diff --git a/.editorconfig b/.editorconfig
index f8245b85c..16d16b3b5 100644
--- a/.editorconfig
+++ b/.editorconfig
@@ -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
diff --git a/.github/workflows/build.yml b/.github/workflows/build.yml
index 22be233e6..a5090e398 100644
--- a/.github/workflows/build.yml
+++ b/.github/workflows/build.yml
@@ -143,6 +143,9 @@ jobs:
           cd build
           ctest --verbose
 
+  # TODO: build with LLAMA_NO_METAL because test-backend-ops fail on "Apple Paravirtual device" and I don't know
+  #       how to debug it.
+  #       ref: https://github.com/ggerganov/llama.cpp/actions/runs/7131777249/job/19420981052#step:5:1124
   macOS-latest-make:
     runs-on: macos-latest
 
@@ -160,14 +163,18 @@ jobs:
       - name: Build
         id: make_build
         run: |
-          make -j $(sysctl -n hw.logicalcpu)
+          LLAMA_NO_METAL=1 make -j $(sysctl -n hw.logicalcpu)
 
       - name: Test
         id: make_test
         run: |
-          make tests -j $(sysctl -n hw.logicalcpu)
-          make test -j $(sysctl -n hw.logicalcpu)
+          LLAMA_NO_METAL=1 make tests -j $(sysctl -n hw.logicalcpu)
+          LLAMA_NO_METAL=1 make test  -j $(sysctl -n hw.logicalcpu)
 
+  # TODO: build with LLAMA_METAL=OFF because test-backend-ops fail on "Apple Paravirtual device" and I don't know
+  #       how to debug it.
+  #       ref: https://github.com/ggerganov/llama.cpp/actions/runs/7132125951/job/19422043567?pr=4359#step:5:6584
+  #       would be great if we fix these
   macOS-latest-cmake:
     runs-on: macos-latest
 
@@ -188,7 +195,7 @@ jobs:
           sysctl -a
           mkdir build
           cd build
-          cmake ..
+          cmake -DLLAMA_METAL=OFF ..
           cmake --build . --config Release -j $(sysctl -n hw.logicalcpu)
 
       - name: Test
diff --git a/.github/workflows/docker.yml b/.github/workflows/docker.yml
index 9c90c77ac..7f4de50ea 100644
--- a/.github/workflows/docker.yml
+++ b/.github/workflows/docker.yml
@@ -52,6 +52,36 @@ jobs:
           username: ${{ github.repository_owner }}
           password: ${{ secrets.GITHUB_TOKEN }}
 
+      # https://github.com/jlumbroso/free-disk-space/tree/54081f138730dfa15788a46383842cd2f914a1be#example
+      - name: Free Disk Space (Ubuntu)
+        uses: jlumbroso/free-disk-space@main
+        with:
+          # this might remove tools that are actually needed,
+          # if set to "true" but frees about 6 GB
+          tool-cache: false
+
+          # all of these default to true, but feel free to set to
+          # "false" if necessary for your workflow
+          android: true
+          dotnet: true
+          haskell: true
+          large-packages: true
+          docker-images: true
+          swap-storage: true
+
+      - name: Determine tag name
+        id: tag
+        shell: bash
+        run: |
+          BUILD_NUMBER="$(git rev-list --count HEAD)"
+          SHORT_HASH="$(git rev-parse --short=7 HEAD)"
+          if [[ "${{ env.BRANCH_NAME }}" == "master" ]]; then
+            echo "name=b${BUILD_NUMBER}" >> $GITHUB_OUTPUT
+          else
+            SAFE_NAME=$(echo "${{ env.BRANCH_NAME }}" | tr '/' '-')
+            echo "name=${SAFE_NAME}-b${BUILD_NUMBER}-${SHORT_HASH}" >> $GITHUB_OUTPUT
+          fi
+
       - name: Build and push Docker image (versioned)
         if: github.event_name == 'push'
         uses: docker/build-push-action@v4
@@ -59,7 +89,7 @@ jobs:
           context: .
           push: true
           platforms: ${{ matrix.config.platforms }}
-          tags: "ghcr.io/ggerganov/llama.cpp:${{ matrix.config.tag }}-${{ env.COMMIT_SHA }}"
+          tags: "ghcr.io/${{ github.repository_owner }}/llama.cpp:${{ matrix.config.tag }}-${{ env.COMMIT_SHA }}"
           file: ${{ matrix.config.dockerfile }}
 
       - name: Build and push Docker image (tagged)
@@ -68,5 +98,5 @@ jobs:
           context: .
           push: ${{ github.event_name == 'push' }}
           platforms: ${{ matrix.config.platforms }}
-          tags: "ghcr.io/ggerganov/llama.cpp:${{ matrix.config.tag }}"
+          tags: "ghcr.io/${{ github.repository_owner }}/llama.cpp:${{ matrix.config.tag }}" , "ghcr.io/${{ github.repository_owner }}/llama.cpp:${{ matrix.config.tag }}-${{ steps.tag.outputs.name }}"
           file: ${{ matrix.config.dockerfile }}
diff --git a/.gitignore b/.gitignore
index 58c483994..def74a1e9 100644
--- a/.gitignore
+++ b/.gitignore
@@ -48,6 +48,7 @@ models-mnt
 /llama-bench
 /llava-cli
 /lookahead
+/lookup
 /main
 /metal
 /perplexity
@@ -101,3 +102,4 @@ poetry.toml
 /tests/test-tokenizer-1-llama
 /tests/test-tokenizer-1-bpe
 /tests/test-rope
+/tests/test-backend-ops
diff --git a/CMakeLists.txt b/CMakeLists.txt
index 0639518de..6fc6508c5 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -91,15 +91,16 @@ 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)
 option(LLAMA_MPI                             "llama: use MPI"                                   OFF)
 option(LLAMA_QKK_64                          "llama: use super-block size of 64 for k-quants"   OFF)
 
-option(LLAMA_BUILD_TESTS                "llama: build tests"    ${LLAMA_STANDALONE})
-option(LLAMA_BUILD_EXAMPLES             "llama: build examples" ${LLAMA_STANDALONE})
-option(LLAMA_BUILD_SERVER               "llama: build server example"                           ON)
+option(LLAMA_BUILD_TESTS                     "llama: build tests"    ${LLAMA_STANDALONE})
+option(LLAMA_BUILD_EXAMPLES                  "llama: build examples" ${LLAMA_STANDALONE})
+option(LLAMA_BUILD_SERVER                    "llama: build server example"                      ON)
 
 # Required for relocatable CMake package
 include(${CMAKE_CURRENT_SOURCE_DIR}/scripts/build-info.cmake)
@@ -291,7 +292,12 @@ if (LLAMA_CUBLAS)
         add_compile_definitions(GGML_CUDA_PEER_MAX_BATCH_SIZE=${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)
@@ -593,6 +648,11 @@ else()
     message(STATUS "Unknown architecture")
 endif()
 
+if (MINGW)
+    # Target Windows 8 for PrefetchVirtualMemory
+    add_compile_definitions(_WIN32_WINNT=0x602)
+endif()
+
 #
 # POSIX conformance
 #
@@ -662,11 +722,11 @@ add_library(ggml OBJECT
             ggml-backend.h
             ggml-quants.c
             ggml-quants.h
-            ${GGML_SOURCES_CUDA} ${GGML_HEADERS_CUDA}
+            ${GGML_SOURCES_CUDA}   ${GGML_HEADERS_CUDA}
             ${GGML_SOURCES_OPENCL} ${GGML_HEADERS_OPENCL}
-            ${GGML_SOURCES_METAL} ${GGML_HEADERS_METAL}
-            ${GGML_SOURCES_MPI} ${GGML_HEADERS_MPI}
-            ${GGML_SOURCES_EXTRA} ${GGML_HEADERS_EXTRA}
+            ${GGML_SOURCES_METAL}  ${GGML_HEADERS_METAL}
+            ${GGML_SOURCES_MPI}    ${GGML_HEADERS_MPI}
+            ${GGML_SOURCES_EXTRA}  ${GGML_HEADERS_EXTRA}
             )
 
 target_include_directories(ggml PUBLIC . ${LLAMA_EXTRA_INCLUDES})
diff --git a/Makefile b/Makefile
index 3cc932a2e..cb5a4e948 100644
--- a/Makefile
+++ b/Makefile
@@ -2,13 +2,14 @@
 BUILD_TARGETS = \
 	main quantize quantize-stats perplexity embedding vdot q8dot train-text-from-scratch convert-llama2c-to-ggml \
 	simple batched batched-bench save-load-state server gguf llama-bench libllava.a llava-cli baby-llama beam-search  \
-	speculative infill tokenize benchmark-matmult parallel finetune export-lora lookahead tests/test-c.o
+	speculative infill tokenize benchmark-matmult parallel finetune export-lora lookahead lookup tests/test-c.o
 
 # Binaries only useful for tests
 TEST_TARGETS = \
 	tests/test-llama-grammar tests/test-grammar-parser tests/test-double-float tests/test-grad0 tests/test-opt \
 	tests/test-quantize-fns tests/test-quantize-perf tests/test-sampling tests/test-tokenizer-0-llama          \
-	tests/test-tokenizer-0-falcon tests/test-tokenizer-1-llama tests/test-tokenizer-1-bpe tests/test-rope
+	tests/test-tokenizer-0-falcon tests/test-tokenizer-1-llama tests/test-tokenizer-1-bpe tests/test-rope      \
+	tests/test-backend-ops
 
 # Code coverage output files
 COV_TARGETS = *.gcno tests/*.gcno *.gcda tests/*.gcda *.gcov tests/*.gcov lcov-report gcovr-report
@@ -25,20 +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)
@@ -78,7 +65,7 @@ test: $(TEST_TARGETS)
 			./$$test_target; \
 		fi; \
 		if [ $$? -ne 0 ]; then \
-			printf 'Test $$test_target FAILED!\n\n' $$test_target; \
+			printf 'Test %s FAILED!\n\n' $$test_target; \
 			failures=$$(( failures + 1 )); \
 		else \
 			printf 'Test %s passed.\n\n' $$test_target; \
@@ -120,12 +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)
@@ -219,30 +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
@@ -293,8 +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
@@ -305,19 +268,31 @@ ifeq ($(UNAME_M),$(filter $(UNAME_M),x86_64 i686 amd64))
 	#MK_CXXFLAGS += -mssse3
 endif
 
-# The stack is only 16-byte aligned on Windows, so don't let gcc emit aligned moves.
-# https://gcc.gnu.org/bugzilla/show_bug.cgi?id=54412
-# https://github.com/ggerganov/llama.cpp/issues/2922
 ifneq '' '$(findstring mingw,$(shell $(CC) -dumpmachine))'
+	# The stack is only 16-byte aligned on Windows, so don't let gcc emit aligned moves.
+	# https://gcc.gnu.org/bugzilla/show_bug.cgi?id=54412
+	# https://github.com/ggerganov/llama.cpp/issues/2922
 	MK_CFLAGS   += -Xassembler -muse-unaligned-vector-move
 	MK_CXXFLAGS += -Xassembler -muse-unaligned-vector-move
+
+	# Target Windows 8 for PrefetchVirtualMemory
+	MK_CPPFLAGS += -D_WIN32_WINNT=0x602
 endif
 
 ifneq ($(filter aarch64%,$(UNAME_M)),)
 	# Apple M1, M2, etc.
 	# Raspberry Pi 3, 4, Zero 2 (64-bit)
+	# Nvidia Jetson
 	MK_CFLAGS   += -mcpu=native
 	MK_CXXFLAGS += -mcpu=native
+	JETSON_RELEASE_INFO = $(shell jetson_release)
+	ifdef JETSON_RELEASE_INFO
+		ifneq ($(filter TX2%,$(JETSON_RELEASE_INFO)),)
+			JETSON_EOL_MODULE_DETECT = 1
+			CC = aarch64-unknown-linux-gnu-gcc
+			cxx = aarch64-unknown-linux-gnu-g++
+		endif
+	endif
 endif
 
 ifneq ($(filter armv6%,$(UNAME_M)),)
@@ -391,64 +366,74 @@ ifdef LLAMA_BLIS
 endif # LLAMA_BLIS
 
 ifdef LLAMA_CUBLAS
-	MK_CPPFLAGS  += -DGGML_USE_CUBLAS -I/usr/local/cuda/include -I/opt/cuda/include -I$(CUDA_PATH)/targets/x86_64-linux/include
-	MK_LDFLAGS   += -lcublas -lculibos -lcudart -lcublasLt -lpthread -ldl -lrt -L/usr/local/cuda/lib64 -L/opt/cuda/lib64 -L$(CUDA_PATH)/targets/x86_64-linux/lib
+	MK_CPPFLAGS  += -DGGML_USE_CUBLAS -I/usr/local/cuda/include -I/opt/cuda/include -I$(CUDA_PATH)/targets/x86_64-linux/include -I/usr/local/cuda/targets/aarch64-linux/include
+	MK_LDFLAGS   += -lcublas -lculibos -lcudart -lcublasLt -lpthread -ldl -lrt -L/usr/local/cuda/lib64 -L/opt/cuda/lib64 -L$(CUDA_PATH)/targets/x86_64-linux/lib -L/usr/local/cuda/targets/aarch64-linux/lib
 	OBJS         += ggml-cuda.o
-	NVCCFLAGS = --forward-unknown-to-host-compiler -use_fast_math
+	MK_NVCCFLAGS  = -use_fast_math
+ifndef JETSON_EOL_MODULE_DETECT
+	MK_NVCCFLAGS += --forward-unknown-to-host-compiler
+endif # JETSON_EOL_MODULE_DETECT
+
+ifdef LLAMA_DEBUG
+	MK_NVCCFLAGS += -lineinfo
+endif
+
 ifdef LLAMA_CUDA_NVCC
 	NVCC = $(LLAMA_CUDA_NVCC)
 else
 	NVCC = nvcc
 endif #LLAMA_CUDA_NVCC
 ifdef CUDA_DOCKER_ARCH
-	NVCCFLAGS += -Wno-deprecated-gpu-targets -arch=$(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 $@
+ifdef JETSON_EOL_MODULE_DETECT
+	$(NVCC) -I. -Icommon -D_XOPEN_SOURCE=600 -D_GNU_SOURCE -DNDEBUG -DGGML_USE_CUBLAS -I/usr/local/cuda/include -I/opt/cuda/include -I/usr/local/cuda/targets/aarch64-linux/include -std=c++11 -O3 $(NVCCFLAGS) -Xcompiler "$(CUDA_CXXFLAGS)" -c $< -o $@
+else
+	$(NVCC) $(BASE_CXXFLAGS) $(NVCCFLAGS) -Wno-pedantic -Xcompiler "$(CUDA_CXXFLAGS)" -c $< -o $@
+endif # JETSON_EOL_MODULE_DETECT
 endif # LLAMA_CUBLAS
 
 ifdef LLAMA_CLBLAST
@@ -470,13 +455,22 @@ 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                   ?= $(ROCM_PATH)/bin/hipcc
 	LLAMA_CUDA_DMMV_X       ?= 32
 	LLAMA_CUDA_MMV_Y        ?= 1
 	LLAMA_CUDA_KQUANTS_ITER ?= 2
 	MK_CPPFLAGS += -DGGML_USE_HIPBLAS -DGGML_USE_CUBLAS
+ifdef LLAMA_HIP_UMA
+	MK_CPPFLAGS += -DGGML_HIP_UMA
+endif # LLAMA_HIP_UMA
 	MK_LDFLAGS  += -L$(ROCM_PATH)/lib -Wl,-rpath=$(ROCM_PATH)/lib
 	MK_LDFLAGS	+= -lhipblas -lamdhip64 -lrocblas
 	HIPFLAGS    += $(addprefix --offload-arch=,$(GPU_TARGETS))
@@ -510,16 +504,22 @@ 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
@@ -625,7 +625,7 @@ save-load-state: examples/save-load-state/save-load-state.cpp ggml.o llama.o $(C
 server: examples/server/server.cpp examples/server/httplib.h examples/server/json.hpp examples/server/index.html.hpp examples/server/index.js.hpp examples/server/completion.js.hpp examples/llava/clip.cpp examples/llava/clip.h common/stb_image.h ggml.o llama.o $(COMMON_DEPS) grammar-parser.o $(OBJS)
 	$(CXX) $(CXXFLAGS) -Iexamples/server $(filter-out %.h,$(filter-out %.hpp,$^)) -o $@ $(LDFLAGS) $(LWINSOCK2) -Wno-cast-qual
 
-gguf: examples/gguf/gguf.cpp ggml.o llama.o $(OBJS)
+gguf: examples/gguf/gguf.cpp ggml.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
 train-text-from-scratch: examples/train-text-from-scratch/train-text-from-scratch.cpp ggml.o llama.o $(COMMON_DEPS) train.o $(OBJS)
@@ -664,6 +664,9 @@ parallel: examples/parallel/parallel.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 lookahead: examples/lookahead/lookahead.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
+lookup: examples/lookup/lookup.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+
 ifdef LLAMA_METAL
 metal: examples/metal/metal.cpp ggml.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS)
@@ -729,16 +732,16 @@ tests/test-quantize-perf: tests/test-quantize-perf.cpp ggml.o $(OBJS)
 tests/test-sampling: tests/test-sampling.cpp ggml.o llama.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-tests/test-tokenizer-0-falcon: tests/test-tokenizer-0-falcon.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+tests/test-tokenizer-0-falcon: tests/test-tokenizer-0-falcon.cpp ggml.o llama.o $(COMMON_DEPS) console.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-tests/test-tokenizer-0-llama: tests/test-tokenizer-0-llama.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+tests/test-tokenizer-0-llama: tests/test-tokenizer-0-llama.cpp ggml.o llama.o $(COMMON_DEPS) console.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-tests/test-tokenizer-1-bpe: tests/test-tokenizer-1-bpe.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+tests/test-tokenizer-1-bpe: tests/test-tokenizer-1-bpe.cpp ggml.o llama.o $(COMMON_DEPS) console.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-tests/test-tokenizer-1-llama: tests/test-tokenizer-1-llama.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+tests/test-tokenizer-1-llama: tests/test-tokenizer-1-llama.cpp ggml.o llama.o $(COMMON_DEPS) console.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
 tests/test-rope: tests/test-rope.cpp ggml.o $(OBJS)
@@ -746,3 +749,6 @@ tests/test-rope: tests/test-rope.cpp ggml.o $(OBJS)
 
 tests/test-c.o: tests/test-c.c llama.h
 	$(CC) $(CFLAGS) -c $(filter-out %.h,$^) -o $@
+
+tests/test-backend-ops: tests/test-backend-ops.cpp ggml.o $(OBJS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
diff --git a/Package.swift b/Package.swift
index 5b3bd72ca..18d610d69 100644
--- a/Package.swift
+++ b/Package.swift
@@ -2,33 +2,14 @@
 
 import PackageDescription
 
-#if arch(arm) || arch(arm64)
-let platforms: [SupportedPlatform]? = [
-    .macOS(.v12),
-    .iOS(.v14),
-    .watchOS(.v4),
-    .tvOS(.v14)
-]
-let exclude: [String] = []
-let resources: [Resource] = [
-    .process("ggml-metal.metal")
-]
-let additionalSources: [String] = ["ggml-metal.m"]
-let additionalSettings: [CSetting] = [
-    .unsafeFlags(["-fno-objc-arc"]),
-    .define("GGML_USE_METAL")
-]
-#else
-let platforms: [SupportedPlatform]? = nil
-let exclude: [String] = ["ggml-metal.metal"]
-let resources: [Resource] = []
-let additionalSources: [String] = []
-let additionalSettings: [CSetting] = []
-#endif
-
 let package = Package(
     name: "llama",
-    platforms: platforms,
+    platforms: [
+        .macOS(.v12),
+        .iOS(.v14),
+        .watchOS(.v4),
+        .tvOS(.v14)
+    ],
     products: [
         .library(name: "llama", targets: ["llama"]),
     ],
@@ -36,25 +17,30 @@ let package = Package(
         .target(
             name: "llama",
             path: ".",
-            exclude: exclude,
+            exclude: [],
             sources: [
                 "ggml.c",
                 "llama.cpp",
                 "ggml-alloc.c",
                 "ggml-backend.c",
                 "ggml-quants.c",
-            ] + additionalSources,
-            resources: resources,
+                "ggml-metal.m",
+            ],
+            resources: [
+                .process("ggml-metal.metal")
+            ],
             publicHeadersPath: "spm-headers",
             cSettings: [
                 .unsafeFlags(["-Wno-shorten-64-to-32", "-O3", "-DNDEBUG"]),
-                .define("GGML_USE_ACCELERATE")
+                .define("GGML_USE_ACCELERATE"),
+                .unsafeFlags(["-fno-objc-arc"]),
+                .define("GGML_USE_METAL"),
                 // NOTE: NEW_LAPACK will required iOS version 16.4+
                 // We should consider add this in the future when we drop support for iOS 14
                 // (ref: ref: https://developer.apple.com/documentation/accelerate/1513264-cblas_sgemm?language=objc)
                 // .define("ACCELERATE_NEW_LAPACK"),
                 // .define("ACCELERATE_LAPACK_ILP64")
-            ] + additionalSettings,
+            ],
             linkerSettings: [
                 .linkedFramework("Accelerate")
             ]
diff --git a/README.md b/README.md
index dac971ae5..649c3b333 100644
--- a/README.md
+++ b/README.md
@@ -10,9 +10,11 @@ Inference of [LLaMA](https://arxiv.org/abs/2302.13971) model in pure C/C++
 
 ### Hot topics
 
-- Using `llama.cpp` with AWS instances: https://github.com/ggerganov/llama.cpp/discussions/4225
+- 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
 - 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
 
 ----
 
@@ -95,7 +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:**
@@ -110,6 +123,7 @@ as the main playground for developing new features for the [ggml](https://github
 - Clojure: [phronmophobic/llama.clj](https://github.com/phronmophobic/llama.clj)
 - React Native: [mybigday/llama.rn](https://github.com/mybigday/llama.rn)
 - Java: [kherud/java-llama.cpp](https://github.com/kherud/java-llama.cpp)
+- Zig: [deins/llama.cpp.zig](https://github.com/Deins/llama.cpp.zig)
 
 **UI:**
 
@@ -382,6 +396,9 @@ Building the program with BLAS support may lead to some performance improvements
 - #### cuBLAS
 
   This provides BLAS acceleration using the CUDA cores of your Nvidia GPU. Make sure to have the CUDA toolkit installed. You can download it from your Linux distro's package manager (e.g. `apt install nvidia-cuda-toolkit`) or from here: [CUDA Toolkit](https://developer.nvidia.com/cuda-downloads).
+
+  For Jetson user, if you have Jetson Orin, you can try this: [Offical Support](https://www.jetson-ai-lab.com/tutorial_text-generation.html). If you are using an old model(nano/TX2), need some additional operations before compiling.
+
   - Using `make`:
     ```bash
     make LLAMA_CUBLAS=1
@@ -419,14 +436,21 @@ 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 (but enables working with integrated GPUs).
+
+  - Using `make` (example for target gfx1030, build with 16 CPU threads):
+    ```bash
+    make -j16 LLAMA_HIPBLAS=1 LLAMA_HIP_UMA=1 AMDGPU_TARGETS=gxf1030
+    ```
+
+  - Using `CMake` for Windows (using x64 Native Tools Command Prompt for VS, and assuming a gfx1100-compatible AMD GPU):
     ```bash
     set PATH=%HIP_PATH%\bin;%PATH%
     mkdir build
@@ -435,10 +459,11 @@ 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 |
@@ -969,6 +994,8 @@ 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
 
diff --git a/common/common.cpp b/common/common.cpp
index 1dcc235ea..b3425ab09 100644
--- a/common/common.cpp
+++ b/common/common.cpp
@@ -278,8 +278,18 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
                 break;
             }
             params.yarn_beta_slow = std::stof(argv[i]);
-        } else if (arg == "--memory-f32") {
-            params.memory_f16 = false;
+        } else if (arg == "--samplers") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            sparams.samplers_sequence = parse_samplers_input(argv[i]);
+        } else if (arg == "--sampling-seq") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            sparams.samplers_sequence = argv[i];
         } else if (arg == "--top-p") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -498,6 +508,12 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
             params.infill = true;
         } else if (arg == "-dkvc" || arg == "--dump-kv-cache") {
             params.dump_kv_cache = true;
+        } else if (arg == "-nkvo" || arg == "--no-kv-offload") {
+            params.no_kv_offload = true;
+        } else if (arg == "-ctk" || arg == "--cache-type-k") {
+            params.cache_type_k = argv[++i];
+        } else if (arg == "-ctv" || arg == "--cache-type-v") {
+            params.cache_type_v = argv[++i];
         } else if (arg == "--multiline-input") {
             params.multiline_input = true;
         } else if (arg == "--simple-io") {
@@ -640,6 +656,10 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
         } else if (arg == "-h" || arg == "--help") {
             return false;
 
+        } else if (arg == "--version") {
+            fprintf(stderr, "version: %d (%s)\n", LLAMA_BUILD_NUMBER, LLAMA_COMMIT);
+            fprintf(stderr, "built with %s for %s\n", LLAMA_COMPILER, LLAMA_BUILD_TARGET);
+            exit(0);
         } else if (arg == "--random-prompt") {
             params.random_prompt = true;
         } else if (arg == "--in-prefix-bos") {
@@ -678,6 +698,47 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
                 std::istreambuf_iterator<char>(),
                 std::back_inserter(sparams.grammar)
             );
+        } else if (arg == "--override-kv") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            char * sep = strchr(argv[i], '=');
+            if (sep == nullptr || sep - argv[i] >= 128) {
+                fprintf(stderr, "error: Malformed KV override: %s\n", argv[i]);
+                invalid_param = true;
+                break;
+            }
+            struct llama_model_kv_override kvo;
+            std::strncpy(kvo.key, argv[i], sep - argv[i]);
+            kvo.key[sep - argv[i]] = 0;
+            sep++;
+            if (strncmp(sep, "int:", 4) == 0) {
+                sep += 4;
+                kvo.tag = LLAMA_KV_OVERRIDE_INT;
+                kvo.int_value = std::atol(sep);
+            } else if (strncmp(sep, "float:", 6) == 0) {
+                sep += 6;
+                kvo.tag = LLAMA_KV_OVERRIDE_FLOAT;
+                kvo.float_value = std::atof(sep);
+            } else if (strncmp(sep, "bool:", 5) == 0) {
+                sep += 5;
+                kvo.tag = LLAMA_KV_OVERRIDE_BOOL;
+                if (std::strcmp(sep, "true") == 0) {
+                    kvo.bool_value = true;
+                } else if (std::strcmp(sep, "false") == 0) {
+                    kvo.bool_value = false;
+                } else {
+                    fprintf(stderr, "error: Invalid boolean value for KV override: %s\n", argv[i]);
+                    invalid_param = true;
+                    break;
+                }
+            } else {
+                fprintf(stderr, "error: Invalid type for KV override: %s\n", argv[i]);
+                invalid_param = true;
+                break;
+            }
+            params.kv_overrides.push_back(kvo);
 #ifndef LOG_DISABLE_LOGS
         // Parse args for logging parameters
         } else if ( log_param_single_parse( argv[i] ) ) {
@@ -721,6 +782,11 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
         }
     }
 
+    if (!params.kv_overrides.empty()) {
+        params.kv_overrides.emplace_back(llama_model_kv_override());
+        params.kv_overrides.back().key[0] = 0;
+    }
+
     return true;
 }
 
@@ -732,6 +798,7 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     printf("\n");
     printf("options:\n");
     printf("  -h, --help            show this help message and exit\n");
+    printf("      --version         show version and build info\n");
     printf("  -i, --interactive     run in interactive mode\n");
     printf("  --interactive-first   run in interactive mode and wait for input right away\n");
     printf("  -ins, --instruct      run in instruction mode (use with Alpaca models)\n");
@@ -761,6 +828,8 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     printf("  -n N, --n-predict N   number of tokens to predict (default: %d, -1 = infinity, -2 = until context filled)\n", params.n_predict);
     printf("  -c N, --ctx-size N    size of the prompt context (default: %d, 0 = loaded from model)\n", params.n_ctx);
     printf("  -b N, --batch-size N  batch size for prompt processing (default: %d)\n", params.n_batch);
+    printf("  --samplers            samplers that will be used for generation in the order, separated by \';\', for example: \"top_k;tfs;typical;top_p;min_p;temp\"\n");
+    printf("  --sampling-seq        simplified sequence for samplers that will be used (default: %s)\n", sparams.samplers_sequence.c_str());
     printf("  --top-k N             top-k sampling (default: %d, 0 = disabled)\n", sparams.top_k);
     printf("  --top-p N             top-p sampling (default: %.1f, 1.0 = disabled)\n", (double)sparams.top_p);
     printf("  --min-p N             min-p sampling (default: %.1f, 0.0 = disabled)\n", (double)sparams.min_p);
@@ -798,8 +867,6 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     printf("  --yarn-beta-fast N    YaRN: low correction dim or beta (default: %.1f)\n", params.yarn_beta_fast);
     printf("  --ignore-eos          ignore end of stream token and continue generating (implies --logit-bias 2-inf)\n");
     printf("  --no-penalize-nl      do not penalize newline token\n");
-    printf("  --memory-f32          use f32 instead of f16 for memory key+value (default: disabled)\n");
-    printf("                        not recommended: doubles context memory required and no measurable increase in quality\n");
     printf("  --temp N              temperature (default: %.1f)\n", (double)sparams.temp);
     printf("  --logits-all          return logits for all tokens in the batch (default: disabled)\n");
     printf("  --hellaswag           compute HellaSwag score over random tasks from datafile supplied with -f\n");
@@ -840,6 +907,12 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     printf("  --verbose-prompt      print prompt before generation\n");
     printf("  -dkvc, --dump-kv-cache\n");
     printf("                        verbose print of the KV cache\n");
+    printf("  -nkvo, --no-kv-offload\n");
+    printf("                        disable KV offload\n");
+    printf("  -ctk TYPE, --cache-type-k TYPE\n");
+    printf("                        KV cache data type for K (default: %s)\n", params.cache_type_k.c_str());
+    printf("  -ctv TYPE, --cache-type-v TYPE\n");
+    printf("                        KV cache data type for V (default: %s)\n", params.cache_type_v.c_str());
     printf("  --simple-io           use basic IO for better compatibility in subprocesses and limited consoles\n");
     printf("  --lora FNAME          apply LoRA adapter (implies --no-mmap)\n");
     printf("  --lora-scaled FNAME S apply LoRA adapter with user defined scaling S (implies --no-mmap)\n");
@@ -847,9 +920,12 @@ 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");
+    printf("                        advanced option to override model metadata by key. may be specified multiple times.\n");
+    printf("                        types: int, float, bool. example: --override-kv tokenizer.ggml.add_bos_token=bool:false\n");
     printf("\n");
 #ifndef LOG_DISABLE_LOGS
     log_print_usage();
@@ -886,6 +962,48 @@ std::string gpt_random_prompt(std::mt19937 & rng) {
     GGML_UNREACHABLE();
 }
 
+//
+// String parsing
+//
+
+std::string parse_samplers_input(std::string input) {
+    std::string output = "";
+    // since samplers names are written multiple ways
+    // make it ready for both system names and input names
+    std::unordered_map<std::string, char> samplers_symbols {
+        {"top_k",      'k'},
+        {"top-k",      'k'},
+        {"top_p",      'p'},
+        {"top-p",      'p'},
+        {"nucleus",    'p'},
+        {"typical_p",  'y'},
+        {"typical-p",  'y'},
+        {"typical",    'y'},
+        {"min_p",      'm'},
+        {"min-p",      'm'},
+        {"tfs_z",      'f'},
+        {"tfs-z",      'f'},
+        {"tfs",        'f'},
+        {"temp",       't'},
+        {"temperature",'t'}
+    };
+    // expected format example: "temp;top_k;tfs_z;typical_p;top_p;min_p"
+    size_t separator = input.find(';');
+    while (separator != input.npos) {
+        std::string name = input.substr(0,separator);
+        input = input.substr(separator+1);
+        separator = input.find(';');
+
+        if (samplers_symbols.find(name) != samplers_symbols.end()) {
+            output += samplers_symbols[name];
+        }
+    }
+    if (samplers_symbols.find(input) != samplers_symbols.end()) {
+        output += samplers_symbols[input];
+    }
+    return output;
+}
+
 //
 // Model utils
 //
@@ -900,10 +1018,39 @@ struct llama_model_params llama_model_params_from_gpt_params(const gpt_params &
     mparams.tensor_split    = params.tensor_split;
     mparams.use_mmap        = params.use_mmap;
     mparams.use_mlock       = params.use_mlock;
+    if (params.kv_overrides.empty()) {
+        mparams.kv_overrides = NULL;
+    } else {
+        GGML_ASSERT(params.kv_overrides.back().key[0] == 0 && "KV overrides not terminated with empty key");
+        mparams.kv_overrides = params.kv_overrides.data();
+    }
 
     return mparams;
 }
 
+static ggml_type kv_cache_type_from_str(const std::string & s) {
+    if (s == "f16") {
+        return GGML_TYPE_F16;
+    }
+    if (s == "q8_0") {
+        return GGML_TYPE_Q8_0;
+    }
+    if (s == "q4_0") {
+        return GGML_TYPE_Q4_0;
+    }
+    if (s == "q4_1") {
+        return GGML_TYPE_Q4_1;
+    }
+    if (s == "q5_0") {
+        return GGML_TYPE_Q5_0;
+    }
+    if (s == "q5_1") {
+        return GGML_TYPE_Q5_1;
+    }
+
+    throw std::runtime_error("Invalid cache type: " + s);
+}
+
 struct llama_context_params llama_context_params_from_gpt_params(const gpt_params & params) {
     auto cparams = llama_context_default_params();
 
@@ -913,7 +1060,6 @@ struct llama_context_params llama_context_params_from_gpt_params(const gpt_param
     cparams.n_threads_batch   = params.n_threads_batch == -1 ? params.n_threads : params.n_threads_batch;
     cparams.mul_mat_q         = params.mul_mat_q;
     cparams.seed              = params.seed;
-    cparams.f16_kv            = params.memory_f16;
     cparams.logits_all        = params.logits_all;
     cparams.embedding         = params.embedding;
     cparams.rope_scaling_type = params.rope_scaling_type;
@@ -924,6 +1070,10 @@ struct llama_context_params llama_context_params_from_gpt_params(const gpt_param
     cparams.yarn_beta_fast    = params.yarn_beta_fast;
     cparams.yarn_beta_slow    = params.yarn_beta_slow;
     cparams.yarn_orig_ctx     = params.yarn_orig_ctx;
+    cparams.offload_kqv       = !params.no_kv_offload;
+
+    cparams.type_k = kv_cache_type_from_str(params.cache_type_k);
+    cparams.type_v = kv_cache_type_from_str(params.cache_type_v);
 
     return cparams;
 }
@@ -1336,7 +1486,6 @@ void dump_non_result_info_yaml(FILE * stream, const gpt_params & params, const l
     }
     fprintf(stream, "lora_base: %s\n", params.lora_base.c_str());
     fprintf(stream, "main_gpu: %d # default: 0\n", params.main_gpu);
-    fprintf(stream, "memory_f32: %s # default: false\n", !params.memory_f16 ? "true" : "false");
     fprintf(stream, "mirostat: %d # default: 0 (disabled)\n", sparams.mirostat);
     fprintf(stream, "mirostat_ent: %f # default: 5.0\n", sparams.mirostat_tau);
     fprintf(stream, "mirostat_lr: %f # default: 0.1\n", sparams.mirostat_eta);
diff --git a/common/common.h b/common/common.h
index 2f6fe48ab..9659aa045 100644
--- a/common/common.h
+++ b/common/common.h
@@ -51,7 +51,7 @@ struct gpt_params {
     int32_t n_ctx                           = 512;   // context size
     int32_t n_batch                         = 512;   // batch size for prompt processing (must be >=32 to use BLAS)
     int32_t n_keep                          = 0;     // number of tokens to keep from initial prompt
-    int32_t n_draft                         = 16;    // number of tokens to draft during speculative decoding
+    int32_t n_draft                         = 8;     // number of tokens to draft during speculative decoding
     int32_t n_chunks                        = -1;    // max number of chunks to process (-1 = unlimited)
     int32_t n_parallel                      = 1;     // number of parallel sequences to decode
     int32_t n_sequences                     = 1;     // number of sequences to decode
@@ -86,6 +86,8 @@ struct gpt_params {
     std::vector<std::string> antiprompt; // string upon seeing which more user input is prompted
     std::string logdir            = "";  // directory in which to save YAML log files
 
+    std::vector<llama_model_kv_override> kv_overrides;
+
     // TODO: avoid tuple, use struct
     std::vector<std::tuple<std::string, float>> lora_adapter; // lora adapter path with user defined scale
     std::string lora_base  = "";                              // base model path for the lora adapter
@@ -98,7 +100,6 @@ struct gpt_params {
     size_t hellaswag_tasks = 400;   // number of tasks to use when computing the HellaSwag score
 
     bool mul_mat_q         = true;  // if true, use mul_mat_q kernels instead of cuBLAS
-    bool memory_f16        = true;  // use f16 instead of f32 for memory kv
     bool random_prompt     = false; // do not randomize prompt if none provided
     bool use_color         = false; // use color to distinguish generations and inputs
     bool interactive       = false; // interactive mode
@@ -123,10 +124,14 @@ struct gpt_params {
     bool verbose_prompt    = false; // print prompt tokens before generation
     bool infill            = false; // use infill mode
     bool dump_kv_cache     = false; // dump the KV cache contents for debugging purposes
+    bool no_kv_offload     = false; // disable KV offloading
+
+    std::string cache_type_k = "f16"; // KV cache data type for the K
+    std::string cache_type_v = "f16"; // KV cache data type for the V
 
     // multimodal models (see examples/llava)
     std::string mmproj = ""; // path to multimodal projector
-    std::string image = ""; // path to an image file
+    std::string image  = ""; // path to an image file
 };
 
 bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params);
@@ -141,6 +146,12 @@ std::string gpt_random_prompt(std::mt19937 & rng);
 
 void process_escapes(std::string& input);
 
+//
+// String parsing
+//
+
+std::string parse_samplers_input(std::string input);
+
 //
 // Model utils
 //
@@ -229,3 +240,4 @@ void dump_kv_cache_view(const llama_kv_cache_view & view, int row_size = 80);
 
 // Dump the KV cache view showing individual sequences in each cell (long output).
 void dump_kv_cache_view_seqs(const llama_kv_cache_view & view, int row_size = 40);
+
diff --git a/common/grammar-parser.cpp b/common/grammar-parser.cpp
index ff51cc803..bf89a96f3 100644
--- a/common/grammar-parser.cpp
+++ b/common/grammar-parser.cpp
@@ -190,7 +190,7 @@ namespace grammar_parser {
                 pos = parse_space(pos + 1, is_nested);
             } else if (*pos == '*' || *pos == '+' || *pos == '?') { // repetition operator
                 if (last_sym_start == out_elements.size()) {
-                    throw std::runtime_error(std::string("expecting preceeding item to */+/? at ") + pos);
+                    throw std::runtime_error(std::string("expecting preceding item to */+/? at ") + pos);
                 }
 
                 // apply transformation to previous symbol (last_sym_start to end) according to
diff --git a/common/log.h b/common/log.h
index c0e814861..e4e1b9f4f 100644
--- a/common/log.h
+++ b/common/log.h
@@ -61,13 +61,13 @@
 //  #define LOG_TARGET stderr
 //  #include "log.h"
 //
-//  The log target can also be redirected to a diffrent function
+//  The log target can also be redirected to a different function
 //  like so:
 //
-//  #define LOG_TARGET log_handler_diffrent()
+//  #define LOG_TARGET log_handler_different()
 //  #include "log.h"
 //
-//  FILE* log_handler_diffrent()
+//  FILE* log_handler_different()
 //  {
 //      return stderr;
 //  }
@@ -421,7 +421,7 @@ inline FILE *log_handler2_impl(bool change = false, LogTriState append = LogTriS
 
 // Disables logs entirely at runtime.
 //  Makes LOG() and LOG_TEE() produce no output,
-//  untill enabled back.
+//  until enabled back.
 #define log_disable() log_disable_impl()
 
 // INTERNAL, DO NOT USE
diff --git a/common/sampling.cpp b/common/sampling.cpp
index 6344c29da..5b15204be 100644
--- a/common/sampling.cpp
+++ b/common/sampling.cpp
@@ -99,6 +99,56 @@ std::string llama_sampling_print(const llama_sampling_params & params) {
     return std::string(result);
 }
 
+std::string llama_sampling_order_print(const llama_sampling_params & params) {
+    std::string result = "CFG -> Penalties ";
+    if (params.mirostat == 0) {
+        for (auto s : params.samplers_sequence) {
+            switch (s) {
+                case 'k': result += "-> top_k "; break;
+                case 'f': result += "-> tfs_z "; break;
+                case 'y': result += "-> typical_p "; break;
+                case 'p': result += "-> top_p "; break;
+                case 'm': result += "-> min_p "; break;
+                case 't': result += "-> temp "; break;
+                default : break;
+            }
+        }
+    } else {
+        result += "-> mirostat ";
+    }
+
+    return result;
+}
+
+// no reasons to expose this function in header
+static void sampler_queue(
+                   struct llama_context * ctx_main,
+            const llama_sampling_params & params,
+                 llama_token_data_array & cur_p,
+                                 size_t & min_keep) {
+    const int n_vocab = llama_n_vocab(llama_get_model(ctx_main));
+
+    const float         temp              = params.temp;
+    const int32_t       top_k             = params.top_k <= 0 ? n_vocab : params.top_k;
+    const float         top_p             = params.top_p;
+    const float         min_p             = params.min_p;
+    const float         tfs_z             = params.tfs_z;
+    const float         typical_p         = params.typical_p;
+    const std::string & samplers_sequence = params.samplers_sequence;
+
+    for (auto s : samplers_sequence) {
+        switch (s){
+            case 'k': llama_sample_top_k    (ctx_main, &cur_p, top_k,     min_keep); break;
+            case 'f': llama_sample_tail_free(ctx_main, &cur_p, tfs_z,     min_keep); break;
+            case 'y': llama_sample_typical  (ctx_main, &cur_p, typical_p, min_keep); break;
+            case 'p': llama_sample_top_p    (ctx_main, &cur_p, top_p,     min_keep); break;
+            case 'm': llama_sample_min_p    (ctx_main, &cur_p, min_p,     min_keep); break;
+            case 't': llama_sample_temp     (ctx_main, &cur_p, temp); break;
+            default : break;
+        }
+    }
+}
+
 static llama_token llama_sampling_sample_impl(
                   struct llama_sampling_context * ctx_sampling,
                   struct llama_context * ctx_main,
@@ -110,11 +160,6 @@ static llama_token llama_sampling_sample_impl(
     const int n_vocab = llama_n_vocab(llama_get_model(ctx_main));
 
     const float   temp            = params.temp;
-    const int32_t top_k           = params.top_k <= 0 ? n_vocab : params.top_k;
-    const float   top_p           = params.top_p;
-    const float   min_p           = params.min_p;
-    const float   tfs_z           = params.tfs_z;
-    const float   typical_p       = params.typical_p;
     const int32_t penalty_last_n  = params.penalty_last_n < 0 ? params.n_prev : params.penalty_last_n;
     const float   penalty_repeat  = params.penalty_repeat;
     const float   penalty_freq    = params.penalty_freq;
@@ -199,12 +244,7 @@ static llama_token llama_sampling_sample_impl(
             // temperature sampling
             size_t min_keep = std::max(1, params.n_probs);
 
-            llama_sample_top_k    (ctx_main, &cur_p, top_k,     min_keep);
-            llama_sample_tail_free(ctx_main, &cur_p, tfs_z,     min_keep);
-            llama_sample_typical  (ctx_main, &cur_p, typical_p, min_keep);
-            llama_sample_top_p    (ctx_main, &cur_p, top_p,     min_keep);
-            llama_sample_min_p    (ctx_main, &cur_p, min_p,     min_keep);
-            llama_sample_temp     (ctx_main, &cur_p, temp);
+            sampler_queue(ctx_main, params, cur_p, min_keep);
 
             id = llama_sample_token(ctx_main, &cur_p);
 
diff --git a/common/sampling.h b/common/sampling.h
index 4a8c522b6..7dbb865ac 100644
--- a/common/sampling.h
+++ b/common/sampling.h
@@ -10,22 +10,23 @@
 
 // sampling parameters
 typedef struct llama_sampling_params {
-    int32_t n_prev            = 64;    // number of previous tokens to remember
-    int32_t n_probs           = 0;     // if greater than 0, output the probabilities of top n_probs tokens.
-    int32_t top_k             = 40;    // <= 0 to use vocab size
-    float   top_p             = 0.95f; // 1.0 = disabled
-    float   min_p             = 0.05f; // 0.0 = disabled
-    float   tfs_z             = 1.00f; // 1.0 = disabled
-    float   typical_p         = 1.00f; // 1.0 = disabled
-    float   temp              = 0.80f; // 1.0 = disabled
-    int32_t penalty_last_n    = 64;    // last n tokens to penalize (0 = disable penalty, -1 = context size)
-    float   penalty_repeat    = 1.10f; // 1.0 = disabled
-    float   penalty_freq      = 0.00f; // 0.0 = disabled
-    float   penalty_present   = 0.00f; // 0.0 = disabled
-    int32_t mirostat          = 0;     // 0 = disabled, 1 = mirostat, 2 = mirostat 2.0
-    float   mirostat_tau      = 5.00f; // target entropy
-    float   mirostat_eta      = 0.10f; // learning rate
-    bool    penalize_nl       = true;  // consider newlines as a repeatable token
+    int32_t     n_prev                = 64;       // number of previous tokens to remember
+    int32_t     n_probs               = 0;        // if greater than 0, output the probabilities of top n_probs tokens.
+    int32_t     top_k                 = 40;       // <= 0 to use vocab size
+    float       top_p                 = 0.95f;    // 1.0 = disabled
+    float       min_p                 = 0.05f;    // 0.0 = disabled
+    float       tfs_z                 = 1.00f;    // 1.0 = disabled
+    float       typical_p             = 1.00f;    // 1.0 = disabled
+    float       temp                  = 0.80f;    // 1.0 = disabled
+    int32_t     penalty_last_n        = 64;       // last n tokens to penalize (0 = disable penalty, -1 = context size)
+    float       penalty_repeat        = 1.10f;    // 1.0 = disabled
+    float       penalty_freq          = 0.00f;    // 0.0 = disabled
+    float       penalty_present       = 0.00f;    // 0.0 = disabled
+    int32_t     mirostat              = 0;        // 0 = disabled, 1 = mirostat, 2 = mirostat 2.0
+    float       mirostat_tau          = 5.00f;    // target entropy
+    float       mirostat_eta          = 0.10f;    // learning rate
+    bool        penalize_nl           = true;     // consider newlines as a repeatable token
+    std::string samplers_sequence     = "kfypmt"; // top_k, tail_free, typical_p, top_p, min_p, temp
 
     std::string grammar;  // optional BNF-like grammar to constrain sampling
 
@@ -80,6 +81,9 @@ std::string llama_sampling_prev_str(llama_sampling_context * ctx_sampling, llama
 // Print sampling parameters into a string
 std::string llama_sampling_print(const llama_sampling_params & params);
 
+// Print sampling order into a string
+std::string llama_sampling_order_print(const llama_sampling_params & params);
+
 // this is a common sampling function used across the examples for convenience
 // it can serve as a starting point for implementing your own sampling function
 // Note: When using multiple sequences, it is the caller's responsibility to call
diff --git a/common/train.cpp b/common/train.cpp
index 773e2c59c..dcf9614e4 100644
--- a/common/train.cpp
+++ b/common/train.cpp
@@ -71,7 +71,7 @@ void free_random_uniform_distribution(struct random_uniform_distribution * rnd)
 
 struct ggml_tensor * randomize_tensor_normal(struct ggml_tensor * tensor, struct random_normal_distribution * rnd) {
     float scale = 1.0f; // xavier
-    switch (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];
diff --git a/convert-hf-to-gguf.py b/convert-hf-to-gguf.py
index bced1f561..e71a96c48 100755
--- a/convert-hf-to-gguf.py
+++ b/convert-hf-to-gguf.py
@@ -77,8 +77,18 @@ class Model:
             self.gguf_writer.add_embedding_length(n_embd)
         if (n_ff := self.hparams.get("intermediate_size")) is not None:
             self.gguf_writer.add_feed_forward_length(n_ff)
-        if (n_head := self.hparams.get("num_attention_head")) is not None:
+        if (n_head := self.hparams.get("num_attention_heads")) is not None:
             self.gguf_writer.add_head_count(n_head)
+        if (n_head_kv := self.hparams.get("num_key_value_heads")) is not None:
+            self.gguf_writer.add_head_count_kv(n_head_kv)
+
+        if (n_rms_eps := self.hparams.get("rms_norm_eps")) is not None:
+            self.gguf_writer.add_layer_norm_rms_eps(n_rms_eps)
+        if (n_experts := self.hparams.get("num_local_experts")) is not None:
+            self.gguf_writer.add_expert_count(n_experts)
+        if (n_experts_used := self.hparams.get("num_experts_per_tok")) is not None:
+            self.gguf_writer.add_expert_used_count(n_experts_used)
+
         self.gguf_writer.add_parallel_residual(self.hparams.get("use_parallel_residual", True))
 
     def write_tensors(self):
@@ -170,6 +180,10 @@ class Model:
             return StableLMModel
         if model_architecture == "QWenLMHeadModel":
             return QwenModel
+        if model_architecture == "MixtralForCausalLM":
+            return MixtralModel
+        if model_architecture == "PhiForCausalLM":
+            return Phi2Model
         return Model
 
     def _is_model_safetensors(self) -> bool:
@@ -207,6 +221,10 @@ class Model:
             return gguf.MODEL_ARCH.STABLELM
         if arch == "QWenLMHeadModel":
             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!')
 
