llama : allow getting n_batch from llama_context in c api (#4540 )

* allowed getting n_batch from llama_context in c api * changed to use `uint32_t` instead of `int` * changed to use `uint32_t` instead of `int` in `llama_n_ctx` * Update llama.h --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
metal : fix ggml_metal_log vargs (#4373 )
2026-07-10 22:45:53 +02:00 · 2023-12-21 21:57:48 +02:00 · 2023-12-21 21:55:02 +02:00 · 2023-12-21 21:45:32 +02:00 · 2023-12-21 20:13:25 +01:00 · 2023-12-21 19:55:34 +02:00
38 changed files with 3866 additions and 1409 deletions
@@ -15,8 +15,14 @@ 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
@@ -91,6 +91,7 @@ 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"                                 ${LLAMA_METAL_DEFAULT})
 option(LLAMA_METAL_NDEBUG                    "llama: disable Metal debugging"                   OFF)
@@ -291,7 +292,12 @@ if (LLAMA_CUBLAS)
        add_compile_definitions(GGML_CUDA_PEER_MAX_BATCH_SIZE=${LLAMA_CUDA_PEER_MAX_BATCH_SIZE})

        if (LLAMA_STATIC)
-            set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} CUDA::cudart_static CUDA::cublas_static CUDA::cublasLt_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()
        else()
            set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} CUDA::cudart CUDA::cublas CUDA::cublasLt)
        endif()
@@ -372,6 +378,9 @@ if (LLAMA_HIPBLAS)
    if (${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)
@@ -397,57 +406,102 @@ if (LLAMA_HIPBLAS)
    endif()
 endif()

-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(host_cxx_flags "")
+function(get_flags CCID CCVER)
+    set(C_FLAGS "")
+    set(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)
+    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 (
-                (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)
-
-            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()
+        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()
-    else()
-        # todo : msvc
    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}>")
+    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(C_FLAGS   ${WARNING_FLAGS} ${C_FLAGS})
+        set(CXX_FLAGS ${WARNING_FLAGS} ${CXX_FLAGS})
+
+        get_flags(${CMAKE_CXX_COMPILER_ID} ${CMAKE_CXX_COMPILER_VERSION})
+
+        add_compile_options("$<$<COMPILE_LANGUAGE:C>:${C_FLAGS};${GF_C_FLAGS}>"
+                            "$<$<COMPILE_LANGUAGE:CXX>:${CXX_FLAGS};${GF_CXX_FLAGS}>")
+    else()
+        # todo : msvc
+        set(C_FLAGS   "")
+        set(CXX_FLAGS "")
+    endif()
 endif()

-if (NOT MSVC)
-    set(cuda_flags -Wno-pedantic)
-endif()
-set(cuda_flags ${cxx_flags} -use_fast_math ${cuda_flags})
+if (LLAMA_CUBLAS)
+    set(CUDA_FLAGS ${CXX_FLAGS} -use_fast_math)
+    if (NOT MSVC)
+        set(CUDA_FLAGS ${CUDA_FLAGS} -Wno-pedantic)
+    endif()

-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()
+    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()

-add_compile_options("$<$<COMPILE_LANGUAGE:CUDA>:${cuda_flags}>")
+        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}>")
+endif()

 if (WIN32)
    add_compile_definitions(_CRT_SECURE_NO_WARNINGS)
@@ -471,6 +525,7 @@ 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)
@@ -26,20 +26,6 @@ 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)
@@ -121,12 +107,12 @@ 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
-MK_HOST_CXXFLAGS += -Ofast
-MK_CUDA_CXXFLAGS += -O3
+MK_CFLAGS     += -Ofast
+HOST_CXXFLAGS += -Ofast
+MK_NVCCFLAGS  += -O3
 else
-MK_CFLAGS        += -O3
-MK_CXXFLAGS      += -O3
+MK_CFLAGS     += -O3
+MK_CXXFLAGS   += -O3
 endif

 # clock_gettime came in POSIX.1b (1993)
@@ -220,30 +206,6 @@ 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
@@ -294,8 +256,8 @@ 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
-	MK_HOST_CXXFLAGS += -march=native -mtune=native
+	MK_CFLAGS     += -march=native -mtune=native
+	HOST_CXXFLAGS += -march=native -mtune=native

 	# Usage AVX-only
 	#MK_CFLAGS   += -mfma -mf16c -mavx
@@ -398,10 +360,10 @@ 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
-	NVCCFLAGS = --forward-unknown-to-host-compiler -use_fast_math
+	MK_NVCCFLAGS  = --forward-unknown-to-host-compiler -use_fast_math

 ifdef LLAMA_DEBUG
-	NVCCFLAGS += -lineinfo
+	MK_NVCCFLAGS += -lineinfo
 endif

 ifdef LLAMA_CUDA_NVCC
@@ -410,54 +372,52 @@ else
 	NVCC = nvcc
 endif #LLAMA_CUDA_NVCC
 ifdef CUDA_DOCKER_ARCH
-	NVCCFLAGS += -Wno-deprecated-gpu-targets -arch=$(CUDA_DOCKER_ARCH)
-else ifdef CUDA_POWER_ARCH
-	NVCCFLAGS +=
-else
-	NVCCFLAGS += -arch=native
+	MK_NVCCFLAGS += -Wno-deprecated-gpu-targets -arch=$(CUDA_DOCKER_ARCH)
+else ifndef CUDA_POWER_ARCH
+	MK_NVCCFLAGS += -arch=native
 endif # CUDA_DOCKER_ARCH
 ifdef LLAMA_CUDA_FORCE_DMMV
-	NVCCFLAGS += -DGGML_CUDA_FORCE_DMMV
+	MK_NVCCFLAGS += -DGGML_CUDA_FORCE_DMMV
 endif # LLAMA_CUDA_FORCE_DMMV
 ifdef LLAMA_CUDA_FORCE_MMQ
-	NVCCFLAGS += -DGGML_CUDA_FORCE_MMQ
+	MK_NVCCFLAGS += -DGGML_CUDA_FORCE_MMQ
 endif # LLAMA_CUDA_FORCE_MMQ
 ifdef LLAMA_CUDA_DMMV_X
-	NVCCFLAGS += -DGGML_CUDA_DMMV_X=$(LLAMA_CUDA_DMMV_X)
+	MK_NVCCFLAGS += -DGGML_CUDA_DMMV_X=$(LLAMA_CUDA_DMMV_X)
 else
-	NVCCFLAGS += -DGGML_CUDA_DMMV_X=32
+	MK_NVCCFLAGS += -DGGML_CUDA_DMMV_X=32
 endif # LLAMA_CUDA_DMMV_X
 ifdef LLAMA_CUDA_MMV_Y
-	NVCCFLAGS += -DGGML_CUDA_MMV_Y=$(LLAMA_CUDA_MMV_Y)
+	MK_NVCCFLAGS += -DGGML_CUDA_MMV_Y=$(LLAMA_CUDA_MMV_Y)
 else ifdef LLAMA_CUDA_DMMV_Y
-	NVCCFLAGS += -DGGML_CUDA_MMV_Y=$(LLAMA_CUDA_DMMV_Y) # for backwards compatibility
+	MK_NVCCFLAGS += -DGGML_CUDA_MMV_Y=$(LLAMA_CUDA_DMMV_Y) # for backwards compatibility
 else
-	NVCCFLAGS += -DGGML_CUDA_MMV_Y=1
+	MK_NVCCFLAGS += -DGGML_CUDA_MMV_Y=1
 endif # LLAMA_CUDA_MMV_Y
 ifdef LLAMA_CUDA_F16
-	NVCCFLAGS += -DGGML_CUDA_F16
+	MK_NVCCFLAGS += -DGGML_CUDA_F16
 endif # LLAMA_CUDA_F16
 ifdef LLAMA_CUDA_DMMV_F16
-	NVCCFLAGS += -DGGML_CUDA_F16
+	MK_NVCCFLAGS += -DGGML_CUDA_F16
 endif # LLAMA_CUDA_DMMV_F16
 ifdef LLAMA_CUDA_KQUANTS_ITER
-	NVCCFLAGS += -DK_QUANTS_PER_ITERATION=$(LLAMA_CUDA_KQUANTS_ITER)
+	MK_NVCCFLAGS += -DK_QUANTS_PER_ITERATION=$(LLAMA_CUDA_KQUANTS_ITER)
 else
-	NVCCFLAGS += -DK_QUANTS_PER_ITERATION=2
+	MK_NVCCFLAGS += -DK_QUANTS_PER_ITERATION=2
 endif
 ifdef LLAMA_CUDA_PEER_MAX_BATCH_SIZE
-	NVCCFLAGS += -DGGML_CUDA_PEER_MAX_BATCH_SIZE=$(LLAMA_CUDA_PEER_MAX_BATCH_SIZE)
+	MK_NVCCFLAGS += -DGGML_CUDA_PEER_MAX_BATCH_SIZE=$(LLAMA_CUDA_PEER_MAX_BATCH_SIZE)
 else
-	NVCCFLAGS += -DGGML_CUDA_PEER_MAX_BATCH_SIZE=128
+	MK_NVCCFLAGS += -DGGML_CUDA_PEER_MAX_BATCH_SIZE=128
 endif # LLAMA_CUDA_PEER_MAX_BATCH_SIZE
 #ifdef LLAMA_CUDA_CUBLAS
-#	NVCCFLAGS += -DGGML_CUDA_CUBLAS
+#	MK_NVCCFLAGS += -DGGML_CUDA_CUBLAS
 #endif # LLAMA_CUDA_CUBLAS
 ifdef LLAMA_CUDA_CCBIN
-	NVCCFLAGS += -ccbin $(LLAMA_CUDA_CCBIN)
+	MK_NVCCFLAGS += -ccbin $(LLAMA_CUDA_CCBIN)
 endif
 ggml-cuda.o: ggml-cuda.cu ggml-cuda.h
-	$(NVCC) $(NVCCFLAGS) -c $< -o $@
+	$(NVCC) $(BASE_CXXFLAGS) $(NVCCFLAGS) -Wno-pedantic -Xcompiler "$(CUDA_CXXFLAGS)" -c $< -o $@
 endif # LLAMA_CUBLAS

 ifdef LLAMA_CLBLAST
@@ -479,9 +439,15 @@ ggml-opencl.o: ggml-opencl.cpp ggml-opencl.h
 endif # LLAMA_CLBLAST

 ifdef LLAMA_HIPBLAS
-	ROCM_PATH	?= /opt/rocm
-	HIPCC	    ?= $(ROCM_PATH)/bin/hipcc
-	GPU_TARGETS ?= $(shell $(ROCM_PATH)/llvm/bin/amdgpu-arch)
+
+	ifeq ($(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
 	LLAMA_CUDA_DMMV_X       ?= 32
 	LLAMA_CUDA_MMV_Y        ?= 1
 	LLAMA_CUDA_KQUANTS_ITER ?= 2
@@ -519,16 +485,22 @@ ggml-mpi.o: ggml-mpi.c ggml-mpi.h
 	$(CC) $(CFLAGS) -c $< -o $@
 endif # LLAMA_MPI

-# combine build flags with cmdline overrides
-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)
+GF_CC := $(CC)
+include scripts/get-flags.mk

-# save CXXFLAGS before we add host-only options
-NVCCFLAGS := $(NVCCFLAGS) $(CXXFLAGS) $(CUDA_CXXFLAGS) -Wno-pedantic -Xcompiler "$(HOST_CXXFLAGS)"
-override CXXFLAGS += $(HOST_CXXFLAGS)
+# 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)
+
+# 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

 #
 # Print build information
@@ -10,11 +10,11 @@ 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

 ----

@@ -97,7 +97,18 @@ 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:**
@@ -421,14 +432,15 @@ Building the program with BLAS support may lead to some performance improvements
    ```bash
    make LLAMA_HIPBLAS=1
    ```
-  - Using `CMake` for Linux:
+  - Using `CMake` for Linux (assuming a gfx1030-compatible AMD GPU):
    ```bash
-    mkdir build
-    cd build
-    CC=/opt/rocm/llvm/bin/clang CXX=/opt/rocm/llvm/bin/clang++ cmake .. -DLLAMA_HIPBLAS=ON
-    cmake --build .
+    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
    ```
-  - Using `CMake` for Windows (using x64 Native Tools Command Prompt for VS):
+    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.
+  - Using `CMake` for Windows (using x64 Native Tools Command Prompt for VS, and assuming a gfx1100-compatible AMD GPU):
    ```bash
    set PATH=%HIP_PATH%\bin;%PATH%
    mkdir build
@@ -437,10 +449,11 @@ 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 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 (e.g. gfx1030, gfx1031, or gfx1035) 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 |
@@ -971,6 +984,8 @@ 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

@@ -920,7 +920,7 @@ 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 (default: %s)\n", params.model.c_str());
+    printf("                        draft model for speculative decoding\n");
    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");
@@ -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 (tensor->n_dims) {
+    switch (ggml_n_dims(tensor)) {
        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 (tensor->n_dims) {
+    switch (ggml_n_dims(tensor)) {
        case 1:
            for (int i0 = 0; i0 < tensor->ne[0]; i0++) {
                float * dst = (float *) ((char *) tensor->data + i0*tensor->nb[0]);
@@ -183,25 +183,27 @@ 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);
@@ -225,8 +227,8 @@ int64_t get_example_targets_batch(
    bool                   sample_random_offsets
 ) {
    GGML_ASSERT(samples_count > 0);
-    GGML_ASSERT(tokens_input->n_dims  == 2);
-    GGML_ASSERT(target_probs->n_dims  == 3);
+    GGML_ASSERT(ggml_is_matrix(tokens_input));
+    GGML_ASSERT(ggml_is_3d(target_probs));
    int64_t n_vocab  = target_probs->ne[0];
    int64_t n_tokens = tokens_input->ne[0];
    int64_t n_batch  = tokens_input->ne[1];
@@ -182,6 +182,8 @@ class Model:
            return QwenModel
        if model_architecture == "MixtralForCausalLM":
            return MixtralModel
+        if model_architecture == "PhiForCausalLM":
+            return Phi2Model
        return Model

    def _is_model_safetensors(self) -> bool:
@@ -221,6 +223,8 @@ 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!')

