-- [x] [Yi models](https://huggingface.co/models?search=01-ai/Yi)
 - [X] [StableLM-3b-4e1t](https://github.com/ggerganov/llama.cpp/pull/3586)
-- [x] [Deepseek models](https://huggingface.co/models?search=deepseek-ai/deepseek)
-- [x] [Qwen models](https://huggingface.co/models?search=Qwen/Qwen)
-- [x] [Mixtral MoE](https://huggingface.co/models?search=mistral-ai/Mixtral)
-
-**Multimodal models:**
-
-- [x] [Llava 1.5 models](https://huggingface.co/collections/liuhaotian/llava-15-653aac15d994e992e2677a7e)
-- [x] [Bakllava](https://huggingface.co/models?search=SkunkworksAI/Bakllava)
-- [x] [Obsidian](https://huggingface.co/NousResearch/Obsidian-3B-V0.5)
-- [x] [ShareGPT4V](https://huggingface.co/models?search=Lin-Chen/ShareGPT4V)
 
 
 **Bindings:**
@@ -123,7 +111,6 @@ as the main playground for developing new features for the [ggml](https://github
 - Clojure: [phronmophobic/llama.clj](https://github.com/phronmophobic/llama.clj)
 - React Native: [mybigday/llama.rn](https://github.com/mybigday/llama.rn)
 - Java: [kherud/java-llama.cpp](https://github.com/kherud/java-llama.cpp)
-- Zig: [deins/llama.cpp.zig](https://github.com/Deins/llama.cpp.zig)
 
 **UI:**
 
@@ -433,21 +420,14 @@ Building the program with BLAS support may lead to some performance improvements
     ```bash
     make LLAMA_HIPBLAS=1
     ```
-  - Using `CMake` for Linux (assuming a gfx1030-compatible AMD GPU):
+  - Using `CMake` for Linux:
     ```bash
-    CC=/opt/rocm/llvm/bin/clang CXX=/opt/rocm/llvm/bin/clang++ \
-        cmake -H. -Bbuild -DLLAMA_HIPBLAS=ON -DAMDGPU_TARGETS=gfx1030 -DCMAKE_BUILD_TYPE=Release \
-        && cmake --build build -- -j 16
+    mkdir build
+    cd build
+    CC=/opt/rocm/llvm/bin/clang CXX=/opt/rocm/llvm/bin/clang++ cmake .. -DLLAMA_HIPBLAS=ON
+    cmake --build .
     ```
-    On Linux it is also possible to use unified memory architecture (UMA) to share main memory between the CPU and integrated GPU by setting `-DLLAMA_HIP_UMA=ON"`.
-    However, this hurts performance for non-integrated GPUs (but enables working with integrated GPUs).
-
-  - Using `make` (example for target gfx1030, build with 16 CPU threads):
-    ```bash
-    make -j16 LLAMA_HIPBLAS=1 LLAMA_HIP_UMA=1 AMDGPU_TARGETS=gxf1030
-    ```
-
-  - Using `CMake` for Windows (using x64 Native Tools Command Prompt for VS, and assuming a gfx1100-compatible AMD GPU):
+  - Using `CMake` for Windows (using x64 Native Tools Command Prompt for VS):
     ```bash
     set PATH=%HIP_PATH%\bin;%PATH%
     mkdir build
@@ -456,11 +436,10 @@ Building the program with BLAS support may lead to some performance improvements
     cmake --build .
     ```
     Make sure that `AMDGPU_TARGETS` is set to the GPU arch you want to compile for. The above example uses `gfx1100` that corresponds to Radeon RX 7900XTX/XT/GRE. You can find a list of targets [here](https://llvm.org/docs/AMDGPUUsage.html#processors)
-    Find your gpu version string by matching the most significant version information from `rocminfo | grep gfx | head -1 | awk '{print $2}'` with the list of processors, e.g. `gfx1035` maps to `gfx1030`.
 
 
   The environment variable [`HIP_VISIBLE_DEVICES`](https://rocm.docs.amd.com/en/latest/understand/gpu_isolation.html#hip-visible-devices) can be used to specify which GPU(s) will be used.
-  If your GPU is not officially supported you can use the environment variable [`HSA_OVERRIDE_GFX_VERSION`] set to a similar GPU, for example 10.3.0 on RDNA2 (e.g. gfx1030, gfx1031, or gfx1035) or 11.0.0 on RDNA3.
+  If your GPU is not officially supported you can use the environment variable [`HSA_OVERRIDE_GFX_VERSION`] set to a similar GPU, for example 10.3.0 on RDNA2 or 11.0.0 on RDNA3.
   The following compilation options are also available to tweak performance (yes, they refer to CUDA, not HIP, because it uses the same code as the cuBLAS version above):
 
   | Option                  | Legal values           | Default | Description |
@@ -991,8 +970,6 @@ docker run --gpus all -v /path/to/models:/models local/llama.cpp:light-cuda -m /
 - There are no strict rules for the code style, but try to follow the patterns in the code (indentation, spaces, etc.). Vertical alignment makes things more readable and easier to batch edit
 - Clean-up any trailing whitespaces, use 4 spaces for indentation, brackets on the same line, `void * ptr`, `int & a`
 - See [good first issues](https://github.com/ggerganov/llama.cpp/issues?q=is%3Aissue+is%3Aopen+label%3A%22good+first+issue%22) for tasks suitable for first contributions
-- Tensors store data in row-major order. We refer to dimension 0 as columns, 1 as rows, 2 as matrices
-- Matrix multiplication is unconventional: [`z = ggml_mul_mat(ctx, x, y)`](https://github.com/ggerganov/llama.cpp/blob/880e352277fc017df4d5794f0c21c44e1eae2b84/ggml.h#L1058-L1064) means `zT = x @ yT`
 
 ### Docs
 

common/common.cpp

@@ -656,10 +656,6 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
         } else if (arg == "-h" || arg == "--help") {
             return false;
 
-        } else if (arg == "--version") {
-            fprintf(stderr, "version: %d (%s)\n", LLAMA_BUILD_NUMBER, LLAMA_COMMIT);
-            fprintf(stderr, "built with %s for %s\n", LLAMA_COMPILER, LLAMA_BUILD_TARGET);
-            exit(0);
         } else if (arg == "--random-prompt") {
             params.random_prompt = true;
         } else if (arg == "--in-prefix-bos") {
@@ -798,7 +794,6 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     printf("\n");
     printf("options:\n");
     printf("  -h, --help            show this help message and exit\n");
-    printf("      --version         show version and build info\n");
     printf("  -i, --interactive     run in interactive mode\n");
     printf("  --interactive-first   run in interactive mode and wait for input right away\n");
     printf("  -ins, --instruct      run in instruction mode (use with Alpaca models)\n");
@@ -920,7 +915,7 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     printf("  -m FNAME, --model FNAME\n");
     printf("                        model path (default: %s)\n", params.model.c_str());
     printf("  -md FNAME, --model-draft FNAME\n");
-    printf("                        draft model for speculative decoding\n");
+    printf("                        draft model for speculative decoding (default: %s)\n", params.model.c_str());
     printf("  -ld LOGDIR, --logdir LOGDIR\n");
     printf("                        path under which to save YAML logs (no logging if unset)\n");
     printf("  --override-kv KEY=TYPE:VALUE\n");

common/log.h

@@ -61,13 +61,13 @@
 //  #define LOG_TARGET stderr
 //  #include "log.h"
 //
-//  The log target can also be redirected to a different function
+//  The log target can also be redirected to a diffrent function
 //  like so:
 //
-//  #define LOG_TARGET log_handler_different()
+//  #define LOG_TARGET log_handler_diffrent()
 //  #include "log.h"
 //
-//  FILE* log_handler_different()
+//  FILE* log_handler_diffrent()
 //  {
 //      return stderr;
 //  }
@@ -421,7 +421,7 @@ inline FILE *log_handler2_impl(bool change = false, LogTriState append = LogTriS
 
 // Disables logs entirely at runtime.
 //  Makes LOG() and LOG_TEE() produce no output,
-//  until enabled back.
+//  untill enabled back.
 #define log_disable() log_disable_impl()
 
 // INTERNAL, DO NOT USE

common/train.cpp

@@ -71,7 +71,7 @@ void free_random_uniform_distribution(struct random_uniform_distribution * rnd)
 
 struct ggml_tensor * randomize_tensor_normal(struct ggml_tensor * tensor, struct random_normal_distribution * rnd) {
     float scale = 1.0f; // xavier
-    switch (ggml_n_dims(tensor)) {
+    switch (tensor->n_dims) {
         case 1:
             scale /= sqrtf((float) tensor->ne[0]);
             for (int i0 = 0; i0 < tensor->ne[0]; i0++) {
@@ -119,7 +119,7 @@ struct ggml_tensor * randomize_tensor_normal(struct ggml_tensor * tensor, struct
 }
 
 struct ggml_tensor * randomize_tensor_uniform(struct ggml_tensor * tensor, struct random_uniform_distribution * rnd) {
-    switch (ggml_n_dims(tensor)) {
+    switch (tensor->n_dims) {
         case 1:
             for (int i0 = 0; i0 < tensor->ne[0]; i0++) {
                 float * dst = (float *) ((char *) tensor->data + i0*tensor->nb[0]);
@@ -183,27 +183,25 @@ float fclamp(const float v, const float min, const float max) {
 }
 
 void assert_shape_1d(struct ggml_tensor * tensor, int64_t ne0) {
+    GGML_ASSERT(tensor->n_dims == 1);
     GGML_ASSERT(tensor->ne[0] == ne0);
-    GGML_ASSERT(tensor->ne[1] == 1);
-    GGML_ASSERT(tensor->ne[2] == 1);
-    GGML_ASSERT(tensor->ne[3] == 1);
 }
 
 void assert_shape_2d(struct ggml_tensor * tensor, int64_t ne0, int64_t ne1) {
+    GGML_ASSERT(tensor->n_dims == 2);
     GGML_ASSERT(tensor->ne[0] == ne0);
     GGML_ASSERT(tensor->ne[1] == ne1);
-    GGML_ASSERT(tensor->ne[2] == 1);
-    GGML_ASSERT(tensor->ne[3] == 1);
 }
 
 void assert_shape_3d(struct ggml_tensor * tensor, int64_t ne0, int64_t ne1, int64_t ne2) {
+    GGML_ASSERT(tensor->n_dims == 3);
     GGML_ASSERT(tensor->ne[0] == ne0);
     GGML_ASSERT(tensor->ne[1] == ne1);
     GGML_ASSERT(tensor->ne[2] == ne2);
-    GGML_ASSERT(tensor->ne[3] == 1);
 }
 
 void assert_shape_4d(struct ggml_tensor * tensor, int64_t ne0, int64_t ne1, int64_t ne2, int64_t ne3) {
+    GGML_ASSERT(tensor->n_dims == 4);
     GGML_ASSERT(tensor->ne[0] == ne0);
     GGML_ASSERT(tensor->ne[1] == ne1);
     GGML_ASSERT(tensor->ne[2] == ne2);
@@ -227,8 +225,8 @@ int64_t get_example_targets_batch(
     bool                   sample_random_offsets
 ) {
     GGML_ASSERT(samples_count > 0);
-    GGML_ASSERT(ggml_is_matrix(tokens_input));
-    GGML_ASSERT(ggml_is_3d(target_probs));
+    GGML_ASSERT(tokens_input->n_dims  == 2);
+    GGML_ASSERT(target_probs->n_dims  == 3);
     int64_t n_vocab  = target_probs->ne[0];
     int64_t n_tokens = tokens_input->ne[0];
     int64_t n_batch  = tokens_input->ne[1];

convert-hf-to-gguf.py

@@ -182,8 +182,6 @@ class Model:
             return QwenModel
         if model_architecture == "MixtralForCausalLM":
             return MixtralModel
-        if model_architecture == "PhiForCausalLM":
-            return Phi2Model
         return Model
 
     def _is_model_safetensors(self) -> bool:
@@ -223,8 +221,6 @@ class Model:
             return gguf.MODEL_ARCH.QWEN
         if arch == "MixtralForCausalLM":
             return gguf.MODEL_ARCH.LLAMA
-        if arch == "PhiForCausalLM":
-            return gguf.MODEL_ARCH.PHI2
 
         raise NotImplementedError(f'Architecture "{arch}" not supported!')
 
@@ -984,24 +980,6 @@ class QwenModel(Model):
             print(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
             self.gguf_writer.add_tensor(new_name, data)
 
-
-class Phi2Model(Model):
-    def set_gguf_parameters(self):
-        block_count = self.hparams["n_layer"]
-
-        self.gguf_writer.add_name("Phi2")
-        self.gguf_writer.add_context_length(self.hparams["n_positions"])
-        self.gguf_writer.add_embedding_length(self.hparams["n_embd"])
-        self.gguf_writer.add_feed_forward_length(4 * self.hparams["n_embd"])
-        self.gguf_writer.add_block_count(block_count)
-        self.gguf_writer.add_head_count(self.hparams["n_head"])
-        self.gguf_writer.add_head_count_kv(self.hparams["n_head"])
-        self.gguf_writer.add_layer_norm_eps(self.hparams["layer_norm_epsilon"])
-        self.gguf_writer.add_rope_dimension_count(self.hparams["rotary_dim"])
-        self.gguf_writer.add_file_type(self.ftype)
-        self.gguf_writer.add_add_bos_token(False)
-
-
 ###### CONVERSION LOGIC ######
 
 

convert-lora-to-ggml.py

@@ -3,6 +3,7 @@ from __future__ import annotations
 
 import json
 import os
+import re
 import struct
 import sys
 from typing import Any, BinaryIO, Sequence
@@ -10,15 +11,43 @@ from typing import Any, BinaryIO, Sequence
 import numpy as np
 import torch
 
-from pathlib import Path
-if 'NO_LOCAL_GGUF' not in os.environ:
-    sys.path.insert(1, str(Path(__file__).parent / 'gguf-py' / 'gguf'))
-import gguf
-
-
 NUMPY_TYPE_TO_FTYPE: dict[str, int] = {"float32": 0, "float16": 1}
 
 
+HF_SUBLAYER_TO_GGML = {
+    "self_attn.q_proj": "attn_q",
+    "self_attn.k_proj": "attn_k",
+    "self_attn.v_proj": "attn_v",
+    "self_attn.o_proj": "attn_output",
+    "mlp.gate_proj": "ffn_gate",
+    "mlp.down_proj": "ffn_down",
+    "mlp.up_proj": "ffn_up",
+    "input_layernorm": "attn_norm",
+    "post_attention_layernorm": "ffn_norm",
+}
+
+
+def translate_tensor_name(t: str) -> str:
+    match = re.match(r".*layers\.(\d+)\.(\w+\.\w+)\.lora_(A|B)\.weight", t)
+    if match:
+        nn = match.group(1)
+        sub_layer = match.group(2)
+        lora_type = match.group(3)
+
+        sub_layer_renamed = HF_SUBLAYER_TO_GGML.get(sub_layer)
+        if sub_layer_renamed is None:
+            print(f"Error: unrecognized sub-layer {sub_layer} in tensor {t}")
+            sys.exit(1)
+
+        output_string = (
+            f"blk.{nn}.{HF_SUBLAYER_TO_GGML[sub_layer]}.weight.lora{lora_type}"
+        )
+        return output_string
+    else:
+        print(f"Error: unrecognized tensor {t}")
+        sys.exit(1)
+
+
 def write_file_header(fout: BinaryIO, params: dict[str, Any]) -> None:
     fout.write(b"ggla"[::-1])  # magic (ggml lora)
     fout.write(struct.pack("i", 1))  # file version
@@ -32,7 +61,9 @@ def write_file_header(fout: BinaryIO, params: dict[str, Any]) -> None:
     fout.write(struct.pack("i", int(params["lora_alpha"])))
 
 
-def write_tensor_header(fout: BinaryIO, name: str, shape: Sequence[int], data_type: np.dtype[Any]) -> None:
+def write_tensor_header(
+    self, name: str, shape: Sequence[int], data_type: np.dtype[Any]
+) -> None:
     sname = name.encode("utf-8")
     fout.write(
         struct.pack(
@@ -47,12 +78,11 @@ def write_tensor_header(fout: BinaryIO, name: str, shape: Sequence[int], data_ty
     fout.seek((fout.tell() + 31) & -32)
 
 
-if len(sys.argv) < 2:
-    print(f"Usage: python {sys.argv[0]} <path> [arch]")
+if len(sys.argv) != 2:
+    print(f"Usage: python {sys.argv[0]} <path>")
     print(
         "Path must contain HuggingFace PEFT LoRA files 'adapter_config.json' and 'adapter_model.bin'"
     )
-    print(f"Arch must be one of {list(gguf.MODEL_ARCH_NAMES.values())} (default: llama)")
     sys.exit(1)
 
 input_json = os.path.join(sys.argv[1], "adapter_config.json")
@@ -60,14 +90,6 @@ input_model = os.path.join(sys.argv[1], "adapter_model.bin")
 output_path = os.path.join(sys.argv[1], "ggml-adapter-model.bin")
 
 model = torch.load(input_model, map_location="cpu")
-arch_name = sys.argv[2] if len(sys.argv) == 3 else "llama"
-
-if arch_name not in gguf.MODEL_ARCH_NAMES.values():
-    print(f"Error: unsupported architecture {arch_name}")
-    sys.exit(1)
-
-arch = list(gguf.MODEL_ARCH_NAMES.keys())[list(gguf.MODEL_ARCH_NAMES.values()).index(arch_name)]
-name_map = gguf.TensorNameMap(arch, 200) # 200 layers ought to be enough for anyone
 
 with open(input_json, "r") as f:
     params = json.load(f)
@@ -95,7 +117,6 @@ with open(output_path, "wb") as fout:
 
     write_file_header(fout, params)
     for k, v in model.items():
-        orig_k = k
         if k.endswith(".default.weight"):
             k = k.replace(".default.weight", ".weight")
         if k in ["llama_proj.weight", "llama_proj.bias"]:
@@ -108,32 +129,7 @@ with open(output_path, "wb") as fout:
             v = v.float()
 
         t = v.detach().numpy()
-
-        prefix = "base_model.model."
-        if k.startswith(prefix):
-            k = k[len(prefix) :]
-
-        lora_suffixes = (".lora_A.weight", ".lora_B.weight")
-        if k.endswith(lora_suffixes):
-            suffix = k[-len(lora_suffixes[0]):]
-            k = k[: -len(lora_suffixes[0])]
-        else:
-            print(f"Error: unrecognized tensor name {orig_k}")
-            sys.exit(1)
-
-        tname = name_map.get_name(k)
-        if tname is None:
-            print(f"Error: could not map tensor name {orig_k}")
-            print(" Note: the arch parameter must be specified if the model is not llama")
-            sys.exit(1)
-
-        if suffix == ".lora_A.weight":
-            tname += ".weight.loraA"
-        elif suffix == ".lora_B.weight":
-            tname += ".weight.loraB"
-        else:
-            assert False
-
+        tname = translate_tensor_name(k)
         print(f"{k} => {tname} {t.shape} {t.dtype} {t.nbytes/1024/1024:.2f}MB")
         write_tensor_header(fout, tname, t.shape, t.dtype)
         t.tofile(fout)

convert.py

@@ -10,7 +10,6 @@ import itertools
 import json
 import math
 import mmap
-import os
 import pickle
 import re
 import signal
@@ -19,15 +18,15 @@ import sys
 import time
 import zipfile
 from abc import ABCMeta, abstractmethod
-from collections import OrderedDict
 from concurrent.futures import ProcessPoolExecutor, ThreadPoolExecutor
 from dataclasses import dataclass
 from pathlib import Path
-from typing import IO, TYPE_CHECKING, Any, Callable, Iterable, Literal, Optional, TypeVar, cast
+from typing import IO, TYPE_CHECKING, Any, Callable, Iterable, Literal, TypeVar
 
 import numpy as np
 from sentencepiece import SentencePieceProcessor
 
+import os
 if 'NO_LOCAL_GGUF' not in os.environ:
     sys.path.insert(1, str(Path(__file__).parent / 'gguf-py'))
 import gguf
@@ -328,138 +327,127 @@ class Params:
         return params
 
 
-class VocabLoader:
-    def __init__(self, params: Params, fname_tokenizer: Path) -> None:
-        try:
-            from transformers import AutoTokenizer
-        except ImportError as e:
-            raise ImportError(
-                "To use VocabLoader, please install the `transformers` package. "
-                "You can install it with `pip install transformers`."
-            ) from e
+#
+# vocab
+#
 
-        try:
-            self.tokenizer = AutoTokenizer.from_pretrained(str(fname_tokenizer), trust_remote_code=True)
-        except ValueError:
-            self.tokenizer = AutoTokenizer.from_pretrained(str(fname_tokenizer), use_fast=False, trust_remote_code=True)
-
-        self.added_tokens_dict: OrderedDict[str, int] = OrderedDict()
-
-        for tok, tokidx in sorted(self.tokenizer.get_added_vocab().items(), key=lambda x: x[1]):
-            if tokidx >= params.n_vocab or tokidx < self.tokenizer.vocab_size:
-                continue
-
-            self.added_tokens_dict[tok] = tokidx
-
-        self.unk_token_id: int = self.tokenizer.unk_token_id
-        self.specials: dict[str, int] = {
-            tok: self.tokenizer.get_vocab()[tok]
-            for tok in self.tokenizer.all_special_tokens
-        }
-        self.special_ids: set[int] = set(self.tokenizer.all_special_ids)
-        self.vocab_size_base: int = self.tokenizer.vocab_size
-        self.vocab_size: int = self.vocab_size_base + len(self.added_tokens_dict)
-        self.fname_tokenizer: Path = fname_tokenizer
-
-        vocab_file = "tokenizer.model"
-        path_candidate = find_vocab_file_path(self.fname_tokenizer, vocab_file)
-        if path_candidate is not None:
-            self.spm = SentencePieceProcessor(str(path_candidate))
-            print(self.spm.vocab_size(), self.vocab_size_base)
+class BpeVocab:
+    def __init__(self, fname_tokenizer: Path, fname_added_tokens: Path | None) -> None:
+        self.bpe_tokenizer = json.loads(open(str(fname_tokenizer), encoding="utf-8").read())
+        added_tokens: dict[str, int]
+        if fname_added_tokens is not None:
+            # FIXME: Verify that added tokens here _cannot_ overlap with the main vocab.
+            added_tokens = json.load(open(fname_added_tokens, encoding="utf-8"))
         else:
-            self.spm = None
+            # Fall back to trying to find the added tokens in tokenizer.json
+            tokenizer_json_file = fname_tokenizer.parent / 'tokenizer.json'
+            if not tokenizer_json_file.is_file():
+                added_tokens = {}
+            else:
+                tokenizer_json = json.load(open(tokenizer_json_file, encoding="utf-8"))
+                added_tokens = dict(
+                    (item['content'], item['id'])
+                    for item in tokenizer_json.get('added_tokens', [])
+                    # Added tokens here can be duplicates of the main vocabulary.
+                    if item['content'] not in self.bpe_tokenizer)
 
-    def hf_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]:
-        tokenizer = self.tokenizer
-        reverse_vocab = {id: encoded_tok for encoded_tok, id in tokenizer.get_vocab().items()}
-        added_tokens_ids = set(self.added_tokens_dict.values())
+        vocab_size: int = len(self.bpe_tokenizer)
+        expected_ids    = list(range(vocab_size, vocab_size + len(added_tokens)))
+        actual_ids      = sorted(added_tokens.values())
+        if expected_ids != actual_ids:
+            expected_end_id = vocab_size + len(actual_ids) - 1
+            raise Exception(f"Expected the {len(actual_ids)} added token ID(s) to be sequential in the range {vocab_size} - {expected_end_id}; got {actual_ids}")
 
-        for i in range(self.vocab_size_base):
-            if i in added_tokens_ids:
-                continue
+        items = sorted(added_tokens.items(), key=lambda text_idx: text_idx[1])
+        self.added_tokens_list    = [text for (text, idx) in items]
+        self.vocab_size_base: int = vocab_size
+        self.vocab_size: int      = self.vocab_size_base + len(self.added_tokens_list)
+        self.fname_tokenizer      = fname_tokenizer
+        self.fname_added_tokens   = fname_added_tokens
 
-            text = reverse_vocab[i].encode("utf-8")
-            yield text, self.get_token_score(i), self.get_token_type(i)
+    def bpe_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]:
+        tokenizer = self.bpe_tokenizer
+        reverse_vocab = {id: encoded_tok for encoded_tok, id in tokenizer.items()}
 
-    def get_token_type(self, token_id: int) -> gguf.TokenType:
-        toktype = gguf.TokenType.NORMAL
-
-        if self.spm is not None and token_id < self.spm.vocab_size():
-            if self.spm.is_unknown(token_id):
-                toktype = gguf.TokenType.UNKNOWN
-            if self.spm.is_control(token_id):
-                toktype = gguf.TokenType.CONTROL
-            if self.spm.is_unused(token_id):
-                toktype = gguf.TokenType.UNUSED
-            if self.spm.is_byte(token_id):
-                toktype = gguf.TokenType.BYTE
-        else:
-            if token_id == self.unk_token_id:
-                toktype = gguf.TokenType.UNKNOWN
-            if token_id in self.special_ids:
-                toktype = gguf.TokenType.CONTROL
-
-        return toktype
-
-    def get_token_score(self, token_id: int) -> float:
-        if self.spm is not None and token_id < self.spm.vocab_size():
-            return cast(float, self.spm.get_score(token_id))
-        return 0.0
+        for i, _ in enumerate(tokenizer):
+            yield reverse_vocab[i], 0.0, gguf.TokenType.NORMAL
 
     def added_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]:
-
-        for text in self.added_tokens_dict:
-            if text in self.specials:
-
-                toktype = self.get_token_type(self.specials[text])
-                score = self.get_token_score(self.specials[text])
-
-            else:
-                toktype = gguf.TokenType.USER_DEFINED
-                score = -1000.0
-
-            yield text.encode("utf-8"), score, toktype
-
-    def has_newline_token(self) -> bool:
-        return '<0x0A>' in self.tokenizer.vocab or '\n' in self.tokenizer.vocab
+        for text in self.added_tokens_list:
+            score = -1000.0
+            yield text.encode("utf-8"), score, gguf.TokenType.CONTROL
 
     def all_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]:
-        yield from self.hf_tokens()
+        yield from self.bpe_tokens()
         yield from self.added_tokens()
 
-    def get_vocab_type(self) -> str:
-        path_candidates = []
-        vocab_file = "tokenizer.model"
-        path_candidates.append(vocab_file)
-        path_candidate = find_vocab_file_path(self.fname_tokenizer, vocab_file)
-        if path_candidate is not None:
-            return "llama"
+    def __repr__(self) -> str:
+        return f"<BpeVocab with {self.vocab_size_base} base tokens and {len(self.added_tokens_list)} added tokens>"
 
-        vocab_file = "vocab.json"
-        path_candidates.append(vocab_file)
-        path_candidate = find_vocab_file_path(self.fname_tokenizer, vocab_file)
-        if path_candidate is not None:
-            return "gpt2"
 
-        vocab_file = "tokenizer.json"
-        path_candidates.append(vocab_file)
-        path_candidate = find_vocab_file_path(self.fname_tokenizer, vocab_file)
-        if path_candidate:
-            if not self.has_newline_token():
-                return "gpt2"
-            return "llama"
+class SentencePieceVocab:
+    def __init__(self, fname_tokenizer: Path, fname_added_tokens: Path | None) -> None:
+        self.sentencepiece_tokenizer = SentencePieceProcessor(str(fname_tokenizer))
+        added_tokens: dict[str, int]
+        if fname_added_tokens is not None:
+            added_tokens = json.load(open(fname_added_tokens, encoding="utf-8"))
+        else:
+            added_tokens = {}
 
-        raise FileNotFoundError(
-            f"Could not find {path_candidates} in {self.fname_tokenizer} or its parent; "
-            "if it's in another directory, pass the directory as --vocab-dir"
-        )
+        vocab_size: int = self.sentencepiece_tokenizer.vocab_size()
+
+        new_tokens       = {id: piece for piece, id in added_tokens.items() if id >= vocab_size}
+        expected_new_ids = list(range(vocab_size, vocab_size + len(new_tokens)))
+        actual_new_ids   = sorted(new_tokens.keys())
+
+        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]]:
+        tokenizer = self.sentencepiece_tokenizer
+        for i in range(tokenizer.vocab_size()):
+            piece = tokenizer.id_to_piece(i)
+            text: bytes = piece.encode("utf-8")
+            score: float = tokenizer.get_score(i)
+
+            toktype = gguf.TokenType.NORMAL
+            if tokenizer.is_unknown(i):
+                toktype = gguf.TokenType.UNKNOWN
+            if tokenizer.is_control(i):
+                toktype = gguf.TokenType.CONTROL
+
+            # NOTE: I think added_tokens are user defined.
+            # ref: https://github.com/google/sentencepiece/blob/master/src/sentencepiece_model.proto
+            # if tokenizer.is_user_defined(i): toktype = gguf.TokenType.USER_DEFINED
+
+            if tokenizer.is_unused(i):
+                toktype = gguf.TokenType.UNUSED
+            if tokenizer.is_byte(i):
+                toktype = gguf.TokenType.BYTE
+
+            yield text, score, toktype
+
+    def added_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]:
+        for text in self.added_tokens_list:
+            score = -1000.0
+            yield text.encode("utf-8"), score, gguf.TokenType.USER_DEFINED
+
+    def all_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]:
+        yield from self.sentencepiece_tokens()
+        yield from self.added_tokens()
 
     def __repr__(self) -> str:
-        return f"<VocabLoader with {self.vocab_size_base} base tokens and {len(self.added_tokens_dict)} added tokens>"
+        return f"<SentencePieceVocab with {self.vocab_size_base} base tokens and {len(self.added_tokens_list)} added tokens>"
 
 
-Vocab: TypeAlias = 'VocabLoader'
-
+Vocab: TypeAlias = 'BpeVocab | SentencePieceVocab'
 
 #
 # data loading
@@ -627,7 +615,7 @@ def merge_multifile_models(models_plus: list[ModelPlus]) -> ModelPlus:
 
     if any("model.embed_tokens.weight" in mp.model for mp in models_plus):
         # Transformers models put different tensors in different files, but
-        # don't split individual tensors between files.
+        # don't split indivdual tensors between files.
         model: LazyModel = {}
         for mp in models_plus:
             model.update(mp.model)
@@ -720,7 +708,7 @@ class LazyUnpickler(pickle.Unpickler):
         return func(*args)
 
     CLASSES: dict[tuple[str, str], Any] = {
-        # getattr used here as a workaround for mypy not being smart enough to determine
+        # getattr used here as a workaround for mypy not being smart enough to detrmine
         # the staticmethods have a __func__ attribute.
         ('torch._tensor', '_rebuild_from_type_v2'): getattr(rebuild_from_type_v2, '__func__'),
         ('torch._utils', '_rebuild_tensor_v2'): getattr(lazy_rebuild_tensor_v2, '__func__'),
@@ -836,27 +824,20 @@ def bounded_parallel_map(func: Callable[[In], Out], iterable: Iterable[In], conc
             yield result
 
 
-def check_vocab_size(params: Params, vocab: Vocab, pad_vocab: bool = False) -> None:
+def check_vocab_size(params: Params, vocab: Vocab) -> None:
     if params.n_vocab != vocab.vocab_size:
-        if params.n_vocab == vocab.vocab_size:
+        assert isinstance(vocab, BpeVocab) or isinstance(vocab, SentencePieceVocab)
+        if params.n_vocab == vocab.vocab_size_base:
             print("Ignoring added_tokens.json since model matches vocab size without it.")
-            vocab.added_tokens_dict = OrderedDict()
-            vocab.vocab_size = vocab.vocab_size
-            return
-
-        if pad_vocab and params.n_vocab > vocab.vocab_size:
-            pad_count = params.n_vocab - vocab.vocab_size
-            print(f'Padding vocab with {pad_count} token(s) - <dummy00001> through <dummy{pad_count:05}>')
-            for i in range(1, (params.n_vocab - vocab.vocab_size) + 1):
-                vocab.added_tokens_dict[f'<dummy{i:05}>'] = -1
-            vocab.vocab_size = params.n_vocab
+            vocab.added_tokens_list = []
+            vocab.vocab_size = vocab.vocab_size_base
             return
         msg = f"Vocab size mismatch (model has {params.n_vocab}, but {vocab.fname_tokenizer}"
+        if vocab.fname_added_tokens is not None:
+            msg += f" combined with {vocab.fname_added_tokens}"
         msg += f" has {vocab.vocab_size})."
-        if vocab.vocab_size < params.n_vocab < vocab.vocab_size + 20:
+        if vocab.vocab_size < params.n_vocab < vocab.vocab_size + 20 and vocab.fname_added_tokens is None:
             msg += f"  Most likely you are missing added_tokens.json (should be in {vocab.fname_tokenizer.parent})."
-        if vocab.vocab_size < params.n_vocab:
-            msg += " Possibly try using the --padvocab option."
         raise Exception(msg)
 
 
@@ -920,8 +901,12 @@ class OutputFile:
             scores.append(score)
             toktypes.append(toktype)
 
-        vocab_type = vocab.get_vocab_type()
-        self.gguf.add_tokenizer_model(vocab_type)
+        if isinstance(vocab, SentencePieceVocab):
+            self.gguf.add_tokenizer_model("llama")
+        elif isinstance(vocab, BpeVocab):
+            self.gguf.add_tokenizer_model("gpt2")
+        else:
+            raise ValueError('Unknown vocab type: Not BpeVocab or SentencePieceVocab')
         self.gguf.add_token_list(tokens)
         self.gguf.add_token_scores(scores)
         self.gguf.add_token_types(toktypes)
@@ -947,12 +932,8 @@ class OutputFile:
         self.gguf.close()
 
     @staticmethod
-    def write_vocab_only(
-        fname_out: Path, params: Params, vocab: Vocab, svocab: gguf.SpecialVocab,
-        endianess: gguf.GGUFEndian = gguf.GGUFEndian.LITTLE,
-        pad_vocab: bool = False,
-    ) -> None:
-        check_vocab_size(params, vocab, pad_vocab = pad_vocab)
+    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, endianess=endianess)
 
@@ -979,13 +960,8 @@ 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,
-        endianess: gguf.GGUFEndian = gguf.GGUFEndian.LITTLE,
-        pad_vocab: bool = False,
-    ) -> None:
-        check_vocab_size(params, vocab, pad_vocab = pad_vocab)
+    def write_all(fname_out: Path, ftype: GGMLFileType, params: Params, model: LazyModel, vocab: Vocab, svocab: gguf.SpecialVocab, concurrency: int = DEFAULT_CONCURRENCY, endianess: gguf.GGUFEndian = gguf.GGUFEndian.LITTLE) -> None:
+        check_vocab_size(params, vocab)
 
         of = OutputFile(fname_out, endianess=endianess)
 
@@ -1143,17 +1119,35 @@ def load_some_model(path: Path) -> ModelPlus:
     return model_plus
 
 
-def find_vocab_file_path(path: Path, vocab_file: str) -> Optional[Path]:
-    path2 = path / vocab_file
-    # Use `.parent` instead of /.. to handle the symlink case better.
-    path3 = path.parent / vocab_file
+def load_vocab(path: Path, vocabtype: str | None) -> Vocab:
+    # Be extra-friendly and accept either a file or a directory.  Also, if it's
+    # a directory, it might be the model directory, and tokenizer.model might
+    # be in the parent of that.
+    if path.is_dir():
+        vocab_file = "tokenizer.model"
+        if vocabtype == 'bpe':
+            vocab_file = "vocab.json"
+        path2 = path / vocab_file
+        # Use `.parent` instead of /.. to handle the symlink case better.
+        path3 = path.parent / vocab_file
+        if path2.exists():
+            path = path2
+        elif path3.exists():
+            path = path3
+        else:
+            raise FileNotFoundError(
+                f"Could not find {vocab_file} in {path} or its parent; "
+                "if it's in another directory, pass the directory as --vocab-dir")
 