@@ -837,6 +855,11 @@ class StableLMModel(Model):
         self.gguf_writer.add_layer_norm_eps(1e-5)
 
 
+class MixtralModel(Model):
+    def set_vocab(self):
+        self._set_vocab_sentencepiece()
+
+
 class QwenModel(Model):
     @staticmethod
     def token_bytes_to_string(b):
@@ -961,6 +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 ######
 
 
diff --git a/convert-lora-to-ggml.py b/convert-lora-to-ggml.py
index a937410dd..53bb8a3d9 100755
--- a/convert-lora-to-ggml.py
+++ b/convert-lora-to-ggml.py
@@ -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)
diff --git a/convert.py b/convert.py
index 6e95d6cb3..7a3cd615e 100755
--- a/convert.py
+++ b/convert.py
@@ -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
@@ -42,6 +43,7 @@ NDArray: TypeAlias = 'np.ndarray[Any, Any]'
 ARCH = gguf.MODEL_ARCH.LLAMA
 
 DEFAULT_CONCURRENCY = 8
+
 #
 # data types
 #
@@ -62,10 +64,10 @@ class UnquantizedDataType(DataType):
     pass
 
 
-DT_F16  = UnquantizedDataType('F16', dtype = np.dtype(np.float16), valid_conversions = ['F32', 'Q8_0'])
-DT_F32  = UnquantizedDataType('F32', dtype = np.dtype(np.float32), valid_conversions = ['F16', 'Q8_0'])
-DT_I32  = UnquantizedDataType('I32', dtype = np.dtype(np.int16), valid_conversions = [])
-DT_BF16 = UnquantizedDataType('BF16', dtype = np.dtype(np.uint16), valid_conversions = ['F32', 'F16', 'Q8_0'])
+DT_F16  = UnquantizedDataType('F16',  dtype = np.dtype(np.float16), valid_conversions = ['F32', 'Q8_0'])
+DT_F32  = UnquantizedDataType('F32',  dtype = np.dtype(np.float32), valid_conversions = ['F16', 'Q8_0'])
+DT_I32  = UnquantizedDataType('I32',  dtype = np.dtype(np.int16),   valid_conversions = [])
+DT_BF16 = UnquantizedDataType('BF16', dtype = np.dtype(np.uint16),  valid_conversions = ['F32', 'F16', 'Q8_0'])
 
 
 @dataclass(frozen=True)
@@ -151,14 +153,16 @@ GGML_FILE_TYPE_TO_DATA_TYPE: dict[GGMLFileType, DataType] = {
 
 @dataclass
 class Params:
-    n_vocab:    int
-    n_embd:     int
-    n_layer:    int
-    n_ctx:      int
-    n_ff:       int
-    n_head:     int
-    n_head_kv:  int
-    f_norm_eps: float
+    n_vocab:        int
+    n_embd:         int
+    n_layer:        int
+    n_ctx:          int
+    n_ff:           int
+    n_head:         int
+    n_head_kv:      int
+    n_experts:      int | None = None
+    n_experts_used: int | None = None
+    f_norm_eps:     float | None = None
 
     rope_scaling_type: gguf.RopeScalingType | None = None
     f_rope_freq_base: float | None = None
@@ -233,6 +237,13 @@ class Params:
             raise Exception("failed to guess 'n_ctx'. This model is unknown or unsupported.\n"
                             "Suggestion: provide 'config.json' of the model in the same directory containing model files.")
 
+        n_experts      = None
+        n_experts_used = None
+
+        if "num_local_experts" in config:
+            n_experts = config["num_local_experts"]
+            n_experts_used = config["num_experts_per_tok"]
+
         return Params(
             n_vocab           = config["vocab_size"],
             n_embd            = config["hidden_size"],
@@ -241,6 +252,8 @@ class Params:
             n_ff              = config["intermediate_size"],
             n_head            = (n_head := config["num_attention_heads"]),
             n_head_kv         = config.get("num_key_value_heads", n_head),
+            n_experts         = n_experts,
+            n_experts_used    = n_experts_used,
             f_norm_eps        = config["rms_norm_eps"],
             f_rope_freq_base  = config.get("rope_theta"),
             rope_scaling_type = rope_scaling_type,
@@ -255,8 +268,15 @@ class Params:
     def loadOriginalParamsJson(model: LazyModel, config_path: Path) -> Params:
         config = json.load(open(config_path))
 
+        n_experts      = None
+        n_experts_used = None
+        f_rope_freq_base = None
+
         # hack to determine LLaMA v1 vs v2 vs CodeLlama
-        if config.get("rope_theta") == 1000000:
+        if config.get("moe"):
+            # Mixtral
+            n_ctx = 32768
+        elif config.get("rope_theta") == 1000000:
             # CodeLlama
             n_ctx = 16384
         elif config["norm_eps"] == 1e-05:
@@ -266,16 +286,27 @@ class Params:
             # LLaMA v1
             n_ctx = 2048
 
+        if "layers.0.feed_forward.w1.weight" in model:
+            n_ff = model["layers.0.feed_forward.w1.weight"].shape[0]
+
+        if config.get("moe"):
+            n_ff = model["layers.0.feed_forward.experts.0.w1.weight"].shape[0]
+            n_experts      = config["moe"]["num_experts"]
+            n_experts_used = config["moe"]["num_experts_per_tok"]
+            f_rope_freq_base = 1e6
+
         return Params(
             n_vocab          = model["tok_embeddings.weight"].shape[0],
             n_embd           = config["dim"],
             n_layer          = config["n_layers"],
             n_ctx            = n_ctx,
-            n_ff             = model["layers.0.feed_forward.w1.weight"].shape[0],
+            n_ff             = n_ff,
             n_head           = (n_head := config["n_heads"]),
             n_head_kv        = config.get("n_kv_heads", n_head),
+            n_experts        = n_experts,
+            n_experts_used   = n_experts_used,
             f_norm_eps       = config["norm_eps"],
-            f_rope_freq_base = config.get("rope_theta"),
+            f_rope_freq_base = config.get("rope_theta", f_rope_freq_base),
         )
 
     @staticmethod
@@ -297,127 +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
@@ -585,7 +627,7 @@ def merge_multifile_models(models_plus: list[ModelPlus]) -> ModelPlus:
 
     if any("model.embed_tokens.weight" in mp.model for mp in models_plus):
         # Transformers models put different tensors in different files, but
-        # don't split indivdual tensors between files.
+        # don't split individual tensors between files.
         model: LazyModel = {}
         for mp in models_plus:
             model.update(mp.model)
@@ -678,7 +720,7 @@ class LazyUnpickler(pickle.Unpickler):
         return func(*args)
 
     CLASSES: dict[tuple[str, str], Any] = {
-        # getattr used here as a workaround for mypy not being smart enough to detrmine
+        # getattr used here as a workaround for mypy not being smart enough to determine
         # the staticmethods have a __func__ attribute.
         ('torch._tensor', '_rebuild_from_type_v2'): getattr(rebuild_from_type_v2, '__func__'),
         ('torch._utils', '_rebuild_tensor_v2'): getattr(lazy_rebuild_tensor_v2, '__func__'),
@@ -794,20 +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)
 
 
@@ -832,7 +881,17 @@ class OutputFile:
         self.gguf.add_rope_dimension_count(params.n_embd // params.n_head)
         self.gguf.add_head_count          (params.n_head)
         self.gguf.add_head_count_kv       (params.n_head_kv)
-        self.gguf.add_layer_norm_rms_eps  (params.f_norm_eps)
+
+        if params.n_experts:
+            self.gguf.add_expert_count(params.n_experts)
+
+        if params.n_experts_used:
+            self.gguf.add_expert_used_count(params.n_experts_used)
+
+        if params.f_norm_eps:
+            self.gguf.add_layer_norm_rms_eps(params.f_norm_eps)
+        else:
+            raise ValueError('f_norm_eps is None')
 
         if params.f_rope_freq_base is not None:
             self.gguf.add_rope_freq_base(params.f_rope_freq_base)
@@ -861,12 +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)
@@ -892,8 +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)
 
@@ -920,8 +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)
 
@@ -956,7 +1020,7 @@ class OutputFile:
 
 
 def pick_output_type(model: LazyModel, output_type_str: str | None) -> GGMLFileType:
-    wq_type = model[gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.ATTN_Q].format(bid=0) +".weight"].data_type
+    wq_type = model[gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.ATTN_Q].format(bid=0) + ".weight"].data_type
 
     if output_type_str == "f32" or (output_type_str is None and wq_type == DT_F32):
         return GGMLFileType.AllF32
@@ -1079,35 +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:
@@ -1145,11 +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:
@@ -1192,12 +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
 
@@ -1205,12 +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)
@@ -1220,7 +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}")
 
 
diff --git a/examples/CMakeLists.txt b/examples/CMakeLists.txt
index 6744944fd..4cc13d6e9 100644
--- a/examples/CMakeLists.txt
+++ b/examples/CMakeLists.txt
@@ -33,6 +33,7 @@ else()
     add_subdirectory(simple)
     add_subdirectory(speculative)
     add_subdirectory(lookahead)
+    add_subdirectory(lookup)
     add_subdirectory(train-text-from-scratch)
     if (LLAMA_METAL)
         add_subdirectory(metal)
diff --git a/examples/baby-llama/baby-llama.cpp b/examples/baby-llama/baby-llama.cpp
index 8155101d0..e7d2ad592 100644
--- a/examples/baby-llama/baby-llama.cpp
+++ b/examples/baby-llama/baby-llama.cpp
@@ -575,10 +575,7 @@ static struct ggml_tensor * forward(
 
             // KQ_scaled = KQ / sqrt(n_embd/n_head)
             // KQ_scaled shape [n_past + N, N, n_head, 1]
-            struct ggml_tensor * KQ_scaled =
-                ggml_scale(ctx0,
-                        KQ,
-                        ggml_new_f32(ctx0, 1.0f/sqrtf(float(n_embd)/n_head)));
+            struct ggml_tensor * KQ_scaled = ggml_scale(ctx0, KQ, 1.0f/sqrtf(float(n_embd)/n_head));
 
             // KQ_masked = mask_past(KQ_scaled)
             // KQ_masked shape [n_past + N, N, n_head, 1]
@@ -844,10 +841,7 @@ static struct ggml_tensor * forward_batch(
 
             // KQ_scaled = KQ / sqrt(n_embd/n_head)
             // KQ_scaled shape [n_past + N, N, n_head, n_batch]
-            struct ggml_tensor * KQ_scaled =
-                ggml_scale(ctx0,
-                        KQ,
-                        ggml_new_f32(ctx0, 1.0f/sqrtf(float(n_embd)/n_head)));
+            struct ggml_tensor * KQ_scaled = ggml_scale(ctx0, KQ, 1.0f/sqrtf(float(n_embd)/n_head));
             assert_shape_4d(KQ_scaled, n_past + N, N, n_head, n_batch);
 
             // KQ_masked = mask_past(KQ_scaled)
@@ -1131,10 +1125,7 @@ static struct ggml_tensor * forward_lora(
 
             // KQ_scaled = KQ / sqrt(n_embd/n_head)
             // KQ_scaled shape [n_past + N, N, n_head, 1]
-            struct ggml_tensor * KQ_scaled =
-                ggml_scale(ctx0,
-                        KQ,
-                        ggml_new_f32(ctx0, 1.0f/sqrtf(float(n_embd)/n_head)));
+            struct ggml_tensor * KQ_scaled = ggml_scale(ctx0, KQ, 1.0f/sqrtf(float(n_embd)/n_head));
 
             // KQ_masked = mask_past(KQ_scaled)
             // KQ_masked shape [n_past + N, N, n_head, 1]
@@ -1258,9 +1249,9 @@ 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 +1283,9 @@ 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 +1325,7 @@ 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 +1377,8 @@ 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]);
diff --git a/examples/batched.swift/Sources/main.swift b/examples/batched.swift/Sources/main.swift
index ce9d80d9b..4d0005349 100644
--- a/examples/batched.swift/Sources/main.swift
+++ b/examples/batched.swift/Sources/main.swift
@@ -215,9 +215,10 @@ print("decoded \(n_decode) tokens in \(String(format: "%.2f", Double(t_main_end
 llama_print_timings(context)
 
 private func tokenize(text: String, add_bos: Bool) -> [llama_token] {
-    let n_tokens = text.count + (add_bos ? 1 : 0)
+    let utf8Count = text.utf8.count
+    let n_tokens = utf8Count + (add_bos ? 1 : 0)
     let tokens = UnsafeMutablePointer<llama_token>.allocate(capacity: n_tokens)
-    let tokenCount = llama_tokenize(model, text, Int32(text.count), tokens, Int32(n_tokens), add_bos, /*special tokens*/ false)
+    let tokenCount = llama_tokenize(model, text, Int32(utf8Count), tokens, Int32(n_tokens), add_bos, /*special tokens*/ false)
     var swiftTokens: [llama_token] = []
     for i in 0 ..< tokenCount {
         swiftTokens.append(tokens[Int(i)])
diff --git a/examples/benchmark/benchmark-matmult.cpp b/examples/benchmark/benchmark-matmult.cpp
index 284733b10..434e1d6bd 100644
--- a/examples/benchmark/benchmark-matmult.cpp
+++ b/examples/benchmark/benchmark-matmult.cpp
@@ -129,13 +129,13 @@ int main(int argc, char ** argv)  {
     const ggml_type qtype = GGML_TYPE_Q4_1;
 
     size_t ctx_size = 0;
-    ctx_size += 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));
diff --git a/examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp b/examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp
index cae3bf3c3..4d41e1779 100644
--- a/examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp
+++ b/examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp
@@ -427,7 +427,7 @@ static void print_row(struct ggml_tensor * probs, int i) {
 }
 
 static void print_matrix(struct ggml_tensor * probs) {
-    assert(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++){
diff --git a/examples/export-lora/export-lora.cpp b/examples/export-lora/export-lora.cpp
index c8754ce70..58fbe204d 100644
--- a/examples/export-lora/export-lora.cpp
+++ b/examples/export-lora/export-lora.cpp
@@ -309,7 +309,7 @@ static struct ggml_cgraph * build_graph_lora(
 ) {
     struct ggml_tensor * ab = ggml_mul_mat(ctx, lora_a, lora_b);
     if (scaling != 1.0f) {
-        ab = ggml_scale(ctx, ab, ggml_new_f32(ctx, scaling));
+        ab = ggml_scale(ctx, ab, scaling);
     }
     struct ggml_tensor * res = ggml_add_inplace(ctx, tensor, ab);
 
diff --git a/examples/finetune/finetune.cpp b/examples/finetune/finetune.cpp
index af46e44a6..7b1333a9d 100644
--- a/examples/finetune/finetune.cpp
+++ b/examples/finetune/finetune.cpp
@@ -269,7 +269,7 @@ static void load_model_hparams_gguf(struct gguf_context * ctx, struct my_llama_h
     float rope_freq_scale = 1.0f;
     GGUF_GET_KEY(ctx, hparams->f_norm_rms_eps, gguf_get_val_f32, GGUF_TYPE_FLOAT32, false, kv(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS));
     GGUF_GET_KEY(ctx, hparams->rope_freq_base, gguf_get_val_f32, GGUF_TYPE_FLOAT32, false, kv(LLM_KV_ROPE_FREQ_BASE));
-    GGUF_GET_KEY(ctx, rope_freq_scale, gguf_get_val_f32, GGUF_TYPE_FLOAT32, false, kv(LLM_KV_ROPE_SCALE_LINEAR));
+    GGUF_GET_KEY(ctx, rope_freq_scale,         gguf_get_val_f32, GGUF_TYPE_FLOAT32, false, kv(LLM_KV_ROPE_SCALE_LINEAR));
     if (rope_freq_scale != 1.0f) {
         hparams->rope_freq_scale = 1.0f / rope_freq_scale;
     }
@@ -612,6 +612,7 @@ static struct ggml_tensor * llama_build_lora_finetune_graphs(
     const int n_rot       = hparams.n_embd_head();
     const int n_embd_head = hparams.n_embd_head();
     const int n_embd_gqa  = hparams.n_embd_gqa();
+
     const float rms_norm_eps    = hparams.f_norm_rms_eps;
     const float rope_freq_base  = hparams.rope_freq_base;
     const float rope_freq_scale = hparams.rope_freq_scale;
@@ -680,10 +681,7 @@ static struct ggml_tensor * llama_build_lora_finetune_graphs(
         checkpoints.push_back(t01);
     }
 
-    struct ggml_tensor * kv_scale = NULL;
-    if (!enable_flash_attn) {
-        kv_scale = ggml_new_f32(ctx, 1.0f/sqrtf(float(n_embd)/n_head));
-    }
+    const float kv_scale = 1.0f/sqrtf(float(n_embd)/n_head);
 
     for (int il = 0; il < n_layer; ++il) {
         struct my_llama_layer & layer = model->layers[il];
@@ -781,32 +779,32 @@ static struct ggml_tensor * llama_build_lora_finetune_graphs(
     // make sure some tensors are not reallocated by inserting new temporary nodes depending on them
     int n_leafs_before = gb->n_leafs;
     int n_nodes_before = gb->n_nodes;
-    struct ggml_tensor * one = ggml_new_f32(ctx, 1.0f);
+
     // output tensors
-    ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, t35, one));
-    ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, t36, one));
+    ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, t35, 1.0f));
+    ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, t36, 1.0f));
     // input gradient
-    ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, t36->grad, one));
+    ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, t36->grad, 1.0f));
     GGML_ASSERT(t36->grad->data == NULL && t36->grad->view_src == NULL);
     ggml_allocr_alloc(alloc, t36->grad);
     // KQ_pos
-    ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, KQ_pos, one));
+    ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, KQ_pos, 1.0f));
 
     // make sure base model tensors data cannot be used in viewable operations
-    ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, model->tok_embeddings, one));
-    ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, model->norm, one));
-    ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, model->output, one));
+    ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, model->tok_embeddings, 1.0f));
+    ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, model->norm, 1.0f));
+    ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, model->output, 1.0f));
     for (int il = 0; il < n_layer; ++il) {
         struct my_llama_layer & layer = model->layers[il];
-        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.attention_norm, one));
-        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.ffn_norm, one));
-        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.wq, one));
-        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.wk, one));
-        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.wv, one));
-        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.wo, one));
-        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.w1, one));
-        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.w2, one));
-        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.w3, one));
+        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.attention_norm, 1.0f));
+        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.ffn_norm, 1.0f));
+        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.wq, 1.0f));
+        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.wk, 1.0f));
+        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.wv, 1.0f));
+        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.wo, 1.0f));
+        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.w1, 1.0f));
+        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.w2, 1.0f));
+        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.w3, 1.0f));
     }
 
     // allocating checkpoints in one block to reduce memory fragmentation
@@ -1110,7 +1108,7 @@ 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 +1618,6 @@ 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 +1721,9 @@ 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 +1750,7 @@ 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 +1783,7 @@ 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 +1799,8 @@ 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;
diff --git a/examples/gguf/CMakeLists.txt b/examples/gguf/CMakeLists.txt
index 7d1806af3..6481f087b 100644
--- a/examples/gguf/CMakeLists.txt
+++ b/examples/gguf/CMakeLists.txt
@@ -1,5 +1,5 @@
 set(TARGET gguf)
 add_executable(${TARGET} gguf.cpp)
 install(TARGETS ${TARGET} RUNTIME)
-target_link_libraries(${TARGET} PRIVATE llama ${CMAKE_THREAD_LIBS_INIT})
+target_link_libraries(${TARGET} PRIVATE ggml ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_11)
diff --git a/examples/gguf/gguf.cpp b/examples/gguf/gguf.cpp
index 9ab63a293..e67be4fb2 100644
--- a/examples/gguf/gguf.cpp
+++ b/examples/gguf/gguf.cpp
@@ -1,5 +1,4 @@
 #include "ggml.h"
-#include "llama.h"
 
 #include <cstdio>
 #include <cinttypes>
@@ -195,7 +194,7 @@ 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;
diff --git a/examples/llama-bench/llama-bench.cpp b/examples/llama-bench/llama-bench.cpp
index 9bd82d565..6617c050d 100644
--- a/examples/llama-bench/llama-bench.cpp
+++ b/examples/llama-bench/llama-bench.cpp
@@ -53,6 +53,13 @@ static std::vector<T> split(const std::string & str, char delim) {
     return values;
 }
 
+template<typename T, typename F>
+static std::vector<std::string> transform_to_str(const std::vector<T> & values, F f) {
+    std::vector<std::string> str_values;
+    std::transform(values.begin(), values.end(), std::back_inserter(str_values), f);
+    return str_values;
+}
+
 template<typename T>
 static T avg(const std::vector<T> & v) {
     if (v.empty()) {
@@ -126,7 +133,8 @@ struct cmd_params {
     std::vector<int> n_prompt;
     std::vector<int> n_gen;
     std::vector<int> n_batch;
-    std::vector<bool> f32_kv;
+    std::vector<ggml_type> type_k;
+    std::vector<ggml_type> type_v;
     std::vector<int> n_threads;
     std::vector<int> n_gpu_layers;
     std::vector<int> main_gpu;
@@ -142,7 +150,8 @@ static const cmd_params cmd_params_defaults = {
     /* n_prompt      */ {512},
     /* n_gen         */ {128},
     /* n_batch       */ {512},
-    /* f32_kv        */ {false},
+    /* type_k        */ {GGML_TYPE_F16},
+    /* type_v        */ {GGML_TYPE_F16},
     /* n_threads     */ {get_num_physical_cores()},
     /* n_gpu_layers  */ {99},
     /* main_gpu      */ {0},
@@ -162,7 +171,8 @@ static void print_usage(int /* argc */, char ** argv) {
     printf("  -p, --n-prompt <n>                (default: %s)\n", join(cmd_params_defaults.n_prompt, ",").c_str());
     printf("  -n, --n-gen <n>                   (default: %s)\n", join(cmd_params_defaults.n_gen, ",").c_str());
     printf("  -b, --batch-size <n>              (default: %s)\n", join(cmd_params_defaults.n_batch, ",").c_str());
-    printf("  --memory-f32 <0|1>                (default: %s)\n", join(cmd_params_defaults.f32_kv, ",").c_str());
+    printf("  -ctk <t>, --cache-type-k <t>      (default: %s)\n", join(transform_to_str(cmd_params_defaults.type_k, ggml_type_name), ",").c_str());
+    printf("  -ctv <t>, --cache-type-v <t>      (default: %s)\n", join(transform_to_str(cmd_params_defaults.type_v, ggml_type_name), ",").c_str());
     printf("  -t, --threads <n>                 (default: %s)\n", join(cmd_params_defaults.n_threads, ",").c_str());
     printf("  -ngl, --n-gpu-layers <n>          (default: %s)\n", join(cmd_params_defaults.n_gpu_layers, ",").c_str());
     printf("  -mg, --main-gpu <i>               (default: %s)\n", join(cmd_params_defaults.main_gpu, ",").c_str());
@@ -173,9 +183,32 @@ static void print_usage(int /* argc */, char ** argv) {
     printf("  -v, --verbose                     (default: %s)\n", cmd_params_defaults.verbose ? "1" : "0");
     printf("\n");
     printf("Multiple values can be given for each parameter by separating them with ',' or by specifying the parameter multiple times.\n");
-
 }
 
+static ggml_type ggml_type_from_name(const std::string & s) {
+    if (s == "f16") {
+        return GGML_TYPE_F16;
+    }
+    if (s == "q8_0") {
+        return GGML_TYPE_Q8_0;
+    }
+    if (s == "q4_0") {
+        return GGML_TYPE_Q4_0;
+    }
+    if (s == "q4_1") {
+        return GGML_TYPE_Q4_1;
+    }
+    if (s == "q5_0") {
+        return GGML_TYPE_Q5_0;
+    }
+    if (s == "q5_1") {
+        return GGML_TYPE_Q5_1;
+    }
+
+    return GGML_TYPE_COUNT;
+}
+
+
 static cmd_params parse_cmd_params(int argc, char ** argv) {
     cmd_params params;
     std::string arg;
@@ -224,13 +257,38 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
             }
             auto p = split<int>(argv[i], split_delim);
             params.n_batch.insert(params.n_batch.end(), p.begin(), p.end());
-        } else if (arg == "--memory-f32") {
+        } else if (arg == "-ctk" || arg == "--cache-type-k") {
             if (++i >= argc) {
                 invalid_param = true;
                 break;
             }
-            auto p = split<int>(argv[i], split_delim);
-            params.f32_kv.insert(params.f32_kv.end(), p.begin(), p.end());
+            auto p = split<std::string>(argv[i], split_delim);
+            std::vector<ggml_type> types;
+            for (const auto & t : p) {
+                ggml_type gt = ggml_type_from_name(t);
+                if (gt == GGML_TYPE_COUNT) {
+                    invalid_param = true;
+                    break;
+                }
+                types.push_back(gt);
+            }
+            params.type_k.insert(params.type_k.end(), types.begin(), types.end());
+        } else if (arg == "-ctv" || arg == "--cache-type-v") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            auto p = split<std::string>(argv[i], split_delim);
+            std::vector<ggml_type> types;
+            for (const auto & t : p) {
+                ggml_type gt = ggml_type_from_name(t);
+                if (gt == GGML_TYPE_COUNT) {
+                    invalid_param = true;
+                    break;
+                }
+                types.push_back(gt);
+            }
+            params.type_v.insert(params.type_v.end(), types.begin(), types.end());
         } else if (arg == "-t" || arg == "--threads") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -321,7 +379,8 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
     if (params.n_prompt.empty())     { params.n_prompt = cmd_params_defaults.n_prompt; }
     if (params.n_gen.empty())        { params.n_gen = cmd_params_defaults.n_gen; }
     if (params.n_batch.empty())      { params.n_batch = cmd_params_defaults.n_batch; }
-    if (params.f32_kv.empty())       { params.f32_kv = cmd_params_defaults.f32_kv; }
+    if (params.type_k.empty())       { params.type_k = cmd_params_defaults.type_k; }
+    if (params.type_v.empty())       { params.type_v = cmd_params_defaults.type_v; }
     if (params.n_gpu_layers.empty()) { params.n_gpu_layers = cmd_params_defaults.n_gpu_layers; }
     if (params.main_gpu.empty())     { params.main_gpu = cmd_params_defaults.main_gpu; }
     if (params.mul_mat_q.empty())    { params.mul_mat_q = cmd_params_defaults.mul_mat_q; }
@@ -336,7 +395,8 @@ struct cmd_params_instance {
     int n_prompt;
     int n_gen;
     int n_batch;
-    bool f32_kv;
+    ggml_type type_k;
+    ggml_type type_v;
     int n_threads;
     int n_gpu_layers;
     int main_gpu;
@@ -365,7 +425,8 @@ struct cmd_params_instance {
 
         cparams.n_ctx = n_prompt + n_gen;
         cparams.n_batch = n_batch;
-        cparams.f16_kv = !f32_kv;
+        cparams.type_k = type_k;
+        cparams.type_v = type_v;
         cparams.mul_mat_q = mul_mat_q;
 
         return cparams;
@@ -380,7 +441,8 @@ static std::vector<cmd_params_instance> get_cmd_params_instances_int(const cmd_p
     for (const auto & mg : params.main_gpu)
     for (const auto & ts : params.tensor_split)
     for (const auto & nb : params.n_batch)
-    for (const auto & fk : params.f32_kv)
+    for (const auto & tk : params.type_k)
+    for (const auto & tv : params.type_v)
     for (const auto & mmq : params.mul_mat_q)
     for (const auto & nt : params.n_threads) {
         cmd_params_instance instance = {
@@ -388,7 +450,8 @@ static std::vector<cmd_params_instance> get_cmd_params_instances_int(const cmd_p
             /* .n_prompt     = */ n_prompt,
             /* .n_gen        = */ n_gen,
             /* .n_batch      = */ nb,
-            /* .f32_kv       = */ fk,
+            /* .type_k       = */ tk,
+            /* .type_v       = */ tv,
             /* .n_threads    = */ nt,
             /* .n_gpu_layers = */ nl,
             /* .main_gpu     = */ mg,
@@ -410,7 +473,8 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
     for (const auto & mg : params.main_gpu)
     for (const auto & ts : params.tensor_split)
     for (const auto & nb : params.n_batch)
-    for (const auto & fk : params.f32_kv)
+    for (const auto & tk : params.type_k)
+    for (const auto & tv : params.type_v)
     for (const auto & mmq : params.mul_mat_q)
     for (const auto & nt : params.n_threads) {
         for (const auto & n_prompt : params.n_prompt) {
@@ -422,7 +486,8 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
                 /* .n_prompt     = */ n_prompt,
                 /* .n_gen        = */ 0,
                 /* .n_batch      = */ nb,
-                /* .f32_kv       = */ fk,
+                /* .type_k       = */ tk,
+                /* .type_v       = */ tv,
                 /* .n_threads    = */ nt,
                 /* .n_gpu_layers = */ nl,
                 /* .main_gpu     = */ mg,
@@ -441,7 +506,8 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
                 /* .n_prompt     = */ 0,
                 /* .n_gen        = */ n_gen,
                 /* .n_batch      = */ nb,
-                /* .f32_kv       = */ fk,
+                /* .type_k       = */ tk,
+                /* .type_v       = */ tv,
                 /* .n_threads    = */ nt,
                 /* .n_gpu_layers = */ nl,
                 /* .main_gpu     = */ mg,
@@ -489,7 +555,8 @@ struct test {
     uint64_t model_n_params;
     int n_batch;
     int n_threads;
-    bool f32_kv;
+    ggml_type type_k;
+    ggml_type type_v;
     int n_gpu_layers;
     int main_gpu;
     bool mul_mat_q;
@@ -508,7 +575,8 @@ struct test {
         model_n_params = llama_model_n_params(lmodel);
         n_batch = inst.n_batch;
         n_threads = inst.n_threads;
-        f32_kv = inst.f32_kv;
+        type_k = inst.type_k;
+        type_v = inst.type_v;
         n_gpu_layers = inst.n_gpu_layers;
         main_gpu = inst.main_gpu;
         mul_mat_q = inst.mul_mat_q;
@@ -571,7 +639,7 @@ struct test {
             "cuda", "opencl", "metal", "gpu_blas", "blas",
             "cpu_info", "gpu_info",
             "model_filename", "model_type", "model_size", "model_n_params",
-            "n_batch", "n_threads", "f16_kv",
+            "n_batch", "n_threads", "type_k", "type_v",
             "n_gpu_layers", "main_gpu", "mul_mat_q", "tensor_split",
             "n_prompt", "n_gen", "test_time",
             "avg_ns", "stddev_ns",
@@ -621,7 +689,7 @@ struct test {
             std::to_string(cuda), std::to_string(opencl), std::to_string(metal), std::to_string(gpu_blas), std::to_string(blas),
             cpu_info, gpu_info,
             model_filename, model_type, std::to_string(model_size), std::to_string(model_n_params),
-            std::to_string(n_batch), std::to_string(n_threads), std::to_string(!f32_kv),
+            std::to_string(n_batch), std::to_string(n_threads), ggml_type_name(type_k), ggml_type_name(type_v),
             std::to_string(n_gpu_layers), std::to_string(main_gpu), std::to_string(mul_mat_q), tensor_split_str,
             std::to_string(n_prompt), std::to_string(n_gen), test_time,
             std::to_string(avg_ns()), std::to_string(stdev_ns()),
@@ -805,8 +873,11 @@ struct markdown_printer : public printer {
         if (params.n_batch.size() > 1 || params.n_batch != cmd_params_defaults.n_batch) {
             fields.push_back("n_batch");
         }
-        if (params.f32_kv.size() > 1 || params.f32_kv != cmd_params_defaults.f32_kv) {
-            fields.push_back("f16_kv");
+        if (params.type_k.size() > 1 || params.type_k != cmd_params_defaults.type_k) {
+            fields.push_back("type_k");
+        }
+        if (params.type_v.size() > 1 || params.type_v != cmd_params_defaults.type_v) {
+            fields.push_back("type_v");
         }
         if (params.main_gpu.size() > 1 || params.main_gpu != cmd_params_defaults.main_gpu) {
             fields.push_back("main_gpu");
diff --git a/examples/llama.swiftui/.gitignore b/examples/llama.swiftui/.gitignore
index 9bce6af39..e585a2a4f 100644
--- a/examples/llama.swiftui/.gitignore
+++ b/examples/llama.swiftui/.gitignore
@@ -1 +1,2 @@
 xcuserdata
+xcshareddata
diff --git a/examples/llama.swiftui/llama.cpp.swift/LibLlama.swift b/examples/llama.swiftui/llama.cpp.swift/LibLlama.swift
index 09b36d9e6..464fb3277 100644
--- a/examples/llama.swiftui/llama.cpp.swift/LibLlama.swift
+++ b/examples/llama.swiftui/llama.cpp.swift/LibLlama.swift
@@ -6,14 +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]
 
-    var n_len: Int32 = 512
+    /// This variable is used to store temporarily invalid cchars
+    private var temporary_invalid_cchars: [CChar]
+
+    var n_len: Int32 = 64
     var n_cur: Int32 = 0
+
     var n_decode: Int32 = 0
 
     init(model: OpaquePointer, context: OpaquePointer) {
@@ -21,28 +41,38 @@ actor LlamaContext {
         self.context = context
         self.tokens_list = []
         self.batch = llama_batch_init(512, 0, 1)
+        self.temporary_invalid_cchars = []
     }
 
     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 {
@@ -53,6 +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;
     }
@@ -61,6 +111,7 @@ actor LlamaContext {
         print("attempting to complete \"\(text)\"")
 
         tokens_list = tokenize(text: text, add_bos: true)
+        temporary_invalid_cchars = []
 
         let n_ctx = llama_n_ctx(context)
         let n_kv_req = tokens_list.count + (Int(n_len) - tokens_list.count)
@@ -72,19 +123,14 @@ actor LlamaContext {
         }
 
         for id in tokens_list {
-            print(token_to_piece(token: id))
+            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
 
@@ -115,25 +161,33 @@ actor LlamaContext {
 
         if new_token_id == llama_token_eos(context) || n_cur == n_len {
             print("\n")
-            return ""
+            let new_token_str = String(cString: temporary_invalid_cchars + [0])
+            temporary_invalid_cchars.removeAll()
+            return new_token_str
         }
 
-        let new_token_str = token_to_piece(token: new_token_id)
+        let new_token_cchars = token_to_piece(token: new_token_id)
+        temporary_invalid_cchars.append(contentsOf: new_token_cchars)
+        let new_token_str: String
+        if let string = String(validatingUTF8: temporary_invalid_cchars + [0]) {
+            temporary_invalid_cchars.removeAll()
+            new_token_str = string
+        } else if (0 ..< temporary_invalid_cchars.count).contains(where: {$0 != 0 && String(validatingUTF8: Array(temporary_invalid_cchars.suffix($0)) + [0]) != nil}) {
+            // in this case, at least the suffix of the temporary_invalid_cchars can be interpreted as UTF8 string
+            let string = String(cString: temporary_invalid_cchars + [0])
+            temporary_invalid_cchars.removeAll()
+            new_token_str = string
+        } else {
+            new_token_str = ""
+        }
         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!")
@@ -142,14 +196,113 @@ 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 n_tokens = text.count + (add_bos ? 1 : 0)
+        let utf8Count = text.utf8.count
+        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(text.count), tokens, Int32(n_tokens), add_bos, false)
+        let tokenCount = llama_tokenize(model, text, Int32(utf8Count), tokens, Int32(n_tokens), add_bos, false)
 
         var swiftTokens: [llama_token] = []
         for i in 0..<tokenCount {
@@ -161,7 +314,8 @@ actor LlamaContext {
         return swiftTokens
     }
 
-    private func token_to_piece(token: llama_token) -> String {
+    /// - note: The result does not contain null-terminator
+    private func token_to_piece(token: llama_token) -> [CChar] {
         let result = UnsafeMutablePointer<Int8>.allocate(capacity: 8)
         result.initialize(repeating: Int8(0), count: 8)
         defer {
@@ -175,10 +329,12 @@ actor LlamaContext {
             defer {
                 newResult.deallocate()
             }
-            _ = llama_token_to_piece(model, token, newResult, -nTokens)
-            return String(cString: newResult)
+            let nNewTokens = llama_token_to_piece(model, token, newResult, -nTokens)
+            let bufferPointer = UnsafeBufferPointer(start: newResult, count: Int(nNewTokens))
+            return Array(bufferPointer)
         } else {
-            return String(cString: result)
+            let bufferPointer = UnsafeBufferPointer(start: result, count: Int(nTokens))
+            return Array(bufferPointer)
         }
     }
 }
diff --git a/examples/llama.swiftui/llama.swiftui.xcodeproj/project.pbxproj b/examples/llama.swiftui/llama.swiftui.xcodeproj/project.pbxproj
index bc1fd15ce..2e6159928 100644
--- a/examples/llama.swiftui/llama.swiftui.xcodeproj/project.pbxproj
+++ b/examples/llama.swiftui/llama.swiftui.xcodeproj/project.pbxproj
@@ -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 */;
 }
diff --git a/examples/llama.swiftui/llama.swiftui/Models/LlamaState.swift b/examples/llama.swiftui/llama.swiftui/Models/LlamaState.swift
index babc60cdc..3393eb242 100644
--- a/examples/llama.swiftui/llama.swiftui/Models/LlamaState.swift
+++ b/examples/llama.swiftui/llama.swiftui/Models/LlamaState.swift
@@ -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 = ""
+    }
 }
diff --git a/examples/llama.swiftui/llama.swiftui/UI/ContentView.swift b/examples/llama.swiftui/llama.swiftui/UI/ContentView.swift
index 0bd16a806..c78f107b3 100644
--- a/examples/llama.swiftui/llama.swiftui/UI/ContentView.swift
+++ b/examples/llama.swiftui/llama.swiftui/UI/ContentView.swift
@@ -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()
+//}
diff --git a/examples/llama.swiftui/llama.swiftui/UI/DownloadButton.swift b/examples/llama.swiftui/llama.swiftui/UI/DownloadButton.swift
new file mode 100644
index 000000000..4bd75cb69
--- /dev/null
+++ b/examples/llama.swiftui/llama.swiftui/UI/DownloadButton.swift
@@ -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"
+//    )
+// }
diff --git a/examples/llava/clip.cpp b/examples/llava/clip.cpp
index fc0656c23..f06ec400d 100644
--- a/examples/llava/clip.cpp
+++ b/examples/llava/clip.cpp
@@ -330,12 +330,6 @@ static ggml_cgraph * clip_image_build_graph(const clip_ctx * ctx, const clip_ima
                               ggml_repeat(ctx0, model.pre_ln_b, embeddings));
     }
 
-    struct ggml_tensor * KQ_scale = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, 1);
-    ggml_allocr_alloc(ctx->alloc, KQ_scale);
-    if (!ggml_allocr_is_measure(ctx->alloc)) {
-        ggml_set_f32(KQ_scale, 1.0f / sqrt((float)d_head));
-    }
-
     // loop over layers
     for (int il = 0; il < n_layer - 1; il++) {
         struct ggml_tensor * cur = embeddings; // embeddings = residual, cur = hidden_states
@@ -356,7 +350,7 @@ static ggml_cgraph * clip_image_build_graph(const clip_ctx * ctx, const clip_ima
             struct ggml_tensor * Q =
                 ggml_add(ctx0, ggml_repeat(ctx0, model.layers[il].q_b, cur), ggml_mul_mat(ctx0, model.layers[il].q_w, cur));
 
-            Q = ggml_scale_inplace(ctx0, Q, KQ_scale);
+            Q = ggml_scale_inplace(ctx0, Q, 1.0f / sqrt((float)d_head));
             Q = ggml_reshape_4d(ctx0, Q, d_head, n_head, num_positions, batch_size);
             Q = ggml_cont(ctx0, ggml_permute(ctx0, Q, 0, 2, 1, 3));
             Q = ggml_reshape_3d(ctx0, Q, d_head, num_positions, n_head * batch_size);
@@ -514,7 +508,7 @@ 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);
             }
         }
     }
@@ -739,7 +733,7 @@ bool clip_image_preprocess(const clip_ctx * ctx, const clip_image_u8 * img, clip
         temp->ny = longer_side;
         temp->size = 3 * longer_side * longer_side;
         temp->data = new uint8_t[temp->size]();
-        uint8_t bc[3] = {122, 116, 104}; // bakground color in RGB from LLaVA
+        uint8_t bc[3] = {122, 116, 104}; // background color in RGB from LLaVA
 
         // fill with background color
         for (size_t i = 0; i < temp->size; i++) {
@@ -962,7 +956,7 @@ 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 +1029,7 @@ 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);
     }
 
diff --git a/examples/llava/convert-image-encoder-to-gguf.py b/examples/llava/convert-image-encoder-to-gguf.py
index 729aaef8f..03688e0ea 100644
--- a/examples/llava/convert-image-encoder-to-gguf.py
+++ b/examples/llava/convert-image-encoder-to-gguf.py
@@ -51,7 +51,7 @@ def bytes_to_unicode():
     The reversible bpe codes work on unicode strings.
     This means you need a large # of unicode characters in your vocab if you want to avoid UNKs.
     When you're at something like a 10B token dataset you end up needing around 5K for decent coverage.
-    This is a signficant percentage of your normal, say, 32K bpe vocab.
+    This is a significant percentage of your normal, say, 32K bpe vocab.
     To avoid that, we want lookup tables between utf-8 bytes and unicode strings.
     And avoids mapping to whitespace/control characters the bpe code barfs on.
     """
diff --git a/examples/lookahead/README.md b/examples/lookahead/README.md
index 252a6689e..a69a471b4 100644
--- a/examples/lookahead/README.md
+++ b/examples/lookahead/README.md
@@ -1,6 +1,6 @@
 # llama.cpp/examples/lookahead
 