@@ -980,6 +984,24 @@ 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 ######


@@ -3,7 +3,6 @@ from __future__ import annotations

 import json
 import os
-import re
 import struct
 import sys
 from typing import Any, BinaryIO, Sequence
@@ -11,43 +10,15 @@ 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
@@ -61,9 +32,7 @@ def write_file_header(fout: BinaryIO, params: dict[str, Any]) -> None:
    fout.write(struct.pack("i", int(params["lora_alpha"])))


-def write_tensor_header(
-    self, name: str, shape: Sequence[int], data_type: np.dtype[Any]
-) -> None:
+def write_tensor_header(fout: BinaryIO, name: str, shape: Sequence[int], data_type: np.dtype[Any]) -> None:
    sname = name.encode("utf-8")
    fout.write(
        struct.pack(
@@ -78,11 +47,12 @@ def write_tensor_header(
    fout.seek((fout.tell() + 31) & -32)


-if len(sys.argv) != 2:
-    print(f"Usage: python {sys.argv[0]} <path>")
+if len(sys.argv) < 2:
+    print(f"Usage: python {sys.argv[0]} <path> [arch]")
    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")
@@ -90,6 +60,14 @@ 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)
@@ -117,6 +95,7 @@ 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"]:
@@ -129,7 +108,32 @@ with open(output_path, "wb") as fout:
            v = v.float()

        t = v.detach().numpy()
-        tname = translate_tensor_name(k)
+
+        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
+
        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)
@@ -10,6 +10,7 @@ import itertools
 import json
 import math
 import mmap
+import os
 import pickle
 import re
 import signal
@@ -18,15 +19,15 @@ 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, TypeVar
+from typing import IO, TYPE_CHECKING, Any, Callable, Iterable, Literal, Optional, TypeVar, cast

 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
@@ -327,127 +328,138 @@ class Params:
        return params


-#
-# vocab
-#
+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

-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"))
+        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)
        else:
-            # 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)
+            self.spm = None

-        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}")
+    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())

-        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
+        for i in range(self.vocab_size_base):
+            if i in added_tokens_ids:
+                continue

-    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()}
+            text = reverse_vocab[i].encode("utf-8")
+            yield text, self.get_token_score(i), self.get_token_type(i)

-        for i, _ in enumerate(tokenizer):
-            yield reverse_vocab[i], 0.0, gguf.TokenType.NORMAL
+    def get_token_type(self, token_id: int) -> gguf.TokenType:
+        toktype = gguf.TokenType.NORMAL

-    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.CONTROL
-
-    def all_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]:
-        yield from self.bpe_tokens()
-        yield from self.added_tokens()
-
-    def __repr__(self) -> str:
-        return f"<BpeVocab with {self.vocab_size_base} base tokens and {len(self.added_tokens_list)} added tokens>"
-
-
-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 = {}
-
-        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):
+        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 tokenizer.is_control(i):
+            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

-            # 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
+        return toktype

-            if tokenizer.is_unused(i):
-                toktype = gguf.TokenType.UNUSED
-            if tokenizer.is_byte(i):
-                toktype = gguf.TokenType.BYTE
-
-            yield text, score, 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

    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
+
+        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

    def all_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]:
-        yield from self.sentencepiece_tokens()
+        yield from self.hf_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"
+
+        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"
+
+        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"
+        )
+
    def __repr__(self) -> str:
-        return f"<SentencePieceVocab with {self.vocab_size_base} base tokens and {len(self.added_tokens_list)} added tokens>"
+        return f"<VocabLoader with {self.vocab_size_base} base tokens and {len(self.added_tokens_dict)} added tokens>"


-Vocab: TypeAlias = 'BpeVocab | SentencePieceVocab'
+Vocab: TypeAlias = 'VocabLoader'
+

 #
 # data loading
@@ -824,20 +836,27 @@ def bounded_parallel_map(func: Callable[[In], Out], iterable: Iterable[In], conc
            yield result


-def check_vocab_size(params: Params, vocab: Vocab) -> None:
+def check_vocab_size(params: Params, vocab: Vocab, pad_vocab: bool = False) -> None:
    if params.n_vocab != vocab.vocab_size:
-        assert isinstance(vocab, BpeVocab) or isinstance(vocab, SentencePieceVocab)
-        if params.n_vocab == vocab.vocab_size_base:
+        if params.n_vocab == vocab.vocab_size:
            print("Ignoring added_tokens.json since model matches vocab size without it.")
-            vocab.added_tokens_list = []
-            vocab.vocab_size = vocab.vocab_size_base
+            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
            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 and vocab.fname_added_tokens is None:
+        if vocab.vocab_size < params.n_vocab < vocab.vocab_size + 20:
            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)


@@ -901,12 +920,8 @@ class OutputFile:
            scores.append(score)
            toktypes.append(toktype)

-        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')
+        vocab_type = vocab.get_vocab_type()
+        self.gguf.add_tokenizer_model(vocab_type)
        self.gguf.add_token_list(tokens)
        self.gguf.add_token_scores(scores)
        self.gguf.add_token_types(toktypes)
@@ -932,8 +947,12 @@ 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) -> None:
-        check_vocab_size(params, vocab)
+    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)

        of = OutputFile(fname_out, endianess=endianess)

@@ -960,8 +979,13 @@ 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) -> None:
-        check_vocab_size(params, 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,
+        pad_vocab: bool = False,
+    ) -> None:
+        check_vocab_size(params, vocab, pad_vocab = pad_vocab)

        of = OutputFile(fname_out, endianess=endianess)

@@ -1119,35 +1143,17 @@ def load_some_model(path: Path) -> ModelPlus:
    return model_plus


-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")
+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

-    print(f"Loading vocab file '{path}', type '{vocabtype}'")
+    if path2.exists():
+        return path2
+    if path3.exists():
+        return path3

-    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}")
+    return None


 def default_outfile(model_paths: list[Path], file_type: GGMLFileType) -> Path:
@@ -1185,11 +1191,11 @@ 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, *.safetensors)")
-    parser.add_argument("--vocabtype",   choices=["spm", "bpe"], help="vocab format (default: spm)", default="spm")
+    parser.add_argument("model",         type=Path,              help="directory containing model file, or model file itself (*.pth, *.pt, *.bin)")
    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:
@@ -1232,12 +1238,13 @@ 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 = load_vocab(args.vocab_dir or args.model, args.vocabtype)
+        vocab = VocabLoader(params, args.vocab_dir or args.model)
        special_vocab = gguf.SpecialVocab(model_plus.paths[0].parent,
-                                          load_merges = args.vocabtype == 'bpe',
+                                          load_merges = True,
                                          n_vocab = vocab.vocab_size)
        outfile = args.outfile
-        OutputFile.write_vocab_only(outfile, params, vocab, special_vocab)
+        OutputFile.write_vocab_only(outfile, params, vocab, special_vocab,
+                                    endianess = endianess, pad_vocab = args.padvocab)
        print(f"Wrote {outfile}")
        return

@@ -1245,12 +1252,15 @@ 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 = load_vocab(vocab_dir, args.vocabtype)
+        vocab = VocabLoader(params, vocab_dir)
+
    # FIXME: Try to respect vocab_dir somehow?
+    print(f"Vocab info: {vocab}")
    special_vocab = gguf.SpecialVocab(model_plus.paths[0].parent,
-                                      load_merges = args.vocabtype == 'bpe',
+                                      load_merges = True,
                                      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)
@@ -1260,7 +1270,8 @@ 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)
+    OutputFile.write_all(outfile, ftype, params, model, vocab, special_vocab,
+                         concurrency = args.concurrency, endianess = endianess, pad_vocab = args.padvocab)
    print(f"Wrote {outfile}")


@@ -1258,9 +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(logits->n_dims == 2);
-    assert(probs->n_dims == 2);
-    assert(best_samples->n_dims == 1);
+    assert(ggml_is_matrix(logits));
+    assert(ggml_is_matrix(probs));
+    assert(ggml_is_vector(best_samples));
    assert(logits->ne[1] == best_samples->ne[0]);
    assert(logits->ne[0] == probs->ne[0]);
    assert(logits->ne[1] == probs->ne[1]);
@@ -1292,9 +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(best_samples->n_dims == 2);
-    GGML_ASSERT(logits->n_dims == 3);
-    GGML_ASSERT(probs->n_dims == 3);
+    GGML_ASSERT(ggml_is_matrix(best_samples));
+    GGML_ASSERT(ggml_is_3d(logits));
+    GGML_ASSERT(ggml_is_3d(probs));
    int n_tokens = best_samples->ne[0];
    int n_batch  = best_samples->ne[1];
    int n_vocab  = logits->ne[0];
@@ -1334,7 +1334,7 @@ static void print_row(struct ggml_tensor * probs, int i) {
 }

 static void print_matrix(struct ggml_tensor * probs) {
-    assert(probs->n_dims == 2);
+    assert(ggml_is_matrix(probs));
    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);
@@ -1386,8 +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(tokens_input->n_dims == 2);
-    GGML_ASSERT(     targets->n_dims == 3);
+    GGML_ASSERT(ggml_is_matrix(tokens_input));
+    GGML_ASSERT(ggml_is_3d(targets));
    int n_tokens = tokens_input->ne[0];
    int n_batch  = tokens_input->ne[1];
    GGML_ASSERT(n_tokens == targets->ne[1]);
@@ -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 += 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 += 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 += 1024*1024*16;

    printf("Allocating Memory of size %zi bytes, %zi MB\n",ctx_size, (ctx_size/1024/1024));
@@ -427,7 +427,7 @@ static void print_row(struct ggml_tensor * probs, int i) {
 }

 static void print_matrix(struct ggml_tensor * probs) {
-    assert(probs->n_dims == 2);
+    assert(ggml_is_matrix(probs));
    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 (gg_weights->n_dims){
+    switch (ggml_n_dims(gg_weights)) {
        case 1:
            ct = 0;
            for (int i0 = 0; i0 < gg_weights->ne[0]; i0++){
@@ -1110,7 +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 = tensor->n_dims;
+    uint32_t nd = ggml_n_dims(tensor);
    uint32_t ne[4] = { (uint32_t)tensor->ne[0],
                       (uint32_t)tensor->ne[1],
                       (uint32_t)tensor->ne[2],
@@ -1620,8 +1620,6 @@ 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);

@@ -1725,10 +1723,9 @@ int main(int argc, char ** argv) {

    // allocate input tensors
    mem_input_data.resize(max_input_size);
-    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);
+    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);

    // context for compute tensors without their data
    const size_t estimated_compute_size_wo_data = (
@@ -1755,7 +1752,7 @@ 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);
-        alloc = ggml_allocr_new_measure(tensor_alignment);
+        ggml_allocr_t 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);
@@ -1788,7 +1785,7 @@ int main(int argc, char ** argv) {
    // allocate compute tensors
    mem_compute_data.resize(max_compute_size);
    ctx_compute = ggml_init(ctx_compute_params);
-    alloc = ggml_allocr_new(mem_compute_data.data(), mem_compute_data.size(), tensor_alignment);
+    ggml_allocr_t 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);
@@ -1804,6 +1801,8 @@ 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;
@@ -195,7 +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, cur->n_dims, cur->name, cur->data);
+            printf("%s: tensor[%d]: n_dims = %d, name = %s, data = %p\n", __func__, i, ggml_n_dims(cur), cur->name, cur->data);

            // print first 10 elements
            const float * data = (const float *) cur->data;
@@ -1 +1,2 @@
 xcuserdata
+xcshareddata
@@ -6,16 +6,34 @@ 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 = 512
+    var n_len: Int32 = 64
    var n_cur: Int32 = 0
+
    var n_decode: Int32 = 0

    init(model: OpaquePointer, context: OpaquePointer) {
@@ -27,25 +45,34 @@ actor LlamaContext {
    }

    deinit {
+        llama_batch_free(batch)
        llama_free(context)
        llama_free_model(model)
        llama_backend_free()
    }

-    static func createContext(path: String) throws -> LlamaContext {
+    static func create_context(path: String) throws -> LlamaContext {
        llama_backend_init(false)
-        let model_params = llama_model_default_params()
+        var 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 = 8
-        ctx_params.n_threads_batch = 8
+        ctx_params.n_threads       = UInt32(n_threads)
+        ctx_params.n_threads_batch = UInt32(n_threads)

        let context = llama_new_context_with_model(model, ctx_params)
        guard let context else {
@@ -56,6 +83,26 @@ 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;
    }
@@ -79,16 +126,11 @@ actor LlamaContext {
            print(String(cString: token_to_piece(token: id) + [0]))
        }

-        // batch = llama_batch_init(512, 0) // done in init()
-        batch.n_tokens = Int32(tokens_list.count)
+        llama_batch_clear(&batch)

-        for i1 in 0..<batch.n_tokens {
+        for i1 in 0..<tokens_list.count {
            let i = Int(i1)
-            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
+            llama_batch_add(&batch, tokens_list[i], Int32(i), [0], false)
        }
        batch.logits[Int(batch.n_tokens) - 1] = 1 // true

@@ -141,18 +183,11 @@ actor LlamaContext {
        print(new_token_str)
        // tokens_list.append(new_token_id)

-        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
+        llama_batch_clear(&batch)
+        llama_batch_add(&batch, new_token_id, n_cur, [0], true)

        n_decode += 1
-
-        n_cur += 1
+        n_cur    += 1

        if llama_decode(context, batch) != 0 {
            print("failed to evaluate llama!")
@@ -161,14 +196,111 @@ 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)
+        let n_tokens = utf8Count + (add_bos ? 1 : 0) + 1
        let tokens = UnsafeMutablePointer<llama_token>.allocate(capacity: n_tokens)
        let tokenCount = llama_tokenize(model, text, Int32(utf8Count), tokens, Int32(n_tokens), add_bos, false)