-    if path2.exists():
-        return path2
-    if path3.exists():
-        return path3
+    print(f"Loading vocab file '{path}', type '{vocabtype}'")
 
-    return None
+    added_tokens_path = path.parent / "added_tokens.json"
+    if vocabtype == "bpe":
+        return BpeVocab(path, added_tokens_path if added_tokens_path.exists() else None)
+    elif vocabtype == "spm":
+        return SentencePieceVocab(path, added_tokens_path if added_tokens_path.exists() else None)
+    else:
+        raise ValueError(f"Unsupported vocabulary type {vocabtype}")
 
 
 def default_outfile(model_paths: list[Path], file_type: GGMLFileType) -> Path:
@@ -1191,11 +1185,11 @@ def main(args_in: list[str] | None = None) -> None:
     parser.add_argument("--outtype",     choices=output_choices, help="output format - note: q8_0 may be very slow (default: f16 or f32 based on input)")
     parser.add_argument("--vocab-dir",   type=Path,              help="directory containing tokenizer.model, if separate from model file")
     parser.add_argument("--outfile",     type=Path,              help="path to write to; default: based on input")
-    parser.add_argument("model",         type=Path,              help="directory containing model file, or model file itself (*.pth, *.pt, *.bin)")
+    parser.add_argument("model",         type=Path,              help="directory containing model file, or model file itself (*.pth, *.pt, *.bin, *.safetensors)")
+    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)
     parser.add_argument("--bigendian",   action="store_true",    help="model is executed on big endian machine")
-    parser.add_argument("--padvocab", action="store_true", help="add pad tokens when model vocab expects more than tokenizer metadata provides")
 
     args = parser.parse_args(args_in)
     if args.dump_single:
@@ -1238,13 +1232,12 @@ def main(args_in: list[str] | None = None) -> None:
         if not args.outfile:
             raise ValueError("need --outfile if using --vocab-only")
         # FIXME: Try to respect vocab_dir somehow?
-        vocab = VocabLoader(params, args.vocab_dir or args.model)
+        vocab = load_vocab(args.vocab_dir or args.model, args.vocabtype)
         special_vocab = gguf.SpecialVocab(model_plus.paths[0].parent,
-                                          load_merges = True,
+                                          load_merges = args.vocabtype == 'bpe',
                                           n_vocab = vocab.vocab_size)
         outfile = args.outfile
-        OutputFile.write_vocab_only(outfile, params, vocab, special_vocab,
-                                    endianess = endianess, pad_vocab = args.padvocab)
+        OutputFile.write_vocab_only(outfile, params, vocab, special_vocab)
         print(f"Wrote {outfile}")
         return
 
@@ -1252,15 +1245,12 @@ def main(args_in: list[str] | None = None) -> None:
         vocab = model_plus.vocab
     else:
         vocab_dir = args.vocab_dir if args.vocab_dir else model_plus.paths[0].parent
-        vocab = VocabLoader(params, vocab_dir)
-
+        vocab = load_vocab(vocab_dir, args.vocabtype)
     # FIXME: Try to respect vocab_dir somehow?
-    print(f"Vocab info: {vocab}")
     special_vocab = gguf.SpecialVocab(model_plus.paths[0].parent,
-                                      load_merges = True,
+                                      load_merges = args.vocabtype == 'bpe',
                                       n_vocab = vocab.vocab_size)
 
-    print(f"Special vocab info: {special_vocab}")
     model   = model_plus.model
     model   = convert_model_names(model, params)
     ftype   = pick_output_type(model, args.outtype)
@@ -1270,8 +1260,7 @@ 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, endianess = endianess, pad_vocab = args.padvocab)
+    OutputFile.write_all(outfile, ftype, params, model, vocab, special_vocab, concurrency = args.concurrency, endianess=endianess)
     print(f"Wrote {outfile}")
 
 

examples/baby-llama/baby-llama.cpp

@@ -575,7 +575,10 @@ static struct ggml_tensor * forward(
 
             // KQ_scaled = KQ / sqrt(n_embd/n_head)
             // KQ_scaled shape [n_past + N, N, n_head, 1]
-            struct ggml_tensor * KQ_scaled = ggml_scale(ctx0, KQ, 1.0f/sqrtf(float(n_embd)/n_head));
+            struct ggml_tensor * KQ_scaled =
+                ggml_scale(ctx0,
+                        KQ,
+                        ggml_new_f32(ctx0, 1.0f/sqrtf(float(n_embd)/n_head)));
 
             // KQ_masked = mask_past(KQ_scaled)
             // KQ_masked shape [n_past + N, N, n_head, 1]
@@ -841,7 +844,10 @@ static struct ggml_tensor * forward_batch(
 
             // KQ_scaled = KQ / sqrt(n_embd/n_head)
             // KQ_scaled shape [n_past + N, N, n_head, n_batch]
-            struct ggml_tensor * KQ_scaled = ggml_scale(ctx0, KQ, 1.0f/sqrtf(float(n_embd)/n_head));
+            struct ggml_tensor * KQ_scaled =
+                ggml_scale(ctx0,
+                        KQ,
+                        ggml_new_f32(ctx0, 1.0f/sqrtf(float(n_embd)/n_head)));
             assert_shape_4d(KQ_scaled, n_past + N, N, n_head, n_batch);
 
             // KQ_masked = mask_past(KQ_scaled)
@@ -1125,7 +1131,10 @@ static struct ggml_tensor * forward_lora(
 
             // KQ_scaled = KQ / sqrt(n_embd/n_head)
             // KQ_scaled shape [n_past + N, N, n_head, 1]
-            struct ggml_tensor * KQ_scaled = ggml_scale(ctx0, KQ, 1.0f/sqrtf(float(n_embd)/n_head));
+            struct ggml_tensor * KQ_scaled =
+                ggml_scale(ctx0,
+                        KQ,
+                        ggml_new_f32(ctx0, 1.0f/sqrtf(float(n_embd)/n_head)));
 
             // KQ_masked = mask_past(KQ_scaled)
             // KQ_masked shape [n_past + N, N, n_head, 1]
@@ -1249,9 +1258,9 @@ static struct ggml_tensor * forward_lora(
 }
 
 static void sample_softmax(struct ggml_tensor * logits, struct ggml_tensor * probs, struct ggml_tensor * best_samples) {
-    assert(ggml_is_matrix(logits));
-    assert(ggml_is_matrix(probs));
-    assert(ggml_is_vector(best_samples));
+    assert(logits->n_dims == 2);
+    assert(probs->n_dims == 2);
+    assert(best_samples->n_dims == 1);
     assert(logits->ne[1] == best_samples->ne[0]);
     assert(logits->ne[0] == probs->ne[0]);
     assert(logits->ne[1] == probs->ne[1]);
@@ -1283,9 +1292,9 @@ static void sample_softmax_batch(
     struct ggml_context * ctx, struct ggml_tensor * logits, struct ggml_tensor * probs,
     struct ggml_tensor * best_samples
 ) {
-    GGML_ASSERT(ggml_is_matrix(best_samples));
-    GGML_ASSERT(ggml_is_3d(logits));
-    GGML_ASSERT(ggml_is_3d(probs));
+    GGML_ASSERT(best_samples->n_dims == 2);
+    GGML_ASSERT(logits->n_dims == 3);
+    GGML_ASSERT(probs->n_dims == 3);
     int n_tokens = best_samples->ne[0];
     int n_batch  = best_samples->ne[1];
     int n_vocab  = logits->ne[0];
@@ -1325,7 +1334,7 @@ static void print_row(struct ggml_tensor * probs, int i) {
 }
 
 static void print_matrix(struct ggml_tensor * probs) {
-    assert(ggml_is_matrix(probs));
+    assert(probs->n_dims == 2);
     for (int i = 0; i < probs->ne[1]; ++i) {
         for (int k = 0; k < probs->ne[0]; ++k) {
             float p = ggml_get_f32_1d(probs, i*probs->ne[0] + k);
@@ -1377,8 +1386,8 @@ static void get_example_targets(int example_id, struct ggml_tensor * tokens_inpu
 static void get_example_targets_batch(
     struct ggml_context * ctx, int example_id, struct ggml_tensor * tokens_input, struct ggml_tensor * targets
 ) {
-    GGML_ASSERT(ggml_is_matrix(tokens_input));
-    GGML_ASSERT(ggml_is_3d(targets));
+    GGML_ASSERT(tokens_input->n_dims == 2);
+    GGML_ASSERT(     targets->n_dims == 3);
     int n_tokens = tokens_input->ne[0];
     int n_batch  = tokens_input->ne[1];
     GGML_ASSERT(n_tokens == targets->ne[1]);

examples/benchmark/benchmark-matmult.cpp

@@ -129,13 +129,13 @@ int main(int argc, char ** argv)  {
     const ggml_type qtype = GGML_TYPE_Q4_1;
 
     size_t ctx_size = 0;
-    ctx_size += ggml_row_size(GGML_TYPE_F32, sizex*sizey);
-    ctx_size += ggml_row_size(GGML_TYPE_F32, sizex*sizey);
-    ctx_size += ggml_row_size(GGML_TYPE_F32, sizex*sizez);
-    ctx_size += ggml_row_size(qtype,         sizex*sizey);
-    ctx_size += ggml_row_size(qtype,         sizex*sizey);
-    ctx_size += ggml_row_size(GGML_TYPE_F32, sizex*sizey); // BLAS
-    ctx_size += ggml_row_size(GGML_TYPE_F32, sizex*sizey); // BLAS
+    ctx_size += sizex*sizey*ggml_type_sizef(GGML_TYPE_F32);
+    ctx_size += sizex*sizey*ggml_type_sizef(GGML_TYPE_F32);
+    ctx_size += sizex*sizez*ggml_type_sizef(GGML_TYPE_F32);
+    ctx_size += sizex*sizey*ggml_type_sizef(qtype);
+    ctx_size += sizex*sizey*ggml_type_sizef(qtype);
+    ctx_size += sizex*sizey*ggml_type_sizef(GGML_TYPE_F32); // BLAS
+    ctx_size += sizex*sizey*ggml_type_sizef(GGML_TYPE_F32); // BLAS
     ctx_size += 1024*1024*16;
 
     printf("Allocating Memory of size %zi bytes, %zi MB\n",ctx_size, (ctx_size/1024/1024));

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

@@ -427,7 +427,7 @@ static void print_row(struct ggml_tensor * probs, int i) {
 }
 
 static void print_matrix(struct ggml_tensor * probs) {
-    assert(ggml_is_matrix(probs));
+    assert(probs->n_dims == 2);
     for (int i = 0; i < probs->ne[1]; ++i) {
         for (int k = 0; k < probs->ne[0]; ++k) {
             float p = get_f32_2d(probs, k, i);
@@ -639,7 +639,7 @@ static void load_vocab(const char *filename, Config *config, struct llama_vocab
 
 static void convert_weights_ak_to_gg(struct ggml_tensor * gg_weights, const float * karpathy_weights) {
     int ct;
-    switch (ggml_n_dims(gg_weights)) {
+    switch (gg_weights->n_dims){
         case 1:
             ct = 0;
             for (int i0 = 0; i0 < gg_weights->ne[0]; i0++){

examples/export-lora/export-lora.cpp

@@ -309,7 +309,7 @@ static struct ggml_cgraph * build_graph_lora(
 ) {
     struct ggml_tensor * ab = ggml_mul_mat(ctx, lora_a, lora_b);
     if (scaling != 1.0f) {
-        ab = ggml_scale(ctx, ab, scaling);
+        ab = ggml_scale(ctx, ab, ggml_new_f32(ctx, scaling));
     }
     struct ggml_tensor * res = ggml_add_inplace(ctx, tensor, ab);
 

examples/finetune/finetune.cpp

@@ -269,7 +269,7 @@ static void load_model_hparams_gguf(struct gguf_context * ctx, struct my_llama_h
     float rope_freq_scale = 1.0f;
     GGUF_GET_KEY(ctx, hparams->f_norm_rms_eps, gguf_get_val_f32, GGUF_TYPE_FLOAT32, false, kv(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS));
     GGUF_GET_KEY(ctx, hparams->rope_freq_base, gguf_get_val_f32, GGUF_TYPE_FLOAT32, false, kv(LLM_KV_ROPE_FREQ_BASE));
-    GGUF_GET_KEY(ctx, rope_freq_scale,         gguf_get_val_f32, GGUF_TYPE_FLOAT32, false, kv(LLM_KV_ROPE_SCALE_LINEAR));
+    GGUF_GET_KEY(ctx, rope_freq_scale, gguf_get_val_f32, GGUF_TYPE_FLOAT32, false, kv(LLM_KV_ROPE_SCALE_LINEAR));
     if (rope_freq_scale != 1.0f) {
         hparams->rope_freq_scale = 1.0f / rope_freq_scale;
     }
@@ -612,7 +612,6 @@ static struct ggml_tensor * llama_build_lora_finetune_graphs(
     const int n_rot       = hparams.n_embd_head();
     const int n_embd_head = hparams.n_embd_head();
     const int n_embd_gqa  = hparams.n_embd_gqa();
-
     const float rms_norm_eps    = hparams.f_norm_rms_eps;
     const float rope_freq_base  = hparams.rope_freq_base;
     const float rope_freq_scale = hparams.rope_freq_scale;
@@ -681,7 +680,10 @@ static struct ggml_tensor * llama_build_lora_finetune_graphs(
         checkpoints.push_back(t01);
     }
 
-    const float kv_scale = 1.0f/sqrtf(float(n_embd)/n_head);
+    struct ggml_tensor * kv_scale = NULL;
+    if (!enable_flash_attn) {
+        kv_scale = ggml_new_f32(ctx, 1.0f/sqrtf(float(n_embd)/n_head));
+    }
 
     for (int il = 0; il < n_layer; ++il) {
         struct my_llama_layer & layer = model->layers[il];
@@ -779,32 +781,32 @@ static struct ggml_tensor * llama_build_lora_finetune_graphs(
     // make sure some tensors are not reallocated by inserting new temporary nodes depending on them
     int n_leafs_before = gb->n_leafs;
     int n_nodes_before = gb->n_nodes;
-
+    struct ggml_tensor * one = ggml_new_f32(ctx, 1.0f);
     // output tensors
-    ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, t35, 1.0f));
-    ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, t36, 1.0f));
+    ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, t35, one));
+    ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, t36, one));
     // input gradient
-    ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, t36->grad, 1.0f));
+    ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, t36->grad, one));
     GGML_ASSERT(t36->grad->data == NULL && t36->grad->view_src == NULL);
     ggml_allocr_alloc(alloc, t36->grad);
     // KQ_pos
-    ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, KQ_pos, 1.0f));
+    ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, KQ_pos, one));
 
     // make sure base model tensors data cannot be used in viewable operations
-    ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, model->tok_embeddings, 1.0f));
-    ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, model->norm, 1.0f));
-    ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, model->output, 1.0f));
+    ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, model->tok_embeddings, one));
+    ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, model->norm, one));
+    ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, model->output, one));
     for (int il = 0; il < n_layer; ++il) {
         struct my_llama_layer & layer = model->layers[il];
-        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.attention_norm, 1.0f));
-        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.ffn_norm, 1.0f));
-        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.wq, 1.0f));
-        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.wk, 1.0f));
-        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.wv, 1.0f));
-        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.wo, 1.0f));
-        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.w1, 1.0f));
-        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.w2, 1.0f));
-        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.w3, 1.0f));
+        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.attention_norm, one));
+        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.ffn_norm, one));
+        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.wq, one));
+        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.wk, one));
+        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.wv, one));
+        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.wo, one));
+        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.w1, one));
+        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.w2, one));
+        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.w3, one));
     }
 
     // allocating checkpoints in one block to reduce memory fragmentation
@@ -1108,7 +1110,7 @@ static void write_tensor(struct llama_file * file, struct ggml_tensor * tensor,
         name = ggml_get_name(tensor);
     }
     uint32_t name_len = strlen(name);
-    uint32_t nd = ggml_n_dims(tensor);
+    uint32_t nd = tensor->n_dims;
     uint32_t ne[4] = { (uint32_t)tensor->ne[0],
                        (uint32_t)tensor->ne[1],
                        (uint32_t)tensor->ne[2],
@@ -1618,6 +1620,8 @@ int main(int argc, char ** argv) {
     opt->params.adam.gclip              = params.common.adam_gclip;
     opt->params.adam.eps_f              = params.common.adam_eps_f;
 
+    ggml_allocr * alloc = NULL;
+
     printf("%s: init model\n", __func__);
     bool existed = load_checkpoint_lora_file(params.common.fn_checkpoint_in, &model, &lora, train);
 
@@ -1721,9 +1725,10 @@ int main(int argc, char ** argv) {
 
     // allocate input tensors
     mem_input_data.resize(max_input_size);
-    ggml_allocr_t alloc_inps = ggml_allocr_new(mem_input_data.data(), mem_input_data.size(), tensor_alignment);
-    ggml_allocr_alloc(alloc_inps, tokens_input);
-    ggml_allocr_alloc(alloc_inps, target_probs);
+    alloc = ggml_allocr_new(mem_input_data.data(), mem_input_data.size(), tensor_alignment);
+    ggml_allocr_alloc(alloc, tokens_input);
+    ggml_allocr_alloc(alloc, target_probs);
+    ggml_allocr_free(alloc);
 
     // context for compute tensors without their data
     const size_t estimated_compute_size_wo_data = (
@@ -1750,7 +1755,7 @@ int main(int argc, char ** argv) {
     // find best evaluation order
     for (unsigned order = 0; order < (unsigned) GGML_CGRAPH_EVAL_ORDER_COUNT; ++order) {
         ctx_compute = ggml_init(ctx_compute_params);
-        ggml_allocr_t alloc = ggml_allocr_new_measure(tensor_alignment);
+        alloc = ggml_allocr_new_measure(tensor_alignment);
         gf = ggml_new_graph_custom(ctx_compute, LLAMA_TRAIN_MAX_NODES, true);
         gf->order = (enum ggml_cgraph_eval_order) order;
         gb = ggml_new_graph_custom(ctx_compute, LLAMA_TRAIN_MAX_NODES, true);
@@ -1783,7 +1788,7 @@ int main(int argc, char ** argv) {
     // allocate compute tensors
     mem_compute_data.resize(max_compute_size);
     ctx_compute = ggml_init(ctx_compute_params);
-    ggml_allocr_t alloc = ggml_allocr_new(mem_compute_data.data(), mem_compute_data.size(), tensor_alignment);
+    alloc = ggml_allocr_new(mem_compute_data.data(), mem_compute_data.size(), tensor_alignment);
     gf = ggml_new_graph_custom(ctx_compute, LLAMA_TRAIN_MAX_NODES, true);
     gf->order = best_order;
     gb = ggml_new_graph_custom(ctx_compute, LLAMA_TRAIN_MAX_NODES, true);
@@ -1799,8 +1804,6 @@ int main(int argc, char ** argv) {
         params.common.use_checkpointing
     );
     ggml_allocr_free(alloc);
-    ggml_allocr_free(alloc_inps);
-
 
     // tokenize data
     std::vector<llama_token> train_tokens;

examples/gguf/CMakeLists.txt

@@ -1,5 +1,5 @@
 set(TARGET gguf)
 add_executable(${TARGET} gguf.cpp)
 install(TARGETS ${TARGET} RUNTIME)
-target_link_libraries(${TARGET} PRIVATE ggml ${CMAKE_THREAD_LIBS_INIT})
+target_link_libraries(${TARGET} PRIVATE llama ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_11)

examples/gguf/gguf.cpp

@@ -1,4 +1,5 @@
 #include "ggml.h"
+#include "llama.h"
 
 #include <cstdio>
 #include <cinttypes>
@@ -194,7 +195,7 @@ static bool gguf_ex_read_1(const std::string & fname) {
 
             struct ggml_tensor * cur = ggml_get_tensor(ctx_data, name);
 
-            printf("%s: tensor[%d]: n_dims = %d, name = %s, data = %p\n", __func__, i, ggml_n_dims(cur), cur->name, cur->data);
+            printf("%s: tensor[%d]: n_dims = %d, name = %s, data = %p\n", __func__, i, cur->n_dims, cur->name, cur->data);
 
             // print first 10 elements
             const float * data = (const float *) cur->data;

examples/llama.swiftui/.gitignore

@@ -1,2 +1 @@
 xcuserdata
-xcshareddata

examples/llama.swiftui/llama.cpp.swift/LibLlama.swift

@@ -6,34 +6,16 @@ enum LlamaError: Error {
     case couldNotInitializeContext
 }
 
-func llama_batch_clear(_ batch: inout llama_batch) {
-    batch.n_tokens = 0
-}
-
-func llama_batch_add(_ batch: inout llama_batch, _ id: llama_token, _ pos: llama_pos, _ seq_ids: [llama_seq_id], _ logits: Bool) {
-    batch.token   [Int(batch.n_tokens)] = id
-    batch.pos     [Int(batch.n_tokens)] = pos
-    batch.n_seq_id[Int(batch.n_tokens)] = Int32(seq_ids.count)
-    for i in 0..<seq_ids.count {
-        batch.seq_id[Int(batch.n_tokens)]![Int(i)] = seq_ids[i]
-    }
-    batch.logits  [Int(batch.n_tokens)] = logits ? 1 : 0
-
-    batch.n_tokens += 1
-}
-
 actor LlamaContext {
     private var model: OpaquePointer
     private var context: OpaquePointer
     private var batch: llama_batch
     private var tokens_list: [llama_token]
-
     /// This variable is used to store temporarily invalid cchars
     private var temporary_invalid_cchars: [CChar]
 
-    var n_len: Int32 = 64
+    var n_len: Int32 = 512
     var n_cur: Int32 = 0
-
     var n_decode: Int32 = 0
 
     init(model: OpaquePointer, context: OpaquePointer) {
@@ -45,34 +27,25 @@ actor LlamaContext {
     }
 
     deinit {
-        llama_batch_free(batch)
         llama_free(context)
         llama_free_model(model)
         llama_backend_free()
     }
 
-    static func create_context(path: String) throws -> LlamaContext {
+    static func createContext(path: String) throws -> LlamaContext {
         llama_backend_init(false)
-        var model_params = llama_model_default_params()
+        let model_params = llama_model_default_params()
 
-#if targetEnvironment(simulator)
-        model_params.n_gpu_layers = 0
-        print("Running on simulator, force use n_gpu_layers = 0")
-#endif
         let model = llama_load_model_from_file(path, model_params)
         guard let model else {
             print("Could not load model at \(path)")
             throw LlamaError.couldNotInitializeContext
         }
-
-        let n_threads = max(1, min(8, ProcessInfo.processInfo.processorCount - 2))
-        print("Using \(n_threads) threads")
-
         var ctx_params = llama_context_default_params()
-        ctx_params.seed  = 1234
+        ctx_params.seed = 1234
         ctx_params.n_ctx = 2048
-        ctx_params.n_threads       = UInt32(n_threads)
-        ctx_params.n_threads_batch = UInt32(n_threads)
+        ctx_params.n_threads = 8
+        ctx_params.n_threads_batch = 8
 
         let context = llama_new_context_with_model(model, ctx_params)
         guard let context else {
@@ -83,26 +56,6 @@ actor LlamaContext {
         return LlamaContext(model: model, context: context)
     }
 
-    func model_info() -> String {
-        let result = UnsafeMutablePointer<Int8>.allocate(capacity: 256)
-        result.initialize(repeating: Int8(0), count: 256)
-        defer {
-            result.deallocate()
-        }
-
-        // TODO: this is probably very stupid way to get the string from C
-
-        let nChars = llama_model_desc(model, result, 256)
-        let bufferPointer = UnsafeBufferPointer(start: result, count: Int(nChars))
-
-        var SwiftString = ""
-        for char in bufferPointer {
-            SwiftString.append(Character(UnicodeScalar(UInt8(char))))
-        }
-
-        return SwiftString
-    }
-
     func get_n_tokens() -> Int32 {
         return batch.n_tokens;
     }
@@ -126,11 +79,16 @@ actor LlamaContext {
             print(String(cString: token_to_piece(token: id) + [0]))
         }
 
-        llama_batch_clear(&batch)
+        // batch = llama_batch_init(512, 0) // done in init()
+        batch.n_tokens = Int32(tokens_list.count)
 
-        for i1 in 0..<tokens_list.count {
+        for i1 in 0..<batch.n_tokens {
             let i = Int(i1)
-            llama_batch_add(&batch, tokens_list[i], Int32(i), [0], false)
+            batch.token[i] = tokens_list[i]
+            batch.pos[i] = i1
+            batch.n_seq_id[Int(i)] = 1
+            batch.seq_id[Int(i)]![0] = 0
+            batch.logits[i] = 0
         }
         batch.logits[Int(batch.n_tokens) - 1] = 1 // true
 
@@ -183,11 +141,18 @@ actor LlamaContext {
         print(new_token_str)
         // tokens_list.append(new_token_id)
 
-        llama_batch_clear(&batch)
-        llama_batch_add(&batch, new_token_id, n_cur, [0], true)
+        batch.n_tokens = 0
+
+        batch.token[Int(batch.n_tokens)] = new_token_id
+        batch.pos[Int(batch.n_tokens)] = n_cur
+        batch.n_seq_id[Int(batch.n_tokens)] = 1
+        batch.seq_id[Int(batch.n_tokens)]![0] = 0
+        batch.logits[Int(batch.n_tokens)] = 1 // true
+        batch.n_tokens += 1
 
         n_decode += 1
-        n_cur    += 1
+
+        n_cur += 1
 
         if llama_decode(context, batch) != 0 {
             print("failed to evaluate llama!")
@@ -196,111 +161,14 @@ actor LlamaContext {
         return new_token_str
     }
 
-    func bench(pp: Int, tg: Int, pl: Int, nr: Int = 1) -> String {
-        var pp_avg: Double = 0
-        var tg_avg: Double = 0
-
-        var pp_std: Double = 0
-        var tg_std: Double = 0
-
-        for _ in 0..<nr {
-            // bench prompt processing
-
-            llama_batch_clear(&batch)
-
-            let n_tokens = pp
-
-            for i in 0..<n_tokens {
-                llama_batch_add(&batch, 0, Int32(i), [0], false)
-            }
-            batch.logits[Int(batch.n_tokens) - 1] = 1 // true
-
-            llama_kv_cache_clear(context)
-
-            let t_pp_start = ggml_time_us()
-
-            if llama_decode(context, batch) != 0 {
-                print("llama_decode() failed during prompt")
-            }
-
-            let t_pp_end = ggml_time_us()
-
-            // bench text generation
-
-            llama_kv_cache_clear(context)
-
-            let t_tg_start = ggml_time_us()
-
-            for i in 0..<tg {
-                llama_batch_clear(&batch)
-
-                for j in 0..<pl {
-                    llama_batch_add(&batch, 0, Int32(i), [Int32(j)], true)
-                }
-
-                if llama_decode(context, batch) != 0 {
-                    print("llama_decode() failed during text generation")
-                }
-            }
-
-            let t_tg_end = ggml_time_us()
-
-            llama_kv_cache_clear(context)
-
-            let t_pp = Double(t_pp_end - t_pp_start) / 1000000.0
-            let t_tg = Double(t_tg_end - t_tg_start) / 1000000.0
-
-            let speed_pp = Double(pp)    / t_pp
-            let speed_tg = Double(pl*tg) / t_tg
-
-            pp_avg += speed_pp
-            tg_avg += speed_tg
-
-            pp_std += speed_pp * speed_pp
-            tg_std += speed_tg * speed_tg
-
-            print("pp \(speed_pp) t/s, tg \(speed_tg) t/s")
-        }
-
-        pp_avg /= Double(nr)
-        tg_avg /= Double(nr)
-
-        if nr > 1 {
-            pp_std = sqrt(pp_std / Double(nr - 1) - pp_avg * pp_avg * Double(nr) / Double(nr - 1))
-            tg_std = sqrt(tg_std / Double(nr - 1) - tg_avg * tg_avg * Double(nr) / Double(nr - 1))
-        } else {
-            pp_std = 0
-            tg_std = 0
-        }
-
-        let model_desc     = model_info();
-        let model_size     = String(format: "%.2f GiB", Double(llama_model_size(model)) / 1024.0 / 1024.0 / 1024.0);
-        let model_n_params = String(format: "%.2f B", Double(llama_model_n_params(model)) / 1e9);
-        let backend        = "Metal";
-        let pp_avg_str     = String(format: "%.2f", pp_avg);
-        let tg_avg_str     = String(format: "%.2f", tg_avg);
-        let pp_std_str     = String(format: "%.2f", pp_std);
-        let tg_std_str     = String(format: "%.2f", tg_std);
-
-        var result = ""
-
-        result += String("| model | size | params | backend | test | t/s |\n")
-        result += String("| --- | --- | --- | --- | --- | --- |\n")
-        result += String("| \(model_desc) | \(model_size) | \(model_n_params) | \(backend) | pp \(pp) | \(pp_avg_str) ± \(pp_std_str) |\n")
-        result += String("| \(model_desc) | \(model_size) | \(model_n_params) | \(backend) | tg \(tg) | \(tg_avg_str) ± \(tg_std_str) |\n")
-
-        return result;
-    }
-
     func clear() {
         tokens_list.removeAll()
         temporary_invalid_cchars.removeAll()
-        llama_kv_cache_clear(context)
     }
 
     private func tokenize(text: String, add_bos: Bool) -> [llama_token] {
         let utf8Count = text.utf8.count
-        let n_tokens = utf8Count + (add_bos ? 1 : 0) + 1
+        let n_tokens = utf8Count + (add_bos ? 1 : 0)
         let tokens = UnsafeMutablePointer<llama_token>.allocate(capacity: n_tokens)
         let tokenCount = llama_tokenize(model, text, Int32(utf8Count), tokens, Int32(n_tokens), add_bos, false)
 

examples/llama.swiftui/llama.swiftui.xcodeproj/project.pbxproj

@@ -1,483 +1,481 @@
 // !$*UTF8*$!
 {
-	archiveVersion = 1;
-	classes = {
-	};
-	objectVersion = 56;
-	objects = {
+    archiveVersion = 1;
+    classes = {
+    };
+    objectVersion = 56;
+    objects = {
 
 /* Begin PBXBuildFile section */
-		542376082B0D9BFB008E6A1C /* ggml-quants.c in Sources */ = {isa = PBXBuildFile; fileRef = 542376072B0D9BFB008E6A1C /* ggml-quants.c */; settings = {COMPILER_FLAGS = "-O3"; }; };
-		5423760B2B0D9C4B008E6A1C /* ggml-backend.c in Sources */ = {isa = PBXBuildFile; fileRef = 5423760A2B0D9C4B008E6A1C /* ggml-backend.c */; settings = {COMPILER_FLAGS = "-O3"; }; };
-		542378792ACE3F3500834A7B /* ggml-metal.metal in Resources */ = {isa = PBXBuildFile; fileRef = 549479C82AC9E10B00E0F78B /* ggml-metal.metal */; };
-		542EA09D2AC8723900A8AEE9 /* ggml.c in Sources */ = {isa = PBXBuildFile; fileRef = 542EA09B2AC8723900A8AEE9 /* ggml.c */; settings = {COMPILER_FLAGS = "-DGGML_USE_ACCELERATE -DGGML_USE_METAL -DGGML_USE_K_QUANTS -O3"; }; };
-		542EA0A02AC8725700A8AEE9 /* ggml-alloc.c in Sources */ = {isa = PBXBuildFile; fileRef = 542EA09F2AC8725700A8AEE9 /* ggml-alloc.c */; settings = {COMPILER_FLAGS = "-O3"; }; };
-		542EA0A32AC8729100A8AEE9 /* llama.cpp in Sources */ = {isa = PBXBuildFile; fileRef = 542EA0A12AC8729100A8AEE9 /* llama.cpp */; settings = {COMPILER_FLAGS = "-DGGML_USE_K_QUANTS -DGGML_USE_METAL -O3"; }; };
-		549479CB2AC9E16000E0F78B /* Metal.framework in Frameworks */ = {isa = PBXBuildFile; fileRef = 549479CA2AC9E16000E0F78B /* Metal.framework */; };
-		549479CD2AC9E42A00E0F78B /* ggml-metal.m in Sources */ = {isa = PBXBuildFile; fileRef = 549479C52AC9E0F200E0F78B /* ggml-metal.m */; settings = {COMPILER_FLAGS = "-fno-objc-arc -DGGML_SWIFT -DGGML_USE_METAL -O3"; }; };
-		7FA3D2B32B2EA2F600543F92 /* DownloadButton.swift in Sources */ = {isa = PBXBuildFile; fileRef = 7FA3D2B22B2EA2F600543F92 /* DownloadButton.swift */; };
-		8A1C83772AC328BD0096AF73 /* llama_swiftuiApp.swift in Sources */ = {isa = PBXBuildFile; fileRef = 8A1C83762AC328BD0096AF73 /* llama_swiftuiApp.swift */; };
-		8A1C83792AC328BD0096AF73 /* ContentView.swift in Sources */ = {isa = PBXBuildFile; fileRef = 8A1C83782AC328BD0096AF73 /* ContentView.swift */; };
-		8A1C837B2AC328BE0096AF73 /* Assets.xcassets in Resources */ = {isa = PBXBuildFile; fileRef = 8A1C837A2AC328BE0096AF73 /* Assets.xcassets */; };
-		8A1C837E2AC328BE0096AF73 /* Preview Assets.xcassets in Resources */ = {isa = PBXBuildFile; fileRef = 8A1C837D2AC328BE0096AF73 /* Preview Assets.xcassets */; };
-		8A39BE0A2AC7601100BFEB40 /* Accelerate.framework in Frameworks */ = {isa = PBXBuildFile; fileRef = 8A39BE092AC7601000BFEB40 /* Accelerate.framework */; };
-		8A3F84242AC4C891005E2EE8 /* models in Resources */ = {isa = PBXBuildFile; fileRef = 8A3F84232AC4C891005E2EE8 /* models */; };
-		8A907F332AC7138A006146EA /* LibLlama.swift in Sources */ = {isa = PBXBuildFile; fileRef = 8A907F322AC7134E006146EA /* LibLlama.swift */; };
-		8A9F7C4D2AC332EE008AE1EA /* LlamaState.swift in Sources */ = {isa = PBXBuildFile; fileRef = 8A9F7C4C2AC332EE008AE1EA /* LlamaState.swift */; };
+        542376082B0D9BFB008E6A1C /* ggml-quants.c in Sources */ = {isa = PBXBuildFile; fileRef = 542376072B0D9BFB008E6A1C /* ggml-quants.c */; };
+        5423760B2B0D9C4B008E6A1C /* ggml-backend.c in Sources */ = {isa = PBXBuildFile; fileRef = 5423760A2B0D9C4B008E6A1C /* ggml-backend.c */; };
+        542378792ACE3F3500834A7B /* ggml-metal.metal in Resources */ = {isa = PBXBuildFile; fileRef = 549479C82AC9E10B00E0F78B /* ggml-metal.metal */; };
+        542EA09D2AC8723900A8AEE9 /* ggml.c in Sources */ = {isa = PBXBuildFile; fileRef = 542EA09B2AC8723900A8AEE9 /* ggml.c */; settings = {COMPILER_FLAGS = "-DGGML_USE_ACCELERATE -DGGML_USE_METAL -DGGML_USE_K_QUANTS -O3"; }; };
+        542EA0A02AC8725700A8AEE9 /* ggml-alloc.c in Sources */ = {isa = PBXBuildFile; fileRef = 542EA09F2AC8725700A8AEE9 /* ggml-alloc.c */; };
+        542EA0A32AC8729100A8AEE9 /* llama.cpp in Sources */ = {isa = PBXBuildFile; fileRef = 542EA0A12AC8729100A8AEE9 /* llama.cpp */; settings = {COMPILER_FLAGS = "-DGGML_USE_K_QUANTS -DGGML_USE_METAL -O3"; }; };
+        549479CB2AC9E16000E0F78B /* Metal.framework in Frameworks */ = {isa = PBXBuildFile; fileRef = 549479CA2AC9E16000E0F78B /* Metal.framework */; };
+        549479CD2AC9E42A00E0F78B /* ggml-metal.m in Sources */ = {isa = PBXBuildFile; fileRef = 549479C52AC9E0F200E0F78B /* ggml-metal.m */; settings = {COMPILER_FLAGS = "-fno-objc-arc -DGGML_SWIFT -DGGML_USE_METAL -O3"; }; };
+        8A1C83772AC328BD0096AF73 /* llama_swiftuiApp.swift in Sources */ = {isa = PBXBuildFile; fileRef = 8A1C83762AC328BD0096AF73 /* llama_swiftuiApp.swift */; };
+        8A1C83792AC328BD0096AF73 /* ContentView.swift in Sources */ = {isa = PBXBuildFile; fileRef = 8A1C83782AC328BD0096AF73 /* ContentView.swift */; };
+        8A1C837B2AC328BE0096AF73 /* Assets.xcassets in Resources */ = {isa = PBXBuildFile; fileRef = 8A1C837A2AC328BE0096AF73 /* Assets.xcassets */; };
+        8A1C837E2AC328BE0096AF73 /* Preview Assets.xcassets in Resources */ = {isa = PBXBuildFile; fileRef = 8A1C837D2AC328BE0096AF73 /* Preview Assets.xcassets */; };
+        8A39BE0A2AC7601100BFEB40 /* Accelerate.framework in Frameworks */ = {isa = PBXBuildFile; fileRef = 8A39BE092AC7601000BFEB40 /* Accelerate.framework */; };
+        8A3F84242AC4C891005E2EE8 /* models in Resources */ = {isa = PBXBuildFile; fileRef = 8A3F84232AC4C891005E2EE8 /* models */; };
+        8A907F332AC7138A006146EA /* LibLlama.swift in Sources */ = {isa = PBXBuildFile; fileRef = 8A907F322AC7134E006146EA /* LibLlama.swift */; };
+        8A9F7C4D2AC332EE008AE1EA /* LlamaState.swift in Sources */ = {isa = PBXBuildFile; fileRef = 8A9F7C4C2AC332EE008AE1EA /* LlamaState.swift */; };
 /* End PBXBuildFile section */
 