-Demonstartion of lookahead decoding technique:
+Demonstration of lookahead decoding technique:
 
 https://lmsys.org/blog/2023-11-21-lookahead-decoding/
 
diff --git a/examples/lookup/CMakeLists.txt b/examples/lookup/CMakeLists.txt
new file mode 100644
index 000000000..c060b8f56
--- /dev/null
+++ b/examples/lookup/CMakeLists.txt
@@ -0,0 +1,5 @@
+set(TARGET lookup)
+add_executable(${TARGET} lookup.cpp)
+install(TARGETS ${TARGET} RUNTIME)
+target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_11)
diff --git a/examples/lookup/README.md b/examples/lookup/README.md
new file mode 100644
index 000000000..5bfb0de93
--- /dev/null
+++ b/examples/lookup/README.md
@@ -0,0 +1,13 @@
+# llama.cpp/examples/lookup
+
+Demonstration of Prompt Lookup Decoding
+
+https://github.com/apoorvumang/prompt-lookup-decoding
+
+The key parameters for lookup decoding are `ngram_min`, `ngram_max` and `n_draft`. The first two determine the size of the ngrams to search for in the prompt for a match. The latter specifies how many subsequent tokens to draft if a match is found.
+
+More info:
+
+https://github.com/ggerganov/llama.cpp/pull/4484
+https://github.com/ggerganov/llama.cpp/issues/4226
+
diff --git a/examples/lookup/lookup.cpp b/examples/lookup/lookup.cpp
new file mode 100644
index 000000000..d8de7dd38
--- /dev/null
+++ b/examples/lookup/lookup.cpp
@@ -0,0 +1,230 @@
+#include "common.h"
+#include "llama.h"
+
+#include <cmath>
+#include <cstdio>
+#include <string>
+#include <vector>
+
+int main(int argc, char ** argv){
+    gpt_params params;
+
+    if (!gpt_params_parse(argc, argv, params)) {
+        return 1;
+    }
+
+    // max/min n-grams size to search for in prompt
+    const int ngram_max = 4;
+    const int ngram_min = 1;
+
+    // length of the candidate / draft sequence, if match is found
+    const int n_draft = params.n_draft;
+
+    const bool dump_kv_cache = params.dump_kv_cache;
+
+#ifndef LOG_DISABLE_LOGS
+    log_set_target(log_filename_generator("lookup", "log"));
+    LOG_TEE("Log start\n");
+    log_dump_cmdline(argc, argv);
+#endif // LOG_DISABLE_LOGS
+
+    // init llama.cpp
+    llama_backend_init(params.numa);
+
+    llama_model * model = NULL;
+    llama_context * ctx = NULL;
+
+    // load the model
+    std::tie(model, ctx) = llama_init_from_gpt_params(params);
+
+    // tokenize the prompt
+    const bool add_bos = llama_should_add_bos_token(model);
+    LOG("add_bos tgt: %d\n", add_bos);
+
+    std::vector<llama_token> inp;
+    inp = ::llama_tokenize(ctx, params.prompt, add_bos, true);
+
+    const int max_context_size     = llama_n_ctx(ctx);
+    const int max_tokens_list_size = max_context_size - 4;
+
+    if ((int) inp.size() > max_tokens_list_size) {
+        fprintf(stderr, "%s: error: prompt too long (%d tokens, max %d)\n", __func__, (int) inp.size(), max_tokens_list_size);
+        return 1;
+    }
+
+    fprintf(stderr, "\n\n");
+
+    for (auto id : inp) {
+        fprintf(stderr, "%s", llama_token_to_piece(ctx, id).c_str());
+    }
+
+    fflush(stderr);
+
+    const int n_input = inp.size();
+
+    const auto t_enc_start = ggml_time_us();
+
+    llama_decode(ctx, llama_batch_get_one( inp.data(), n_input - 1, 0,           0));
+    llama_decode(ctx, llama_batch_get_one(&inp.back(),           1, n_input - 1, 0));
+
+    const auto t_enc_end = ggml_time_us();
+
+    int n_predict = 0;
+    int n_drafted = 0;
+    int n_accept  = 0;
+
+    int n_past = inp.size();
+
+    bool has_eos = false;
+
+    struct llama_sampling_context * ctx_sampling = llama_sampling_init(params.sparams);
+
+    std::vector<llama_token> draft;
+
+    llama_batch batch_tgt = llama_batch_init(params.n_ctx, 0, 1);
+
+    // debug
+    struct llama_kv_cache_view kvc_view = llama_kv_cache_view_init(ctx, 1);
+
+    const auto t_dec_start = ggml_time_us();
+
+    while (true) {
+        // debug
+        if (dump_kv_cache) {
+            llama_kv_cache_view_update(ctx, &kvc_view);
+            dump_kv_cache_view_seqs(kvc_view, 40);
+        }
+
+        // print current draft sequence
+        LOG("drafted %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, draft).c_str());
+
+        int i_dft = 0;
+        while (true) {
+            // sample from the target model
+            llama_token id = llama_sampling_sample(ctx_sampling, ctx, NULL, i_dft);
+
+            llama_sampling_accept(ctx_sampling, ctx, id, true);
+
+            const std::string token_str = llama_token_to_piece(ctx, id);
+
+            if (!params.use_color) {
+                printf("%s", token_str.c_str());
+            }
+
+            if (id == llama_token_eos(model)) {
+                has_eos = true;
+            }
+
+            ++n_predict;
+
+            // check if the target token matches the draft
+            if (i_dft < (int) draft.size() && id == draft[i_dft]) {
+                LOG("the sampled target token matches the %dth drafted token (%d, '%s') - accepted\n", i_dft, id, token_str.c_str());
+                ++n_accept;
+                ++n_past;
+                ++i_dft;
+                inp.push_back(id);
+
+                if (params.use_color) {
+                    // color accepted draft token
+                    printf("\033[34m%s\033[0m", token_str.c_str());
+                    fflush(stdout);
+                }
+                continue;
+            }
+
+            if (params.use_color) {
+                printf("%s", token_str.c_str());
+            }
+            fflush(stdout);
+
+
+            LOG("the sampled target token (%d, '%s') did not match, or we ran out of drafted tokens\n", id, token_str.c_str());
+
+            draft.clear();
+            draft.push_back(id);
+            inp.push_back(id);
+            break;
+        }
+
+        if ((params.n_predict > 0 && n_predict > params.n_predict) || has_eos) {
+            break;
+        }
+
+        // KV cache management
+        // clean the cache of draft tokens that weren't accepted
+        llama_kv_cache_seq_rm(ctx, 0, n_past, -1);
+
+        llama_batch_clear(batch_tgt);
+        llama_batch_add(batch_tgt, draft[0], n_past, { 0 }, true);
+
+        // generate n_pred tokens through prompt lookup
+        auto prompt_lookup = [&]() -> void {
+            int inp_size = inp.size();
+            for (int ngram_size = ngram_max ; ngram_size > ngram_min; --ngram_size){
+                const llama_token * ngram = &inp[inp_size - ngram_size];
+
+                for (int i = 0; i <= (int) inp_size - (ngram_size * 2); ++i) {
+                    bool match = true;
+                    for (int j = 0; j < ngram_size; ++j) {
+                        if (inp[i + j] != ngram[j]) {
+                            match = false;
+                            break;
+                        }
+                    }
+
+                    if (match) {
+                        const int startIdx = i + ngram_size;
+                        const int endIdx = startIdx + n_draft;
+                        if (endIdx < inp_size) {
+                            for (int j = startIdx; j < endIdx; ++j) {
+                                LOG(" - draft candidate %d: %d\n", j, inp[j]);
+                                draft.push_back(inp[j]);
+                                llama_batch_add(batch_tgt, inp[j], n_past + (j - startIdx) + 1, { 0 }, true);
+                                ++n_drafted;
+                            }
+                            return;
+                        }
+                    }
+                }
+            }
+            return;
+        };
+
+        prompt_lookup();
+
+        llama_decode(ctx, batch_tgt);
+        ++n_past;
+
+        draft.erase(draft.begin());
+    }
+
+    auto t_dec_end = ggml_time_us();
+
+    LOG_TEE("\n\n");
+
+    LOG_TEE("encoded %4d tokens in %8.3f seconds, speed: %8.3f t/s\n", n_input,   (t_enc_end - t_enc_start) / 1e6f, inp.size() / ((t_enc_end - t_enc_start) / 1e6f));
+    LOG_TEE("decoded %4d tokens in %8.3f seconds, speed: %8.3f t/s\n", n_predict, (t_dec_end - t_dec_start) / 1e6f, n_predict  / ((t_dec_end - t_dec_start) / 1e6f));
+
+    LOG_TEE("\n");
+    LOG_TEE("n_draft   = %d\n", n_draft);
+    LOG_TEE("n_predict = %d\n", n_predict);
+    LOG_TEE("n_drafted = %d\n", n_drafted);
+    LOG_TEE("n_accept  = %d\n", n_accept);
+    LOG_TEE("accept    = %.3f%%\n", 100.0f * n_accept / n_drafted);
+
+    LOG_TEE("\ntarget:\n");
+    llama_print_timings(ctx);
+
+    llama_sampling_free(ctx_sampling);
+    llama_batch_free(batch_tgt);
+
+    llama_free(ctx);
+    llama_free_model(model);
+
+    llama_backend_free();
+
+    fprintf(stderr, "\n\n");
+
+    return 0;
+}
diff --git a/examples/main/main.cpp b/examples/main/main.cpp
index c5cdfbf21..c096f110b 100644
--- a/examples/main/main.cpp
+++ b/examples/main/main.cpp
@@ -437,6 +437,7 @@ int main(int argc, char ** argv) {
         }
     }
     LOG_TEE("sampling: \n%s\n", llama_sampling_print(sparams).c_str());
+    LOG_TEE("sampling order: \n%s\n", llama_sampling_order_print(sparams).c_str());
     LOG_TEE("generate: n_ctx = %d, n_batch = %d, n_predict = %d, n_keep = %d\n", n_ctx, params.n_batch, params.n_predict, params.n_keep);
     LOG_TEE("\n\n");
 
diff --git a/examples/quantize-stats/quantize-stats.cpp b/examples/quantize-stats/quantize-stats.cpp
index 271282477..773024160 100644
--- a/examples/quantize-stats/quantize-stats.cpp
+++ b/examples/quantize-stats/quantize-stats.cpp
@@ -321,7 +321,6 @@ int main(int argc, char ** argv) {
         auto cparams = llama_context_default_params();
         cparams.n_ctx      = 256;
         cparams.seed       = 1;
-        cparams.f16_kv     = false;
 
         ctx = llama_new_context_with_model(model, cparams);
 
diff --git a/examples/server/README.md b/examples/server/README.md
index cfc220f58..0751b9612 100644
--- a/examples/server/README.md
+++ b/examples/server/README.md
@@ -222,7 +222,7 @@ node index.js
 
     `content`: Set the text to process.
 
-    **POST** `/infill`: For code infilling. Takes a prefix and a suffix and returns the predicted completion as stream.
+-   **POST** `/infill`: For code infilling. Takes a prefix and a suffix and returns the predicted completion as stream.
 
     *Options:*
 
diff --git a/examples/server/json.hpp b/examples/server/json.hpp
index 4d1a37ad7..ea945f346 100644
--- a/examples/server/json.hpp
+++ b/examples/server/json.hpp
@@ -11227,7 +11227,7 @@ class binary_reader
                 }
                 if (is_ndarray) // ndarray dimensional vector can only contain integers, and can not embed another array
                 {
-                    return sax->parse_error(chars_read, get_token_string(), parse_error::create(113, chars_read, exception_message(input_format, "ndarray dimentional vector is not allowed", "size"), nullptr));
+                    return sax->parse_error(chars_read, get_token_string(), parse_error::create(113, chars_read, exception_message(input_format, "ndarray dimensional vector is not allowed", "size"), nullptr));
                 }
                 std::vector<size_t> dim;
                 if (JSON_HEDLEY_UNLIKELY(!get_ubjson_ndarray_size(dim)))
diff --git a/examples/server/public/completion.js b/examples/server/public/completion.js
index b9c442509..6e2b99565 100644
--- a/examples/server/public/completion.js
+++ b/examples/server/public/completion.js
@@ -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,
   });
@@ -114,7 +115,7 @@ export async function* llama(prompt, params = {}, config = {}) {
   return content;
 }
 
-// Call llama, return an event target that you can subcribe to
+// Call llama, return an event target that you can subscribe to
 //
 // Example:
 //
diff --git a/examples/server/public/index.html b/examples/server/public/index.html
index 175c52478..07d779d20 100644
--- a/examples/server/public/index.html
+++ b/examples/server/public/index.html
@@ -223,7 +223,7 @@
       repeat_last_n: 256, // 0 = disable penalty, -1 = context size
       repeat_penalty: 1.18, // 1.0 = disabled
       top_k: 40, // <= 0 to use vocab size
-      top_p: 0.5, // 1.0 = disabled
+      top_p: 0.95, // 1.0 = disabled
       min_p: 0.05, // 0 = disabled
       tfs_z: 1.0, // 1.0 = disabled
       typical_p: 1.0, // 1.0 = disabled
@@ -235,10 +235,11 @@
       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 borwser LocalStorage */
+    /* START: Support for storing prompt templates and parameters in browsers LocalStorage */
 
     const local_storage_storageKey = "llamacpp_server_local_storage";
 
@@ -282,7 +283,7 @@
     let importedTemplates = local_storage_getDataAsObject('user_templates')
 
     if (importedTemplates) {
-      // saved templates were successfuly imported.
+      // saved templates were successfully imported.
 
       console.log('Processing saved templates and updating default template')
       params.value = { ...params.value, image_data: [] };
@@ -303,7 +304,7 @@
     }
 
     function userTemplateResetToDefault() {
-      console.log('Reseting themplate to default')
+      console.log('Resetting template to default')
       selectedUserTemplate.value.name = 'default';
       selectedUserTemplate.value.data = savedUserTemplates.value['default'];
     }
@@ -762,7 +763,7 @@
 
           <fieldset class="two">
             ${IntField({ label: "Predictions", max: 2048, min: -1, name: "n_predict", value: params.value.n_predict })}
-            ${FloatField({ label: "Temperature", max: 1.5, min: 0.0, name: "temperature", step: 0.01, value: params.value.temperature })}
+            ${FloatField({ label: "Temperature", max: 2.0, min: 0.0, name: "temperature", step: 0.01, value: params.value.temperature })}
             ${FloatField({ label: "Penalize repeat sequence", max: 2.0, min: 0.0, name: "repeat_penalty", step: 0.01, value: params.value.repeat_penalty })}
             ${IntField({ label: "Consider N tokens for penalize", max: 2048, min: 0, name: "repeat_last_n", value: params.value.repeat_last_n })}
             ${IntField({ label: "Top-K sampling", max: 100, min: -1, name: "top_k", value: params.value.top_k })}
@@ -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>
       `
diff --git a/examples/server/server.cpp b/examples/server/server.cpp
index 911f7bbe1..04038530f 100644
--- a/examples/server/server.cpp
+++ b/examples/server/server.cpp
@@ -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)
@@ -2108,10 +2119,6 @@ static void server_params_parse(int argc, char **argv, server_params &sparams,
             }
             params.yarn_beta_slow = std::stof(argv[i]);
         }
-        else if (arg == "--memory-f32" || arg == "--memory_f32")
-        {
-            params.memory_f16 = false;
-        }
         else if (arg == "--threads" || arg == "-t")
         {
             if (++i >= argc)
@@ -2386,7 +2393,9 @@ json oaicompat_completion_params_parse(
     llama_params["__oaicompat"] = true;
 
     // Map OpenAI parameters to llama.cpp parameters
+    llama_params["model"]             = json_value(body, "model", std::string("uknown"));
     llama_params["prompt"]            = format_chatml(body["messages"]); // OpenAI 'messages' to llama.cpp 'prompt'
+    llama_params["cache_prompt"]      = json_value(body, "cache_prompt", false);
     llama_params["temperature"]       = json_value(body, "temperature", 0.8);
     llama_params["top_k"]             = json_value(body, "top_k", 40);
     llama_params["top_p"]             = json_value(body, "top_p", 0.95);
@@ -2405,7 +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());
     }
 
@@ -2636,6 +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;
@@ -2672,6 +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"}});
@@ -2679,28 +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;
             });
 
@@ -2711,23 +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 {
@@ -2798,12 +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);
@@ -2817,10 +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 {
@@ -2872,8 +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)) {
@@ -2881,12 +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 {
@@ -2935,11 +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)
             {
@@ -2950,7 +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)
@@ -2964,7 +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)
@@ -2981,7 +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);
@@ -3002,19 +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;
                 }
             });
@@ -3035,11 +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([&]()
             {
diff --git a/examples/simple/simple.cpp b/examples/simple/simple.cpp
index 374aef6f1..9cfde8308 100644
--- a/examples/simple/simple.cpp
+++ b/examples/simple/simple.cpp
@@ -75,7 +75,7 @@ int main(int argc, char ** argv) {
     // make sure the KV cache is big enough to hold all the prompt and generated tokens
     if (n_kv_req > n_ctx) {
         LOG_TEE("%s: error: n_kv_req > n_ctx, the required KV cache size is not big enough\n", __func__);
-        LOG_TEE("%s:        either reduce n_parallel or increase n_ctx\n", __func__);
+        LOG_TEE("%s:        either reduce n_len or increase n_ctx\n", __func__);
         return 1;
     }
 
diff --git a/examples/speculative/README.md b/examples/speculative/README.md
index d88fd3790..814efa592 100644
--- a/examples/speculative/README.md
+++ b/examples/speculative/README.md
@@ -1,6 +1,6 @@
 # llama.cpp/examples/speculative
 
-Demonstartion of speculative decoding and tree-based speculative decoding techniques
+Demonstration of speculative decoding and tree-based speculative decoding techniques
 
 More info:
 
diff --git a/examples/speculative/speculative.cpp b/examples/speculative/speculative.cpp
index ace755c51..20f1fb5bf 100644
--- a/examples/speculative/speculative.cpp
+++ b/examples/speculative/speculative.cpp
@@ -203,8 +203,9 @@ int main(int argc, char ** argv) {
 
             const std::string token_str = llama_token_to_piece(ctx_tgt, id);
 
-            printf("%s", token_str.c_str());
-            fflush(stdout);
+            if (!params.use_color) {
+                printf("%s", token_str.c_str());
+            }
 
             if (id == llama_token_eos(model_tgt)) {
                 has_eos = true;
@@ -236,10 +237,18 @@ int main(int argc, char ** argv) {
                     ++n_past_tgt;
                     ++n_past_dft;
                     ++i_dft;
-
+                    if (params.use_color) {
+                        // Color token according to its origin sequence
+                        printf("\u001b[%dm%s\u001b[37m", (36 - s_keep % 6), token_str.c_str());
+                        fflush(stdout);
+                    }
                     continue;
                 }
             }
+            if (params.use_color) {
+                printf("%s", token_str.c_str());
+            }
+            fflush(stdout);
 
             LOG("the sampled target token (%d, '%s') did not match, or we ran out of drafted tokens\n", id, token_str.c_str());
 
@@ -419,7 +428,7 @@ int main(int argc, char ** argv) {
             ++n_past_tgt;
         }
 
-        // the first token is always proposed by the traget model before the speculation loop so we erase it here
+        // the first token is always proposed by the target model before the speculation loop so we erase it here
         for (int s = 0; s < n_seq_dft; ++s) {
             if (!drafts[s].active) {
                 continue;
diff --git a/examples/train-text-from-scratch/train-text-from-scratch.cpp b/examples/train-text-from-scratch/train-text-from-scratch.cpp
index f049a3923..4a9a2340b 100644
--- a/examples/train-text-from-scratch/train-text-from-scratch.cpp
+++ b/examples/train-text-from-scratch/train-text-from-scratch.cpp
@@ -369,10 +369,7 @@ static struct ggml_tensor * llama_build_train_graphs(
     checkpoints.push_back(t00);
     checkpoints.push_back(t01);
 
-    struct ggml_tensor * kv_scale = NULL;
-    if (!enable_flash_attn) {
-        kv_scale = ggml_new_f32(ctx, 1.0f/sqrtf(float(n_embd)/n_head));
-    }
+    const float kv_scale = 1.0f/sqrtf(float(n_embd)/n_head);
 
     for (int il = 0; il < n_layer; ++il) {
         struct my_llama_layer & layer = model->layers[il];
@@ -444,14 +441,13 @@ static struct ggml_tensor * llama_build_train_graphs(
         // make sure some tensors are not reallocated by inserting new temporary nodes depending on them
         int n_leafs_before = gb->n_leafs;
         int n_nodes_before = gb->n_nodes;
-        struct ggml_tensor * one = ggml_new_f32(ctx, 1.0f);
         // output tensors
-        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, t35, one));
-        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, t36, one));
+        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, t35, 1.0f));
+        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, t36, 1.0f));
         // input gradient
-        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, t36->grad, one));
+        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, t36->grad, 1.0f));
         // KQ_pos
-        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, KQ_pos, one));
+        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, KQ_pos, 1.0f));
         GGML_ASSERT(t36->grad->data == NULL && t36->grad->view_src == NULL);
 
         ggml_allocr_alloc(alloc, t36->grad);
@@ -1295,10 +1291,6 @@ int main(int argc, char ** argv) {
         opt_cb_data.last_save_iter = opt->iter;
     }
 
-    if (alloc) {
-        ggml_allocr_free(alloc);
-    }
-
     ggml_free(opt->ctx);
     free_train_state(train);
     ggml_free(model.ctx);
diff --git a/ggml-alloc.c b/ggml-alloc.c
index 0d4e12ae9..a27dd54b0 100644
--- a/ggml-alloc.c
+++ b/ggml-alloc.c
@@ -72,7 +72,7 @@ static void remove_allocated_tensor(ggml_tallocr_t alloc, struct ggml_tensor * t
 
 // check if a tensor is allocated by this buffer
 static bool ggml_tallocr_is_own(ggml_tallocr_t alloc, const struct ggml_tensor * tensor) {
-    return tensor->buffer == alloc->buffer;
+    return tensor->buffer == alloc->buffer && (!tensor->view_src || tensor->view_src->buffer == alloc->buffer);
 }
 
 static bool ggml_is_view(struct ggml_tensor * t) {
@@ -168,10 +168,6 @@ static void ggml_tallocr_free_tensor(ggml_tallocr_t alloc, struct ggml_tensor *
     size = aligned_offset(NULL, size, alloc->alignment);
     AT_PRINTF("%s: freeing %s at %p (%zu bytes) - n_free_blocks = %d\n", __func__, tensor->name, ptr, size, alloc->n_free_blocks);
 
-    if (!alloc->measure) {
-        ggml_backend_buffer_free_tensor(alloc->buffer, tensor);
-    }
-
 #ifdef GGML_ALLOCATOR_DEBUG
     remove_allocated_tensor(alloc, tensor);
 #endif
@@ -237,7 +233,7 @@ void ggml_tallocr_reset(ggml_tallocr_t alloc) {
 }
 
 ggml_tallocr_t ggml_tallocr_new(void * data, size_t size, size_t alignment) {
-    struct ggml_backend_buffer * buffer = ggml_backend_cpu_buffer_from_ptr(NULL, data, size);
+    struct ggml_backend_buffer * buffer = ggml_backend_cpu_buffer_from_ptr(data, size);
 
     ggml_tallocr_t alloc = (ggml_tallocr_t)malloc(sizeof(struct ggml_tallocr));
 
@@ -449,17 +445,15 @@ static ggml_tallocr_t node_tallocr(ggml_gallocr_t galloc, struct ggml_tensor * n
 static void init_view(ggml_gallocr_t galloc, struct ggml_tensor * view, bool update_backend) {
     ggml_tallocr_t alloc = node_tallocr(galloc, view);
 
-    //printf("init_view: %s from src %s\n", view->name, view->view_src->name);
     GGML_ASSERT(view->view_src != NULL && view->view_src->data != NULL);
     if (update_backend) {
         view->backend = view->view_src->backend;
     }
-    view->buffer  = view->view_src->buffer;
+    // views are initialized in the alloc buffer rather than the view_src buffer
+    view->buffer  = alloc->buffer;
     view->data    = (char *)view->view_src->data + view->view_offs;
 
-    // FIXME: the view should be initialized by the owning buffer, but currently this breaks the CUDA backend
-    // due to the ggml_tensor_extra_gpu ring buffer overwriting the KV cache extras
-    assert(ggml_tallocr_is_measure(alloc) || !view->buffer || view->buffer->backend == alloc->buffer->backend);
+    assert(ggml_tallocr_is_measure(alloc) || !view->buffer || view->buffer->buft == alloc->buffer->buft);
 
     if (!alloc->measure) {
         ggml_backend_buffer_init_tensor(alloc->buffer, view);
@@ -741,6 +735,10 @@ void ggml_allocr_set_parse_seq(ggml_allocr_t alloc, const int * list, int n) {
 }
 
 void ggml_allocr_free(ggml_allocr_t alloc) {
+    if (alloc == NULL) {
+        return;
+    }
+
     ggml_gallocr_free(alloc->galloc);
     ggml_tallocr_free(alloc->talloc);
     free(alloc);
@@ -765,3 +763,48 @@ size_t ggml_allocr_max_size(ggml_allocr_t alloc) {
 size_t ggml_allocr_alloc_graph(ggml_allocr_t alloc, struct ggml_cgraph * graph) {
     return ggml_gallocr_alloc_graph(alloc->galloc, alloc->talloc, graph);
 }
+
+// utils
+ggml_backend_buffer_t ggml_backend_alloc_ctx_tensors_from_buft(struct ggml_context * ctx, ggml_backend_buffer_type_t buft) {
+    GGML_ASSERT(ggml_get_no_alloc(ctx) == true);
+
+    size_t alignment = ggml_backend_buft_get_alignment(buft);
+
+    size_t nbytes = 0;
+    for (struct ggml_tensor * t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) {
+        if (t->data == NULL && t->view_src == NULL) {
+            nbytes += GGML_PAD(ggml_backend_buft_get_alloc_size(buft, t), alignment);
+        }
+    }
+
+    if (nbytes == 0) {
+        // all the tensors in the context are already allocated
+        return NULL;
+    }
+
+    ggml_backend_buffer_t buffer = ggml_backend_buft_alloc_buffer(buft, nbytes);
+    ggml_tallocr_t tallocr = ggml_tallocr_new_from_buffer(buffer);
+
+    for (struct ggml_tensor * t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) {
+        if (t->data == NULL) {
+            if (t->view_src == NULL) {
+                ggml_tallocr_alloc(tallocr, t);
+            } else {
+                ggml_backend_view_init(buffer, t);
+            }
+        } else {
+            if (t->view_src != NULL) {
+                // view of a pre-allocated tensor
+                ggml_backend_view_init(buffer, t);
+            }
+        }
+    }
+
+    ggml_tallocr_free(tallocr);
+
+    return buffer;
+}
+
+ggml_backend_buffer_t ggml_backend_alloc_ctx_tensors(struct ggml_context * ctx, ggml_backend_t backend) {
+    return ggml_backend_alloc_ctx_tensors_from_buft(ctx, ggml_backend_get_default_buffer_type(backend));
+}
diff --git a/ggml-alloc.h b/ggml-alloc.h
index dde2a06bf..64a412468 100644
--- a/ggml-alloc.h
+++ b/ggml-alloc.h
@@ -8,6 +8,7 @@ extern "C" {
 
 struct ggml_backend;
 struct ggml_backend_buffer;
+struct ggml_backend_buffer_type;
 
 //
 // Legacy API
@@ -42,7 +43,7 @@ GGML_API size_t ggml_allocr_alloc_graph(ggml_allocr_t alloc, struct ggml_cgraph
 // ggml-backend v2 API
 //
 
-// Seperate tensor and graph allocator objects
+// Separate tensor and graph allocator objects
 // This is necessary for multi-backend allocation because the graph allocator needs to use multiple tensor allocators
 // The original API is kept as a wrapper around the new API
 
@@ -80,6 +81,12 @@ GGML_API void   ggml_gallocr_alloc_graph_n(
                     struct ggml_hash_set hash_set,
                     ggml_tallocr_t * hash_node_talloc);
 
+
+// Utils
+// Create a buffer and allocate all the tensors in a ggml_context
+GGML_API struct ggml_backend_buffer * ggml_backend_alloc_ctx_tensors_from_buft(struct ggml_context * ctx, struct ggml_backend_buffer_type * buft);
+GGML_API struct ggml_backend_buffer * ggml_backend_alloc_ctx_tensors(struct ggml_context * ctx, struct ggml_backend * backend);
+
 #ifdef  __cplusplus
 }
 #endif
diff --git a/ggml-backend-impl.h b/ggml-backend-impl.h
index 211e3d424..05859935a 100644
--- a/ggml-backend-impl.h
+++ b/ggml-backend-impl.h
@@ -12,31 +12,54 @@ extern "C" {
     // Backend buffer
     //
 
+    // buffer type
+    typedef void * ggml_backend_buffer_type_context_t;
+
+    struct ggml_backend_buffer_type_i {
+        ggml_backend_buffer_t (*alloc_buffer)    (ggml_backend_buffer_type_t buft, size_t size);
+        size_t                (*get_alignment)   (ggml_backend_buffer_type_t buft); // tensor alignment
+        size_t                (*get_alloc_size)  (ggml_backend_buffer_type_t buft, struct ggml_tensor * tensor); // data size needed to allocate the tensor, including padding
+        bool                  (*supports_backend)(ggml_backend_buffer_type_t buft, ggml_backend_t backend); // check if the buffer type is usable by the backend
+        // check if tensor data is in host memory
+        // should be equivalent to supports_backend(buft, ggml_backend_cpu_init())
+        bool                  (*is_host)         (ggml_backend_buffer_type_t buft);
+    };
+
+    struct ggml_backend_buffer_type {
+        struct ggml_backend_buffer_type_i  iface;
+        ggml_backend_buffer_type_context_t context;
+    };
+
+    // buffer
     typedef void * ggml_backend_buffer_context_t;
 
     struct ggml_backend_buffer_i {
-        void   (*free_buffer)   (ggml_backend_buffer_t buffer);
-        void * (*get_base)      (ggml_backend_buffer_t buffer); // get base pointer
-        size_t (*get_alloc_size)(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor); // pre-allocation callback
-        void   (*init_tensor)   (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor); // post-allocation callback
-        void   (*free_tensor)   (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor); // pre-free callback
+        void   (*free_buffer)    (ggml_backend_buffer_t buffer);
+        //void     (*reset)      (ggml_backend_buffer_t buffer); // reset any internal state due to tensor initialization, such as tensor extras
+        void * (*get_base)       (ggml_backend_buffer_t buffer);
+        void   (*init_tensor)    (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
+        void   (*set_tensor)     (ggml_backend_buffer_t buffer,       struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
+        void   (*get_tensor)     (ggml_backend_buffer_t buffer, const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
+        // (optional) copy tensor between different buffer-type, allow for single-copy tranfers
+        void   (*cpy_tensor_from)(ggml_backend_buffer_t buffer, struct ggml_tensor * src, struct ggml_tensor * dst);
+        void   (*cpy_tensor_to)  (ggml_backend_buffer_t buffer, struct ggml_tensor * src, struct ggml_tensor * dst);
+        void   (*clear)          (ggml_backend_buffer_t buffer, uint8_t value);
     };
 
     struct ggml_backend_buffer {
-        struct ggml_backend_buffer_i iface;
-
-        ggml_backend_t                backend;
+        struct ggml_backend_buffer_i  iface;
+        ggml_backend_buffer_type_t    buft;
         ggml_backend_buffer_context_t context;
-
         size_t size;
     };
 
-    GGML_API ggml_backend_buffer_t ggml_backend_buffer_init(
-            struct ggml_backend                  * backend,
+    ggml_backend_buffer_t ggml_backend_buffer_init(
+                   ggml_backend_buffer_type_t      buft,
             struct ggml_backend_buffer_i           iface,
                    ggml_backend_buffer_context_t   context,
                    size_t                          size);
 
+
     //
     // Backend
     //
@@ -49,20 +72,17 @@ extern "C" {
         void (*free)(ggml_backend_t backend);
 
         // buffer allocation
-        ggml_backend_buffer_t (*alloc_buffer)(ggml_backend_t backend, size_t size);
+        ggml_backend_buffer_type_t (*get_default_buffer_type)(ggml_backend_t backend);
 
-        // get buffer alignment
-        size_t (*get_alignment)(ggml_backend_t backend);
-
-        // tensor data access
-        // these functions can be asynchronous, helper functions are provided for synchronous access that automatically call synchronize
+        // (optional) asynchroneous tensor data access
         void (*set_tensor_async)(ggml_backend_t backend,       struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
         void (*get_tensor_async)(ggml_backend_t backend, const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
-        void (*synchronize)     (ggml_backend_t backend);
 
-        // (optional) copy tensor between different backends, allow for single-copy tranfers
-        void (*cpy_tensor_from)(ggml_backend_t backend, struct ggml_tensor * src, struct ggml_tensor * dst);
-        void (*cpy_tensor_to)  (ggml_backend_t backend, struct ggml_tensor * src, struct ggml_tensor * dst);
+        // (optional) asynchroneous tensor copy
+        void (*cpy_tensor_from_async)(ggml_backend_t backend, struct ggml_tensor * src, struct ggml_tensor * dst);
+        void (*cpy_tensor_to_async)  (ggml_backend_t backend, struct ggml_tensor * src, struct ggml_tensor * dst);
+
+        void (*synchronize)(ggml_backend_t backend);
 
         // compute graph with a plan
         ggml_backend_graph_plan_t (*graph_plan_create) (ggml_backend_t backend, struct ggml_cgraph * cgraph);
@@ -82,6 +102,15 @@ extern "C" {
         ggml_backend_context_t context;
     };
 
+
+    //
+    // Backend registry
+    //
+
+    typedef ggml_backend_t (*ggml_backend_init_fn)(const char * params, void * user_data);
+
+    void ggml_backend_register(const char * name, ggml_backend_init_fn init_fn, ggml_backend_buffer_type_t default_buffer_type, void * user_data);
+
 #ifdef  __cplusplus
 }
 #endif
diff --git a/ggml-backend.c b/ggml-backend.c
index f6e5fceed..0c8c9ec43 100644
--- a/ggml-backend.c
+++ b/ggml-backend.c
@@ -9,14 +9,43 @@
 #include <stdlib.h>
 #include <string.h>
 
-#define UNUSED GGML_UNUSED
 
 #define MAX(a, b) ((a) > (b) ? (a) : (b))
 
+
+// backend buffer type
+
+ggml_backend_buffer_t ggml_backend_buft_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
+    return buft->iface.alloc_buffer(buft, size);
+}
+
+size_t ggml_backend_buft_get_alignment(ggml_backend_buffer_type_t buft) {
+    return buft->iface.get_alignment(buft);
+}
+
+size_t ggml_backend_buft_get_alloc_size(ggml_backend_buffer_type_t buft, struct ggml_tensor * tensor) {
+    // get_alloc_size is optional, defaults to ggml_nbytes
+    if (buft->iface.get_alloc_size) {
+        return buft->iface.get_alloc_size(buft, tensor);
+    }
+    return ggml_nbytes(tensor);
+}
+
+bool ggml_backend_buft_supports_backend(ggml_backend_buffer_type_t buft, ggml_backend_t backend) {
+    return buft->iface.supports_backend(buft, backend);
+}
+
+bool ggml_backend_buft_is_host(ggml_backend_buffer_type_t buft) {
+    if (buft->iface.is_host) {
+        return buft->iface.is_host(buft);
+    }
+    return false;
+}
+
 // backend buffer
 
 ggml_backend_buffer_t ggml_backend_buffer_init(
-        struct ggml_backend                  * backend,
+               ggml_backend_buffer_type_t      buft,
         struct ggml_backend_buffer_i           iface,
                ggml_backend_buffer_context_t   context,
                size_t                          size) {
@@ -26,7 +55,7 @@ ggml_backend_buffer_t ggml_backend_buffer_init(
 
     (*buffer) = (struct ggml_backend_buffer) {
         /* .interface = */ iface,
-        /* .backend   = */ backend,
+        /* .buft      = */ buft,
         /* .context   = */ context,
         /* .size      = */ size,
     };
@@ -45,10 +74,6 @@ void ggml_backend_buffer_free(ggml_backend_buffer_t buffer) {
     free(buffer);
 }
 
-size_t ggml_backend_buffer_get_alignment(ggml_backend_buffer_t buffer) {
-    return ggml_backend_get_alignment(buffer->backend);
-}
-
 size_t ggml_backend_buffer_get_size(ggml_backend_buffer_t buffer) {
     return buffer->size;
 }
@@ -61,14 +86,6 @@ void * ggml_backend_buffer_get_base(ggml_backend_buffer_t buffer) {
     return base;
 }
 
-size_t ggml_backend_buffer_get_alloc_size(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) {
-    // get_alloc_size is optional, defaults to ggml_nbytes
-    if (buffer->iface.get_alloc_size) {
-        return buffer->iface.get_alloc_size(buffer, tensor);
-    }
-    return ggml_nbytes(tensor);
-}
-
 void ggml_backend_buffer_init_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) {
     // init_tensor is optional
     if (buffer->iface.init_tensor) {
@@ -76,19 +93,28 @@ void ggml_backend_buffer_init_tensor(ggml_backend_buffer_t buffer, struct ggml_t
     }
 }
 
-void ggml_backend_buffer_free_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) {
-    // free_tensor is optional
-    if (buffer->iface.free_tensor) {
-        buffer->iface.free_tensor(buffer, tensor);
-    }
+size_t ggml_backend_buffer_get_alignment (ggml_backend_buffer_t buffer) {
+    return ggml_backend_buft_get_alignment(ggml_backend_buffer_type(buffer));
+}
+
+size_t ggml_backend_buffer_get_alloc_size(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) {
+    return ggml_backend_buft_get_alloc_size(ggml_backend_buffer_type(buffer), tensor);
+}
+
+void ggml_backend_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
+    buffer->iface.clear(buffer, value);
+}
+
+bool ggml_backend_buffer_is_host(ggml_backend_buffer_t buffer) {
+    return ggml_backend_buft_is_host(ggml_backend_buffer_type(buffer));
+}
+
+ggml_backend_buffer_type_t ggml_backend_buffer_type(ggml_backend_buffer_t buffer) {
+    return buffer->buft;
 }
 
 // backend
 
-ggml_backend_t ggml_get_backend(const struct ggml_tensor * tensor) {
-    return tensor->buffer ? tensor->buffer->backend : NULL;
-}
-
 const char * ggml_backend_name(ggml_backend_t backend) {
     if (backend == NULL) {
         return "NULL";
@@ -104,43 +130,53 @@ void ggml_backend_free(ggml_backend_t backend) {
     backend->iface.free(backend);
 }
 
+ggml_backend_buffer_type_t ggml_backend_get_default_buffer_type(ggml_backend_t backend) {
+    return backend->iface.get_default_buffer_type(backend);
+}
+
 ggml_backend_buffer_t ggml_backend_alloc_buffer(ggml_backend_t backend, size_t size) {
-    return backend->iface.alloc_buffer(backend, size);
+    return ggml_backend_buft_alloc_buffer(ggml_backend_get_default_buffer_type(backend), size);
 }
 
 size_t ggml_backend_get_alignment(ggml_backend_t backend) {
-    return backend->iface.get_alignment(backend);
+    return ggml_backend_buft_get_alignment(ggml_backend_get_default_buffer_type(backend));
 }
 
-void ggml_backend_tensor_set_async(struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
-    ggml_get_backend(tensor)->iface.set_tensor_async(ggml_get_backend(tensor), tensor, data, offset, size);
+void ggml_backend_tensor_set_async(ggml_backend_t backend, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
+    GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
+    GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor write out of bounds");
+
+    backend->iface.set_tensor_async(backend, tensor, data, offset, size);
 }
 
-void ggml_backend_tensor_get_async(const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) {
-    ggml_get_backend(tensor)->iface.get_tensor_async(ggml_get_backend(tensor), tensor, data, offset, size);
+void ggml_backend_tensor_get_async(ggml_backend_t backend, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) {
+    GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
+    GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor read out of bounds");
+
+    backend->iface.get_tensor_async(backend, tensor, data, offset, size);
 }
 
 void ggml_backend_tensor_set(struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
-    ggml_backend_t backend = ggml_get_backend(tensor);
-
     GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
-    GGML_ASSERT(backend != NULL && "tensor backend not set");
+    GGML_ASSERT(tensor->buffer != NULL && "tensor buffer not set");
+    GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor write out of bounds");
 
-    backend->iface.set_tensor_async(backend, tensor, data, offset, size);
-    backend->iface.synchronize(backend);
+    tensor->buffer->iface.set_tensor(tensor->buffer, tensor, data, offset, size);
 }
 
 void ggml_backend_tensor_get(const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) {
-    ggml_backend_t backend = ggml_get_backend(tensor);
-
     GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
-    GGML_ASSERT(backend != NULL && "tensor backend not set");
+    GGML_ASSERT(tensor->buffer != NULL && "tensor buffer not set");
+    GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor read out of bounds");
 
-    backend->iface.get_tensor_async(backend, tensor, data, offset, size);
-    backend->iface.synchronize(backend);
+    tensor->buffer->iface.get_tensor(tensor->buffer, tensor, data, offset, size);
 }
 
 void ggml_backend_synchronize(ggml_backend_t backend) {
+    if (backend->iface.synchronize == NULL) {
+        return;
+    }
+
     backend->iface.synchronize(backend);
 }
 
@@ -154,10 +190,16 @@ void ggml_backend_graph_plan_free(ggml_backend_t backend, ggml_backend_graph_pla
 
 void ggml_backend_graph_plan_compute(ggml_backend_t backend, ggml_backend_graph_plan_t plan) {
     backend->iface.graph_plan_compute(backend, plan);
+
+    // TODO: optional sync
+    ggml_backend_synchronize(backend);
 }
 
 void ggml_backend_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
     backend->iface.graph_compute(backend, cgraph);
+
+    // TODO: optional sync
+    ggml_backend_synchronize(backend);
 }
 
 bool ggml_backend_supports_op(ggml_backend_t backend, const struct ggml_tensor * op) {
@@ -194,14 +236,15 @@ void ggml_backend_tensor_copy(struct ggml_tensor * src, struct ggml_tensor * dst
 
     // TODO: allow backends to support copy to/from same backend
 
-    if (ggml_get_backend(dst)->iface.cpy_tensor_from != NULL) {
-        ggml_get_backend(dst)->iface.cpy_tensor_from(ggml_get_backend(dst)->context, src, dst);
-    } else if (ggml_get_backend(src)->iface.cpy_tensor_to != NULL) {
-        ggml_get_backend(src)->iface.cpy_tensor_to(ggml_get_backend(src)->context, src, dst);
+    if (dst->buffer->iface.cpy_tensor_from != NULL) {
+        dst->buffer->iface.cpy_tensor_from(dst->buffer, src, dst);
+    } else if (src->buffer->iface.cpy_tensor_to != NULL) {
+        src->buffer->iface.cpy_tensor_to(src->buffer, src, dst);
     } else {
         // shouldn't be hit when copying from/to CPU
         #ifndef NDEBUG
-        fprintf(stderr, "ggml_backend_tensor_copy: neither cpy_tensor_from nor cpy_tensor_to are implemented for backends %s and %s, falling back to get/set\n", ggml_backend_name(src->buffer->backend), ggml_backend_name(dst->buffer->backend));
+        fprintf(stderr, "ggml_backend_tensor_copy: neither cpy_tensor_from nor cpy_tensor_to "
+                        "are implemented for %s and %s, falling back to get/set\n", src->name, dst->name);
         #endif
         size_t nbytes = ggml_nbytes(src);
         void * data = malloc(nbytes);
@@ -211,8 +254,291 @@ void ggml_backend_tensor_copy(struct ggml_tensor * src, struct ggml_tensor * dst
     }
 }
 
+// backend registry
+
+#define GGML_MAX_BACKENDS_REG 16
+
+struct ggml_backend_reg {
+    char name[128];
+    ggml_backend_init_fn init_fn;
+    ggml_backend_buffer_type_t default_buffer_type;
+    void * user_data;
+};
+
+static struct ggml_backend_reg ggml_backend_registry[GGML_MAX_BACKENDS_REG];
+static size_t ggml_backend_registry_count = 0;
+
+static ggml_backend_t ggml_backend_reg_cpu_init(const char * params, void * user_data);
+
+static void ggml_backend_registry_init(void) {
+    static bool initialized = false;
+
+    if (initialized) {
+        return;
+    }
+
+    initialized = true;
+
+    ggml_backend_register("CPU", ggml_backend_reg_cpu_init, ggml_backend_cpu_buffer_type(), NULL);
+
+    // add forward decls here to avoid including the backend headers
+#ifdef GGML_USE_CUBLAS
+    extern void ggml_backend_cuda_reg_devices(void);
+    ggml_backend_cuda_reg_devices();
+#endif
+
+#ifdef GGML_USE_METAL
+    extern ggml_backend_t ggml_backend_reg_metal_init(const char * params, void * user_data);
+    extern ggml_backend_buffer_type_t ggml_backend_metal_buffer_type(void);
+    ggml_backend_register("Metal", ggml_backend_reg_metal_init, ggml_backend_metal_buffer_type(), NULL);
+#endif
+}
+
+void ggml_backend_register(const char * name, ggml_backend_init_fn init_fn, ggml_backend_buffer_type_t default_buffer_type, void * user_data) {
+    GGML_ASSERT(ggml_backend_registry_count < GGML_MAX_BACKENDS_REG);
+
+    int id = ggml_backend_registry_count;
+
+    ggml_backend_registry[id] = (struct ggml_backend_reg) {
+        /* .name                = */ {0},
+        /* .fn                  = */ init_fn,
+        /* .default_buffer_type = */ default_buffer_type,
+        /* .user_data           = */ user_data,
+    };
+
+    snprintf(ggml_backend_registry[id].name, sizeof(ggml_backend_registry[id].name), "%s", name);
+
+#ifndef NDEBUG
+    fprintf(stderr, "%s: registered backend %s\n", __func__, name);
+#endif
+
+    ggml_backend_registry_count++;
+}
+
+size_t ggml_backend_reg_get_count(void) {
+    ggml_backend_registry_init();
+
+    return ggml_backend_registry_count;
+}
+
+size_t ggml_backend_reg_find_by_name(const char * name) {
+    ggml_backend_registry_init();
+
+    for (size_t i = 0; i < ggml_backend_registry_count; i++) {
+        // TODO: case insensitive in a portable way
+        if (strcmp(ggml_backend_registry[i].name, name) == 0) {
+            return i;
+        }
+    }
+    return SIZE_MAX;
+}
+
+// init from backend:params string
+ggml_backend_t ggml_backend_reg_init_backend_from_str(const char * backend_str) {
+    ggml_backend_registry_init();
+
+    const char * params = strchr(backend_str, ':');
+    char backend_name[128];
+    if (params == NULL) {
+        strcpy(backend_name, backend_str);
+        params = "";
+    } else {
+        strncpy(backend_name, backend_str, params - backend_str);
+        backend_name[params - backend_str] = '\0';
+        params++;
+    }
+
+    size_t backend_i = ggml_backend_reg_find_by_name(backend_name);
+    if (backend_i == SIZE_MAX) {
+        fprintf(stderr, "%s: backend %s not found\n", __func__, backend_name);
+        return NULL;
+    }
+
+    return ggml_backend_reg_init_backend(backend_i, params);
+}
+
+const char * ggml_backend_reg_get_name(size_t i) {
+    ggml_backend_registry_init();
+
+    GGML_ASSERT(i < ggml_backend_registry_count);
+    return ggml_backend_registry[i].name;
+}
+
+ggml_backend_t ggml_backend_reg_init_backend(size_t i, const char * params) {
+    ggml_backend_registry_init();
+
+    GGML_ASSERT(i < ggml_backend_registry_count);
+    return ggml_backend_registry[i].init_fn(params, ggml_backend_registry[i].user_data);
+}
+
+ggml_backend_buffer_type_t ggml_backend_reg_get_default_buffer_type(size_t i) {
+    ggml_backend_registry_init();
+
+    GGML_ASSERT(i < ggml_backend_registry_count);
+    return ggml_backend_registry[i].default_buffer_type;
+}
+
+ggml_backend_buffer_t ggml_backend_reg_alloc_buffer(size_t i, size_t size) {
+    ggml_backend_registry_init();
+
+    GGML_ASSERT(i < ggml_backend_registry_count);
+    return ggml_backend_buft_alloc_buffer(ggml_backend_registry[i].default_buffer_type, size);
+}
+
 // backend CPU
 