@@ -1,481 +1,483 @@
 // !$*UTF8*$!
 {
-    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 */; };
-        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 */; };
+		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 */; };
 /* 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; };
-        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>"; };
+		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>"; };
 /* 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 = (
-                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>";
-        };
+		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>";
+		};
 /* 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 */,
-                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 */,
+				7FA3D2B32B2EA2F600543F92 /* DownloadButton.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 */;
 }
@@ -3,24 +3,26 @@ import Foundation
@MainActor
 class LlamaState: ObservableObject {
    @Published var messageLog = ""
+    @Published var cacheCleared = false

    private var llamaContext: LlamaContext?
-    private var modelUrl: URL? {
-        Bundle.main.url(forResource: "q8_0", withExtension: "gguf", subdirectory: "models")
+    private var defaultModelUrl: URL? {
+        Bundle.main.url(forResource: "ggml-model", withExtension: "gguf", subdirectory: "models")
        // Bundle.main.url(forResource: "llama-2-7b-chat", withExtension: "Q2_K.gguf", subdirectory: "models")
    }
+
    init() {
        do {
-            try loadModel()
+            try loadModel(modelUrl: defaultModelUrl)
        } catch {
            messageLog += "Error!\n"
        }
    }

-    private func loadModel() throws {
+    func loadModel(modelUrl: URL?) throws {
        messageLog += "Loading model...\n"
        if let modelUrl {
-            llamaContext = try LlamaContext.createContext(path: modelUrl.path())
+            llamaContext = try LlamaContext.create_context(path: modelUrl.path())
            messageLog += "Loaded model \(modelUrl.lastPathComponent)\n"
        } else {
            messageLog += "Could not locate model\n"
@@ -31,7 +33,7 @@ class LlamaState: ObservableObject {
        guard let llamaContext else {
            return
        }
-        messageLog += "Attempting to complete text...\n"
+
        await llamaContext.completion_init(text: text)
        messageLog += "\(text)"

@@ -42,4 +44,42 @@ 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 = ""
+    }
 }
@@ -5,24 +5,132 @@ 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) {
+            ScrollView(.vertical, showsIndicators: true) {
                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: 200)
+                .frame(height: 80)
                .padding()
                .border(Color.gray, width: 0.5)
-            Button(action: {
-                sendText()
-            }) {
-                Text("Send")
-                    .padding()
-                    .background(Color.blue)
-                    .foregroundColor(.white)
-                    .cornerRadius(8)
+
+            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))
            }
        }
        .padding()
@@ -34,9 +142,20 @@ struct ContentView: View {
            multiLineText = ""
        }
    }
+
+    func bench() {
+        Task {
+            await llamaState.bench()
+        }
+    }
+
+    func clear() {
+        Task {
+            await llamaState.clear()
+        }
+    }
 }
-/*
-#Preview {
-    ContentView()
-}
-*/
+
+//#Preview {
+//    ContentView()
+//}
@@ -0,0 +1,122 @@
+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"
+//    )
+// }
@@ -514,7 +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,
-                       cur->n_dims, cur->name, tensor_size, padded_size, offset);
+                       ggml_n_dims(cur), cur->name, tensor_size, padded_size, offset);
            }
        }
    }
@@ -962,7 +962,7 @@ bool clip_model_quantize(const char * fname_inp, const char * fname_out, const i
        }

        // quantize only 2D tensors
-        quantize &= (cur->n_dims == 2);
+        quantize &= (ggml_n_dims(cur) == 2);

        if (quantize) {
            new_type = type;
@@ -1035,7 +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(), cur->n_dims, quantize,
+        printf("%s: n_dims = %d | quantize=%d | size = %f MB -> %f MB\n", name.c_str(), ggml_n_dims(cur), quantize,
               orig_size / 1024.0 / 1024.0, new_size / 1024.0 / 1024.0);
    }

@@ -34,7 +34,8 @@ export async function* llama(prompt, params = {}, config = {}) {
    headers: {
      'Connection': 'keep-alive',
      'Content-Type': 'application/json',
-      'Accept': 'text/event-stream'
+      'Accept': 'text/event-stream',
+      ...(params.api_key ? {'Authorization': `Bearer ${params.api_key}`} : {})
    },
    signal: controller.signal,
  });
@@ -235,7 +235,8 @@
      grammar: '',
      n_probs: 0, // no completion_probabilities,
      image_data: [],
-      cache_prompt: true
+      cache_prompt: true,
+      api_key: ''
    })

    /* START: Support for storing prompt templates and parameters in browsers LocalStorage */
@@ -790,6 +791,10 @@
            <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>
      `
@@ -10,7 +10,8 @@
 // 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"

@@ -36,6 +37,7 @@ 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;
@@ -376,7 +378,6 @@ 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;
@@ -425,7 +426,6 @@ 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;
@@ -992,35 +992,36 @@ struct llama_server_context
        slot.generated_text += token_str;
        slot.has_next_token = true;

-        if (slot.multibyte_pending > 0)
+        // 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)
        {
-            slot.multibyte_pending -= token_str.size();
-        }
-        else if (token_str.size() == 1)
-        {
-            const char c = token_str[0];
-            // 2-byte characters: 110xxxxx 10xxxxxx
+            unsigned char c = slot.generated_text[slot.generated_text.size() - i];
+            if ((c & 0xC0) == 0x80)
+            {
+                // continuation byte: 10xxxxxx
+                continue;
+            }
            if ((c & 0xE0) == 0xC0)
            {
-                slot.multibyte_pending = 1;
-                // 3-byte characters: 1110xxxx 10xxxxxx 10xxxxxx
+                // 2-byte character: 110xxxxx ...
+                incomplete = i < 2;
            }
            else if ((c & 0xF0) == 0xE0)
            {
-                slot.multibyte_pending = 2;
-                // 4-byte characters: 11110xxx 10xxxxxx 10xxxxxx 10xxxxxx
+                // 3-byte character: 1110xxxx ...
+                incomplete = i < 3;
            }
            else if ((c & 0xF8) == 0xF0)
            {
-                slot.multibyte_pending = 3;
-            }
-            else
-            {
-                slot.multibyte_pending = 0;
+                // 4-byte character: 11110xxx ...
+                incomplete = i < 4;
            }
+            // else 1-byte character or invalid byte
+            break;
        }

-        if (slot.multibyte_pending == 0)
+        if (!incomplete)
        {
            size_t pos = std::min(slot.sent_count, slot.generated_text.size());
            const std::string str_test = slot.generated_text.substr(pos);
@@ -1055,7 +1056,7 @@ struct llama_server_context
            }
        }

-        if (slot.multibyte_pending > 0 && !slot.has_next_token)
+        if (incomplete)
        {
            slot.has_next_token = true;
        }
@@ -1954,6 +1955,7 @@ 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);
@@ -2003,6 +2005,15 @@ 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)
@@ -2403,7 +2414,7 @@ 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 (llama_params.count("grammar") != 0) {
+    if (body.count("grammar") != 0) {
        llama_params["grammar"] = json_value(body, "grammar", json::object());
    }

@@ -2634,6 +2645,9 @@ 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;
@@ -2670,6 +2684,32 @@ 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"}});
@@ -2677,28 +2717,28 @@ 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");
+                res.set_content(reinterpret_cast<const char*>(&index_html), index_html_len, "text/html; charset=utf-8");
                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");
+                res.set_content(reinterpret_cast<const char *>(&index_js), index_js_len, "text/javascript; charset=utf-8");
                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");
+                res.set_content(reinterpret_cast<const char*>(&completion_js), completion_js_len, "application/javascript; charset=utf-8");
                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");
+                res.set_content(reinterpret_cast<const char*>(&json_schema_to_grammar_mjs), json_schema_to_grammar_mjs_len, "application/javascript; charset=utf-8");
                return false;
            });

@@ -2709,23 +2749,26 @@ 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");
+                res.set_content(data.dump(), "application/json; charset=utf-8");
            });

-    svr.Post("/completion", [&llama](const httplib::Request &req, httplib::Response &res)
+    svr.Post("/completion", [&llama, &validate_api_key](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");
+                        res.set_content(result.result_json.dump(-1, ' ', false, json::error_handler_t::replace), "application/json; charset=utf-8");
                    }
                    else
                    {
                        res.status = 404;
-                        res.set_content(result.result_json["content"], "text/plain");
+                        res.set_content(result.result_json["content"], "text/plain; charset=utf-8");
                        return;
                    }
                } else {
@@ -2796,12 +2839,15 @@ int main(int argc, char **argv)
                    }}
                };

-                res.set_content(models.dump(), "application/json");
+                res.set_content(models.dump(), "application/json; charset=utf-8");
            });

    // TODO: add mount point without "/v1" prefix -- how?
-    svr.Post("/v1/chat/completions", [&llama](const httplib::Request &req, httplib::Response &res)
+    svr.Post("/v1/chat/completions", [&llama, &validate_api_key](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);
@@ -2815,10 +2861,10 @@ int main(int argc, char **argv)

                        res.set_content(oaicompat_result.dump(-1, ' ', false,
                                            json::error_handler_t::replace),
-                                            "application/json");
+                                            "application/json; charset=utf-8");
                    } else {
                        res.status = 500;
-                        res.set_content(result.result_json["content"], "text/plain");
+                        res.set_content(result.result_json["content"], "text/plain; charset=utf-8");
                        return;
                    }
                } else {
@@ -2870,8 +2916,11 @@ int main(int argc, char **argv)
                }
            });

-    svr.Post("/infill", [&llama](const httplib::Request &req, httplib::Response &res)
+    svr.Post("/infill", [&llama, &validate_api_key](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)) {
@@ -2879,12 +2928,12 @@ 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");
+                        res.set_content(result.result_json.dump(-1, ' ', false, json::error_handler_t::replace), "application/json; charset=utf-8");
                    }
                    else
                    {
                        res.status = 404;
-                        res.set_content(result.result_json["content"], "text/plain");
+                        res.set_content(result.result_json["content"], "text/plain; charset=utf-8");
                        return;
                    }
                } else {
@@ -2933,11 +2982,11 @@ 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");
+                return res.set_content(data.dump(), "application/json; charset=utf-8");
            });

    svr.Options(R"(/.*)", [](const httplib::Request &, httplib::Response &res)
-                { return res.set_content("", "application/json"); });
+                { return res.set_content("", "application/json; charset=utf-8"); });

    svr.Post("/tokenize", [&llama](const httplib::Request &req, httplib::Response &res)
            {
@@ -2948,7 +2997,7 @@ 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");
+                return res.set_content(data.dump(), "application/json; charset=utf-8");
            });

    svr.Post("/detokenize", [&llama](const httplib::Request &req, httplib::Response &res)
@@ -2962,7 +3011,7 @@ int main(int argc, char **argv)
                }

                const json data = format_detokenized_response(content);
-                return res.set_content(data.dump(), "application/json");
+                return res.set_content(data.dump(), "application/json; charset=utf-8");
            });

    svr.Post("/embedding", [&llama](const httplib::Request &req, httplib::Response &res)
@@ -2979,7 +3028,7 @@ 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");
+                return res.set_content(result.result_json.dump(), "application/json; charset=utf-8");
            });

    svr.set_logger(log_server_request);
@@ -3000,19 +3049,23 @@ int main(int argc, char **argv)
                {
                    snprintf(buf, sizeof(buf), fmt, "Unknown Exception");
                }
-                res.set_content(buf, "text/plain");
+                res.set_content(buf, "text/plain; charset=utf-8");
                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");
+                    res.set_content("Invalid request", "text/plain; charset=utf-8");
                }
-                else if (res.status != 500)
+                else if (res.status == 404)
                {
-                    res.set_content("File Not Found", "text/plain");
+                    res.set_content("File Not Found", "text/plain; charset=utf-8");
                    res.status = 404;
                }
            });
@@ -3033,11 +3086,15 @@ 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);

-    LOG_INFO("HTTP server listening", {
-                                          {"hostname", sparams.hostname},
-                                          {"port", sparams.port},
-                                      });
+    std::unordered_map<std::string, std::string> log_data;
+    log_data["hostname"] = sparams.hostname;
+    log_data["port"] = std::to_string(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([&]()
            {
@@ -66,9 +66,11 @@ 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(silu);
+    GGML_METAL_DECL_KERNEL(tanh);
    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);
@@ -86,6 +88,7 @@ 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);
@@ -145,8 +148,11 @@ 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);
@@ -334,9 +340,11 @@ 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(silu);
+        GGML_METAL_ADD_KERNEL(tanh);
        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);
@@ -354,6 +362,7 @@ 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);
@@ -415,8 +424,11 @@ 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);
@@ -450,9 +462,11 @@ 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(silu);
+    GGML_METAL_DEL_KERNEL(tanh);
    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);
@@ -470,6 +484,7 @@ 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);
@@ -531,8 +546,11 @@ 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);
@@ -843,9 +861,11 @@ 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_SILU:
+                case GGML_UNARY_OP_TANH:
                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;
@@ -853,11 +873,11 @@ 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_PERMUTE:
        case GGML_OP_TRANSPOSE:
-        case GGML_OP_GET_ROWS:
+        case GGML_OP_PERMUTE:
        case GGML_OP_CONCAT:
        case GGML_OP_ADD:
+        case GGML_OP_ACC:
        case GGML_OP_MUL:
        case GGML_OP_DIV:
        case GGML_OP_SCALE:
@@ -865,11 +885,15 @@ 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;
@@ -902,8 +926,9 @@ 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[0] % 4 == 0;
+                return op->ne[3] == 1;
            }
        default:
            return false;
@@ -979,7 +1004,10 @@ void ggml_metal_graph_compute(
                        } break;
                }

-                GGML_ASSERT(ggml_metal_supports_op(dst));
+                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");
+                }

                const int64_t  ne00 = src0 ? src0->ne[0] : 0;
                const int64_t  ne01 = src0 ? src0->ne[1] : 0;
@@ -1076,6 +1104,8 @@ 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;
@@ -1134,7 +1164,8 @@ 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:&nb   length:sizeof(nb)   atIndex:27];
+                            [encoder setBytes:&offs length:sizeof(offs) atIndex:27];
+                            [encoder setBytes:&nb   length:sizeof(nb)   atIndex:28];