 /* Begin PBXFileReference section */
-		542376062B0D9BEA008E6A1C /* ggml-quants.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; name = "ggml-quants.h"; path = "../../ggml-quants.h"; sourceTree = "<group>"; };
-		542376072B0D9BFB008E6A1C /* ggml-quants.c */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.c; name = "ggml-quants.c"; path = "../../ggml-quants.c"; sourceTree = "<group>"; };
-		542376092B0D9C40008E6A1C /* ggml-backend.h */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.c.h; name = "ggml-backend.h"; path = "../../ggml-backend.h"; sourceTree = "<group>"; };
-		5423760A2B0D9C4B008E6A1C /* ggml-backend.c */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.c; name = "ggml-backend.c"; path = "../../ggml-backend.c"; sourceTree = "<group>"; };
-		542EA09B2AC8723900A8AEE9 /* ggml.c */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.c; name = ggml.c; path = ../../ggml.c; sourceTree = "<group>"; };
-		542EA09C2AC8723900A8AEE9 /* ggml.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; name = ggml.h; path = ../../ggml.h; sourceTree = "<group>"; };
-		542EA09E2AC8725700A8AEE9 /* ggml-alloc.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; name = "ggml-alloc.h"; path = "../../ggml-alloc.h"; sourceTree = "<group>"; };
-		542EA09F2AC8725700A8AEE9 /* ggml-alloc.c */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.c; name = "ggml-alloc.c"; path = "../../ggml-alloc.c"; sourceTree = "<group>"; };
-		542EA0A12AC8729100A8AEE9 /* llama.cpp */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.cpp.cpp; name = llama.cpp; path = ../../llama.cpp; sourceTree = "<group>"; };
-		542EA0A22AC8729100A8AEE9 /* llama.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; name = llama.h; path = ../../llama.h; sourceTree = "<group>"; };
-		549479C52AC9E0F200E0F78B /* ggml-metal.m */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.objc; name = "ggml-metal.m"; path = "../../ggml-metal.m"; sourceTree = "<group>"; };
-		549479C62AC9E0F200E0F78B /* ggml-metal.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; name = "ggml-metal.h"; path = "../../ggml-metal.h"; sourceTree = "<group>"; };
-		549479C82AC9E10B00E0F78B /* ggml-metal.metal */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.metal; name = "ggml-metal.metal"; path = "../../ggml-metal.metal"; sourceTree = "<group>"; };
-		549479CA2AC9E16000E0F78B /* Metal.framework */ = {isa = PBXFileReference; lastKnownFileType = wrapper.framework; name = Metal.framework; path = System/Library/Frameworks/Metal.framework; sourceTree = SDKROOT; };
-		7FA3D2B22B2EA2F600543F92 /* DownloadButton.swift */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.swift; path = DownloadButton.swift; sourceTree = "<group>"; };
-		8A08D20A2AC73B1500FE6CD4 /* bridging-header.h */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.c.h; path = "bridging-header.h"; sourceTree = "<group>"; };
-		8A1C83732AC328BD0096AF73 /* llama.swiftui.app */ = {isa = PBXFileReference; explicitFileType = wrapper.application; includeInIndex = 0; path = llama.swiftui.app; sourceTree = BUILT_PRODUCTS_DIR; };
-		8A1C83762AC328BD0096AF73 /* llama_swiftuiApp.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = llama_swiftuiApp.swift; sourceTree = "<group>"; };
-		8A1C83782AC328BD0096AF73 /* ContentView.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = ContentView.swift; sourceTree = "<group>"; };
-		8A1C837A2AC328BE0096AF73 /* Assets.xcassets */ = {isa = PBXFileReference; lastKnownFileType = folder.assetcatalog; path = Assets.xcassets; sourceTree = "<group>"; };
-		8A1C837D2AC328BE0096AF73 /* Preview Assets.xcassets */ = {isa = PBXFileReference; lastKnownFileType = folder.assetcatalog; path = "Preview Assets.xcassets"; sourceTree = "<group>"; };
-		8A39BE092AC7601000BFEB40 /* Accelerate.framework */ = {isa = PBXFileReference; lastKnownFileType = wrapper.framework; name = Accelerate.framework; path = System/Library/Frameworks/Accelerate.framework; sourceTree = SDKROOT; };
-		8A3F84232AC4C891005E2EE8 /* models */ = {isa = PBXFileReference; lastKnownFileType = folder; name = models; path = llama.swiftui/Resources/models; sourceTree = "<group>"; };
-		8A907F322AC7134E006146EA /* LibLlama.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = LibLlama.swift; sourceTree = "<group>"; };
-		8A9F7C4C2AC332EE008AE1EA /* LlamaState.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = LlamaState.swift; sourceTree = "<group>"; };
+        542376062B0D9BEA008E6A1C /* ggml-quants.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; name = "ggml-quants.h"; path = "../../ggml-quants.h"; sourceTree = "<group>"; };
+        542376072B0D9BFB008E6A1C /* ggml-quants.c */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.c; name = "ggml-quants.c"; path = "../../ggml-quants.c"; sourceTree = "<group>"; };
+        542376092B0D9C40008E6A1C /* ggml-backend.h */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.c.h; name = "ggml-backend.h"; path = "../../ggml-backend.h"; sourceTree = "<group>"; };
+        5423760A2B0D9C4B008E6A1C /* ggml-backend.c */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.c; name = "ggml-backend.c"; path = "../../ggml-backend.c"; sourceTree = "<group>"; };
+        542EA09B2AC8723900A8AEE9 /* ggml.c */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.c; name = ggml.c; path = ../../ggml.c; sourceTree = "<group>"; };
+        542EA09C2AC8723900A8AEE9 /* ggml.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; name = ggml.h; path = ../../ggml.h; sourceTree = "<group>"; };
+        542EA09E2AC8725700A8AEE9 /* ggml-alloc.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; name = "ggml-alloc.h"; path = "../../ggml-alloc.h"; sourceTree = "<group>"; };
+        542EA09F2AC8725700A8AEE9 /* ggml-alloc.c */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.c; name = "ggml-alloc.c"; path = "../../ggml-alloc.c"; sourceTree = "<group>"; };
+        542EA0A12AC8729100A8AEE9 /* llama.cpp */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.cpp.cpp; name = llama.cpp; path = ../../llama.cpp; sourceTree = "<group>"; };
+        542EA0A22AC8729100A8AEE9 /* llama.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; name = llama.h; path = ../../llama.h; sourceTree = "<group>"; };
+        549479C52AC9E0F200E0F78B /* ggml-metal.m */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.objc; name = "ggml-metal.m"; path = "../../ggml-metal.m"; sourceTree = "<group>"; };
+        549479C62AC9E0F200E0F78B /* ggml-metal.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; name = "ggml-metal.h"; path = "../../ggml-metal.h"; sourceTree = "<group>"; };
+        549479C82AC9E10B00E0F78B /* ggml-metal.metal */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.metal; name = "ggml-metal.metal"; path = "../../ggml-metal.metal"; sourceTree = "<group>"; };
+        549479CA2AC9E16000E0F78B /* Metal.framework */ = {isa = PBXFileReference; lastKnownFileType = wrapper.framework; name = Metal.framework; path = System/Library/Frameworks/Metal.framework; sourceTree = SDKROOT; };
+        8A08D20A2AC73B1500FE6CD4 /* bridging-header.h */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.c.h; path = "bridging-header.h"; sourceTree = "<group>"; };
+        8A1C83732AC328BD0096AF73 /* llama.swiftui.app */ = {isa = PBXFileReference; explicitFileType = wrapper.application; includeInIndex = 0; path = llama.swiftui.app; sourceTree = BUILT_PRODUCTS_DIR; };
+        8A1C83762AC328BD0096AF73 /* llama_swiftuiApp.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = llama_swiftuiApp.swift; sourceTree = "<group>"; };
+        8A1C83782AC328BD0096AF73 /* ContentView.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = ContentView.swift; sourceTree = "<group>"; };
+        8A1C837A2AC328BE0096AF73 /* Assets.xcassets */ = {isa = PBXFileReference; lastKnownFileType = folder.assetcatalog; path = Assets.xcassets; sourceTree = "<group>"; };
+        8A1C837D2AC328BE0096AF73 /* Preview Assets.xcassets */ = {isa = PBXFileReference; lastKnownFileType = folder.assetcatalog; path = "Preview Assets.xcassets"; sourceTree = "<group>"; };
+        8A39BE092AC7601000BFEB40 /* Accelerate.framework */ = {isa = PBXFileReference; lastKnownFileType = wrapper.framework; name = Accelerate.framework; path = System/Library/Frameworks/Accelerate.framework; sourceTree = SDKROOT; };
+        8A3F841F2AC4C824005E2EE8 /* llama-2-7b-chat.Q2_K.gguf */ = {isa = PBXFileReference; lastKnownFileType = file; path = "llama-2-7b-chat.Q2_K.gguf"; sourceTree = "<group>"; };
+        8A3F84232AC4C891005E2EE8 /* models */ = {isa = PBXFileReference; lastKnownFileType = folder; name = models; path = llama.swiftui/Resources/models; sourceTree = "<group>"; };
+        8A907F322AC7134E006146EA /* LibLlama.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = LibLlama.swift; sourceTree = "<group>"; };
+        8A9F7C4C2AC332EE008AE1EA /* LlamaState.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = LlamaState.swift; sourceTree = "<group>"; };
 /* End PBXFileReference section */
 
 /* Begin PBXFrameworksBuildPhase section */
-		8A1C83702AC328BD0096AF73 /* Frameworks */ = {
-			isa = PBXFrameworksBuildPhase;
-			buildActionMask = 2147483647;
-			files = (
-				549479CB2AC9E16000E0F78B /* Metal.framework in Frameworks */,
-				8A39BE0A2AC7601100BFEB40 /* Accelerate.framework in Frameworks */,
-			);
-			runOnlyForDeploymentPostprocessing = 0;
-		};
+        8A1C83702AC328BD0096AF73 /* Frameworks */ = {
+            isa = PBXFrameworksBuildPhase;
+            buildActionMask = 2147483647;
+            files = (
+                549479CB2AC9E16000E0F78B /* Metal.framework in Frameworks */,
+                8A39BE0A2AC7601100BFEB40 /* Accelerate.framework in Frameworks */,
+            );
+            runOnlyForDeploymentPostprocessing = 0;
+        };
 /* End PBXFrameworksBuildPhase section */
 
 /* Begin PBXGroup section */
-		8A08D1F62AC7383900FE6CD4 /* llama.cpp */ = {
-			isa = PBXGroup;
-			children = (
-				5423760A2B0D9C4B008E6A1C /* ggml-backend.c */,
-				542376092B0D9C40008E6A1C /* ggml-backend.h */,
-				542376062B0D9BEA008E6A1C /* ggml-quants.h */,
-				542376072B0D9BFB008E6A1C /* ggml-quants.c */,
-				549479C82AC9E10B00E0F78B /* ggml-metal.metal */,
-				549479C62AC9E0F200E0F78B /* ggml-metal.h */,
-				549479C52AC9E0F200E0F78B /* ggml-metal.m */,
-				542EA09B2AC8723900A8AEE9 /* ggml.c */,
-				542EA09C2AC8723900A8AEE9 /* ggml.h */,
-				542EA09F2AC8725700A8AEE9 /* ggml-alloc.c */,
-				542EA09E2AC8725700A8AEE9 /* ggml-alloc.h */,
-				542EA0A12AC8729100A8AEE9 /* llama.cpp */,
-				542EA0A22AC8729100A8AEE9 /* llama.h */,
-			);
-			name = llama.cpp;
-			sourceTree = "<group>";
-		};
-		8A1C836A2AC328BD0096AF73 = {
-			isa = PBXGroup;
-			children = (
-				8A08D1F62AC7383900FE6CD4 /* llama.cpp */,
-				8A907F312AC7134E006146EA /* llama.cpp.swift */,
-				8A3F84232AC4C891005E2EE8 /* models */,
-				8A1C83752AC328BD0096AF73 /* llama.swiftui */,
-				8A1C83742AC328BD0096AF73 /* Products */,
-				8A39BE082AC7601000BFEB40 /* Frameworks */,
-			);
-			sourceTree = "<group>";
-		};
-		8A1C83742AC328BD0096AF73 /* Products */ = {
-			isa = PBXGroup;
-			children = (
-				8A1C83732AC328BD0096AF73 /* llama.swiftui.app */,
-			);
-			name = Products;
-			sourceTree = "<group>";
-		};
-		8A1C83752AC328BD0096AF73 /* llama.swiftui */ = {
-			isa = PBXGroup;
-			children = (
-				8A3F84102AC4BD85005E2EE8 /* Resources */,
-				8A9F7C4B2AC332DC008AE1EA /* Models */,
-				8A9F7C4A2AC332BF008AE1EA /* UI */,
-				8A1C83762AC328BD0096AF73 /* llama_swiftuiApp.swift */,
-				8A1C837A2AC328BE0096AF73 /* Assets.xcassets */,
-				8A1C837C2AC328BE0096AF73 /* Preview Content */,
-			);
-			path = llama.swiftui;
-			sourceTree = "<group>";
-		};
-		8A1C837C2AC328BE0096AF73 /* Preview Content */ = {
-			isa = PBXGroup;
-			children = (
-				8A1C837D2AC328BE0096AF73 /* Preview Assets.xcassets */,
-			);
-			path = "Preview Content";
-			sourceTree = "<group>";
-		};
-		8A39BE082AC7601000BFEB40 /* Frameworks */ = {
-			isa = PBXGroup;
-			children = (
-				549479CA2AC9E16000E0F78B /* Metal.framework */,
-				8A39BE092AC7601000BFEB40 /* Accelerate.framework */,
-			);
-			name = Frameworks;
-			sourceTree = "<group>";
-		};
-		8A3F84102AC4BD85005E2EE8 /* Resources */ = {
-			isa = PBXGroup;
-			children = (
-				8A3F84112AC4BD8C005E2EE8 /* models */,
-			);
-			path = Resources;
-			sourceTree = "<group>";
-		};
-		8A3F84112AC4BD8C005E2EE8 /* models */ = {
-			isa = PBXGroup;
-			children = (
-			);
-			path = models;
-			sourceTree = "<group>";
-		};
-		8A907F312AC7134E006146EA /* llama.cpp.swift */ = {
-			isa = PBXGroup;
-			children = (
-				8A08D20A2AC73B1500FE6CD4 /* bridging-header.h */,
-				8A907F322AC7134E006146EA /* LibLlama.swift */,
-			);
-			path = llama.cpp.swift;
-			sourceTree = "<group>";
-		};
-		8A9F7C4A2AC332BF008AE1EA /* UI */ = {
-			isa = PBXGroup;
-			children = (
-				7FA3D2B22B2EA2F600543F92 /* DownloadButton.swift */,
-				8A1C83782AC328BD0096AF73 /* ContentView.swift */,
-			);
-			path = UI;
-			sourceTree = "<group>";
-		};
-		8A9F7C4B2AC332DC008AE1EA /* Models */ = {
-			isa = PBXGroup;
-			children = (
-				8A9F7C4C2AC332EE008AE1EA /* LlamaState.swift */,
-			);
-			path = Models;
-			sourceTree = "<group>";
-		};
+        8A08D1F62AC7383900FE6CD4 /* llama.cpp */ = {
+            isa = PBXGroup;
+            children = (
+                5423760A2B0D9C4B008E6A1C /* ggml-backend.c */,
+                542376092B0D9C40008E6A1C /* ggml-backend.h */,
+                542376062B0D9BEA008E6A1C /* ggml-quants.h */,
+                542376072B0D9BFB008E6A1C /* ggml-quants.c */,
+                549479C82AC9E10B00E0F78B /* ggml-metal.metal */,
+                549479C62AC9E0F200E0F78B /* ggml-metal.h */,
+                549479C52AC9E0F200E0F78B /* ggml-metal.m */,
+                542EA09B2AC8723900A8AEE9 /* ggml.c */,
+                542EA09C2AC8723900A8AEE9 /* ggml.h */,
+                542EA09F2AC8725700A8AEE9 /* ggml-alloc.c */,
+                542EA09E2AC8725700A8AEE9 /* ggml-alloc.h */,
+                542EA0A12AC8729100A8AEE9 /* llama.cpp */,
+                542EA0A22AC8729100A8AEE9 /* llama.h */,
+            );
+            name = llama.cpp;
+            sourceTree = "<group>";
+        };
+        8A1C836A2AC328BD0096AF73 = {
+            isa = PBXGroup;
+            children = (
+                8A08D1F62AC7383900FE6CD4 /* llama.cpp */,
+                8A907F312AC7134E006146EA /* llama.cpp.swift */,
+                8A3F84232AC4C891005E2EE8 /* models */,
+                8A1C83752AC328BD0096AF73 /* llama.swiftui */,
+                8A1C83742AC328BD0096AF73 /* Products */,
+                8A39BE082AC7601000BFEB40 /* Frameworks */,
+            );
+            sourceTree = "<group>";
+        };
+        8A1C83742AC328BD0096AF73 /* Products */ = {
+            isa = PBXGroup;
+            children = (
+                8A1C83732AC328BD0096AF73 /* llama.swiftui.app */,
+            );
+            name = Products;
+            sourceTree = "<group>";
+        };
+        8A1C83752AC328BD0096AF73 /* llama.swiftui */ = {
+            isa = PBXGroup;
+            children = (
+                8A3F84102AC4BD85005E2EE8 /* Resources */,
+                8A9F7C4B2AC332DC008AE1EA /* Models */,
+                8A9F7C4A2AC332BF008AE1EA /* UI */,
+                8A1C83762AC328BD0096AF73 /* llama_swiftuiApp.swift */,
+                8A1C837A2AC328BE0096AF73 /* Assets.xcassets */,
+                8A1C837C2AC328BE0096AF73 /* Preview Content */,
+            );
+            path = llama.swiftui;
+            sourceTree = "<group>";
+        };
+        8A1C837C2AC328BE0096AF73 /* Preview Content */ = {
+            isa = PBXGroup;
+            children = (
+                8A1C837D2AC328BE0096AF73 /* Preview Assets.xcassets */,
+            );
+            path = "Preview Content";
+            sourceTree = "<group>";
+        };
+        8A39BE082AC7601000BFEB40 /* Frameworks */ = {
+            isa = PBXGroup;
+            children = (
+                549479CA2AC9E16000E0F78B /* Metal.framework */,
+                8A39BE092AC7601000BFEB40 /* Accelerate.framework */,
+            );
+            name = Frameworks;
+            sourceTree = "<group>";
+        };
+        8A3F84102AC4BD85005E2EE8 /* Resources */ = {
+            isa = PBXGroup;
+            children = (
+                8A3F84112AC4BD8C005E2EE8 /* models */,
+            );
+            path = Resources;
+            sourceTree = "<group>";
+        };
+        8A3F84112AC4BD8C005E2EE8 /* models */ = {
+            isa = PBXGroup;
+            children = (
+                8A3F841F2AC4C824005E2EE8 /* llama-2-7b-chat.Q2_K.gguf */,
+            );
+            path = models;
+            sourceTree = "<group>";
+        };
+        8A907F312AC7134E006146EA /* llama.cpp.swift */ = {
+            isa = PBXGroup;
+            children = (
+                8A08D20A2AC73B1500FE6CD4 /* bridging-header.h */,
+                8A907F322AC7134E006146EA /* LibLlama.swift */,
+            );
+            path = llama.cpp.swift;
+            sourceTree = "<group>";
+        };
+        8A9F7C4A2AC332BF008AE1EA /* UI */ = {
+            isa = PBXGroup;
+            children = (
+                8A1C83782AC328BD0096AF73 /* ContentView.swift */,
+            );
+            path = UI;
+            sourceTree = "<group>";
+        };
+        8A9F7C4B2AC332DC008AE1EA /* Models */ = {
+            isa = PBXGroup;
+            children = (
+                8A9F7C4C2AC332EE008AE1EA /* LlamaState.swift */,
+            );
+            path = Models;
+            sourceTree = "<group>";
+        };
 /* End PBXGroup section */
 
 /* Begin PBXNativeTarget section */
-		8A1C83722AC328BD0096AF73 /* llama.swiftui */ = {
-			isa = PBXNativeTarget;
-			buildConfigurationList = 8A1C83812AC328BE0096AF73 /* Build configuration list for PBXNativeTarget "llama.swiftui" */;
-			buildPhases = (
-				8A1C836F2AC328BD0096AF73 /* Sources */,
-				8A1C83702AC328BD0096AF73 /* Frameworks */,
-				8A1C83712AC328BD0096AF73 /* Resources */,
-			);
-			buildRules = (
-			);
-			dependencies = (
-			);
-			name = llama.swiftui;
-			packageProductDependencies = (
-			);
-			productName = llama.swiftui;
-			productReference = 8A1C83732AC328BD0096AF73 /* llama.swiftui.app */;
-			productType = "com.apple.product-type.application";
-		};
+        8A1C83722AC328BD0096AF73 /* llama.swiftui */ = {
+            isa = PBXNativeTarget;
+            buildConfigurationList = 8A1C83812AC328BE0096AF73 /* Build configuration list for PBXNativeTarget "llama.swiftui" */;
+            buildPhases = (
+                8A1C836F2AC328BD0096AF73 /* Sources */,
+                8A1C83702AC328BD0096AF73 /* Frameworks */,
+                8A1C83712AC328BD0096AF73 /* Resources */,
+            );
+            buildRules = (
+            );
+            dependencies = (
+            );
+            name = llama.swiftui;
+            packageProductDependencies = (
+            );
+            productName = llama.swiftui;
+            productReference = 8A1C83732AC328BD0096AF73 /* llama.swiftui.app */;
+            productType = "com.apple.product-type.application";
+        };
 /* End PBXNativeTarget section */
 
 /* Begin PBXProject section */
-		8A1C836B2AC328BD0096AF73 /* Project object */ = {
-			isa = PBXProject;
-			attributes = {
-				BuildIndependentTargetsInParallel = 1;
-				LastSwiftUpdateCheck = 1500;
-				LastUpgradeCheck = 1500;
-				TargetAttributes = {
-					8A1C83722AC328BD0096AF73 = {
-						CreatedOnToolsVersion = 15.0;
-						LastSwiftMigration = 1500;
-					};
-				};
-			};
-			buildConfigurationList = 8A1C836E2AC328BD0096AF73 /* Build configuration list for PBXProject "llama.swiftui" */;
-			compatibilityVersion = "Xcode 14.0";
-			developmentRegion = en;
-			hasScannedForEncodings = 0;
-			knownRegions = (
-				en,
-				Base,
-			);
-			mainGroup = 8A1C836A2AC328BD0096AF73;
-			packageReferences = (
-			);
-			productRefGroup = 8A1C83742AC328BD0096AF73 /* Products */;
-			projectDirPath = "";
-			projectRoot = "";
-			targets = (
-				8A1C83722AC328BD0096AF73 /* llama.swiftui */,
-			);
-		};
+        8A1C836B2AC328BD0096AF73 /* Project object */ = {
+            isa = PBXProject;
+            attributes = {
+                BuildIndependentTargetsInParallel = 1;
+                LastSwiftUpdateCheck = 1500;
+                LastUpgradeCheck = 1500;
+                TargetAttributes = {
+                    8A1C83722AC328BD0096AF73 = {
+                        CreatedOnToolsVersion = 15.0;
+                        LastSwiftMigration = 1500;
+                    };
+                };
+            };
+            buildConfigurationList = 8A1C836E2AC328BD0096AF73 /* Build configuration list for PBXProject "llama.swiftui" */;
+            compatibilityVersion = "Xcode 14.0";
+            developmentRegion = en;
+            hasScannedForEncodings = 0;
+            knownRegions = (
+                en,
+                Base,
+            );
+            mainGroup = 8A1C836A2AC328BD0096AF73;
+            packageReferences = (
+            );
+            productRefGroup = 8A1C83742AC328BD0096AF73 /* Products */;
+            projectDirPath = "";
+            projectRoot = "";
+            targets = (
+                8A1C83722AC328BD0096AF73 /* llama.swiftui */,
+            );
+        };
 /* End PBXProject section */
 
 /* Begin PBXResourcesBuildPhase section */
-		8A1C83712AC328BD0096AF73 /* Resources */ = {
-			isa = PBXResourcesBuildPhase;
-			buildActionMask = 2147483647;
-			files = (
-				542378792ACE3F3500834A7B /* ggml-metal.metal in Resources */,
-				8A3F84242AC4C891005E2EE8 /* models in Resources */,
-				8A1C837E2AC328BE0096AF73 /* Preview Assets.xcassets in Resources */,
-				8A1C837B2AC328BE0096AF73 /* Assets.xcassets in Resources */,
-			);
-			runOnlyForDeploymentPostprocessing = 0;
-		};
+        8A1C83712AC328BD0096AF73 /* Resources */ = {
+            isa = PBXResourcesBuildPhase;
+            buildActionMask = 2147483647;
+            files = (
+                542378792ACE3F3500834A7B /* ggml-metal.metal in Resources */,
+                8A3F84242AC4C891005E2EE8 /* models in Resources */,
+                8A1C837E2AC328BE0096AF73 /* Preview Assets.xcassets in Resources */,
+                8A1C837B2AC328BE0096AF73 /* Assets.xcassets in Resources */,
+            );
+            runOnlyForDeploymentPostprocessing = 0;
+        };
 /* End PBXResourcesBuildPhase section */
 
 /* Begin PBXSourcesBuildPhase section */
-		8A1C836F2AC328BD0096AF73 /* Sources */ = {
-			isa = PBXSourcesBuildPhase;
-			buildActionMask = 2147483647;
-			files = (
-				542376082B0D9BFB008E6A1C /* ggml-quants.c in Sources */,
-				549479CD2AC9E42A00E0F78B /* ggml-metal.m in Sources */,
-				542EA09D2AC8723900A8AEE9 /* ggml.c in Sources */,
-				8A907F332AC7138A006146EA /* LibLlama.swift in Sources */,
-				542EA0A32AC8729100A8AEE9 /* llama.cpp in Sources */,
-				8A9F7C4D2AC332EE008AE1EA /* LlamaState.swift in Sources */,
-				8A1C83792AC328BD0096AF73 /* ContentView.swift in Sources */,
-				8A1C83772AC328BD0096AF73 /* llama_swiftuiApp.swift in Sources */,
-				7FA3D2B32B2EA2F600543F92 /* DownloadButton.swift in Sources */,
-				542EA0A02AC8725700A8AEE9 /* ggml-alloc.c in Sources */,
-				5423760B2B0D9C4B008E6A1C /* ggml-backend.c in Sources */,
-			);
-			runOnlyForDeploymentPostprocessing = 0;
-		};
+        8A1C836F2AC328BD0096AF73 /* Sources */ = {
+            isa = PBXSourcesBuildPhase;
+            buildActionMask = 2147483647;
+            files = (
+                542376082B0D9BFB008E6A1C /* ggml-quants.c in Sources */,
+                549479CD2AC9E42A00E0F78B /* ggml-metal.m in Sources */,
+                542EA09D2AC8723900A8AEE9 /* ggml.c in Sources */,
+                8A907F332AC7138A006146EA /* LibLlama.swift in Sources */,
+                542EA0A32AC8729100A8AEE9 /* llama.cpp in Sources */,
+                8A9F7C4D2AC332EE008AE1EA /* LlamaState.swift in Sources */,
+                8A1C83792AC328BD0096AF73 /* ContentView.swift in Sources */,
+                8A1C83772AC328BD0096AF73 /* llama_swiftuiApp.swift in Sources */,
+                542EA0A02AC8725700A8AEE9 /* ggml-alloc.c in Sources */,
+                5423760B2B0D9C4B008E6A1C /* ggml-backend.c in Sources */,
+            );
+            runOnlyForDeploymentPostprocessing = 0;
+        };
 /* End PBXSourcesBuildPhase section */
 