+static void * ggml_backend_cpu_buffer_get_base(ggml_backend_buffer_t buffer) {
+    return (void *)buffer->context;
+}
+
+static void ggml_backend_cpu_buffer_free_buffer(ggml_backend_buffer_t buffer) {
+    free(buffer->context);
+}
+
+static void ggml_backend_cpu_buffer_set_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
+    GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor write out of bounds");
+    GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
+
+    memcpy((char *)tensor->data + offset, data, size);
+
+    GGML_UNUSED(buffer);
+}
+
+static void ggml_backend_cpu_buffer_get_tensor(ggml_backend_buffer_t buffer, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) {
+    GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor read out of bounds");
+    GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
+
+    memcpy(data, (const char *)tensor->data + offset, size);
+
+    GGML_UNUSED(buffer);
+}
+
+static void ggml_backend_cpu_buffer_cpy_tensor_from(ggml_backend_buffer_t buffer, struct ggml_tensor * src, struct ggml_tensor * dst) {
+    ggml_backend_tensor_get(src, dst->data, 0, ggml_nbytes(src));
+
+    GGML_UNUSED(buffer);
+}
+
+static void ggml_backend_cpu_buffer_cpy_tensor_to(ggml_backend_buffer_t buffer, struct ggml_tensor * src, struct ggml_tensor * dst) {
+    ggml_backend_tensor_set(dst, src->data, 0, ggml_nbytes(src));
+
+    GGML_UNUSED(buffer);
+}
+
+static void ggml_backend_cpu_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
+    memset(buffer->context, value, buffer->size);
+}
+
+static struct ggml_backend_buffer_i cpu_backend_buffer_i = {
+    /* .free_buffer     = */ ggml_backend_cpu_buffer_free_buffer,
+    /* .get_base        = */ ggml_backend_cpu_buffer_get_base,
+    /* .init_tensor     = */ NULL, // no initialization required
+    /* .set_tensor      = */ ggml_backend_cpu_buffer_set_tensor,
+    /* .get_tensor      = */ ggml_backend_cpu_buffer_get_tensor,
+    /* .cpy_tensor_from = */ ggml_backend_cpu_buffer_cpy_tensor_from,
+    /* .cpy_tensor_to   = */ ggml_backend_cpu_buffer_cpy_tensor_to,
+    /* .clear           = */ ggml_backend_cpu_buffer_clear,
+};
+
+// for buffers from ptr, free is not called
+static struct ggml_backend_buffer_i cpu_backend_buffer_i_from_ptr = {
+    /* .free_buffer     = */ NULL, // ptr is not owned by the buffer, so it does not need to be freed
+    /* .get_base        = */ ggml_backend_cpu_buffer_get_base,
+    /* .init_tensor     = */ NULL, // no initialization required
+    /* .set_tensor      = */ ggml_backend_cpu_buffer_set_tensor,
+    /* .get_tensor      = */ ggml_backend_cpu_buffer_get_tensor,
+    /* .cpy_tensor_from = */ ggml_backend_cpu_buffer_cpy_tensor_from,
+    /* .cpy_tensor_to   = */ ggml_backend_cpu_buffer_cpy_tensor_to,
+    /* .clear           = */ ggml_backend_cpu_buffer_clear,
+};
+
+static const size_t TENSOR_ALIGNMENT = 64; // should be enough for AVX 512
+
+static ggml_backend_buffer_t ggml_backend_cpu_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
+    size += TENSOR_ALIGNMENT;   // malloc may return an address that is not aligned
+    void * data = malloc(size); // TODO: maybe use GGML_ALIGNED_MALLOC?
+
+    GGML_ASSERT(data != NULL && "failed to allocate buffer");
+
+    return ggml_backend_buffer_init(buft, cpu_backend_buffer_i, data, size);
+}
+
+static size_t ggml_backend_cpu_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
+    return TENSOR_ALIGNMENT;
+
+    GGML_UNUSED(buft);
+}
+
+static bool ggml_backend_cpu_buffer_type_supports_backend(ggml_backend_buffer_type_t buft, ggml_backend_t backend) {
+    return ggml_backend_is_cpu(backend);
+
+    GGML_UNUSED(buft);
+}
+
+static bool ggml_backend_cpu_buffer_type_is_host(ggml_backend_buffer_type_t buft) {
+    return true;
+
+    GGML_UNUSED(buft);
+}
+
+ggml_backend_buffer_type_t ggml_backend_cpu_buffer_type(void) {
+    static struct ggml_backend_buffer_type ggml_backend_cpu_buffer_type = {
+        /* .iface = */ {
+            /* .alloc_buffer     = */ ggml_backend_cpu_buffer_type_alloc_buffer,
+            /* .get_alignment    = */ ggml_backend_cpu_buffer_type_get_alignment,
+            /* .get_alloc_size   = */ NULL, // defaults to ggml_nbytes
+            /* .supports_backend = */ ggml_backend_cpu_buffer_type_supports_backend,
+            /* .is_host          = */ ggml_backend_cpu_buffer_type_is_host,
+        },
+        /* .context = */ NULL,
+    };
+
+    return &ggml_backend_cpu_buffer_type;
+}
+
+#ifdef GGML_USE_CPU_HBM
+
+// buffer type HBM
+
+#include <hbwmalloc.h>
+
+static void ggml_backend_cpu_hbm_buffer_free_buffer(ggml_backend_buffer_t buffer) {
+    hbw_free(buffer->context);
+}
+
+static ggml_backend_buffer_t ggml_backend_cpu_hbm_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
+    //void * ptr = hbw_malloc(size);
+    void * ptr;
+    int result = hbw_posix_memalign(&ptr, ggml_backend_cpu_buffer_type_get_alignment(buft), size);
+    if (result != 0) {
+        fprintf(stderr, "failed to allocate HBM buffer of size %zu\n", size);
+        return NULL;
+    }
+
+    // FIXME: this is a hack to avoid having to implement a new buffer type
+    ggml_backend_buffer_t buffer = ggml_backend_cpu_buffer_from_ptr(ptr, size);
+    buffer->buft = buft;
+    buffer->iface.free_buffer = ggml_backend_cpu_hbm_buffer_free_buffer;
+
+    return buffer;
+}
+
+ggml_backend_buffer_type_t ggml_backend_cpu_hbm_buffer_type() {
+    static struct ggml_backend_buffer_type ggml_backend_cpu_buffer_type_hbm = {
+        /* .iface    = */ {
+            /* .alloc_buffer     = */ ggml_backend_cpu_hbm_buffer_type_alloc_buffer,
+            /* .get_alignment    = */ ggml_backend_cpu_buffer_type_get_alignment,
+            /* .get_alloc_size   = */ NULL, // defaults to ggml_nbytes
+            /* .supports_backend = */ ggml_backend_cpu_buffer_type_supports_backend,
+            /* .is_host          = */ ggml_backend_cpu_buffer_type_is_host,
+        },
+        /* .context  = */ NULL,
+    };
+
+    return &ggml_backend_cpu_buffer_type_hbm;
+}
+#endif
+
 struct ggml_backend_cpu_context {
     int n_threads;
     void * work_data;
@@ -222,7 +548,7 @@ struct ggml_backend_cpu_context {
 static const char * ggml_backend_cpu_name(ggml_backend_t backend) {
     return "CPU";
 
-    UNUSED(backend);
+    GGML_UNUSED(backend);
 }
 
 static void ggml_backend_cpu_free(ggml_backend_t backend) {
@@ -232,80 +558,10 @@ static void ggml_backend_cpu_free(ggml_backend_t backend) {
     free(backend);
 }
 
-static void * ggml_backend_cpu_buffer_get_base(ggml_backend_buffer_t buffer) {
-    return (void *)buffer->context;
-}
+static ggml_backend_buffer_type_t ggml_backend_cpu_get_default_buffer_type(ggml_backend_t backend) {
+    return ggml_backend_cpu_buffer_type();
 
-static void ggml_backend_cpu_buffer_free_buffer(ggml_backend_buffer_t buffer) {
-    free(buffer->context);
-    UNUSED(buffer);
-}
-
-static struct ggml_backend_buffer_i cpu_backend_buffer_i = {
-    /* .free_buffer    = */ ggml_backend_cpu_buffer_free_buffer,
-    /* .get_base       = */ ggml_backend_cpu_buffer_get_base,
-    /* .get_alloc_size = */ NULL, // defaults to ggml_nbytes
-    /* .init_tensor    = */ NULL, // no initialization required
-    /* .free_tensor    = */ NULL, // no cleanup required
-};
-
-// for buffers from ptr, free is not called
-static struct ggml_backend_buffer_i cpu_backend_buffer_i_from_ptr = {
-    /* .free_buffer    = */ NULL, // ptr is not owned by the buffer, so it does not need to be freed
-    /* .get_base       = */ ggml_backend_cpu_buffer_get_base,
-    /* .get_alloc_size = */ NULL, // defaults to ggml_nbytes
-    /* .init_tensor    = */ NULL,
-    /* .free_tensor    = */ NULL,
-};
-
-static const size_t TENSOR_ALIGNMENT = 64; // should be enough for AVX 512
-
-static ggml_backend_buffer_t ggml_backend_cpu_alloc_buffer(ggml_backend_t backend, size_t size) {
-    size += TENSOR_ALIGNMENT;   // malloc may return an address that is not aligned
-    void * data = malloc(size); // TODO: maybe use GGML_ALIGNED_MALLOC?
-
-    GGML_ASSERT(data != NULL && "failed to allocate buffer");
-
-    return ggml_backend_buffer_init(backend, cpu_backend_buffer_i, data, size);
-}
-
-static size_t ggml_backend_cpu_get_alignment(ggml_backend_t backend) {
-    return TENSOR_ALIGNMENT;
-    UNUSED(backend);
-}
-
-static void ggml_backend_cpu_set_tensor_async(ggml_backend_t backend, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
-    GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor write out of bounds");
-    GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
-
-    memcpy((char *)tensor->data + offset, data, size);
-
-    UNUSED(backend);
-}
-
-static void ggml_backend_cpu_get_tensor_async(ggml_backend_t backend, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) {
-    GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor read out of bounds");
-    GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
-
-    memcpy(data, (const char *)tensor->data + offset, size);
-
-    UNUSED(backend);
-}
-
-static void ggml_backend_cpu_synchronize(ggml_backend_t backend) {
-    UNUSED(backend);
-}
-
-static void ggml_backend_cpu_cpy_tensor_from(ggml_backend_t backend, struct ggml_tensor * src, struct ggml_tensor * dst) {
-    ggml_backend_tensor_get(src, dst->data, 0, ggml_nbytes(src));
-
-    UNUSED(backend);
-}
-
-static void ggml_backend_cpu_cpy_tensor_to(ggml_backend_t backend, struct ggml_tensor * src, struct ggml_tensor * dst) {
-    ggml_backend_tensor_set(dst, src->data, 0, ggml_nbytes(src));
-
-    UNUSED(backend);
+    GGML_UNUSED(backend);
 }
 
 struct ggml_backend_plan_cpu {
@@ -319,7 +575,7 @@ static ggml_backend_graph_plan_t ggml_backend_cpu_graph_plan_create(ggml_backend
     struct ggml_backend_plan_cpu * cpu_plan = malloc(sizeof(struct ggml_backend_plan_cpu));
 
     cpu_plan->cplan = ggml_graph_plan(cgraph, cpu_ctx->n_threads);
-    cpu_plan->cgraph = *cgraph;
+    cpu_plan->cgraph = *cgraph; // FIXME: deep copy
 
     if (cpu_plan->cplan.work_size > 0) {
         cpu_plan->cplan.work_data = malloc(cpu_plan->cplan.work_size);
@@ -334,7 +590,7 @@ static void ggml_backend_cpu_graph_plan_free(ggml_backend_t backend, ggml_backen
     free(cpu_plan->cplan.work_data);
     free(cpu_plan);
 
-    UNUSED(backend);
+    GGML_UNUSED(backend);
 }
 
 static void ggml_backend_cpu_graph_plan_compute(ggml_backend_t backend, ggml_backend_graph_plan_t plan) {
@@ -342,7 +598,7 @@ static void ggml_backend_cpu_graph_plan_compute(ggml_backend_t backend, ggml_bac
 
     ggml_graph_compute(&cpu_plan->cgraph, &cpu_plan->cplan);
 
-    UNUSED(backend);
+    GGML_UNUSED(backend);
 }
 
 static void ggml_backend_cpu_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
@@ -363,25 +619,25 @@ static void ggml_backend_cpu_graph_compute(ggml_backend_t backend, struct ggml_c
 
 static bool ggml_backend_cpu_supports_op(ggml_backend_t backend, const struct ggml_tensor * op) {
     return true;
-    UNUSED(backend);
-    UNUSED(op);
+
+    GGML_UNUSED(backend);
+    GGML_UNUSED(op);
 }
 
 static struct ggml_backend_i cpu_backend_i = {
-    /* .get_name            = */ ggml_backend_cpu_name,
-    /* .free                = */ ggml_backend_cpu_free,
-    /* .alloc_buffer        = */ ggml_backend_cpu_alloc_buffer,
-    /* .get_alignment       = */ ggml_backend_cpu_get_alignment,
-    /* .set_tensor_async    = */ ggml_backend_cpu_set_tensor_async,
-    /* .get_tensor_async    = */ ggml_backend_cpu_get_tensor_async,
-    /* .synchronize         = */ ggml_backend_cpu_synchronize,
-    /* .cpy_tensor_from     = */ ggml_backend_cpu_cpy_tensor_from,
-    /* .cpy_tensor_to       = */ ggml_backend_cpu_cpy_tensor_to,
-    /* .graph_plan_create   = */ ggml_backend_cpu_graph_plan_create,
-    /* .graph_plan_free     = */ ggml_backend_cpu_graph_plan_free,
-    /* .graph_plan_compute  = */ ggml_backend_cpu_graph_plan_compute,
-    /* .graph_compute       = */ ggml_backend_cpu_graph_compute,
-    /* .supports_op         = */ ggml_backend_cpu_supports_op,
+    /* .get_name                = */ ggml_backend_cpu_name,
+    /* .free                    = */ ggml_backend_cpu_free,
+    /* .get_default_buffer_type = */ ggml_backend_cpu_get_default_buffer_type,
+    /* .set_tensor_async        = */ NULL,
+    /* .get_tensor_async        = */ NULL,
+    /* .cpy_tensor_from_async   = */ NULL,
+    /* .cpy_tensor_to_async     = */ NULL,
+    /* .synchronize             = */ NULL,
+    /* .graph_plan_create       = */ ggml_backend_cpu_graph_plan_create,
+    /* .graph_plan_free         = */ ggml_backend_cpu_graph_plan_free,
+    /* .graph_plan_compute      = */ ggml_backend_cpu_graph_plan_compute,
+    /* .graph_compute           = */ ggml_backend_cpu_graph_compute,
+    /* .supports_op             = */ ggml_backend_cpu_supports_op,
 };
 
 ggml_backend_t ggml_backend_cpu_init(void) {
@@ -411,10 +667,18 @@ void ggml_backend_cpu_set_n_threads(ggml_backend_t backend_cpu, int n_threads) {
     ctx->n_threads = n_threads;
 }
 
-ggml_backend_buffer_t ggml_backend_cpu_buffer_from_ptr(ggml_backend_t backend_cpu, void * ptr, size_t size) {
-    return ggml_backend_buffer_init(backend_cpu, cpu_backend_buffer_i_from_ptr, ptr, size);
+ggml_backend_buffer_t ggml_backend_cpu_buffer_from_ptr(void * ptr, size_t size) {
+    return ggml_backend_buffer_init(ggml_backend_cpu_buffer_type(), cpu_backend_buffer_i_from_ptr, ptr, size);
 }
 
+static ggml_backend_t ggml_backend_reg_cpu_init(const char * params, void * user_data) {
+    return ggml_backend_cpu_init();
+
+    GGML_UNUSED(params);
+    GGML_UNUSED(user_data);
+}
+
+
 // scheduler
 
 #define GGML_MAX_BACKENDS 4
@@ -427,7 +691,7 @@ struct ggml_backend_sched_split {
     int i_end;
     struct ggml_tensor * inputs[GGML_MAX_SPLIT_INPUTS];
     int n_inputs;
-    struct ggml_cgraph * graph;
+    struct ggml_cgraph graph;
 };
 
 struct ggml_backend_sched {
@@ -453,7 +717,7 @@ struct ggml_backend_sched {
     #else
     __attribute__((aligned(GGML_MEM_ALIGN)))
     #endif
-    char context_buffer[GGML_MAX_SPLITS*GGML_MAX_SPLIT_INPUTS*sizeof(struct ggml_tensor) + GGML_MAX_SPLITS*sizeof(struct ggml_cgraph)];
+    char context_buffer[GGML_MAX_SPLITS*GGML_MAX_SPLIT_INPUTS*sizeof(struct ggml_tensor) + sizeof(struct ggml_cgraph)];
 };
 
 #define hash_id(node) ggml_hash_find_or_insert(sched->hash_set, node)
@@ -482,23 +746,57 @@ static int sched_allocr_prio(ggml_backend_sched_t sched, ggml_tallocr_t allocr)
     return INT_MAX;
 }
 
+static ggml_backend_t get_buffer_backend(ggml_backend_sched_t sched, ggml_backend_buffer_t buffer) {
+    if (buffer == NULL) {
+        return NULL;
+    }
+    // find highest prio backend that supports the buffer type
+    for (int i = 0; i < sched->n_backends; i++) {
+        if (ggml_backend_buft_supports_backend(buffer->buft, sched->backends[i])) {
+            return sched->backends[i];
+        }
+    }
+    GGML_ASSERT(false && "tensor buffer type not supported by any backend");
+}
+
+static ggml_backend_t get_allocr_backend(ggml_backend_sched_t sched, ggml_tallocr_t allocr) {
+    if (allocr == NULL) {
+        return NULL;
+    }
+    // find highest prio backend that supports the buffer type
+    for (int i = 0; i < sched->n_backends; i++) {
+        if (sched->tallocs[i] == allocr) {
+            return sched->backends[i];
+        }
+    }
+    GGML_UNREACHABLE();
+}
+
+#if 0
+static char causes[GGML_DEFAULT_GRAPH_SIZE*8 + GGML_MAX_SPLITS*GGML_MAX_SPLIT_INPUTS][128]; // debug, remove
+#define SET_CAUSE(node, ...) sprintf(causes[hash_id(node)], __VA_ARGS__)
+#define GET_CAUSE(node) causes[hash_id(node)]
+#else
+#define SET_CAUSE(node, ...)
+#define GET_CAUSE(node) ""
+#endif
+
 // returns the backend that should be used for the node based on the current locations
-char causes[GGML_DEFAULT_GRAPH_SIZE*4 + GGML_MAX_SPLITS*GGML_MAX_SPLIT_INPUTS][128]; // debug, remove
 static ggml_backend_t sched_backend_from_cur(ggml_backend_sched_t sched, struct ggml_tensor * node) {
     // if the dst tensor is already allocated in a buffer, we must assume that it is critical to keep it there
     // ie. kv cache updates
     // note that this doesn't allow fallback to CPU. need to add output tensors to the splits to copy the data back to the original backend.
     // dst
-    ggml_backend_t cur_backend = ggml_get_backend(node);
+    ggml_backend_t cur_backend = get_buffer_backend(sched, node->buffer);
     if (cur_backend != NULL) {
-        sprintf(causes[hash_id(node)], "1.dst");
+        SET_CAUSE(node, "1.dst");
         return cur_backend;
     }
 
     // view_src
-    if (node->view_src != NULL && ggml_get_backend(node->view_src) != NULL) {
-        sprintf(causes[hash_id(node)], "1.vsrc");
-        return ggml_get_backend(node->view_src);
+    if (node->view_src != NULL && get_buffer_backend(sched, node->view_src->buffer) != NULL) {
+        SET_CAUSE(node, "1.vsrc");
+        return get_buffer_backend(sched, node->view_src->buffer);
     }
 
     // src
@@ -510,7 +808,7 @@ static ggml_backend_t sched_backend_from_cur(ggml_backend_sched_t sched, struct
         if (src == NULL) {
             break;
         }
-        ggml_backend_t src_backend = ggml_get_backend(src);
+        ggml_backend_t src_backend = get_buffer_backend(sched, src->buffer);
         if (src_backend != NULL) {
             int src_prio = sched_backend_prio(sched, src_backend);
             size_t src_size = ggml_nbytes(src);
@@ -518,7 +816,7 @@ static ggml_backend_t sched_backend_from_cur(ggml_backend_sched_t sched, struct
                 cur_prio = src_prio;
                 cur_size = src_size;
                 cur_backend = src_backend;
-                sprintf(causes[hash_id(node)], "1.src%d", i);
+                SET_CAUSE(node, "1.src%d", i);
             }
         }
     }
@@ -539,10 +837,12 @@ static void sched_print_assignments(ggml_backend_sched_t sched, struct ggml_cgra
     int cur_split = 0;
     for (int i = 0; i < graph->n_nodes; i++) {
         if (cur_split < sched->n_splits && i == sched->splits[cur_split].i_start) {
-            ggml_backend_t split_backend = ggml_tallocr_get_buffer(sched->splits[cur_split].tallocr)->backend;
-            fprintf(stderr, "\n## SPLIT #%d: %s # %d inputs: ", cur_split, ggml_backend_name(split_backend), sched->splits[cur_split].n_inputs);
+            ggml_backend_t split_backend = get_allocr_backend(sched, sched->splits[cur_split].tallocr);
+            fprintf(stderr, "\n## SPLIT #%d: %s # %d inputs: ", cur_split, ggml_backend_name(split_backend),
+                sched->splits[cur_split].n_inputs);
             for (int j = 0; j < sched->splits[cur_split].n_inputs; j++) {
-                fprintf(stderr, "[%s (%5.5s)] ", sched->splits[cur_split].inputs[j]->name, fmt_size(ggml_nbytes(sched->splits[cur_split].inputs[j])));
+                fprintf(stderr, "[%s (%5.5s)] ", sched->splits[cur_split].inputs[j]->name,
+                    fmt_size(ggml_nbytes(sched->splits[cur_split].inputs[j])));
             }
             fprintf(stderr, "\n");
             cur_split++;
@@ -552,16 +852,18 @@ static void sched_print_assignments(ggml_backend_sched_t sched, struct ggml_cgra
             continue;
         }
         ggml_tallocr_t node_allocr = node_allocr(node);
-        ggml_backend_t node_backend = node_allocr ? ggml_tallocr_get_buffer(node_allocr)->backend : NULL;
-        fprintf(stderr, "node #%3d (%10.10s): %20.20s (%4.4s) [%4.4s %8.8s]:", i, ggml_op_name(node->op), node->name, fmt_size(ggml_nbytes(node)), node_allocr ? ggml_backend_name(node_backend) : "NULL", causes[hash_id(node)]);
+        ggml_backend_t node_backend = node_allocr ? get_allocr_backend(sched, node_allocr) : NULL; // FIXME:
+        fprintf(stderr, "node #%3d (%10.10s): %20.20s (%4.4s) [%4.4s %8.8s]:", i, ggml_op_name(node->op), node->name,
+            fmt_size(ggml_nbytes(node)), node_allocr ? ggml_backend_name(node_backend) : "NULL", GET_CAUSE(node));
         for (int j = 0; j < GGML_MAX_SRC; j++) {
             struct ggml_tensor * src = node->src[j];
             if (src == NULL) {
                 break;
             }
             ggml_tallocr_t src_allocr = node_allocr(src);
-            ggml_backend_t src_backend = src_allocr ? ggml_tallocr_get_buffer(src_allocr)->backend : NULL;
-            fprintf(stderr, " %20.20s (%4.4s) [%4.4s %8.8s]", src->name, fmt_size(ggml_nbytes(src)), src_backend ? ggml_backend_name(src_backend) : "NULL", causes[hash_id(src)]);
+            ggml_backend_t src_backend = src_allocr ? get_allocr_backend(sched, src_allocr) : NULL;
+            fprintf(stderr, " %20.20s (%4.4s) [%4.4s %8.8s]", src->name,
+                fmt_size(ggml_nbytes(src)), src_backend ? ggml_backend_name(src_backend) : "NULL", GET_CAUSE(src));
         }
         fprintf(stderr, "\n");
     }
@@ -587,9 +889,9 @@ static void sched_split_graph(ggml_backend_sched_t sched, struct ggml_cgraph * g
     sched->n_splits = 0;
 
     struct ggml_init_params params = {
-        /*.mem_size =   */ sizeof(sched->context_buffer),
-        /*.mem_buffer = */ sched->context_buffer,
-        /*.no_alloc =   */ true
+        /* .mem_size =   */ sizeof(sched->context_buffer),
+        /* .mem_buffer = */ sched->context_buffer,
+        /* .no_alloc =   */ true
     };
 
     if (sched->ctx != NULL) {
@@ -605,9 +907,9 @@ static void sched_split_graph(ggml_backend_sched_t sched, struct ggml_cgraph * g
             // do not overwrite user assignments
             continue;
         }
-        ggml_backend_t leaf_backend = ggml_get_backend(leaf);
+        ggml_backend_t leaf_backend = get_buffer_backend(sched, leaf->buffer);
         if (leaf_backend == NULL && leaf->view_src != NULL) {
-            leaf_backend = ggml_get_backend(leaf->view_src);
+            leaf_backend = get_buffer_backend(sched, leaf->view_src->buffer);
         }
         if (leaf_backend != NULL) {
             node_allocr(leaf) = ggml_backend_sched_get_tallocr(sched, leaf_backend);
@@ -649,7 +951,7 @@ static void sched_split_graph(ggml_backend_sched_t sched, struct ggml_cgraph * g
                         cur_prio = src_prio;
                         cur_size = src_size;
                         node_allocr = src_allocr;
-                        sprintf(causes[hash_id(node)], "2.src%d", j);
+                        SET_CAUSE(node, "2.src%d", j);
                     }
                 }
             }
@@ -733,7 +1035,7 @@ static void sched_split_graph(ggml_backend_sched_t sched, struct ggml_cgraph * g
                     struct ggml_tensor * tensor_copy = ggml_dup_tensor_layout(sched->ctx, src);
                     sched->node_copies[id][cur_backend_id] = tensor_copy;
                     node_allocr(tensor_copy) = cur_allocr;
-                    ggml_backend_t backend = ggml_tallocr_get_buffer(cur_allocr)->backend;
+                    ggml_backend_t backend = get_allocr_backend(sched, cur_allocr);
                     ggml_format_name(tensor_copy, "%s#%s", ggml_backend_name(backend), src->name);
                 }
                 node->src[j] = sched->node_copies[id][cur_backend_id];
@@ -761,8 +1063,8 @@ static void sched_split_graph(ggml_backend_sched_t sched, struct ggml_cgraph * g
             ggml_tallocr_t src_allocr = node_allocr(src);
             if (src_allocr != node_allocr /* && src_backend != NULL */) { // ignore nulls for now
                 fprintf(stderr, "!!!! %s has backend %s, src %d (%s) has backend %s\n",
-                    node->name, node_allocr ? ggml_backend_name(ggml_tallocr_get_buffer(node_allocr)->backend) : "NULL",
-                    j, src->name, src_allocr ? ggml_backend_name(ggml_tallocr_get_buffer(src_allocr)->backend) : "NULL");
+                    node->name, node_allocr ? ggml_backend_name(get_allocr_backend(sched, node_allocr)) : "NULL",
+                    j, src->name, src_allocr ? ggml_backend_name(get_allocr_backend(sched, src_allocr)) : "NULL");
             }
         }
     }
@@ -773,7 +1075,7 @@ static void sched_split_graph(ggml_backend_sched_t sched, struct ggml_cgraph * g
     struct ggml_cgraph * graph_copy = ggml_new_graph_custom(sched->ctx, graph->n_nodes + sched->n_splits*GGML_MAX_SPLIT_INPUTS, false);
     for (int i = 0; i < sched->n_splits; i++) {
         struct ggml_backend_sched_split * split = &sched->splits[i];
-        split->graph = ggml_graph_view(sched->ctx, graph, split->i_start, split->i_end);
+        split->graph = ggml_graph_view(graph, split->i_start, split->i_end);
 
         // add inputs to the graph copy so that they are allocated by ggml-alloc at the start of the split
         for (int j = 0; j < split->n_inputs; j++) {
@@ -806,31 +1108,29 @@ static void sched_compute_splits(ggml_backend_sched_t sched) {
 
     for (int i = 0; i < sched->n_splits; i++) {
         struct ggml_backend_sched_split * split = &splits[i];
-        ggml_backend_t split_backend = ggml_tallocr_get_buffer(split->tallocr)->backend;
+        ggml_backend_t split_backend = get_allocr_backend(sched, split->tallocr);
         int split_backend_id = sched_backend_prio(sched, split_backend);
 
         // copy the input tensors to the split backend
         uint64_t copy_start_us = ggml_time_us();
         for (int j = 0; j < split->n_inputs; j++) {
-            struct ggml_tensor * input_cpy = sched->node_copies[hash_id(split->inputs[j])][sched_backend_prio(sched, split_backend)];
-            if (split->inputs[j]->buffer == NULL) {
-                if (split->inputs[j]->view_src == NULL) {
-                    fprintf(stderr, "input %s has no buffer and no view_src\n", split->inputs[j]->name);
+            struct ggml_tensor * input = split->inputs[j];
+            struct ggml_tensor * input_cpy = sched->node_copies[hash_id(input)][sched_backend_prio(sched, split_backend)];
+            if (input->buffer == NULL) {
+                if (input->view_src == NULL) {
+                    fprintf(stderr, "input %s has no buffer and no view_src\n", input->name);
                     exit(1);
                 }
-                struct ggml_tensor * view = split->inputs[j];
-                view->backend = view->view_src->backend;
-                view->buffer  = view->view_src->buffer;
-                view->data    = (char *)view->view_src->data + view->view_offs;
-                ggml_backend_buffer_init_tensor(ggml_backend_sched_get_buffer(sched, view->buffer->backend), view);
+                // FIXME: may need to use the sched buffer instead
+                ggml_backend_view_init(input->view_src->buffer, input);
             }
             if (input_cpy->buffer == NULL) {
                 fprintf(stderr, "input_cpy %s has no buffer\n", input_cpy->name);
                 exit(1);
             }
-            GGML_ASSERT(split->inputs[j]->buffer->backend != input_cpy->buffer->backend);
-            GGML_ASSERT(input_cpy->buffer->backend == split_backend);
-            ggml_backend_tensor_copy(split->inputs[j], input_cpy);
+            //GGML_ASSERT(input->buffer->backend != input_cpy->buffer->backend);
+            //GGML_ASSERT(input_cpy->buffer->backend == split_backend);
+            ggml_backend_tensor_copy(input, input_cpy);
         }
         // ggml_backend_synchronize(split_backend);
         int64_t copy_end_us = ggml_time_us();
@@ -843,7 +1143,7 @@ static void sched_compute_splits(ggml_backend_sched_t sched) {
 #endif
 
         uint64_t compute_start_us = ggml_time_us();
-        ggml_backend_graph_compute(split_backend, split->graph);
+        ggml_backend_graph_compute(split_backend, &split->graph);
         // ggml_backend_synchronize(split_backend);
         uint64_t compute_end_us = ggml_time_us();
         compute_us[split_backend_id] += compute_end_us - compute_start_us;
@@ -872,8 +1172,6 @@ ggml_backend_sched_t ggml_backend_sched_new(ggml_backend_t * backends, int n_bac
     struct ggml_backend_sched * sched = malloc(sizeof(struct ggml_backend_sched));
     memset(sched, 0, sizeof(struct ggml_backend_sched));
 
-    fprintf(stderr, "ggml_backend_sched size: %lu KB\n", sizeof(struct ggml_backend_sched)/1024);
-
     sched->n_backends = n_backends;
     for (int i = 0; i < n_backends; i++) {
         sched->backends[i] = backends[i];
@@ -948,3 +1246,182 @@ void ggml_backend_sched_set_node_backend(ggml_backend_sched_t sched, struct ggml
     GGML_ASSERT(backend_index >= 0 && backend_index < sched->n_backends);
     node_allocr(node) = sched->tallocs[backend_index];
 }
+
+// utils
+void ggml_backend_view_init(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) {
+    GGML_ASSERT(tensor->buffer == NULL);
+    //GGML_ASSERT(tensor->data == NULL); // views of pre-allocted tensors may have the data set, but still need to be initialized
+    GGML_ASSERT(tensor->view_src != NULL);
+    GGML_ASSERT(tensor->view_src->buffer != NULL);
+    GGML_ASSERT(tensor->view_src->data != NULL);
+
+    tensor->buffer = buffer;
+    tensor->data = (char *)tensor->view_src->data + tensor->view_offs;
+    tensor->backend = tensor->view_src->backend;
+    ggml_backend_buffer_init_tensor(buffer, tensor);
+}
+
+void ggml_backend_tensor_alloc(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, void * addr) {
+    GGML_ASSERT(tensor->buffer == NULL);
+    GGML_ASSERT(tensor->data == NULL);
+    GGML_ASSERT(tensor->view_src == NULL);
+    GGML_ASSERT(addr >= ggml_backend_buffer_get_base(buffer));
+    GGML_ASSERT((char *)addr + ggml_backend_buffer_get_alloc_size(buffer, tensor) <=
+                (char *)ggml_backend_buffer_get_base(buffer) + ggml_backend_buffer_get_size(buffer));
+
+    tensor->buffer = buffer;
+    tensor->data = addr;
+    ggml_backend_buffer_init_tensor(buffer, tensor);
+}
+
+static struct ggml_tensor * graph_dup_tensor(struct ggml_hash_set hash_set, struct ggml_tensor ** node_copies,
+    struct ggml_context * ctx_allocated, struct ggml_context * ctx_unallocated, struct ggml_tensor * src) {
+
+    GGML_ASSERT(src != NULL);
+    GGML_ASSERT(src->data && "graph must be allocated");
+
+    size_t id = ggml_hash_insert(hash_set, src);
+    if (id == GGML_HASHTABLE_ALREADY_EXISTS) {
+        return node_copies[ggml_hash_find(hash_set, src)];
+    }
+
+    struct ggml_tensor * dst = ggml_dup_tensor_layout(src->data && !src->view_src ? ctx_allocated : ctx_unallocated, src);
+    if (src->view_src != NULL) {
+        dst->view_src = graph_dup_tensor(hash_set, node_copies, ctx_allocated, ctx_unallocated, src->view_src);
+        dst->view_offs = src->view_offs;
+    }
+    dst->op = src->op;
+    memcpy(dst->op_params, src->op_params, sizeof(dst->op_params));
+    ggml_set_name(dst, src->name);
+
+    // copy src
+    for (int i = 0; i < GGML_MAX_SRC; i++) {
+        struct ggml_tensor * s = src->src[i];
+        if (s == NULL) {
+            break;
+        }
+        dst->src[i] = graph_dup_tensor(hash_set, node_copies, ctx_allocated, ctx_unallocated, s);
+    }
+
+    node_copies[id] = dst;
+    return dst;
+}
+
+static void graph_init_tensor(struct ggml_hash_set hash_set, struct ggml_tensor ** node_copies, bool * node_init, struct ggml_tensor * src) {
+    size_t id = ggml_hash_find(hash_set, src);
+    if (node_init[id]) {
+        return;
+    }
+    node_init[id] = true;
+
+    struct ggml_tensor * dst = node_copies[id];
+    if (dst->view_src != NULL) {
+        ggml_backend_view_init(dst->view_src->buffer, dst);
+    }
+    else {
+        ggml_backend_tensor_copy(src, dst);
+    }
+
+    // init src
+    for (int i = 0; i < GGML_MAX_SRC; i++) {
+        struct ggml_tensor * s = src->src[i];
+        if (s == NULL) {
+            break;
+        }
+        graph_init_tensor(hash_set, node_copies, node_init, s);
+    }
+}
+
+struct ggml_backend_graph_copy ggml_backend_graph_copy(ggml_backend_t backend, struct ggml_cgraph * graph) {
+    struct ggml_hash_set hash_set = {
+        /* .size = */ graph->visited_hash_table.size,
+        /* .keys = */ calloc(sizeof(hash_set.keys[0]) * graph->visited_hash_table.size, 1)
+    };
+    struct ggml_tensor ** node_copies = calloc(sizeof(node_copies[0]) * hash_set.size, 1);
+    bool * node_init = calloc(sizeof(node_init[0]) * hash_set.size, 1);
+
+    struct ggml_init_params params = {
+        /* .mem_size   = */ ggml_tensor_overhead()*hash_set.size + ggml_graph_overhead_custom(graph->size, false),
+        /* .mem_buffer = */ NULL,
+        /* .no_alloc   = */ true
+    };
+
+    struct ggml_context * ctx_allocated = ggml_init(params);
+    struct ggml_context * ctx_unallocated = ggml_init(params);
+
+    // dup nodes
+    for (int i = 0; i < graph->n_nodes; i++) {
+        struct ggml_tensor * node = graph->nodes[i];
+        graph_dup_tensor(hash_set, node_copies, ctx_allocated, ctx_unallocated, node);
+    }
+
+    // allocate nodes
+    ggml_backend_buffer_t buffer = ggml_backend_alloc_ctx_tensors(ctx_allocated, backend);
+
+    //printf("copy buffer size: %zu MB\n", ggml_backend_buffer_get_size(buffer) / 1024 / 1024);
+
+    // copy data and init views
+    for (int i = 0; i < graph->n_nodes; i++) {
+        struct ggml_tensor * node = graph->nodes[i];
+        graph_init_tensor(hash_set, node_copies, node_init, node);
+    }
+
+    // build graph copy
+    struct ggml_cgraph * graph_copy = ggml_new_graph_custom(ctx_allocated, graph->size, false);
+    for (int i = 0; i < graph->n_nodes; i++) {
+        struct ggml_tensor * node = graph->nodes[i];
+        struct ggml_tensor * node_copy = node_copies[ggml_hash_find(hash_set, node)];
+        graph_copy->nodes[i] = node_copy;
+    }
+    graph_copy->n_nodes = graph->n_nodes;
+
+    free(hash_set.keys);
+    free(node_copies);
+    free(node_init);
+
+    return (struct ggml_backend_graph_copy) {
+        /* .buffer           = */ buffer,
+        /* .ctx_allocated    = */ ctx_allocated,
+        /* .ctx_unallocated  = */ ctx_unallocated,
+        /* .graph            = */ graph_copy,
+    };
+}
+
+void ggml_backend_graph_copy_free(struct ggml_backend_graph_copy copy) {
+    ggml_backend_buffer_free(copy.buffer);
+    ggml_free(copy.ctx_allocated);
+    ggml_free(copy.ctx_unallocated);
+}
+
+void ggml_backend_compare_graph_backend(ggml_backend_t backend1, ggml_backend_t backend2, struct ggml_cgraph * graph, ggml_backend_eval_callback callback, void * user_data) {
+    struct ggml_backend_graph_copy copy = ggml_backend_graph_copy(backend2, graph);
+    struct ggml_cgraph * g1 = graph;
+    struct ggml_cgraph * g2 = copy.graph;
+
+    assert(g1->n_nodes == g2->n_nodes);
+
+    for (int i = 0; i < g1->n_nodes; i++) {
+        //printf("eval %d/%d\n", i, g1->n_nodes);
+        struct ggml_tensor * t1 = g1->nodes[i];
+        struct ggml_tensor * t2 = g2->nodes[i];
+
+        assert(t1->op == t2->op && ggml_are_same_layout(t1, t2));
+
+        struct ggml_cgraph g1v = ggml_graph_view(g1, i, i + 1);
+        struct ggml_cgraph g2v = ggml_graph_view(g2, i, i + 1);
+
+        ggml_backend_graph_compute(backend1, &g1v);
+        ggml_backend_graph_compute(backend2, &g2v);
+
+        if (ggml_is_view_op(t1->op)) {
+            continue;
+        }
+
+        // compare results, calculate rms etc
+        if (!callback(i, t1, t2, user_data)) {
+            break;
+        }
+    }
+
+    ggml_backend_graph_copy_free(copy);
+}
diff --git a/ggml-backend.h b/ggml-backend.h
index 966687320..a9d2fddd7 100644
--- a/ggml-backend.h
+++ b/ggml-backend.h
@@ -7,41 +7,47 @@
 extern "C" {
 #endif
 
+    typedef struct ggml_backend_buffer_type * ggml_backend_buffer_type_t;
+    typedef struct ggml_backend_buffer * ggml_backend_buffer_t;
+    typedef struct ggml_backend * ggml_backend_t;
+    typedef void * ggml_backend_graph_plan_t;
+
     //
     // Backend buffer
     //
 
-    struct ggml_backend_buffer;
-    typedef struct ggml_backend_buffer * ggml_backend_buffer_t;
+    // buffer type
+    GGML_API ggml_backend_buffer_t ggml_backend_buft_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size);
+    GGML_API size_t ggml_backend_buft_get_alignment (ggml_backend_buffer_type_t buft);
+    GGML_API size_t ggml_backend_buft_get_alloc_size(ggml_backend_buffer_type_t buft, struct ggml_tensor * tensor);
+    GGML_API bool ggml_backend_buft_supports_backend(ggml_backend_buffer_type_t buft, ggml_backend_t backend);
+    GGML_API bool ggml_backend_buft_is_host         (ggml_backend_buffer_type_t buft);
 
-    // backend buffer functions
+    // buffer
     GGML_API void   ggml_backend_buffer_free          (ggml_backend_buffer_t buffer);
-    GGML_API size_t ggml_backend_buffer_get_alignment (ggml_backend_buffer_t buffer);
     GGML_API void * ggml_backend_buffer_get_base      (ggml_backend_buffer_t buffer);
     GGML_API size_t ggml_backend_buffer_get_size      (ggml_backend_buffer_t buffer);
-    GGML_API size_t ggml_backend_buffer_get_alloc_size(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
     GGML_API void   ggml_backend_buffer_init_tensor   (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
-    GGML_API void   ggml_backend_buffer_free_tensor   (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
+    GGML_API size_t ggml_backend_buffer_get_alignment (ggml_backend_buffer_t buffer);
+    GGML_API size_t ggml_backend_buffer_get_alloc_size(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
+    GGML_API void   ggml_backend_buffer_clear         (ggml_backend_buffer_t buffer, uint8_t value);
+    GGML_API bool   ggml_backend_buffer_is_host       (ggml_backend_buffer_t buffer);
+    GGML_API ggml_backend_buffer_type_t ggml_backend_buffer_type(ggml_backend_buffer_t buffer);
 
     //
     // Backend
     //
 
-    struct ggml_backend;
-    typedef struct ggml_backend * ggml_backend_t;
-    typedef void * ggml_backend_graph_plan_t;
-
-    GGML_API ggml_backend_t ggml_get_backend(const struct ggml_tensor * tensor);
 
     GGML_API const char * ggml_backend_name(ggml_backend_t backend);
     GGML_API void         ggml_backend_free(ggml_backend_t backend);
 
-    GGML_API ggml_backend_buffer_t ggml_backend_alloc_buffer(ggml_backend_t backend, size_t size);
+    GGML_API ggml_backend_buffer_type_t ggml_backend_get_default_buffer_type(ggml_backend_t backend);
+    GGML_API ggml_backend_buffer_t      ggml_backend_alloc_buffer(ggml_backend_t backend, size_t size);
+    GGML_API size_t                     ggml_backend_get_alignment(ggml_backend_t backend);
 
-    GGML_API size_t ggml_backend_get_alignment(ggml_backend_t backend);
-
-    GGML_API void ggml_backend_tensor_set_async(      struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
-    GGML_API void ggml_backend_tensor_get_async(const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
+    GGML_API void ggml_backend_tensor_set_async(ggml_backend_t backend,       struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
+    GGML_API void ggml_backend_tensor_get_async(ggml_backend_t backend, const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
 
     GGML_API void ggml_backend_tensor_set(      struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
     GGML_API void ggml_backend_tensor_get(const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
@@ -57,6 +63,7 @@ extern "C" {
 
     // tensor copy between different backends
     GGML_API void ggml_backend_tensor_copy(struct ggml_tensor * src, struct ggml_tensor * dst);
+    GGML_API void ggml_backend_tensor_copy_async(ggml_backend_t backend, struct ggml_tensor * src, struct ggml_tensor * dst); // automatic fallback to sync copy
 
     //
     // CPU backend
@@ -68,8 +75,27 @@ extern "C" {
     GGML_API void ggml_backend_cpu_set_n_threads(ggml_backend_t backend_cpu, int n_threads);
 
     // Create a backend buffer from an existing pointer
-    GGML_API ggml_backend_buffer_t ggml_backend_cpu_buffer_from_ptr(ggml_backend_t backend_cpu, void * ptr, size_t size);
+    GGML_API ggml_backend_buffer_t ggml_backend_cpu_buffer_from_ptr(void * ptr, size_t size);
 