                            if (bcast_row) {
                                const int64_t n = ggml_nelements(dst)/4;
@@ -1146,6 +1177,86 @@ 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));
@@ -1169,16 +1280,15 @@ void ggml_metal_graph_compute(
                        } break;
                    case GGML_OP_UNARY:
                        switch (ggml_get_unary_op(gf->nodes[i])) {
-                            case GGML_UNARY_OP_SILU:
+                            case GGML_UNARY_OP_TANH:
                                {
-                                    [encoder setComputePipelineState:ctx->pipeline_silu];
+                                    [encoder setComputePipelineState:ctx->pipeline_tanh];
                                    [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)];
+                                    [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
                                } break;
                            case GGML_UNARY_OP_RELU:
                                {
@@ -1199,6 +1309,28 @@ 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:
@@ -1837,6 +1969,38 @@ 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;
@@ -2006,6 +2170,65 @@ 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);
@@ -2027,6 +2250,22 @@ 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:
@@ -79,6 +79,7 @@ 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]]) {
@@ -90,9 +91,9 @@ 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;
+    device const char * src0_ptr = src0 + i03*nb03 + i02*nb02 + i01*nb01 + offs;
    device const char * src1_ptr = src1 + i13*nb13 + i12*nb12 + i11*nb11;
-    device       char * dst_ptr  = dst  + i03*nb3  + i02*nb2  + i01*nb1;
+    device       char * dst_ptr  = dst  + i03*nb3  + i02*nb2  + i01*nb1  + offs;

    for (int i0 = tpitg.x; i0 < ne0; i0 += ntg.x) {
        const int i10 = i0 % ne10;
@@ -204,7 +205,7 @@ kernel void kernel_add_row(
        device const float4 * src0,
        device const float4 * src1,
        device       float4 * dst,
-        constant    int64_t & nb [[buffer(27)]],
+        constant    int64_t & nb [[buffer(28)]],
        uint tpig[[thread_position_in_grid]]) {
    dst[tpig] = src0[tpig] + src1[tpig % nb];
 }
@@ -213,7 +214,7 @@ kernel void kernel_mul_row(
        device const float4 * src0,
        device const float4 * src1,
        device       float4 * dst,
-        constant    int64_t & nb  [[buffer(27)]],
+        constant    int64_t & nb  [[buffer(28)]],
        uint tpig[[thread_position_in_grid]]) {
    dst[tpig] = src0[tpig] * src1[tpig % nb];
 }
@@ -222,7 +223,7 @@ kernel void kernel_div_row(
        device const float4 * src0,
        device const float4 * src1,
        device       float4 * dst,
-        constant    int64_t & nb  [[buffer(27)]],
+        constant    int64_t & nb  [[buffer(28)]],
        uint tpig[[thread_position_in_grid]]) {
    dst[tpig] = src0[tpig] / src1[tpig % nb];
 }
@@ -243,6 +244,47 @@ 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,
@@ -251,13 +293,6 @@ 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,
@@ -313,22 +348,6 @@ 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,
@@ -650,6 +669,94 @@ 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
@@ -1595,8 +1702,9 @@ kernel void kernel_rope(
            dst_data[1] = x0*sin_theta + x1*cos_theta;
        }
    } else {
-        for (int64_t ib = 0; ib < ne0/n_dims; ++ib) {
-            for (int64_t ic = 2*tiitg; ic < n_dims; ic += 2*tptg.x) {
+        for (int64_t ic = 2*tiitg; ic < ne0; ic += 2*tptg.x) {
+            if (ic < n_dims) {
+                const int64_t ib = 0;

                // simplified from `(ib * n_dims + ic) * inv_ndims`
                const float cur_rot = inv_ndims*ic - ib;
@@ -1615,6 +1723,14 @@ 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];
            }
        }
    }
@@ -1656,6 +1772,97 @@ 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,
@@ -1708,6 +1915,14 @@ 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,
@@ -2066,9 +2281,9 @@ 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,
@@ -2105,7 +2320,7 @@ 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;

@@ -3315,10 +3530,10 @@ 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 half d = xb->d;
-    const half min = xb->dmin;
+    const float d = xb->d;
+    const float min = xb->dmin;
    device const uint8_t * q = (device const uint8_t *)xb->qs;
-    half dl, ml;
+    float dl, ml;
    uint8_t sc = xb->scales[il];

 #if QK_K == 256
@@ -3388,10 +3603,10 @@ 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 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 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];
 #else
    q = q + 16 * (il&1);
    device const uint8_t * s = xb->scales;
@@ -3418,13 +3633,13 @@ 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 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 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 ushort mask = il<2 ? 0x0F : 0xF0;
-    const half qh_val = il<2 ? 16.h : 256.h;
+    const ushort mask  = il<2 ? 0x0F : 0xF0;
+    const float qh_val = il<2 ? 16.f : 256.f;
    for (int i = 0; i < 16; ++i) {
        reg[i/4][i%4] = dl * ((q[i] & mask) + (qh[i] & ul ? qh_val : 0)) - ml;
    }
@@ -303,7 +303,7 @@ extern "C" {

 #if defined(__ARM_NEON) && defined(__CUDACC__)
    typedef half ggml_fp16_t;
-#elif defined(__ARM_NEON)
+#elif defined(__ARM_NEON) && !defined(_MSC_VER)
    typedef __fp16 ggml_fp16_t;
 #else
    typedef uint16_t ggml_fp16_t;
@@ -343,6 +343,12 @@ 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,
@@ -423,7 +429,9 @@ 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,
@@ -463,7 +471,6 @@ extern "C" {
        GGML_UNARY_OP_GELU,
        GGML_UNARY_OP_GELU_QUICK,
        GGML_UNARY_OP_SILU,
-        GGML_UNARY_OP_LEAKY,

        GGML_UNARY_OP_COUNT,
    };
@@ -501,7 +508,6 @@ 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)
@@ -533,7 +539,7 @@ extern "C" {

        void * extra; // extra things e.g. for ggml-cuda.cu

-        char padding[12];
+        char padding[8];
    };

    static const size_t GGML_TENSOR_SIZE = sizeof(struct ggml_tensor);
@@ -638,11 +644,14 @@ 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 float   ggml_type_sizef(enum ggml_type type); // ggml_type_size()/ggml_blck_size() as float
+    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 const char * ggml_type_name(enum ggml_type type);
    GGML_API const char * ggml_op_name  (enum ggml_op   op);
@@ -661,6 +670,11 @@ 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);

@@ -793,6 +807,9 @@ 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,
@@ -957,15 +974,14 @@ extern "C" {
            struct ggml_context * ctx,
            struct ggml_tensor  * a);

-    GGML_API struct ggml_tensor * ggml_leaky(
+    GGML_API struct ggml_tensor * ggml_leaky_relu(
            struct ggml_context * ctx,
-            struct ggml_tensor  * a);
+            struct ggml_tensor  * a, float negative_slope, bool inplace);

    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);
@@ -1047,6 +1063,12 @@ 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(
@@ -1551,6 +1573,15 @@ 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,
@@ -95,6 +95,7 @@ class MODEL_ARCH(IntEnum):
    BLOOM     = auto()
    STABLELM  = auto()
    QWEN      = auto()
+    PHI2      = auto()


 class MODEL_TENSOR(IntEnum):
@@ -140,6 +141,7 @@ 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] = {
@@ -350,6 +352,17 @@ 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
 }

@@ -17,6 +17,7 @@ class TensorNameMap:
            "tok_embeddings",                            # llama-pth
            "embeddings.word_embeddings",                # bert
            "language_model.embedding.word_embeddings",  # persimmon
+            "transformer.embd.wte",                      # phi2
        ),

        # Token type embeddings
@@ -41,6 +42,7 @@ 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
@@ -53,6 +55,7 @@ class TensorNameMap:
            "transformer.norm_f",                      # mpt
            "ln_f",                                    # refact bloom qwen
            "language_model.encoder.final_layernorm",  # persimmon
+            "lm_head.ln",                              # phi2
        ),

        # Rope frequencies
@@ -75,6 +78,7 @@ 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
@@ -90,6 +94,7 @@ 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
@@ -128,6 +133,7 @@ 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
@@ -167,6 +173,7 @@ 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: (
@@ -198,6 +205,7 @@ 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: (
@@ -84,7 +84,7 @@ class SpecialVocab:
        merges_file = path / 'merges.txt'
        if not merges_file.is_file():
            return False
-        with open(merges_file, 'r') as fp:
+        with open(merges_file, 'r', encoding = 'utf-8') as fp:
            first_line = next(fp, '').strip()
            if not first_line.startswith('#'):
                fp.seek(0)
@@ -109,8 +109,10 @@ class SpecialVocab:
        return True

    def _set_special_token(self, typ: str, tid: Any) -> None:
-        if not isinstance(tid, int) or tid < 0:
+        if not isinstance(tid, int):
            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
@@ -195,6 +195,7 @@ enum llm_arch {
    LLM_ARCH_BLOOM,
    LLM_ARCH_STABLELM,
    LLM_ARCH_QWEN,
+    LLM_ARCH_PHI2,
    LLM_ARCH_UNKNOWN,
 };

@@ -212,6 +213,7 @@ static std::map<llm_arch, std::string> LLM_ARCH_NAMES = {
    { LLM_ARCH_BLOOM,           "bloom"     },
    { LLM_ARCH_STABLELM,        "stablelm"  },
    { LLM_ARCH_QWEN,            "qwen"      },
+    { LLM_ARCH_PHI2,            "phi2"      },
 };

 enum llm_kv {
@@ -550,6 +552,19 @@ static std::map<llm_arch, std::map<llm_tensor, std::string>> LLM_TENSOR_NAMES =
            { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
        },
    },
+    {
+        LLM_ARCH_PHI2,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,     "output_norm" },
+            { LLM_TENSOR_OUTPUT,          "output" },
+            { LLM_TENSOR_ATTN_NORM,       "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_QKV,        "blk.%d.attn_qkv" },
+            { LLM_TENSOR_ATTN_OUT,        "blk.%d.attn_output" },
+            { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
+        },
+    },

    {
        LLM_ARCH_UNKNOWN,
@@ -1420,6 +1435,7 @@ struct llama_model {
    struct ggml_tensor * output_norm;
    struct ggml_tensor * output_norm_b;
    struct ggml_tensor * output;
+    struct ggml_tensor * output_b;

    std::vector<llama_layer> layers;

@@ -1505,6 +1521,10 @@ struct llama_context {

    // decode output (2-dimensional array: [n_tokens][n_vocab])
    std::vector<float> logits;
+#ifndef NDEBUG
+    // guard against access to unset logits
+    std::vector<bool>  logits_valid;
+#endif
    bool logits_all = false;

    // input embedding (1-dimensional array: [n_embd])
@@ -1555,7 +1575,7 @@ static bool llama_kv_cache_init(
    cache.cells.clear();
    cache.cells.resize(n_ctx);

-    cache.buf.resize(n_elements*(ggml_type_sizef(ktype) + ggml_type_sizef(vtype)) + 2u*n_layer*ggml_tensor_overhead());
+    cache.buf.resize(ggml_row_size(ktype, n_elements) + ggml_row_size(vtype, n_elements) + 2u*n_layer*ggml_tensor_overhead());
    memset(cache.buf.data, 0, cache.buf.size);

    struct ggml_init_params params;
@@ -1933,7 +1953,7 @@ namespace GGUFMeta {
                target = override->bool_value;
                return true;
            }
-            return true;
+            return false;
        }

        template<typename OT>
@@ -2063,7 +2083,7 @@ struct llama_model_loader {
                    type_max   = meta->type;
                }

-                LLAMA_LOG_INFO("%s: - tensor %4d: %32s %-8s [ %s ]\n", __func__, i, name, ggml_type_name(meta->type), llama_format_tensor_shape(meta).c_str());
+                // LLAMA_LOG_INFO("%s: - tensor %4d: %32s %-8s [ %s ]\n", __func__, i, name, ggml_type_name(meta->type), llama_format_tensor_shape(meta).c_str());
            }

            switch (type_max) {
@@ -2393,25 +2413,25 @@ static std::string llama_model_ftype_name(llama_ftype ftype) {

    switch (ftype) {
        case LLAMA_FTYPE_ALL_F32:     return "all F32";
-        case LLAMA_FTYPE_MOSTLY_F16:  return "mostly F16";
-        case LLAMA_FTYPE_MOSTLY_Q4_0: return "mostly Q4_0";
-        case LLAMA_FTYPE_MOSTLY_Q4_1: return "mostly Q4_1";
+        case LLAMA_FTYPE_MOSTLY_F16:  return "F16";
+        case LLAMA_FTYPE_MOSTLY_Q4_0: return "Q4_0";
+        case LLAMA_FTYPE_MOSTLY_Q4_1: return "Q4_1";
        case LLAMA_FTYPE_MOSTLY_Q4_1_SOME_F16:
-                                      return "mostly Q4_1, some F16";
-        case LLAMA_FTYPE_MOSTLY_Q5_0: return "mostly Q5_0";
-        case LLAMA_FTYPE_MOSTLY_Q5_1: return "mostly Q5_1";
-        case LLAMA_FTYPE_MOSTLY_Q8_0: return "mostly Q8_0";
+                                      return "Q4_1, some F16";
+        case LLAMA_FTYPE_MOSTLY_Q5_0: return "Q5_0";
+        case LLAMA_FTYPE_MOSTLY_Q5_1: return "Q5_1";
+        case LLAMA_FTYPE_MOSTLY_Q8_0: return "Q8_0";