 /* Begin XCBuildConfiguration section */
-		8A1C837F2AC328BE0096AF73 /* Debug */ = {
-			isa = XCBuildConfiguration;
-			buildSettings = {
-				ALWAYS_SEARCH_USER_PATHS = NO;
-				ASSETCATALOG_COMPILER_GENERATE_SWIFT_ASSET_SYMBOL_EXTENSIONS = YES;
-				CLANG_ANALYZER_NONNULL = YES;
-				CLANG_ANALYZER_NUMBER_OBJECT_CONVERSION = YES_AGGRESSIVE;
-				CLANG_CXX_LANGUAGE_STANDARD = "gnu++20";
-				CLANG_ENABLE_MODULES = YES;
-				CLANG_ENABLE_OBJC_ARC = YES;
-				CLANG_ENABLE_OBJC_WEAK = YES;
-				CLANG_WARN_BLOCK_CAPTURE_AUTORELEASING = YES;
-				CLANG_WARN_BOOL_CONVERSION = YES;
-				CLANG_WARN_COMMA = YES;
-				CLANG_WARN_CONSTANT_CONVERSION = YES;
-				CLANG_WARN_DEPRECATED_OBJC_IMPLEMENTATIONS = YES;
-				CLANG_WARN_DIRECT_OBJC_ISA_USAGE = YES_ERROR;
-				CLANG_WARN_DOCUMENTATION_COMMENTS = YES;
-				CLANG_WARN_EMPTY_BODY = YES;
-				CLANG_WARN_ENUM_CONVERSION = YES;
-				CLANG_WARN_INFINITE_RECURSION = YES;
-				CLANG_WARN_INT_CONVERSION = YES;
-				CLANG_WARN_NON_LITERAL_NULL_CONVERSION = YES;
-				CLANG_WARN_OBJC_IMPLICIT_RETAIN_SELF = YES;
-				CLANG_WARN_OBJC_LITERAL_CONVERSION = YES;
-				CLANG_WARN_OBJC_ROOT_CLASS = YES_ERROR;
-				CLANG_WARN_QUOTED_INCLUDE_IN_FRAMEWORK_HEADER = YES;
-				CLANG_WARN_RANGE_LOOP_ANALYSIS = YES;
-				CLANG_WARN_STRICT_PROTOTYPES = YES;
-				CLANG_WARN_SUSPICIOUS_MOVE = YES;
-				CLANG_WARN_UNGUARDED_AVAILABILITY = YES_AGGRESSIVE;
-				CLANG_WARN_UNREACHABLE_CODE = YES;
-				CLANG_WARN__DUPLICATE_METHOD_MATCH = YES;
-				COPY_PHASE_STRIP = NO;
-				DEBUG_INFORMATION_FORMAT = dwarf;
-				ENABLE_STRICT_OBJC_MSGSEND = YES;
-				ENABLE_TESTABILITY = YES;
-				ENABLE_USER_SCRIPT_SANDBOXING = YES;
-				GCC_C_LANGUAGE_STANDARD = gnu17;
-				GCC_DYNAMIC_NO_PIC = NO;
-				GCC_NO_COMMON_BLOCKS = YES;
-				GCC_OPTIMIZATION_LEVEL = 0;
-				GCC_PREPROCESSOR_DEFINITIONS = (
-					"DEBUG=1",
-					"$(inherited)",
-				);
-				GCC_WARN_64_TO_32_BIT_CONVERSION = YES;
-				GCC_WARN_ABOUT_RETURN_TYPE = YES_ERROR;
-				GCC_WARN_UNDECLARED_SELECTOR = YES;
-				GCC_WARN_UNINITIALIZED_AUTOS = YES_AGGRESSIVE;
-				GCC_WARN_UNUSED_FUNCTION = YES;
-				GCC_WARN_UNUSED_VARIABLE = YES;
-				IPHONEOS_DEPLOYMENT_TARGET = 17.0;
-				LOCALIZATION_PREFERS_STRING_CATALOGS = YES;
-				MTL_ENABLE_DEBUG_INFO = INCLUDE_SOURCE;
-				MTL_FAST_MATH = YES;
-				ONLY_ACTIVE_ARCH = YES;
-				SDKROOT = iphoneos;
-				SWIFT_ACTIVE_COMPILATION_CONDITIONS = "DEBUG $(inherited)";
-				SWIFT_OPTIMIZATION_LEVEL = "-Onone";
-			};
-			name = Debug;
-		};
-		8A1C83802AC328BE0096AF73 /* Release */ = {
-			isa = XCBuildConfiguration;
-			buildSettings = {
-				ALWAYS_SEARCH_USER_PATHS = NO;
-				ASSETCATALOG_COMPILER_GENERATE_SWIFT_ASSET_SYMBOL_EXTENSIONS = YES;
-				CLANG_ANALYZER_NONNULL = YES;
-				CLANG_ANALYZER_NUMBER_OBJECT_CONVERSION = YES_AGGRESSIVE;
-				CLANG_CXX_LANGUAGE_STANDARD = "gnu++20";
-				CLANG_ENABLE_MODULES = YES;
-				CLANG_ENABLE_OBJC_ARC = YES;
-				CLANG_ENABLE_OBJC_WEAK = YES;
-				CLANG_WARN_BLOCK_CAPTURE_AUTORELEASING = YES;
-				CLANG_WARN_BOOL_CONVERSION = YES;
-				CLANG_WARN_COMMA = YES;
-				CLANG_WARN_CONSTANT_CONVERSION = YES;
-				CLANG_WARN_DEPRECATED_OBJC_IMPLEMENTATIONS = YES;
-				CLANG_WARN_DIRECT_OBJC_ISA_USAGE = YES_ERROR;
-				CLANG_WARN_DOCUMENTATION_COMMENTS = YES;
-				CLANG_WARN_EMPTY_BODY = YES;
-				CLANG_WARN_ENUM_CONVERSION = YES;
-				CLANG_WARN_INFINITE_RECURSION = YES;
-				CLANG_WARN_INT_CONVERSION = YES;
-				CLANG_WARN_NON_LITERAL_NULL_CONVERSION = YES;
-				CLANG_WARN_OBJC_IMPLICIT_RETAIN_SELF = YES;
-				CLANG_WARN_OBJC_LITERAL_CONVERSION = YES;
-				CLANG_WARN_OBJC_ROOT_CLASS = YES_ERROR;
-				CLANG_WARN_QUOTED_INCLUDE_IN_FRAMEWORK_HEADER = YES;
-				CLANG_WARN_RANGE_LOOP_ANALYSIS = YES;
-				CLANG_WARN_STRICT_PROTOTYPES = YES;
-				CLANG_WARN_SUSPICIOUS_MOVE = YES;
-				CLANG_WARN_UNGUARDED_AVAILABILITY = YES_AGGRESSIVE;
-				CLANG_WARN_UNREACHABLE_CODE = YES;
-				CLANG_WARN__DUPLICATE_METHOD_MATCH = YES;
-				COPY_PHASE_STRIP = NO;
-				DEBUG_INFORMATION_FORMAT = "dwarf-with-dsym";
-				ENABLE_NS_ASSERTIONS = NO;
-				ENABLE_STRICT_OBJC_MSGSEND = YES;
-				ENABLE_USER_SCRIPT_SANDBOXING = YES;
-				GCC_C_LANGUAGE_STANDARD = gnu17;
-				GCC_NO_COMMON_BLOCKS = YES;
-				GCC_WARN_64_TO_32_BIT_CONVERSION = YES;
-				GCC_WARN_ABOUT_RETURN_TYPE = YES_ERROR;
-				GCC_WARN_UNDECLARED_SELECTOR = YES;
-				GCC_WARN_UNINITIALIZED_AUTOS = YES_AGGRESSIVE;
-				GCC_WARN_UNUSED_FUNCTION = YES;
-				GCC_WARN_UNUSED_VARIABLE = YES;
-				IPHONEOS_DEPLOYMENT_TARGET = 17.0;
-				LOCALIZATION_PREFERS_STRING_CATALOGS = YES;
-				MTL_ENABLE_DEBUG_INFO = NO;
-				MTL_FAST_MATH = YES;
-				SDKROOT = iphoneos;
-				SWIFT_COMPILATION_MODE = wholemodule;
-				VALIDATE_PRODUCT = YES;
-			};
-			name = Release;
-		};
-		8A1C83822AC328BE0096AF73 /* Debug */ = {
-			isa = XCBuildConfiguration;
-			buildSettings = {
-				ASSETCATALOG_COMPILER_APPICON_NAME = AppIcon;
-				ASSETCATALOG_COMPILER_GLOBAL_ACCENT_COLOR_NAME = AccentColor;
-				CLANG_ENABLE_MODULES = YES;
-				CODE_SIGN_STYLE = Automatic;
-				CURRENT_PROJECT_VERSION = 1;
-				DEVELOPMENT_ASSET_PATHS = "\"llama.swiftui/Preview Content\"";
-				DEVELOPMENT_TEAM = STLSG3FG8Q;
-				ENABLE_PREVIEWS = YES;
-				GENERATE_INFOPLIST_FILE = YES;
-				INFOPLIST_KEY_UIApplicationSceneManifest_Generation = YES;
-				INFOPLIST_KEY_UIApplicationSupportsIndirectInputEvents = YES;
-				INFOPLIST_KEY_UILaunchScreen_Generation = YES;
-				INFOPLIST_KEY_UISupportedInterfaceOrientations_iPad = "UIInterfaceOrientationPortrait UIInterfaceOrientationPortraitUpsideDown UIInterfaceOrientationLandscapeLeft UIInterfaceOrientationLandscapeRight";
-				INFOPLIST_KEY_UISupportedInterfaceOrientations_iPhone = "UIInterfaceOrientationPortrait UIInterfaceOrientationLandscapeLeft UIInterfaceOrientationLandscapeRight";
-				IPHONEOS_DEPLOYMENT_TARGET = 16.0;
-				LD_RUNPATH_SEARCH_PATHS = (
-					"$(inherited)",
-					"@executable_path/Frameworks",
-				);
-				MARKETING_VERSION = 1.0;
-				PRODUCT_BUNDLE_IDENTIFIER = "com.bachittle.llama-swift";
-				PRODUCT_NAME = "$(TARGET_NAME)";
-				SWIFT_EMIT_LOC_STRINGS = YES;
-				SWIFT_OBJC_BRIDGING_HEADER = "llama.cpp.swift/bridging-header.h";
-				SWIFT_OPTIMIZATION_LEVEL = "-Onone";
-				SWIFT_VERSION = 5.0;
-				TARGETED_DEVICE_FAMILY = "1,2";
-			};
-			name = Debug;
-		};
-		8A1C83832AC328BE0096AF73 /* Release */ = {
-			isa = XCBuildConfiguration;
-			buildSettings = {
-				ASSETCATALOG_COMPILER_APPICON_NAME = AppIcon;
-				ASSETCATALOG_COMPILER_GLOBAL_ACCENT_COLOR_NAME = AccentColor;
-				CLANG_ENABLE_MODULES = YES;
-				CODE_SIGN_STYLE = Automatic;
-				CURRENT_PROJECT_VERSION = 1;
-				DEVELOPMENT_ASSET_PATHS = "\"llama.swiftui/Preview Content\"";
-				DEVELOPMENT_TEAM = STLSG3FG8Q;
-				ENABLE_PREVIEWS = YES;
-				GENERATE_INFOPLIST_FILE = YES;
-				INFOPLIST_KEY_UIApplicationSceneManifest_Generation = YES;
-				INFOPLIST_KEY_UIApplicationSupportsIndirectInputEvents = YES;
-				INFOPLIST_KEY_UILaunchScreen_Generation = YES;
-				INFOPLIST_KEY_UISupportedInterfaceOrientations_iPad = "UIInterfaceOrientationPortrait UIInterfaceOrientationPortraitUpsideDown UIInterfaceOrientationLandscapeLeft UIInterfaceOrientationLandscapeRight";
-				INFOPLIST_KEY_UISupportedInterfaceOrientations_iPhone = "UIInterfaceOrientationPortrait UIInterfaceOrientationLandscapeLeft UIInterfaceOrientationLandscapeRight";
-				IPHONEOS_DEPLOYMENT_TARGET = 16.0;
-				LD_RUNPATH_SEARCH_PATHS = (
-					"$(inherited)",
-					"@executable_path/Frameworks",
-				);
-				MARKETING_VERSION = 1.0;
-				PRODUCT_BUNDLE_IDENTIFIER = "com.bachittle.llama-swift";
-				PRODUCT_NAME = "$(TARGET_NAME)";
-				SWIFT_EMIT_LOC_STRINGS = YES;
-				SWIFT_OBJC_BRIDGING_HEADER = "llama.cpp.swift/bridging-header.h";
-				SWIFT_VERSION = 5.0;
-				TARGETED_DEVICE_FAMILY = "1,2";
-			};
-			name = Release;
-		};
+        8A1C837F2AC328BE0096AF73 /* Debug */ = {
+            isa = XCBuildConfiguration;
+            buildSettings = {
+                ALWAYS_SEARCH_USER_PATHS = NO;
+                ASSETCATALOG_COMPILER_GENERATE_SWIFT_ASSET_SYMBOL_EXTENSIONS = YES;
+                CLANG_ANALYZER_NONNULL = YES;
+                CLANG_ANALYZER_NUMBER_OBJECT_CONVERSION = YES_AGGRESSIVE;
+                CLANG_CXX_LANGUAGE_STANDARD = "gnu++20";
+                CLANG_ENABLE_MODULES = YES;
+                CLANG_ENABLE_OBJC_ARC = YES;
+                CLANG_ENABLE_OBJC_WEAK = YES;
+                CLANG_WARN_BLOCK_CAPTURE_AUTORELEASING = YES;
+                CLANG_WARN_BOOL_CONVERSION = YES;
+                CLANG_WARN_COMMA = YES;
+                CLANG_WARN_CONSTANT_CONVERSION = YES;
+                CLANG_WARN_DEPRECATED_OBJC_IMPLEMENTATIONS = YES;
+                CLANG_WARN_DIRECT_OBJC_ISA_USAGE = YES_ERROR;
+                CLANG_WARN_DOCUMENTATION_COMMENTS = YES;
+                CLANG_WARN_EMPTY_BODY = YES;
+                CLANG_WARN_ENUM_CONVERSION = YES;
+                CLANG_WARN_INFINITE_RECURSION = YES;
+                CLANG_WARN_INT_CONVERSION = YES;
+                CLANG_WARN_NON_LITERAL_NULL_CONVERSION = YES;
+                CLANG_WARN_OBJC_IMPLICIT_RETAIN_SELF = YES;
+                CLANG_WARN_OBJC_LITERAL_CONVERSION = YES;
+                CLANG_WARN_OBJC_ROOT_CLASS = YES_ERROR;
+                CLANG_WARN_QUOTED_INCLUDE_IN_FRAMEWORK_HEADER = YES;
+                CLANG_WARN_RANGE_LOOP_ANALYSIS = YES;
+                CLANG_WARN_STRICT_PROTOTYPES = YES;
+                CLANG_WARN_SUSPICIOUS_MOVE = YES;
+                CLANG_WARN_UNGUARDED_AVAILABILITY = YES_AGGRESSIVE;
+                CLANG_WARN_UNREACHABLE_CODE = YES;
+                CLANG_WARN__DUPLICATE_METHOD_MATCH = YES;
+                COPY_PHASE_STRIP = NO;
+                DEBUG_INFORMATION_FORMAT = dwarf;
+                ENABLE_STRICT_OBJC_MSGSEND = YES;
+                ENABLE_TESTABILITY = YES;
+                ENABLE_USER_SCRIPT_SANDBOXING = YES;
+                GCC_C_LANGUAGE_STANDARD = gnu17;
+                GCC_DYNAMIC_NO_PIC = NO;
+                GCC_NO_COMMON_BLOCKS = YES;
+                GCC_OPTIMIZATION_LEVEL = 0;
+                GCC_PREPROCESSOR_DEFINITIONS = (
+                    "DEBUG=1",
+                    "$(inherited)",
+                );
+                GCC_WARN_64_TO_32_BIT_CONVERSION = YES;
+                GCC_WARN_ABOUT_RETURN_TYPE = YES_ERROR;
+                GCC_WARN_UNDECLARED_SELECTOR = YES;
+                GCC_WARN_UNINITIALIZED_AUTOS = YES_AGGRESSIVE;
+                GCC_WARN_UNUSED_FUNCTION = YES;
+                GCC_WARN_UNUSED_VARIABLE = YES;
+                IPHONEOS_DEPLOYMENT_TARGET = 17.0;
+                LOCALIZATION_PREFERS_STRING_CATALOGS = YES;
+                MTL_ENABLE_DEBUG_INFO = INCLUDE_SOURCE;
+                MTL_FAST_MATH = YES;
+                ONLY_ACTIVE_ARCH = YES;
+                SDKROOT = iphoneos;
+                SWIFT_ACTIVE_COMPILATION_CONDITIONS = "DEBUG $(inherited)";
+                SWIFT_OPTIMIZATION_LEVEL = "-Onone";
+            };
+            name = Debug;
+        };
+        8A1C83802AC328BE0096AF73 /* Release */ = {
+            isa = XCBuildConfiguration;
+            buildSettings = {
+                ALWAYS_SEARCH_USER_PATHS = NO;
+                ASSETCATALOG_COMPILER_GENERATE_SWIFT_ASSET_SYMBOL_EXTENSIONS = YES;
+                CLANG_ANALYZER_NONNULL = YES;
+                CLANG_ANALYZER_NUMBER_OBJECT_CONVERSION = YES_AGGRESSIVE;
+                CLANG_CXX_LANGUAGE_STANDARD = "gnu++20";
+                CLANG_ENABLE_MODULES = YES;
+                CLANG_ENABLE_OBJC_ARC = YES;
+                CLANG_ENABLE_OBJC_WEAK = YES;
+                CLANG_WARN_BLOCK_CAPTURE_AUTORELEASING = YES;
+                CLANG_WARN_BOOL_CONVERSION = YES;
+                CLANG_WARN_COMMA = YES;
+                CLANG_WARN_CONSTANT_CONVERSION = YES;
+                CLANG_WARN_DEPRECATED_OBJC_IMPLEMENTATIONS = YES;
+                CLANG_WARN_DIRECT_OBJC_ISA_USAGE = YES_ERROR;
+                CLANG_WARN_DOCUMENTATION_COMMENTS = YES;
+                CLANG_WARN_EMPTY_BODY = YES;
+                CLANG_WARN_ENUM_CONVERSION = YES;
+                CLANG_WARN_INFINITE_RECURSION = YES;
+                CLANG_WARN_INT_CONVERSION = YES;
+                CLANG_WARN_NON_LITERAL_NULL_CONVERSION = YES;
+                CLANG_WARN_OBJC_IMPLICIT_RETAIN_SELF = YES;
+                CLANG_WARN_OBJC_LITERAL_CONVERSION = YES;
+                CLANG_WARN_OBJC_ROOT_CLASS = YES_ERROR;
+                CLANG_WARN_QUOTED_INCLUDE_IN_FRAMEWORK_HEADER = YES;
+                CLANG_WARN_RANGE_LOOP_ANALYSIS = YES;
+                CLANG_WARN_STRICT_PROTOTYPES = YES;
+                CLANG_WARN_SUSPICIOUS_MOVE = YES;
+                CLANG_WARN_UNGUARDED_AVAILABILITY = YES_AGGRESSIVE;
+                CLANG_WARN_UNREACHABLE_CODE = YES;
+                CLANG_WARN__DUPLICATE_METHOD_MATCH = YES;
+                COPY_PHASE_STRIP = NO;
+                DEBUG_INFORMATION_FORMAT = "dwarf-with-dsym";
+                ENABLE_NS_ASSERTIONS = NO;
+                ENABLE_STRICT_OBJC_MSGSEND = YES;
+                ENABLE_USER_SCRIPT_SANDBOXING = YES;
+                GCC_C_LANGUAGE_STANDARD = gnu17;
+                GCC_NO_COMMON_BLOCKS = YES;
+                GCC_WARN_64_TO_32_BIT_CONVERSION = YES;
+                GCC_WARN_ABOUT_RETURN_TYPE = YES_ERROR;
+                GCC_WARN_UNDECLARED_SELECTOR = YES;
+                GCC_WARN_UNINITIALIZED_AUTOS = YES_AGGRESSIVE;
+                GCC_WARN_UNUSED_FUNCTION = YES;
+                GCC_WARN_UNUSED_VARIABLE = YES;
+                IPHONEOS_DEPLOYMENT_TARGET = 17.0;
+                LOCALIZATION_PREFERS_STRING_CATALOGS = YES;
+                MTL_ENABLE_DEBUG_INFO = NO;
+                MTL_FAST_MATH = YES;
+                SDKROOT = iphoneos;
+                SWIFT_COMPILATION_MODE = wholemodule;
+                VALIDATE_PRODUCT = YES;
+            };
+            name = Release;
+        };
+        8A1C83822AC328BE0096AF73 /* Debug */ = {
+            isa = XCBuildConfiguration;
+            buildSettings = {
+                ASSETCATALOG_COMPILER_APPICON_NAME = AppIcon;
+                ASSETCATALOG_COMPILER_GLOBAL_ACCENT_COLOR_NAME = AccentColor;
+                CLANG_ENABLE_MODULES = YES;
+                CODE_SIGN_STYLE = Automatic;
+                CURRENT_PROJECT_VERSION = 1;
+                DEVELOPMENT_ASSET_PATHS = "\"llama.swiftui/Preview Content\"";
+                DEVELOPMENT_TEAM = STLSG3FG8Q;
+                ENABLE_PREVIEWS = YES;
+                GENERATE_INFOPLIST_FILE = YES;
+                INFOPLIST_KEY_UIApplicationSceneManifest_Generation = YES;
+                INFOPLIST_KEY_UIApplicationSupportsIndirectInputEvents = YES;
+                INFOPLIST_KEY_UILaunchScreen_Generation = YES;
+                INFOPLIST_KEY_UISupportedInterfaceOrientations_iPad = "UIInterfaceOrientationPortrait UIInterfaceOrientationPortraitUpsideDown UIInterfaceOrientationLandscapeLeft UIInterfaceOrientationLandscapeRight";
+                INFOPLIST_KEY_UISupportedInterfaceOrientations_iPhone = "UIInterfaceOrientationPortrait UIInterfaceOrientationLandscapeLeft UIInterfaceOrientationLandscapeRight";
+                IPHONEOS_DEPLOYMENT_TARGET = 16.0;
+                LD_RUNPATH_SEARCH_PATHS = (
+                    "$(inherited)",
+                    "@executable_path/Frameworks",
+                );
+                MARKETING_VERSION = 1.0;
+                PRODUCT_BUNDLE_IDENTIFIER = "com.bachittle.llama-swift";
+                PRODUCT_NAME = "$(TARGET_NAME)";
+                SWIFT_EMIT_LOC_STRINGS = YES;
+                SWIFT_OBJC_BRIDGING_HEADER = "llama.cpp.swift/bridging-header.h";
+                SWIFT_OPTIMIZATION_LEVEL = "-Onone";
+                SWIFT_VERSION = 5.0;
+                TARGETED_DEVICE_FAMILY = "1,2";
+            };
+            name = Debug;
+        };
+        8A1C83832AC328BE0096AF73 /* Release */ = {
+            isa = XCBuildConfiguration;
+            buildSettings = {
+                ASSETCATALOG_COMPILER_APPICON_NAME = AppIcon;
+                ASSETCATALOG_COMPILER_GLOBAL_ACCENT_COLOR_NAME = AccentColor;
+                CLANG_ENABLE_MODULES = YES;
+                CODE_SIGN_STYLE = Automatic;
+                CURRENT_PROJECT_VERSION = 1;
+                DEVELOPMENT_ASSET_PATHS = "\"llama.swiftui/Preview Content\"";
+                DEVELOPMENT_TEAM = STLSG3FG8Q;
+                ENABLE_PREVIEWS = YES;
+                GENERATE_INFOPLIST_FILE = YES;
+                INFOPLIST_KEY_UIApplicationSceneManifest_Generation = YES;
+                INFOPLIST_KEY_UIApplicationSupportsIndirectInputEvents = YES;
+                INFOPLIST_KEY_UILaunchScreen_Generation = YES;
+                INFOPLIST_KEY_UISupportedInterfaceOrientations_iPad = "UIInterfaceOrientationPortrait UIInterfaceOrientationPortraitUpsideDown UIInterfaceOrientationLandscapeLeft UIInterfaceOrientationLandscapeRight";
+                INFOPLIST_KEY_UISupportedInterfaceOrientations_iPhone = "UIInterfaceOrientationPortrait UIInterfaceOrientationLandscapeLeft UIInterfaceOrientationLandscapeRight";
+                IPHONEOS_DEPLOYMENT_TARGET = 16.0;
+                LD_RUNPATH_SEARCH_PATHS = (
+                    "$(inherited)",
+                    "@executable_path/Frameworks",
+                );
+                MARKETING_VERSION = 1.0;
+                PRODUCT_BUNDLE_IDENTIFIER = "com.bachittle.llama-swift";
+                PRODUCT_NAME = "$(TARGET_NAME)";
+                SWIFT_EMIT_LOC_STRINGS = YES;
+                SWIFT_OBJC_BRIDGING_HEADER = "llama.cpp.swift/bridging-header.h";
+                SWIFT_VERSION = 5.0;
+                TARGETED_DEVICE_FAMILY = "1,2";
+            };
+            name = Release;
+        };
 /* End XCBuildConfiguration section */
 
 /* Begin XCConfigurationList section */
-		8A1C836E2AC328BD0096AF73 /* Build configuration list for PBXProject "llama.swiftui" */ = {
-			isa = XCConfigurationList;
-			buildConfigurations = (
-				8A1C837F2AC328BE0096AF73 /* Debug */,
-				8A1C83802AC328BE0096AF73 /* Release */,
-			);
-			defaultConfigurationIsVisible = 0;
-			defaultConfigurationName = Release;
-		};
-		8A1C83812AC328BE0096AF73 /* Build configuration list for PBXNativeTarget "llama.swiftui" */ = {
-			isa = XCConfigurationList;
-			buildConfigurations = (
-				8A1C83822AC328BE0096AF73 /* Debug */,
-				8A1C83832AC328BE0096AF73 /* Release */,
-			);
-			defaultConfigurationIsVisible = 0;
-			defaultConfigurationName = Release;
-		};
+        8A1C836E2AC328BD0096AF73 /* Build configuration list for PBXProject "llama.swiftui" */ = {
+            isa = XCConfigurationList;
+            buildConfigurations = (
+                8A1C837F2AC328BE0096AF73 /* Debug */,
+                8A1C83802AC328BE0096AF73 /* Release */,
+            );
+            defaultConfigurationIsVisible = 0;
+            defaultConfigurationName = Release;
+        };
+        8A1C83812AC328BE0096AF73 /* Build configuration list for PBXNativeTarget "llama.swiftui" */ = {
+            isa = XCConfigurationList;
+            buildConfigurations = (
+                8A1C83822AC328BE0096AF73 /* Debug */,
+                8A1C83832AC328BE0096AF73 /* Release */,
+            );
+            defaultConfigurationIsVisible = 0;
+            defaultConfigurationName = Release;
+        };
 /* End XCConfigurationList section */
-	};
-	rootObject = 8A1C836B2AC328BD0096AF73 /* Project object */;
+    };
+    rootObject = 8A1C836B2AC328BD0096AF73 /* Project object */;
 }

examples/llama.swiftui/llama.swiftui/Models/LlamaState.swift

@@ -3,26 +3,24 @@ import Foundation
 @MainActor
 class LlamaState: ObservableObject {
     @Published var messageLog = ""
-    @Published var cacheCleared = false
 
     private var llamaContext: LlamaContext?
-    private var defaultModelUrl: URL? {
-        Bundle.main.url(forResource: "ggml-model", withExtension: "gguf", subdirectory: "models")
+    private var modelUrl: URL? {
+        Bundle.main.url(forResource: "q8_0", withExtension: "gguf", subdirectory: "models")
         // Bundle.main.url(forResource: "llama-2-7b-chat", withExtension: "Q2_K.gguf", subdirectory: "models")
     }
-
     init() {
         do {
-            try loadModel(modelUrl: defaultModelUrl)
+            try loadModel()
         } catch {
             messageLog += "Error!\n"
         }
     }
 
-    func loadModel(modelUrl: URL?) throws {
+    private func loadModel() throws {
         messageLog += "Loading model...\n"
         if let modelUrl {
-            llamaContext = try LlamaContext.create_context(path: modelUrl.path())
+            llamaContext = try LlamaContext.createContext(path: modelUrl.path())
             messageLog += "Loaded model \(modelUrl.lastPathComponent)\n"
         } else {
             messageLog += "Could not locate model\n"
@@ -33,7 +31,7 @@ class LlamaState: ObservableObject {
         guard let llamaContext else {
             return
         }
-
+        messageLog += "Attempting to complete text...\n"
         await llamaContext.completion_init(text: text)
         messageLog += "\(text)"
 
@@ -44,42 +42,4 @@ class LlamaState: ObservableObject {
         await llamaContext.clear()
         messageLog += "\n\ndone\n"
     }
-
-    func bench() async {
-        guard let llamaContext else {
-            return
-        }
-
-        messageLog += "\n"
-        messageLog += "Running benchmark...\n"
-        messageLog += "Model info: "
-        messageLog += await llamaContext.model_info() + "\n"
-
-        let t_start = DispatchTime.now().uptimeNanoseconds
-        await llamaContext.bench(pp: 8, tg: 4, pl: 1) // heat up
-        let t_end = DispatchTime.now().uptimeNanoseconds
-
-        let t_heat = Double(t_end - t_start) / 1_000_000_000.0
-        messageLog += "Heat up time: \(t_heat) seconds, please wait...\n"
-
-        // if more than 5 seconds, then we're probably running on a slow device
-        if t_heat > 5.0 {
-            messageLog += "Heat up time is too long, aborting benchmark\n"
-            return
-        }
-
-        let result = await llamaContext.bench(pp: 512, tg: 128, pl: 1, nr: 3)
-
-        messageLog += "\(result)"
-        messageLog += "\n"
-    }
-
-    func clear() async {
-        guard let llamaContext else {
-            return
-        }
-
-        await llamaContext.clear()
-        messageLog = ""
-    }
 }

examples/llama.swiftui/llama.swiftui/UI/ContentView.swift

@@ -5,132 +5,24 @@ struct ContentView: View {
 
     @State private var multiLineText = ""
 
-    private static func cleanupModelCaches() {
-        // Delete all models (*.gguf)
-        let fileManager = FileManager.default
-        let documentsUrl =  FileManager.default.urls(for: .documentDirectory, in: .userDomainMask)[0]
-        do {
-            let fileURLs = try fileManager.contentsOfDirectory(at: documentsUrl, includingPropertiesForKeys: nil)
-            for fileURL in fileURLs {
-                if fileURL.pathExtension == "gguf" {
-                    try fileManager.removeItem(at: fileURL)
-                }
-            }
-        } catch {
-            print("Error while enumerating files \(documentsUrl.path): \(error.localizedDescription)")
-        }
-    }
-
     var body: some View {
         VStack {
-            ScrollView(.vertical, showsIndicators: true) {
+            ScrollView(.vertical) {
                 Text(llamaState.messageLog)
-                .font(.system(size: 12))
-                .frame(maxWidth: .infinity, alignment: .leading)
-                .padding()
-                .onTapGesture {
-                    UIApplication.shared.sendAction(#selector(UIResponder.resignFirstResponder), to: nil, from: nil, for: nil)
-                }
             }
 
             TextEditor(text: $multiLineText)
-                .frame(height: 80)
+                .frame(height: 200)
                 .padding()
                 .border(Color.gray, width: 0.5)
-
-            HStack {
-                Button("Send") {
-                    sendText()
-                }
-                .padding(8)
-                .background(Color.blue)
-                .foregroundColor(.white)
-                .cornerRadius(8)
-
-                Button("Bench") {
-                    bench()
-                }
-                .padding(8)
-                .background(Color.blue)
-                .foregroundColor(.white)
-                .cornerRadius(8)
-
-                Button("Clear") {
-                    clear()
-                }
-                .padding(8)
-                .background(Color.blue)
-                .foregroundColor(.white)
-                .cornerRadius(8)
-
-                Button("Copy") {
-                    UIPasteboard.general.string = llamaState.messageLog
-                }
-                .padding(8)
-                .background(Color.blue)
-                .foregroundColor(.white)
-                .cornerRadius(8)
-            }
-
-            VStack {
-                DownloadButton(
-                    llamaState: llamaState,
-                    modelName: "TinyLlama-1.1B (Q4_0, 0.6 GiB)",
-                    modelUrl: "https://huggingface.co/TheBloke/TinyLlama-1.1B-1T-OpenOrca-GGUF/resolve/main/tinyllama-1.1b-1t-openorca.Q4_0.gguf?download=true",
-                    filename: "tinyllama-1.1b-1t-openorca.Q4_0.gguf"
-                )
-                .font(.system(size: 12))
-                .padding(.top, 4)
-                .frame(maxWidth: .infinity, alignment: .leading)
-
-                DownloadButton(
-                    llamaState: llamaState,
-                    modelName: "TinyLlama-1.1B (Q8_0, 1.1 GiB)",
-                    modelUrl: "https://huggingface.co/TheBloke/TinyLlama-1.1B-1T-OpenOrca-GGUF/resolve/main/tinyllama-1.1b-1t-openorca.Q8_0.gguf?download=true",
-                    filename: "tinyllama-1.1b-1t-openorca.Q8_0.gguf"
-                )
-                .font(.system(size: 12))
-
-                DownloadButton(
-                    llamaState: llamaState,
-                    modelName: "TinyLlama-1.1B (F16, 2.2 GiB)",
-                    modelUrl: "https://huggingface.co/ggml-org/models/resolve/main/tinyllama-1.1b/ggml-model-f16.gguf?download=true",
-                    filename: "tinyllama-1.1b-f16.gguf"
-                )
-                .font(.system(size: 12))
-                .frame(maxWidth: .infinity, alignment: .leading)
-
-                DownloadButton(
-                    llamaState: llamaState,
-                    modelName: "Phi-2.7B (Q4_0, 1.6 GiB)",
-                    modelUrl: "https://huggingface.co/ggml-org/models/resolve/main/phi-2/ggml-model-q4_0.gguf?download=true",
-                    filename: "phi-2-q4_0.gguf"
-                )
-                .font(.system(size: 12))
-
-                DownloadButton(
-                    llamaState: llamaState,
-                    modelName: "Phi-2.7B (Q8_0, 2.8 GiB)",
-                    modelUrl: "https://huggingface.co/ggml-org/models/resolve/main/phi-2/ggml-model-q8_0.gguf?download=true",
-                    filename: "phi-2-q8_0.gguf"
-                )
-                .font(.system(size: 12))
-                .frame(maxWidth: .infinity, alignment: .leading)
-
-                DownloadButton(
-                    llamaState: llamaState,
-                    modelName: "Mistral-7B-v0.1 (Q4_0, 3.8 GiB)",
-                    modelUrl: "https://huggingface.co/TheBloke/Mistral-7B-v0.1-GGUF/resolve/main/mistral-7b-v0.1.Q4_0.gguf?download=true",
-                    filename: "mistral-7b-v0.1.Q4_0.gguf"
-                )
-                .font(.system(size: 12))
-
-                Button("Clear downloaded models") {
-                    ContentView.cleanupModelCaches()
-                    llamaState.cacheCleared = true
-                }
-                .padding(8)
-                .font(.system(size: 12))
+            Button(action: {
+                sendText()
+            }) {
+                Text("Send")
+                    .padding()
+                    .background(Color.blue)
+                    .foregroundColor(.white)
+                    .cornerRadius(8)
             }
         }
         .padding()
@@ -142,20 +34,9 @@ struct ContentView: View {
             multiLineText = ""
         }
     }
-
-    func bench() {
-        Task {
-            await llamaState.bench()
-        }
-    }
-
-    func clear() {
-        Task {
-            await llamaState.clear()
-        }
-    }
 }
-
-//#Preview {
-//    ContentView()
-//}
+/*
+#Preview {
+    ContentView()
+}
+*/

examples/llama.swiftui/llama.swiftui/UI/DownloadButton.swift

@@ -1,122 +0,0 @@
-import SwiftUI
-
-struct DownloadButton: View {
-    @ObservedObject private var llamaState: LlamaState
-    private var modelName: String
-    private var modelUrl: String
-    private var filename: String
-
-    @State private var status: String
-
-    @State private var downloadTask: URLSessionDownloadTask?
-    @State private var progress = 0.0
-    @State private var observation: NSKeyValueObservation?
-
-    private static func getFileURL(filename: String) -> URL {
-        FileManager.default.urls(for: .documentDirectory, in: .userDomainMask)[0].appendingPathComponent(filename)
-    }
-
-    private func checkFileExistenceAndUpdateStatus() {
-    }
-
-    init(llamaState: LlamaState, modelName: String, modelUrl: String, filename: String) {
-        self.llamaState = llamaState
-        self.modelName = modelName
-        self.modelUrl = modelUrl
-        self.filename = filename
-
-        let fileURL = DownloadButton.getFileURL(filename: filename)
-        status = FileManager.default.fileExists(atPath: fileURL.path) ? "downloaded" : "download"
-    }
-
-    private func download() {
-        status = "downloading"
-        print("Downloading model \(modelName) from \(modelUrl)")
-        guard let url = URL(string: modelUrl) else { return }
-        let fileURL = DownloadButton.getFileURL(filename: filename)
-
-        downloadTask = URLSession.shared.downloadTask(with: url) { temporaryURL, response, error in
-            if let error = error {
-                print("Error: \(error.localizedDescription)")
-                return
-            }
-
-            guard let response = response as? HTTPURLResponse, (200...299).contains(response.statusCode) else {
-                print("Server error!")
-                return
-            }
-
-            do {
-                if let temporaryURL = temporaryURL {
-                    try FileManager.default.copyItem(at: temporaryURL, to: fileURL)
-                    print("Writing to \(filename) completed")
-
-                    llamaState.cacheCleared = false
-
-                    status = "downloaded"
-                }
-            } catch let err {
-                print("Error: \(err.localizedDescription)")
-            }
-        }
-
-        observation = downloadTask?.progress.observe(\.fractionCompleted) { progress, _ in
-            self.progress = progress.fractionCompleted
-        }
-
-        downloadTask?.resume()
-    }
-
-    var body: some View {
-        VStack {
-            if status == "download" {
-                Button(action: download) {
-                    Text("Download " + modelName)
-                }
-            } else if status == "downloading" {
-                Button(action: {
-                    downloadTask?.cancel()
-                    status = "download"
-                }) {
-                    Text("\(modelName) (Downloading \(Int(progress * 100))%)")
-                }
-            } else if status == "downloaded" {
-                Button(action: {
-                    let fileURL = DownloadButton.getFileURL(filename: filename)
-                    if !FileManager.default.fileExists(atPath: fileURL.path) {
-                        download()
-                        return
-                    }
-                    do {
-                        try llamaState.loadModel(modelUrl: fileURL)
-                    } catch let err {
-                        print("Error: \(err.localizedDescription)")
-                    }
-                }) {
-                    Text("\(modelName) (Downloaded)")
-                }
-            } else {
-                Text("Unknown status")
-            }
-        }
-        .onDisappear() {
-            downloadTask?.cancel()
-        }
-        .onChange(of: llamaState.cacheCleared) { newValue in
-            if newValue {
-                downloadTask?.cancel()
-                let fileURL = DownloadButton.getFileURL(filename: filename)
-                status = FileManager.default.fileExists(atPath: fileURL.path) ? "downloaded" : "download"
-            }
-        }
-    }
-}
-
-// #Preview {
-//    DownloadButton(
-//        llamaState: LlamaState(),
-//        modelName: "TheBloke / TinyLlama-1.1B-1T-OpenOrca-GGUF (Q4_0)",
-//        modelUrl: "https://huggingface.co/TheBloke/TinyLlama-1.1B-1T-OpenOrca-GGUF/resolve/main/tinyllama-1.1b-1t-openorca.Q4_0.gguf?download=true",
-//        filename: "tinyllama-1.1b-1t-openorca.Q4_0.gguf"
-//    )
-// }

examples/llava/clip.cpp

@@ -330,6 +330,12 @@ static ggml_cgraph * clip_image_build_graph(const clip_ctx * ctx, const clip_ima
                               ggml_repeat(ctx0, model.pre_ln_b, embeddings));
     }
 
+    struct ggml_tensor * KQ_scale = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, 1);
+    ggml_allocr_alloc(ctx->alloc, KQ_scale);
+    if (!ggml_allocr_is_measure(ctx->alloc)) {
+        ggml_set_f32(KQ_scale, 1.0f / sqrt((float)d_head));
+    }
+
     // loop over layers
     for (int il = 0; il < n_layer - 1; il++) {
         struct ggml_tensor * cur = embeddings; // embeddings = residual, cur = hidden_states
@@ -350,7 +356,7 @@ static ggml_cgraph * clip_image_build_graph(const clip_ctx * ctx, const clip_ima
             struct ggml_tensor * Q =
                 ggml_add(ctx0, ggml_repeat(ctx0, model.layers[il].q_b, cur), ggml_mul_mat(ctx0, model.layers[il].q_w, cur));
 
-            Q = ggml_scale_inplace(ctx0, Q, 1.0f / sqrt((float)d_head));
+            Q = ggml_scale_inplace(ctx0, Q, KQ_scale);
             Q = ggml_reshape_4d(ctx0, Q, d_head, n_head, num_positions, batch_size);
             Q = ggml_cont(ctx0, ggml_permute(ctx0, Q, 0, 2, 1, 3));
             Q = ggml_reshape_3d(ctx0, Q, d_head, num_positions, n_head * batch_size);
@@ -508,7 +514,7 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
             ctx_size += padded_size;
             if (verbosity >= 3) {
                 printf("%s: tensor[%d]: n_dims = %d, name = %s, tensor_size=%zu, padded_size=%zu, offset=%zu\n", __func__, i,
-                       ggml_n_dims(cur), cur->name, tensor_size, padded_size, offset);
+                       cur->n_dims, cur->name, tensor_size, padded_size, offset);
             }
         }
     }
@@ -733,7 +739,7 @@ bool clip_image_preprocess(const clip_ctx * ctx, const clip_image_u8 * img, clip
         temp->ny = longer_side;
         temp->size = 3 * longer_side * longer_side;
         temp->data = new uint8_t[temp->size]();
-        uint8_t bc[3] = {122, 116, 104}; // background color in RGB from LLaVA
+        uint8_t bc[3] = {122, 116, 104}; // bakground color in RGB from LLaVA
 
         // fill with background color
         for (size_t i = 0; i < temp->size; i++) {
@@ -956,7 +962,7 @@ bool clip_model_quantize(const char * fname_inp, const char * fname_out, const i
         }
 
         // quantize only 2D tensors
-        quantize &= (ggml_n_dims(cur) == 2);
+        quantize &= (cur->n_dims == 2);
 
         if (quantize) {
             new_type = type;
@@ -1029,7 +1035,7 @@ bool clip_model_quantize(const char * fname_inp, const char * fname_out, const i
             fout.put(0);
         }
 
-        printf("%s: n_dims = %d | quantize=%d | size = %f MB -> %f MB\n", name.c_str(), ggml_n_dims(cur), quantize,
+        printf("%s: n_dims = %d | quantize=%d | size = %f MB -> %f MB\n", name.c_str(), cur->n_dims, quantize,
                orig_size / 1024.0 / 1024.0, new_size / 1024.0 / 1024.0);
     }
 

examples/llava/convert-image-encoder-to-gguf.py

@@ -51,7 +51,7 @@ def bytes_to_unicode():
     The reversible bpe codes work on unicode strings.
     This means you need a large # of unicode characters in your vocab if you want to avoid UNKs.
     When you're at something like a 10B token dataset you end up needing around 5K for decent coverage.
-    This is a significant percentage of your normal, say, 32K bpe vocab.
+    This is a signficant percentage of your normal, say, 32K bpe vocab.
     To avoid that, we want lookup tables between utf-8 bytes and unicode strings.
     And avoids mapping to whitespace/control characters the bpe code barfs on.
     """

examples/lookahead/README.md

@@ -1,6 +1,6 @@
 # llama.cpp/examples/lookahead
 
-Demonstration of lookahead decoding technique:
+Demonstartion of lookahead decoding technique:
 
 https://lmsys.org/blog/2023-11-21-lookahead-decoding/
 

examples/server/README.md

@@ -222,7 +222,7 @@ node index.js
 
     `content`: Set the text to process.
 