+    GGML_API ggml_backend_buffer_type_t ggml_backend_cpu_buffer_type(void);
+
+#ifdef GGML_USE_CPU_HBM
+    GGML_API ggml_backend_buffer_type_t ggml_backend_cpu_hbm_buffer_type(void);
+#endif
+
+    //
+    // Backend registry
+    //
+
+    // The backend registry is a registry of all the available backends, and allows initializing backends in a generic way
+
+    GGML_API size_t                     ggml_backend_reg_get_count(void);
+    GGML_API size_t                     ggml_backend_reg_find_by_name(const char * name);
+    GGML_API ggml_backend_t             ggml_backend_reg_init_backend_from_str(const char * backend_str); // str is name[:params]
+    GGML_API const char *               ggml_backend_reg_get_name(size_t i);
+    GGML_API ggml_backend_t             ggml_backend_reg_init_backend(size_t i, const char * params); // params is backend-specific
+    GGML_API ggml_backend_buffer_type_t ggml_backend_reg_get_default_buffer_type(size_t i);
+    GGML_API ggml_backend_buffer_t      ggml_backend_reg_alloc_buffer(size_t i, size_t size);
 
     //
     // Backend scheduler
@@ -131,6 +157,32 @@ extern "C" {
             ggml_backend_sched_t sched,
             struct ggml_cgraph * graph);
 
+
+    //
+    // Utils
+    //
+
+    struct ggml_backend_graph_copy {
+        ggml_backend_buffer_t buffer;
+        struct ggml_context * ctx_allocated;
+        struct ggml_context * ctx_unallocated;
+        struct ggml_cgraph * graph;
+    };
+
+    // Copy a graph to a different backend
+    GGML_API struct ggml_backend_graph_copy ggml_backend_graph_copy(ggml_backend_t backend, struct ggml_cgraph * graph);
+    GGML_API void                           ggml_backend_graph_copy_free(struct ggml_backend_graph_copy copy);
+
+    typedef bool (*ggml_backend_eval_callback)(int node_index, struct ggml_tensor * t1, struct ggml_tensor * t2, void * user_data);
+
+    // Compare the output of two backends
+    GGML_API void ggml_backend_compare_graph_backend(ggml_backend_t backend1, ggml_backend_t backend2, struct ggml_cgraph * graph, ggml_backend_eval_callback callback, void * user_data);
+
+    // Tensor initialization
+    GGML_API void ggml_backend_tensor_alloc(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, void * addr);
+    GGML_API void ggml_backend_view_init(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
+
+
 #ifdef  __cplusplus
 }
 #endif
diff --git a/ggml-cuda.cu b/ggml-cuda.cu
index 9019a849f..490081cac 100644
--- a/ggml-cuda.cu
+++ b/ggml-cuda.cu
@@ -1,12 +1,15 @@
 #include <algorithm>
+#include <assert.h>
+#include <atomic>
 #include <cinttypes>
 #include <cstddef>
 #include <cstdint>
+#include <float.h>
 #include <limits>
 #include <stdint.h>
 #include <stdio.h>
-#include <atomic>
-#include <assert.h>
+#include <vector>
+
 
 #if defined(GGML_USE_HIPBLAS)
 #include <hip/hip_runtime.h>
@@ -28,6 +31,7 @@
 #define CUDA_R_16F  HIPBLAS_R_16F
 #define CUDA_R_32F  HIPBLAS_R_32F
 #define __shfl_xor_sync(mask, var, laneMask, width) __shfl_xor(var, laneMask, width)
+#define cublasComputeType_t hipblasDatatype_t //deprecated, new hipblasComputeType_t not in 5.6
 #define cublasCreate hipblasCreate
 #define cublasGemmEx hipblasGemmEx
 #define cublasGemmBatchedEx hipblasGemmBatchedEx
@@ -37,6 +41,7 @@
 #define cublasSetStream hipblasSetStream
 #define cublasSgemm hipblasSgemm
 #define cublasStatus_t hipblasStatus_t
+#define cudaDataType_t hipblasDatatype_t //deprecated, new hipblasDatatype not in 5.6
 #define cudaDeviceCanAccessPeer hipDeviceCanAccessPeer
 #define cudaDeviceDisablePeerAccess hipDeviceDisablePeerAccess
 #define cudaDeviceEnablePeerAccess hipDeviceEnablePeerAccess
@@ -55,8 +60,13 @@
 #define cudaGetDeviceProperties hipGetDeviceProperties
 #define cudaGetErrorString hipGetErrorString
 #define cudaGetLastError hipGetLastError
+#ifdef GGML_HIP_UMA
+#define cudaMalloc hipMallocManaged
+#define cudaMallocHost(ptr, size) hipHostMalloc(ptr, size)
+#else
 #define cudaMalloc hipMalloc
 #define cudaMallocHost(ptr, size) hipHostMalloc(ptr, size, hipHostMallocDefault)
+#endif
 #define cudaMemcpy hipMemcpy
 #define cudaMemcpy2DAsync hipMemcpy2DAsync
 #define cudaMemcpyAsync hipMemcpyAsync
@@ -69,15 +79,24 @@
 #define cudaOccupancyMaxPotentialBlockSize hipOccupancyMaxPotentialBlockSize
 #define cudaSetDevice hipSetDevice
 #define cudaStreamCreateWithFlags hipStreamCreateWithFlags
+#define cudaStreamFireAndForget hipStreamFireAndForget
 #define cudaStreamNonBlocking hipStreamNonBlocking
 #define cudaStreamSynchronize hipStreamSynchronize
 #define cudaStreamWaitEvent(stream, event, flags) hipStreamWaitEvent(stream, event, flags)
 #define cudaStream_t hipStream_t
 #define cudaSuccess hipSuccess
+#define __trap abort
 #else
 #include <cuda_runtime.h>
 #include <cublas_v2.h>
 #include <cuda_fp16.h>
+// CUDA 10.2 does not have these macro definitions.
+#ifndef CUBLAS_TF32_TENSOR_OP_MATH
+#define CUBLAS_TF32_TENSOR_OP_MATH CUBLAS_TENSOR_OP_MATH
+#define CUBLAS_COMPUTE_16F CUDA_R_16F
+#define CUBLAS_COMPUTE_32F CUDA_R_32F
+#define cublasComputeType_t cudaDataType_t
+#endif
 #endif // defined(GGML_USE_HIPBLAS)
 
 #include "ggml-cuda.h"
@@ -190,7 +209,7 @@ static_assert(sizeof(half) == sizeof(ggml_fp16_t), "wrong fp16 size");
             fprintf(stderr, "\nCUDA error %d at %s:%d: %s\n", err_, __FILE__, __LINE__, \
                 cudaGetErrorString(err_));                                              \
             fprintf(stderr, "current device: %d\n", id);                                \
-            exit(1);                                                                    \
+            GGML_ASSERT(!"CUDA error");                                                 \
         }                                                                               \
     } while (0)
 
@@ -204,7 +223,7 @@ static_assert(sizeof(half) == sizeof(ggml_fp16_t), "wrong fp16 size");
             fprintf(stderr, "\ncuBLAS error %d at %s:%d: %s\n",                         \
                     err_, __FILE__, __LINE__, cublasGetStatusString(err_));             \
             fprintf(stderr, "current device: %d\n", id);                                \
-            exit(1);                                                                    \
+            GGML_ASSERT(!"cuBLAS error");                                               \
         }                                                                               \
     } while (0)
 #else
@@ -216,7 +235,7 @@ static_assert(sizeof(half) == sizeof(ggml_fp16_t), "wrong fp16 size");
             cudaGetDevice(&id);                                                         \
             fprintf(stderr, "\ncuBLAS error %d at %s:%d\n", err_, __FILE__, __LINE__);  \
             fprintf(stderr, "current device: %d\n", id);                                \
-            exit(1);                                                                    \
+            GGML_ASSERT(!"cuBLAS error");                                               \
         }                                                                               \
     } while (0)
 #endif // CUDART_VERSION >= 11
@@ -433,10 +452,9 @@ static_assert(sizeof(block_q6_K) == sizeof(ggml_fp16_t) + 13*QK_K/16, "wrong q6_
 #define WARP_SIZE 32
 #define MATRIX_ROW_PADDING 512 // last row of quant. matrices is a multiple of this to avoid out-of-bounds memory accesses
 
-#define CUDA_ADD_BLOCK_SIZE 256
-#define CUDA_MUL_BLOCK_SIZE 256
 #define CUDA_GELU_BLOCK_SIZE 256
 #define CUDA_SILU_BLOCK_SIZE 256
+#define CUDA_TANH_BLOCK_SIZE 256
 #define CUDA_RELU_BLOCK_SIZE 256
 #define CUDA_SQR_BLOCK_SIZE 256
 #define CUDA_CPY_BLOCK_SIZE 32
@@ -449,6 +467,11 @@ static_assert(sizeof(block_q6_K) == sizeof(ggml_fp16_t) + 13*QK_K/16, "wrong q6_
 #define CUDA_QUANTIZE_BLOCK_SIZE 256
 #define CUDA_DEQUANTIZE_BLOCK_SIZE 256
 #define CUDA_GET_ROWS_BLOCK_SIZE 256
+#define CUDA_UPSCALE_BLOCK_SIZE 256
+#define CUDA_CONCAT_BLOCK_SIZE 256
+#define CUDA_PAD_BLOCK_SIZE 256
+#define CUDA_ACC_BLOCK_SIZE 256
+#define CUDA_IM2COL_BLOCK_SIZE 256
 
 // dmmv = dequantize_mul_mat_vec
 #ifndef GGML_CUDA_DMMV_X
@@ -502,6 +525,14 @@ static size_t g_scratch_offset = 0;
 
 static cublasHandle_t g_cublas_handles[GGML_CUDA_MAX_DEVICES] = {nullptr};
 
+[[noreturn]]
+static __device__ void bad_arch() {
+    printf("ERROR: ggml-cuda was compiled without support for the current GPU architecture.\n");
+    __trap();
+
+    (void) bad_arch; // suppress unused function warning
+}
+
 static __device__ __forceinline__ float warp_reduce_sum(float x) {
 #pragma unroll
     for (int mask = 16; mask > 0; mask >>= 1) {
@@ -527,40 +558,105 @@ static __device__ __forceinline__ float warp_reduce_max(float x) {
     return x;
 }
 
-static __global__ void add_f32(const float * x, const float * y, float * dst, const int kx, const int ky) {
-    const int i = blockDim.x*blockIdx.x + threadIdx.x;
-
-    if (i >= kx) {
-        return;
-    }
-    dst[i] = x[i] + y[i%ky];
+static __device__ __forceinline__ float op_repeat(const float a, const float b) {
+    return b;
 }
 
-static __global__ void add_f16_f32_f16(const half * x, const float * y, half * dst, const int k) {
-    const int i = blockDim.x*blockIdx.x + threadIdx.x;
-
-    if (i >= k) {
-        return;
-    }
-    dst[i] = __hadd(x[i], __float2half(y[i]));
+static __device__ __forceinline__ float op_add(const float a, const float b) {
+    return a + b;
 }
 
-static __global__ void add_f16_f32_f32(const half * x, const float * y, float * dst, const int k) {
-    const int i = blockDim.x*blockIdx.x + threadIdx.x;
-
-    if (i >= k) {
-        return;
-    }
-    dst[i] = __half2float(x[i]) + y[i];
+static __device__ __forceinline__ float op_mul(const float a, const float b) {
+    return a * b;
 }
 
-static __global__ void mul_f32(const float * x, const float * y, float * dst, const int kx, const int ky) {
-    const int i = blockDim.x*blockIdx.x + threadIdx.x;
+static __device__ __forceinline__ float op_div(const float a, const float b) {
+    return a / b;
+}
 
-    if (i >= kx) {
+template<float (*bin_op)(const float, const float), typename src0_t, typename src1_t, typename dst_t>
+static __global__ void k_bin_bcast(const src0_t * src0, const src1_t * src1, dst_t * dst,
+        int ne0, int ne1, int ne2, int ne3,
+        int ne10, int ne11, int ne12, int ne13,
+        /*int s0, */ int s1,  int s2,  int s3,
+        /*int s10,*/ int s11, int s12, int s13) {
+    const int i0s = blockDim.x*blockIdx.x + threadIdx.x;
+    const int i1 = (blockDim.y*blockIdx.y + threadIdx.y);
+    const int i2 = (blockDim.z*blockIdx.z + threadIdx.z) / ne3;
+    const int i3 = (blockDim.z*blockIdx.z + threadIdx.z) % ne3;
+
+    if (i0s >= ne0 || i1 >= ne1 || i2 >= ne2 || i3 >= ne3) {
         return;
     }
-    dst[i] = x[i] * y[i%ky];
+
+    const int i11 = i1 % ne11;
+    const int i12 = i2 % ne12;
+    const int i13 = i3 % ne13;
+
+    const size_t i_src0 = i3*s3 + i2*s2 + i1*s1;
+    const size_t i_src1 = i13*s13 + i12*s12 + i11*s11;
+    const size_t i_dst  = i_src0;
+
+    const src0_t * src0_row = src0 + i_src0;
+    const src1_t * src1_row = src1 + i_src1;
+    dst_t * dst_row = dst + i_dst;
+
+    for (int i0 = i0s; i0 < ne0; i0 += blockDim.x*gridDim.x) {
+        const int i10 = i0 % ne10;
+        dst_row[i0] = (dst_t)bin_op(src0 ? (float)src0_row[i0] : 0.0f, (float)src1_row[i10]);
+    }
+}
+
+template<float (*bin_op)(const float, const float), typename src0_t, typename src1_t, typename dst_t>
+static __global__ void k_bin_bcast_unravel(const src0_t * src0, const src1_t * src1, dst_t * dst,
+        int ne0, int ne1, int ne2, int ne3,
+        int ne10, int ne11, int ne12, int ne13,
+        /*int s0, */ int s1,  int s2,  int s3,
+        /*int s10,*/ int s11, int s12, int s13) {
+
+    const int i = blockDim.x*blockIdx.x + threadIdx.x;
+
+    const int i3 = i/(ne2*ne1*ne0);
+    const int i2 = (i/(ne1*ne0)) % ne2;
+    const int i1 = (i/ne0) % ne1;
+    const int i0 = i % ne0;
+
+    if (i0 >= ne0 || i1 >= ne1 || i2 >= ne2 || i3 >= ne3) {
+        return;
+    }
+
+    const int i11 = i1 % ne11;
+    const int i12 = i2 % ne12;
+    const int i13 = i3 % ne13;
+
+    const size_t i_src0 = i3*s3 + i2*s2 + i1*s1;
+    const size_t i_src1 = i13*s13 + i12*s12 + i11*s11;
+    const size_t i_dst  = i_src0;
+
+    const src0_t * src0_row = src0 + i_src0;
+    const src1_t * src1_row = src1 + i_src1;
+    dst_t * dst_row = dst + i_dst;
+
+    const int i10 = i0 % ne10;
+    dst_row[i0] = (dst_t)bin_op(src0 ? (float)src0_row[i0] : 0.0f, (float)src1_row[i10]);
+}
+
+static __global__ void acc_f32(const float * x, const float * y, float * dst, const int ne,
+    const int ne10, const int ne11, const int ne12,
+    const int nb1, const int nb2, int offset) {
+    const int i = blockDim.x * blockIdx.x + threadIdx.x;
+    if (i >= ne) {
+        return;
+    }
+    int src1_idx = i - offset;
+    int oz = src1_idx / nb2;
+    int oy = (src1_idx - (oz * nb2)) / nb1;
+    int ox = src1_idx % nb1;
+    if (src1_idx >= 0 && ox < ne10 && oy < ne11 && oz < ne12) {
+        dst[i] = x[i] + y[ox + oy * ne10 + oz * ne10 * ne11];
+    } else {
+        dst[i] = x[i];
+    }
 }
 
 static __global__ void gelu_f32(const float * x, float * dst, const int k) {
@@ -585,6 +681,23 @@ static __global__ void silu_f32(const float * x, float * dst, const int k) {
     dst[i] = x[i] / (1.0f + expf(-x[i]));
 }
 
+static __global__ void gelu_quick_f32(const float *x, float *dst, int k) {
+    const float GELU_QUICK_COEF = -1.702f;
+    const int i  = blockDim.x*blockIdx.x + threadIdx.x;
+    if (i >= k) {
+        return;
+    }
+    dst[i] = x[i] * (1.0f / (1.0f + expf(GELU_QUICK_COEF * x[i])));
+}
+
+static __global__ void tanh_f32(const float *x, float *dst, int k) {
+    const int i  = blockDim.x*blockIdx.x + threadIdx.x;
+    if (i >= k) {
+        return;
+    }
+    dst[i] = tanhf(x[i]);
+}
+
 static __global__ void relu_f32(const float * x, float * dst, const int k) {
     const int i = blockDim.x*blockIdx.x + threadIdx.x;
 
@@ -594,6 +707,14 @@ static __global__ void relu_f32(const float * x, float * dst, const int k) {
     dst[i] = fmaxf(x[i], 0);
 }
 
+static __global__ void leaky_relu_f32(const float *x, float *dst, const int k, const float negative_slope) {
+    const int i  = blockDim.x*blockIdx.x + threadIdx.x;
+    if (i >= k) {
+        return;
+    }
+    dst[i] = fmaxf(x[i], 0) + fminf(x[i], 0.0f) * negative_slope;
+}
+
 static __global__ void sqr_f32(const float * x, float * dst, const int k) {
     const int i = blockDim.x*blockIdx.x + threadIdx.x;
 
@@ -604,12 +725,10 @@ static __global__ void sqr_f32(const float * x, float * dst, const int k) {
 }
 
 template <int block_size>
-static __global__ void norm_f32(const float * x, float * dst, const int ncols) {
+static __global__ void norm_f32(const float * x, float * dst, const int ncols, const float eps) {
     const int row = blockIdx.x*blockDim.y + threadIdx.y;
     const int tid = threadIdx.x;
 
-    const float eps = 1e-5f;
-
     float2 mean_var = make_float2(0.f, 0.f);
 
     for (int col = tid; col < ncols; col += block_size) {
@@ -641,6 +760,132 @@ static __global__ void norm_f32(const float * x, float * dst, const int ncols) {
     }
 }
 
+static __global__ void concat_f32(const float  *x,const float  *y, float *dst, const int ne0, const int ne02) {
+    int nidx = threadIdx.x + blockIdx.x * blockDim.x;
+    if (nidx >= ne0) {
+        return;
+    }
+    // operation
+    int offset_dst =
+        nidx +
+        blockIdx.y * ne0 +
+        blockIdx.z * ne0 * gridDim.y;
+    if (blockIdx.z < ne02) { // src0
+        int offset_src =
+            nidx +
+            blockIdx.y * ne0 +
+            blockIdx.z * ne0 * gridDim.y;
+            dst[offset_dst] = x[offset_src];
+    } else {
+        int offset_src =
+            nidx +
+            blockIdx.y * ne0 +
+            (blockIdx.z - ne02) * ne0 *  gridDim.y;
+            dst[offset_dst] = y[offset_src];
+    }
+}
+
+static __global__ void upscale_f32(const float  *x, float *dst, const int ne00, const int nb02, const int scale_factor) {
+    int ne0 = ne00 * scale_factor;
+    int nidx = threadIdx.x + blockIdx.x * blockDim.x;
+    if (nidx >= ne0) {
+        return;
+    }
+    // operation
+    int i00 = nidx / scale_factor;
+    int i01 = blockIdx.y / scale_factor;
+    int offset_src =
+        i00 +
+        i01 * ne00 +
+        blockIdx.z * nb02;
+    int offset_dst =
+        nidx +
+        blockIdx.y * ne0 +
+        blockIdx.z * ne0 * gridDim.y;
+    dst[offset_dst] = x[offset_src];
+}
+
+static __global__ void pad_f32(const float  *x, float *dst, const int ne0, const int ne00, const int ne01, const int ne02) {
+    int nidx = threadIdx.x + blockIdx.x * blockDim.x;
+    if (nidx >= ne0) {
+        return;
+    }
+
+    // operation
+    int offset_dst =
+        nidx +
+        blockIdx.y * ne0 +
+        blockIdx.z * ne0 * gridDim.y;
+    if (nidx < ne00 && blockIdx.y < ne01 && blockIdx.z < ne02) {
+        int offset_src =
+            nidx +
+            blockIdx.y * ne00 +
+            blockIdx.z * ne00 * ne01;
+            dst[offset_dst] = x[offset_src];
+    } else {
+        dst[offset_dst] = 0.0f;
+    }
+}
+
+template <int block_size>
+static __global__ void group_norm_f32(const float * x, float * dst, const int group_size, const int ne_elements, const float eps) {
+    int start = blockIdx.x * group_size;
+    int end = start + group_size;
+
+    start += threadIdx.x;
+
+    if (end >= ne_elements) {
+        end = ne_elements;
+    }
+
+    float tmp = 0.0f; // partial sum for thread in warp
+
+    for (int j = start; j < end; j += block_size) {
+        tmp += x[j];
+    }
+
+    tmp = warp_reduce_sum(tmp);
+    if (block_size > WARP_SIZE) {
+        __shared__ float s_sum[32];
+        int warp_id = threadIdx.x / WARP_SIZE;
+        int lane_id = threadIdx.x % WARP_SIZE;
+        if (lane_id == 0) {
+            s_sum[warp_id] = tmp;
+        }
+        __syncthreads();
+        tmp = s_sum[lane_id];
+        tmp = warp_reduce_sum(tmp);
+    }
+
+    float mean = tmp / group_size;
+    tmp = 0.0f;
+
+    for (int j = start; j < end; j += block_size) {
+        float xi = x[j] - mean;
+        dst[j] = xi;
+        tmp += xi * xi;
+    }
+
+    tmp = warp_reduce_sum(tmp);
+    if (block_size > WARP_SIZE) {
+        __shared__ float s_sum[32];
+        int warp_id = threadIdx.x / WARP_SIZE;
+        int lane_id = threadIdx.x % WARP_SIZE;
+        if (lane_id == 0) {
+            s_sum[warp_id] = tmp;
+        }
+        __syncthreads();
+        tmp = s_sum[lane_id];
+        tmp = warp_reduce_sum(tmp);
+    }
+
+    float variance = tmp / group_size;
+    float scale = rsqrtf(variance + eps);
+    for (int j = start; j < end; j += block_size) {
+        dst[j] *= scale;
+    }
+}
+
 template <int block_size>
 static __global__ void rms_norm_f32(const float * x, float * dst, const int ncols, const float eps) {
     const int row = blockIdx.x*blockDim.y + threadIdx.y;
@@ -1639,31 +1884,65 @@ static __global__ void quantize_q8_1(const float * __restrict__ x, void * __rest
 }
 
 template<int qk, int qr, dequantize_kernel_t dequantize_kernel, typename dst_t>
-static __global__ void k_get_rows(const void * x, const int32_t * y, dst_t * dst, const int ncols) {
-    const int col = (blockIdx.x*blockDim.x + threadIdx.x)*2;
-    const int row = blockDim.y*blockIdx.y + threadIdx.y;
+static __global__ void k_get_rows(
+            const void * src0, const int32_t * src1, dst_t * dst,
+            int64_t ne00, /*int64_t ne01, int64_t ne02, int64_t ne03,*/
+            /*int64_t ne10, int64_t ne11,*/ int64_t ne12, /*int64_t ne13,*/
+            /*size_t s0,*/ size_t s1, size_t s2, size_t s3,
+            /*size_t nb00,*/ size_t nb01, size_t nb02, size_t nb03,
+            size_t s10, size_t s11, size_t s12/*, size_t s13*/) {
 
-    if (col >= ncols) {
+    const int i00 = (blockIdx.x*blockDim.x + threadIdx.x)*2;
+    const int i10 = blockDim.y*blockIdx.y + threadIdx.y;
+    const int i11 = (blockIdx.z*blockDim.z + threadIdx.z)/ne12;
+    const int i12 = (blockIdx.z*blockDim.z + threadIdx.z)%ne12;
+
+    if (i00 >= ne00) {
         return;
     }
 
-    const int r = y[row];
+    const int i01 = src1[i10*s10 + i11*s11 + i12*s12];
 
-    // copy x[r*ncols + col] to dst[row*ncols + col]
-    const int xi = r*ncols + col;
-    const int di = row*ncols + col;
+    dst_t * dst_row = dst + i10*s1 + i11*s2 + i12*s3;
+    const void * src0_row = (const char *)src0 + i01*nb01 + i11*nb02 + i12*nb03;
 
-    const int ib = xi/qk; // block index
-    const int iqs = (xi%qk)/qr; // quant index
-    const int iybs = di - di%qk; // y block start index
+    const int ib = i00/qk; // block index
+    const int iqs = (i00%qk)/qr; // quant index
+    const int iybs = i00 - i00%qk; // dst block start index
     const int y_offset = qr == 1 ? 1 : qk/2;
 
     // dequantize
     dfloat2 v;
-    dequantize_kernel(x, ib, iqs, v);
+    dequantize_kernel(src0_row, ib, iqs, v);
 
-    dst[iybs + iqs + 0]        = v.x;
-    dst[iybs + iqs + y_offset] = v.y;
+    dst_row[iybs + iqs + 0]        = v.x;
+    dst_row[iybs + iqs + y_offset] = v.y;
+}
+
+template<typename src0_t, typename dst_t>
+static __global__ void k_get_rows_float(
+            const src0_t * src0, const int32_t * src1, dst_t * dst,
+            int64_t ne00, /*int64_t ne01, int64_t ne02, int64_t ne03,*/
+            /*int64_t ne10, int64_t ne11,*/ int64_t ne12, /*int64_t ne13,*/
+            /*size_t s0,*/ size_t s1, size_t s2, size_t s3,
+            /*size_t nb00,*/ size_t nb01, size_t nb02, size_t nb03,
+            size_t s10, size_t s11, size_t s12/*, size_t s13*/) {
+
+    const int i00 = blockIdx.x*blockDim.x + threadIdx.x;
+    const int i10 = blockDim.y*blockIdx.y + threadIdx.y;
+    const int i11 = (blockIdx.z*blockDim.z + threadIdx.z)/ne12;
+    const int i12 = (blockIdx.z*blockDim.z + threadIdx.z)%ne12;
+
+    if (i00 >= ne00) {
+        return;
+    }
+
+    const int i01 = src1[i10*s10 + i11*s11 + i12*s12];
+
+    dst_t * dst_row = dst + i10*s1 + i11*s2 + i12*s3;
+    const src0_t * src0_row = (const src0_t *)((const char *)src0 + i01*nb01 + i11*nb02 + i12*nb03);
+
+    dst_row[i00] = src0_row[i00];
 }
 
 template <int qk, int qr, dequantize_kernel_t dequantize_kernel, typename dst_t>
@@ -1714,8 +1993,7 @@ template <int vdr> static __device__ __forceinline__ float vec_dot_q4_0_q8_1_imp
     // second part effectively subtracts 8 from each quant value
     return d4 * (sumi * ds8f.x - (8*vdr/QI4_0) * ds8f.y);
 #else
-    assert(false);
-    return 0.0f; // only to satisfy the compiler
+    bad_arch();
 #endif // __CUDA_ARCH__ >= MIN_CC_DP4A
 }
 
@@ -1752,8 +2030,7 @@ template <int vdr> static __device__ __forceinline__ float vec_dot_q4_1_q8_1_imp
     // scale second part of sum by QI8_1/(vdr * QR4_1) to compensate for multiple threads adding it
     return sumi * d4d8 + m4s8 / (QI8_1 / (vdr * QR4_1));
 #else
-    assert(false);
-    return 0.0f; // only to satisfy the compiler
+    bad_arch();
 #endif // __CUDA_ARCH__ >= MIN_CC_DP4A
 }
 
@@ -1788,8 +2065,7 @@ template <int vdr> static __device__ __forceinline__ float vec_dot_q5_0_q8_1_imp
     // second part effectively subtracts 16 from each quant value
     return d5 * (sumi * ds8f.x - (16*vdr/QI5_0) * ds8f.y);
 #else
-    assert(false);
-    return 0.0f; // only to satisfy the compiler
+    bad_arch();
 #endif // __CUDA_ARCH__ >= MIN_CC_DP4A
 }
 
@@ -1834,8 +2110,7 @@ template <int vdr> static __device__ __forceinline__ float vec_dot_q5_1_q8_1_imp
     return sumi*d5d8 + m5s8 / (QI5_1 / vdr);
 
 #else
-    assert(false);
-    return 0.0f; // only to satisfy the compiler
+    bad_arch();
 #endif // __CUDA_ARCH__ >= MIN_CC_DP4A
 }
 
@@ -1856,8 +2131,7 @@ template <int vdr> static __device__ __forceinline__ float vec_dot_q8_0_q8_1_imp
 
     return d8_0*d8_1 * sumi;
 #else
-    assert(false);
-    return 0.0f; // only to satisfy the compiler
+    bad_arch();
 #endif // __CUDA_ARCH__ >= MIN_CC_DP4A
 }
 
@@ -1887,8 +2161,7 @@ template <int vdr> static __device__ __forceinline__ float vec_dot_q8_1_q8_1_imp
     // scale second part of sum by QI8_1/ vdr to compensate for multiple threads adding it
     return sumi*d8d8 + m8s8 / (QI8_1 / vdr);
 #else
-    assert(false);
-    return 0.0f; // only to satisfy the compiler
+    bad_arch();
 #endif // __CUDA_ARCH__ >= MIN_CC_DP4A
 }
 
@@ -1923,8 +2196,7 @@ static __device__ __forceinline__ float vec_dot_q2_K_q8_1_impl_mmvq(
 
     return dm2f.x*sumf_d - dm2f.y*sumf_m;
 #else
-    assert(false);
-    return 0.0f; // only to satisfy the compiler
+    bad_arch();
 #endif // __CUDA_ARCH__ >= MIN_CC_DP4A
 }
 
@@ -1961,8 +2233,7 @@ static __device__ __forceinline__ float vec_dot_q2_K_q8_1_impl_mmq(
 
     return d8 * (dm2f.x*sumi_d - dm2f.y*sumi_m);
 #else
-    assert(false);
-    return 0.0f; // only to satisfy the compiler
+    bad_arch();
 #endif // __CUDA_ARCH__ >= MIN_CC_DP4A
 }
 
@@ -2002,8 +2273,7 @@ static __device__ __forceinline__ float vec_dot_q3_K_q8_1_impl_mmvq(
 
     return d3 * sumf;
 #else
-    assert(false);
-    return 0.0f; // only to satisfy the compiler
+    bad_arch();
 #endif // __CUDA_ARCH__ >= MIN_CC_DP4A
 }
 
@@ -2028,8 +2298,7 @@ static __device__ __forceinline__ float vec_dot_q3_K_q8_1_impl_mmq(
 
     return d3*d8 * sumi;
 #else
-    assert(false);
-    return 0.0f; // only to satisfy the compiler
+    bad_arch();
 #endif // __CUDA_ARCH__ >= MIN_CC_DP4A
 }
 
@@ -2062,8 +2331,7 @@ static __device__ __forceinline__ float vec_dot_q4_K_q8_1_impl_vmmq(
     return dm4f.x*sumf_d - dm4f.y*sumf_m;
 
 #else
-    assert(false);
-    return 0.0f; // only to satisfy the compiler
+    bad_arch();
 #endif // __CUDA_ARCH__ >= MIN_CC_DP4A
 }
 
@@ -2096,8 +2364,7 @@ static __device__ __forceinline__ float vec_dot_q4_K_q8_1_impl_mmq(
     return dm4f.x*sumf_d - dm4f.y*sumf_m;
 
 #else
-    assert(false);
-    return 0.0f; // only to satisfy the compiler
+    bad_arch();
 #endif // __CUDA_ARCH__ >= MIN_CC_DP4A
 }
 
@@ -2137,8 +2404,7 @@ static __device__ __forceinline__ float vec_dot_q5_K_q8_1_impl_vmmq(
     return dm5f.x*sumf_d - dm5f.y*sumf_m;
 
 #else
-    assert(false);
-    return 0.0f; // only to satisfy the compiler
+    bad_arch();
 #endif // __CUDA_ARCH__ >= MIN_CC_DP4A
 }
 
@@ -2171,8 +2437,7 @@ static __device__ __forceinline__ float vec_dot_q5_K_q8_1_impl_mmq(
     return dm4f.x*sumf_d - dm4f.y*sumf_m;
 
 #else
-    assert(false);
-    return 0.0f; // only to satisfy the compiler
+    bad_arch();
 #endif // __CUDA_ARCH__ >= MIN_CC_DP4A
 }
 
@@ -2202,8 +2467,7 @@ static __device__ __forceinline__ float vec_dot_q6_K_q8_1_impl_mmvq(
 
     return d*sumf;
 #else
-    assert(false);
-    return 0.0f; // only to satisfy the compiler
+    bad_arch();
 #endif // __CUDA_ARCH__ >= MIN_CC_DP4A
 }
 
@@ -2234,8 +2498,7 @@ static __device__ __forceinline__ float vec_dot_q6_K_q8_1_impl_mmq(
     return d6 * sumf_d;
 
 #else
-    assert(false);
-    return 0.0f; // only to satisfy the compiler
+    bad_arch();
 #endif // __CUDA_ARCH__ >= MIN_CC_DP4A
 }
 
@@ -3101,8 +3364,7 @@ static __device__ __forceinline__ float vec_dot_q4_K_q8_1(
     return dall * sumf_d - dmin * sumf_m;
 
 #else
-    assert(false);
-    return 0.0f; // only to satisfy the compiler
+    bad_arch();
 #endif // __CUDA_ARCH__ >= MIN_CC_DP4A
 
 #endif
@@ -3285,8 +3547,7 @@ static __device__ __forceinline__ float vec_dot_q5_K_q8_1(
     return d * sumf_d;
 
 #else
-    assert(false);
-    return 0.0f; // only to satisfy the compiler
+    bad_arch();
 #endif // __CUDA_ARCH__ >= MIN_CC_DP4A
 
 #endif
@@ -3696,7 +3957,7 @@ template <bool need_check> static __global__ void
         (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
 #else
     (void) vec_dot_q4_0_q8_1_mul_mat;
-    assert(false);
+    bad_arch();
 #endif // __CUDA_ARCH__ >= CC_VOLTA
 }
 
@@ -3765,7 +4026,7 @@ template <bool need_check> static __global__ void
         (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
 #else
     (void) vec_dot_q4_1_q8_1_mul_mat;
-    assert(false);
+    bad_arch();
 #endif // __CUDA_ARCH__ >= CC_VOLTA
 }
 
@@ -3832,7 +4093,7 @@ template <bool need_check> static __global__ void
         (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
 #else
     (void) vec_dot_q5_0_q8_1_mul_mat;
-    assert(false);
+    bad_arch();
 #endif // __CUDA_ARCH__ >= CC_VOLTA
 }
 
@@ -3899,7 +4160,7 @@ mul_mat_q5_1(
         (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
 #else
     (void) vec_dot_q5_1_q8_1_mul_mat;
-    assert(false);
+    bad_arch();
 #endif // __CUDA_ARCH__ >= CC_VOLTA
 }
 
@@ -3966,7 +4227,7 @@ template <bool need_check> static __global__ void
         (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
 #else
     (void) vec_dot_q8_0_q8_1_mul_mat;
-    assert(false);
+    bad_arch();
 #endif // __CUDA_ARCH__ >= CC_VOLTA
 }
 
@@ -4033,7 +4294,7 @@ mul_mat_q2_K(
         (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
 #else
     (void) vec_dot_q2_K_q8_1_mul_mat;
-    assert(false);
+    bad_arch();
 #endif // __CUDA_ARCH__ >= CC_VOLTA
 }
 
@@ -4102,7 +4363,7 @@ template <bool need_check> static __global__ void
         (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
 #else
     (void) vec_dot_q3_K_q8_1_mul_mat;
-    assert(false);
+    bad_arch();
 #endif // __CUDA_ARCH__ >= CC_VOLTA
 }
 
@@ -4171,7 +4432,7 @@ template <bool need_check> static __global__ void
         (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
 #else
     (void) vec_dot_q4_K_q8_1_mul_mat;
-    assert(false);
+    bad_arch();
 #endif // __CUDA_ARCH__ >= CC_VOLTA
 }
 
@@ -4238,7 +4499,7 @@ mul_mat_q5_K(
         (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
 #else
     (void) vec_dot_q5_K_q8_1_mul_mat;
-    assert(false);
+    bad_arch();
 #endif // __CUDA_ARCH__ >= CC_VOLTA
 }
 
@@ -4307,7 +4568,7 @@ template <bool need_check> static __global__ void
         (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
 #else
     (void) vec_dot_q6_K_q8_1_mul_mat;
-    assert(false);
+    bad_arch();
 #endif // __CUDA_ARCH__ >= CC_VOLTA
 }
 
@@ -4559,6 +4820,116 @@ static __global__ void cpy_f32_f16(const char * cx, char * cdst, const int ne,
     cpy_1(cx + x_offset, cdst + dst_offset);
 }
 
+static __device__ void cpy_blck_f32_q8_0(const char * cxi, char * cdsti) {
+    const float * xi = (const float *) cxi;
+    block_q8_0 * dsti = (block_q8_0 *) cdsti;
+
+    float amax = 0.0f; // absolute max
+
+    for (int j = 0; j < QK8_0; j++) {
+        const float v = xi[j];
+        amax = fmaxf(amax, fabsf(v));
+    }
+
+    const float d = amax / ((1 << 7) - 1);
+    const float id = d ? 1.0f/d : 0.0f;
+
+    dsti->d = d;
+
+    for (int j = 0; j < QK8_0; ++j) {
+        const float x0 = xi[j]*id;
+
+        dsti->qs[j] = roundf(x0);
+    }
+}
+
+static __device__ void cpy_blck_f32_q4_0(const char * cxi, char * cdsti) {
+    const float * xi = (const float *) cxi;
+    block_q4_0 * dsti = (block_q4_0 *) cdsti;
+
+    float amax = 0.0f;
+    float vmax = 0.0f;
+
+    for (int j = 0; j < QK4_0; ++j) {
+        const float v = xi[j];
+        if (amax < fabsf(v)) {
+            amax = fabsf(v);
+            vmax = v;
+        }
+    }
+
+    const float d  = vmax / -8;
+    const float id = d ? 1.0f/d : 0.0f;
+
+    dsti->d = d;
+
+    for (int j = 0; j < QK4_0/2; ++j) {
+        const float x0 = xi[0       + j]*id;
+        const float x1 = xi[QK4_0/2 + j]*id;
+
+        const uint8_t xi0 = min(15, (int8_t)(x0 + 8.5f));
+        const uint8_t xi1 = min(15, (int8_t)(x1 + 8.5f));
+
+        dsti->qs[j]  = xi0;
+        dsti->qs[j] |= xi1 << 4;
+    }
+}
+
+static __device__ void cpy_blck_f32_q4_1(const char * cxi, char * cdsti) {
+    const float * xi = (const float *) cxi;
+    block_q4_1 * dsti = (block_q4_1 *) cdsti;
+
+    float vmin = FLT_MAX;
+    float vmax = -FLT_MAX;
+
+    for (int j = 0; j < QK4_1; ++j) {
+        const float v = xi[j];
+
+        if (v < vmin) vmin = v;
+        if (v > vmax) vmax = v;
+    }
+
+    const float d  = (vmax - vmin) / ((1 << 4) - 1);
+    const float id = d ? 1.0f/d : 0.0f;
+
+    dsti->dm.x = d;
+    dsti->dm.y = vmin;
+
+    for (int j = 0; j < QK4_1/2; ++j) {
+        const float x0 = (xi[0       + j] - vmin)*id;
+        const float x1 = (xi[QK4_1/2 + j] - vmin)*id;
+
+        const uint8_t xi0 = min(15, (int8_t)(x0 + 0.5f));
+        const uint8_t xi1 = min(15, (int8_t)(x1 + 0.5f));
+
+        dsti->qs[j]  = xi0;
+        dsti->qs[j] |= xi1 << 4;
+    }
+}
+
+template <cpy_kernel_t cpy_blck, int qk>
+static __global__ void cpy_f32_q(const char * cx, char * cdst, const int ne,
+                                 const int ne00, const int ne01, const int nb00, const int nb01, const int nb02,
+                                 const int ne10, const int ne11, const int nb10, const int nb11, const int nb12) {
+    const int i = (blockDim.x*blockIdx.x + threadIdx.x)*qk;
+
+    if (i >= ne) {
+        return;
+    }
+
+    const int i02 = i / (ne00*ne01);
+    const int i01 = (i - i02*ne01*ne00) / ne00;
+    const int i00 = (i - i02*ne01*ne00 - i01*ne00);
+    const int x_offset = i00*nb00 + i01*nb01 + i02*nb02;
+
+    const int i12 = i / (ne10*ne11);
+    const int i11 = (i - i12*ne10*ne11) / ne10;
+    const int i10 = (i - i12*ne10*ne11 - i11*ne10)/qk;
+    const int dst_offset = i10*nb10 + i11*nb11 + i12*nb12;
+
+    cpy_blck(cx + x_offset, cdst + dst_offset);
+}
+
 static __device__ float rope_yarn_ramp(const float low, const float high, const int i0) {
     const float y = (i0 / 2 - low) / max(0.001f, high - low);
     return 1.0f - min(1.0f, max(0.0f, y));
@@ -4632,7 +5003,16 @@ static __global__ void rope_neox(
     const int ib = col / n_dims;
     const int ic = col % n_dims;
 
-    const int i = row*ncols + ib*n_dims + ic/2;
+    if (ib > 0) {
+        const int i = row*ncols + ib*n_dims + ic;
+
+        dst[i + 0] = x[i + 0];
+        dst[i + 1] = x[i + 1];
+
+        return;
+    }
+
+    const int i  = row*ncols + ib*n_dims + ic/2;
     const int i2 = row/p_delta_rows;
 
     float cur_rot = inv_ndims * ic - ib;
@@ -4713,6 +5093,65 @@ static __global__ void alibi_f32(const float * x, float * dst, const int ncols,
     dst[i] = col * m_k + x[i];
 }
 
+static __global__ void k_sum_rows_f32(const float * x, float * dst, const int ncols) {
+    const int row = blockIdx.y;
+    const int col = threadIdx.x;
+
+    float sum = 0.0f;
+    for (int i = col; i < ncols; i += blockDim.x) {
+        sum += x[row * ncols + i];
+    }
+
+    sum = warp_reduce_sum(sum);
+
+    if (col == 0) {
+        dst[row] = sum;
+    }
+}
+
+template<typename T>
+static inline __device__ void swap(T & a, T & b) {
+    T tmp = a;
+    a = b;
+    b = tmp;
+}
+
+template<ggml_sort_order order>
+static __global__ void k_argsort_f32_i32(const float * x, int * dst, const int ncols) {
+    // bitonic sort
+    int col = threadIdx.x;
+    int row = blockIdx.y;
+
+    if (col >= ncols) return;
+
+    const float * x_row = x + row * ncols;
+    int * dst_row = dst + row * ncols;
+
+    // initialize indices
+    if (col < ncols) {
+        dst_row[col] = col;
+    }
+    __syncthreads();
+
+    for (int k = 2; k <= ncols; k *= 2) {
+        for (int j = k / 2; j > 0; j /= 2) {
+            int ixj = col ^ j;
+            if (ixj > col) {
+                if ((col & k) == 0) {
+                    if (order == GGML_SORT_ASC ? x_row[dst_row[col]] > x_row[dst_row[ixj]] : x_row[dst_row[col]] < x_row[dst_row[ixj]]) {
+                        swap(dst_row[col], dst_row[ixj]);
+                    }
+                } else {
+                    if (order == GGML_SORT_ASC ? x_row[dst_row[col]] < x_row[dst_row[ixj]] : x_row[dst_row[col]] > x_row[dst_row[ixj]]) {
+                        swap(dst_row[col], dst_row[ixj]);
+                    }
+                }
+            }
+            __syncthreads();
+        }
+    }
+}
+
 static __global__ void diag_mask_inf_f32(const float * x, float * dst, const int ncols, const int rows_per_channel, const int n_past) {
     const int col = blockDim.y*blockIdx.y + threadIdx.y;
     const int row = blockDim.x*blockIdx.x + threadIdx.x;
@@ -4722,8 +5161,9 @@ static __global__ void diag_mask_inf_f32(const float * x, float * dst, const int
     }
 
     const int i = row*ncols + col;
-    // dst[i] = col > n_past + row ? -INFINITY : x[i];
-    dst[i] = x[i] - (col > n_past + row % rows_per_channel) * INT_MAX; // equivalent within rounding error but slightly faster on GPU
+    //dst[i] = col > (n_past + row % rows_per_channel) ? -INFINITY : x[i];
+    //dst[i] = x[i] - (col > n_past + row % rows_per_channel) * INT_MAX; // equivalent within rounding error but slightly faster on GPU
+    dst[i] = x[i] - (col > n_past + row % rows_per_channel) * FLT_MAX;
 }
 
 static __global__ void soft_max_f32(const float * x, const float * y, float * dst, const int ncols, const int nrows_y, const float scale) {
@@ -4820,49 +5260,220 @@ static __global__ void clamp_f32(const float * x, float * dst, const float min,
 
 static  __global__ void im2col_f32_f16(
         const float * x, half * dst,
-        int ofs0, int ofs1, int IW, int IH, int CHW,
+        int offset_delta, int IW, int IH, int OW, int KW, int KH, int pelements, int CHW,
         int s0, int s1, int p0, int p1, int d0, int d1) {
-    const int iiw = blockIdx.z * s0 + threadIdx.z * d0 - p0;
-    const int iih = blockIdx.y * s1 + threadIdx.y * d1 - p1;
+    const int i = threadIdx.x + blockIdx.x * blockDim.x;
+    if (i >= pelements) {
+        return;
+    }
 