        // K-quants
-        case LLAMA_FTYPE_MOSTLY_Q2_K:   return "mostly Q2_K";
-        case LLAMA_FTYPE_MOSTLY_Q3_K_S: return "mostly Q3_K - Small";
-        case LLAMA_FTYPE_MOSTLY_Q3_K_M: return "mostly Q3_K - Medium";
-        case LLAMA_FTYPE_MOSTLY_Q3_K_L: return "mostly Q3_K - Large";
-        case LLAMA_FTYPE_MOSTLY_Q4_K_S: return "mostly Q4_K - Small";
-        case LLAMA_FTYPE_MOSTLY_Q4_K_M: return "mostly Q4_K - Medium";
-        case LLAMA_FTYPE_MOSTLY_Q5_K_S: return "mostly Q5_K - Small";
-        case LLAMA_FTYPE_MOSTLY_Q5_K_M: return "mostly Q5_K - Medium";
-        case LLAMA_FTYPE_MOSTLY_Q6_K:   return "mostly Q6_K";
+        case LLAMA_FTYPE_MOSTLY_Q2_K:   return "Q2_K";
+        case LLAMA_FTYPE_MOSTLY_Q3_K_S: return "Q3_K - Small";
+        case LLAMA_FTYPE_MOSTLY_Q3_K_M: return "Q3_K - Medium";
+        case LLAMA_FTYPE_MOSTLY_Q3_K_L: return "Q3_K - Large";
+        case LLAMA_FTYPE_MOSTLY_Q4_K_S: return "Q4_K - Small";
+        case LLAMA_FTYPE_MOSTLY_Q4_K_M: return "Q4_K - Medium";
+        case LLAMA_FTYPE_MOSTLY_Q5_K_S: return "Q5_K - Small";
+        case LLAMA_FTYPE_MOSTLY_Q5_K_M: return "Q5_K - Medium";
+        case LLAMA_FTYPE_MOSTLY_Q6_K:   return "Q6_K";

        default: return "unknown, may not work";
    }
@@ -2529,6 +2549,7 @@ static void llm_load_hparams(
                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);

                switch (hparams.n_layer) {
+                    case 22: model.type = e_model::MODEL_1B; break;
                    case 26: model.type = e_model::MODEL_3B; break;
                    case 32: model.type = e_model::MODEL_7B; break;
                    case 40: model.type = e_model::MODEL_13B; break;
@@ -2630,6 +2651,15 @@ static void llm_load_hparams(
                    default: model.type = e_model::MODEL_UNKNOWN;
                }
            } break;
+        case LLM_ARCH_PHI2:
+            {
+                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps);
+
+                switch (hparams.n_layer) {
+                    case 32: model.type = e_model::MODEL_3B; break;
+                    default: model.type = e_model::MODEL_UNKNOWN;
+                }
+            } break;

        default: (void)0;
    }
@@ -2982,7 +3012,7 @@ static void llm_load_tensors(

    (void) main_gpu;

-    enum ggml_backend_type llama_backend_offload = GGML_BACKEND_CPU;
+    enum ggml_backend_type llama_backend_offload       = GGML_BACKEND_CPU;
    enum ggml_backend_type llama_backend_offload_split = GGML_BACKEND_CPU;

 #ifdef GGML_USE_CUBLAS
@@ -3625,7 +3655,73 @@ static void llm_load_tensors(
                        }
                    }
                } break;
+            case LLM_ARCH_PHI2:
+                {
+                    model.tok_embd = ml.create_tensor(ctx, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, GGML_BACKEND_CPU);

+                    // output
+                    {
+                        ggml_backend_type backend_norm;
+                        ggml_backend_type backend_output;
+
+                        if (n_gpu_layers > int(n_layer)) {
+                            backend_norm   = llama_backend_offload;
+                            backend_output = llama_backend_offload;
+                        } else {
+                            backend_norm   = GGML_BACKEND_CPU;
+                            backend_output = GGML_BACKEND_CPU;
+                        }
+
+                        model.output_norm   = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd},          backend_norm);
+                        model.output_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT_NORM, "bias"),   {n_embd},          backend_norm);
+                        model.output        = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, backend_output);
+                        model.output_b      = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT,      "bias"),   {n_vocab},         backend_output);
+
+                        if (backend_norm == GGML_BACKEND_GPU) {
+                            vram_weights += ggml_nbytes(model.output_norm);
+                            vram_weights += ggml_nbytes(model.output_norm_b);
+                            vram_weights += ggml_nbytes(model.output);
+                            vram_weights += ggml_nbytes(model.output_b);
+                        }
+                    }
+
+                    const uint32_t n_ff = hparams.n_ff;
+
+                    const int i_gpu_start = n_layer - n_gpu_layers;
+
+                    model.layers.resize(n_layer);
+
+                    for (uint32_t i = 0; i < n_layer; ++i) {
+                        const ggml_backend_type backend       = int(i) < i_gpu_start ? GGML_BACKEND_CPU : llama_backend_offload; // NOLINT
+                        const ggml_backend_type backend_split = int(i) < i_gpu_start ? GGML_BACKEND_CPU : llama_backend_offload_split; // NOLINT
+
+                        auto & layer = model.layers[i];
+
+                        layer.attn_norm   = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, backend);
+                        layer.attn_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_NORM, "bias", i),   {n_embd}, backend);
+
+                        layer.wqkv = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}, backend_split);
+                        layer.bqkv = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_QKV, "bias", i),   {n_embd + 2*n_embd_gqa},         backend);
+
+                        layer.wo   = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, backend_split);
+                        layer.bo   = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_OUT, "bias", i),   {n_embd},         backend);
+
+                        layer.ffn_down   = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, backend_split);
+                        layer.ffn_down_b = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "bias", i),   {n_embd},       backend);
+
+                        layer.ffn_up   = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, backend_split);
+                        layer.ffn_up_b = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP, "bias", i),   {n_ff},         backend);
+
+                        if (backend == GGML_BACKEND_GPU) {
+                            vram_weights +=
+                                ggml_nbytes(layer.attn_norm) + ggml_nbytes(layer.attn_norm_b) +
+                                ggml_nbytes(layer.wqkv)      + ggml_nbytes(layer.bqkv)        +
+                                ggml_nbytes(layer.wo)        + ggml_nbytes(layer.bo)          +
+                                ggml_nbytes(layer.ffn_up)    + ggml_nbytes(layer.ffn_up_b)    +
+                                ggml_nbytes(layer.ffn_down)  + ggml_nbytes(layer.ffn_down_b);
+                        }
+                    }
+                } break;
            default:
                throw std::runtime_error("unknown architecture");
        }
@@ -3822,8 +3918,8 @@ static void llm_build_k_shift(
            ggml_rope_custom_inplace(ctx,
                    ggml_view_3d(ctx, kv.k_l[il],
                        n_embd_head, n_head_kv, n_ctx,
-                        ggml_type_sizef(kv.k_l[il]->type)*n_embd_head,
-                        ggml_type_sizef(kv.k_l[il]->type)*n_embd_gqa,
+                        ggml_row_size(kv.k_l[il]->type, n_embd_head),
+                        ggml_row_size(kv.k_l[il]->type, n_embd_gqa),
                        0),
                    K_shift, n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
                    ext_factor, attn_factor, beta_fast, beta_slow);
@@ -3852,7 +3948,7 @@ static void llm_build_kv_store(
    cb(v_cur_t, "v_cur_t", il);

    struct ggml_tensor * k_cache_view = ggml_view_1d(ctx, kv.k_l[il], n_tokens*n_embd_gqa,
-            (ggml_type_sizef(kv.k_l[il]->type)*n_embd_gqa)*kv_head);
+            (ggml_row_size(kv.k_l[il]->type, n_embd_gqa))*kv_head);
    cb(k_cache_view, "k_cache_view", il);

    struct ggml_tensor * v_cache_view = ggml_view_2d(ctx, kv.v_l[il], n_tokens, n_embd_gqa,
@@ -3986,6 +4082,7 @@ static struct ggml_tensor * llm_build_ffn(
 // if max_alibi_bias > 0 then apply ALiBi
 static struct ggml_tensor * llm_build_kqv(
        struct ggml_context * ctx,
+          const llama_model & model,
        const llama_hparams & hparams,
       const llama_kv_cache & kv,
         struct ggml_tensor * wo,
@@ -3997,6 +4094,7 @@ static struct ggml_tensor * llm_build_kqv(
                    int32_t   n_tokens,
                    int32_t   n_kv,
                    float     max_alibi_bias,
+                    float     scale,
         const llm_build_cb & cb,
                    int       il) {
    const int64_t n_embd      = hparams.n_embd;
@@ -4011,14 +4109,20 @@ static struct ggml_tensor * llm_build_kqv(
    struct ggml_tensor * k =
        ggml_view_3d(ctx, kv.k_l[il],
                n_embd_head, n_kv, n_head_kv,
-                ggml_type_sizef(kv.k_l[il]->type)*n_embd_gqa,
-                ggml_type_sizef(kv.k_l[il]->type)*n_embd_head,
+                ggml_row_size(kv.k_l[il]->type, n_embd_gqa),
+                ggml_row_size(kv.k_l[il]->type, n_embd_head),
                0);
    cb(k, "k", il);

    struct ggml_tensor * kq = ggml_mul_mat(ctx, k, q);
    cb(kq, "kq", il);

+    if (model.arch == LLM_ARCH_PHI2) {
+        // for this arch, we need to perform the KQ multiplication with F32 precision, otherwise we get NaNs
+        // ref: https://github.com/ggerganov/llama.cpp/pull/4490#issuecomment-1859055847
+        ggml_mul_mat_set_prec(kq, GGML_PREC_F32);
+    }
+
    if (max_alibi_bias > 0.0f) {
        // temporary branch until we figure out how to handle ggml_alibi through ggml_add
        kq = ggml_scale(ctx, kq, kq_scale);
@@ -4038,7 +4142,7 @@ static struct ggml_tensor * llm_build_kqv(
        kq = ggml_soft_max(ctx, kq);
        cb(kq, "kq_soft_max", il);
    } else {
-        kq = ggml_soft_max_ext(ctx, kq, kq_mask, 1.0f/sqrtf(float(n_embd_head)));
+        kq = ggml_soft_max_ext(ctx, kq, kq_mask, scale);
        cb(kq, "kq_soft_max_ext", il);
    }

@@ -4245,9 +4349,9 @@ struct llm_build_context {

                llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il);

-                cur = llm_build_kqv(ctx0, hparams, kv_self,
+                cur = llm_build_kqv(ctx0, model, hparams, kv_self,
                        model.layers[il].wo, model.layers[il].bo,
-                        Qcur, KQ_scale, KQ_mask, n_ctx, n_tokens, n_kv, -1.0f, cb, il);
+                        Qcur, KQ_scale, KQ_mask, n_ctx, n_tokens, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il);
                cb(cur, "kqv_out", il);
            }

@@ -4428,9 +4532,9 @@ struct llm_build_context {
                // apply ALiBi for 13B model
                const float max_alibi_bias = model.type == MODEL_13B ? 8.0f : -1.0f;

-                cur = llm_build_kqv(ctx0, hparams, kv_self,
+                cur = llm_build_kqv(ctx0, model, hparams, kv_self,
                        model.layers[il].wo, NULL,
-                        Qcur, KQ_scale, KQ_mask, n_ctx, n_tokens, n_kv, max_alibi_bias, cb, il);
+                        Qcur, KQ_scale, KQ_mask, n_ctx, n_tokens, n_kv, max_alibi_bias, 1.0f/sqrtf(float(n_embd_head)), cb, il);
                cb(cur, "kqv_out", il);
            }

@@ -4552,9 +4656,9 @@ struct llm_build_context {

                llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il);

-                cur = llm_build_kqv(ctx0, hparams, kv_self,
+                cur = llm_build_kqv(ctx0, model, hparams, kv_self,
                        model.layers[il].wo, NULL,
-                        Qcur, KQ_scale, KQ_mask, n_ctx, n_tokens, n_kv, -1.0f, cb, il);
+                        Qcur, KQ_scale, KQ_mask, n_ctx, n_tokens, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il);
                cb(cur, "kqv_out", il);
            }

@@ -4652,9 +4756,9 @@ struct llm_build_context {

                llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il);

-                cur = llm_build_kqv(ctx0, hparams, kv_self,
+                cur = llm_build_kqv(ctx0, model, hparams, kv_self,
                        model.layers[il].wo, model.layers[il].bo,
-                        Qcur, KQ_scale, KQ_mask, n_ctx, n_tokens, n_kv, -1.0f, cb, il);
+                        Qcur, KQ_scale, KQ_mask, n_ctx, n_tokens, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il);
                cb(cur, "kqv_out", il);
            }

@@ -4861,9 +4965,9 @@ struct llm_build_context {
                llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il);

                // TODO: not tested, could be broken
-                cur = llm_build_kqv(ctx0, hparams, kv_self,
+                cur = llm_build_kqv(ctx0, model, hparams, kv_self,
                        model.layers[il].wo, model.layers[il].bo,
-                        Q, KQ_scale, KQ_mask, n_ctx, n_tokens, n_kv, -1.0f, cb, il);
+                        Q, KQ_scale, KQ_mask, n_ctx, n_tokens, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il);
                cb(cur, "kqv_out", il);
            }

@@ -4952,9 +5056,9 @@ struct llm_build_context {

                llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il);

-                cur = llm_build_kqv(ctx0, hparams, kv_self,
+                cur = llm_build_kqv(ctx0, model, hparams, kv_self,
                        model.layers[il].wo, NULL,
-                        Qcur, KQ_scale, KQ_mask, n_ctx, n_tokens, n_kv, 8.0f, cb, il);
+                        Qcur, KQ_scale, KQ_mask, n_ctx, n_tokens, n_kv, 8.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il);
                cb(cur, "kqv_out", il);
            }

@@ -5049,9 +5153,9 @@ struct llm_build_context {

                llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il);

-                cur = llm_build_kqv(ctx0, hparams, kv_self,
+                cur = llm_build_kqv(ctx0, model, hparams, kv_self,
                        model.layers[il].wo, model.layers[il].bo,
-                        Qcur, KQ_scale, KQ_mask, n_ctx, n_tokens, n_kv, 8.0f, cb, il);
+                        Qcur, KQ_scale, KQ_mask, n_ctx, n_tokens, n_kv, 8.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il);
                cb(cur, "kqv_out", il);
            }