--   **POST** `/infill`: For code infilling. Takes a prefix and a suffix and returns the predicted completion as stream.
+    **POST** `/infill`: For code infilling. Takes a prefix and a suffix and returns the predicted completion as stream.
 
     *Options:*
 

examples/server/json.hpp

@@ -11227,7 +11227,7 @@ class binary_reader
                 }
                 if (is_ndarray) // ndarray dimensional vector can only contain integers, and can not embed another array
                 {
-                    return sax->parse_error(chars_read, get_token_string(), parse_error::create(113, chars_read, exception_message(input_format, "ndarray dimensional vector is not allowed", "size"), nullptr));
+                    return sax->parse_error(chars_read, get_token_string(), parse_error::create(113, chars_read, exception_message(input_format, "ndarray dimentional vector is not allowed", "size"), nullptr));
                 }
                 std::vector<size_t> dim;
                 if (JSON_HEDLEY_UNLIKELY(!get_ubjson_ndarray_size(dim)))

examples/server/public/completion.js

@@ -34,8 +34,7 @@ export async function* llama(prompt, params = {}, config = {}) {
     headers: {
       'Connection': 'keep-alive',
       'Content-Type': 'application/json',
-      'Accept': 'text/event-stream',
-      ...(params.api_key ? {'Authorization': `Bearer ${params.api_key}`} : {})
+      'Accept': 'text/event-stream'
     },
     signal: controller.signal,
   });
@@ -115,7 +114,7 @@ export async function* llama(prompt, params = {}, config = {}) {
   return content;
 }
 
-// Call llama, return an event target that you can subscribe to
+// Call llama, return an event target that you can subcribe to
 //
 // Example:
 //

examples/server/public/index.html

@@ -223,7 +223,7 @@
       repeat_last_n: 256, // 0 = disable penalty, -1 = context size
       repeat_penalty: 1.18, // 1.0 = disabled
       top_k: 40, // <= 0 to use vocab size
-      top_p: 0.95, // 1.0 = disabled
+      top_p: 0.5, // 1.0 = disabled
       min_p: 0.05, // 0 = disabled
       tfs_z: 1.0, // 1.0 = disabled
       typical_p: 1.0, // 1.0 = disabled
@@ -235,11 +235,10 @@
       grammar: '',
       n_probs: 0, // no completion_probabilities,
       image_data: [],
-      cache_prompt: true,
-      api_key: ''
+      cache_prompt: true
     })
 
-    /* START: Support for storing prompt templates and parameters in browsers LocalStorage */
+    /* START: Support for storing prompt templates and parameters in borwser LocalStorage */
 
     const local_storage_storageKey = "llamacpp_server_local_storage";
 
@@ -283,7 +282,7 @@
     let importedTemplates = local_storage_getDataAsObject('user_templates')
 
     if (importedTemplates) {
-      // saved templates were successfully imported.
+      // saved templates were successfuly imported.
 
       console.log('Processing saved templates and updating default template')
       params.value = { ...params.value, image_data: [] };
@@ -304,7 +303,7 @@
     }
 
     function userTemplateResetToDefault() {
-      console.log('Resetting template to default')
+      console.log('Reseting themplate to default')
       selectedUserTemplate.value.name = 'default';
       selectedUserTemplate.value.data = savedUserTemplates.value['default'];
     }
@@ -763,7 +762,7 @@
 
           <fieldset class="two">
             ${IntField({ label: "Predictions", max: 2048, min: -1, name: "n_predict", value: params.value.n_predict })}
-            ${FloatField({ label: "Temperature", max: 2.0, 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 })}
             ${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 })}
@@ -791,10 +790,6 @@
             <fieldset>
               ${IntField({ label: "Show Probabilities", max: 10, min: 0, name: "n_probs", value: params.value.n_probs })}
             </fieldset>
-            <fieldset>
-              <label for="api_key">API Key</label>
-              <input type="text" name="api_key" value="${params.value.api_key}" placeholder="Enter API key" oninput=${updateParams} />
-            </fieldset>
           </details>
         </form>
       `

examples/server/server.cpp

@@ -10,8 +10,7 @@
 // crash the server in debug mode, otherwise send an http 500 error
 #define CPPHTTPLIB_NO_EXCEPTIONS 1
 #endif
-// increase max payload length to allow use of larger context size
-#define CPPHTTPLIB_FORM_URL_ENCODED_PAYLOAD_MAX_LENGTH 1048576
+
 #include "httplib.h"
 #include "json.hpp"
 
@@ -37,7 +36,6 @@ using json = nlohmann::json;
 struct server_params
 {
     std::string hostname = "127.0.0.1";
-    std::string api_key;
     std::string public_path = "examples/server/public";
     int32_t port = 8080;
     int32_t read_timeout = 600;
@@ -378,6 +376,7 @@ struct llama_client_slot
 
     int32_t num_prompt_tokens           = 0;
     int32_t num_prompt_tokens_processed = 0;
+    int32_t multibyte_pending           = 0;
 
     json prompt;
     std::string generated_text;
@@ -426,6 +425,7 @@ struct llama_client_slot
         stopped_word           = false;
         stopped_limit          = false;
         stopping_word          = "";
+        multibyte_pending      = 0;
         n_past                 = 0;
         sent_count             = 0;
         sent_token_probs_index = 0;
@@ -992,36 +992,35 @@ struct llama_server_context
         slot.generated_text += token_str;
         slot.has_next_token = true;
 
-        // check if there is incomplete UTF-8 character at the end
-        bool incomplete = false;
-        for (unsigned i = 1; i < 5 && i <= slot.generated_text.size(); ++i)
+        if (slot.multibyte_pending > 0)
         {
-            unsigned char c = slot.generated_text[slot.generated_text.size() - i];
-            if ((c & 0xC0) == 0x80)
-            {
-                // continuation byte: 10xxxxxx
-                continue;
-            }
+            slot.multibyte_pending -= token_str.size();
+        }
+        else if (token_str.size() == 1)
+        {
+            const char c = token_str[0];
+            // 2-byte characters: 110xxxxx 10xxxxxx
             if ((c & 0xE0) == 0xC0)
             {
-                // 2-byte character: 110xxxxx ...
-                incomplete = i < 2;
+                slot.multibyte_pending = 1;
+                // 3-byte characters: 1110xxxx 10xxxxxx 10xxxxxx
             }
             else if ((c & 0xF0) == 0xE0)
             {
-                // 3-byte character: 1110xxxx ...
-                incomplete = i < 3;
+                slot.multibyte_pending = 2;
+                // 4-byte characters: 11110xxx 10xxxxxx 10xxxxxx 10xxxxxx
             }
             else if ((c & 0xF8) == 0xF0)
             {
-                // 4-byte character: 11110xxx ...
-                incomplete = i < 4;
+                slot.multibyte_pending = 3;
+            }
+            else
+            {
+                slot.multibyte_pending = 0;
             }
-            // else 1-byte character or invalid byte
-            break;
         }
 
-        if (!incomplete)
+        if (slot.multibyte_pending == 0)
         {
             size_t pos = std::min(slot.sent_count, slot.generated_text.size());
             const std::string str_test = slot.generated_text.substr(pos);
@@ -1056,7 +1055,7 @@ struct llama_server_context
             }
         }
 
-        if (incomplete)
+        if (slot.multibyte_pending > 0 && !slot.has_next_token)
         {
             slot.has_next_token = true;
         }
@@ -1955,7 +1954,6 @@ static void server_print_usage(const char *argv0, const gpt_params &params,
     printf("  --host                ip address to listen (default  (default: %s)\n", sparams.hostname.c_str());
     printf("  --port PORT           port to listen (default  (default: %d)\n", sparams.port);
     printf("  --path PUBLIC_PATH    path from which to serve static files (default %s)\n", sparams.public_path.c_str());
-    printf("  --api-key API_KEY     optional api key to enhance server security. If set, requests must include this key for access.\n");
     printf("  -to N, --timeout N    server read/write timeout in seconds (default: %d)\n", sparams.read_timeout);
     printf("  --embedding           enable embedding vector output (default: %s)\n", params.embedding ? "enabled" : "disabled");
     printf("  -np N, --parallel N   number of slots for process requests (default: %d)\n", params.n_parallel);
@@ -2005,15 +2003,6 @@ static void server_params_parse(int argc, char **argv, server_params &sparams,
             }
             sparams.public_path = argv[i];
         }
-        else if (arg == "--api-key")
-        {
-            if (++i >= argc)
-            {
-                invalid_param = true;
-                break;
-            }
-            sparams.api_key = argv[i];
-        }
         else if (arg == "--timeout" || arg == "-to")
         {
             if (++i >= argc)
@@ -2393,7 +2382,6 @@ json oaicompat_completion_params_parse(
     llama_params["__oaicompat"] = true;
 
     // Map OpenAI parameters to llama.cpp parameters
-    llama_params["model"]             = json_value(body, "model", std::string("uknown"));
     llama_params["prompt"]            = format_chatml(body["messages"]); // OpenAI 'messages' to llama.cpp 'prompt'
     llama_params["cache_prompt"]      = json_value(body, "cache_prompt", false);
     llama_params["temperature"]       = json_value(body, "temperature", 0.8);
@@ -2414,7 +2402,7 @@ json oaicompat_completion_params_parse(
     llama_params["ignore_eos"]        = json_value(body, "ignore_eos", false);
     llama_params["tfs_z"]             = json_value(body, "tfs_z", 0.0);
 
-    if (body.count("grammar") != 0) {
+    if (llama_params.count("grammar") != 0) {
         llama_params["grammar"] = json_value(body, "grammar", json::object());
     }
 
@@ -2645,9 +2633,6 @@ static void append_to_generated_text_from_generated_token_probs(llama_server_con
 
 int main(int argc, char **argv)
 {
-#if SERVER_VERBOSE != 1
-    log_disable();
-#endif
     // own arguments required by this example
     gpt_params params;
     server_params sparams;
@@ -2684,32 +2669,6 @@ int main(int argc, char **argv)
 
     httplib::Server svr;
 
-    // Middleware for API key validation
-    auto validate_api_key = [&sparams](const httplib::Request &req, httplib::Response &res) -> bool {
-        // If API key is not set, skip validation
-        if (sparams.api_key.empty()) {
-            return true;
-        }
-
-        // Check for API key in the header
-        auto auth_header = req.get_header_value("Authorization");
-        std::string prefix = "Bearer ";
-        if (auth_header.substr(0, prefix.size()) == prefix) {
-            std::string received_api_key = auth_header.substr(prefix.size());
-            if (received_api_key == sparams.api_key) {
-                return true; // API key is valid
-            }
-        }
-
-        // API key is invalid or not provided
-        res.set_content("Unauthorized: Invalid API Key", "text/plain; charset=utf-8");
-        res.status = 401; // Unauthorized
-
-        LOG_WARNING("Unauthorized: Invalid API Key", {});
-
-        return false;
-    };
-
     svr.set_default_headers({{"Server", "llama.cpp"},
                              {"Access-Control-Allow-Origin", "*"},
                              {"Access-Control-Allow-Headers", "content-type"}});
@@ -2717,28 +2676,28 @@ int main(int argc, char **argv)
     // this is only called if no index.html is found in the public --path
     svr.Get("/", [](const httplib::Request &, httplib::Response &res)
             {
-                res.set_content(reinterpret_cast<const char*>(&index_html), index_html_len, "text/html; charset=utf-8");
+                res.set_content(reinterpret_cast<const char*>(&index_html), index_html_len, "text/html");
                 return false;
             });
 
     // this is only called if no index.js is found in the public --path
     svr.Get("/index.js", [](const httplib::Request &, httplib::Response &res)
             {
-                res.set_content(reinterpret_cast<const char *>(&index_js), index_js_len, "text/javascript; charset=utf-8");
+                res.set_content(reinterpret_cast<const char *>(&index_js), index_js_len, "text/javascript");
                 return false;
             });
 
     // this is only called if no index.html is found in the public --path
     svr.Get("/completion.js", [](const httplib::Request &, httplib::Response &res)
             {
-                res.set_content(reinterpret_cast<const char*>(&completion_js), completion_js_len, "application/javascript; charset=utf-8");
+                res.set_content(reinterpret_cast<const char*>(&completion_js), completion_js_len, "application/javascript");
                 return false;
             });
 
     // this is only called if no index.html is found in the public --path
     svr.Get("/json-schema-to-grammar.mjs", [](const httplib::Request &, httplib::Response &res)
             {
-                res.set_content(reinterpret_cast<const char*>(&json_schema_to_grammar_mjs), json_schema_to_grammar_mjs_len, "application/javascript; charset=utf-8");
+                res.set_content(reinterpret_cast<const char*>(&json_schema_to_grammar_mjs), json_schema_to_grammar_mjs_len, "application/javascript");
                 return false;
             });
 
@@ -2749,26 +2708,23 @@ int main(int argc, char **argv)
                     { "user_name",      llama.name_user.c_str() },
                     { "assistant_name", llama.name_assistant.c_str() }
                 };
-                res.set_content(data.dump(), "application/json; charset=utf-8");
+                res.set_content(data.dump(), "application/json");
             });
 
-    svr.Post("/completion", [&llama, &validate_api_key](const httplib::Request &req, httplib::Response &res)
+    svr.Post("/completion", [&llama](const httplib::Request &req, httplib::Response &res)
             {
-                if (!validate_api_key(req, res)) {
-                    return;
-                }
                 json data = json::parse(req.body);
                 const int task_id = llama.request_completion(data, false, false, -1);
                 if (!json_value(data, "stream", false)) {
                     std::string completion_text;
                     task_result result = llama.next_result(task_id);
                     if (!result.error && result.stop) {
-                        res.set_content(result.result_json.dump(-1, ' ', false, json::error_handler_t::replace), "application/json; charset=utf-8");
+                        res.set_content(result.result_json.dump(-1, ' ', false, json::error_handler_t::replace), "application/json");
                     }
                     else
                     {
                         res.status = 404;
-                        res.set_content(result.result_json["content"], "text/plain; charset=utf-8");
+                        res.set_content(result.result_json["content"], "text/plain");
                         return;
                     }
                 } else {
@@ -2839,15 +2795,12 @@ int main(int argc, char **argv)
                     }}
                 };
 
-                res.set_content(models.dump(), "application/json; charset=utf-8");
+                res.set_content(models.dump(), "application/json");
             });
 
     // TODO: add mount point without "/v1" prefix -- how?
-    svr.Post("/v1/chat/completions", [&llama, &validate_api_key](const httplib::Request &req, httplib::Response &res)
+    svr.Post("/v1/chat/completions", [&llama](const httplib::Request &req, httplib::Response &res)
             {
-                if (!validate_api_key(req, res)) {
-                    return;
-                }
                 json data = oaicompat_completion_params_parse(json::parse(req.body));
 
                 const int task_id = llama.request_completion(data, false, false, -1);
@@ -2861,10 +2814,10 @@ int main(int argc, char **argv)
 
                         res.set_content(oaicompat_result.dump(-1, ' ', false,
                                             json::error_handler_t::replace),
-                                            "application/json; charset=utf-8");
+                                            "application/json");
                     } else {
                         res.status = 500;
-                        res.set_content(result.result_json["content"], "text/plain; charset=utf-8");
+                        res.set_content(result.result_json["content"], "text/plain");
                         return;
                     }
                 } else {
@@ -2916,11 +2869,8 @@ int main(int argc, char **argv)
                 }
             });
 
-    svr.Post("/infill", [&llama, &validate_api_key](const httplib::Request &req, httplib::Response &res)
+    svr.Post("/infill", [&llama](const httplib::Request &req, httplib::Response &res)
             {
-                if (!validate_api_key(req, res)) {
-                    return;
-                }
                 json data = json::parse(req.body);
                 const int task_id = llama.request_completion(data, true, false, -1);
                 if (!json_value(data, "stream", false)) {
@@ -2928,12 +2878,12 @@ int main(int argc, char **argv)
                     task_result result = llama.next_result(task_id);
                     if (!result.error && result.stop)
                     {
-                        res.set_content(result.result_json.dump(-1, ' ', false, json::error_handler_t::replace), "application/json; charset=utf-8");
+                        res.set_content(result.result_json.dump(-1, ' ', false, json::error_handler_t::replace), "application/json");
                     }
                     else
                     {
                         res.status = 404;
-                        res.set_content(result.result_json["content"], "text/plain; charset=utf-8");
+                        res.set_content(result.result_json["content"], "text/plain");
                         return;
                     }
                 } else {
@@ -2982,11 +2932,11 @@ int main(int argc, char **argv)
     svr.Get("/model.json", [&llama](const httplib::Request &, httplib::Response &res)
             {
                 const json data = llama.get_model_props();
-                return res.set_content(data.dump(), "application/json; charset=utf-8");
+                return res.set_content(data.dump(), "application/json");
             });
 
     svr.Options(R"(/.*)", [](const httplib::Request &, httplib::Response &res)
-                { return res.set_content("", "application/json; charset=utf-8"); });
+                { return res.set_content("", "application/json"); });
 
     svr.Post("/tokenize", [&llama](const httplib::Request &req, httplib::Response &res)
             {
@@ -2997,7 +2947,7 @@ int main(int argc, char **argv)
                     tokens = llama.tokenize(body["content"], false);
                 }
                 const json data = format_tokenizer_response(tokens);
-                return res.set_content(data.dump(), "application/json; charset=utf-8");
+                return res.set_content(data.dump(), "application/json");
             });
 
     svr.Post("/detokenize", [&llama](const httplib::Request &req, httplib::Response &res)
@@ -3011,7 +2961,7 @@ int main(int argc, char **argv)
                 }
 
                 const json data = format_detokenized_response(content);
-                return res.set_content(data.dump(), "application/json; charset=utf-8");
+                return res.set_content(data.dump(), "application/json");
             });
 
     svr.Post("/embedding", [&llama](const httplib::Request &req, httplib::Response &res)
@@ -3028,7 +2978,7 @@ int main(int argc, char **argv)
                 }
                 const int task_id = llama.request_completion({ {"prompt", prompt}, { "n_predict", 0} }, false, true, -1);
                 task_result result = llama.next_result(task_id);
-                return res.set_content(result.result_json.dump(), "application/json; charset=utf-8");
+                return res.set_content(result.result_json.dump(), "application/json");
             });
 
     svr.set_logger(log_server_request);
@@ -3049,23 +2999,19 @@ int main(int argc, char **argv)
                 {
                     snprintf(buf, sizeof(buf), fmt, "Unknown Exception");
                 }
-                res.set_content(buf, "text/plain; charset=utf-8");
+                res.set_content(buf, "text/plain");
                 res.status = 500;
             });
 
     svr.set_error_handler([](const httplib::Request &, httplib::Response &res)
             {
-                if (res.status == 401)
-                {
-                    res.set_content("Unauthorized", "text/plain; charset=utf-8");
-                }
                 if (res.status == 400)
                 {
-                    res.set_content("Invalid request", "text/plain; charset=utf-8");
+                    res.set_content("Invalid request", "text/plain");
                 }
-                else if (res.status == 404)
+                else if (res.status != 500)
                 {
-                    res.set_content("File Not Found", "text/plain; charset=utf-8");
+                    res.set_content("File Not Found", "text/plain");
                     res.status = 404;
                 }
             });
@@ -3086,15 +3032,11 @@ int main(int argc, char **argv)
     // to make it ctrl+clickable:
     LOG_TEE("\nllama server listening at http://%s:%d\n\n", sparams.hostname.c_str(), sparams.port);
 
-    std::unordered_map<std::string, std::string> log_data;
-    log_data["hostname"] = sparams.hostname;
-    log_data["port"] = std::to_string(sparams.port);
+    LOG_INFO("HTTP server listening", {
+                                          {"hostname", sparams.hostname},
+                                          {"port", sparams.port},
+                                      });
 
-    if (!sparams.api_key.empty()) {
-        log_data["api_key"] = "api_key: ****" + sparams.api_key.substr(sparams.api_key.length() - 4);
-    }
-
-    LOG_INFO("HTTP server listening", log_data);
     // run the HTTP server in a thread - see comment below
     std::thread t([&]()
             {

examples/speculative/README.md

@@ -1,6 +1,6 @@
 # llama.cpp/examples/speculative
 
-Demonstration of speculative decoding and tree-based speculative decoding techniques
+Demonstartion of speculative decoding and tree-based speculative decoding techniques
 
 More info:
 

examples/speculative/speculative.cpp

@@ -428,7 +428,7 @@ int main(int argc, char ** argv) {
             ++n_past_tgt;
         }
 
-        // the first token is always proposed by the target model before the speculation loop so we erase it here
+        // 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;

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

@@ -369,7 +369,10 @@ static struct ggml_tensor * llama_build_train_graphs(
     checkpoints.push_back(t00);
     checkpoints.push_back(t01);
 
-    const float kv_scale = 1.0f/sqrtf(float(n_embd)/n_head);
+    struct ggml_tensor * kv_scale = NULL;
+    if (!enable_flash_attn) {
+        kv_scale = ggml_new_f32(ctx, 1.0f/sqrtf(float(n_embd)/n_head));
+    }
 
     for (int il = 0; il < n_layer; ++il) {
         struct my_llama_layer & layer = model->layers[il];
@@ -441,13 +444,14 @@ static struct ggml_tensor * llama_build_train_graphs(
         // make sure some tensors are not reallocated by inserting new temporary nodes depending on them
         int n_leafs_before = gb->n_leafs;
         int n_nodes_before = gb->n_nodes;
+        struct ggml_tensor * one = ggml_new_f32(ctx, 1.0f);
         // output tensors
-        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, t35, 1.0f));
-        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, t36, 1.0f));
+        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, t35, one));
+        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, t36, one));
         // input gradient
-        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, t36->grad, 1.0f));
+        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, t36->grad, one));
         // KQ_pos
-        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, KQ_pos, 1.0f));
+        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, KQ_pos, one));
         GGML_ASSERT(t36->grad->data == NULL && t36->grad->view_src == NULL);
 
         ggml_allocr_alloc(alloc, t36->grad);

ggml-alloc.c

@@ -449,10 +449,11 @@ static void init_view(ggml_gallocr_t galloc, struct ggml_tensor * view, bool upd
     if (update_backend) {
         view->backend = view->view_src->backend;
     }
-    // views are initialized in the alloc buffer rather than the view_src buffer
-    view->buffer  = alloc->buffer;
+    view->buffer  = view->view_src->buffer;
     view->data    = (char *)view->view_src->data + view->view_offs;
 
+    // FIXME: the view should be initialized by the owning buffer, but currently this breaks the CUDA backend
+    // due to the ggml_tensor_extra_gpu ring buffer overwriting the KV cache extras
     assert(ggml_tallocr_is_measure(alloc) || !view->buffer || view->buffer->buft == alloc->buffer->buft);
 
     if (!alloc->measure) {
@@ -735,10 +736,6 @@ void ggml_allocr_set_parse_seq(ggml_allocr_t alloc, const int * list, int n) {
 }
 
 void ggml_allocr_free(ggml_allocr_t alloc) {
-    if (alloc == NULL) {
-        return;
-    }
-
     ggml_gallocr_free(alloc->galloc);
     ggml_tallocr_free(alloc->talloc);
     free(alloc);
@@ -778,7 +775,7 @@ ggml_backend_buffer_t ggml_backend_alloc_ctx_tensors_from_buft(struct ggml_conte
     }
 
     if (nbytes == 0) {
-        // all the tensors in the context are already allocated
+        fprintf(stderr, "%s: no tensors to allocate\n", __func__);
         return NULL;
     }
 
@@ -792,11 +789,6 @@ ggml_backend_buffer_t ggml_backend_alloc_ctx_tensors_from_buft(struct ggml_conte
             } else {
                 ggml_backend_view_init(buffer, t);
             }
-        } else {
-            if (t->view_src != NULL) {
-                // view of a pre-allocated tensor
-                ggml_backend_view_init(buffer, t);
-            }
         }
     }
 

ggml-alloc.h

@@ -43,7 +43,7 @@ GGML_API size_t ggml_allocr_alloc_graph(ggml_allocr_t alloc, struct ggml_cgraph
 // ggml-backend v2 API
 //
 
-// Separate tensor and graph allocator objects
+// Seperate tensor and graph allocator objects
 // This is necessary for multi-backend allocation because the graph allocator needs to use multiple tensor allocators
 // The original API is kept as a wrapper around the new API
 

ggml-backend-impl.h

@@ -20,9 +20,6 @@ extern "C" {
         size_t                (*get_alignment)   (ggml_backend_buffer_type_t buft); // tensor alignment
         size_t                (*get_alloc_size)  (ggml_backend_buffer_type_t buft, struct ggml_tensor * tensor); // data size needed to allocate the tensor, including padding
         bool                  (*supports_backend)(ggml_backend_buffer_type_t buft, ggml_backend_t backend); // check if the buffer type is usable by the backend
-        // check if tensor data is in host memory
-        // should be equivalent to supports_backend(buft, ggml_backend_cpu_init())
-        bool                  (*is_host)         (ggml_backend_buffer_type_t buft);
     };
 
     struct ggml_backend_buffer_type {
@@ -34,16 +31,15 @@ extern "C" {
     typedef void * ggml_backend_buffer_context_t;
 
     struct ggml_backend_buffer_i {
-        void   (*free_buffer)    (ggml_backend_buffer_t buffer);
+        void     (*free_buffer)(ggml_backend_buffer_t buffer);
         //void     (*reset)      (ggml_backend_buffer_t buffer); // reset any internal state due to tensor initialization, such as tensor extras
-        void * (*get_base)       (ggml_backend_buffer_t buffer);
-        void   (*init_tensor)    (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
-        void   (*set_tensor)     (ggml_backend_buffer_t buffer,       struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
-        void   (*get_tensor)     (ggml_backend_buffer_t buffer, const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
+        void *   (*get_base)   (ggml_backend_buffer_t buffer);
+        void     (*init_tensor)(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
+        void     (*set_tensor) (ggml_backend_buffer_t buffer,       struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
+        void     (*get_tensor) (ggml_backend_buffer_t buffer, const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
         // (optional) copy tensor between different buffer-type, allow for single-copy tranfers
-        void   (*cpy_tensor_from)(ggml_backend_buffer_t buffer, struct ggml_tensor * src, struct ggml_tensor * dst);
-        void   (*cpy_tensor_to)  (ggml_backend_buffer_t buffer, struct ggml_tensor * src, struct ggml_tensor * dst);
-        void   (*clear)          (ggml_backend_buffer_t buffer, uint8_t value);
+        void (*cpy_tensor_from)(ggml_backend_buffer_t buffer, struct ggml_tensor * src, struct ggml_tensor * dst);
+        void (*cpy_tensor_to)  (ggml_backend_buffer_t buffer, struct ggml_tensor * src, struct ggml_tensor * dst);
     };
 
     struct ggml_backend_buffer {
@@ -82,7 +78,7 @@ extern "C" {
         void (*cpy_tensor_from_async)(ggml_backend_t backend, struct ggml_tensor * src, struct ggml_tensor * dst);
         void (*cpy_tensor_to_async)  (ggml_backend_t backend, struct ggml_tensor * src, struct ggml_tensor * dst);
 
-        void (*synchronize)(ggml_backend_t backend);
+        void (*synchronize)     (ggml_backend_t backend);
 
         // compute graph with a plan
         ggml_backend_graph_plan_t (*graph_plan_create) (ggml_backend_t backend, struct ggml_cgraph * cgraph);

ggml-backend.c

@@ -35,13 +35,6 @@ bool ggml_backend_buft_supports_backend(ggml_backend_buffer_type_t buft, ggml_ba
     return buft->iface.supports_backend(buft, backend);
 }
 
-bool ggml_backend_buft_is_host(ggml_backend_buffer_type_t buft) {
-    if (buft->iface.is_host) {
-        return buft->iface.is_host(buft);
-    }
-    return false;
-}
-
 // backend buffer
 
 ggml_backend_buffer_t ggml_backend_buffer_init(
@@ -101,14 +94,6 @@ size_t ggml_backend_buffer_get_alloc_size(ggml_backend_buffer_t buffer, struct g
     return ggml_backend_buft_get_alloc_size(ggml_backend_buffer_type(buffer), tensor);
 }
 
-void ggml_backend_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
-    buffer->iface.clear(buffer, value);
-}
-
-bool ggml_backend_buffer_is_host(ggml_backend_buffer_t buffer) {
-    return ggml_backend_buft_is_host(ggml_backend_buffer_type(buffer));
-}
-
 ggml_backend_buffer_type_t ggml_backend_buffer_type(ggml_backend_buffer_t buffer) {
     return buffer->buft;
 }
@@ -393,6 +378,7 @@ static void * ggml_backend_cpu_buffer_get_base(ggml_backend_buffer_t buffer) {
 
 static void ggml_backend_cpu_buffer_free_buffer(ggml_backend_buffer_t buffer) {
     free(buffer->context);
+    GGML_UNUSED(buffer);
 }
 
 static void ggml_backend_cpu_buffer_set_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
@@ -425,10 +411,6 @@ static void ggml_backend_cpu_buffer_cpy_tensor_to(ggml_backend_buffer_t buffer,
     GGML_UNUSED(buffer);
 }
 
-static void ggml_backend_cpu_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
-    memset(buffer->context, value, buffer->size);
-}
-
 static struct ggml_backend_buffer_i cpu_backend_buffer_i = {
     /* .free_buffer     = */ ggml_backend_cpu_buffer_free_buffer,
     /* .get_base        = */ ggml_backend_cpu_buffer_get_base,
@@ -437,7 +419,6 @@ static struct ggml_backend_buffer_i cpu_backend_buffer_i = {
     /* .get_tensor      = */ ggml_backend_cpu_buffer_get_tensor,
     /* .cpy_tensor_from = */ ggml_backend_cpu_buffer_cpy_tensor_from,
     /* .cpy_tensor_to   = */ ggml_backend_cpu_buffer_cpy_tensor_to,
-    /* .clear           = */ ggml_backend_cpu_buffer_clear,
 };
 
 // for buffers from ptr, free is not called
@@ -449,7 +430,6 @@ static struct ggml_backend_buffer_i cpu_backend_buffer_i_from_ptr = {
     /* .get_tensor      = */ ggml_backend_cpu_buffer_get_tensor,
     /* .cpy_tensor_from = */ ggml_backend_cpu_buffer_cpy_tensor_from,
     /* .cpy_tensor_to   = */ ggml_backend_cpu_buffer_cpy_tensor_to,
-    /* .clear           = */ ggml_backend_cpu_buffer_clear,
 };
 
 static const size_t TENSOR_ALIGNMENT = 64; // should be enough for AVX 512
@@ -475,70 +455,20 @@ static bool ggml_backend_cpu_buffer_type_supports_backend(ggml_backend_buffer_ty
     GGML_UNUSED(buft);
 }
 
-static bool ggml_backend_cpu_buffer_type_is_host(ggml_backend_buffer_type_t buft) {
-    return true;
-
-    GGML_UNUSED(buft);
-}
-
 ggml_backend_buffer_type_t ggml_backend_cpu_buffer_type(void) {
-    static struct ggml_backend_buffer_type ggml_backend_cpu_buffer_type = {
+    static struct ggml_backend_buffer_type ggml_backend_buffer_type_cpu = {
         /* .iface = */ {
             /* .alloc_buffer     = */ ggml_backend_cpu_buffer_type_alloc_buffer,
             /* .get_alignment    = */ ggml_backend_cpu_buffer_type_get_alignment,
             /* .get_alloc_size   = */ NULL, // defaults to ggml_nbytes
             /* .supports_backend = */ ggml_backend_cpu_buffer_type_supports_backend,
-            /* .is_host          = */ ggml_backend_cpu_buffer_type_is_host,
         },
         /* .context = */ NULL,
     };
 