-    const int offset_dst =
-        (threadIdx.x * gridDim.y * gridDim.z + blockIdx.y * gridDim.z + blockIdx.z) * CHW +
-        (blockIdx.x * (blockDim.y * blockDim.z) + threadIdx.y * blockDim.z + threadIdx.z);
+    const int ksize = OW * (KH > 1 ? KW : 1);
+    const int kx = i / ksize;
+    const int kd = kx * ksize;
+    const int ky = (i - kd) / OW;
+    const int ix = i % OW;
+
+    const int64_t iiw = ix * s0 + kx * d0 - p0;
+    const int64_t iih = blockIdx.y * s1 + ky * d1 - p1;
+
+    const int64_t offset_dst =
+        (blockIdx.y * OW + ix) * CHW +
+        (blockIdx.z * (KW * KH) + ky * KW + kx);
 
     if (iih < 0 || iih >= IH || iiw < 0 || iiw >= IW) {
         dst[offset_dst] = __float2half(0.0f);
     } else {
-        const int offset_src =  threadIdx.x * ofs0 + blockIdx.x * ofs1;
+        const int64_t offset_src = blockIdx.z * offset_delta;
         dst[offset_dst] = __float2half(x[offset_src + iih * IW + iiw]);
     }
 }
 
 template<int qk, int qr, dequantize_kernel_t dq>
-static void get_rows_cuda(const void * x, const int32_t * y, float * dst, const int nrows, const int ncols, cudaStream_t stream) {
+static void get_rows_cuda(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
+                            const void * src0_dd, const int32_t * src1_dd, float * dst_dd, cudaStream_t stream) {
+
+    GGML_TENSOR_BINARY_OP_LOCALS
+
     const dim3 block_dims(CUDA_GET_ROWS_BLOCK_SIZE, 1, 1);
-    const int block_num_x = (ncols + 2*CUDA_GET_ROWS_BLOCK_SIZE - 1) / (2*CUDA_GET_ROWS_BLOCK_SIZE);
-    const dim3 block_nums(block_num_x, nrows, 1);
-    k_get_rows<qk, qr, dq><<<block_nums, block_dims, 0, stream>>>(x, y, dst, ncols);
+    const int block_num_x = (ne00 + 2*CUDA_GET_ROWS_BLOCK_SIZE - 1) / (2*CUDA_GET_ROWS_BLOCK_SIZE);
+    const dim3 block_nums(block_num_x, ne10, ne11*ne12);
+
+    // strides in elements
+    //const size_t s0 = nb0 / ggml_element_size(dst);
+    const size_t s1 = nb1 / ggml_element_size(dst);
+    const size_t s2 = nb2 / ggml_element_size(dst);
+    const size_t s3 = nb3 / ggml_element_size(dst);
+
+    const size_t s10 = nb10 / ggml_element_size(src1);
+    const size_t s11 = nb11 / ggml_element_size(src1);
+    const size_t s12 = nb12 / ggml_element_size(src1);
+    //const size_t s13 = nb13 / ggml_element_size(src1);
+
+    GGML_ASSERT(ne00 % 2 == 0);
+
+    k_get_rows<qk, qr, dq><<<block_nums, block_dims, 0, stream>>>(
+            src0_dd, src1_dd, dst_dd,
+            ne00, /*ne01, ne02, ne03,*/
+            /*ne10, ne11,*/ ne12, /*ne13,*/
+            /* s0,*/ s1, s2, s3,
+            /* nb00,*/ nb01, nb02, nb03,
+            s10, s11, s12/*, s13*/);
+
+    (void) dst;
 }
 
-static void add_f32_cuda(const float * x, const float * y, float * dst, const int kx, const int ky, cudaStream_t stream) {
-    const int num_blocks = (kx + CUDA_ADD_BLOCK_SIZE - 1) / CUDA_ADD_BLOCK_SIZE;
-    add_f32<<<num_blocks, CUDA_ADD_BLOCK_SIZE, 0, stream>>>(x, y, dst, kx, ky);
+template<typename src0_t>
+static void get_rows_cuda_float(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
+                                const src0_t * src0_dd, const int32_t * src1_dd, float * dst_dd, cudaStream_t stream) {
+
+    GGML_TENSOR_BINARY_OP_LOCALS
+
+    const dim3 block_dims(CUDA_GET_ROWS_BLOCK_SIZE, 1, 1);
+    const int block_num_x = (ne00 + CUDA_GET_ROWS_BLOCK_SIZE - 1) / CUDA_GET_ROWS_BLOCK_SIZE;
+    const dim3 block_nums(block_num_x, ne10, ne11*ne12);
+
+    // strides in elements
+    //const size_t s0 = nb0 / ggml_element_size(dst);
+    const size_t s1 = nb1 / ggml_element_size(dst);
+    const size_t s2 = nb2 / ggml_element_size(dst);
+    const size_t s3 = nb3 / ggml_element_size(dst);
+
+    const size_t s10 = nb10 / ggml_element_size(src1);
+    const size_t s11 = nb11 / ggml_element_size(src1);
+    const size_t s12 = nb12 / ggml_element_size(src1);
+    //const size_t s13 = nb13 / ggml_element_size(src1);
+
+    k_get_rows_float<<<block_nums, block_dims, 0, stream>>>(
+            src0_dd, src1_dd, dst_dd,
+            ne00, /*ne01, ne02, ne03,*/
+            /*ne10, ne11,*/ ne12, /*ne13,*/
+            /* s0,*/ s1, s2, s3,
+            /* nb00,*/ nb01, nb02, nb03,
+            s10, s11, s12/*, s13*/);
+
+    (void) dst;
 }
 
-static void add_f16_f32_f16_cuda(const half * x, const float * y, half * dst, const int k, cudaStream_t stream) {
-    const int num_blocks = (k + CUDA_ADD_BLOCK_SIZE - 1) / CUDA_ADD_BLOCK_SIZE;
-    add_f16_f32_f16<<<num_blocks, CUDA_ADD_BLOCK_SIZE, 0, stream>>>(x, y, dst, k);
-}
+template<float (*bin_op)(const float, const float)>
+struct bin_bcast_cuda {
+    template<typename src0_t, typename src1_t, typename dst_t>
+    void operator()(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst,
+            const src0_t * src0_dd, const src1_t * src1_dd, dst_t * dst_dd,
+            cudaStream_t stream) {
 
-static void add_f16_f32_f32_cuda(const half * x, const float * y, float * dst, const int k, cudaStream_t stream) {
-    const int num_blocks = (k + CUDA_ADD_BLOCK_SIZE - 1) / CUDA_ADD_BLOCK_SIZE;
-    add_f16_f32_f32<<<num_blocks, CUDA_ADD_BLOCK_SIZE, 0, stream>>>(x, y, dst, k);
-}
+        GGML_TENSOR_BINARY_OP_LOCALS
 
-static void mul_f32_cuda(const float * x, const float * y, float * dst, const int kx, const int ky, cudaStream_t stream) {
-    const int num_blocks = (kx + CUDA_MUL_BLOCK_SIZE - 1) / CUDA_MUL_BLOCK_SIZE;
-    mul_f32<<<num_blocks, CUDA_MUL_BLOCK_SIZE, 0, stream>>>(x, y, dst, kx, ky);
+        int nr0 = ne10/ne0;
+        int nr1 = ne11/ne1;
+        int nr2 = ne12/ne2;
+        int nr3 = ne13/ne3;
+
+        int nr[4] = { nr0, nr1, nr2, nr3 };
+
+        // collapse dimensions until first broadcast dimension
+        int64_t cne0[] = {ne0, ne1, ne2, ne3};
+        int64_t cne1[] = {ne10, ne11, ne12, ne13};
+        size_t cnb0[] = {nb0, nb1, nb2, nb3};
+        size_t cnb1[] = {nb10, nb11, nb12, nb13};
+        auto collapse = [](int64_t cne[]) {
+            cne[0] *= cne[1];
+            cne[1] = cne[2];
+            cne[2] = cne[3];
+            cne[3] = 1;
+        };
+
+        auto collapse_nb = [](size_t cnb[], int64_t cne[]) {
+            cnb[1] *= cne[1];
+            cnb[2] *= cne[2];
+            cnb[3] *= cne[3];
+        };
+
+        for (int i = 0; i < 4; i++) {
+            if (nr[i] != 1) {
+                break;
+            }
+            if (i > 0) {
+                collapse_nb(cnb0, cne0);
+                collapse_nb(cnb1, cne1);
+                collapse(cne0);
+                collapse(cne1);
+            }
+        }
+        {
+            int64_t ne0 = cne0[0];
+            int64_t ne1 = cne0[1];
+            int64_t ne2 = cne0[2];
+            int64_t ne3 = cne0[3];
+
+            int64_t ne10 = cne1[0];
+            int64_t ne11 = cne1[1];
+            int64_t ne12 = cne1[2];
+            int64_t ne13 = cne1[3];
+
+            size_t nb0 = cnb0[0];
+            size_t nb1 = cnb0[1];
+            size_t nb2 = cnb0[2];
+            size_t nb3 = cnb0[3];
+
+            size_t nb10 = cnb1[0];
+            size_t nb11 = cnb1[1];
+            size_t nb12 = cnb1[2];
+            size_t nb13 = cnb1[3];
+
+            size_t s0 = nb0 / sizeof(dst_t);
+            size_t s1 = nb1 / sizeof(dst_t);
+            size_t s2 = nb2 / sizeof(dst_t);
+            size_t s3 = nb3 / sizeof(dst_t);
+
+            size_t s10 = nb10 / sizeof(src1_t);
+            size_t s11 = nb11 / sizeof(src1_t);
+            size_t s12 = nb12 / sizeof(src1_t);
+            size_t s13 = nb13 / sizeof(src1_t);
+
+            GGML_ASSERT(s0 == 1);
+            GGML_ASSERT(s10 == 1);
+
+            const int block_size = 128;
+
+            int64_t hne0 = std::max(ne0/2LL, 1LL);
+
+            dim3 block_dims;
+            block_dims.x = std::min<unsigned int>(hne0, block_size);
+            block_dims.y = std::min<unsigned int>(ne1, block_size / block_dims.x);
+            block_dims.z = std::min(std::min<unsigned int>(ne2*ne3, block_size / block_dims.x / block_dims.y), 64U);
+
+            dim3 block_nums(
+                (hne0 + block_dims.x - 1) / block_dims.x,
+                (ne1 + block_dims.y - 1) / block_dims.y,
+                (ne2*ne3 + block_dims.z - 1) / block_dims.z
+            );
+
+            if (block_nums.z > 65535) {
+                // this is the maximum number of blocks in z direction, fallback to 1D grid kernel
+                int block_num = (ne0*ne1*ne2*ne3 + block_size - 1) / block_size;
+                k_bin_bcast_unravel<bin_op><<<block_num, block_size, 0, stream>>>(
+                    src0_dd, src1_dd, dst_dd,
+                    ne0, ne1, ne2, ne3,
+                    ne10, ne11, ne12, ne13,
+                    /* s0, */ s1, s2, s3,
+                    /* s10, */ s11, s12, s13);
+            } else {
+                k_bin_bcast<bin_op><<<block_nums, block_dims, 0, stream>>>(
+                    src0_dd, src1_dd, dst_dd,
+                    ne0, ne1, ne2, ne3,
+                    ne10, ne11, ne12, ne13,
+                    /* s0, */ s1, s2, s3,
+                    /* s10, */ s11, s12, s13);
+            }
+        }
+    }
+};
+
+static void acc_f32_cuda(const float * x, const float * y, float * dst, const int n_elements,
+    const int ne10, const int ne11, const int ne12,
+    const int nb1, const int nb2, const int offset, cudaStream_t stream) {
+    int num_blocks = (n_elements + CUDA_ACC_BLOCK_SIZE - 1) / CUDA_ACC_BLOCK_SIZE;
+    acc_f32<<<num_blocks, CUDA_ACC_BLOCK_SIZE, 0, stream>>>(x, y, dst, n_elements, ne10, ne11, ne12, nb1, nb2, offset);
 }
 
 static void gelu_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) {
@@ -4875,27 +5486,74 @@ static void silu_f32_cuda(const float * x, float * dst, const int k, cudaStream_
     silu_f32<<<num_blocks, CUDA_SILU_BLOCK_SIZE, 0, stream>>>(x, dst, k);
 }
 
+static void gelu_quick_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) {
+    const int num_blocks = (k + CUDA_GELU_BLOCK_SIZE - 1) / CUDA_GELU_BLOCK_SIZE;
+    gelu_quick_f32<<<num_blocks, CUDA_GELU_BLOCK_SIZE, 0, stream>>>(x, dst, k);
+}
+
+static void tanh_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) {
+    const int num_blocks = (k + CUDA_TANH_BLOCK_SIZE - 1) / CUDA_TANH_BLOCK_SIZE;
+    tanh_f32<<<num_blocks, CUDA_TANH_BLOCK_SIZE, 0, stream>>>(x, dst, k);
+}
+
 static void relu_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) {
     const int num_blocks = (k + CUDA_RELU_BLOCK_SIZE - 1) / CUDA_RELU_BLOCK_SIZE;
     relu_f32<<<num_blocks, CUDA_RELU_BLOCK_SIZE, 0, stream>>>(x, dst, k);
 }
 
+static void leaky_relu_f32_cuda(const float * x, float * dst, const int k, const float negative_slope, cudaStream_t stream) {
+    const int num_blocks = (k + CUDA_RELU_BLOCK_SIZE - 1) / CUDA_RELU_BLOCK_SIZE;
+    leaky_relu_f32<<<num_blocks, CUDA_RELU_BLOCK_SIZE, 0, stream>>>(x, dst, k, negative_slope);
+}
+
 static void sqr_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) {
     const int num_blocks = (k + CUDA_SQR_BLOCK_SIZE - 1) / CUDA_SQR_BLOCK_SIZE;
     sqr_f32<<<num_blocks, CUDA_SQR_BLOCK_SIZE, 0, stream>>>(x, dst, k);
 }
 
-static void norm_f32_cuda(const float * x, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
+static void norm_f32_cuda(const float * x, float * dst, const int ncols, const int nrows, const float eps, cudaStream_t stream) {
     GGML_ASSERT(ncols % WARP_SIZE == 0);
     if (ncols < 1024) {
         const dim3 block_dims(WARP_SIZE, 1, 1);
-        norm_f32<WARP_SIZE><<<nrows, block_dims, 0, stream>>>(x, dst, ncols);
+        norm_f32<WARP_SIZE><<<nrows, block_dims, 0, stream>>>(x, dst, ncols, eps);
     } else {
         const dim3 block_dims(1024, 1, 1);
-        norm_f32<1024><<<nrows, block_dims, 0, stream>>>(x, dst, ncols);
+        norm_f32<1024><<<nrows, block_dims, 0, stream>>>(x, dst, ncols, eps);
     }
 }
 
+static void group_norm_f32_cuda(const float * x, float * dst, const int num_groups, const int group_size, const int ne_elements, cudaStream_t stream) {
+    static const float eps = 1e-6f;
+    if (group_size < 1024) {
+        const dim3 block_dims(WARP_SIZE, 1, 1);
+        group_norm_f32<WARP_SIZE><<<num_groups, block_dims, 0, stream>>>(x, dst, group_size, ne_elements, eps);
+    } else {
+        const dim3 block_dims(1024, 1, 1);
+        group_norm_f32<1024><<<num_groups, block_dims, 0, stream>>>(x, dst, group_size, ne_elements, eps);
+    }
+}
+
+static void concat_f32_cuda(const float * x, const float * y, float * dst, const int ne0, int ne1, int ne2, int ne02, cudaStream_t stream) {
+    int num_blocks = (ne0 + CUDA_CONCAT_BLOCK_SIZE - 1) / CUDA_CONCAT_BLOCK_SIZE;
+    dim3 gridDim(num_blocks, ne1, ne2);
+    concat_f32<<<gridDim, CUDA_CONCAT_BLOCK_SIZE, 0, stream>>>(x, y, dst, ne0, ne02);
+}
+
+static void upscale_f32_cuda(const float * x, float * dst, const int ne00, const int ne01, const int ne02, const int scale_factor, cudaStream_t stream) {
+    int ne0 = (ne00 * scale_factor);
+    int num_blocks = (ne0 + CUDA_UPSCALE_BLOCK_SIZE - 1) / CUDA_UPSCALE_BLOCK_SIZE;
+    dim3 gridDim(num_blocks, (ne01 * scale_factor), ne02);
+    upscale_f32<<<gridDim, CUDA_UPSCALE_BLOCK_SIZE, 0, stream>>>(x, dst, ne00, ne00 * ne01, scale_factor);
+}
+
+static void pad_f32_cuda(const float * x, float * dst,
+    const int ne00, const int ne01, const int ne02,
+    const int ne0, const int ne1, const int ne2, cudaStream_t stream) {
+    int num_blocks = (ne0 + CUDA_PAD_BLOCK_SIZE - 1) / CUDA_PAD_BLOCK_SIZE;
+    dim3 gridDim(num_blocks, ne1, ne2);
+    pad_f32<<<gridDim, CUDA_PAD_BLOCK_SIZE, 0, stream>>>(x, dst, ne0, ne00, ne01, ne02);
+}
+
 static void rms_norm_f32_cuda(const float * x, float * dst, const int ncols, const int nrows, const float eps, cudaStream_t stream) {
     GGML_ASSERT(ncols % WARP_SIZE == 0);
     if (ncols < 1024) {
@@ -4914,34 +5572,10 @@ static void quantize_row_q8_1_cuda(const float * x, void * vy, const int kx, con
     quantize_q8_1<<<num_blocks, block_size, 0, stream>>>(x, vy, kx, kx_padded);
 }
 
-template<typename dst_t>
-static void dequantize_row_q4_0_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
+template <int qk, int qr, dequantize_kernel_t dequantize_kernel, typename dst_t>
+static void dequantize_block_cuda(const void * __restrict__ vx, dst_t * __restrict__ y, const int k, cudaStream_t stream) {
     const int num_blocks = (k + CUDA_DEQUANTIZE_BLOCK_SIZE - 1) / CUDA_DEQUANTIZE_BLOCK_SIZE;
-    dequantize_block<QK4_0, QR4_0, dequantize_q4_0><<<num_blocks, CUDA_DEQUANTIZE_BLOCK_SIZE, 0, stream>>>(vx, y, k);
-}
-
-template<typename dst_t>
-static void dequantize_row_q4_1_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
-    const int num_blocks = (k + CUDA_DEQUANTIZE_BLOCK_SIZE - 1) / CUDA_DEQUANTIZE_BLOCK_SIZE;
-    dequantize_block<QK4_1, QR4_1, dequantize_q4_1><<<num_blocks, CUDA_DEQUANTIZE_BLOCK_SIZE, 0, stream>>>(vx, y, k);
-}
-
-template<typename dst_t>
-static void dequantize_row_q5_0_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
-    const int num_blocks = (k + CUDA_DEQUANTIZE_BLOCK_SIZE - 1) / CUDA_DEQUANTIZE_BLOCK_SIZE;
-    dequantize_block<QK5_0, QR5_0, dequantize_q5_0><<<num_blocks, CUDA_DEQUANTIZE_BLOCK_SIZE, 0, stream>>>(vx, y, k);
-}
-
-template<typename dst_t>
-static void dequantize_row_q5_1_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
-    const int num_blocks = (k + CUDA_DEQUANTIZE_BLOCK_SIZE - 1) / CUDA_DEQUANTIZE_BLOCK_SIZE;
-    dequantize_block<QK5_1, QR5_1, dequantize_q5_1><<<num_blocks, CUDA_DEQUANTIZE_BLOCK_SIZE, 0, stream>>>(vx, y, k);
-}
-
-template<typename dst_t>
-static void dequantize_row_q8_0_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
-    const int num_blocks = (k + CUDA_DEQUANTIZE_BLOCK_SIZE - 1) / CUDA_DEQUANTIZE_BLOCK_SIZE;
-    dequantize_block<QK8_0, QR8_0, dequantize_q8_0><<<num_blocks, CUDA_DEQUANTIZE_BLOCK_SIZE, 0, stream>>>(vx, y, k);
+    dequantize_block<qk, qr, dequantize_kernel><<<num_blocks, CUDA_DEQUANTIZE_BLOCK_SIZE, 0, stream>>>(vx, y, k);
 }
 
 template<typename dst_t>
@@ -4990,6 +5624,64 @@ static void dequantize_row_q6_K_cuda(const void * vx, dst_t * y, const int k, cu
 #endif
 }
 
+static to_fp16_cuda_t ggml_get_to_fp16_cuda(ggml_type type) {
+    switch (type) {
+        case GGML_TYPE_Q4_0:
+            return dequantize_block_cuda<QK4_0, QR4_0, dequantize_q4_0>;
+        case GGML_TYPE_Q4_1:
+            return dequantize_block_cuda<QK4_1, QR4_1, dequantize_q4_1>;
+        case GGML_TYPE_Q5_0:
+            return dequantize_block_cuda<QK5_0, QR5_0, dequantize_q5_0>;
+        case GGML_TYPE_Q5_1:
+            return dequantize_block_cuda<QK5_1, QR5_1, dequantize_q5_1>;
+        case GGML_TYPE_Q8_0:
+            return dequantize_block_cuda<QK8_0, QR8_0, dequantize_q8_0>;
+        case GGML_TYPE_Q2_K:
+            return dequantize_row_q2_K_cuda;
+        case GGML_TYPE_Q3_K:
+            return dequantize_row_q3_K_cuda;
+        case GGML_TYPE_Q4_K:
+            return dequantize_row_q4_K_cuda;
+        case GGML_TYPE_Q5_K:
+            return dequantize_row_q5_K_cuda;
+        case GGML_TYPE_Q6_K:
+            return dequantize_row_q6_K_cuda;
+        case GGML_TYPE_F32:
+            return dequantize_block_cuda<1, 1, convert_f32>;
+        default:
+            return nullptr;
+    }
+}
+
+static to_fp32_cuda_t ggml_get_to_fp32_cuda(ggml_type type) {
+    switch (type) {
+        case GGML_TYPE_Q4_0:
+            return dequantize_block_cuda<QK4_0, QR4_0, dequantize_q4_0>;
+        case GGML_TYPE_Q4_1:
+            return dequantize_block_cuda<QK4_1, QR4_1, dequantize_q4_1>;
+        case GGML_TYPE_Q5_0:
+            return dequantize_block_cuda<QK5_0, QR5_0, dequantize_q5_0>;
+        case GGML_TYPE_Q5_1:
+            return dequantize_block_cuda<QK5_1, QR5_1, dequantize_q5_1>;
+        case GGML_TYPE_Q8_0:
+            return dequantize_block_cuda<QK8_0, QR8_0, dequantize_q8_0>;
+        case GGML_TYPE_Q2_K:
+            return dequantize_row_q2_K_cuda;
+        case GGML_TYPE_Q3_K:
+            return dequantize_row_q3_K_cuda;
+        case GGML_TYPE_Q4_K:
+            return dequantize_row_q4_K_cuda;
+        case GGML_TYPE_Q5_K:
+            return dequantize_row_q5_K_cuda;
+        case GGML_TYPE_Q6_K:
+            return dequantize_row_q6_K_cuda;
+        case GGML_TYPE_F16:
+            return dequantize_block_cuda<1, 1, convert_f16>;
+        default:
+            return nullptr;
+    }
+}
+
 static void dequantize_mul_mat_vec_q4_0_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
     GGML_ASSERT(ncols % GGML_CUDA_DMMV_X == 0);
     const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
@@ -5078,6 +5770,15 @@ static void dequantize_mul_mat_vec_q6_K_cuda(const void * vx, const float * y, f
     dequantize_mul_mat_vec_q6_k<<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
 }
 
+static void convert_mul_mat_vec_f16_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
+    GGML_ASSERT(ncols % GGML_CUDA_DMMV_X == 0);
+    const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
+    const dim3 block_nums(block_num_y, 1, 1);
+    const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
+    dequantize_mul_mat_vec<1, 1, convert_f16>
+        <<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
+}
+
 static void mul_mat_vec_q4_0_q8_1_cuda(const void * vx, const void * vy, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
     GGML_ASSERT(ncols % QK4_0 == 0);
     const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
@@ -5168,83 +5869,6 @@ static void mul_mat_vec_q6_K_q8_1_cuda(const void * vx, const void * vy, float *
         <<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols, nrows);
 }
 
-static void convert_fp16_to_fp32_cuda(const void * vx, float * y, const int k, cudaStream_t stream) {
-    const int num_blocks = (k + CUDA_DEQUANTIZE_BLOCK_SIZE - 1) / CUDA_DEQUANTIZE_BLOCK_SIZE;
-    dequantize_block<1, 1, convert_f16><<<num_blocks, CUDA_DEQUANTIZE_BLOCK_SIZE, 0, stream>>>(vx, y, k);
-}
-
-static void convert_fp32_to_fp16_cuda(const void * vx, half * y, const int k, cudaStream_t stream) {
-    const int num_blocks = (k + CUDA_QUANTIZE_BLOCK_SIZE - 1) / CUDA_QUANTIZE_BLOCK_SIZE;
-    dequantize_block<1, 1, convert_f32><<<num_blocks, CUDA_DEQUANTIZE_BLOCK_SIZE, 0, stream>>>(vx, y, k);
-}
-
-static void convert_mul_mat_vec_f16_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
-    GGML_ASSERT(ncols % GGML_CUDA_DMMV_X == 0);
-    const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
-    const dim3 block_nums(block_num_y, 1, 1);
-    const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
-    dequantize_mul_mat_vec<1, 1, convert_f16>
-        <<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
-}
-
-static to_fp16_cuda_t ggml_get_to_fp16_cuda(ggml_type type) {
-    switch (type) {
-        case GGML_TYPE_Q4_0:
-            return dequantize_row_q4_0_cuda;
-        case GGML_TYPE_Q4_1:
-            return dequantize_row_q4_1_cuda;
-        case GGML_TYPE_Q5_0:
-            return dequantize_row_q5_0_cuda;
-        case GGML_TYPE_Q5_1:
-            return dequantize_row_q5_1_cuda;
-        case GGML_TYPE_Q8_0:
-            return dequantize_row_q8_0_cuda;
-        case GGML_TYPE_Q2_K:
-            return dequantize_row_q2_K_cuda;
-        case GGML_TYPE_Q3_K:
-            return dequantize_row_q3_K_cuda;
-        case GGML_TYPE_Q4_K:
-            return dequantize_row_q4_K_cuda;
-        case GGML_TYPE_Q5_K:
-            return dequantize_row_q5_K_cuda;
-        case GGML_TYPE_Q6_K:
-            return dequantize_row_q6_K_cuda;
-        case GGML_TYPE_F32:
-            return convert_fp32_to_fp16_cuda;
-        default:
-            return nullptr;
-    }
-}
-
-static to_fp32_cuda_t ggml_get_to_fp32_cuda(ggml_type type) {
-    switch (type) {
-        case GGML_TYPE_Q4_0:
-            return dequantize_row_q4_0_cuda;
-        case GGML_TYPE_Q4_1:
-            return dequantize_row_q4_1_cuda;
-        case GGML_TYPE_Q5_0:
-            return dequantize_row_q5_0_cuda;
-        case GGML_TYPE_Q5_1:
-            return dequantize_row_q5_1_cuda;
-        case GGML_TYPE_Q8_0:
-            return dequantize_row_q8_0_cuda;
-        case GGML_TYPE_Q2_K:
-            return dequantize_row_q2_K_cuda;
-        case GGML_TYPE_Q3_K:
-            return dequantize_row_q3_K_cuda;
-        case GGML_TYPE_Q4_K:
-            return dequantize_row_q4_K_cuda;
-        case GGML_TYPE_Q5_K:
-            return dequantize_row_q5_K_cuda;
-        case GGML_TYPE_Q6_K:
-            return dequantize_row_q6_K_cuda;
-        case GGML_TYPE_F16:
-            return convert_fp16_to_fp32_cuda;
-        default:
-            return nullptr;
-    }
-}
-
 static void ggml_mul_mat_q4_0_q8_1_cuda(
     const void * vx, const void * vy, float * dst, const int ncols_x, const int nrows_x,
     const int ncols_y, const int nrows_y, const int nrows_dst, cudaStream_t stream) {
@@ -5737,6 +6361,39 @@ static void ggml_cpy_f32_f16_cuda(
         (cx, cdst, ne, ne00, ne01, nb00, nb01, nb02, ne10, ne11, nb10, nb11, nb12);
 }
 
+static void ggml_cpy_f32_q8_0_cuda(
+    const char * cx, char * cdst, const int ne,
+    const int ne00, const int ne01, const int nb00, const int nb01, const int nb02,
+    const int ne10, const int ne11, const int nb10, const int nb11, const int nb12, cudaStream_t stream) {
+
+    GGML_ASSERT(ne % QK8_0 == 0);
+    const int num_blocks = ne / QK8_0;
+    cpy_f32_q<cpy_blck_f32_q8_0, QK8_0><<<num_blocks, 1, 0, stream>>>
+        (cx, cdst, ne, ne00, ne01, nb00, nb01, nb02, ne10, ne11, nb10, nb11, nb12);
+}
+
+static void ggml_cpy_f32_q4_0_cuda(
+    const char * cx, char * cdst, const int ne,
+    const int ne00, const int ne01, const int nb00, const int nb01, const int nb02,
+    const int ne10, const int ne11, const int nb10, const int nb11, const int nb12, cudaStream_t stream) {
+
+    GGML_ASSERT(ne % QK4_0 == 0);
+    const int num_blocks = ne / QK4_0;
+    cpy_f32_q<cpy_blck_f32_q4_0, QK4_0><<<num_blocks, 1, 0, stream>>>
+        (cx, cdst, ne, ne00, ne01, nb00, nb01, nb02, ne10, ne11, nb10, nb11, nb12);
+}
+
+static void ggml_cpy_f32_q4_1_cuda(
+    const char * cx, char * cdst, const int ne,
+    const int ne00, const int ne01, const int nb00, const int nb01, const int nb02,
+    const int ne10, const int ne11, const int nb10, const int nb11, const int nb12, cudaStream_t stream) {
+
+    GGML_ASSERT(ne % QK4_1 == 0);
+    const int num_blocks = ne / QK4_1;
+    cpy_f32_q<cpy_blck_f32_q4_1, QK4_1><<<num_blocks, 1, 0, stream>>>
+        (cx, cdst, ne, ne00, ne01, nb00, nb01, nb02, ne10, ne11, nb10, nb11, nb12);
+}
+
 static void ggml_cpy_f16_f16_cuda(
     const char * cx, char * cdst, const int ne,
     const int ne00, const int ne01, const int nb00, const int nb01, const int nb02,
@@ -5823,6 +6480,27 @@ static void alibi_f32_cuda(const float * x, float * dst, const int ncols, const
     alibi_f32<<<block_nums, block_dims, 0, stream>>>(x, dst, ncols, k_rows, n_heads_log2_floor, m0, m1);
 }
 
+static void sum_rows_f32_cuda(const float * x, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
+    const dim3 block_dims(WARP_SIZE, 1, 1);
+    const dim3 block_nums(1, nrows, 1);
+    k_sum_rows_f32<<<block_nums, block_dims, 0, stream>>>(x, dst, ncols);
+}
+
+static void argsort_f32_i32_cuda(const float * x, int * dst, const int ncols, const int nrows, ggml_sort_order order, cudaStream_t stream) {
+    // bitonic sort requires ncols to be power of 2
+    GGML_ASSERT((ncols & (ncols - 1)) == 0);
+
+    const dim3 block_dims(ncols, 1, 1);
+    const dim3 block_nums(1, nrows, 1);
+    if (order == GGML_SORT_ASC) {
+        k_argsort_f32_i32<GGML_SORT_ASC><<<block_nums, block_dims, 0, stream>>>(x, dst, ncols);
+    } else if (order == GGML_SORT_DESC) {
+        k_argsort_f32_i32<GGML_SORT_DESC><<<block_nums, block_dims, 0, stream>>>(x, dst, ncols);
+    } else {
+        GGML_ASSERT(false);
+    }
+}
+
 static void diag_mask_inf_f32_cuda(const float * x, float * dst, const int ncols_x, const int nrows_x, const int rows_per_channel, const int n_past, cudaStream_t stream) {
     const dim3 block_dims(1, CUDA_DIAG_MASK_INF_BLOCK_SIZE, 1);
     const int block_num_x = (ncols_x + CUDA_DIAG_MASK_INF_BLOCK_SIZE - 1) / CUDA_DIAG_MASK_INF_BLOCK_SIZE;
@@ -5838,13 +6516,14 @@ static void soft_max_f32_cuda(const float * x, const float * y, float * dst, con
     soft_max_f32<<<block_nums, block_dims, 0, stream>>>(x, y, dst, ncols_x, nrows_y, scale);
 }
 
-static void im2col_f32_f16_cuda(const float * x, half * dst,
-    int OH, int IW, int IH, int OW, int IC,
-    int KH, int KW, int N,  int ofs0, int ofs1,
-    int s0, int s1, int p0, int p1, int d0, int d1, cudaStream_t stream) {
-    dim3 block_nums(IC, OH, OW);
-    dim3 block_dims(N,  KH, KW);
-    im2col_f32_f16<<<block_nums, block_dims, 0, stream>>>(x, dst, ofs0, ofs1, IW, IH, (IC * KH * KW), s0, s1, p0, p1, d0, d1);
+static void im2col_f32_f16_cuda(const float* x, half* dst,
+    int IW, int IH, int OW, int OH, int KW, int KH, int IC,
+    int offset_delta,
+    int s0,int s1,int p0,int p1,int d0,int d1, cudaStream_t stream) {
+    const int parallel_elements = OW * KW * KH;
+    const int num_blocks = (parallel_elements + CUDA_IM2COL_BLOCK_SIZE - 1) / CUDA_IM2COL_BLOCK_SIZE;
+    dim3 block_nums(num_blocks, OH, IC);
+    im2col_f32_f16<<<block_nums, CUDA_IM2COL_BLOCK_SIZE, 0, stream>>>(x, dst, offset_delta, IW, IH, OW, KW, KH, parallel_elements, (IC * KH * KW), s0, s1, p0, p1, d0, d1);
 }
 
 // buffer pool for cuda
@@ -5915,7 +6594,7 @@ static void * ggml_cuda_pool_malloc(size_t size, size_t * actual_size) {
         return ptr;
     }
 #ifdef DEBUG_CUDA_MALLOC
-    fprintf(stderr, "%s: %d buffers, max_size = %u MiB, tot_size = %u MiB, requested %u MiB\n", __func__, nnz,
+    fprintf(stderr, "%s: %d buffers, max_size = %u MB, tot_size = %u MB, requested %u MB\n", __func__, nnz,
             (uint32_t)(max_size/1024/1024), (uint32_t)(tot_size/1024/1024), (uint32_t)(size/1024/1024));
 #endif
     void * ptr;
@@ -6053,7 +6732,7 @@ void * ggml_cuda_host_malloc(size_t size) {
         // The allocation error can be bypassed. A null ptr will assigned out of this function.
         // This can fixed the OOM error in WSL.
         cudaGetLastError();
-        fprintf(stderr, "WARNING: failed to allocate %.2f MiB of pinned memory: %s\n",
+        fprintf(stderr, "WARNING: failed to allocate %.2f MB of pinned memory: %s\n",
             size/1024.0/1024.0, cudaGetErrorString(err));
         return nullptr;
     }
@@ -6098,75 +6777,18 @@ static cudaError_t ggml_cuda_cpy_tensor_2d(
     const char * x = src_ptr + i1_low*nb1 + i2*nb2 + i3*nb3;
     if (nb0 == ts && nb1 == ts*ne0/bs) {
         return cudaMemcpyAsync(dst_ptr, x, i1_diff*nb1, kind, stream);
-    }
-    if (nb0 == ts) {
+    } else if (nb0 == ts) {
         return cudaMemcpy2DAsync(dst_ptr, ts*ne0/bs, x, nb1, ts*ne0/bs, i1_diff, kind, stream);
-    }
-    for (int64_t i1 = 0; i1 < i1_diff; i1++) {
-        const void * rx = (const void *) ((const char *) x + i1*nb1);
-        void * rd = (void *) (dst_ptr + i1*ts*ne0/bs);
-        // pretend the row is a matrix with cols=1
-        cudaError_t r = cudaMemcpy2DAsync(rd, ts/bs, rx, nb0, ts/bs, ne0, kind, stream);
-        if (r != cudaSuccess) { return r; }
-    }
-    return cudaSuccess;
-}
-
-static void ggml_cuda_op_repeat(
-    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
-    const float * src0_d, const float * src1_d, float * dst_d, const cudaStream_t & stream) {
-    // guaranteed to be an integer due to the check in ggml_can_repeat
-    const int64_t ne0 = dst->ne[0];
-    const int64_t ne1 = dst->ne[1];
-    const int64_t ne2 = dst->ne[2];
-    const int64_t ne3 = dst->ne[3];
-
-    const int64_t ne00 = src0->ne[0];
-    const int64_t ne01 = src0->ne[1];
-    const int64_t ne02 = src0->ne[2];
-    const int64_t ne03 = src0->ne[3];
-
-    const size_t nb0 = dst->nb[0];
-    const size_t nb1 = dst->nb[1];
-    const size_t nb2 = dst->nb[2];
-    const size_t nb3 = dst->nb[3];
-
-    const size_t nb00 = src0->nb[0];
-    const size_t nb01 = src0->nb[1];
-    const size_t nb02 = src0->nb[2];
-    const size_t nb03 = src0->nb[3];
-
-    const int nr0 = (int)(ne0/ne00);
-    const int nr1 = (int)(ne1/ne01);
-    const int nr2 = (int)(ne2/ne02);
-    const int nr3 = (int)(ne3/ne03);
-
-    // TODO: support for transposed / permuted tensors
-    GGML_ASSERT(nb0  == sizeof(float));
-    GGML_ASSERT(nb00 == sizeof(float));
-
-    // TODO: very inefficient, implement in a kernel, or fewer cudaMemcpyAsync calls for contiguous tensors
-    for                         (int i3 = 0; i3 < nr3;  i3++) {
-        for                     (int k3 = 0; k3 < ne03; k3++) {
-            for                 (int i2 = 0; i2 < nr2;  i2++) {
-                for             (int k2 = 0; k2 < ne02; k2++) {
-                    for         (int i1 = 0; i1 < nr1;  i1++) {
-                        for     (int k1 = 0; k1 < ne01; k1++) {
-                            for (int i0 = 0; i0 < nr0;  i0++) {
-                                CUDA_CHECK(cudaMemcpyAsync(
-                                              (char *)  dst_d + (i3*ne03 + k3)*nb3  + (i2*ne02 + k2)*nb2  + (i1*ne01 + k1)*nb1  + (i0*ne00)*nb0,
-                                        (const char *) src0_d + (          k3)*nb03 + (          k2)*nb02 + (          k1)*nb01,
-                                        ne00*nb0, cudaMemcpyDeviceToDevice, stream));
-                            }
-                        }
-                    }
-                }
-            }
+    } else {
+        for (int64_t i1 = 0; i1 < i1_diff; i1++) {
+            const void * rx = (const void *) ((const char *) x + i1*nb1);
+            void * rd = (void *) (dst_ptr + i1*ts*ne0/bs);
+            // pretend the row is a matrix with cols=1
+            cudaError_t r = cudaMemcpy2DAsync(rd, ts/bs, rx, nb0, ts/bs, ne0, kind, stream);
+            if (r != cudaSuccess) return r;
         }
+        return cudaSuccess;
     }
-
-    (void) src1;
-    (void) src1_d;
 }
 
 static void ggml_cuda_op_get_rows(
@@ -6175,65 +6797,96 @@ static void ggml_cuda_op_get_rows(
 
     GGML_ASSERT(src1->type == GGML_TYPE_I32);
     GGML_ASSERT(dst->type == GGML_TYPE_F32);
-    GGML_ASSERT(ggml_is_contiguous(src0));
-    GGML_ASSERT(ggml_is_contiguous(src1));
-    GGML_ASSERT(ggml_is_contiguous(dst));
 
-    const int ncols = src0->ne[0];
-    const int nrows = ggml_nelements(src1);
+    GGML_ASSERT(src0->nb[0] == ggml_type_size(src0->type));
+    GGML_ASSERT(src1->nb[0] == ggml_type_size(src1->type));
+    GGML_ASSERT(dst->nb[0] == ggml_type_size(dst->type));
 
     const int32_t * src1_i32 = (const int32_t *) src1_d;
 
     switch (src0->type) {
         case GGML_TYPE_F16:
-            get_rows_cuda<1, 1, convert_f16>(src0_d, src1_i32, dst_d, nrows, ncols, stream);
+            get_rows_cuda_float(src0, src1, dst, (const half *)src0_d, src1_i32, dst_d, stream);
             break;
         case GGML_TYPE_F32:
-            get_rows_cuda<1, 1, convert_f32>(src0_d, src1_i32, dst_d, nrows, ncols, stream);
+            get_rows_cuda_float(src0, src1, dst, src0_d, src1_i32, dst_d, stream);
             break;
         case GGML_TYPE_Q4_0:
-            get_rows_cuda<QK4_0, QR4_0, dequantize_q4_0>(src0_d, src1_i32, dst_d, nrows, ncols, stream);
+            get_rows_cuda<QK4_0, QR4_0, dequantize_q4_0>(src0, src1, dst, src0_d, src1_i32, dst_d, stream);
             break;
         case GGML_TYPE_Q4_1:
-            get_rows_cuda<QK4_1, QR4_1, dequantize_q4_1>(src0_d, src1_i32, dst_d, nrows, ncols, stream);
+            get_rows_cuda<QK4_1, QR4_1, dequantize_q4_1>(src0, src1, dst, src0_d, src1_i32, dst_d, stream);
             break;
         case GGML_TYPE_Q5_0:
-            get_rows_cuda<QK5_0, QR5_0, dequantize_q5_0>(src0_d, src1_i32, dst_d, nrows, ncols, stream);
+            get_rows_cuda<QK5_0, QR5_0, dequantize_q5_0>(src0, src1, dst, src0_d, src1_i32, dst_d, stream);
             break;
         case GGML_TYPE_Q5_1:
-            get_rows_cuda<QK5_1, QR5_1, dequantize_q5_1>(src0_d, src1_i32, dst_d, nrows, ncols, stream);
+            get_rows_cuda<QK5_1, QR5_1, dequantize_q5_1>(src0, src1, dst, src0_d, src1_i32, dst_d, stream);
             break;
         case GGML_TYPE_Q8_0:
-            get_rows_cuda<QK8_0, QR8_0, dequantize_q8_0>(src0_d, src1_i32, dst_d, nrows, ncols, stream);
+            get_rows_cuda<QK8_0, QR8_0, dequantize_q8_0>(src0, src1, dst, src0_d, src1_i32, dst_d, stream);
             break;
         default:
             // TODO: k-quants
+            fprintf(stderr, "%s: unsupported type: %s\n", __func__, ggml_type_name(src0->type));
             GGML_ASSERT(false);
             break;
     }
 }
 
+template<class op>
+inline void ggml_cuda_op_bin_bcast(
+    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
+    const float * src0_dd, const float * src1_dd, float * dst_dd, const cudaStream_t & main_stream) {
+
+    GGML_ASSERT(src1->type == GGML_TYPE_F32);
+
+    if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
+        op()(src0, src1, dst, src0_dd, src1_dd, dst_dd, main_stream);
+    } else if (src0->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F16) {
+        op()(src0, src1, dst, (const half *) src0_dd, src1_dd, (half *) dst_dd, main_stream);
+    } else if (src0->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F32) {
+        op()(src0, src1, dst, (const half *) src0_dd, src1_dd, dst_dd, main_stream);
+    } else {
+        fprintf(stderr, "%s: unsupported types: dst: %s, src0: %s, src1: %s\n", __func__,
+            ggml_type_name(dst->type), ggml_type_name(src0->type), ggml_type_name(src1->type));
+        GGML_ASSERT(false);
+    }
+}
+
+static void ggml_cuda_op_repeat(
+    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
+    const float * src0_d, const float * src1_d, float * dst_d, const cudaStream_t & main_stream) {
+
+    ggml_cuda_op_bin_bcast<bin_bcast_cuda<op_repeat>>(dst, src0, dst, nullptr, src0_d, dst_d, main_stream);
+
+    (void) src1;
+    (void) src1_d;
+}
+
 inline void ggml_cuda_op_add(
     const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
     const float * src0_dd, const float * src1_dd, float * dst_dd, const cudaStream_t & main_stream) {
 
+    ggml_cuda_op_bin_bcast<bin_bcast_cuda<op_add>>(src0, src1, dst, src0_dd, src1_dd, dst_dd, main_stream);
+}
+
+inline void ggml_cuda_op_acc(
+    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
+    const float * src0_dd, const float * src1_dd, float * dst_dd, const cudaStream_t & main_stream) {
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
     GGML_ASSERT(src1->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->ne[3] == 1); // just 3D tensors supported
 
-    const int64_t ne10 = src1->ne[0];
-    const int64_t ne11 = src1->ne[1];
+    int nb1 = dst->op_params[0] / 4; // 4 bytes of float32
+    int nb2 = dst->op_params[1] / 4; // 4 bytes of float32
+    // int nb3 = dst->op_params[2] / 4; // 4 bytes of float32 - unused
+    int offset = dst->op_params[3] / 4; // offset in bytes
 
-    if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
-        add_f32_cuda(src0_dd, src1_dd, dst_dd, ggml_nelements(src0), ne10*ne11, main_stream);
-    } else if (src0->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F16) {
-        add_f16_f32_f16_cuda((const half *) src0_dd, src1_dd, (half *) dst_dd, ggml_nelements(src0), main_stream);
-    } else if (src0->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F32) {
-        add_f16_f32_f32_cuda((const half *) src0_dd, src1_dd, dst_dd, ggml_nelements(src0), main_stream);
-    } else {
-        fprintf(stderr, "src0->type: %d  dst->type: %d\n", src0->type, dst->type);
-        GGML_ASSERT(false);
-    }
+    acc_f32_cuda(src0_dd, src1_dd, dst_dd, ggml_nelements(dst), src1->ne[0], src1->ne[1], src1->ne[2], nb1, nb2, offset, main_stream);
 
-    (void) src1;
     (void) dst;
 }
 
@@ -6241,16 +6894,14 @@ inline void ggml_cuda_op_mul(
     const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
     const float * src0_dd, const float * src1_dd, float * dst_dd, const cudaStream_t & main_stream) {
 
-    GGML_ASSERT(src0->type == GGML_TYPE_F32);
-    GGML_ASSERT(src1->type == GGML_TYPE_F32);
-    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+    ggml_cuda_op_bin_bcast<bin_bcast_cuda<op_mul>>(src0, src1, dst, src0_dd, src1_dd, dst_dd, main_stream);
+}
 
-    const int64_t ne10 = src1->ne[0];
-    const int64_t ne11 = src1->ne[1];
+inline void ggml_cuda_op_div(
+    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
+    const float * src0_dd, const float * src1_dd, float * dst_dd, const cudaStream_t & main_stream) {
 