@@ -5143,9 +5247,9 @@ struct llm_build_context {

                llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il);

-                cur = llm_build_kqv(ctx0, hparams, kv_self,
+                cur = llm_build_kqv(ctx0, model, hparams, kv_self,
                        model.layers[il].wo, NULL,
-                        Qcur, KQ_scale, KQ_mask, n_ctx, n_tokens, n_kv, hparams.f_max_alibi_bias, cb, il);
+                        Qcur, KQ_scale, KQ_mask, n_ctx, n_tokens, n_kv, hparams.f_max_alibi_bias, 1.0f/sqrtf(float(n_embd_head)), cb, il);
                cb(cur, "kqv_out", il);
            }

@@ -5256,9 +5360,9 @@ struct llm_build_context {

                llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il);

-                cur = llm_build_kqv(ctx0, hparams, kv_self,
+                cur = llm_build_kqv(ctx0, model, hparams, kv_self,
                        model.layers[il].wo, NULL,
-                        Qcur, KQ_scale, KQ_mask, n_ctx, n_tokens, n_kv, -1.0f, cb, il);
+                        Qcur, KQ_scale, KQ_mask, n_ctx, n_tokens, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il);
                cb(cur, "kqv_out", il);
            }

@@ -5315,15 +5419,15 @@ struct llm_build_context {
        cb(inpL, "inp_embd", -1);

        // inp_pos - contains the positions
-        struct ggml_tensor * inp_pos= ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens);
+        struct ggml_tensor * inp_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens);
        cb(inp_pos, "inp_pos", -1);

        // KQ_scale
-        struct ggml_tensor * KQ_scale= ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, 1);
+        struct ggml_tensor * KQ_scale = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, 1);
        cb(KQ_scale, "KQ_scale", -1);

        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
-        struct ggml_tensor * KQ_mask= ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_kv, n_tokens, 1);
+        struct ggml_tensor * KQ_mask = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_kv, n_tokens, 1);
        cb(KQ_mask, "KQ_mask", -1);

        // shift the entire K-cache if needed
@@ -5373,9 +5477,9 @@ struct llm_build_context {

                llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il);

-                cur = llm_build_kqv(ctx0, hparams, kv_self,
+                cur = llm_build_kqv(ctx0, model, hparams, kv_self,
                        model.layers[il].wo, NULL,
-                        Qcur, KQ_scale, KQ_mask, n_ctx, n_tokens, n_kv, -1.0f, cb, il);
+                        Qcur, KQ_scale, KQ_mask, n_ctx, n_tokens, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il);
                cb(cur, "kqv_out", il);
            }

@@ -5417,6 +5521,122 @@ struct llm_build_context {

        ggml_build_forward_expand(gf, cur);

+        return gf;
+    }
+    struct ggml_cgraph * build_phi2() {
+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
+
+        struct ggml_tensor * cur;
+        struct ggml_tensor * attn_norm_output;
+        struct ggml_tensor * ffn_output;
+        struct ggml_tensor * inpL;
+
+        inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, cb);
+        cb(inpL, "inp_embd", -1);
+
+        // inp_pos - contains the positions
+        struct ggml_tensor * inp_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens);
+        cb(inp_pos, "inp_pos", -1);
+
+        // Q_scale
+        struct ggml_tensor * Q_scale = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, 1);
+        cb(Q_scale, "Q_scale", -1);
+
+        // KQ_scale
+        struct ggml_tensor * KQ_scale = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, 1);
+        cb(KQ_scale, "KQ_scale", -1);
+
+        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
+        struct ggml_tensor * KQ_mask = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_kv, n_tokens, 1);
+        cb(KQ_mask, "KQ_mask", -1);
+
+        // shift the entire K-cache if needed
+        if (do_rope_shift) {
+            llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, LLM_ROPE_NEOX, n_ctx, n_embd_head, freq_base, freq_scale, cb);
+        }
+
+        for (int il = 0; il < n_layer; ++il) {
+            attn_norm_output = llm_build_norm(ctx0, inpL, hparams,
+                    model.layers[il].attn_norm,
+                    model.layers[il].attn_norm_b,
+                    LLM_NORM, cb, il);
+            cb(attn_norm_output, "attn_norm", il);
+
+            // self-attention
+            {
+                cur = ggml_mul_mat(ctx0, model.layers[il].wqkv, attn_norm_output);
+                cb(cur, "wqkv", il);
+
+                cur = ggml_add(ctx0, cur, model.layers[il].bqkv);
+                cb(cur, "bqkv", il);
+
+                struct ggml_tensor * Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd,     n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd)));
+                struct ggml_tensor * Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd)));
+                struct ggml_tensor * Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa)));
+
+                cb(Qcur, "Qcur", il);
+                cb(Kcur, "Kcur", il);
+                cb(Vcur, "Vcur", il);
+
+                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+
+                Qcur = ggml_rope_custom(
+                    ctx0, Qcur, inp_pos, hparams.n_rot, 2, 0, n_orig_ctx,
+                    freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow
+                );
+                cb(Qcur, "Qcur", il);
+
+                Qcur = ggml_scale(ctx0, Qcur, Q_scale);
+                cb(Qcur, "Qcur", il);
+
+                Kcur = ggml_rope_custom(
+                    ctx0, Kcur, inp_pos, hparams.n_rot, 2, 0, n_orig_ctx,
+                    freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow
+                );
+                cb(Kcur, "Kcur", il);
+
+                llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il);
+
+                cur = llm_build_kqv(ctx0, model, hparams, kv_self,
+                        model.layers[il].wo, model.layers[il].bo,
+                        Qcur, KQ_scale, KQ_mask, n_ctx, n_tokens, n_kv, -1.0f, 1.0f, cb, il);
+                cb(cur, "kqv_out", il);
+            }
+
+            // FF
+            {
+                ffn_output = llm_build_ffn(ctx0, attn_norm_output,
+                        model.layers[il].ffn_up,   model.layers[il].ffn_up_b,
+                        NULL,                      NULL,
+                        model.layers[il].ffn_down, model.layers[il].ffn_down_b,
+                        LLM_FFN_GELU, LLM_FFN_SEQ, cb, il);
+                cb(ffn_output, "ffn_out", il);
+            }
+
+            cur = ggml_add(ctx0, cur, ffn_output);
+            cb(cur, "l_out", il);
+
+            cur = ggml_add(ctx0, cur, inpL);
+            cb(cur, "l_out", il);
+
+            inpL = cur;
+        }
+
+        cur = llm_build_norm(ctx0, inpL, hparams,
+                model.output_norm,
+                model.output_norm_b,
+                LLM_NORM, cb, -1);
+        cb(cur, "result_norm", -1);
+
+        cur = ggml_mul_mat(ctx0, model.output, cur);
+        cb(cur, "result_output_no_bias", -1);
+
+        cur = ggml_add(ctx0, cur, model.output_b);
+        cb(cur, "result_output", -1);
+
+        ggml_build_forward_expand(gf, cur);
+
        return gf;
    }
 };
@@ -5432,7 +5652,7 @@ enum llm_offload_func_e {
    OFFLOAD_FUNC_FRC, // force offload
    OFFLOAD_FUNC_KQV,
    OFFLOAD_FUNC_NR,
-    OFFLOAD_FUNC_EMB,
+    OFFLOAD_FUNC_EMB, // embeddings
    OFFLOAD_FUNC_OUT,
 };

@@ -5517,6 +5737,7 @@ static const std::unordered_map<const char *, llm_offload_func_e> k_offload_map
    { "pos_embd",                   OFFLOAD_FUNC_NR  },

    { "inp_pos",                    OFFLOAD_FUNC_FRC }, // this is often used for KQ ops (e.g. rope)
+    { "Q_scale",                    OFFLOAD_FUNC_FRC },
    { "KQ_scale",                   OFFLOAD_FUNC_FRC },
    { "KQ_mask",                    OFFLOAD_FUNC_FRC },
    { "K_shift",                    OFFLOAD_FUNC_FRC },
@@ -5601,6 +5822,7 @@ static const std::unordered_map<const char *, llm_offload_func_e> k_offload_map
    { "l_out",                      OFFLOAD_FUNC     },

    { "result_norm",                OFFLOAD_FUNC_EMB },
+    { "result_output_no_bias",      OFFLOAD_FUNC_EMB },
    { "result_output",              OFFLOAD_FUNC_OUT },
 };

@@ -5618,6 +5840,7 @@ static struct ggml_cgraph * llama_build_graph(
    bool alloc_inp_tokens   = false;
    bool alloc_inp_embd     = false;
    bool alloc_inp_pos      = false;
+    bool alloc_inp_Q_scale  = false;
    bool alloc_inp_KQ_scale = false;
    bool alloc_inp_KQ_mask  = false;
    bool alloc_inp_K_shift  = false;
@@ -5685,7 +5908,7 @@ static struct ggml_cgraph * llama_build_graph(
            alloc_inp_pos = true;
        }

-        if (!alloc_inp_KQ_scale && strcmp(name, "KQ_scale") == 0) {
+        if (!alloc_inp_Q_scale && strcmp(name, "Q_scale") == 0) {
            ggml_allocr_alloc(lctx.alloc, cur);

            if (!ggml_allocr_is_measure(lctx.alloc)) {
@@ -5693,6 +5916,23 @@ static struct ggml_cgraph * llama_build_graph(
                ggml_set_f32(cur, 1.0f/sqrtf(float(n_embd_head)));
            }

+            alloc_inp_Q_scale = true;
+        }
+
+        if (!alloc_inp_KQ_scale && strcmp(name, "KQ_scale") == 0) {
+            ggml_allocr_alloc(lctx.alloc, cur);
+
+            if (!ggml_allocr_is_measure(lctx.alloc)) {
+                const int64_t n_embd_head = model.hparams.n_embd_head();
+                if (model.arch == LLM_ARCH_PHI2) {
+                    // with phi2, we scale the Q to avoid precision issues
+                    // ref: https://github.com/ml-explore/mlx-examples/blob/08e862336ade809bc37d1035f94b359e7d1a5152/phi2/phi2.py#L64-L66
+                    ggml_set_f32(cur, 1.0f);
+                } else {
+                    ggml_set_f32(cur, 1.0f/sqrtf(float(n_embd_head)));
+                }
+            }
+
            alloc_inp_KQ_scale = true;
        }

@@ -5917,6 +6157,10 @@ static struct ggml_cgraph * llama_build_graph(
            {
                result = llm.build_qwen();
            } break;
+        case LLM_ARCH_PHI2:
+            {
+                result = llm.build_phi2();
+            } break;
        default:
            GGML_ASSERT(false);
    }
@@ -6050,12 +6294,16 @@ static int llama_decode_internal(

    ggml_allocr_alloc_graph(lctx.alloc, gf);

-    struct ggml_tensor * res        = gf->nodes[gf->n_nodes - 1];
+    // the output is always the last tensor in the graph
+    struct ggml_tensor * res = gf->nodes[gf->n_nodes - 1];
+    GGML_ASSERT(strcmp(res->name, "result_output") == 0);
+
+    // the embeddings could be the second to last tensor, or the third to last tensor
    struct ggml_tensor * embeddings = gf->nodes[gf->n_nodes - 2];
-
-    GGML_ASSERT(strcmp(res->name,        "result_output") == 0);
-    GGML_ASSERT(strcmp(embeddings->name, "result_norm")   == 0);
-
+    if (strcmp(embeddings->name, "result_norm") != 0) {
+        embeddings = gf->nodes[gf->n_nodes - 3];
+        GGML_ASSERT(strcmp(embeddings->name, "result_norm") == 0);
+    }

 #ifdef GGML_USE_CUBLAS
    for (int i = 0; i < gf->n_leafs; i++) {
@@ -6150,6 +6398,14 @@ static int llama_decode_internal(
    {
        auto & logits_out = lctx.logits;

+#ifndef NDEBUG
+        auto & logits_valid = lctx.logits_valid;
+        logits_valid.clear();
+        logits_valid.resize(n_tokens);
+
+        logits_out.clear();
+#endif
+
        if (batch.logits) {
            logits_out.resize(n_vocab * n_tokens);
            for (uint32_t i = 0; i < n_tokens; i++) {
@@ -6157,13 +6413,22 @@ static int llama_decode_internal(
                    continue;
                }
                memcpy(logits_out.data() + (n_vocab*i), (float *) ggml_get_data(res) + (n_vocab*i), sizeof(float)*n_vocab);
+#ifndef NDEBUG
+                logits_valid[i] = true;
+#endif
            }
        } else if (lctx.logits_all) {
            logits_out.resize(n_vocab * n_tokens);
            memcpy(logits_out.data(), (float *) ggml_get_data(res), sizeof(float)*n_vocab*n_tokens);
+#ifndef NDEBUG
+            std::fill(logits_valid.begin(), logits_valid.end(), true);
+#endif
        } else {
            logits_out.resize(n_vocab);
            memcpy(logits_out.data(), (float *) ggml_get_data(res) + (n_vocab*(n_tokens - 1)), sizeof(float)*n_vocab);
+#ifndef NDEBUG
+            logits_valid[0] = true;
+#endif
        }
    }

@@ -8471,7 +8736,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
        bool quantize = name.rfind("weight") == name.size() - 6; // ends with 'weight'?