-    return &ggml_backend_cpu_buffer_type;
+    return &ggml_backend_buffer_type_cpu;
 }
 
-#ifdef GGML_USE_CPU_HBM
-
-// buffer type HBM
-
-#include <hbwmalloc.h>
-
-static void ggml_backend_cpu_hbm_buffer_free_buffer(ggml_backend_buffer_t buffer) {
-    hbw_free(buffer->context);
-}
-
-static ggml_backend_buffer_t ggml_backend_cpu_hbm_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
-    //void * ptr = hbw_malloc(size);
-    void * ptr;
-    int result = hbw_posix_memalign(&ptr, ggml_backend_cpu_buffer_type_get_alignment(buft), size);
-    if (result != 0) {
-        fprintf(stderr, "failed to allocate HBM buffer of size %zu\n", size);
-        return NULL;
-    }
-
-    // FIXME: this is a hack to avoid having to implement a new buffer type
-    ggml_backend_buffer_t buffer = ggml_backend_cpu_buffer_from_ptr(ptr, size);
-    buffer->buft = buft;
-    buffer->iface.free_buffer = ggml_backend_cpu_hbm_buffer_free_buffer;
-
-    return buffer;
-}
-
-ggml_backend_buffer_type_t ggml_backend_cpu_hbm_buffer_type() {
-    static struct ggml_backend_buffer_type ggml_backend_cpu_buffer_type_hbm = {
-        /* .iface    = */ {
-            /* .alloc_buffer     = */ ggml_backend_cpu_hbm_buffer_type_alloc_buffer,
-            /* .get_alignment    = */ ggml_backend_cpu_buffer_type_get_alignment,
-            /* .get_alloc_size   = */ NULL, // defaults to ggml_nbytes
-            /* .supports_backend = */ ggml_backend_cpu_buffer_type_supports_backend,
-            /* .is_host          = */ ggml_backend_cpu_buffer_type_is_host,
-        },
-        /* .context  = */ NULL,
-    };
-
-    return &ggml_backend_cpu_buffer_type_hbm;
-}
-#endif
-
 struct ggml_backend_cpu_context {
     int n_threads;
     void * work_data;
@@ -575,7 +505,7 @@ static ggml_backend_graph_plan_t ggml_backend_cpu_graph_plan_create(ggml_backend
     struct ggml_backend_plan_cpu * cpu_plan = malloc(sizeof(struct ggml_backend_plan_cpu));
 
     cpu_plan->cplan = ggml_graph_plan(cgraph, cpu_ctx->n_threads);
-    cpu_plan->cgraph = *cgraph; // FIXME: deep copy
+    cpu_plan->cgraph = *cgraph;
 
     if (cpu_plan->cplan.work_size > 0) {
         cpu_plan->cplan.work_data = malloc(cpu_plan->cplan.work_size);
@@ -1250,7 +1180,7 @@ void ggml_backend_sched_set_node_backend(ggml_backend_sched_t sched, struct ggml
 // utils
 void ggml_backend_view_init(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) {
     GGML_ASSERT(tensor->buffer == NULL);
-    //GGML_ASSERT(tensor->data == NULL); // views of pre-allocted tensors may have the data set, but still need to be initialized
+    GGML_ASSERT(tensor->data == NULL);
     GGML_ASSERT(tensor->view_src != NULL);
     GGML_ASSERT(tensor->view_src->buffer != NULL);
     GGML_ASSERT(tensor->view_src->data != NULL);

ggml-backend.h

@@ -21,7 +21,6 @@ extern "C" {
     GGML_API size_t ggml_backend_buft_get_alignment (ggml_backend_buffer_type_t buft);
     GGML_API size_t ggml_backend_buft_get_alloc_size(ggml_backend_buffer_type_t buft, struct ggml_tensor * tensor);
     GGML_API bool ggml_backend_buft_supports_backend(ggml_backend_buffer_type_t buft, ggml_backend_t backend);
-    GGML_API bool ggml_backend_buft_is_host         (ggml_backend_buffer_type_t buft);
 
     // buffer
     GGML_API void   ggml_backend_buffer_free          (ggml_backend_buffer_t buffer);
@@ -30,8 +29,6 @@ extern "C" {
     GGML_API void   ggml_backend_buffer_init_tensor   (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
     GGML_API size_t ggml_backend_buffer_get_alignment (ggml_backend_buffer_t buffer);
     GGML_API size_t ggml_backend_buffer_get_alloc_size(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
-    GGML_API void   ggml_backend_buffer_clear         (ggml_backend_buffer_t buffer, uint8_t value);
-    GGML_API bool   ggml_backend_buffer_is_host       (ggml_backend_buffer_t buffer);
     GGML_API ggml_backend_buffer_type_t ggml_backend_buffer_type(ggml_backend_buffer_t buffer);
 
     //
@@ -79,10 +76,6 @@ extern "C" {
 
     GGML_API ggml_backend_buffer_type_t ggml_backend_cpu_buffer_type(void);
 
-#ifdef GGML_USE_CPU_HBM
-    GGML_API ggml_backend_buffer_type_t ggml_backend_cpu_hbm_buffer_type(void);
-#endif
-
     //
     // Backend registry
     //

ggml-metal.h

@@ -98,10 +98,7 @@ GGML_API ggml_backend_t ggml_backend_metal_init(void);
 
 GGML_API bool ggml_backend_is_metal(ggml_backend_t backend);
 
-GGML_API ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void * data, size_t size, size_t max_size);
-
 GGML_API void ggml_backend_metal_set_n_cb(ggml_backend_t backend, int n_cb);
-
 GGML_API ggml_backend_buffer_type_t ggml_backend_metal_buffer_type(void);
 
 // helper to check if the device supports a specific family

ggml-metal.m

@@ -66,11 +66,9 @@ struct ggml_metal_context {
     GGML_METAL_DECL_KERNEL(div_row);
     GGML_METAL_DECL_KERNEL(scale);
     GGML_METAL_DECL_KERNEL(scale_4);
-    GGML_METAL_DECL_KERNEL(tanh);
+    GGML_METAL_DECL_KERNEL(silu);
     GGML_METAL_DECL_KERNEL(relu);
     GGML_METAL_DECL_KERNEL(gelu);
-    GGML_METAL_DECL_KERNEL(gelu_quick);
-    GGML_METAL_DECL_KERNEL(silu);
     GGML_METAL_DECL_KERNEL(soft_max);
     GGML_METAL_DECL_KERNEL(soft_max_4);
     GGML_METAL_DECL_KERNEL(diag_mask_inf);
@@ -88,7 +86,6 @@ struct ggml_metal_context {
     GGML_METAL_DECL_KERNEL(get_rows_q5_K);
     GGML_METAL_DECL_KERNEL(get_rows_q6_K);
     GGML_METAL_DECL_KERNEL(rms_norm);
-    GGML_METAL_DECL_KERNEL(group_norm);
     GGML_METAL_DECL_KERNEL(norm);
     GGML_METAL_DECL_KERNEL(mul_mv_f32_f32);
     GGML_METAL_DECL_KERNEL(mul_mv_f16_f16);
@@ -148,11 +145,8 @@ struct ggml_metal_context {
     GGML_METAL_DECL_KERNEL(rope_f16);
     GGML_METAL_DECL_KERNEL(alibi_f32);
     GGML_METAL_DECL_KERNEL(im2col_f16);
-    GGML_METAL_DECL_KERNEL(upscale_f32);
-    GGML_METAL_DECL_KERNEL(pad_f32);
     GGML_METAL_DECL_KERNEL(argsort_f32_i32_asc);
     GGML_METAL_DECL_KERNEL(argsort_f32_i32_desc);
-    GGML_METAL_DECL_KERNEL(leaky_relu_f32);
     GGML_METAL_DECL_KERNEL(cpy_f32_f16);
     GGML_METAL_DECL_KERNEL(cpy_f32_f32);
     GGML_METAL_DECL_KERNEL(cpy_f32_q8_0);
@@ -180,15 +174,7 @@ struct ggml_metal_context {
 @implementation GGMLMetalClass
 @end
 
-
-static void ggml_metal_default_log_callback(enum ggml_log_level level, const char * msg, void * user_data) {
-    fprintf(stderr, "%s", msg);
-
-    UNUSED(level);
-    UNUSED(user_data);
-}
-
-ggml_log_callback ggml_metal_log_callback = ggml_metal_default_log_callback;
+ggml_log_callback ggml_metal_log_callback = NULL;
 void * ggml_metal_log_user_data = NULL;
 
 void ggml_metal_log_set_callback(ggml_log_callback log_callback, void * user_data) {
@@ -348,11 +334,9 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
         GGML_METAL_ADD_KERNEL(div_row);
         GGML_METAL_ADD_KERNEL(scale);
         GGML_METAL_ADD_KERNEL(scale_4);
-        GGML_METAL_ADD_KERNEL(tanh);
+        GGML_METAL_ADD_KERNEL(silu);
         GGML_METAL_ADD_KERNEL(relu);
         GGML_METAL_ADD_KERNEL(gelu);
-        GGML_METAL_ADD_KERNEL(gelu_quick);
-        GGML_METAL_ADD_KERNEL(silu);
         GGML_METAL_ADD_KERNEL(soft_max);
         GGML_METAL_ADD_KERNEL(soft_max_4);
         GGML_METAL_ADD_KERNEL(diag_mask_inf);
@@ -370,7 +354,6 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
         GGML_METAL_ADD_KERNEL(get_rows_q5_K);
         GGML_METAL_ADD_KERNEL(get_rows_q6_K);
         GGML_METAL_ADD_KERNEL(rms_norm);
-        GGML_METAL_ADD_KERNEL(group_norm);
         GGML_METAL_ADD_KERNEL(norm);
         GGML_METAL_ADD_KERNEL(mul_mv_f32_f32);
         GGML_METAL_ADD_KERNEL(mul_mv_f16_f16);
@@ -432,11 +415,8 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
         GGML_METAL_ADD_KERNEL(rope_f16);
         GGML_METAL_ADD_KERNEL(alibi_f32);
         GGML_METAL_ADD_KERNEL(im2col_f16);
-        GGML_METAL_ADD_KERNEL(upscale_f32);
-        GGML_METAL_ADD_KERNEL(pad_f32);
         GGML_METAL_ADD_KERNEL(argsort_f32_i32_asc);
         GGML_METAL_ADD_KERNEL(argsort_f32_i32_desc);
-        GGML_METAL_ADD_KERNEL(leaky_relu_f32);
         GGML_METAL_ADD_KERNEL(cpy_f32_f16);
         GGML_METAL_ADD_KERNEL(cpy_f32_f32);
         GGML_METAL_ADD_KERNEL(cpy_f32_q8_0);
@@ -470,11 +450,9 @@ void ggml_metal_free(struct ggml_metal_context * ctx) {
     GGML_METAL_DEL_KERNEL(div_row);
     GGML_METAL_DEL_KERNEL(scale);
     GGML_METAL_DEL_KERNEL(scale_4);
-    GGML_METAL_DEL_KERNEL(tanh);
+    GGML_METAL_DEL_KERNEL(silu);
     GGML_METAL_DEL_KERNEL(relu);
     GGML_METAL_DEL_KERNEL(gelu);
-    GGML_METAL_DEL_KERNEL(gelu_quick);
-    GGML_METAL_DEL_KERNEL(silu);
     GGML_METAL_DEL_KERNEL(soft_max);
     GGML_METAL_DEL_KERNEL(soft_max_4);
     GGML_METAL_DEL_KERNEL(diag_mask_inf);
@@ -492,7 +470,6 @@ void ggml_metal_free(struct ggml_metal_context * ctx) {
     GGML_METAL_DEL_KERNEL(get_rows_q5_K);
     GGML_METAL_DEL_KERNEL(get_rows_q6_K);
     GGML_METAL_DEL_KERNEL(rms_norm);
-    GGML_METAL_DEL_KERNEL(group_norm);
     GGML_METAL_DEL_KERNEL(norm);
     GGML_METAL_DEL_KERNEL(mul_mv_f32_f32);
     GGML_METAL_DEL_KERNEL(mul_mv_f16_f16);
@@ -554,11 +531,8 @@ void ggml_metal_free(struct ggml_metal_context * ctx) {
     GGML_METAL_DEL_KERNEL(rope_f16);
     GGML_METAL_DEL_KERNEL(alibi_f32);
     GGML_METAL_DEL_KERNEL(im2col_f16);
-    GGML_METAL_DEL_KERNEL(upscale_f32);
-    GGML_METAL_DEL_KERNEL(pad_f32);
     GGML_METAL_DEL_KERNEL(argsort_f32_i32_asc);
     GGML_METAL_DEL_KERNEL(argsort_f32_i32_desc);
-    GGML_METAL_DEL_KERNEL(leaky_relu_f32);
     GGML_METAL_DEL_KERNEL(cpy_f32_f16);
     GGML_METAL_DEL_KERNEL(cpy_f32_f32);
     GGML_METAL_DEL_KERNEL(cpy_f32_q8_0);
@@ -615,24 +589,12 @@ int * ggml_metal_get_concur_list(struct ggml_metal_context * ctx) {
 }
 
 // temporarily defined here for compatibility between ggml-backend and the old API
-
-struct ggml_backend_metal_buffer {
-    void   * data;
-    size_t   size;
+struct ggml_backend_metal_buffer_context {
+    void * data;
 
     id<MTLBuffer> metal;
 };
 
-struct ggml_backend_metal_buffer_context {
-    void * all_data;
-    size_t all_size;
-    bool owned;
-
-    // multiple buffers are used only to avoid the maximum buffer size limitation when using mmap
-    int n_buffers;
-    struct ggml_backend_metal_buffer buffers[GGML_METAL_MAX_BUFFERS];
-};
-
 // finds the Metal buffer that contains the tensor data on the GPU device
 // the assumption is that there is 1-to-1 mapping between the host and device memory buffers, so we can find the
 // Metal buffer based on the host memory pointer
@@ -642,29 +604,17 @@ static id<MTLBuffer> ggml_metal_get_buffer(struct ggml_metal_context * ctx, stru
 
     const int64_t tsize = ggml_nbytes(t);
 
-    ggml_backend_buffer_t buffer = t->view_src ? t->view_src->buffer : t->buffer;
-
     // compatibility with ggml-backend
-    if (buffer && buffer->buft == ggml_backend_metal_buffer_type()) {
-        struct ggml_backend_metal_buffer_context * buf_ctx = (struct ggml_backend_metal_buffer_context *) buffer->context;
+    if (t->buffer && t->buffer->buft == ggml_backend_metal_buffer_type()) {
+        struct ggml_backend_metal_buffer_context * buf_ctx = (struct ggml_backend_metal_buffer_context *) t->buffer->context;
 
-        // find the view that contains the tensor fully
-        for (int i = 0; i < buf_ctx->n_buffers; ++i) {
-            const int64_t ioffs = (int64_t) t->data - (int64_t) buf_ctx->buffers[i].data;
+        const int64_t ioffs = (int64_t) t->data - (int64_t) buf_ctx->data;
 
-            //GGML_METAL_LOG_INFO("ioffs = %10ld, tsize = %10ld, sum = %10ld, buf_ctx->buffers[%d].size = %10ld\n", ioffs, tsize, ioffs + tsize, i, buf_ctx->buffers[i].size);
-            if (ioffs >= 0 && ioffs + tsize <= (int64_t) buf_ctx->buffers[i].size) {
-                *offs = (size_t) ioffs;
+        GGML_ASSERT(ioffs >= 0 && ioffs + tsize <= (int64_t) t->buffer->size);
 
-                //GGML_METAL_LOG_INFO("%s: tensor '%16s', offs = %8ld\n", __func__, t->name, *offs);
+        *offs = (size_t) ioffs;
 
-                return buf_ctx->buffers[i].metal;
-            }
-        }
-
-        GGML_METAL_LOG_ERROR("%s: error: tensor '%s' buffer is nil\n", __func__, t->name);
-
-        return nil;
+        return buf_ctx->metal;
     }
 
     // find the view that contains the tensor fully
@@ -893,11 +843,9 @@ static bool ggml_metal_supports_op(const struct ggml_tensor * op) {
     switch (op->op) {
         case GGML_OP_UNARY:
             switch (ggml_get_unary_op(op)) {
-                case GGML_UNARY_OP_TANH:
+                case GGML_UNARY_OP_SILU:
                 case GGML_UNARY_OP_RELU:
                 case GGML_UNARY_OP_GELU:
-                case GGML_UNARY_OP_GELU_QUICK:
-                case GGML_UNARY_OP_SILU:
                     return true;
                 default:
                     return false;
@@ -905,11 +853,11 @@ static bool ggml_metal_supports_op(const struct ggml_tensor * op) {
         case GGML_OP_NONE:
         case GGML_OP_RESHAPE:
         case GGML_OP_VIEW:
-        case GGML_OP_TRANSPOSE:
         case GGML_OP_PERMUTE:
+        case GGML_OP_TRANSPOSE:
+        case GGML_OP_GET_ROWS:
         case GGML_OP_CONCAT:
         case GGML_OP_ADD:
-        case GGML_OP_ACC:
         case GGML_OP_MUL:
         case GGML_OP_DIV:
         case GGML_OP_SCALE:
@@ -917,15 +865,11 @@ static bool ggml_metal_supports_op(const struct ggml_tensor * op) {
         case GGML_OP_SUM_ROWS:
         case GGML_OP_SOFT_MAX:
         case GGML_OP_RMS_NORM:
-        case GGML_OP_GROUP_NORM:
         case GGML_OP_NORM:
         case GGML_OP_ALIBI:
         case GGML_OP_ROPE:
         case GGML_OP_IM2COL:
-        case GGML_OP_UPSCALE:
-        case GGML_OP_PAD:
         case GGML_OP_ARGSORT:
-        case GGML_OP_LEAKY_RELU:
         case GGML_OP_MUL_MAT:
         case GGML_OP_MUL_MAT_ID:
             return true;
@@ -958,9 +902,8 @@ static bool ggml_metal_supports_op(const struct ggml_tensor * op) {
                 };
             }
         case GGML_OP_DIAG_MASK_INF:
-        case GGML_OP_GET_ROWS:
             {
-                return op->ne[3] == 1;
+                return op->ne[0] % 4 == 0;
             }
         default:
             return false;
@@ -1036,10 +979,7 @@ void ggml_metal_graph_compute(
                         } break;
                 }
 
-                if (!ggml_metal_supports_op(dst)) {
-                    GGML_METAL_LOG_ERROR("%s: error: unsupported op '%s'\n", __func__, ggml_op_desc(dst));
-                    GGML_ASSERT(!"unsupported op");
-                }
+                GGML_ASSERT(ggml_metal_supports_op(dst));
 
                 const int64_t  ne00 = src0 ? src0->ne[0] : 0;
                 const int64_t  ne01 = src0 ? src0->ne[1] : 0;
@@ -1136,8 +1076,6 @@ void ggml_metal_graph_compute(
                     case GGML_OP_MUL:
                     case GGML_OP_DIV:
                         {
-                            const size_t offs = 0;
-
                             bool bcast_row = false;
 
                             int64_t nb = ne00;
@@ -1196,8 +1134,7 @@ void ggml_metal_graph_compute(
                             [encoder setBytes:&nb1  length:sizeof(nb1)  atIndex:24];
                             [encoder setBytes:&nb2  length:sizeof(nb2)  atIndex:25];
                             [encoder setBytes:&nb3  length:sizeof(nb3)  atIndex:26];
-                            [encoder setBytes:&offs length:sizeof(offs) atIndex:27];
-                            [encoder setBytes:&nb   length:sizeof(nb)   atIndex:28];
+                            [encoder setBytes:&nb   length:sizeof(nb)   atIndex:27];
 
                             if (bcast_row) {
                                 const int64_t n = ggml_nelements(dst)/4;
@@ -1209,91 +1146,11 @@ void ggml_metal_graph_compute(
                                 [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
                             }
                         } break;
-                    case GGML_OP_ACC:
-                        {
-                            GGML_ASSERT(src0t == GGML_TYPE_F32);
-                            GGML_ASSERT(src1t == GGML_TYPE_F32);
-                            GGML_ASSERT(dstt  == GGML_TYPE_F32);
-
-                            GGML_ASSERT(ggml_is_contiguous(src0));
-                            GGML_ASSERT(ggml_is_contiguous(src1));
-
-                            const size_t pnb1 = ((int32_t *) dst->op_params)[0];
-                            const size_t pnb2 = ((int32_t *) dst->op_params)[1];
-                            const size_t pnb3 = ((int32_t *) dst->op_params)[2];
-                            const size_t offs = ((int32_t *) dst->op_params)[3];
-
-                            const bool inplace = (bool) ((int32_t *) dst->op_params)[4];
-
-                            if (!inplace) {
-                                // run a separete kernel to cpy src->dst
-                                // not sure how to avoid this
-                                // TODO: make a simpler cpy_bytes kernel
-
-                                const int nth = MIN(1024, ne00);
-
-                                [encoder setComputePipelineState:ctx->pipeline_cpy_f32_f32];
-                                [encoder setBuffer:id_src0 offset:offs_src0        atIndex:0];
-                                [encoder setBuffer:id_dst  offset:offs_dst         atIndex:1];
-                                [encoder setBytes:&ne00    length:sizeof( int64_t) atIndex:2];
-                                [encoder setBytes:&ne01    length:sizeof( int64_t) atIndex:3];
-                                [encoder setBytes:&ne02    length:sizeof( int64_t) atIndex:4];
-                                [encoder setBytes:&ne03    length:sizeof( int64_t) atIndex:5];
-                                [encoder setBytes:&nb00    length:sizeof(uint64_t) atIndex:6];
-                                [encoder setBytes:&nb01    length:sizeof(uint64_t) atIndex:7];
-                                [encoder setBytes:&nb02    length:sizeof(uint64_t) atIndex:8];
-                                [encoder setBytes:&nb03    length:sizeof(uint64_t) atIndex:9];
-                                [encoder setBytes:&ne0     length:sizeof( int64_t) atIndex:10];
-                                [encoder setBytes:&ne1     length:sizeof( int64_t) atIndex:11];
-                                [encoder setBytes:&ne2     length:sizeof( int64_t) atIndex:12];
-                                [encoder setBytes:&ne3     length:sizeof( int64_t) atIndex:13];
-                                [encoder setBytes:&nb0     length:sizeof(uint64_t) atIndex:14];
-                                [encoder setBytes:&nb1     length:sizeof(uint64_t) atIndex:15];
-                                [encoder setBytes:&nb2     length:sizeof(uint64_t) atIndex:16];
-                                [encoder setBytes:&nb3     length:sizeof(uint64_t) atIndex:17];
-
-                                [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
-                            }
-
-                            [encoder setComputePipelineState:ctx->pipeline_add];
-                            [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
-                            [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
-                            [encoder setBuffer:id_dst  offset:offs_dst  atIndex:2];
-                            [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:3];
-                            [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:4];
-                            [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:5];
-                            [encoder setBytes:&ne03 length:sizeof(ne03) atIndex:6];
-                            [encoder setBytes:&nb00 length:sizeof(nb00) atIndex:7];
-                            [encoder setBytes:&pnb1 length:sizeof(pnb1) atIndex:8];
-                            [encoder setBytes:&pnb2 length:sizeof(pnb2) atIndex:9];
-                            [encoder setBytes:&pnb3 length:sizeof(pnb3) atIndex:10];
-                            [encoder setBytes:&ne10 length:sizeof(ne10) atIndex:11];
-                            [encoder setBytes:&ne11 length:sizeof(ne11) atIndex:12];
-                            [encoder setBytes:&ne12 length:sizeof(ne12) atIndex:13];
-                            [encoder setBytes:&ne13 length:sizeof(ne13) atIndex:14];
-                            [encoder setBytes:&nb10 length:sizeof(nb10) atIndex:15];
-                            [encoder setBytes:&nb11 length:sizeof(nb11) atIndex:16];
-                            [encoder setBytes:&nb12 length:sizeof(nb12) atIndex:17];
-                            [encoder setBytes:&nb13 length:sizeof(nb13) atIndex:18];
-                            [encoder setBytes:&ne0  length:sizeof(ne0)  atIndex:19];
-                            [encoder setBytes:&ne1  length:sizeof(ne1)  atIndex:20];
-                            [encoder setBytes:&ne2  length:sizeof(ne2)  atIndex:21];
-                            [encoder setBytes:&ne3  length:sizeof(ne3)  atIndex:22];
-                            [encoder setBytes:&nb0  length:sizeof(nb0)  atIndex:23];
-                            [encoder setBytes:&pnb1 length:sizeof(pnb1) atIndex:24];
-                            [encoder setBytes:&pnb2 length:sizeof(pnb2) atIndex:25];
-                            [encoder setBytes:&pnb3 length:sizeof(pnb3) atIndex:26];
-                            [encoder setBytes:&offs length:sizeof(offs) atIndex:27];
-
-                            const int nth = MIN(1024, ne0);
-
-                            [encoder dispatchThreadgroups:MTLSizeMake(ne11, ne12, ne13) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
-                        } break;
                     case GGML_OP_SCALE:
                         {
                             GGML_ASSERT(ggml_is_contiguous(src0));
 
-                            const float scale = *(const float *) dst->op_params;
+                            const float scale = *(const float *) src1->data;
 
                             int64_t n = ggml_nelements(dst);
 
@@ -1304,23 +1161,24 @@ void ggml_metal_graph_compute(
                                 [encoder setComputePipelineState:ctx->pipeline_scale];
                             }
 
-                            [encoder setBuffer:id_src0   offset:offs_src0 atIndex:0];
-                            [encoder setBuffer:id_dst    offset:offs_dst  atIndex:1];
+                            [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];
 
                             [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
                         } break;
                     case GGML_OP_UNARY:
                         switch (ggml_get_unary_op(gf->nodes[i])) {
-                            case GGML_UNARY_OP_TANH:
+                            case GGML_UNARY_OP_SILU:
                                 {
-                                    [encoder setComputePipelineState:ctx->pipeline_tanh];
+                                    [encoder setComputePipelineState:ctx->pipeline_silu];
                                     [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
                                     [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
 
                                     const int64_t n = ggml_nelements(dst);
+                                    GGML_ASSERT(n % 4 == 0);
 
-                                    [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                                    [encoder dispatchThreadgroups:MTLSizeMake(n/4, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
                                 } break;
                             case GGML_UNARY_OP_RELU:
                                 {
@@ -1341,28 +1199,6 @@ void ggml_metal_graph_compute(
                                     const int64_t n = ggml_nelements(dst);
                                     GGML_ASSERT(n % 4 == 0);
 
-                                    [encoder dispatchThreadgroups:MTLSizeMake(n/4, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
-                                } break;
-                            case GGML_UNARY_OP_GELU_QUICK:
-                                {
-                                    [encoder setComputePipelineState:ctx->pipeline_gelu_quick];
-                                    [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
-                                    [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
-
-                                    const int64_t n = ggml_nelements(dst);
-                                    GGML_ASSERT(n % 4 == 0);
-
-                                    [encoder dispatchThreadgroups:MTLSizeMake(n/4, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
-                                } break;
-                            case GGML_UNARY_OP_SILU:
-                                {
-                                    [encoder setComputePipelineState:ctx->pipeline_silu];
-                                    [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
-                                    [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
-
-                                    const int64_t n = ggml_nelements(dst);
-                                    GGML_ASSERT(n % 4 == 0);
-
                                     [encoder dispatchThreadgroups:MTLSizeMake(n/4, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
                                 } break;
                             default:
@@ -2001,38 +1837,6 @@ void ggml_metal_graph_compute(
 
                             [encoder dispatchThreadgroups:MTLSizeMake(nrows, 1, 1) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
                         } break;
-                    case GGML_OP_GROUP_NORM:
-                        {
-                            GGML_ASSERT(ne00 % 4 == 0);
-
-                            //float eps;
-                            //memcpy(&eps, dst->op_params, sizeof(float));
-
-                            const float eps = 1e-6f; // TODO: temporarily hardcoded
-
-                            const int32_t n_groups = ((int32_t *) dst->op_params)[0];
-
-                            int nth = 32; // SIMD width
-
-                            //while (nth < ne00/4 && nth < 1024) {
-                            //    nth *= 2;
-                            //}
-
-                            [encoder setComputePipelineState:ctx->pipeline_group_norm];
-                            [encoder setBuffer:id_src0  offset:offs_src0        atIndex:0];
-                            [encoder setBuffer:id_dst   offset:offs_dst         atIndex:1];
-                            [encoder setBytes:&ne00     length:sizeof( int64_t) atIndex:2];
-                            [encoder setBytes:&ne01     length:sizeof( int64_t) atIndex:3];
-                            [encoder setBytes:&ne02     length:sizeof( int64_t) atIndex:4];
-                            [encoder setBytes:&nb00     length:sizeof(uint64_t) atIndex:5];
-                            [encoder setBytes:&nb01     length:sizeof(uint64_t) atIndex:6];
-                            [encoder setBytes:&nb02     length:sizeof(uint64_t) atIndex:7];
-                            [encoder setBytes:&n_groups length:sizeof( int32_t) atIndex:8];
-                            [encoder setBytes:&eps      length:sizeof(   float) atIndex:9];
-                            [encoder setThreadgroupMemoryLength:32*sizeof(float) atIndex:0];
-
-                            [encoder dispatchThreadgroups:MTLSizeMake(n_groups, 1, 1) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
-                        } break;
                     case GGML_OP_NORM:
                         {
                             float eps;
@@ -2202,65 +2006,6 @@ void ggml_metal_graph_compute(
 
                             [encoder dispatchThreadgroups:MTLSizeMake(IC, OH, OW) threadsPerThreadgroup:MTLSizeMake(N, KH, KW)];
                         } break;
-                    case GGML_OP_UPSCALE:
-                        {
-                            GGML_ASSERT(src0->type == GGML_TYPE_F32);
-
-                            const int sf = dst->op_params[0];
-
-                            [encoder setComputePipelineState:ctx->pipeline_upscale_f32];
-                            [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
-                            [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
-                            [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 setBytes:&ne03 length:sizeof(ne03) atIndex:5];
-                            [encoder setBytes:&nb00 length:sizeof(nb00) atIndex:6];
-                            [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:7];
-                            [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:8];
-                            [encoder setBytes:&nb03 length:sizeof(nb03) atIndex:9];
-                            [encoder setBytes:&ne0  length:sizeof(ne0)  atIndex:10];
-                            [encoder setBytes:&ne1  length:sizeof(ne1)  atIndex:11];
-                            [encoder setBytes:&ne2  length:sizeof(ne2)  atIndex:12];
-                            [encoder setBytes:&ne3  length:sizeof(ne3)  atIndex:13];
-                            [encoder setBytes:&nb0  length:sizeof(nb0)  atIndex:14];
-                            [encoder setBytes:&nb1  length:sizeof(nb1)  atIndex:15];
-                            [encoder setBytes:&nb2  length:sizeof(nb2)  atIndex:16];
-                            [encoder setBytes:&nb3  length:sizeof(nb3)  atIndex:17];
-                            [encoder setBytes:&sf   length:sizeof(sf)   atIndex:18];
-
-                            const int nth = MIN(1024, ne0);
-
-                            [encoder dispatchThreadgroups:MTLSizeMake(ne1, ne2, ne3) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
-                        } break;
-                    case GGML_OP_PAD:
-                        {
-                            GGML_ASSERT(src0->type == GGML_TYPE_F32);
-
-                            [encoder setComputePipelineState:ctx->pipeline_pad_f32];
-                            [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
-                            [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
-                            [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 setBytes:&ne03 length:sizeof(ne03) atIndex:5];
-                            [encoder setBytes:&nb00 length:sizeof(nb00) atIndex:6];
-                            [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:7];
-                            [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:8];
-                            [encoder setBytes:&nb03 length:sizeof(nb03) atIndex:9];
-                            [encoder setBytes:&ne0  length:sizeof(ne0)  atIndex:10];
-                            [encoder setBytes:&ne1  length:sizeof(ne1)  atIndex:11];
-                            [encoder setBytes:&ne2  length:sizeof(ne2)  atIndex:12];
-                            [encoder setBytes:&ne3  length:sizeof(ne3)  atIndex:13];
-                            [encoder setBytes:&nb0  length:sizeof(nb0)  atIndex:14];
-                            [encoder setBytes:&nb1  length:sizeof(nb1)  atIndex:15];
-                            [encoder setBytes:&nb2  length:sizeof(nb2)  atIndex:16];
-                            [encoder setBytes:&nb3  length:sizeof(nb3)  atIndex:17];
-
-                            const int nth = MIN(1024, ne0);
-
-                            [encoder dispatchThreadgroups:MTLSizeMake(ne1, ne2, ne3) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
-                        } break;
                     case GGML_OP_ARGSORT:
                         {
                             GGML_ASSERT(src0->type == GGML_TYPE_F32);
@@ -2282,22 +2027,6 @@ void ggml_metal_graph_compute(
 
                             [encoder dispatchThreadgroups:MTLSizeMake(1, nrows, 1) threadsPerThreadgroup:MTLSizeMake(ne00, 1, 1)];
                         } break;
-                    case GGML_OP_LEAKY_RELU:
-                        {
-                            GGML_ASSERT(src0->type == GGML_TYPE_F32);
-
-                            float slope;
-                            memcpy(&slope, dst->op_params, sizeof(float));
-
-                            [encoder setComputePipelineState:ctx->pipeline_leaky_relu_f32];
-                            [encoder setBuffer:id_src0 offset:offs_src0   atIndex:0];
-                            [encoder setBuffer:id_dst  offset:offs_dst    atIndex:1];
-                            [encoder setBytes:&slope length:sizeof(slope) atIndex:2];
-
-                            const int64_t n = ggml_nelements(dst);
-
-                            [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
-                        } break;
                     case GGML_OP_DUP:
                     case GGML_OP_CPY:
                     case GGML_OP_CONT:
@@ -2393,7 +2122,6 @@ void ggml_metal_graph_compute(
 
 // backend interface
 
-// default buffer
 static id<MTLDevice> g_backend_device = nil;
 static int g_backend_device_ref_count = 0;
 
@@ -2421,31 +2149,34 @@ static void ggml_backend_metal_free_device(void) {
 static void * ggml_backend_metal_buffer_get_base(ggml_backend_buffer_t buffer) {
     struct ggml_backend_metal_buffer_context * ctx = (struct ggml_backend_metal_buffer_context *)buffer->context;
 
-    return ctx->all_data;
+    return ctx->data;
 }
 
 static void ggml_backend_metal_buffer_free_buffer(ggml_backend_buffer_t buffer) {
     struct ggml_backend_metal_buffer_context * ctx = (struct ggml_backend_metal_buffer_context *)buffer->context;
 
-    for (int i = 0; i < ctx->n_buffers; i++) {
-        [ctx->buffers[i].metal release];
-    }
+    [ctx->metal release];
     ggml_backend_metal_free_device();
 
-    if (ctx->owned) {
-        free(ctx->all_data);
-    }
-
+    free(ctx->data);
     free(ctx);
+
+    UNUSED(buffer);
 }
 
 static void ggml_backend_metal_buffer_set_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
+    GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor write out of bounds");
+    GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
+
     memcpy((char *)tensor->data + offset, data, size);
 