-    mul_f32_cuda(src0_dd, src1_dd, dst_dd, ggml_nelements(src0), ne10*ne11, main_stream);
-
-    (void) dst;
+    ggml_cuda_op_bin_bcast<bin_bcast_cuda<op_div>>(src0, src1, dst, src0_dd, src1_dd, dst_dd, main_stream);
 }
 
 inline void ggml_cuda_op_gelu(
@@ -6281,6 +6932,34 @@ inline void ggml_cuda_op_silu(
     (void) src1_dd;
 }
 
+inline void ggml_cuda_op_gelu_quick(
+    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
+    const float * src0_dd, const float * src1_dd, float * dst_dd, const cudaStream_t & main_stream) {
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+    gelu_quick_f32_cuda(src0_dd, dst_dd, ggml_nelements(src0), main_stream);
+
+    (void) src1;
+    (void) dst;
+    (void) src1_dd;
+}
+
+inline void ggml_cuda_op_tanh(
+    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
+    const float * src0_dd, const float * src1_dd, float * dst_dd, const cudaStream_t & main_stream) {
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+    tanh_f32_cuda(src0_dd, dst_dd, ggml_nelements(src0), main_stream);
+
+    (void) src1;
+    (void) dst;
+    (void) src1_dd;
+}
+
 inline void ggml_cuda_op_relu(
     const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
     const float * src0_dd, const float * src1_dd, float * dst_dd, const cudaStream_t & main_stream) {
@@ -6295,6 +6974,23 @@ inline void ggml_cuda_op_relu(
     (void) src1_dd;
 }
 
+inline void ggml_cuda_op_leaky_relu(
+    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
+    const float * src0_dd, const float * src1_dd, float * dst_dd, const cudaStream_t & main_stream) {
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+    float negative_slope;
+    memcpy(&negative_slope, dst->op_params, sizeof(float));
+
+    leaky_relu_f32_cuda(src0_dd, dst_dd, ggml_nelements(src0), negative_slope, main_stream);
+
+    (void) src1;
+    (void) dst;
+    (void) src1_dd;
+}
+
 inline void ggml_cuda_op_sqr(
     const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
     const float * src0_dd, const float * src1_dd, float * dst_dd, const cudaStream_t & main_stream) {
@@ -6319,7 +7015,77 @@ inline void ggml_cuda_op_norm(
     const int64_t ne00 = src0->ne[0];
     const int64_t nrows = ggml_nrows(src0);
 
-    norm_f32_cuda(src0_dd, dst_dd, ne00, nrows, main_stream);
+    float eps;
+    memcpy(&eps, dst->op_params, sizeof(float));
+
+    norm_f32_cuda(src0_dd, dst_dd, ne00, nrows, eps, main_stream);
+
+    (void) src1;
+    (void) dst;
+    (void) src1_dd;
+}
+
+
+inline void ggml_cuda_op_group_norm(
+    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
+    const float * src0_dd, const float * src1_dd, float * dst_dd, const cudaStream_t & main_stream) {
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+    int num_groups = dst->op_params[0];
+    int group_size = src0->ne[0] * src0->ne[1] * ((src0->ne[2] + num_groups - 1) / num_groups);
+    group_norm_f32_cuda(src0_dd, dst_dd, num_groups, group_size, src0->ne[0] * src0->ne[1] * src0->ne[2], main_stream);
+
+    (void) src1;
+    (void) dst;
+    (void) src1_dd;
+}
+
+inline void ggml_cuda_op_concat(
+    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
+    const float * src0_dd, const float * src1_dd, float * dst_dd, const cudaStream_t & main_stream) {
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT(src1->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type == GGML_TYPE_F32);
+
+    for (int i3 = 0; i3 < dst->ne[3]; i3++) {
+        concat_f32_cuda(src0_dd + i3 * (src0->nb[3] / 4), src1_dd + i3 * (src1->nb[3] / 4), dst_dd + i3 * (dst->nb[3] / 4), dst->ne[0], dst->ne[1], dst->ne[2], src0->ne[2], main_stream);
+    }
+
+    (void) src1;
+    (void) dst;
+}
+
+inline void ggml_cuda_op_upscale(
+    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
+    const float * src0_dd, const float * src1_dd, float * dst_dd, const cudaStream_t & main_stream) {
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type == GGML_TYPE_F32);
+    GGML_ASSERT(src0->ne[3] == 1 && dst->ne[3] == 1); // just 3D tensors
+
+    const int scale_factor = dst->op_params[0];
+
+    upscale_f32_cuda(src0_dd, dst_dd, src0->ne[0], src0->ne[1], src0->ne[2], scale_factor, main_stream);
+
+    (void) src1;
+    (void) dst;
+    (void) src1_dd;
+}
+
+inline void ggml_cuda_op_pad(
+    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
+    const float * src0_dd, const float * src1_dd, float * dst_dd, const cudaStream_t & main_stream) {
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type == GGML_TYPE_F32);
+    GGML_ASSERT(src0->ne[3] == 1 && dst->ne[3] == 1); // just 3D tensors
+
+    pad_f32_cuda(src0_dd, dst_dd,
+        src0->ne[0], src0->ne[1], src0->ne[2],
+        dst->ne[0], dst->ne[1], dst->ne[2], main_stream);
 
     (void) src1;
     (void) dst;
@@ -6474,6 +7240,8 @@ inline void ggml_cuda_op_mul_mat_vec_q(
     const char * src1_ddq_i, float * dst_dd_i, const int64_t row_low, const int64_t row_high, const int64_t src1_ncols,
     const int64_t src1_padded_row_size, const cudaStream_t & stream) {
 
+    GGML_ASSERT(ggml_nrows(src1) == 1);
+
     const int64_t ne00 = src0->ne[0];
     const int64_t row_diff = row_high - row_low;
 
@@ -6533,7 +7301,8 @@ inline void ggml_cuda_op_dequantize_mul_mat_vec(
     size_t ash;
     dfloat * src1_dfloat = nullptr; // dfloat == half
 
-    bool src1_convert_f16 = src0->type == GGML_TYPE_Q4_0 || src0->type == GGML_TYPE_Q4_1 ||
+    bool src1_convert_f16 =
+        src0->type == GGML_TYPE_Q4_0 || src0->type == GGML_TYPE_Q4_1 ||
         src0->type == GGML_TYPE_Q5_0 || src0->type == GGML_TYPE_Q5_1 ||
         src0->type == GGML_TYPE_Q8_0 || src0->type == GGML_TYPE_F16;
 
@@ -6624,7 +7393,7 @@ inline void ggml_cuda_op_mul_mat_cublas(
 
     const int compute_capability = g_compute_capabilities[id];
 
-    if (compute_capability >= CC_VOLTA && (src0->type == GGML_TYPE_F16 || ggml_is_quantized(src0->type)) && ggml_is_contiguous(src0) && row_diff == src0->ne[1]) {
+    if (compute_capability >= CC_VOLTA && (src0->type == GGML_TYPE_F16 || ggml_is_quantized(src0->type)) && ggml_is_contiguous(src0) && row_diff == src0->ne[1] && dst->op_params[0] == GGML_PREC_DEFAULT) {
         // convert src0 and src1 to fp16, multiply as fp16, convert dst to fp32
         half * src0_as_f16 = nullptr;
         size_t src0_as = 0;
@@ -6837,7 +7606,6 @@ inline void ggml_cuda_op_im2col(
 
     const bool is_2D = ((const int32_t*)(dst->op_params))[6] == 1;
 
-    const int64_t N  = src1->ne[is_2D ? 3 : 2];
     const int64_t IC = src1->ne[is_2D ? 2 : 1];
     const int64_t IH = is_2D ? src1->ne[1] : 1;
     const int64_t IW =         src1->ne[0];
@@ -6848,17 +7616,51 @@ inline void ggml_cuda_op_im2col(
     const int64_t OH = is_2D ? dst->ne[2] : 1;
     const int64_t OW =         dst->ne[1];
 
-    const size_t ofs0 = src1->nb[is_2D ? 3 : 2] / 4; // nb is byte offset, src is type float32
-    const size_t ofs1 = src1->nb[is_2D ? 2 : 1] / 4; // nb is byte offset, src is type float32
+    const size_t delta_offset = src1->nb[is_2D ? 2 : 1] / 4; // nb is byte offset, src is type float32
 
-    im2col_f32_f16_cuda(src1_dd, (half*) dst_dd,
-        OH, IW, IH, OW, IC, KH, KW, N,
-        ofs0, ofs1, s0, s1, p0, p1, d0, d1, main_stream);
+    im2col_f32_f16_cuda(src1_dd, (half*) dst_dd, IW, IH, OW, OH, KW, KH, IC, delta_offset, s0, s1, p0, p1, d0, d1, main_stream);
 
     (void) src0;
     (void) src0_dd;
 }
 
+
+inline void ggml_cuda_op_sum_rows(
+    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
+    const float * src0_dd, const float * src1_dd, float * dst_dd, const cudaStream_t & main_stream) {
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+    const int64_t ncols = src0->ne[0];
+    const int64_t nrows = ggml_nrows(src0);
+
+    sum_rows_f32_cuda(src0_dd, dst_dd, ncols, nrows, main_stream);
+
+    (void) src1;
+    (void) dst;
+    (void) src1_dd;
+}
+
+inline void ggml_cuda_op_argsort(
+    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
+    const float * src0_dd, const float * src1_dd, float * dst_dd, const cudaStream_t & main_stream) {
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_I32);
+
+    const int64_t ncols = src0->ne[0];
+    const int64_t nrows = ggml_nrows(src0);
+
+    enum ggml_sort_order order = (enum ggml_sort_order) dst->op_params[0];
+
+    argsort_f32_i32_cuda(src0_dd, (int *)dst_dd, ncols, nrows, order, main_stream);
+
+    (void) src1;
+    (void) dst;
+    (void) src1_dd;
+}
+
 inline void ggml_cuda_op_diag_mask_inf(
     const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
     const float * src0_dd, const float * src1_dd, float * dst_dd, const cudaStream_t & main_stream) {
@@ -6905,17 +7707,10 @@ inline void ggml_cuda_op_scale(
     const float * src0_dd, const float * src1_dd, float * dst_dd, const cudaStream_t & main_stream) {
 
     GGML_ASSERT(src0->type == GGML_TYPE_F32);
-    GGML_ASSERT(src1->type == GGML_TYPE_F32);
     GGML_ASSERT( dst->type == GGML_TYPE_F32);
 
     float scale;
-    // HACK: support for ggml backend interface
-    if (src1->backend == GGML_BACKEND_CPU) {
-        scale = ((float *) src1->data)[0];
-    } else {
-        // TODO: pass pointer to kernel instead of copying to host
-        CUDA_CHECK(cudaMemcpy(&scale, src1->data, sizeof(float), cudaMemcpyDeviceToHost));
-    }
+    memcpy(&scale, dst->op_params, sizeof(float));
 
     scale_f32_cuda(src0_dd, dst_dd, scale, ggml_nelements(src0), main_stream);
     CUDA_CHECK(cudaGetLastError());
@@ -6962,8 +7757,6 @@ static void ggml_cuda_op_flatten(const ggml_tensor * src0, const ggml_tensor * s
     const bool src1_on_device = use_src1 && src1->backend == GGML_BACKEND_GPU;
     const bool  dst_on_device =              dst->backend == GGML_BACKEND_GPU;
 
-    const bool src1_stays_on_host = use_src1 && dst->op == GGML_OP_SCALE;
-
     // dd = data device
     float * src0_ddf = nullptr;
     float * src1_ddf = nullptr;
@@ -6984,7 +7777,7 @@ static void ggml_cuda_op_flatten(const ggml_tensor * src0, const ggml_tensor * s
         CUDA_CHECK(ggml_cuda_cpy_tensor_2d(src0_ddf, src0, 0, 0, 0, nrows0, main_stream));
     }
 
-    if (use_src1 && !src1_stays_on_host) {
+    if (use_src1) {
         if (src1_on_device) {
             src1_ddf = (float *) src1_extra->data_device[g_main_device];
         } else {
@@ -7032,6 +7825,11 @@ static void ggml_cuda_set_peer_access(const int n_tokens) {
     }
 
 #ifdef NDEBUG
+    for (int id = 0; id < g_device_count; ++id) {
+        CUDA_CHECK(ggml_cuda_set_device(id));
+        CUDA_CHECK(cudaDeviceSynchronize());
+    }
+
     for (int id = 0; id < g_device_count; ++id) {
         CUDA_CHECK(ggml_cuda_set_device(id));
 
@@ -7067,7 +7865,7 @@ static void ggml_cuda_op_mul_mat(
     const int64_t ne01 = src0->ne[1];
     const int64_t ne02 = src0->ne[2];
     const int64_t ne03 = src0->ne[3];
-    // const int64_t nrows0 = ggml_nrows(src0);
+    const int64_t nrows0 = ggml_nrows(src0);
 
     const int64_t ne10 = src1->ne[0];
     const int64_t ne11 = src1->ne[1];
@@ -7083,8 +7881,6 @@ static void ggml_cuda_op_mul_mat(
     const int nb2 = dst->nb[2];
     const int nb3 = dst->nb[3];
 
-    ggml_cuda_set_peer_access(ne11);
-
     GGML_ASSERT(dst->backend != GGML_BACKEND_GPU_SPLIT);
     GGML_ASSERT(src1->backend != GGML_BACKEND_GPU_SPLIT);
 
@@ -7103,10 +7899,9 @@ static void ggml_cuda_op_mul_mat(
 
     const bool src0_on_device = src0->backend == GGML_BACKEND_GPU || src0->backend == GGML_BACKEND_GPU_SPLIT;
     const bool src0_is_contiguous = ggml_is_contiguous(src0);
-
     const bool src1_is_contiguous = ggml_is_contiguous(src1);
-    const int64_t src1_padded_col_size = ne10 % MATRIX_ROW_PADDING == 0 ?
-        ne10 : ne10 - ne10 % MATRIX_ROW_PADDING + MATRIX_ROW_PADDING;
+
+    const int64_t src1_padded_col_size = GGML_PAD(ne10, MATRIX_ROW_PADDING);
 
     const bool split = src0->backend == GGML_BACKEND_GPU_SPLIT;
     GGML_ASSERT(!(split && ne02 > 1));
@@ -7142,12 +7937,16 @@ static void ggml_cuda_op_mul_mat(
 
             if (id != 0) {
                 row_low[id]  = ne01*g_tensor_split[id];
-                row_low[id] -= row_low[id] % rounding;
+                if (row_low[id] < ne01) {
+                    row_low[id] -= row_low[id] % rounding;
+                }
             }
 
             if (id != g_device_count - 1) {
                 row_high[id]  = ne01*g_tensor_split[id + 1];
-                row_high[id] -= row_high[id] % rounding;
+                if (row_high[id] < ne01) {
+                    row_high[id] -= row_high[id] % rounding;
+                }
             }
         }
     }
@@ -7231,7 +8030,7 @@ static void ggml_cuda_op_mul_mat(
                 const size_t src1_ddq_i_offset = (i0*ne11 + src1_col_0) * src1_padded_col_size*q8_1_ts/q8_1_bs;
 
                 // for split tensors the data begins at i0 == i0_offset_low
-                char  *  src0_dd_i =  src0_dd[id] + (i0/i02_divisor) * ne01*ne00*src0_ts/src0_bs;
+                char  *  src0_dd_i =  src0_dd[id] + (i0/i02_divisor) * (ne01*ne00*src0_ts)/src0_bs;
                 float * src1_ddf_i = src1_ddf[id] + (i0*ne11 + src1_col_0) * ne10;
                 char  * src1_ddq_i = src1_ddq[id] +  src1_ddq_i_offset;
                 float *   dst_dd_i =   dst_dd[id] + (i0*ne1  + src1_col_0) * (dst_on_device ? ne0 : row_diff);
@@ -7372,10 +8171,18 @@ static void ggml_cuda_add(const ggml_tensor * src0, const ggml_tensor * src1, gg
     ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_add);
 }
 
+static void ggml_cuda_acc(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_acc);
+}
+
 static void ggml_cuda_mul(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_mul);
 }
 
+static void ggml_cuda_div(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_div);
+}
+
 static void ggml_cuda_gelu(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_gelu);
 }
@@ -7384,10 +8191,22 @@ static void ggml_cuda_silu(const ggml_tensor * src0, const ggml_tensor * src1, g
     ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_silu);
 }
 
+static void ggml_cuda_gelu_quick(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_gelu_quick);
+}
+
+static void ggml_cuda_tanh(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_tanh);
+}
+
 static void ggml_cuda_relu(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_relu);
 }
 
+static void ggml_cuda_leaky_relu(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_leaky_relu);
+}
+
 static void ggml_cuda_sqr(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_sqr);
 }
@@ -7396,12 +8215,28 @@ static void ggml_cuda_norm(const ggml_tensor * src0, const ggml_tensor * src1, g
     ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_norm);
 }
 
+static void ggml_cuda_group_norm(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_group_norm);
+}
+
+static void ggml_cuda_concat(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_concat);
+}
+
+static void ggml_cuda_upscale(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_upscale);
+}
+
+static void ggml_cuda_pad(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_pad);
+}
+
 static void ggml_cuda_rms_norm(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_rms_norm);
 }
 
 bool ggml_cuda_can_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst) {
-    if (!g_cublas_loaded) { return false; }
+    if (!g_cublas_loaded) return false;
 
     const int64_t ne10 = src1->ne[0];
 
@@ -7479,28 +8314,28 @@ static void ggml_cuda_mul_mat_vec_nc(const ggml_tensor * src0, const ggml_tensor
     ggml_mul_mat_vec_nc_f16_f32_cuda(src0_ddq, src1_ddf, dst_ddf, ne00, ne01, row_stride_x, ne02, ne12, channel_stride_x, main_stream);
 }
 
-__global__ static void k_compute_batched_ptrs(
-        const half * src0_as_f16, const half * src1_as_f16, half * dst_f16,
+static __global__ void k_compute_batched_ptrs(
+        const half * src0_as_f16, const half * src1_as_f16, char * dst,
         const void ** ptrs_src, void ** ptrs_dst,
-        int ne12, int ne13,
-        int ne23,
-        int nb02, int nb03,
-        int nb12, int nb13,
-        int nb2, int nb3,
-        int r2, int r3) {
-    int i13 = blockIdx.x * blockDim.x + threadIdx.x;
-    int i12 = blockIdx.y * blockDim.y + threadIdx.y;
+        int64_t ne12, int64_t ne13,
+        int64_t ne23,
+        size_t  nb02, size_t  nb03,
+        size_t  nb12, size_t  nb13,
+        size_t  nbd2, size_t  nbd3,
+        int64_t r2,   int64_t r3) {
+    int64_t i13 = blockIdx.x * blockDim.x + threadIdx.x;
+    int64_t i12 = blockIdx.y * blockDim.y + threadIdx.y;
 
     if (i13 >= ne13 || i12 >= ne12) {
         return;
     }
 
-    int i03 = i13 / r3;
-    int i02 = i12 / r2;
+    int64_t i03 = i13 / r3;
+    int64_t i02 = i12 / r2;
 
     ptrs_src[0*ne23 + i12 + i13*ne12] = (const char *) src0_as_f16 + i02*nb02   + i03*nb03;
     ptrs_src[1*ne23 + i12 + i13*ne12] = (const char *) src1_as_f16 + i12*nb12/2 + i13*nb13/2;
-    ptrs_dst[0*ne23 + i12 + i13*ne12] = (      char *)     dst_f16 + i12* nb2/2 + i13* nb3/2;
+    ptrs_dst[0*ne23 + i12 + i13*ne12] = (      char *)         dst + i12*nbd2   + i13*nbd3;
 }
 
 static void ggml_cuda_mul_mat_mat_batched_cublas(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
@@ -7535,9 +8370,7 @@ static void ggml_cuda_mul_mat_mat_batched_cublas(const ggml_tensor * src0, const
     CUDA_CHECK(ggml_cuda_set_device(g_main_device));
     cudaStream_t main_stream = g_cudaStreams[g_main_device][0];
 
-    int id;
-    CUDA_CHECK(cudaGetDevice(&id));
-    CUBLAS_CHECK(cublasSetStream(g_cublas_handles[id], main_stream));
+    CUBLAS_CHECK(cublasSetStream(g_cublas_handles[g_main_device], main_stream));
 
     ggml_tensor_extra_gpu * src0_extra = (ggml_tensor_extra_gpu *) src0->extra;
     void * src0_ddq = src0_extra->data_device[g_main_device];
@@ -7558,7 +8391,41 @@ static void ggml_cuda_mul_mat_mat_batched_cublas(const ggml_tensor * src0, const
     to_fp16_cuda(src1_ddf, src1_as_f16, ne1, main_stream);
 
     size_t dst_as = 0;
-    half * dst_f16 = (half *) ggml_cuda_pool_malloc(ne * sizeof(half), &dst_as);
+
+    half * dst_f16 = nullptr;
+    char * dst_t   = nullptr;
+
+    cublasComputeType_t cu_compute_type = CUBLAS_COMPUTE_16F;
+    cudaDataType_t      cu_data_type    = CUDA_R_16F;
+
+    // dst strides
+    size_t nbd2 = dst->nb[2];
+    size_t nbd3 = dst->nb[3];
+
+    const half  alpha_f16 = 1.0f;
+    const half  beta_f16  = 0.0f;
+
+    const float alpha_f32 = 1.0f;
+    const float beta_f32  = 0.0f;
+
+    const void * alpha = &alpha_f16;
+    const void * beta  = &beta_f16;
+
+    if (dst->op_params[0] == GGML_PREC_DEFAULT) {
+        dst_f16 = (half *) ggml_cuda_pool_malloc(ne * sizeof(half), &dst_as);
+        dst_t   = (char *) dst_f16;
+
+        nbd2 /= sizeof(float) / sizeof(half);
+        nbd3 /= sizeof(float) / sizeof(half);
+    } else {
+        dst_t = (char *) dst_ddf;
+
+        cu_compute_type = CUBLAS_COMPUTE_32F;
+        cu_data_type    = CUDA_R_32F;
+
+        alpha = &alpha_f32;
+        beta  = &beta_f32;
+    }
 
     GGML_ASSERT(ne12 % ne02 == 0);
     GGML_ASSERT(ne13 % ne03 == 0);
@@ -7567,9 +8434,6 @@ static void ggml_cuda_mul_mat_mat_batched_cublas(const ggml_tensor * src0, const
     const int64_t r2 = ne12/ne02;
     const int64_t r3 = ne13/ne03;
 
-    const half alpha_f16 = 1.0f;
-    const half beta_f16  = 0.0f;
-
 #if 0
     // use cublasGemmEx
     {
@@ -7579,12 +8443,12 @@ static void ggml_cuda_mul_mat_mat_batched_cublas(const ggml_tensor * src0, const
                 int i02 = i12 / r2;
 
                 CUBLAS_CHECK(
-                        cublasGemmEx(g_cublas_handles[id], CUBLAS_OP_T, CUBLAS_OP_N,
+                        cublasGemmEx(g_cublas_handles[g_main_device], CUBLAS_OP_T, CUBLAS_OP_N,
                             ne01, ne11, ne10,
-                            &alpha_f16, (const char *) src0_as_f16 + i02*src0->nb[2]   + i03*src0->nb[3]  , CUDA_R_16F, nb01/sizeof(half),
-                                        (const char *) src1_as_f16 + i12*src1->nb[2]/2 + i13*src1->nb[3]/2, CUDA_R_16F, nb11/sizeof(float),
-                            &beta_f16,  (      char *)     dst_f16 + i12* dst->nb[2]/2 + i13* dst->nb[3]/2, CUDA_R_16F, ne01,
-                            CUBLAS_COMPUTE_16F,
+                            alpha, (const char *) src0_as_f16 + i02*src0->nb[2]   + i03*src0->nb[3]  , CUDA_R_16F,   nb01/sizeof(half),
+                                   (const char *) src1_as_f16 + i12*src1->nb[2]/2 + i13*src1->nb[3]/2, CUDA_R_16F,   nb11/sizeof(float),
+                            beta,  (      char *)       dst_t + i12*nbd2          + i13*nbd3,          cu_data_type, ne01,
+                            cu_compute_type,
                             CUBLAS_GEMM_DEFAULT_TENSOR_OP));
             }
         }
@@ -7594,13 +8458,13 @@ static void ggml_cuda_mul_mat_mat_batched_cublas(const ggml_tensor * src0, const
         // there is no broadcast and src0, src1 are contiguous across dims 2, 3
         // use cublasGemmStridedBatchedEx
         CUBLAS_CHECK(
-        cublasGemmStridedBatchedEx(g_cublas_handles[id], CUBLAS_OP_T, CUBLAS_OP_N,
+        cublasGemmStridedBatchedEx(g_cublas_handles[g_main_device], CUBLAS_OP_T, CUBLAS_OP_N,
                 ne01, ne11, ne10,
-                &alpha_f16, (const char *) src0_as_f16, CUDA_R_16F, nb01/sizeof(half),  src0->nb[2]/sizeof(half),  // strideA
-                            (const char *) src1_as_f16, CUDA_R_16F, nb11/sizeof(float), src1->nb[2]/sizeof(float), // strideB
-                &beta_f16,  (      char *)     dst_f16, CUDA_R_16F, ne01,                dst->nb[2]/sizeof(float), // strideC
+                alpha, (const char *) src0_as_f16, CUDA_R_16F,   nb01/sizeof(half),  src0->nb[2]/sizeof(half),  // strideA
+                       (const char *) src1_as_f16, CUDA_R_16F,   nb11/sizeof(float), src1->nb[2]/sizeof(float), // strideB
+                beta,  (      char *)       dst_t, cu_data_type, ne01,                dst->nb[2]/sizeof(float), // strideC
                 ne12*ne13,
-                CUBLAS_COMPUTE_16F,
+                cu_compute_type,
                 CUBLAS_GEMM_DEFAULT_TENSOR_OP));
     } else {
         // use cublasGemmBatchedEx
@@ -7617,24 +8481,24 @@ static void ggml_cuda_mul_mat_mat_batched_cublas(const ggml_tensor * src0, const
 
         dim3 block_dims(ne13, ne12);
         k_compute_batched_ptrs<<<1, block_dims, 0, main_stream>>>(
-                src0_as_f16, src1_as_f16, dst_f16,
+                src0_as_f16, src1_as_f16, dst_t,
                 ptrs_src, ptrs_dst,
                 ne12, ne13,
                 ne23,
                 nb02, nb03,
                 nb12, nb13,
-                dst->nb[2], dst->nb[3],
+                nbd2, nbd3,
                 r2, r3);
         CUDA_CHECK(cudaGetLastError());
 
         CUBLAS_CHECK(
-        cublasGemmBatchedEx(g_cublas_handles[id], CUBLAS_OP_T, CUBLAS_OP_N,
+        cublasGemmBatchedEx(g_cublas_handles[g_main_device], CUBLAS_OP_T, CUBLAS_OP_N,
                 ne01, ne11, ne10,
-                &alpha_f16, (const void **) (ptrs_src + 0*ne23), CUDA_R_16F, nb01/sizeof(half),
-                            (const void **) (ptrs_src + 1*ne23), CUDA_R_16F, nb11/sizeof(float),
-                &beta_f16,  (      void **) (ptrs_dst + 0*ne23), CUDA_R_16F, ne01,
+                alpha, (const void **) (ptrs_src + 0*ne23), CUDA_R_16F,   nb01/sizeof(half),
+                       (const void **) (ptrs_src + 1*ne23), CUDA_R_16F,   nb11/sizeof(float),
+                beta,  (      void **) (ptrs_dst + 0*ne23), cu_data_type, ne01,
                 ne23,
-                CUBLAS_COMPUTE_16F,
+                cu_compute_type,
                 CUBLAS_GEMM_DEFAULT_TENSOR_OP));
 
         if (ptrs_src_s != 0) {
@@ -7646,11 +8510,14 @@ static void ggml_cuda_mul_mat_mat_batched_cublas(const ggml_tensor * src0, const
     }
 #endif
 
-    const to_fp32_cuda_t to_fp32_cuda = ggml_get_to_fp32_cuda(GGML_TYPE_F16);
-    to_fp32_cuda(dst_f16, dst_ddf, ne, main_stream);
+    if (dst->op_params[0] == GGML_PREC_DEFAULT) {
+        const to_fp32_cuda_t to_fp32_cuda = ggml_get_to_fp32_cuda(GGML_TYPE_F16);
+        to_fp32_cuda(dst_f16, dst_ddf, ne, main_stream);
+
+        ggml_cuda_pool_free(dst_f16, dst_as);
+    }
 
     ggml_cuda_pool_free(src1_as_f16, src1_as);
-    ggml_cuda_pool_free(dst_f16, dst_as);
 }
 
 static void ggml_cuda_mul_mat(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
@@ -7698,10 +8565,11 @@ static void ggml_cuda_mul_mat(const ggml_tensor * src0, const ggml_tensor * src1
 #ifdef GGML_CUDA_FORCE_DMMV
             const bool use_mul_mat_vec_q = false;
 #else
-            const bool use_mul_mat_vec_q = min_compute_capability >= MIN_CC_DP4A && ggml_is_quantized(src0->type);
+            const bool use_mul_mat_vec_q = min_compute_capability >= MIN_CC_DP4A && ggml_is_quantized(src0->type) && ggml_nrows(src1) == 1;
 #endif // GGML_CUDA_FORCE_DMMV
 
             if (use_mul_mat_vec_q) {
+                // NOTE: this kernel does not support ggml_nrows(src1) > 1
                 ggml_cuda_op_mul_mat(src0, src1, dst, ggml_cuda_op_mul_mat_vec_q, true);
             } else {
                 ggml_cuda_op_mul_mat(src0, src1, dst, ggml_cuda_op_dequantize_mul_mat_vec, false);
@@ -7726,6 +8594,328 @@ static void ggml_cuda_mul_mat(const ggml_tensor * src0, const ggml_tensor * src1
     }
 }
 
+#if 0
+template<typename ... Srcs>
+static __global__ void k_compute_batched_ptrs_id(
+        const void ** ptrs_src, void ** ptrs_dst,
+        int ne12, int ne13,
+        int ne23,
+        int nb02, int nb03,
+        int nb12, int nb13,
+        int nb2, int nb3,
+        int r2, int r3,
+        ggml_type src0_type, half * src0_as_f16, int64_t src0_ne,
+        const half * src1_f16, half * dst_f16,
+        const int32_t * ids, const int id,
+        Srcs... src0s) {
+
+    int i = ids[id];
+
+    half * src0_f16;
+    const void * srcs_ar[] = { (const half *) src0s... };
+    if (src0_type == GGML_TYPE_F16) {
+        src0_f16 = (half *) srcs_ar[i];
+    } else {
+        src0_f16 = src0_as_f16;
+        if (threadIdx.x == 0 && threadIdx.y == 0) {
+            const to_fp16_cuda_t to_fp16 = ggml_get_to_fp16_cuda(src0_type);
+            to_fp16(srcs_ar[i], src0_f16, src0_ne, cudaStreamFireAndForget);
+        }
+    }
+
+    int i13 = blockIdx.x * blockDim.x + threadIdx.x;
+    int i12 = blockIdx.y * blockDim.y + threadIdx.y;
+
+    if (i13 >= ne13 || i12 >= ne12) {
+        return;
+    }
+
+    int i03 = i13 / r3;
+    int i02 = i12 / r2;
+
+    ptrs_src[0*ne23 + i12 + i13*ne12] = (const char *) src0_f16 + i02*nb02   + i03*nb03;
+    ptrs_src[1*ne23 + i12 + i13*ne12] = (const char *) src1_f16 + i12*nb12/2 + i13*nb13/2;
+    ptrs_dst[0*ne23 + i12 + i13*ne12] = (      char *)  dst_f16 + i12* nb2/2 + i13* nb3/2;
+}
+
+static void ggml_cuda_mul_mat_id_cublas(ggml_tensor * dst) {
+    const struct ggml_tensor * ids = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
+    const struct ggml_tensor * src00 = dst->src[2];
+
+    const int id = dst->op_params[0];
+
+    GGML_ASSERT(!ggml_is_transposed(src00));
+    GGML_ASSERT(!ggml_is_transposed(src1));
+
+    GGML_ASSERT(src00->backend != GGML_BACKEND_GPU_SPLIT);
+    GGML_ASSERT(src1->type == GGML_TYPE_F32);
+
+    const int64_t ne00 = src00->ne[0]; GGML_UNUSED(ne00);
+    const int64_t ne01 = src00->ne[1];
+    const int64_t ne02 = src00->ne[2];
+    const int64_t ne03 = src00->ne[3];
+
+    //const int64_t nb01 = src00->nb[1];
+    const int64_t nb02 = src00->nb[2]; GGML_UNUSED(nb02);
+    const int64_t nb03 = src00->nb[3]; GGML_UNUSED(nb03);
+
+    const int64_t ne10 = src1->ne[0];
+    const int64_t ne11 = src1->ne[1];
+    const int64_t ne12 = src1->ne[2];
+    const int64_t ne13 = src1->ne[3];
+
+    //const int64_t nb11 = src1->nb[1];
+    const int64_t nb12 = src1->nb[2]; GGML_UNUSED(nb12);
+    const int64_t nb13 = src1->nb[3]; GGML_UNUSED(nb13);
+
+    const int64_t ne1 = ggml_nelements(src1);
+    const int64_t ne  = ggml_nelements(dst);
+
+    CUDA_CHECK(ggml_cuda_set_device(g_main_device));
+    cudaStream_t main_stream = g_cudaStreams[g_main_device][0];
+
+    CUBLAS_CHECK(cublasSetStream(g_cublas_handles[g_main_device], main_stream));
+
+    //ggml_tensor_extra_gpu * src0_extra = (ggml_tensor_extra_gpu *) src0->extra;
+    //void * src0_ddq = src0_extra->data_device[g_main_device];
+    //half * src0_as_f16 = (half *) src0_ddq;
+
+    ggml_tensor_extra_gpu * src1_extra = (ggml_tensor_extra_gpu *) src1->extra;
+    float * src1_ddf = (float *) src1_extra->data_device[g_main_device];
+
+    ggml_tensor_extra_gpu * dst_extra = (ggml_tensor_extra_gpu *) dst->extra;
+    float * dst_ddf = (float *) dst_extra->data_device[g_main_device];
+
+    // convert src1 to fp16
+    const to_fp16_cuda_t to_fp16_cuda = ggml_get_to_fp16_cuda(src1->type);
+    GGML_ASSERT(to_fp16_cuda != nullptr);
+
+    size_t src1_as = 0;
+    half * src1_as_f16 = (half *) ggml_cuda_pool_malloc(ne1 * sizeof(half), &src1_as);
+    to_fp16_cuda(src1_ddf, src1_as_f16, ne1, main_stream);
+
+    size_t dst_as = 0;
+    half * dst_f16 = (half *) ggml_cuda_pool_malloc(ne * sizeof(half), &dst_as);
+
+    GGML_ASSERT(ne12 % ne02 == 0);
+    GGML_ASSERT(ne13 % ne03 == 0);
+
+    // broadcast factors
+    const int64_t r2 = ne12/ne02;
+    const int64_t r3 = ne13/ne03;
+
+    const half alpha_f16 = 1.0f;
+    const half beta_f16  = 0.0f;
+
+    // use cublasGemmBatchedEx
+    const int ne23 = ne12*ne13;
+
+    const void ** ptrs_src = nullptr;
+          void ** ptrs_dst = nullptr;
+
+    size_t ptrs_src_s = 0;
+    size_t ptrs_dst_s = 0;
+
+    ptrs_src = (const void **) ggml_cuda_pool_malloc(2*ne23*sizeof(void *), &ptrs_src_s);
+    ptrs_dst = (      void **) ggml_cuda_pool_malloc(1*ne23*sizeof(void *), &ptrs_dst_s);
+
+    int64_t src0_ne = ggml_nelements(src00);
+    half * src0_as_f16 = nullptr;
+    size_t src0_as = 0;
+    if (src00->type != GGML_TYPE_F16) {
+        src0_as_f16 = (half *) ggml_cuda_pool_malloc(src0_ne * sizeof(half), &src0_as);
+    }
+
+    static_assert(GGML_MAX_SRC == 6, "GGML_MAX_SRC == 6");
+    dim3 block_dims(ne13, ne12);
+    k_compute_batched_ptrs_id<<<1, block_dims, 0, main_stream>>>(
+            ptrs_src, ptrs_dst,
+            ne12, ne13,
+            ne23,
+            ne00*ne01*sizeof(half), ne00*ne01*ne02*sizeof(half),
+            nb12, nb13,
+            dst->nb[2], dst->nb[3],
+            r2, r3,
+            src00->type, src0_as_f16, src0_ne,
+            src1_as_f16, dst_f16,
+            (const int *)((ggml_tensor_extra_gpu *)ids->extra)->data_device[g_main_device], id,
+            dst->src[2] ? (const half *)((ggml_tensor_extra_gpu *)dst->src[2]->extra)->data_device[g_main_device] : nullptr,
+            dst->src[3] ? (const half *)((ggml_tensor_extra_gpu *)dst->src[3]->extra)->data_device[g_main_device] : nullptr,
+            dst->src[4] ? (const half *)((ggml_tensor_extra_gpu *)dst->src[4]->extra)->data_device[g_main_device] : nullptr,
+            dst->src[5] ? (const half *)((ggml_tensor_extra_gpu *)dst->src[5]->extra)->data_device[g_main_device] : nullptr
+    );
+    CUDA_CHECK(cudaGetLastError());
+
+    CUBLAS_CHECK(
+    cublasGemmBatchedEx(g_cublas_handles[g_main_device], CUBLAS_OP_T, CUBLAS_OP_N,
+            ne01, ne11, ne10,
+            &alpha_f16, (const void **) (ptrs_src + 0*ne23), CUDA_R_16F, ne00,
+                        (const void **) (ptrs_src + 1*ne23), CUDA_R_16F, ne10,
+            &beta_f16,  (      void **) (ptrs_dst + 0*ne23), CUDA_R_16F, ne01,
+            ne23,
+            CUBLAS_COMPUTE_16F,
+            CUBLAS_GEMM_DEFAULT_TENSOR_OP));
+
+    if (src0_as != 0) {
+        ggml_cuda_pool_free(src0_as_f16, src0_as);
+    }
+    if (ptrs_src_s != 0) {
+        ggml_cuda_pool_free(ptrs_src, ptrs_src_s);
+    }
+    if (ptrs_dst_s != 0) {
+        ggml_cuda_pool_free(ptrs_dst, ptrs_dst_s);
+    }
+
+    const to_fp32_cuda_t to_fp32_cuda = ggml_get_to_fp32_cuda(GGML_TYPE_F16);
+    to_fp32_cuda(dst_f16, dst_ddf, ne, main_stream);
+
+    ggml_cuda_pool_free(src1_as_f16, src1_as);
+    ggml_cuda_pool_free(dst_f16, dst_as);
+}
+#endif
+
+static void ggml_cuda_mul_mat_id(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+#if 0
+    ggml_cuda_mul_mat_id_cublas(dst);
+    // TODO: mmq/mmv support
+#endif
+
+    const int64_t nb11 = src1->nb[1];
+    const int64_t nb1  =  dst->nb[1];
+
+    const struct ggml_tensor * ids = src0;
+    const int32_t id = ((int32_t *) dst->op_params)[0];
+    const int32_t n_as = ((int32_t *) dst->op_params)[1];
+
+    std::vector<char> ids_host(ggml_nbytes(ids));
+
+    const cudaStream_t stream = g_cudaStreams[g_main_device][0];
+
+    if (ids->backend == GGML_BACKEND_GPU) {
+        const char * ids_dev = (const char *)((const ggml_tensor_extra_gpu *)ids->extra)->data_device[g_main_device];
+        CUDA_CHECK(cudaMemcpyAsync(ids_host.data(), ids_dev, ggml_nbytes(ids), cudaMemcpyDeviceToHost, stream));
+        CUDA_CHECK(cudaStreamSynchronize(stream));
+    } else {
+        memcpy(ids_host.data(), ids->data, ggml_nbytes(ids));
+    }
+
+    const ggml_tensor_extra_gpu * src1_extra = (const ggml_tensor_extra_gpu *) src1->extra;
+    const ggml_tensor_extra_gpu * dst_extra = (const ggml_tensor_extra_gpu *) dst->extra;
+
+    ggml_tensor_extra_gpu src1_row_extra;
+    ggml_tensor_extra_gpu dst_row_extra;
+
+    ggml_tensor src1_row = *src1;
+    ggml_tensor dst_row = *dst;
+
+    src1_row.backend = GGML_BACKEND_GPU;
+    dst_row.backend  = GGML_BACKEND_GPU;
+
+    src1_row.extra = &src1_row_extra;
+    dst_row.extra = &dst_row_extra;
+
+    char * src1_original = src1->backend == GGML_BACKEND_CPU ?
+        (char *) src1->data : (char *) src1_extra->data_device[g_main_device];
+    char * dst_original  =  dst->backend == GGML_BACKEND_CPU ?
+        (char *)  dst->data : (char *)  dst_extra->data_device[g_main_device];
+
+    if (src1->ne[1] == 1) {
+        GGML_ASSERT(src1->backend == GGML_BACKEND_GPU);
+        GGML_ASSERT(dst->backend  == GGML_BACKEND_GPU);
+
+        for (int64_t i01 = 0; i01 < ids->ne[1]; i01++) {
+            //int32_t row_id;
+            //CUDA_CHECK(cudaMemcpyAsync(&row_id, ids_dev + i01*ids->nb[1] + id*ids->nb[0], sizeof(int32_t), cudaMemcpyDeviceToHost, g_cudaStreams[g_main_device][0]));
+            //CUDA_CHECK(cudaStreamSynchronize(g_cudaStreams[g_main_device][0]));
+
+            const int32_t row_id = *(const int32_t *) (ids_host.data() + i01*ids->nb[1] + id*ids->nb[0]);
+
+            GGML_ASSERT(row_id >= 0 && row_id < n_as);
+
+            const struct ggml_tensor * src0_row = dst->src[row_id + 2];
+
+            src1_row_extra.data_device[g_main_device] = src1_original + i01*src1->nb[1];
+            src1_row.data = (char *) src1->data + i01*src1->nb[1]; // TODO why is this set?
+
+            dst_row_extra.data_device[g_main_device] = dst_original + i01*dst->nb[1];
+            dst_row.data = (char *) dst->data + i01*dst->nb[1]; // TODO why is this set?
+
+            ggml_cuda_mul_mat(src0_row, &src1_row, &dst_row);
+        }
+    } else {
+        size_t as_src1, as_dst;
+        char * src1_contiguous = (char *) ggml_cuda_pool_malloc(sizeof(float)*ggml_nelements(src1), &as_src1);
+        char *  dst_contiguous = (char *) ggml_cuda_pool_malloc(sizeof(float)*ggml_nelements(dst),  &as_dst);
+
+        src1_row_extra.data_device[g_main_device] = src1_contiguous;
+        dst_row_extra.data_device[g_main_device]  =  dst_contiguous;
+
+        const cudaMemcpyKind src1_kind = src1->backend == GGML_BACKEND_CPU ?
+            cudaMemcpyHostToDevice : cudaMemcpyDeviceToDevice;
+        const cudaMemcpyKind dst_kind  =  dst->backend == GGML_BACKEND_CPU ?
+            cudaMemcpyDeviceToHost : cudaMemcpyDeviceToDevice;
+
+        for (int32_t row_id = 0; row_id < n_as; ++row_id) {
+            const struct ggml_tensor * src0_row = dst->src[row_id + 2];
+
+            int64_t num_src1_rows = 0;
+            for (int64_t i01 = 0; i01 < ids->ne[1]; i01++) {
+                const int32_t row_id_i = *(const int32_t *) (ids_host.data() + i01*ids->nb[1] + id*ids->nb[0]);
+
+                if (row_id_i != row_id) {
+                    continue;
+                }
+
+                GGML_ASSERT(row_id >= 0 && row_id < n_as);
+
+                CUDA_CHECK(cudaMemcpyAsync(src1_contiguous + num_src1_rows*nb11, src1_original + i01*nb11,
+                                        nb11, src1_kind, stream));
+                num_src1_rows++;
+            }
+
+            if (num_src1_rows == 0) {
+                continue;
+            }
+
+            src1_row.ne[1] = num_src1_rows;
+            dst_row.ne[1] = num_src1_rows;
+
+            src1_row.nb[1] = nb11;
+            src1_row.nb[2] = num_src1_rows*nb11;
+            src1_row.nb[3] = num_src1_rows*nb11;
+
+            dst_row.nb[1] = nb1;
+            dst_row.nb[2] = num_src1_rows*nb1;
+            dst_row.nb[3] = num_src1_rows*nb1;
+
+            ggml_cuda_mul_mat(src0_row, &src1_row, &dst_row);
+
+            num_src1_rows = 0;
+            for (int64_t i01 = 0; i01 < ids->ne[1]; i01++) {
+                const int32_t row_id_i = *(const int32_t *) (ids_host.data() + i01*ids->nb[1] + id*ids->nb[0]);
+
+                if (row_id_i != row_id) {
+                    continue;
+                }
+
+                GGML_ASSERT(row_id >= 0 && row_id < n_as);
+
+                CUDA_CHECK(cudaMemcpyAsync(dst_original + i01*nb1, dst_contiguous + num_src1_rows*nb1,
+                                        nb1, dst_kind, stream));
+                num_src1_rows++;
+            }
+        }
+
+        ggml_cuda_pool_free(src1_contiguous, as_src1);
+        ggml_cuda_pool_free(dst_contiguous,  as_dst);
+    }
+
+    if (dst->backend == GGML_BACKEND_CPU) {
+        CUDA_CHECK(cudaStreamSynchronize(stream));
+    }
+}
+
 static void ggml_cuda_scale(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_scale);
 }
@@ -7770,14 +8960,17 @@ static void ggml_cuda_cpy(const ggml_tensor * src0, const ggml_tensor * src1, gg
     char * src1_ddc = (char *) src1_extra->data_device[g_main_device];
 
     if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32) {
-        ggml_cpy_f32_f32_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, nb00, nb01, nb02,
-                              ne10, ne11, nb10, nb11, nb12, main_stream);
+        ggml_cpy_f32_f32_cuda (src0_ddc, src1_ddc, ne, ne00, ne01, nb00, nb01, nb02, ne10, ne11, nb10, nb11, nb12, main_stream);
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F16) {
-        ggml_cpy_f32_f16_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, nb00, nb01, nb02,
-                              ne10, ne11, nb10, nb11, nb12, main_stream);
+        ggml_cpy_f32_f16_cuda (src0_ddc, src1_ddc, ne, ne00, ne01, nb00, nb01, nb02, ne10, ne11, nb10, nb11, nb12, main_stream);
+    } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q8_0) {
+        ggml_cpy_f32_q8_0_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, nb00, nb01, nb02, ne10, ne11, nb10, nb11, nb12, main_stream);
+    } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q4_0) {
+        ggml_cpy_f32_q4_0_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, nb00, nb01, nb02, ne10, ne11, nb10, nb11, nb12, main_stream);
+    } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q4_1) {
+        ggml_cpy_f32_q4_1_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, nb00, nb01, nb02, ne10, ne11, nb10, nb11, nb12, main_stream);
     } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F16) {
-        ggml_cpy_f16_f16_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, nb00, nb01, nb02,
-                              ne10, ne11, nb10, nb11, nb12, main_stream);
+        ggml_cpy_f16_f16_cuda (src0_ddc, src1_ddc, ne, ne00, ne01, nb00, nb01, nb02, ne10, ne11, nb10, nb11, nb12, main_stream);
     } else {
         fprintf(stderr, "%s: unsupported type combination (%s to %s)\n", __func__,
                 ggml_type_name(src0->type), ggml_type_name(src1->type));
@@ -7788,6 +8981,7 @@ static void ggml_cuda_cpy(const ggml_tensor * src0, const ggml_tensor * src1, gg
 }
 
 static void ggml_cuda_dup(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    // TODO: why do we pass dst as src1 here?
     ggml_cuda_cpy(src0, dst, nullptr);
     (void) src1;
 }
@@ -7813,12 +9007,28 @@ static void ggml_cuda_im2col(const ggml_tensor * src0, const ggml_tensor * src1,
     ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_im2col);
 }
 