        // quantize only 2D tensors
-        quantize &= (tensor->n_dims == 2);
+        quantize &= (ggml_n_dims(tensor) == 2);
        quantize &= params->quantize_output_tensor || name != "output.weight";
        quantize &= !params->only_copy;

@@ -8626,53 +8891,60 @@ static int llama_apply_lora_from_file_internal(

    const int64_t t_start_lora_us = ggml_time_us();

-    auto fin = std::ifstream(path_lora, std::ios::binary);
-    if (!fin) {
-        LLAMA_LOG_ERROR("%s: failed to open '%s'\n", __func__, path_lora);
-        return 1;
-    }
+    llama_file fin(path_lora, "rb");

    // verify magic and version
    {
-        uint32_t magic;
-        fin.read((char *) &magic, sizeof(magic));
-        uint32_t format_version;
-        fin.read((char *) &format_version, sizeof(format_version));
+        uint32_t magic = fin.read_u32();
+        if (magic != LLAMA_FILE_MAGIC_GGLA) {
+            LLAMA_LOG_ERROR("%s: bad file magic\n", __func__);
+            return 1;
+        }

+        uint32_t format_version = fin.read_u32();
        if (format_version != 1) {
            LLAMA_LOG_ERROR("%s: unsupported file version\n", __func__ );
            return 1;
        }
    }

-    int32_t lora_r;
-    int32_t lora_alpha;
-    fin.read((char *) &lora_r, sizeof(lora_r));
-    fin.read((char *) &lora_alpha, sizeof(lora_alpha));
+    int32_t lora_r = fin.read_u32();
+    int32_t lora_alpha = fin.read_u32();
    float scaling = scale * (float)lora_alpha / (float)lora_r;

    LLAMA_LOG_INFO("%s: r = %d, alpha = %d, scaling = %.2f\n", __func__, lora_r, lora_alpha, scaling);

+    // create a name -> tensor map of the model to accelerate lookups
+    // find the max tensor size to estimate the required temporary buffer size
+    size_t max_tensor_size = 0;
+    std::unordered_map<std::string, struct ggml_tensor*> model_tensors;
+    for (const auto & kv : model.tensors_by_name) {
+        model_tensors.insert(kv);
+        size_t f32_size = ggml_nelements(kv.second) * sizeof(float);
+        max_tensor_size = std::max(max_tensor_size, f32_size);
+    }
+
    // create a temporary ggml context to store the lora tensors
-    // todo: calculate size from biggest possible tensor
-    std::vector<uint8_t> lora_buf(1024ull * 1024ull * 1024ull);
+    // TODO: use ggml-alloc
+    size_t lora_ctx_size = max_tensor_size * 3;
+    LLAMA_LOG_INFO("%s: allocating %.f MB for lora temporary buffer\n", __func__, lora_ctx_size / 1024.0 / 1024.0);
+    std::vector<uint8_t> lora_buf(lora_ctx_size);
+
    struct ggml_init_params params;
    params.mem_size   = lora_buf.size();
    params.mem_buffer = lora_buf.data();
    params.no_alloc   = false;

-    ggml_context * lora_ctx = ggml_init(params);
-    std::unordered_map<std::string, struct ggml_tensor *> lora_tensors;
+    using unique_context = std::unique_ptr<ggml_context, decltype(&ggml_free)>;

-    // create a name -> tensor map of the model to accelerate lookups
-    std::unordered_map<std::string, struct ggml_tensor*> model_tensors;
-    for (const auto & kv : model.tensors_by_name) {
-        model_tensors.insert(kv);
-    }
+    unique_context lora_ctx(nullptr, ggml_free);
+    lora_ctx.reset(ggml_init(params));
+    std::unordered_map<std::string, struct ggml_tensor *> lora_tensors;

    // load base model
    std::unique_ptr<llama_model_loader> ml;
-    ggml_context * base_ctx = NULL;
+
+    unique_context base_ctx(nullptr, ggml_free);
    std::vector<uint8_t> base_buf;
    if (path_base_model) {
        LLAMA_LOG_INFO("%s: loading base model from '%s'\n", __func__, path_base_model);
@@ -8681,6 +8953,7 @@ static int llama_apply_lora_from_file_internal(
        size_t ctx_size;
        size_t mmapped_size;
        ml->calc_sizes(ctx_size, mmapped_size);
+
        base_buf.resize(ctx_size);

        ggml_init_params base_params;
@@ -8688,9 +8961,9 @@ static int llama_apply_lora_from_file_internal(
        base_params.mem_buffer = base_buf.data();
        base_params.no_alloc   = ml->use_mmap;

-        base_ctx = ggml_init(base_params);
+        base_ctx.reset(ggml_init(base_params));

-        // maybe this should in llama_model_loader
+        // maybe this should be in llama_model_loader
        if (ml->use_mmap) {
            ml->mapping.reset(new llama_mmap(&ml->file, /* prefetch */ 0, ggml_is_numa()));
        }
@@ -8703,27 +8976,35 @@ static int llama_apply_lora_from_file_internal(
    std::vector<uint8_t> work_buffer;

    while (true) {
+        if (fin.tell() == fin.size) {
+            // eof
+            break;
+        }
+
        int32_t n_dims;
-        int32_t length;
+        int32_t name_len;
        int32_t ftype;

-        fin.read(reinterpret_cast<char *>(&n_dims), sizeof(n_dims));
-        fin.read(reinterpret_cast<char *>(&length), sizeof(length));
-        fin.read(reinterpret_cast<char *>(&ftype),  sizeof(ftype));
-        if (fin.eof()) {
-            break;
+        fin.read_raw(&n_dims, sizeof(n_dims));
+        fin.read_raw(&name_len, sizeof(name_len));
+        fin.read_raw(&ftype,  sizeof(ftype));
+
+        if (n_dims != 1 && n_dims != 2) {
+            LLAMA_LOG_ERROR("%s: unsupported tensor dimension %d\n", __func__, n_dims);
+            return 1;
        }

        int32_t ne[2] = { 1, 1 };
        for (int i = 0; i < n_dims; ++i) {
-            fin.read(reinterpret_cast<char *>(&ne[i]), sizeof(ne[i]));
+            fin.read_raw(&ne[i], sizeof(ne[i]));
        }

        std::string name;
        {
+            GGML_ASSERT(name_len <= 1024);
            char buf[1024];
-            fin.read(buf, length);
-            name = std::string(buf, length);
+            fin.read_raw(buf, name_len);
+            name = std::string(buf, name_len);
        }

        // check for lora suffix and get the type of tensor
@@ -8737,7 +9018,7 @@ static int llama_apply_lora_from_file_internal(
        std::string lora_type = name.substr(pos + lora_suffix.length());
        std::string base_name = name;
        base_name.erase(pos);
-        // LLAMA_LOG_INFO("%s: %s => %s (lora type %s) \n", __func__, name.c_str(),base_name.c_str(), lora_type.c_str());
+        // LLAMA_LOG_INFO("%s: %s => %s (lora type %s) \n", __func__, name.c_str(), base_name.c_str(), lora_type.c_str());

        if (model_tensors.find(base_name) == model_tensors.end()) {
            LLAMA_LOG_ERROR("%s: unknown tensor '%s' in lora adapter\n", __func__, name.data());
@@ -8756,22 +9037,15 @@ static int llama_apply_lora_from_file_internal(
                        return false;
                    }
        }
-        ggml_tensor * lora_tensor;
-        if (n_dims == 2) {
-            lora_tensor = ggml_new_tensor_2d(lora_ctx, wtype, ne[0], ne[1]);
-        }
-        else {
-            LLAMA_LOG_ERROR("%s: unsupported tensor dimension %d\n", __func__, n_dims);
-            return 1;
-        }
-        ggml_set_name(lora_tensor, "lora_tensor");
+        ggml_tensor * lora_tensor = ggml_new_tensor_2d(lora_ctx.get(), wtype, ne[0], ne[1]);
+        ggml_set_name(lora_tensor, name.c_str());

        // load tensor data
-        size_t offset = fin.tellg();
+        size_t offset = fin.tell();
        size_t tensor_data_size = ggml_nbytes(lora_tensor);
        offset = (offset + 31) & -32;
-        fin.seekg(offset);
-        fin.read((char*)lora_tensor->data, tensor_data_size);
+        fin.seek(offset, SEEK_SET);
+        fin.read_raw(lora_tensor->data, tensor_data_size);

        lora_tensors[name] = lora_tensor;

@@ -8801,13 +9075,11 @@ static int llama_apply_lora_from_file_internal(

                // load from base model
                if (gguf_find_tensor(ctx_gguf, base_name.c_str()) < 0) {
-                    // TODO: throw
                    LLAMA_LOG_ERROR("%s: error: tensor '%s' not found in base model\n", __func__, base_name.c_str());
                    return 1;
                }

-                // TODO: not tested!! maybe not working!
-                base_t = ml->create_tensor(base_ctx, base_name, { (uint32_t)dest_t->ne[0], (uint32_t)dest_t->ne[1] }, GGML_BACKEND_CPU);
+                base_t = ml->create_tensor(base_ctx.get(), base_name, { dest_t->ne[0], dest_t->ne[1] }, GGML_BACKEND_CPU);
                ml->load_data_for(base_t);
            } else {
                base_t = dest_t;
@@ -8836,43 +9108,45 @@ static int llama_apply_lora_from_file_internal(
            }

            // w = w + BA*s
-            ggml_tensor * BA = ggml_mul_mat(lora_ctx, loraA, loraB);
+            ggml_tensor * BA = ggml_mul_mat(lora_ctx.get(), loraA, loraB);
            offload_func(BA);
            ggml_set_name(BA, "BA");

            if (scaling != 1.0f) {
-                ggml_tensor * scale_tensor = ggml_new_f32(lora_ctx, scaling);
+                ggml_tensor * scale_tensor = ggml_new_f32(lora_ctx.get(), scaling);
                ggml_set_name(scale_tensor, "scale_tensor");

-                BA = ggml_scale_inplace(lora_ctx, BA, scale_tensor);
+                BA = ggml_scale_inplace(lora_ctx.get(), BA, scale_tensor);
                offload_func(BA);
                ggml_set_name(BA, "BA_scaled");
            }

            ggml_tensor * r;
            if (base_t == dest_t) {
-                r = ggml_add_inplace(lora_ctx, dest_t, BA);
+                r = ggml_add_inplace(lora_ctx.get(), dest_t, BA);
                offload_func_force_inplace(r);
                ggml_set_name(r, "r_add_inplace");
            }
            else {
-                r = ggml_add(lora_ctx, base_t, BA);
+                r = ggml_add(lora_ctx.get(), base_t, BA);
                offload_func(r);
                ggml_set_name(r, "r_add");

-                r = ggml_cpy(lora_ctx, r, dest_t);
+                r = ggml_cpy(lora_ctx.get(), r, dest_t);
                offload_func(r);
                ggml_set_name(r, "r_cpy");
            }

-            struct ggml_cgraph * gf = ggml_new_graph(lora_ctx);
+            struct ggml_cgraph * gf = ggml_new_graph(lora_ctx.get());
            ggml_build_forward_expand(gf, r);

            ggml_graph_compute_helper(work_buffer, gf, n_threads);

+            // the tensors in the adapter must be sorted such that loraA and loraB of the same tensor are next to each other
+            GGML_ASSERT(lora_tensors.size() == 2);
+
            // we won't need these tensors again, reset the context to save memory
-            ggml_free(lora_ctx);
-            lora_ctx = ggml_init(params);
+            lora_ctx.reset(ggml_init(params));
            lora_tensors.clear();

            n_tensors++;
@@ -8882,12 +9156,6 @@ static int llama_apply_lora_from_file_internal(
        }
    }

-    // TODO: this should be in a destructor, it will leak on failure
-    ggml_free(lora_ctx);
-    if (base_ctx) {
-        ggml_free(base_ctx);
-    }
-
    const int64_t t_lora_us = ggml_time_us() - t_start_lora_us;
    LLAMA_LOG_INFO(" done (%.2f ms)\n", t_lora_us / 1000.0);

@@ -9264,10 +9532,14 @@ const llama_model * llama_get_model(const struct llama_context * ctx) {
    return &ctx->model;
 }

-int llama_n_ctx(const struct llama_context * ctx) {
+uint32_t llama_n_ctx(const struct llama_context * ctx) {
    return ctx->cparams.n_ctx;
 }

+uint32_t llama_n_batch(const struct llama_context * ctx) {
+    return ctx->cparams.n_batch;
+}
+
 enum llama_vocab_type llama_vocab_type(const struct llama_model * model) {
    return model->vocab.type;
 }
@@ -10052,6 +10324,7 @@ float * llama_get_logits(struct llama_context * ctx) {
 }

 float * llama_get_logits_ith(struct llama_context * ctx, int32_t i) {
+    assert(ctx->logits_valid.at(i));
    return ctx->logits.data() + i*ctx->model.hparams.n_vocab;
 }

@@ -39,6 +39,7 @@

 #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
@@ -313,7 +314,9 @@ extern "C" {

    LLAMA_API const struct llama_model * llama_get_model(const struct llama_context * ctx);

-    LLAMA_API int llama_n_ctx      (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 enum llama_vocab_type llama_vocab_type(const struct llama_model * model);

@@ -1,3 +1,5 @@
 numpy==1.24.4
 sentencepiece==0.1.98
+transformers>=4.34.0
 gguf>=0.1.0
+protobuf>=4.21.0
@@ -0,0 +1,38 @@
+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
@@ -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_type_size(tensor->type)*size/ggml_blck_size(tensor->type));
+        std::vector<uint8_t> dataq(ggml_row_size(tensor->type, size));
        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());
@@ -72,6 +72,8 @@ 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++) {
@@ -85,9 +87,8 @@ 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 (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));
+                    } else if (quantized) {
+                        tt.to_float(&buf[i], vq.data(), bs);
                        tv.insert(tv.end(), vq.begin(), vq.end());
                    } else {
                        GGML_ASSERT(false);
@@ -234,6 +235,11 @@ 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() {}

@@ -268,7 +274,58 @@ 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,
@@ -276,6 +333,11 @@ 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) {
@@ -296,13 +358,20 @@ 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);

@@ -318,9 +387,24 @@ 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
@@ -349,9 +433,10 @@ 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("(%f, %f) ", f1[i], f2[i]);
+                //    printf("%5d %9.6f %9.6f, diff = %9.6f\n", i, f1[i], f2[i], f1[i] - f2[i]);
                //}
                //printf("\n");
+                //exit(1);
                ud->ok = false;
            }
            return true;
@@ -375,6 +460,8 @@ 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 = {
@@ -1135,6 +1222,118 @@ 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;
@@ -1219,11 +1418,6 @@ 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;