     UNUSED(buffer);
 }
 
 static void ggml_backend_metal_buffer_get_tensor(ggml_backend_buffer_t buffer, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) {
+    GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor read out of bounds");
+    GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
+
     memcpy(data, (const char *)tensor->data + offset, size);
 
     UNUSED(buffer);
@@ -2463,13 +2194,7 @@ static void ggml_backend_metal_buffer_cpy_tensor_to(ggml_backend_buffer_t buffer
     UNUSED(buffer);
 }
 
-static void ggml_backend_metal_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
-    struct ggml_backend_metal_buffer_context * ctx = (struct ggml_backend_metal_buffer_context *)buffer->context;
-
-    memset(ctx->all_data, value, ctx->all_size);
-}
-
-static struct ggml_backend_buffer_i ggml_backend_metal_buffer_i = {
+static struct ggml_backend_buffer_i metal_backend_buffer_i = {
     /* .free_buffer     = */ ggml_backend_metal_buffer_free_buffer,
     /* .get_base        = */ ggml_backend_metal_buffer_get_base,
     /* .init_tensor     = */ NULL,
@@ -2477,11 +2202,8 @@ static struct ggml_backend_buffer_i ggml_backend_metal_buffer_i = {
     /* .get_tensor      = */ ggml_backend_metal_buffer_get_tensor,
     /* .cpy_tensor_from = */ ggml_backend_metal_buffer_cpy_tensor_from,
     /* .cpy_tensor_to   = */ ggml_backend_metal_buffer_cpy_tensor_to,
-    /* .clear           = */ ggml_backend_metal_buffer_clear,
 };
 
-// default buffer type
-
 static ggml_backend_buffer_t ggml_backend_metal_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
     struct ggml_backend_metal_buffer_context * ctx = malloc(sizeof(struct ggml_backend_metal_buffer_context));
 
@@ -2492,46 +2214,13 @@ static ggml_backend_buffer_t ggml_backend_metal_buffer_type_alloc_buffer(ggml_ba
         size_aligned += (size_page - (size_aligned % size_page));
     }
 
-    id<MTLDevice> device = ggml_backend_metal_get_device();
-
-    ctx->all_data = ggml_metal_host_malloc(size_aligned);
-    ctx->all_size = size_aligned;
-    ctx->owned = true;
-    ctx->n_buffers = 1;
-
-    ctx->buffers[0].data = ctx->all_data;
-    ctx->buffers[0].size = size;
-    ctx->buffers[0].metal = [device newBufferWithBytesNoCopy:ctx->all_data
+    ctx->data  = ggml_metal_host_malloc(size);
+    ctx->metal = [ggml_backend_metal_get_device() newBufferWithBytesNoCopy:ctx->data
                     length:size_aligned
                     options:MTLResourceStorageModeShared
                     deallocator:nil];
 
-    if (ctx->buffers[0].metal == nil) {
-        GGML_METAL_LOG_ERROR("%s: error: failed to allocate buffer, size = %8.2f MiB\n", __func__, size_aligned / 1024.0 / 1024.0);
-        free(ctx);
-        ggml_backend_metal_free_device();
-        return NULL;
-    }
-
-    GGML_METAL_LOG_INFO("%s: allocated buffer, size = %8.2f MiB", __func__, size_aligned / 1024.0 / 1024.0);
-
-
-#if TARGET_OS_OSX
-    GGML_METAL_LOG_INFO(", (%8.2f / %8.2f)",
-            device.currentAllocatedSize / 1024.0 / 1024.0,
-            device.recommendedMaxWorkingSetSize / 1024.0 / 1024.0);
-
-    if (device.currentAllocatedSize > device.recommendedMaxWorkingSetSize) {
-        GGML_METAL_LOG_WARN("%s: warning: current allocated size is greater than the recommended max working set size\n", __func__);
-    } else {
-        GGML_METAL_LOG_INFO("\n");
-    }
-#else
-    GGML_METAL_LOG_INFO(", (%8.2f)\n", device.currentAllocatedSize / 1024.0 / 1024.0);
-#endif
-
-
-    return ggml_backend_buffer_init(buft, ggml_backend_metal_buffer_i, ctx, size);
+    return ggml_backend_buffer_init(buft, metal_backend_buffer_i, ctx, size);
 }
 
 static size_t ggml_backend_metal_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
@@ -2542,13 +2231,7 @@ static size_t ggml_backend_metal_buffer_type_get_alignment(ggml_backend_buffer_t
 static bool ggml_backend_metal_buffer_type_supports_backend(ggml_backend_buffer_type_t buft, ggml_backend_t backend) {
     return ggml_backend_is_metal(backend) || ggml_backend_is_cpu(backend);
 
-    UNUSED(buft);
-}
-
-static bool ggml_backend_metal_buffer_type_is_host(ggml_backend_buffer_type_t buft) {
-    return true;
-
-    UNUSED(buft);
+    GGML_UNUSED(buft);
 }
 
 ggml_backend_buffer_type_t ggml_backend_metal_buffer_type(void) {
@@ -2558,7 +2241,6 @@ ggml_backend_buffer_type_t ggml_backend_metal_buffer_type(void) {
             /* .get_alignment    = */ ggml_backend_metal_buffer_type_get_alignment,
             /* .get_alloc_size   = */ NULL, // defaults to ggml_nbytes
             /* .supports_backend = */ ggml_backend_metal_buffer_type_supports_backend,
-            /* .is_host          = */ ggml_backend_metal_buffer_type_is_host,
         },
         /* .context = */ NULL,
     };
@@ -2566,87 +2248,6 @@ ggml_backend_buffer_type_t ggml_backend_metal_buffer_type(void) {
     return &ggml_backend_buffer_type_metal;
 }
 
-// buffer from ptr
-
-ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void * data, size_t size, size_t max_size) {
-    struct ggml_backend_metal_buffer_context * ctx = malloc(sizeof(struct ggml_backend_metal_buffer_context));
-
-    ctx->all_data = data;
-    ctx->all_size = size;
-    ctx->owned = false;
-    ctx->n_buffers = 0;
-
-    const size_t size_page = sysconf(_SC_PAGESIZE);
-    size_t size_aligned = size;
-    if ((size_aligned % size_page) != 0) {
-        size_aligned += (size_page - (size_aligned % size_page));
-    }
-
-    id<MTLDevice> device = ggml_backend_metal_get_device();
-
-    // the buffer fits into the max buffer size allowed by the device
-    if (size_aligned <= device.maxBufferLength) {
-        ctx->buffers[ctx->n_buffers].data = data;
-        ctx->buffers[ctx->n_buffers].size = size;
-
-        ctx->buffers[ctx->n_buffers].metal = [device newBufferWithBytesNoCopy:data length:size_aligned options:MTLResourceStorageModeShared deallocator:nil];
-
-        if (ctx->buffers[ctx->n_buffers].metal == nil) {
-            GGML_METAL_LOG_ERROR("%s: error: failed to allocate buffer, size = %8.2f MiB\n", __func__, size_aligned / 1024.0 / 1024.0);
-            return false;
-        }
-
-        GGML_METAL_LOG_INFO("%s: allocated buffer, size = %8.2f MiB", __func__, size_aligned / 1024.0 / 1024.0);
-
-        ++ctx->n_buffers;
-    } else {
-        // this overlap between the views will guarantee that the tensor with the maximum size will fully fit into
-        // one of the views
-        const size_t size_ovlp = ((max_size + size_page - 1) / size_page + 1) * size_page; // round-up 2 pages just in case
-        const size_t size_step = device.maxBufferLength - size_ovlp;
-        const size_t size_view = device.maxBufferLength;
-
-        for (size_t i = 0; i < size; i += size_step) {
-            const size_t size_step_aligned = (i + size_view <= size) ? size_view : (size_aligned - i);
-
-            ctx->buffers[ctx->n_buffers].data = (void *) ((uint8_t *) data + i);
-            ctx->buffers[ctx->n_buffers].size = size_step_aligned;
-
-            ctx->buffers[ctx->n_buffers].metal = [device newBufferWithBytesNoCopy:(void *) ((uint8_t *) data + i) length:size_step_aligned options:MTLResourceStorageModeShared deallocator:nil];
-
-            if (ctx->buffers[ctx->n_buffers].metal == nil) {
-                GGML_METAL_LOG_ERROR("%s: error: failed to allocate buffer, size = %8.2f MiB\n", __func__, size_step_aligned / 1024.0 / 1024.0);
-                return false;
-            }
-
-            GGML_METAL_LOG_INFO("%s: allocated buffer, size = %8.2f MiB, offs = %12ld", __func__, size_step_aligned / 1024.0 / 1024.0, i);
-            if (i + size_step < size) {
-                GGML_METAL_LOG_INFO("\n");
-            }
-
-            ++ctx->n_buffers;
-        }
-    }
-
-#if TARGET_OS_OSX
-    GGML_METAL_LOG_INFO(", (%8.2f / %8.2f)",
-            device.currentAllocatedSize / 1024.0 / 1024.0,
-            device.recommendedMaxWorkingSetSize / 1024.0 / 1024.0);
-
-    if (device.currentAllocatedSize > device.recommendedMaxWorkingSetSize) {
-        GGML_METAL_LOG_WARN("%s: warning: current allocated size is greater than the recommended max working set size\n", __func__);
-    } else {
-        GGML_METAL_LOG_INFO("\n");
-    }
-#else
-    GGML_METAL_LOG_INFO(", (%8.2f)\n", device.currentAllocatedSize / 1024.0 / 1024.0);
-#endif
-
-    return ggml_backend_buffer_init(ggml_backend_metal_buffer_type(), ggml_backend_metal_buffer_i, ctx, size);
-}
-
-// backend
-
 static const char * ggml_backend_metal_name(ggml_backend_t backend) {
     return "Metal";
 
@@ -2659,6 +2260,10 @@ static void ggml_backend_metal_free(ggml_backend_t backend) {
     free(backend);
 }
 
+static void ggml_backend_metal_synchronize(ggml_backend_t backend) {
+    UNUSED(backend);
+}
+
 static ggml_backend_buffer_type_t ggml_backend_metal_get_default_buffer_type(ggml_backend_t backend) {
     return ggml_backend_metal_buffer_type();
 
@@ -2685,15 +2290,25 @@ static struct ggml_backend_i metal_backend_i = {
     /* .get_tensor_async        = */ NULL,
     /* .cpy_tensor_from_async   = */ NULL,
     /* .cpy_tensor_to_async     = */ NULL,
-    /* .synchronize             = */ NULL,
-    /* .graph_plan_create       = */ NULL,
+    /* .synchronize             = */ ggml_backend_metal_synchronize,
+    /* .graph_plan_create       = */ NULL, // the metal implementation does not require creating graph plans atm
     /* .graph_plan_free         = */ NULL,
     /* .graph_plan_compute      = */ NULL,
     /* .graph_compute           = */ ggml_backend_metal_graph_compute,
     /* .supports_op             = */ ggml_backend_metal_supports_op,
 };
 
+// TODO: make a common log callback for all backends in ggml-backend
+static void ggml_backend_log_callback(enum ggml_log_level level, const char * msg, void * user_data) {
+    fprintf(stderr, "%s", msg);
+
+    UNUSED(level);
+    UNUSED(user_data);
+}
+
 ggml_backend_t ggml_backend_metal_init(void) {
+    ggml_metal_log_set_callback(ggml_backend_log_callback, NULL);
+
     struct ggml_metal_context * ctx = ggml_metal_init(GGML_DEFAULT_N_THREADS);
 
     if (ctx == NULL) {

ggml-metal.metal

@@ -79,7 +79,6 @@ kernel void kernel_add(
         constant  int64_t & nb1,
         constant  int64_t & nb2,
         constant  int64_t & nb3,
-        constant  int64_t & offs,
         uint3 tgpig[[threadgroup_position_in_grid]],
         uint3 tpitg[[thread_position_in_threadgroup]],
         uint3   ntg[[threads_per_threadgroup]]) {
@@ -91,9 +90,9 @@ kernel void kernel_add(
     const int64_t i12 = i02 % ne12;
     const int64_t i11 = i01 % ne11;
 
-    device const char * src0_ptr = src0 + i03*nb03 + i02*nb02 + i01*nb01 + offs;
+    device const char * src0_ptr = src0 + i03*nb03 + i02*nb02 + i01*nb01;
     device const char * src1_ptr = src1 + i13*nb13 + i12*nb12 + i11*nb11;
-    device       char * dst_ptr  = dst  + i03*nb3  + i02*nb2  + i01*nb1  + offs;
+    device       char * dst_ptr  = dst  + i03*nb3  + i02*nb2  + i01*nb1;
 
     for (int i0 = tpitg.x; i0 < ne0; i0 += ntg.x) {
         const int i10 = i0 % ne10;
@@ -205,7 +204,7 @@ kernel void kernel_add_row(
         device const float4 * src0,
         device const float4 * src1,
         device       float4 * dst,
-        constant    int64_t & nb [[buffer(28)]],
+        constant    int64_t & nb [[buffer(27)]],
         uint tpig[[thread_position_in_grid]]) {
     dst[tpig] = src0[tpig] + src1[tpig % nb];
 }
@@ -214,7 +213,7 @@ kernel void kernel_mul_row(
         device const float4 * src0,
         device const float4 * src1,
         device       float4 * dst,
-        constant    int64_t & nb  [[buffer(28)]],
+        constant    int64_t & nb  [[buffer(27)]],
         uint tpig[[thread_position_in_grid]]) {
     dst[tpig] = src0[tpig] * src1[tpig % nb];
 }
@@ -223,7 +222,7 @@ kernel void kernel_div_row(
         device const float4 * src0,
         device const float4 * src1,
         device       float4 * dst,
-        constant    int64_t & nb  [[buffer(28)]],
+        constant    int64_t & nb  [[buffer(27)]],
         uint tpig[[thread_position_in_grid]]) {
     dst[tpig] = src0[tpig] / src1[tpig % nb];
 }
@@ -244,47 +243,6 @@ kernel void kernel_scale_4(
     dst[tpig] = src0[tpig] * scale;
 }
 
-kernel void kernel_relu(
-        device const float * src0,
-        device       float * dst,
-        uint tpig[[thread_position_in_grid]]) {
-    dst[tpig] = max(0.0f, src0[tpig]);
-}
-
-kernel void kernel_tanh(
-        device const float * src0,
-        device       float * dst,
-        uint tpig[[thread_position_in_grid]]) {
-    device const float & x = src0[tpig];
-    dst[tpig] = precise::tanh(x);
-}
-
-constant float GELU_COEF_A     = 0.044715f;
-constant float GELU_QUICK_COEF = -1.702f;
-constant float SQRT_2_OVER_PI  = 0.79788456080286535587989211986876f;
-
-kernel void kernel_gelu(
-    device const float4 * src0,
-    device       float4 * dst,
-    uint tpig[[thread_position_in_grid]]) {
-    device const float4 & x = src0[tpig];
-
-    // BEWARE !!!
-    // Simply using "tanh" instead of "precise::tanh" will sometimes results in NaNs!
-    // This was observed with Falcon 7B and 40B models
-    //
-    dst[tpig] = 0.5f*x*(1.0f + precise::tanh(SQRT_2_OVER_PI*x*(1.0f + GELU_COEF_A*x*x)));
-}
-
-kernel void kernel_gelu_quick(
-    device const float4 * src0,
-    device       float4 * dst,
-    uint tpig[[thread_position_in_grid]]) {
-    device const float4 & x = src0[tpig];
-
-    dst[tpig] = x*(1.0f/(1.0f+exp(GELU_QUICK_COEF*x)));
-}
-
 kernel void kernel_silu(
         device const float4 * src0,
         device       float4 * dst,
@@ -293,6 +251,13 @@ kernel void kernel_silu(
     dst[tpig] = x / (1.0f + exp(-x));
 }
 
+kernel void kernel_relu(
+        device const float * src0,
+        device       float * dst,
+        uint tpig[[thread_position_in_grid]]) {
+    dst[tpig] = max(0.0f, src0[tpig]);
+}
+
 kernel void kernel_sqr(
         device const float * src0,
         device       float * dst,
@@ -348,6 +313,22 @@ kernel void kernel_sum_rows(
     dst_row[0] = row_sum;
 }
 
+constant float GELU_COEF_A    = 0.044715f;
+constant float SQRT_2_OVER_PI = 0.79788456080286535587989211986876f;
+
+kernel void kernel_gelu(
+    device const float4 * src0,
+    device       float4 * dst,
+    uint tpig[[thread_position_in_grid]]) {
+    device const float4 & x = src0[tpig];
+
+    // BEWARE !!!
+    // Simply using "tanh" instead of "precise::tanh" will sometimes results in NaNs!
+    // This was observed with Falcon 7B and 40B models
+    //
+    dst[tpig] = 0.5f*x*(1.0f + precise::tanh(SQRT_2_OVER_PI*x*(1.0f + GELU_COEF_A*x*x)));
+}
+
 kernel void kernel_soft_max(
         device const float * src0,
         device const float * src1,
@@ -669,94 +650,6 @@ kernel void kernel_rms_norm(
     }
 }
 
-kernel void kernel_group_norm(
-        device const float * src0,
-        device       float * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant   int64_t & ne02,
-        constant  uint64_t & nb00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
-        constant   int32_t & n_groups,
-        constant     float & eps,
-        threadgroup float  * buf [[threadgroup(0)]],
-        uint tgpig[[threadgroup_position_in_grid]],
-        uint tpitg[[thread_position_in_threadgroup]],
-        uint sgitg[[simdgroup_index_in_threadgroup]],
-        uint tiisg[[thread_index_in_simdgroup]],
-        uint   ntg[[threads_per_threadgroup]]) {
-    const int64_t ne = ne00*ne01*ne02;
-    const int64_t gs = ne00*ne01*((ne02 + n_groups - 1) / n_groups);
-
-    int start = tgpig * gs;
-    int end   = start + gs;
-
-    start += tpitg;
-
-    if (end >= ne) {
-        end = ne;
-    }
-
-    float tmp = 0.0f; // partial sum for thread in warp
-
-    for (int j = start; j < end; j += ntg) {
-        tmp += src0[j];
-    }
-
-    threadgroup_barrier(mem_flags::mem_threadgroup);
-    tmp = simd_sum(tmp);
-    if (ntg > N_SIMDWIDTH) {
-        if (sgitg == 0) {
-            buf[tiisg] = 0.0f;
-        }
-
-        threadgroup_barrier(mem_flags::mem_threadgroup);
-
-        if (tiisg == 0) {
-            buf[sgitg] = tmp;
-        }
-
-        threadgroup_barrier(mem_flags::mem_threadgroup);
-
-        tmp = buf[tiisg];
-        tmp = simd_sum(tmp);
-    }
-
-    const float mean = tmp / gs;
-    tmp = 0.0f;
-
-    for (int j = start; j < end; j += ntg) {
-        float xi = src0[j] - mean;
-        dst[j] = xi;
-        tmp += xi * xi;
-    }
-
-    tmp = simd_sum(tmp);
-    if (ntg > N_SIMDWIDTH) {
-        if (sgitg == 0) {
-            buf[tiisg] = 0.0f;
-        }
-
-        threadgroup_barrier(mem_flags::mem_threadgroup);
-
-        if (tiisg == 0) {
-            buf[sgitg] = tmp;
-        }
-
-        threadgroup_barrier(mem_flags::mem_threadgroup);
-
-        tmp = buf[tiisg];
-        tmp = simd_sum(tmp);
-    }
-
-    const float variance = tmp / gs;
-    const float scale = 1.0f/sqrt(variance + eps);
-    for (int j = start; j < end; j += ntg) {
-        dst[j] *= scale;
-    }
-}
-
 // function for calculate inner product between half a q4_0 block and 16 floats (yl), sumy is SUM(yl[i])
 // il indicates where the q4 quants begin (0 or QK4_0/4)
 // we assume that the yl's have been multiplied with the appropriate scale factor
@@ -1702,9 +1595,8 @@ kernel void kernel_rope(
             dst_data[1] = x0*sin_theta + x1*cos_theta;
         }
     } else {
-        for (int64_t ic = 2*tiitg; ic < ne0; ic += 2*tptg.x) {
-            if (ic < n_dims) {
-                const int64_t ib = 0;
+        for (int64_t ib = 0; ib < ne0/n_dims; ++ib) {
+            for (int64_t ic = 2*tiitg; ic < n_dims; ic += 2*tptg.x) {
 
                 // simplified from `(ib * n_dims + ic) * inv_ndims`
                 const float cur_rot = inv_ndims*ic - ib;
@@ -1723,14 +1615,6 @@ kernel void kernel_rope(
 
                 dst_data[0]        = x0*cos_theta - x1*sin_theta;
                 dst_data[n_dims/2] = x0*sin_theta + x1*cos_theta;
-            } else {
-                const int64_t i0 = ic;
-
-                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);
-
-                dst_data[0] = src[0];
-                dst_data[1] = src[1];
             }
         }
     }
@@ -1772,97 +1656,6 @@ kernel void kernel_im2col_f16(
     }
 }
 
-kernel void kernel_upscale_f32(
-    device  const char * src0,
-    device        char * dst,
-    constant   int64_t & ne00,
-    constant   int64_t & ne01,
-    constant   int64_t & ne02,
-    constant   int64_t & ne03,
-    constant  uint64_t & nb00,
-    constant  uint64_t & nb01,
-    constant  uint64_t & nb02,
-    constant  uint64_t & nb03,
-    constant   int64_t & ne0,
-    constant   int64_t & ne1,
-    constant   int64_t & ne2,
-    constant   int64_t & ne3,
-    constant  uint64_t & nb0,
-    constant  uint64_t & nb1,
-    constant  uint64_t & nb2,
-    constant  uint64_t & nb3,
-    constant   int32_t & sf,
-    uint3 tgpig[[threadgroup_position_in_grid]],
-    uint3 tpitg[[thread_position_in_threadgroup]],
-    uint3   ntg[[threads_per_threadgroup]]) {
-
-    const int64_t i3 = tgpig.z;
-    const int64_t i2 = tgpig.y;
-    const int64_t i1 = tgpig.x;
-
-    const int64_t i03 = i3;
-    const int64_t i02 = i2;
-    const int64_t i01 = i1/sf;
-
-    device const float * src0_ptr = (device const float *) (src0 + i03*nb03 + i02*nb02 + i01*nb01);
-    device       float * dst_ptr  = (device       float *) (dst  +  i3*nb3  +  i2*nb2  +  i1*nb1);
-
-    for (int i0 = tpitg.x; i0 < ne0; i0 += ntg.x) {
-        dst_ptr[i0] = src0_ptr[i0/sf];
-    }
-}
-
-kernel void kernel_pad_f32(
-    device  const char * src0,
-    device        char * dst,
-    constant   int64_t & ne00,
-    constant   int64_t & ne01,
-    constant   int64_t & ne02,
-    constant   int64_t & ne03,
-    constant  uint64_t & nb00,
-    constant  uint64_t & nb01,
-    constant  uint64_t & nb02,
-    constant  uint64_t & nb03,
-    constant   int64_t & ne0,
-    constant   int64_t & ne1,
-    constant   int64_t & ne2,
-    constant   int64_t & ne3,
-    constant  uint64_t & nb0,
-    constant  uint64_t & nb1,
-    constant  uint64_t & nb2,
-    constant  uint64_t & nb3,
-    uint3 tgpig[[threadgroup_position_in_grid]],
-    uint3 tpitg[[thread_position_in_threadgroup]],
-    uint3   ntg[[threads_per_threadgroup]]) {
-
-    const int64_t i3 = tgpig.z;
-    const int64_t i2 = tgpig.y;
-    const int64_t i1 = tgpig.x;
-
-    const int64_t i03 = i3;
-    const int64_t i02 = i2;
-    const int64_t i01 = i1;
-
-    device const float * src0_ptr = (device const float *) (src0 + i03*nb03 + i02*nb02 + i01*nb01);
-    device       float * dst_ptr  = (device       float *) (dst  +  i3*nb3  +  i2*nb2  +  i1*nb1);
-
-    if (i1 < ne01 && i2 < ne02 && i3 < ne03) {
-        for (int i0 = tpitg.x; i0 < ne0; i0 += ntg.x) {
-            if (i0 < ne00) {
-                dst_ptr[i0] = src0_ptr[i0];
-            } else {
-                dst_ptr[i0] = 0.0f;
-            }
-        }
-
-        return;
-    }
-
-    for (int i0 = tpitg.x; i0 < ne0; i0 += ntg.x) {
-        dst_ptr[i0] = 0.0f;
-    }
-}
-
 // bitonic sort implementation following the CUDA kernels as reference
 typedef void (argsort_t)(
         device const float * x,
@@ -1915,14 +1708,6 @@ kernel void kernel_argsort_f32_i32(
 template [[host_name("kernel_argsort_f32_i32_asc")]]  kernel argsort_t kernel_argsort_f32_i32<GGML_SORT_ASC>;
 template [[host_name("kernel_argsort_f32_i32_desc")]] kernel argsort_t kernel_argsort_f32_i32<GGML_SORT_DESC>;
 
-kernel void kernel_leaky_relu_f32(
-        device const float * src0,
-        device       float * dst,
-        constant     float & slope,
-        uint tpig[[thread_position_in_grid]]) {
-    dst[tpig] = src0[tpig] > 0.0f ? src0[tpig] : src0[tpig] * slope;
-}
-
 kernel void kernel_cpy_f16_f16(
         device  const half * src0,
         device        half * dst,
@@ -2281,9 +2066,9 @@ kernel void kernel_cpy_f32_q4_1(
 }
 
 kernel void kernel_concat(
-    device  const char * src0,
-    device  const char * src1,
-    device        char * dst,
+    device const char * src0,
+    device const char * src1,
+    device       char * dst,
     constant   int64_t & ne00,
     constant   int64_t & ne01,
     constant   int64_t & ne02,
@@ -2320,7 +2105,7 @@ kernel void kernel_concat(
     const int64_t i12 = i02 % ne12;
     const int64_t i11 = i01 % ne11;
 
-    device const char * src0_ptr = src0 + i03*nb03 + i02*nb02 + i01*nb01 + tpitg.x*nb00;
+    device const char * src0_ptr = src0 + i03 * nb03 + i02 * nb02 + i01 * nb01 + tpitg.x*nb00;
     device const char * src1_ptr = src1 + i13*nb13 + i12*nb12 + i11*nb11 + tpitg.x*nb10;
     device       char * dst_ptr  = dst  + i03*nb3  + i02*nb2  + i01*nb1  + tpitg.x*nb0;
 
@@ -3530,10 +3315,10 @@ void dequantize_q8_0(device const block_q8_0 *xb, short il, thread type4x4 & reg
 
 template <typename type4x4>
 void dequantize_q2_K(device const block_q2_K *xb, short il, thread type4x4 & reg) {
-    const float d = xb->d;
-    const float min = xb->dmin;
+    const half d = xb->d;
+    const half min = xb->dmin;
     device const uint8_t * q = (device const uint8_t *)xb->qs;
-    float dl, ml;
+    half dl, ml;
     uint8_t sc = xb->scales[il];
 
 #if QK_K == 256
@@ -3603,10 +3388,10 @@ void dequantize_q4_K(device const block_q4_K *xb, short il, thread type4x4 & reg
     q = q + (il/4) * 32 + 16 * (il&1);
     il = il & 3;
     const uchar2 sc = get_scale_min_k4_just2(is, il/2, xb->scales);
-    const float d   = il < 2 ? xb->d : xb->d / 16.h;
-    const float min = xb->dmin;
-    const float dl = d * sc[0];
-    const float ml = min * sc[1];
+    const half d   = il < 2 ? xb->d : xb->d / 16.h;
+    const half min = xb->dmin;
+    const half dl = d * sc[0];
+    const half ml = min * sc[1];
 #else
     q = q + 16 * (il&1);
     device const uint8_t * s = xb->scales;
@@ -3633,13 +3418,13 @@ void dequantize_q5_K(device const block_q5_K *xb, short il, thread type4x4 & reg
     uint8_t ul = 1 << (il/2);
     il = il & 3;
     const uchar2 sc = get_scale_min_k4_just2(is, il/2, xb->scales);
-    const float d = il < 2 ? xb->d : xb->d / 16.h;
-    const float min = xb->dmin;
-    const float dl = d * sc[0];
-    const float ml = min * sc[1];
+    const half d = il < 2 ? xb->d : xb->d / 16.h;
+    const half min = xb->dmin;
+    const half dl = d * sc[0];
+    const half ml = min * sc[1];
 
-    const ushort mask  = il<2 ? 0x0F : 0xF0;
-    const float qh_val = il<2 ? 16.f : 256.f;
+    const ushort mask = il<2 ? 0x0F : 0xF0;
+    const half qh_val = il<2 ? 16.h : 256.h;
     for (int i = 0; i < 16; ++i) {
         reg[i/4][i%4] = dl * ((q[i] & mask) + (qh[i] & ul ? qh_val : 0)) - ml;
     }

ggml-quants.c

@@ -3114,7 +3114,7 @@ void ggml_vec_dot_q5_0_q8_0(const int n, float * restrict s, const void * restri
 
     size_t vl = __riscv_vsetvl_e8m1(qk/2);
 
-    // These temporary registers are for masking and shift operations
+    // These tempory registers are for masking and shift operations
     vuint32m2_t vt_1 = __riscv_vid_v_u32m2(vl);
     vuint32m2_t vt_2 = __riscv_vsll_vv_u32m2(__riscv_vmv_v_x_u32m2(1, vl), vt_1, vl);
 
@@ -4757,7 +4757,7 @@ void ggml_vec_dot_q3_K_q8_K(const int n, float * restrict s, const void * restri
 
             vl = 16;
 
-            // retrieve lane to multiply with scale
+            // retreive lane to multiply with scale
             vint32m2_t aux0_0 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(a0, 0), (scale[0]), vl);
             vint32m2_t aux0_1 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(a0, 1), (scale[1]), vl);
             vint32m2_t aux1_0 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(a1, 0), (scale[2]), vl);

ggml.h

@@ -215,7 +215,7 @@
 #define GGML_QNT_VERSION_FACTOR 1000 // do not change this
 
 #define GGML_MAX_DIMS           4
-#define GGML_MAX_PARAMS         2048
+#define GGML_MAX_PARAMS         1024
 #define GGML_MAX_CONTEXTS       64
 #define GGML_MAX_SRC            10
 #define GGML_MAX_NAME           64
@@ -303,7 +303,7 @@ extern "C" {
 
 #if defined(__ARM_NEON) && defined(__CUDACC__)
     typedef half ggml_fp16_t;
-#elif defined(__ARM_NEON) && !defined(_MSC_VER)
+#elif defined(__ARM_NEON)
     typedef __fp16 ggml_fp16_t;
 #else
     typedef uint16_t ggml_fp16_t;
@@ -343,12 +343,6 @@ extern "C" {
         GGML_TYPE_COUNT,
     };
 
-    // precision
-    enum ggml_prec {
-        GGML_PREC_DEFAULT,
-        GGML_PREC_F32,
-    };
-
     enum ggml_backend_type {
         GGML_BACKEND_CPU = 0,
         GGML_BACKEND_GPU = 10,
@@ -429,9 +423,7 @@ extern "C" {
         GGML_OP_POOL_1D,
         GGML_OP_POOL_2D,
         GGML_OP_UPSCALE, // nearest interpolate
-        GGML_OP_PAD,
         GGML_OP_ARGSORT,
-        GGML_OP_LEAKY_RELU,
 
         GGML_OP_FLASH_ATTN,
         GGML_OP_FLASH_FF,
@@ -471,6 +463,7 @@ extern "C" {
         GGML_UNARY_OP_GELU,
         GGML_UNARY_OP_GELU_QUICK,
         GGML_UNARY_OP_SILU,
+        GGML_UNARY_OP_LEAKY,
 
         GGML_UNARY_OP_COUNT,
     };
@@ -484,8 +477,7 @@ extern "C" {
     enum ggml_log_level {
         GGML_LOG_LEVEL_ERROR = 2,
         GGML_LOG_LEVEL_WARN = 3,
-        GGML_LOG_LEVEL_INFO = 4,
-        GGML_LOG_LEVEL_DEBUG = 5
+        GGML_LOG_LEVEL_INFO = 4
     };
 
     // ggml object
@@ -509,6 +501,7 @@ extern "C" {
 
         struct ggml_backend_buffer * buffer;
 
+        int     n_dims;
         int64_t ne[GGML_MAX_DIMS]; // number of elements
         size_t  nb[GGML_MAX_DIMS]; // stride in bytes:
                                    // nb[0] = ggml_type_size(type)
@@ -540,7 +533,7 @@ extern "C" {
 
         void * extra; // extra things e.g. for ggml-cuda.cu
 
-        char padding[8];
+        char padding[12];
     };
 
     static const size_t GGML_TENSOR_SIZE = sizeof(struct ggml_tensor);
@@ -645,14 +638,11 @@ extern "C" {
     GGML_API int64_t ggml_nrows       (const struct ggml_tensor * tensor);
     GGML_API size_t  ggml_nbytes      (const struct ggml_tensor * tensor);
     GGML_API size_t  ggml_nbytes_pad  (const struct ggml_tensor * tensor); // same as ggml_nbytes() but padded to GGML_MEM_ALIGN
+    GGML_API size_t  ggml_nbytes_split(const struct ggml_tensor * tensor, int nrows_split);
 
-    GGML_API int    ggml_blck_size(enum ggml_type type);
-    GGML_API size_t ggml_type_size(enum ggml_type type);             // size in bytes for all elements in a block
-    GGML_API size_t ggml_row_size (enum ggml_type type, int64_t ne); // size in bytes for all elements in a row
-
-    GGML_DEPRECATED(
-    GGML_API double ggml_type_sizef(enum ggml_type type), // ggml_type_size()/ggml_blck_size() as float
-    "use ggml_row_size() instead");
+    GGML_API int     ggml_blck_size (enum ggml_type type);
+    GGML_API size_t  ggml_type_size (enum ggml_type type); // size in bytes for all elements in a block
+    GGML_API float   ggml_type_sizef(enum ggml_type type); // ggml_type_size()/ggml_blck_size() as float
 
     GGML_API const char * ggml_type_name(enum ggml_type type);
     GGML_API const char * ggml_op_name  (enum ggml_op   op);
@@ -671,11 +661,6 @@ extern "C" {
     GGML_API bool ggml_is_transposed(const struct ggml_tensor * tensor);
     GGML_API bool ggml_is_contiguous(const struct ggml_tensor * tensor);
     GGML_API bool ggml_is_permuted  (const struct ggml_tensor * tensor);
-    GGML_API bool ggml_is_scalar    (const struct ggml_tensor * tensor);
-    GGML_API bool ggml_is_vector    (const struct ggml_tensor * tensor);
-    GGML_API bool ggml_is_matrix    (const struct ggml_tensor * tensor);
-    GGML_API bool ggml_is_3d        (const struct ggml_tensor * tensor);
-    GGML_API int  ggml_n_dims       (const struct ggml_tensor * tensor); // returns 1 for scalars
 
     GGML_API bool ggml_are_same_shape(const struct ggml_tensor * t0, const struct ggml_tensor * t1);
 
@@ -736,8 +721,8 @@ extern "C" {
     GGML_API struct ggml_tensor * ggml_view_tensor(struct ggml_context * ctx, struct ggml_tensor * src);
 
     // Context tensor enumeration and lookup
-    GGML_API struct ggml_tensor * ggml_get_first_tensor(const struct ggml_context * ctx);
-    GGML_API struct ggml_tensor * ggml_get_next_tensor (const struct ggml_context * ctx, struct ggml_tensor * tensor);
+    GGML_API struct ggml_tensor * ggml_get_first_tensor(struct ggml_context * ctx);
+    GGML_API struct ggml_tensor * ggml_get_next_tensor (struct ggml_context * ctx, struct ggml_tensor * tensor);
     GGML_API struct ggml_tensor * ggml_get_tensor(struct ggml_context * ctx, const char * name);
 
     GGML_API struct ggml_tensor * ggml_set_zero(struct ggml_tensor * tensor);
@@ -808,9 +793,6 @@ extern "C" {
             struct ggml_tensor  * a,
             struct ggml_tensor  * b);
 
-    // dst = a
-    // view(dst, nb1, nb2, nb3, offset) += b
-    // return dst
     GGML_API struct ggml_tensor * ggml_acc(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
@@ -975,14 +957,15 @@ extern "C" {
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
 
-    GGML_API struct ggml_tensor * ggml_leaky_relu(
+    GGML_API struct ggml_tensor * ggml_leaky(
             struct ggml_context * ctx,
-            struct ggml_tensor  * a, float negative_slope, bool inplace);
+            struct ggml_tensor  * a);
 
     GGML_API struct ggml_tensor * ggml_relu_inplace(
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
 
+    // TODO: double-check this computation is correct
     GGML_API struct ggml_tensor * ggml_gelu(
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
@@ -1064,12 +1047,6 @@ extern "C" {
             struct ggml_tensor  * a,
             struct ggml_tensor  * b);
 
-    // change the precision of a matrix multiplication
-    // set to GGML_PREC_F32 for higher precision (useful for phi-2)
-    GGML_API void ggml_mul_mat_set_prec(
-            struct ggml_tensor * a,
-            enum ggml_prec       prec);
-
     // indirect matrix multiplication
     //  ggml_mul_mat_id(ctx, as, ids, id, b) ~= ggml_mul_mat(as[ids[id]], b)
     GGML_API struct ggml_tensor * ggml_mul_mat_id(
@@ -1095,13 +1072,13 @@ extern "C" {
     GGML_API struct ggml_tensor * ggml_scale(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
-            float                 s);
+            struct ggml_tensor  * b);
 
     // in-place, returns view(a)
     GGML_API struct ggml_tensor * ggml_scale_inplace(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
-            float                 s);
+            struct ggml_tensor  * b);
 
     // b -> view(a,offset,nb1,nb2,3), return modified a
     GGML_API struct ggml_tensor * ggml_set(
@@ -1574,15 +1551,6 @@ extern "C" {
             struct ggml_tensor  * a,
             int                   scale_factor);
 
-    // pad each dimension with zeros: [x, ..., x] -> [x, ..., x, 0, ..., 0]
-    GGML_API struct ggml_tensor * ggml_pad(
-            struct ggml_context * ctx,
-            struct ggml_tensor  * a,
-            int                  p0,
-            int                  p1,
-            int                  p2,
-            int                  p3);
-
     // sort rows
     enum ggml_sort_order {
         GGML_SORT_ASC,
@@ -2136,11 +2104,10 @@ extern "C" {
     GGML_API const void * gguf_get_arr_data(const struct gguf_context * ctx, int key_id);
     GGML_API const char * gguf_get_arr_str (const struct gguf_context * ctx, int key_id, int i);
 
-    GGML_API int            gguf_get_n_tensors    (const struct gguf_context * ctx);
-    GGML_API int            gguf_find_tensor      (const struct gguf_context * ctx, const char * name);
-    GGML_API size_t         gguf_get_tensor_offset(const struct gguf_context * ctx, int i);
-    GGML_API char *         gguf_get_tensor_name  (const struct gguf_context * ctx, int i);
-    GGML_API enum ggml_type gguf_get_tensor_type  (const struct gguf_context * ctx, int i);
+    GGML_API int    gguf_get_n_tensors    (const struct gguf_context * ctx);
+    GGML_API int    gguf_find_tensor      (const struct gguf_context * ctx, const char * name);
+    GGML_API size_t gguf_get_tensor_offset(const struct gguf_context * ctx, int i);
+    GGML_API char * gguf_get_tensor_name  (const struct gguf_context * ctx, int i);
 
     // overrides existing values or adds a new one
     GGML_API void gguf_set_val_u8  (struct gguf_context * ctx, const char * key, uint8_t  val);

gguf-py/README.md

@@ -3,7 +3,7 @@
 This is a Python package for writing binary files in the [GGUF](https://github.com/ggerganov/ggml/pull/302)
 (GGML Universal File) format.
 
-See [convert-llama-hf-to-gguf.py](https://github.com/ggerganov/llama.cpp/blob/master/convert-hf-to-gguf.py)
+See [convert-llama-hf-to-gguf.py](https://github.com/ggerganov/llama.cpp/blob/master/convert-llama-hf-to-gguf.py)
 as an example for its usage.
 
 ## Installation
@@ -61,7 +61,7 @@ If you want to publish the package manually for any reason, you need to have `tw
 pip install build twine
 ```
 
-Then, follow these steps to release a new version:
+Then, folow these steps to release a new version:
 
 1. Bump the version in `pyproject.toml`.
 2. Build the package:

gguf-py/gguf/constants.py

@@ -95,7 +95,6 @@ class MODEL_ARCH(IntEnum):
     BLOOM     = auto()
     STABLELM  = auto()
     QWEN      = auto()
-    PHI2      = auto()
 
 
 class MODEL_TENSOR(IntEnum):
@@ -141,7 +140,6 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
     MODEL_ARCH.BLOOM:          "bloom",
     MODEL_ARCH.STABLELM:       "stablelm",
     MODEL_ARCH.QWEN:           "qwen",
-    MODEL_ARCH.PHI2:           "phi2",
 }
 
 TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
@@ -352,17 +350,6 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
     MODEL_ARCH.GPT2: [
         # TODO
     ],
-    MODEL_ARCH.PHI2: [
-        MODEL_TENSOR.TOKEN_EMBD,
-        MODEL_TENSOR.OUTPUT_NORM,
-        MODEL_TENSOR.OUTPUT,
-        MODEL_TENSOR.ATTN_NORM,
-        MODEL_TENSOR.ATTN_QKV,
-        MODEL_TENSOR.ATTN_OUT,
-        MODEL_TENSOR.FFN_NORM,
-        MODEL_TENSOR.FFN_DOWN,
-        MODEL_TENSOR.FFN_UP,
-    ]
     # TODO
 }
 

gguf-py/gguf/tensor_mapping.py

@@ -17,7 +17,6 @@ class TensorNameMap:
             "tok_embeddings",                            # llama-pth
             "embeddings.word_embeddings",                # bert
             "language_model.embedding.word_embeddings",  # persimmon
-            "transformer.embd.wte",                      # phi2
         ),
 
         # Token type embeddings
@@ -42,7 +41,6 @@ class TensorNameMap:
             "lm_head",                   # gpt2 mpt falcon llama-hf baichuan qwen
             "output",                    # llama-pth bloom
             "word_embeddings_for_head",  # persimmon
-            "lm_head.linear",            # phi2
         ),
 
         # Output norm
@@ -55,7 +53,6 @@ class TensorNameMap:
             "transformer.norm_f",                      # mpt
             "ln_f",                                    # refact bloom qwen
             "language_model.encoder.final_layernorm",  # persimmon
-            "lm_head.ln",                              # phi2
         ),
 
         # Rope frequencies
@@ -78,7 +75,6 @@ class TensorNameMap:
             "encoder.layer.{bid}.attention.output.LayerNorm",       # bert
             "language_model.encoder.layers.{bid}.input_layernorm",  # persimmon
             "model.layers.{bid}.ln1",                               # yi
-            "transformer.h.{bid}.ln",                               # phi2
         ),
 
         # Attention norm 2
@@ -94,7 +90,6 @@ class TensorNameMap:
             "transformer.h.{bid}.self_attention.query_key_value",                  # falcon
             "h.{bid}.self_attention.query_key_value",                              # bloom
             "language_model.encoder.layers.{bid}.self_attention.query_key_value",  # persimmon
-            "transformer.h.{bid}.mixer.Wqkv",                                      # phi2
         ),
 
         # Attention query
@@ -133,7 +128,6 @@ class TensorNameMap:
             "encoder.layer.{bid}.attention.output.dense",                # bert
             "transformer.h.{bid}.attn.out_proj",                         # gpt-j
             "language_model.encoder.layers.{bid}.self_attention.dense",  # persimmon
-            "transformer.h.{bid}.mixer.out_proj",                        # phi2
         ),
 
         # Rotary embeddings
@@ -173,7 +167,6 @@ class TensorNameMap:
             "transformer.h.{bid}.mlp.fc_in",                          # gpt-j
             "language_model.encoder.layers.{bid}.mlp.dense_h_to_4h",  # persimmon
             "transformer.h.{bid}.mlp.w1",                             # qwen
-            "transformer.h.{bid}.mlp.fc1",                            # phi2
         ),
 
         MODEL_TENSOR.FFN_UP_EXP: (
@@ -205,7 +198,6 @@ class TensorNameMap:
             "encoder.layer.{bid}.output.dense",                       # bert
             "transformer.h.{bid}.mlp.fc_out",                         # gpt-j
             "language_model.encoder.layers.{bid}.mlp.dense_4h_to_h",  # persimmon
-            "transformer.h.{bid}.mlp.fc2",                            # phi2
         ),
 
         MODEL_TENSOR.FFN_DOWN_EXP: (

gguf-py/gguf/vocab.py

@@ -84,7 +84,7 @@ class SpecialVocab:
         merges_file = path / 'merges.txt'
         if not merges_file.is_file():
             return False
-        with open(merges_file, 'r', encoding = 'utf-8') as fp:
+        with open(merges_file, 'r') as fp:
             first_line = next(fp, '').strip()
             if not first_line.startswith('#'):
                 fp.seek(0)
@@ -109,10 +109,8 @@ class SpecialVocab:
         return True
 
     def _set_special_token(self, typ: str, tid: Any) -> None:
-        if not isinstance(tid, int):
+        if not isinstance(tid, int) or tid < 0:
             return
-        if tid < 0:
-            raise ValueError(f'invalid value for special token type {typ}: {tid}')
         if self.n_vocab is None or tid < self.n_vocab:
             if typ in self.special_token_ids:
                 return

llama.h

@@ -39,7 +39,6 @@
 
 #define LLAMA_MAX_RNG_STATE (64*1024)
 
-#define LLAMA_FILE_MAGIC_GGLA 0x67676c61u // 'ggla'
 #define LLAMA_FILE_MAGIC_GGSN 0x6767736eu // 'ggsn'
 
 #define LLAMA_SESSION_MAGIC   LLAMA_FILE_MAGIC_GGSN
@@ -127,7 +126,7 @@ extern "C" {
         bool sorted;
     } llama_token_data_array;
 
-    typedef bool (*llama_progress_callback)(float progress, void *ctx);
+    typedef void (*llama_progress_callback)(float progress, void *ctx);
 
     // Input data for llama_decode
     // A llama_batch object can contain input about one or many sequences
@@ -180,9 +179,7 @@ extern "C" {
         int32_t main_gpu;     // the GPU that is used for scratch and small tensors
         const float * tensor_split; // how to split layers across multiple GPUs (size: LLAMA_MAX_DEVICES)
 
-        // Called with a progress value between 0.0 and 1.0. Pass NULL to disable.
-        // If the provided progress_callback returns true, model loading continues.
-        // If it returns false, model loading is immediately aborted.
+        // called with a progress value between 0 and 1, pass NULL to disable
         llama_progress_callback progress_callback;
 
         // context pointer passed to the progress callback
@@ -219,7 +216,7 @@ extern "C" {
 
         // Keep the booleans together to avoid misalignment during copy-by-value.
         bool mul_mat_q;   // if true, use experimental mul_mat_q kernels (DEPRECATED - always true)
-        bool logits_all;  // the llama_eval() call computes all logits, not just the last one (DEPRECATED - set llama_batch.logits instead)
+        bool logits_all;  // the llama_eval() call computes all logits, not just the last one
         bool embedding;   // embedding mode only
         bool offload_kqv; // whether to offload the KQV ops (including the KV cache) to GPU
     };
@@ -316,9 +313,7 @@ extern "C" {
 
     LLAMA_API const struct llama_model * llama_get_model(const struct llama_context * ctx);
 
-    // TODO: become more consistent with returned int types across the API
-    LLAMA_API uint32_t llama_n_ctx      (const struct llama_context * ctx);
-    LLAMA_API uint32_t llama_n_batch    (const struct llama_context * ctx);
+    LLAMA_API int llama_n_ctx      (const struct llama_context * ctx);
 
     LLAMA_API enum llama_vocab_type llama_vocab_type(const struct llama_model * model);
 

requirements.txt

@@ -1,5 +1,3 @@
 numpy==1.24.4
 sentencepiece==0.1.98
-transformers>=4.34.0
 gguf>=0.1.0
-protobuf>=4.21.0

scripts/get-flags.mk

@@ -1,38 +0,0 @@
-ifeq '' '$(findstring clang,$(shell $(GF_CC) --version))'
-	GF_CC_IS_GCC = 1
-	GF_CC_VER := $(shell { $(GF_CC) -dumpfullversion 2>/dev/null || $(GF_CC) -dumpversion; } | awk -F. '{ printf("%02d%02d%02d", $$1, $$2, $$3) }')
-else
-	GF_CC_IS_CLANG = 1
-	ifeq '' '$(findstring Apple,$(shell $(GF_CC) --version))'
-		GF_CC_IS_LLVM_CLANG = 1
-	else
-		GF_CC_IS_APPLE_CLANG = 1
-	endif
-	GF_CC_VER := \
-		$(shell $(GF_CC) --version | sed -n 's/^.* version \([0-9.]*\).*$$/\1/p' \
-		| awk -F. '{ printf("%02d%02d%02d", $$1, $$2, $$3) }')
-endif
-
-ifeq ($(GF_CC_IS_CLANG), 1)
-	# clang options
-	GF_CFLAGS   = -Wunreachable-code-break -Wunreachable-code-return
-	GF_CXXFLAGS = -Wunreachable-code-break -Wunreachable-code-return -Wmissing-prototypes -Wextra-semi
-
-	ifneq '' '$(and $(GF_CC_IS_LLVM_CLANG),$(filter 1,$(shell expr $(GF_CC_VER) \>= 030800)))'
-		GF_CFLAGS += -Wdouble-promotion
-	endif
-	ifneq '' '$(and $(GF_CC_IS_APPLE_CLANG),$(filter 1,$(shell expr $(GF_CC_VER) \>= 070300)))'
-		GF_CFLAGS += -Wdouble-promotion
-	endif
-else
-	# gcc options
-	GF_CFLAGS   = -Wdouble-promotion
-	GF_CXXFLAGS = -Wno-array-bounds
-
-	ifeq ($(shell expr $(GF_CC_VER) \>= 070100), 1)
-		GF_CXXFLAGS += -Wno-format-truncation
-	endif
-	ifeq ($(shell expr $(GF_CC_VER) \>= 080100), 1)
-		GF_CXXFLAGS += -Wextra-semi
-	endif
-endif

tests/test-backend-ops.cpp

@@ -54,7 +54,7 @@ static void init_tensor_uniform(ggml_tensor * tensor, float min = -1.0f, float m
         ggml_backend_tensor_set(tensor, data.data(), 0, size * sizeof(float));
     } else if (ggml_is_quantized(tensor->type) || tensor->type == GGML_TYPE_F16) {
         GGML_ASSERT(size % ggml_blck_size(tensor->type) == 0);
-        std::vector<uint8_t> dataq(ggml_row_size(tensor->type, size));
+        std::vector<uint8_t> dataq(ggml_type_size(tensor->type)*size/ggml_blck_size(tensor->type));
         int64_t hist[16];
         ggml_quantize_chunk(tensor->type, data.data(), dataq.data(), 0, size, hist);
         ggml_backend_tensor_set(tensor, dataq.data(), 0, dataq.size());
@@ -72,8 +72,6 @@ static std::vector<float> tensor_to_float(const ggml_tensor * t) {
 
     ggml_type_traits_t tt = ggml_internal_get_type_traits(t->type);
     size_t bs = ggml_blck_size(t->type);
-    std::vector<float> vq(ggml_blck_size(t->type));
-    bool quantized = ggml_is_quantized(t->type);
 
     // access elements by index to avoid gaps in views
     for (int64_t i3 = 0; i3 < t->ne[3]; i3++) {
@@ -87,8 +85,9 @@ static std::vector<float> tensor_to_float(const ggml_tensor * t) {
                         tv.push_back(*(float *) &buf[i]);
                     } else if (t->type == GGML_TYPE_I32) {
                         tv.push_back((float)*(int32_t *) &buf[i]);
-                    } else if (quantized) {
-                        tt.to_float(&buf[i], vq.data(), bs);
+                    } else if (ggml_is_quantized(t->type)) {
+                        std::vector<float> vq(ggml_blck_size(t->type));
+                        tt.to_float(&buf[i], vq.data(), ggml_blck_size(t->type));
                         tv.insert(tv.end(), vq.begin(), vq.end());
                     } else {
                         GGML_ASSERT(false);
@@ -235,11 +234,6 @@ static bool ggml_is_view_op(enum ggml_op op) {
     return op == GGML_OP_VIEW || op == GGML_OP_RESHAPE || op == GGML_OP_PERMUTE || op == GGML_OP_TRANSPOSE;
 }
 
-enum test_mode {
-    MODE_TEST,
-    MODE_PERF,
-};
-
 struct test_case {
     virtual ~test_case() {}
 
@@ -274,58 +268,7 @@ struct test_case {
         return size;
     }
 
-    ggml_cgraph * gf = nullptr;
-
-    static const int sentinel_size = 1024;
-
-    test_mode mode;
-
-    std::vector<ggml_tensor *> sentinels;
-
-    void add_sentinel(ggml_context * ctx) {
-        if (mode == MODE_PERF) {
-            return;
-        }
-        ggml_tensor * sentinel = ::ggml_new_tensor_1d(ctx, GGML_TYPE_F32, sentinel_size);
-        ggml_format_name(sentinel, "sent_%zu", sentinels.size());
-        sentinels.push_back(sentinel);
-    }
-
-    // hijack ggml_new_tensor to add sentinels after each tensor to check for overflows in the backend
-
-    ggml_tensor * ggml_new_tensor(ggml_context * ctx, ggml_type type, int n_dims, const int64_t * ne) {
-        ggml_tensor * t = ::ggml_new_tensor(ctx, type, n_dims, ne);
-        add_sentinel(ctx);
-        return t;
-    }
-
-    ggml_tensor * ggml_new_tensor_1d(ggml_context * ctx, ggml_type type, int64_t ne0) {
-        ggml_tensor * t = ::ggml_new_tensor_1d(ctx, type, ne0);
-        add_sentinel(ctx);
-        return t;
-    }
-
-    ggml_tensor * ggml_new_tensor_2d(ggml_context * ctx, ggml_type type, int64_t ne0, int64_t ne1) {
-        ggml_tensor * t = ::ggml_new_tensor_2d(ctx, type, ne0, ne1);
-        add_sentinel(ctx);
-        return t;
-    }
-
-    ggml_tensor * ggml_new_tensor_3d(ggml_context * ctx, ggml_type type, int64_t ne0, int64_t ne1, int64_t ne2) {
-        ggml_tensor * t = ::ggml_new_tensor_3d(ctx, type, ne0, ne1, ne2);
-        add_sentinel(ctx);
-        return t;
-    }
-
-    ggml_tensor * ggml_new_tensor_4d(ggml_context * ctx, ggml_type type, int64_t ne0, int64_t ne1, int64_t ne2, int64_t ne3) {
-        ggml_tensor * t = ::ggml_new_tensor_4d(ctx, type, ne0, ne1, ne2, ne3);
-        add_sentinel(ctx);
-        return t;
-    }
-
     bool eval(ggml_backend_t backend1, ggml_backend_t backend2, const char * op_name) {
-        mode = MODE_TEST;
-
         ggml_init_params params = {
             /* .mem_size = */ ggml_tensor_overhead()*128 + ggml_graph_overhead(),
             /* .mem_base = */ NULL,
@@ -333,11 +276,6 @@ struct test_case {
         };
         ggml_context * ctx = ggml_init(params);
 
-        gf = ggml_new_graph(ctx);
-
-        // pre-graph sentinel
-        add_sentinel(ctx);
-
         ggml_tensor * out = build_graph(ctx);
 
         if (op_name != nullptr && op_desc(out) != op_name) {
@@ -358,20 +296,13 @@ struct test_case {
             }
         }
 
-        // post-graph sentinel
-        add_sentinel(ctx);
-
         // allocate
         ggml_backend_buffer_t buf = ggml_backend_alloc_ctx_tensors(ctx, backend1);
 
         // build graph
+        ggml_cgraph * gf = ggml_new_graph(ctx);
         ggml_build_forward_expand(gf, out);
 
-        // add sentinels as graph nodes so that they are checked in the callback
-        for (ggml_tensor * sentinel : sentinels) {
-            gf->nodes[gf->n_nodes++] = sentinel;
-        }
-
         // randomize tensors
         initialize_tensors(ctx);
 
@@ -387,24 +318,9 @@ struct test_case {
         };
 
         auto callback = [](int index, ggml_tensor * t1, ggml_tensor * t2, void * user_data) -> bool {
-            callback_userdata * ud = (callback_userdata *) user_data;
-
-            if (t1->op == GGML_OP_NONE) {
-                // sentinels must be unchanged
-                std::vector<uint8_t> t1_data(ggml_nbytes(t1));
-                std::vector<uint8_t> t2_data(ggml_nbytes(t2));
-                ggml_backend_tensor_get(t1, t1_data.data(), 0, ggml_nbytes(t1));
-                ggml_backend_tensor_get(t2, t2_data.data(), 0, ggml_nbytes(t2));
-
-                if (memcmp(t1_data.data(), t2_data.data(), ggml_nbytes(t1)) != 0) {
-                    printf("sentinel mismatch: %s ", t1->name);
-                    ud->ok = false;
-                    return true;
-                }
-            }
-
             std::vector<float> f1 = tensor_to_float(t1);
             std::vector<float> f2 = tensor_to_float(t2);
+            callback_userdata * ud = (callback_userdata *) user_data;
 
             for (size_t i = 0; i < f1.size(); i++) {
                 // check for nans
@@ -433,10 +349,9 @@ struct test_case {
             if (err > ud->max_err) {
                 printf("[%s] NMSE = %f ", ggml_op_desc(t1), err);
                 //for (int i = 0; i < f1.size(); i++) {
-                //    printf("%5d %9.6f %9.6f, diff = %9.6f\n", i, f1[i], f2[i], f1[i] - f2[i]);
+                //    printf("(%f, %f) ", f1[i], f2[i]);
                 //}
                 //printf("\n");
-                //exit(1);
                 ud->ok = false;
             }
             return true;
@@ -460,8 +375,6 @@ struct test_case {
     }
 
     bool eval_perf(ggml_backend_t backend, const char * op_name) {
-        mode = MODE_PERF;
-
         static const size_t graph_nodes = 8192;
 
         ggml_init_params params = {
@@ -766,19 +679,18 @@ struct test_bin_bcast : public test_case {
 struct test_scale : public test_case {
     const ggml_type type;
     const std::array<int64_t, 4> ne;
-    float scale;
 
     std::string vars() override {
-        return VARS_TO_STR3(type, ne, scale);
+        return VARS_TO_STR2(type, ne);
     }
 
     test_scale(ggml_type type = GGML_TYPE_F32,
-            std::array<int64_t, 4> ne = {10, 10, 10, 10},
-            float scale = 2.0f)
-        : type(type), ne(ne), scale(scale) {}
+            std::array<int64_t, 4> ne = {10, 10, 10, 10})
+        : type(type), ne(ne) {}
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
         ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data());
+        ggml_tensor * scale = ggml_new_tensor_1d(ctx, type, 1);
         ggml_tensor * out = ggml_scale(ctx, a, scale);
         return out;
     }
@@ -1223,118 +1135,6 @@ struct test_sum_rows : public test_case {
     }
 };
 
-// GGML_OP_UPSCALE
-struct test_upscale : public test_case {
-    const ggml_type type;
-    const std::array<int64_t, 4> ne;
-    const int32_t scale_factor;
-
-    std::string vars() override {
-        return VARS_TO_STR3(type, ne, scale_factor);
-    }
-
-    test_upscale(ggml_type type = GGML_TYPE_F32,
-            std::array<int64_t, 4> ne = {512, 512, 3, 1},
-            int32_t scale_factor = 2)
-        : type(type), ne(ne), scale_factor(scale_factor) {}
-
-    ggml_tensor * build_graph(ggml_context * ctx) override {
-        ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data());
-        ggml_tensor * out = ggml_upscale(ctx, a, scale_factor);
-        return out;
-    }
-};
-
-// GGML_OP_GROUP_NORM
-struct test_group_norm : public test_case {
-    const ggml_type type;
-    const std::array<int64_t, 4> ne;
-    const int32_t num_groups;
-
-    std::string vars() override {
-        return VARS_TO_STR3(type, ne, num_groups);
-    }
-
-    test_group_norm(ggml_type type = GGML_TYPE_F32,
-            std::array<int64_t, 4> ne = {64, 64, 320, 1},
-            int32_t num_groups = 32)
-        : type(type), ne(ne), num_groups(num_groups) {}
-
-    ggml_tensor * build_graph(ggml_context * ctx) override {
-        ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data());
-        ggml_tensor * out = ggml_group_norm(ctx, a, num_groups);
-        return out;
-    }
-};
-
-// GGML_OP_ACC
-struct test_acc : public test_case {
-    const ggml_type type;
-    const std::array<int64_t, 4> ne_a;
-    const std::array<int64_t, 4> ne_b;
-
-    std::string vars() override {
-        return VARS_TO_STR3(type, ne_a, ne_b);
-    }
-
-    test_acc(ggml_type type = GGML_TYPE_F32,
-            std::array<int64_t, 4> ne_a = {1024, 577, 1, 1},
-            std::array<int64_t, 4> ne_b = {1024, 576, 1, 1})
-        : type(type), ne_a(ne_a), ne_b(ne_b) {}
-
-    ggml_tensor * build_graph(ggml_context * ctx) override {
-        ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne_a.data());
-        ggml_tensor * b = ggml_new_tensor(ctx, type, 4, ne_b.data());
-        ggml_tensor * out = ggml_acc(ctx, a, b, a->nb[1], a->nb[2], a->nb[3], b->nb[1]);
-        return out;
-    }
-};
-
-// GGML_OP_PAD
-struct test_pad : public test_case {
-    const ggml_type type;
-    const std::array<int64_t, 4> ne_a;
-    const int pad_0;
-    const int pad_1;
-
-    std::string vars() override {
-        return VARS_TO_STR4(type, ne_a, pad_0, pad_1);
-    }
-
-    test_pad(ggml_type type = GGML_TYPE_F32,
-            std::array<int64_t, 4> ne_a = {512, 512, 1, 1},
-            int pad_0 = 1, int pad_1 = 1)
-        : type(type), ne_a(ne_a), pad_0(pad_0), pad_1(pad_1)  {}
-
-    ggml_tensor * build_graph(ggml_context * ctx) override {
-        ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne_a.data());
-        ggml_tensor * out = ggml_pad(ctx, a, pad_0, pad_1, 0, 0);
-        return out;
-    }
-};
-
-// GGML_OP_LEAKY_RELU
-struct test_leaky_relu : public test_case {
-    const ggml_type type;
-    const std::array<int64_t, 4> ne_a;
-    const float negative_slope;
-
-    std::string vars() override {
-        return VARS_TO_STR3(type, ne_a, negative_slope);
-    }
-
-    test_leaky_relu(ggml_type type = GGML_TYPE_F32,
-            std::array<int64_t, 4> ne_a = {10, 10, 10, 10},
-            float negative_slope = 0.1f)
-        : type(type), ne_a(ne_a), negative_slope(negative_slope)  {}
-
-    ggml_tensor * build_graph(ggml_context * ctx) override {
-        ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne_a.data());
-        ggml_tensor * out = ggml_leaky_relu(ctx, a, negative_slope, true);
-        return out;
-    }
-};
-
 // Mixtral MOE
 struct test_moe : public test_case {
     const int n_experts;
@@ -1419,6 +1219,11 @@ struct test_moe : public test_case {
     }
 };
 
+enum test_mode {
+    MODE_TEST,
+    MODE_PERF,
+};
+
 static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op_name) {
     std::vector<std::unique_ptr<test_case>> test_cases;
 
@@ -1556,7 +1361,6 @@ static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op
         test_cases.emplace_back(new test_rope(type, { 64,   8, 10, 1},  64, 2, 512)); // neox (falcon 40B)
         test_cases.emplace_back(new test_rope(type, { 64, 128, 10, 1},  64, 2, 512)); // neox (falcon 40B)
         test_cases.emplace_back(new test_rope(type, { 80,  32, 10, 1},  20, 2, 512)); // neox (stablelm)
-        test_cases.emplace_back(new test_rope(type, { 80,  32, 10, 1},  32, 2, 512)); // neox (phi-2)
     }
 
     test_cases.emplace_back(new test_alibi());
@@ -1568,16 +1372,12 @@ static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op
         test_cases.emplace_back(new test_argsort(GGML_TYPE_F32, {16, 10, 10, 10}, order));
     }
 
-    test_cases.emplace_back(new test_sum_rows());
-    test_cases.emplace_back(new test_upscale());
-    test_cases.emplace_back(new test_group_norm());
-    test_cases.emplace_back(new test_acc());
-    test_cases.emplace_back(new test_pad());
-    test_cases.emplace_back(new test_leaky_relu());
+    test_cases.emplace_back(new test_sum_rows(GGML_TYPE_F32, {10, 10, 10, 10}));
+    test_cases.emplace_back(new test_sum_rows(GGML_TYPE_F32, {2, 1, 1, 1}));
 
 #if !defined(__SANITIZE_THREAD__)
     // FIXME: these tests use too much memory with thread sanitizer
-    test_cases.emplace_back(new test_moe(8, 2, 1, 4096, 8*1024));
+    test_cases.emplace_back(new test_moe(8, 2, 1, 4096, 14336));
     //test_cases.emplace_back(new test_moe(8, 2, 8, 4096, 14336));
 #endif
 

tests/test-grad0.cpp

@@ -1,4 +1,4 @@
-#define _CRT_SECURE_NO_DEPRECATE // Disables ridiculous "unsafe" warnings on Windows
+#define _CRT_SECURE_NO_DEPRECATE // Disables ridiculous "unsafe" warnigns on Windows
 #include "ggml.h"
 
 #include <cmath>
@@ -881,19 +881,19 @@ int main(int argc, const char ** argv) {
         // scale
         {
             srand(seed);
-            const int nargs = 1;
+            const int nargs = 2;
 
             int64_t ne2[4];
             ne2[0] = 1;
 
             for (int ndims = 1; ndims <= 2; ++ndims) {
+                x[1] = get_random_tensor_f32(ctx0, 1, ne2, -1.0f, 1.0f);
                 x[0] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
 
-                const float s = -1.0f + 2.0f*frand();
-
                 ggml_set_param(ctx0, x[0]);
+                ggml_set_param(ctx0, x[1]);
 
-                struct ggml_tensor * f = ggml_sum(ctx0, ggml_scale(ctx0, x[0], s));
+                struct ggml_tensor * f = ggml_sum(ctx0, ggml_scale(ctx0, x[0], x[1]));
 
                 check_gradient("scale", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, INFINITY);
             }
@@ -1395,7 +1395,7 @@ int main(int argc, const char ** argv) {
                                                 ggml_add1(ctx0,
                                                     ggml_scale(ctx0,
                                                         ggml_soft_max(ctx0, x[0]),
-                                                        1.0f - eps),
+                                                        ggml_new_f32(ctx0, 1.0f - eps)),
                                                     ggml_new_f32(ctx0, eps))));
 
                 check_gradient("softmax", ctx0, x, f, ndims, nargs, 1e-3f, 2e-1f, INFINITY);

tests/test-quantize-perf.cpp

@@ -117,7 +117,7 @@ static void usage(char * argv[]) {
     printf("  --size SIZE           set test size, divisible by 32 (L1_SIZE:%d)\n", L1_SIZE);
     printf("  -3                    use size as L1, L2, L3 sizes (L1:%d L2:%d L3:%d)\n", L1_SIZE, L2_SIZE, L3_SIZE);
     printf("  -4                    use size as L1, L2, L3, MEM sizes (L1:%d L2:%d L3:%d MEM:%d)\n", L1_SIZE, L2_SIZE, L3_SIZE, MEM_SIZE);
-    printf("  --op OP               set test operation as quantize_row_q_reference, quantize_row_q, dequantize_row_q,\n");
+    printf("  --op OP               set test opration as quantize_row_q_reference, quantize_row_q, dequantize_row_q,\n");
     printf("                        quantize_row_q_dot, vec_dot_q (all)\n");
     printf("  --type TYPE           set test type as");
     for (int i = 0; i < GGML_TYPE_COUNT; i++) {
@@ -202,7 +202,7 @@ int main(int argc, char * argv[]) {
             }
             int alignment = std::stoi(argv[i]);
             if (alignment < 0 || alignment > MAX_ALIGNMENT) {
-            fprintf(stderr, "error: alignment-offset must be less than %d\n", MAX_ALIGNMENT);
+            fprintf(stderr, "error: aligment-offset must be less than %d\n", MAX_ALIGNMENT);
                 invalid_param = true;
                 break;
             }
@@ -286,7 +286,7 @@ int main(int argc, char * argv[]) {
                         qfns.from_float_reference(test_data1, test_q1, size);
                         return test_q1[0];
                     };
-                    size_t quantized_size = ggml_row_size(type, size);
+                    size_t quantized_size = size / ggml_blck_size(type) * ggml_type_size(type);
                     benchmark_function(size, quantized_size, iterations, quantize_fn);
                 }
                 printf("\n");
@@ -300,7 +300,7 @@ int main(int argc, char * argv[]) {
                         qfns.from_float(test_data1, test_q1, size);
                         return test_q1[0];
                     };
-                    size_t quantized_size = ggml_row_size(type, size);
+                    size_t quantized_size = size / ggml_blck_size(type) * ggml_type_size(type);
                     benchmark_function(size, quantized_size, iterations, quantize_fn);
                 }
                 printf("\n");
@@ -315,7 +315,7 @@ int main(int argc, char * argv[]) {
                         qfns.to_float(test_q1, test_out, size);
                         return test_out[0];
                     };
-                    size_t quantized_size = ggml_row_size(type, size);
+                    size_t quantized_size = size / ggml_blck_size(type) * ggml_type_size(type);
                     benchmark_function(size, quantized_size, iterations, quantize_fn);
                 }
                 printf("\n");
@@ -330,7 +330,7 @@ int main(int argc, char * argv[]) {
                         vdot.from_float(test_data1, test_q1, size);
                         return test_q1[0];
                     };
-                    size_t quantized_size = ggml_row_size(type, size);
+                    size_t quantized_size = size / ggml_blck_size(type) * ggml_type_size(type);
                     benchmark_function(size, quantized_size, iterations, quantize_fn);
                 }
                 printf("\n");
@@ -347,7 +347,7 @@ int main(int argc, char * argv[]) {
                         qfns.vec_dot(size, &result, test_q1, test_q2);
                         return result;
                     };
-                    size_t quantized_size = ggml_row_size(type, size);
+                    size_t quantized_size = size / ggml_blck_size(type) * ggml_type_size(type);
                     benchmark_function(size, quantized_size, iterations, quantize_fn);
                 }
                 printf("\n");