@@ -1361,6 +1555,7 @@ 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());
@@ -1372,12 +1567,16 @@ 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(GGML_TYPE_F32, {10, 10, 10, 10}));
-    test_cases.emplace_back(new test_sum_rows(GGML_TYPE_F32, {2, 1, 1, 1}));
+    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());

 #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, 14336));
+    test_cases.emplace_back(new test_moe(8, 2, 1, 4096, 8*1024));
    //test_cases.emplace_back(new test_moe(8, 2, 8, 4096, 14336));
 #endif

@@ -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 = size / ggml_blck_size(type) * ggml_type_size(type);
+                    size_t quantized_size = ggml_row_size(type, size);
                    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 = size / ggml_blck_size(type) * ggml_type_size(type);
+                    size_t quantized_size = ggml_row_size(type, size);
                    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 = size / ggml_blck_size(type) * ggml_type_size(type);
+                    size_t quantized_size = ggml_row_size(type, size);
                    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 = size / ggml_blck_size(type) * ggml_type_size(type);
+                    size_t quantized_size = ggml_row_size(type, size);
                    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 = size / ggml_blck_size(type) * ggml_type_size(type);
+                    size_t quantized_size = ggml_row_size(type, size);
                    benchmark_function(size, quantized_size, iterations, quantize_fn);
                }
                printf("\n");
Author	SHA1	Message	Date
Marcus Dunn	31f27758fa	llama : allow getting n_batch from llama_context in c api (#4540 ) * allowed getting n_batch from llama_context in c api * changed to use `uint32_t` instead of `int` * changed to use `uint32_t` instead of `int` in `llama_n_ctx` * Update llama.h --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>	2023-12-21 21:57:48 +02:00
Finn Voorhees	56fa50819f	metal : fix `ggml_metal_log` vargs (#4373 )	2023-12-21 21:55:02 +02:00
Erik Garrison	0f630fbc92	cuda : ROCm AMD Unified Memory Architecture (UMA) handling (#4449 ) * AMD ROCm: handle UMA memory VRAM expansions This resolves #2797 by allowing ROCm AMD GPU users with a UMA to dynamically expand the VRAM allocated to the GPU. Without this, AMD ROCm users with shared CPU/GPU memory usually are stuck with the BIOS-set (or fixed) framebuffer VRAM, making it impossible to load more than 1-2 layers. Note that the model is duplicated in RAM because it's loaded once for the CPU and then copied into a second set of allocations that are managed by the HIP UMA system. We can fix this later. * clarify build process for ROCm on linux with cmake * avoid using deprecated ROCm hipMallocHost * keep simplifying the change required for UMA * cmake: enable UMA-compatible allocation when LLAMA_HIP_UMA=ON	2023-12-21 21:45:32 +02:00
arlo-phoenix	562cf222b5	ggml-cuda: Fix HIP build by adding define for __trap (#4569 ) Regression of `1398823922` HIP doesn't have trap, only abort	2023-12-21 20:13:25 +01:00
Jared Van Bortel	8fe03ffdda	common : remove incorrect --model-draft default (#4568 )	2023-12-21 19:55:34 +02:00
Johannes Gäßler	9154494808	CUDA: mul_mat_id always on GPU for batches >= 32 (#4553 )	2023-12-21 18:42:59 +01:00
Georgi Gerganov	c083718c89	readme : update coding guidelines	2023-12-21 19:27:14 +02:00
howlger	880e352277	py : open merges file as 'utf-8' (#4566 ) Otherwise, on Windows converting bling-phi-2-v0 (<https://huggingface.co/llmware/bling-phi-2-v0>) via convert-hf-to-gguf.py will fail with the following error: ``` Traceback (most recent call last): File "C:\Users\User\git\gguf\convert-hf-to-gguf.py", line 1061, in <module> model_instance.set_vocab() File "C:\Users\User\git\gguf\convert-hf-to-gguf.py", line 52, in set_vocab self._set_vocab_gpt2() File "C:\Users\User\git\gguf\convert-hf-to-gguf.py", line 264, in _set_vocab_gpt2 special_vocab = gguf.SpecialVocab(dir_model, load_merges=True) File "C:\Users\User\git\gguf\gguf\vocab.py", line 33, in __init__ self._load(Path(path)) File "C:\Users\User\git\gguf\gguf\vocab.py", line 81, in _load self._try_load_merges_txt(path) File "C:\Users\User\git\gguf\gguf\vocab.py", line 95, in _try_load_merges_txt for line in fp: File "C:\Users\User\miniconda3\envs\gguf\lib\encodings\cp1252.py", line 23, in decode return codecs.charmap_decode(input,self.errors,decoding_table)[0] UnicodeDecodeError: 'charmap' codec can't decode byte 0x81 in position 1415: character maps to <undefined> ```	2023-12-21 19:07:34 +02:00
bobqianic	66f35a2f48	cuda : better error message for ggml_get_rows (#4561 ) * Update ggml-cuda.cu * Update ggml-cuda.cu * Update ggml-cuda.cu --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>	2023-12-21 19:06:44 +02:00
slaren	1398823922	cuda : replace asserts in wrong architecture checks with __trap (#4556 ) * cuda : replace asserts in wrong architecture checks with __trap * make bad_arch noreturn, remove returns	2023-12-21 18:02:30 +01:00
Johannes Gäßler	d3223afdad	llama : disable per-tensor info prints on model load (#4562 )	2023-12-21 18:34:17 +02:00
LoganDark	1d7a1912ce	Fix access violation in ggml_cuda_free_data if tensor->extra is NULL (#4554 )	2023-12-21 10:59:27 +01:00
Johannes Gäßler	799fc22689	CUDA: Faster Mixtral prompt processing (#4538 ) * CUDA: make MoE tensors contiguous for batch size>1 * Update ggml-cuda.cu Co-authored-by: slaren <slarengh@gmail.com> --------- Co-authored-by: slaren <slarengh@gmail.com>	2023-12-20 15:41:22 +01:00
Eric Sommerlade	328b83de23	ggml : fixed check for _MSC_VER (#4535 ) Co-authored-by: Eric Sommerlade <ersomme@microsoft.com>	2023-12-19 18:17:01 +02:00
arlo-phoenix	a7aee47b98	ggml-cuda: Fix HIP build (#4528 ) regression of #4490 Adds defines for two new datatypes cublasComputeType_t, cudaDataType_t. Currently using deprecated hipblasDatatype_t since newer ones very recent.	2023-12-18 22:33:45 +01:00
Georgi Gerganov	0e18b2e7d0	llama.swiftui : add tinyllama 1.1B F16	2023-12-18 20:17:43 +02:00
Georgi Gerganov	6ff39b129d	llama.swiftui : add more models	2023-12-18 20:05:12 +02:00
Ebey Abraham	b9e74f9bca	llama : add phi-2 + fix NeoX rope + ggml_mul_mat_set_prec (#4490 ) * phi2 implementation * fix breaking change * phi-2 : various fixes * phi-2 : use layer norm eps * py : whitespaces * llama : fix meta KV override bug * convert : phi don't add BOS token * convert : revert "added_tokens_decoder" change * phi-2 : scale Q instead of KQ for better precision * ggml : fix NeoX rope to rotate just first n_dims * cuda : less diff in the rope_neox kernel * ggml : add ggml_mul_mat_set_prec ggml-ci * Update ggml-cuda.cu Co-authored-by: slaren <slarengh@gmail.com> * Update ggml-cuda.cu Co-authored-by: slaren <slarengh@gmail.com> * cuda : ggml_cuda_op_mul_mat_cublas support F32 precision * cuda : remove oboslete comment --------- Co-authored-by: Ebey Abraham <ebeyabraham@microsoft.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: slaren <slarengh@gmail.com>	2023-12-18 19:27:47 +02:00
hankcs	3c04bf6da8	llama : fix try_override for bool_value which always return true (#4519 )	2023-12-18 15:14:58 +02:00
Jared Van Bortel	2994f0c5a2	decode : fix logits_valid for legacy API (#4516 )	2023-12-17 19:39:02 -05:00
Georgi Gerganov	b1306c4394	readme : update hot topics	2023-12-17 20:16:23 +02:00
Georgi Gerganov	800a489e4a	llama.swiftui : add bench functionality (#4483 ) * llama.swiftui : add bench button * llama.swiftui : initial bench functionality * force to use n_gpu_layers on simulator * add download buttons & expose llamaState.loadModel * update project.pbxproj * comment #Preview & fix editorconfig check * gitignore : xcode stuff * llama.swiftui : UX improvements * llama.swiftui : avoid data copy via "downloadTask" * llama.swiftui : remove model from project * llama : remove "mostly" from model infos * llama.swiftui : improve bench --------- Co-authored-by: jhen <developer@jhen.me>	2023-12-17 19:38:41 +02:00
Jared Van Bortel	f7f468a97d	gguf-py : fail fast on nonsensical special token IDs (#4489 )	2023-12-17 10:45:46 -05:00
Matheus Gabriel Alves Silva	919c40660f	build : Check the ROCm installation location (#4485 ) * build : Check the ROCm installation location * more generic approach * fixup! It was returning the path instead of the command output * fixup! Trailing whitespace	2023-12-17 17:23:33 +02:00
slaren	45668633fd	finetune : keep allocs alive until all allocations are done (#4486 )	2023-12-17 16:05:56 +01:00
olexiyb	0ffc92d2d2	server : disable llm logs if SERVER_VERBOSE is off (#3792 )	2023-12-17 17:02:16 +02:00
AdithyanI	8edd2b40fd	server : fix grammar being ignored (#4494 ) Fix bug in identifying the grammar.	2023-12-17 16:57:56 +02:00
Alexey Parfenov	eb16dae7e7	server : fix possible ambiguity in content type charset (#4501 )	2023-12-17 16:56:09 +02:00
mzcu	62bd52b7bf	server : allow requests larger than 8K (#4500 )	2023-12-17 16:54:37 +02:00
Bach Le	5daa5f54fd	Link to cublas dynamically on Windows even with LLAMA_STATIC (#4506 )	2023-12-17 11:57:33 +01:00
slaren	c6c4fc081c	lora : add support for non-llama models (#3333 ) * lora : add support for non-llama models ggml-ci * avoid leaking ggml_context on failure cleanup ggml-ci * lora : allow 1d tensors * lora : include embd and output layers in size calculation * fix style	2023-12-16 18:58:46 +01:00
Jared Van Bortel	8a5be3bd58	llama : sanity checks for access to logits (#4274 ) Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>	2023-12-15 22:16:15 -05:00
ShadovvBeast	88ae8952b6	server : add optional API Key Authentication example (#4441 ) * Add API key authentication for enhanced server-client security * server : to snake_case --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>	2023-12-15 13:49:01 +02:00
slaren	ee4725a686	ggml : group mul_mat_id rows by matrix (cpu only) (#4480 ) * ggml : group mul_mat_id rows by matrix (cpu only) * remove mmid parameters from mm forward * store row groups in wdata and calculate only once in GGML_TASK_INIT ggml-ci	2023-12-15 12:45:50 +01:00
slaren	6744dbe924	ggml : use ggml_row_size where possible (#4472 ) * ggml : use ggml_row_size where possible ggml-ci * ggml : move ggml_nbytes_split to ggml-cuda.cu	2023-12-14 20:05:21 +01:00
slaren	cafcd4f895	ggml : remove n_dims from ggml_tensor (#4469 ) ggml-ci	2023-12-14 16:52:08 +01:00
wonjun Jang	c50e400163	py : add protobuf dependency (#4466 )	2023-12-14 14:44:49 +02:00
LostRuins	20a68a7030	ggml : add ggml_row_size() (fixes llama out of space) (#4461 ) * Fixes "Not enough space in the context's memory pool" encountered on certain models, which seems to be caused by some imprecision related to the automatic casting of floating point values * do not cast to size_t, instead just use doubles * ggml : add ggml_row_size(), deprecate ggml_type_sizef() * ggml : fix row size compute to avoid overflows * tests : fix sizey -> sizez --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>	2023-12-14 14:13:33 +02:00
Georgi Gerganov	55e87c3749	ggml : fix OpenCL broadcast requirement for ggml_mul (close #4453 )	2023-12-14 10:35:29 +02:00
wonjun Jang	873637afc7	convert : support loading vocab from fast tokenizer config (#3633 ) * Add HFVocab into convert.py * Update convert.py * Update convert.py * add bytes_to_unicode function * change add_meta_vocab fucntion * remove debug code * remove byte_encoder * Add newline between classes * Check tokenizer.json when tokenizer.model is not exist. * Move transformers dependency to local code * Add error context with 'raise from' * Add fast tokenizer option to BpeVocab * Update convert.py * Add VocabLoader and remove Vocab class Add transformers dependency * remove added tokens and check newline token to decide spm or bpe * Update convert.py * Add special token type * Update convert.py * Update convert.py * Update convert.py * Fix typo in convert.py * Fix when params.n_vocab < tokenizer vocab size * update vocab class * change funtion name * Remove unused variable/functions, add types to class variable and methods, delete blank liens * fix flake8 warnings * code style cleanup * make mypy happy * change exception --------- Co-authored-by: Jared Van Bortel <jared@nomic.ai>	2023-12-14 10:09:34 +02:00
BarfingLemurs	0353a18401	readme : update supported model list (#4457 )	2023-12-14 09:38:49 +02:00
shibe2	948ff137ec	server : fix handling of characters that span multiple tokens when streaming (#4446 )	2023-12-13 21:57:15 +02:00
Georgi Gerganov	4d98d9a656	sync : ggml (SD ops, tests, kernels) (#4444 ) * sync : ggml (SD ops, tests, kernels) ggml-ci * cuda : restore im2col ggml-ci * metal : fix accuracy of dequantization kernels ggml-ci * cuda : restore correct im2col ggml-ci * metal : try to fix moe test by reducing expert size ggml-ci * cuda : fix bin bcast when src1 and dst have different types ggml-ci --------- Co-authored-by: slaren <slarengh@gmail.com>	2023-12-13 21:54:54 +02:00
Jared Van Bortel	70f806b821	build : detect host compiler and cuda compiler separately (#4414 )	2023-12-13 12:10:10 -05:00