+static void ggml_cuda_sum_rows(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_ASSERT(ggml_is_contiguous(src0));
+    ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_sum_rows);
+}
+
+static void ggml_cuda_argsort(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_ASSERT(ggml_is_contiguous(src0));
+    ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_argsort);
+}
+
 static void ggml_cuda_nop(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     (void) src0;
     (void) src1;
     (void) dst;
 }
 
+static size_t ggml_nbytes_split(const struct ggml_tensor * tensor, int nrows_split) {
+    static_assert(GGML_MAX_DIMS == 4, "GGML_MAX_DIMS is not 4 - update this function");
+
+    return nrows_split*ggml_row_size(tensor->type, tensor->ne[0]);
+}
+
 void ggml_cuda_transform_tensor(void * data, struct ggml_tensor * tensor) {
     const int64_t nrows = ggml_nrows(tensor);
 
@@ -7868,13 +9078,12 @@ void ggml_cuda_transform_tensor(void * data, struct ggml_tensor * tensor) {
 
         // pad last row to a multiple of 512 elements to avoid out-of-bounds memory accesses
         if (ne0 % MATRIX_ROW_PADDING != 0) {
-            size += (MATRIX_ROW_PADDING - ne0 % MATRIX_ROW_PADDING)
-                * ggml_type_size(tensor->type)/ggml_blck_size(tensor->type);
+            size += ggml_row_size(tensor->type, MATRIX_ROW_PADDING - ne0 % MATRIX_ROW_PADDING);
         }
 
         char * buf;
         CUDA_CHECK(cudaMalloc(&buf, size));
-        char * buf_host = (char*)data + offset_split;
+        char * buf_host = (char *)data + offset_split;
 
         // set padding to 0 to avoid possible NaN values
         if (size > original_size) {
@@ -7896,7 +9105,7 @@ void ggml_cuda_transform_tensor(void * data, struct ggml_tensor * tensor) {
 }
 
 void ggml_cuda_free_data(struct ggml_tensor * tensor) {
-    if (!tensor || (tensor->backend != GGML_BACKEND_GPU && tensor->backend != GGML_BACKEND_GPU_SPLIT) ) {
+    if (!tensor || !tensor->extra || (tensor->backend != GGML_BACKEND_GPU && tensor->backend != GGML_BACKEND_GPU_SPLIT) ) {
         return;
     }
 
@@ -8019,11 +9228,10 @@ void ggml_cuda_assign_scratch_offset(struct ggml_tensor * tensor, size_t offset)
 
     ggml_tensor_extra_gpu * extra = ggml_cuda_alloc_temp_tensor_extra();
 
-    const bool inplace = (tensor->src[0] != nullptr && tensor->src[0]->data == tensor->data) ||
-        tensor->op == GGML_OP_VIEW;
+    const bool inplace = tensor->view_src != nullptr;
 
-    if (inplace && (tensor->src[0]->backend == GGML_BACKEND_GPU || tensor->src[0]->backend == GGML_BACKEND_GPU_SPLIT)) {
-        ggml_tensor_extra_gpu * src0_extra = (ggml_tensor_extra_gpu * ) tensor->src[0]->extra;
+    if (inplace && (tensor->view_src->backend == GGML_BACKEND_GPU || tensor->view_src->backend == GGML_BACKEND_GPU_SPLIT)) {
+        ggml_tensor_extra_gpu * src0_extra = (ggml_tensor_extra_gpu * ) tensor->view_src->extra;
         char * src0_ddc = (char *) src0_extra->data_device[g_main_device];
         size_t view_offset = 0;
         if (tensor->op == GGML_OP_VIEW) {
@@ -8068,8 +9276,9 @@ void ggml_cuda_set_main_device(const int main_device) {
                 main_device, g_device_count, g_main_device);
         return;
     }
-    g_main_device = main_device;
-    if (g_device_count > 1) {
+
+    if (g_main_device != main_device && g_device_count > 1) {
+        g_main_device = main_device;
         cudaDeviceProp prop;
         CUDA_CHECK(cudaGetDeviceProperties(&prop, g_main_device));
         fprintf(stderr, "%s: using device %d (%s) as main device\n", __func__, g_main_device, prop.name);
@@ -8095,21 +9304,21 @@ void ggml_cuda_free_scratch() {
 }
 
 bool ggml_cuda_compute_forward(struct ggml_compute_params * params, struct ggml_tensor * tensor) {
-    if (!g_cublas_loaded) { return false; }
+    if (!g_cublas_loaded) return false;
 
     ggml_cuda_func_t func;
     const bool any_on_device = tensor->backend == GGML_BACKEND_GPU
         || (tensor->src[0] != nullptr && (tensor->src[0]->backend == GGML_BACKEND_GPU || tensor->src[0]->backend == GGML_BACKEND_GPU_SPLIT))
         || (tensor->src[1] != nullptr && tensor->src[1]->backend == GGML_BACKEND_GPU);
 
-    if (!any_on_device && tensor->op != GGML_OP_MUL_MAT) {
+    if (!any_on_device && tensor->op != GGML_OP_MUL_MAT && tensor->op != GGML_OP_MUL_MAT_ID) {
         return false;
     }
 
     if (tensor->op == GGML_OP_MUL_MAT) {
         if (tensor->src[0]->ne[3] != tensor->src[1]->ne[3]) {
 #ifndef NDEBUG
-            fprintf(stderr, "%s: cannot compute %s: src0->ne[3] = " PRId64 ", src1->ne[3] = " PRId64 " - fallback to CPU\n", __func__, tensor->name, tensor->src[0]->ne[3], tensor->src[1]->ne[3]);
+            fprintf(stderr, "%s: cannot compute %s: src0->ne[3] = %" PRId64 ", src1->ne[3] = %" PRId64 " - fallback to CPU\n", __func__, tensor->name, tensor->src[0]->ne[3], tensor->src[1]->ne[3]);
 #endif
             return false;
         }
@@ -8128,9 +9337,15 @@ bool ggml_cuda_compute_forward(struct ggml_compute_params * params, struct ggml_
         case GGML_OP_ADD:
             func = ggml_cuda_add;
             break;
+        case GGML_OP_ACC:
+            func = ggml_cuda_acc;
+            break;
         case GGML_OP_MUL:
             func = ggml_cuda_mul;
             break;
+        case GGML_OP_DIV:
+            func = ggml_cuda_div;
+            break;
         case GGML_OP_UNARY:
             switch (ggml_get_unary_op(tensor)) {
                 case GGML_UNARY_OP_GELU:
@@ -8139,15 +9354,37 @@ bool ggml_cuda_compute_forward(struct ggml_compute_params * params, struct ggml_
                 case GGML_UNARY_OP_SILU:
                     func = ggml_cuda_silu;
                     break;
+                case GGML_UNARY_OP_GELU_QUICK:
+                    func = ggml_cuda_gelu_quick;
+                    break;
+                case GGML_UNARY_OP_TANH:
+                    func = ggml_cuda_tanh;
+                    break;
                 case GGML_UNARY_OP_RELU:
                     func = ggml_cuda_relu;
                     break;
                 default:
                     return false;
-            } break;
+            }
+            break;
         case GGML_OP_NORM:
             func = ggml_cuda_norm;
             break;
+        case GGML_OP_GROUP_NORM:
+            func = ggml_cuda_group_norm;
+            break;
+        case GGML_OP_CONCAT:
+            func = ggml_cuda_concat;
+            break;
+        case GGML_OP_UPSCALE:
+            func = ggml_cuda_upscale;
+            break;
+        case GGML_OP_PAD:
+            func = ggml_cuda_pad;
+            break;
+        case GGML_OP_LEAKY_RELU:
+            func = ggml_cuda_leaky_relu;
+            break;
         case GGML_OP_RMS_NORM:
             func = ggml_cuda_rms_norm;
             break;
@@ -8157,6 +9394,12 @@ bool ggml_cuda_compute_forward(struct ggml_compute_params * params, struct ggml_
             }
             func = ggml_cuda_mul_mat;
             break;
+        case GGML_OP_MUL_MAT_ID:
+            if (!any_on_device && !ggml_cuda_can_mul_mat(tensor->src[2], tensor->src[1], tensor)) {
+                return false;
+            }
+            func = ggml_cuda_mul_mat_id;
+            break;
         case GGML_OP_SCALE:
             func = ggml_cuda_scale;
             break;
@@ -8164,9 +9407,6 @@ bool ggml_cuda_compute_forward(struct ggml_compute_params * params, struct ggml_
             func = ggml_cuda_sqr;
             break;
         case GGML_OP_CLAMP:
-            if (!any_on_device) {
-                return false;
-            }
             func = ggml_cuda_clamp;
             break;
         case GGML_OP_CPY:
@@ -8175,6 +9415,7 @@ bool ggml_cuda_compute_forward(struct ggml_compute_params * params, struct ggml_
         case GGML_OP_CONT:
             func = ggml_cuda_dup;
             break;
+        case GGML_OP_NONE:
         case GGML_OP_RESHAPE:
         case GGML_OP_VIEW:
         case GGML_OP_PERMUTE:
@@ -8196,10 +9437,20 @@ bool ggml_cuda_compute_forward(struct ggml_compute_params * params, struct ggml_
         case GGML_OP_IM2COL:
             func = ggml_cuda_im2col;
             break;
+        case GGML_OP_SUM_ROWS:
+            func = ggml_cuda_sum_rows;
+            break;
+        case GGML_OP_ARGSORT:
+            func = ggml_cuda_argsort;
+            break;
         default:
             return false;
     }
 
+    if (tensor->src[0] != nullptr && tensor->src[0]->backend == GGML_BACKEND_GPU_SPLIT) {
+        ggml_cuda_set_peer_access(tensor->src[1]->ne[1]);
+    }
+
     if (params->ith != 0) {
         return true;
     }
@@ -8212,7 +9463,9 @@ bool ggml_cuda_compute_forward(struct ggml_compute_params * params, struct ggml_
 
 int ggml_cuda_get_device_count() {
     int device_count;
-    CUDA_CHECK(cudaGetDeviceCount(&device_count));
+    if (cudaGetDeviceCount(&device_count) != cudaSuccess) {
+        return 0;
+    }
     return device_count;
 }
 
@@ -8228,27 +9481,16 @@ void ggml_cuda_get_device_description(int device, char * description, size_t des
 
 #define UNUSED GGML_UNUSED
 
-struct ggml_backend_context_cuda {
-};
-
-static const char * ggml_backend_cuda_name(ggml_backend_t backend) {
-    return GGML_CUDA_NAME;
-
-    UNUSED(backend);
-}
-
-static void ggml_backend_cuda_free(ggml_backend_t backend) {
-    ggml_backend_context_cuda * cuda_ctx = (ggml_backend_context_cuda *)backend->context;
-    delete cuda_ctx;
-    delete backend;
-}
+// cuda buffer
 
 struct ggml_backend_buffer_context_cuda {
-    void * device;
-
+    int device;
+    void * dev_ptr = nullptr;
     ggml_tensor_extra_gpu * temp_tensor_extras = nullptr;
     size_t temp_tensor_extra_index = 0;
 
+    ggml_backend_buffer_context_cuda(int device, void * dev_ptr) : device(device), dev_ptr(dev_ptr) {}
+
     ~ggml_backend_buffer_context_cuda() {
         delete[] temp_tensor_extras;
     }
@@ -8269,41 +9511,20 @@ struct ggml_backend_buffer_context_cuda {
 
 static void ggml_backend_cuda_buffer_free_buffer(ggml_backend_buffer_t buffer) {
     ggml_backend_buffer_context_cuda * ctx = (ggml_backend_buffer_context_cuda *)buffer->context;
-    CUDA_CHECK(cudaFree(ctx->device));
+    CUDA_CHECK(cudaFree(ctx->dev_ptr));
     delete ctx;
 }
 
 static void * ggml_backend_cuda_buffer_get_base(ggml_backend_buffer_t buffer) {
     ggml_backend_buffer_context_cuda * ctx = (ggml_backend_buffer_context_cuda *)buffer->context;
-    return ctx->device;
-}
-
-static size_t ggml_backend_cuda_buffer_get_alloc_size(ggml_backend_buffer_t buffer, ggml_tensor * tensor) {
-    int64_t row_low = 0;
-    int64_t row_high = ggml_nrows(tensor);
-    int64_t nrows_split = row_high - row_low;
-
-    size_t size = ggml_nbytes_split(tensor, nrows_split);
-
-    int64_t ne0 = tensor->ne[0];
-
-    if (ggml_is_quantized(tensor->type)) {
-        if (ne0 % MATRIX_ROW_PADDING != 0) {
-            size += (MATRIX_ROW_PADDING - ne0 % MATRIX_ROW_PADDING)
-                * ggml_type_size(tensor->type)/ggml_blck_size(tensor->type);
-        }
-    }
-
-    return size;
-
-    UNUSED(buffer);
+    return ctx->dev_ptr;
 }
 
 static void ggml_backend_cuda_buffer_init_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor) {
     ggml_backend_buffer_context_cuda * ctx = (ggml_backend_buffer_context_cuda *)buffer->context;
 
     if (tensor->view_src != NULL && tensor->view_offs == 0) {
-        assert(tensor->view_src->buffer->backend == buffer->backend);
+        assert(tensor->view_src->buffer->buft == buffer->buft);
         tensor->backend = tensor->view_src->backend;
         tensor->extra = tensor->view_src->extra;
         return;
@@ -8311,7 +9532,7 @@ static void ggml_backend_cuda_buffer_init_tensor(ggml_backend_buffer_t buffer, g
 
     ggml_tensor_extra_gpu * extra = ctx->ggml_cuda_alloc_temp_tensor_extra();
 
-    extra->data_device[g_main_device] = tensor->data;
+    extra->data_device[ctx->device] = tensor->data;
 
     tensor->backend = GGML_BACKEND_GPU;
     tensor->extra = extra;
@@ -8323,64 +9544,213 @@ static void ggml_backend_cuda_buffer_init_tensor(ggml_backend_buffer_t buffer, g
         int64_t nrows_split = row_high - row_low;
 
         size_t original_size = ggml_nbytes_split(tensor, nrows_split);
-        size_t padded_size = ggml_backend_cuda_buffer_get_alloc_size(tensor->buffer, tensor);
+        size_t padded_size = ggml_backend_buft_get_alloc_size(buffer->buft, tensor);
 
         if (padded_size > original_size && tensor->view_src == nullptr) {
-            CUDA_CHECK(cudaMemsetAsync((char *)tensor->data + original_size, 0, padded_size - original_size, g_cudaStreams[g_main_device][0]));
+            CUDA_CHECK(cudaMemsetAsync((char *)tensor->data + original_size, 0, padded_size - original_size, g_cudaStreams[ctx->device][0]));
         }
     }
 
     UNUSED(buffer);
 }
 
+static void ggml_backend_cuda_buffer_set_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
+    GGML_ASSERT(tensor->backend == GGML_BACKEND_GPU);
+
+    ggml_backend_buffer_context_cuda * ctx = (ggml_backend_buffer_context_cuda *)buffer->context;
+
+    ggml_cuda_set_device(ctx->device);
+    CUDA_CHECK(cudaDeviceSynchronize());
+
+    CUDA_CHECK(cudaMemcpy((char *)tensor->data + offset, data, size, cudaMemcpyHostToDevice));
+}
+
+static void ggml_backend_cuda_buffer_get_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * tensor, void * data, size_t offset, size_t size) {
+    GGML_ASSERT(tensor->backend == GGML_BACKEND_GPU);
+
+    ggml_backend_buffer_context_cuda * ctx = (ggml_backend_buffer_context_cuda *)buffer->context;
+
+    ggml_cuda_set_device(ctx->device);
+    CUDA_CHECK(cudaDeviceSynchronize());
+
+    CUDA_CHECK(cudaMemcpy(data, (const char *)tensor->data + offset, size, cudaMemcpyDeviceToHost));
+}
+
+static void ggml_backend_cuda_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
+    ggml_backend_buffer_context_cuda * ctx = (ggml_backend_buffer_context_cuda *)buffer->context;
+
+    ggml_cuda_set_device(ctx->device);
+    CUDA_CHECK(cudaDeviceSynchronize());
+
+    CUDA_CHECK(cudaMemset(ctx->dev_ptr, value, buffer->size));
+}
+
 static struct ggml_backend_buffer_i cuda_backend_buffer_interface = {
-    /* .free_buffer    = */ ggml_backend_cuda_buffer_free_buffer,
-    /* .get_base       = */ ggml_backend_cuda_buffer_get_base,
-    /* .get_alloc_size = */ ggml_backend_cuda_buffer_get_alloc_size,
-    /* .init_tensor    = */ ggml_backend_cuda_buffer_init_tensor,
-    /* .free_tensor    = */ NULL,
+    /* .free_buffer     = */ ggml_backend_cuda_buffer_free_buffer,
+    /* .get_base        = */ ggml_backend_cuda_buffer_get_base,
+    /* .init_tensor     = */ ggml_backend_cuda_buffer_init_tensor,
+    /* .set_tensor      = */ ggml_backend_cuda_buffer_set_tensor,
+    /* .get_tensor      = */ ggml_backend_cuda_buffer_get_tensor,
+    /* .cpy_tensor_from = */ NULL,
+    /* .cpy_tensor_to   = */ NULL,
+    /* .clear           = */ ggml_backend_cuda_buffer_clear,
 };
 
-static ggml_backend_buffer_t ggml_backend_cuda_alloc_buffer(ggml_backend_t backend, size_t size) {
-    ggml_cuda_set_device(g_main_device);
+// cuda buffer type
 
-    ggml_backend_buffer_context_cuda * ctx = new ggml_backend_buffer_context_cuda;
+static ggml_backend_buffer_t ggml_backend_cuda_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
+    int device = (int) (intptr_t) buft->context;
+
+    ggml_cuda_set_device(device);
 
     size = std::max(size, (size_t)1); // cudaMalloc returns null for size 0
 
-    ggml_cuda_set_device(g_main_device);
-    CUDA_CHECK(cudaMalloc(&ctx->device, size));
+    void * dev_ptr;
+    CUDA_CHECK(cudaMalloc(&dev_ptr, size));
 
-    return ggml_backend_buffer_init(backend, cuda_backend_buffer_interface, ctx, size);
+    ggml_backend_buffer_context_cuda * ctx = new ggml_backend_buffer_context_cuda(device, dev_ptr);
+
+    return ggml_backend_buffer_init(buft, cuda_backend_buffer_interface, ctx, size);
 }
 
-static size_t ggml_backend_cuda_get_alignment(ggml_backend_t backend) {
+static size_t ggml_backend_cuda_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
     return 128;
+
+    UNUSED(buft);
+}
+
+static size_t ggml_backend_cuda_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft, ggml_tensor * tensor) {
+    int64_t row_low = 0;
+    int64_t row_high = ggml_nrows(tensor);
+    int64_t nrows_split = row_high - row_low;
+
+    size_t size = ggml_nbytes_split(tensor, nrows_split);
+
+    int64_t ne0 = tensor->ne[0];
+
+    if (ggml_is_quantized(tensor->type)) {
+        if (ne0 % MATRIX_ROW_PADDING != 0) {
+            size += ggml_row_size(tensor->type, MATRIX_ROW_PADDING - ne0 % MATRIX_ROW_PADDING);
+        }
+    }
+
+    return size;
+
+    UNUSED(buft);
+}
+
+static bool ggml_backend_cuda_buffer_type_supports_backend(ggml_backend_buffer_type_t buft, ggml_backend_t backend) {
+    return ggml_backend_is_cuda(backend);
+
+    UNUSED(buft);
+}
+
+static ggml_backend_buffer_type_i ggml_backend_cuda_buffer_type_interface = {
+    /* .alloc_buffer     = */ ggml_backend_cuda_buffer_type_alloc_buffer,
+    /* .get_alignment    = */ ggml_backend_cuda_buffer_type_get_alignment,
+    /* .get_alloc_size   = */ ggml_backend_cuda_buffer_type_get_alloc_size,
+    /* .supports_backend = */ ggml_backend_cuda_buffer_type_supports_backend,
+    /* .is_host          = */ nullptr,
+};
+
+ggml_backend_buffer_type_t ggml_backend_cuda_buffer_type(int device) {
+    static struct ggml_backend_buffer_type ggml_backend_cuda_buffer_types[GGML_CUDA_MAX_DEVICES];
+
+    static bool ggml_backend_cuda_buffer_type_initialized = false;
+
+    if (!ggml_backend_cuda_buffer_type_initialized) {
+        for (int i = 0; i < GGML_CUDA_MAX_DEVICES; i++) {
+            ggml_backend_cuda_buffer_types[i] = {
+                /* .iface    = */ ggml_backend_cuda_buffer_type_interface,
+                /* .context  = */ (ggml_backend_buffer_type_context_t) (intptr_t) i,
+            };
+        }
+        ggml_backend_cuda_buffer_type_initialized = true;
+    }
+
+    return &ggml_backend_cuda_buffer_types[device];
+}
+
+// host buffer type
+
+static void ggml_backend_cuda_host_buffer_free_buffer(ggml_backend_buffer_t buffer) {
+    CUDA_CHECK(cudaFreeHost(buffer->context));
+}
+
+static ggml_backend_buffer_t ggml_backend_cuda_host_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
+    void * ptr;
+    CUDA_CHECK(cudaMallocHost(&ptr, size));
+
+    // FIXME: this is a hack to avoid having to implement a new buffer type
+    ggml_backend_buffer_t buffer = ggml_backend_cpu_buffer_from_ptr(ptr, size);
+    buffer->buft = buft;
+    buffer->iface.free_buffer = ggml_backend_cuda_host_buffer_free_buffer;
+
+    return buffer;
+}
+
+ggml_backend_buffer_type_t ggml_backend_cuda_host_buffer_type() {
+    static struct ggml_backend_buffer_type ggml_backend_cuda_buffer_type_host = {
+        /* .iface    = */ {
+            /* .alloc_buffer     = */ ggml_backend_cuda_host_buffer_type_alloc_buffer,
+            /* .get_alignment    = */ ggml_backend_cpu_buffer_type()->iface.get_alignment,
+            /* .get_alloc_size   = */ ggml_backend_cpu_buffer_type()->iface.get_alloc_size,
+            /* .supports_backend = */ ggml_backend_cpu_buffer_type()->iface.supports_backend,
+            /* .is_host          = */ ggml_backend_cpu_buffer_type()->iface.is_host,
+        },
+        /* .context  = */ nullptr,
+    };
+
+    return &ggml_backend_cuda_buffer_type_host;
+}
+
+// backend
+
+struct ggml_backend_context_cuda {
+    int device;
+};
+
+static const char * ggml_backend_cuda_name(ggml_backend_t backend) {
+    return GGML_CUDA_NAME;
+
     UNUSED(backend);
 }
 
+static void ggml_backend_cuda_free(ggml_backend_t backend) {
+    ggml_backend_context_cuda * cuda_ctx = (ggml_backend_context_cuda *)backend->context;
+
+    delete cuda_ctx;
+    delete backend;
+}
+
+static ggml_backend_buffer_type_t ggml_backend_cuda_get_default_buffer_type(ggml_backend_t backend) {
+    ggml_backend_context_cuda * cuda_ctx = (ggml_backend_context_cuda *)backend->context;
+
+    return ggml_backend_cuda_buffer_type(cuda_ctx->device);
+}
+
 static void ggml_backend_cuda_set_tensor_async(ggml_backend_t backend, ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
-    GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor write out of bounds");
-    GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
+    ggml_backend_context_cuda * cuda_ctx = (ggml_backend_context_cuda *)backend->context;
+
+    GGML_ASSERT(tensor->buffer->buft == ggml_backend_cuda_buffer_type(cuda_ctx->device) && "unsupported buffer type");
     GGML_ASSERT(tensor->backend == GGML_BACKEND_GPU);
 
-    CUDA_CHECK(cudaMemcpyAsync((char *)tensor->data + offset, data, size, cudaMemcpyHostToDevice, g_cudaStreams[g_main_device][0]));
-
-    UNUSED(backend);
+    CUDA_CHECK(cudaMemcpyAsync((char *)tensor->data + offset, data, size, cudaMemcpyHostToDevice, g_cudaStreams[cuda_ctx->device][0]));
 }
 
 static void ggml_backend_cuda_get_tensor_async(ggml_backend_t backend, const ggml_tensor * tensor, void * data, size_t offset, size_t size) {
-    GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor read out of bounds");
-    GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
+    ggml_backend_context_cuda * cuda_ctx = (ggml_backend_context_cuda *)backend->context;
+
+    GGML_ASSERT(tensor->buffer->buft == ggml_backend_cuda_buffer_type(cuda_ctx->device) && "unsupported buffer type");
     GGML_ASSERT(tensor->backend == GGML_BACKEND_GPU);
 
-    CUDA_CHECK(cudaMemcpyAsync(data, (const char *)tensor->data + offset, size, cudaMemcpyDeviceToHost, g_cudaStreams[g_main_device][0]));
-
-    UNUSED(backend);
+    CUDA_CHECK(cudaMemcpyAsync(data, (const char *)tensor->data + offset, size, cudaMemcpyDeviceToHost, g_cudaStreams[cuda_ctx->device][0]));
 }
 
 static void ggml_backend_cuda_synchronize(ggml_backend_t backend) {
-    CUDA_CHECK(cudaStreamSynchronize(g_cudaStreams[g_main_device][0]));
+    ggml_backend_context_cuda * cuda_ctx = (ggml_backend_context_cuda *)backend->context;
+
+    CUDA_CHECK(cudaStreamSynchronize(g_cudaStreams[cuda_ctx->device][0]));
 
     UNUSED(backend);
 }
@@ -8394,14 +9764,14 @@ static ggml_backend_graph_plan_t ggml_backend_cuda_graph_plan_create(ggml_backen
     UNUSED(cgraph);
 }
 
-[[noreturn]] static void ggml_backend_cuda_graph_plan_free(ggml_backend_t backend, ggml_backend_graph_plan_t plan) {
+static void ggml_backend_cuda_graph_plan_free(ggml_backend_t backend, ggml_backend_graph_plan_t plan) {
     GGML_ASSERT(!"not implemented");
 
     UNUSED(backend);
     UNUSED(plan);
 }
 
-[[noreturn]] static void ggml_backend_cuda_graph_plan_compute(ggml_backend_t backend, ggml_backend_graph_plan_t plan) {
+static void ggml_backend_cuda_graph_plan_compute(ggml_backend_t backend, ggml_backend_graph_plan_t plan) {
     GGML_ASSERT(!"not implemented");
 
     UNUSED(backend);
@@ -8409,7 +9779,9 @@ static ggml_backend_graph_plan_t ggml_backend_cuda_graph_plan_create(ggml_backen
 }
 
 static void ggml_backend_cuda_graph_compute(ggml_backend_t backend, ggml_cgraph * cgraph) {
-    ggml_cuda_set_device(g_main_device);
+    ggml_backend_context_cuda * cuda_ctx = (ggml_backend_context_cuda *)backend->context;
+
+    ggml_cuda_set_main_device(cuda_ctx->device);
 
     ggml_compute_params params = {};
     params.type = GGML_TASK_COMPUTE;
@@ -8417,13 +9789,18 @@ static void ggml_backend_cuda_graph_compute(ggml_backend_t backend, ggml_cgraph
     for (int i = 0; i < cgraph->n_nodes; i++) {
         ggml_tensor * node = cgraph->nodes[i];
 
-        if (node->op == GGML_OP_RESHAPE || node->op == GGML_OP_TRANSPOSE || node->op == GGML_OP_VIEW || node->op == GGML_OP_PERMUTE) {
+        if (node->op == GGML_OP_RESHAPE || node->op == GGML_OP_TRANSPOSE || node->op == GGML_OP_VIEW || node->op == GGML_OP_PERMUTE)
             continue;
-        }
+
         assert(node->backend == GGML_BACKEND_GPU);
+        assert(node->buffer->buft == ggml_backend_cuda_buffer_type(cuda_ctx->device));
+        assert(node->extra != nullptr);
+
         for (int j = 0; j < GGML_MAX_SRC; j++) {
             if (node->src[j] != nullptr) {
                 assert(node->src[j]->backend == GGML_BACKEND_GPU);
+                assert(node->src[j]->buffer->buft == ggml_backend_cuda_buffer_type(cuda_ctx->device));
+                assert(node->src[j]->extra != nullptr);
             }
         }
 
@@ -8460,27 +9837,143 @@ static void ggml_backend_cuda_graph_compute(ggml_backend_t backend, ggml_cgraph
     UNUSED(backend);
 }
 
+static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, const ggml_tensor * op) {
+    switch (op->op) {
+        case GGML_OP_UNARY:
+            switch (ggml_get_unary_op(op)) {
+                case GGML_UNARY_OP_GELU:
+                case GGML_UNARY_OP_SILU:
+                case GGML_UNARY_OP_RELU:
+                case GGML_UNARY_OP_GELU_QUICK:
+                case GGML_UNARY_OP_TANH:
+                    return true;
+                default:
+                    return false;
+            }
+            break;
+        case GGML_OP_MUL_MAT:
+        case GGML_OP_MUL_MAT_ID:
+            {
+                struct ggml_tensor * a;
+                struct ggml_tensor * b;
+                if (op->op == GGML_OP_MUL_MAT) {
+                    a = op->src[0];
+                    b = op->src[1];
+                } else {
+                    a = op->src[2];
+                    b = op->src[1];
+                }
+                if (a->ne[3] != b->ne[3]) {
+                    return false;
+                }
+                return true;
+            } break;
+        case GGML_OP_GET_ROWS:
+            {
+                switch (op->src[0]->type) {
+                    case GGML_TYPE_F16:
+                    case GGML_TYPE_F32:
+                    case GGML_TYPE_Q4_0:
+                    case GGML_TYPE_Q4_1:
+                    case GGML_TYPE_Q5_0:
+                    case GGML_TYPE_Q5_1:
+                    case GGML_TYPE_Q8_0:
+                        return true;
+                    default:
+                        return false;
+                }
+            } break;
+        case GGML_OP_CPY:
+            {
+                ggml_type src0_type = op->src[0]->type;
+                ggml_type src1_type = op->src[1]->type;
+                if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_F32) {
+                    return true;
+                }
+                if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_F16) {
+                    return true;
+                }
+                if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_Q8_0) {
+                    return true;
+                }
+                if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_Q4_0) {
+                    return true;
+                }
+                if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_Q4_1) {
+                    return true;
+                }
+                if (src0_type == GGML_TYPE_F16 && src1_type == GGML_TYPE_F16) {
+                    return true;
+                }
+                return false;
+            } break;
+        case GGML_OP_NONE:
+        case GGML_OP_RESHAPE:
+        case GGML_OP_VIEW:
+        case GGML_OP_PERMUTE:
+        case GGML_OP_TRANSPOSE:
+        case GGML_OP_NORM:
+        case GGML_OP_REPEAT:
+        case GGML_OP_DUP:
+        case GGML_OP_ADD:
+        case GGML_OP_MUL:
+        case GGML_OP_DIV:
+        case GGML_OP_RMS_NORM:
+        case GGML_OP_SCALE:
+        case GGML_OP_SQR:
+        case GGML_OP_CLAMP:
+        case GGML_OP_CONT:
+        case GGML_OP_DIAG_MASK_INF:
+        case GGML_OP_SOFT_MAX:
+        case GGML_OP_ROPE:
+        case GGML_OP_ALIBI:
+        case GGML_OP_IM2COL:
+        case GGML_OP_SUM_ROWS:
+        case GGML_OP_ARGSORT:
+        case GGML_OP_ACC:
+        case GGML_OP_CONCAT:
+        case GGML_OP_GROUP_NORM:
+        case GGML_OP_UPSCALE:
+        case GGML_OP_PAD:
+        case GGML_OP_LEAKY_RELU:
+            return true;
+        default:
+            return false;
+    }
+
+    UNUSED(backend);
+}
+
 static ggml_backend_i cuda_backend_i = {
-    /* .get_name            = */ ggml_backend_cuda_name,
-    /* .free                = */ ggml_backend_cuda_free,
-    /* .alloc_buffer        = */ ggml_backend_cuda_alloc_buffer,
-    /* .get_alignment       = */ ggml_backend_cuda_get_alignment,
-    /* .set_tensor_async    = */ ggml_backend_cuda_set_tensor_async,
-    /* .get_tensor_async    = */ ggml_backend_cuda_get_tensor_async,
-    /* .synchronize         = */ ggml_backend_cuda_synchronize,
-    /* .cpy_tensor_from     = */ nullptr,
-    /* .cpy_tensor_to       = */ nullptr,
-    /* .graph_plan_create   = */ ggml_backend_cuda_graph_plan_create,
-    /* .graph_plan_free     = */ ggml_backend_cuda_graph_plan_free,
-    /* .graph_plan_compute  = */ ggml_backend_cuda_graph_plan_compute,
-    /* .graph_compute       = */ ggml_backend_cuda_graph_compute,
-    /* .supports_op         = */ nullptr,
+    /* .get_name                = */ ggml_backend_cuda_name,
+    /* .free                    = */ ggml_backend_cuda_free,
+    /* .get_default_buffer_type = */ ggml_backend_cuda_get_default_buffer_type,
+    /* .set_tensor_async        = */ ggml_backend_cuda_set_tensor_async,
+    /* .get_tensor_async        = */ ggml_backend_cuda_get_tensor_async,
+    /* .cpy_tensor_from_async   = */ NULL,
+    /* .cpy_tensor_to_async     = */ NULL,
+    /* .synchronize             = */ ggml_backend_cuda_synchronize,
+    /* .graph_plan_create       = */ ggml_backend_cuda_graph_plan_create,
+    /* .graph_plan_free         = */ ggml_backend_cuda_graph_plan_free,
+    /* .graph_plan_compute      = */ ggml_backend_cuda_graph_plan_compute,
+    /* .graph_compute           = */ ggml_backend_cuda_graph_compute,
+    /* .supports_op             = */ ggml_backend_cuda_supports_op,
 };
 
-ggml_backend_t ggml_backend_cuda_init() {
+ggml_backend_t ggml_backend_cuda_init(int device) {
     ggml_init_cublas(); // TODO: remove from ggml.c
 
-    ggml_backend_context_cuda * ctx = new ggml_backend_context_cuda;
+    if (device < 0 || device >= ggml_cuda_get_device_count()) {
+        fprintf(stderr, "%s: error: invalid device %d\n", __func__, device);
+        return nullptr;
+    }
+
+    // not strictly necessary, but it may reduce the overhead of the first graph_compute
+    ggml_cuda_set_main_device(device);
+
+    ggml_backend_context_cuda * ctx = new ggml_backend_context_cuda {
+        /* .device = */ device
+    };
 
     ggml_backend_t cuda_backend = new ggml_backend {
         /* .interface = */ cuda_backend_i,
@@ -8489,3 +9982,27 @@ ggml_backend_t ggml_backend_cuda_init() {
 
     return cuda_backend;
 }
+
+bool ggml_backend_is_cuda(ggml_backend_t backend) {
+    return backend->iface.get_name == ggml_backend_cuda_name;
+}
+
+static ggml_backend_t ggml_backend_reg_cuda_init(const char * params, void * user_data) {
+    ggml_backend_t cuda_backend = ggml_backend_cuda_init((int) (intptr_t) user_data);
+    return cuda_backend;
+
+    UNUSED(params);
+}
+
+extern "C" int ggml_backend_cuda_reg_devices();
+
+int ggml_backend_cuda_reg_devices() {
+    int device_count = ggml_cuda_get_device_count();
+    //int device_count = 1; // DEBUG: some tools require delaying CUDA initialization
+    for (int i = 0; i < device_count; i++) {
+        char name[128];
+        snprintf(name, sizeof(name), "%s%d", GGML_CUDA_NAME, i);
+        ggml_backend_register(name, ggml_backend_reg_cuda_init, ggml_backend_cuda_buffer_type(i), (void *) (intptr_t) i);
+    }
+    return device_count;
+}
diff --git a/ggml-cuda.h b/ggml-cuda.h
index 528e66c33..cdb0c0c41 100644
--- a/ggml-cuda.h
+++ b/ggml-cuda.h
@@ -49,7 +49,15 @@ GGML_API int    ggml_cuda_get_device_count(void);
 GGML_API void   ggml_cuda_get_device_description(int device, char * description, size_t description_size);
 
 // backend API
-GGML_API ggml_backend_t ggml_backend_cuda_init(void); // TODO: take a list of devices to use
+GGML_API ggml_backend_t ggml_backend_cuda_init(int device);
+
+GGML_API bool ggml_backend_is_cuda(ggml_backend_t backend);
+GGML_API int  ggml_backend_cuda_get_device(ggml_backend_t backend);
+
+GGML_API ggml_backend_buffer_type_t ggml_backend_cuda_buffer_type(int device);
+
+// pinned host buffer for use with CPU backend for faster copies between CPU and GPU
+GGML_API ggml_backend_buffer_type_t ggml_backend_cuda_host_buffer_type(void);
 
 #ifdef  __cplusplus
 }
diff --git a/ggml-impl.h b/ggml-impl.h
index 06c07339e..1f5610a86 100644
--- a/ggml-impl.h
+++ b/ggml-impl.h
@@ -232,7 +232,7 @@ bool   ggml_hash_contains      (const struct ggml_hash_set hash_set, struct ggml
 // returns GGML_HASHTABLE_FULL if table is full, otherwise the current index of the key or where it should be inserted
 size_t ggml_hash_find          (const struct ggml_hash_set hash_set, struct ggml_tensor * key);
 
-// returns GGML_HAHSHTABLE_ALREADY_EXISTS if key already exists, index otherwise, asserts if table is full
+// returns GGML_HASHTABLE_ALREADY_EXISTS if key already exists, index otherwise, asserts if table is full
 size_t ggml_hash_insert        (      struct ggml_hash_set hash_set, struct ggml_tensor * key);
 
 // return index, asserts if table is full
diff --git a/ggml-metal.h b/ggml-metal.h
index be2731f8b..b5e02b668 100644
--- a/ggml-metal.h
+++ b/ggml-metal.h
@@ -98,8 +98,17 @@ GGML_API ggml_backend_t ggml_backend_metal_init(void);
 
 GGML_API bool ggml_backend_is_metal(ggml_backend_t backend);
 
+GGML_API ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void * data, size_t size, size_t max_size);
+
 GGML_API void ggml_backend_metal_set_n_cb(ggml_backend_t backend, int n_cb);
 
+GGML_API ggml_backend_buffer_type_t ggml_backend_metal_buffer_type(void);
+
+// helper to check if the device supports a specific family
+// ideally, the user code should be doing these checks
+// ref: https://developer.apple.com/metal/Metal-Feature-Set-Tables.pdf
+GGML_API bool ggml_backend_metal_supports_family(ggml_backend_t backend, int family);
+
 #ifdef __cplusplus
 }
 #endif
diff --git a/ggml-metal.m b/ggml-metal.m
index 3343bc8a3..51a72ae33 100644
--- a/ggml-metal.m
+++ b/ggml-metal.m
@@ -62,11 +62,15 @@ struct ggml_metal_context {
     GGML_METAL_DECL_KERNEL(add_row); // TODO: avoid this extra kernel, instead extend the "add" kernel to support broadcast
     GGML_METAL_DECL_KERNEL(mul);
     GGML_METAL_DECL_KERNEL(mul_row); // TODO: avoid this extra kernel, instead extend the "mul" kernel to support broadcast
+    GGML_METAL_DECL_KERNEL(div);
+    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);
@@ -84,6 +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);
@@ -100,6 +105,21 @@ struct ggml_metal_context {
     GGML_METAL_DECL_KERNEL(mul_mv_q4_K_f32);
     GGML_METAL_DECL_KERNEL(mul_mv_q5_K_f32);
     GGML_METAL_DECL_KERNEL(mul_mv_q6_K_f32);
+    GGML_METAL_DECL_KERNEL(mul_mv_id_f32_f32);
+    //GGML_METAL_DECL_KERNEL(mul_mv_id_f16_f16);
+    GGML_METAL_DECL_KERNEL(mul_mv_id_f16_f32);
+    //GGML_METAL_DECL_KERNEL(mul_mv_id_f16_f32_1row);
+    //GGML_METAL_DECL_KERNEL(mul_mv_id_f16_f32_l4);
+    GGML_METAL_DECL_KERNEL(mul_mv_id_q4_0_f32);
+    GGML_METAL_DECL_KERNEL(mul_mv_id_q4_1_f32);
+    GGML_METAL_DECL_KERNEL(mul_mv_id_q5_0_f32);
+    GGML_METAL_DECL_KERNEL(mul_mv_id_q5_1_f32);
+    GGML_METAL_DECL_KERNEL(mul_mv_id_q8_0_f32);
+    GGML_METAL_DECL_KERNEL(mul_mv_id_q2_K_f32);
+    GGML_METAL_DECL_KERNEL(mul_mv_id_q3_K_f32);
+    GGML_METAL_DECL_KERNEL(mul_mv_id_q4_K_f32);
+    GGML_METAL_DECL_KERNEL(mul_mv_id_q5_K_f32);
+    GGML_METAL_DECL_KERNEL(mul_mv_id_q6_K_f32);
     GGML_METAL_DECL_KERNEL(mul_mm_f32_f32);
     GGML_METAL_DECL_KERNEL(mul_mm_f16_f32);
     GGML_METAL_DECL_KERNEL(mul_mm_q4_0_f32);
@@ -112,15 +132,39 @@ struct ggml_metal_context {
     GGML_METAL_DECL_KERNEL(mul_mm_q4_K_f32);
     GGML_METAL_DECL_KERNEL(mul_mm_q5_K_f32);
     GGML_METAL_DECL_KERNEL(mul_mm_q6_K_f32);
+    GGML_METAL_DECL_KERNEL(mul_mm_id_f32_f32);
+    GGML_METAL_DECL_KERNEL(mul_mm_id_f16_f32);
+    GGML_METAL_DECL_KERNEL(mul_mm_id_q4_0_f32);
+    GGML_METAL_DECL_KERNEL(mul_mm_id_q4_1_f32);
+    GGML_METAL_DECL_KERNEL(mul_mm_id_q5_0_f32);
+    GGML_METAL_DECL_KERNEL(mul_mm_id_q5_1_f32);
+    GGML_METAL_DECL_KERNEL(mul_mm_id_q8_0_f32);
+    GGML_METAL_DECL_KERNEL(mul_mm_id_q2_K_f32);
+    GGML_METAL_DECL_KERNEL(mul_mm_id_q3_K_f32);
+    GGML_METAL_DECL_KERNEL(mul_mm_id_q4_K_f32);
+    GGML_METAL_DECL_KERNEL(mul_mm_id_q5_K_f32);
+    GGML_METAL_DECL_KERNEL(mul_mm_id_q6_K_f32);
     GGML_METAL_DECL_KERNEL(rope_f32);
     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);
+    GGML_METAL_DECL_KERNEL(cpy_f32_q4_0);
+    GGML_METAL_DECL_KERNEL(cpy_f32_q4_1);
+    //GGML_METAL_DECL_KERNEL(cpy_f32_q5_0);
+    //GGML_METAL_DECL_KERNEL(cpy_f32_q5_1);
     GGML_METAL_DECL_KERNEL(cpy_f16_f16);
+    GGML_METAL_DECL_KERNEL(cpy_f16_f32);
     GGML_METAL_DECL_KERNEL(concat);
     GGML_METAL_DECL_KERNEL(sqr);
+    GGML_METAL_DECL_KERNEL(sum_rows);
 
 #undef GGML_METAL_DECL_KERNEL
 };
@@ -136,7 +180,15 @@ struct ggml_metal_context {
 @implementation GGMLMetalClass
 @end
 
-ggml_log_callback ggml_metal_log_callback = NULL;
+
+static void ggml_metal_default_log_callback(enum ggml_log_level level, const char * msg, void * user_data) {
+    fprintf(stderr, "%s", msg);
+
+    UNUSED(level);
+    UNUSED(user_data);
+}
+
+ggml_log_callback ggml_metal_log_callback = ggml_metal_default_log_callback;
 void * ggml_metal_log_user_data = NULL;
 
 void ggml_metal_log_set_callback(ggml_log_callback log_callback, void * user_data) {
@@ -155,6 +207,8 @@ static void ggml_metal_log(enum ggml_log_level level, const char * format, ...){
             ggml_metal_log_callback(level, buffer, ggml_metal_log_user_data);
         } else {
             char* buffer2 = malloc(len+1);
+            va_end(args);
+            va_start(args, format);
             vsnprintf(buffer2, len+1, format, args);
             buffer2[len] = 0;
             ggml_metal_log_callback(level, buffer2, ggml_metal_log_user_data);
@@ -164,12 +218,10 @@ static void ggml_metal_log(enum ggml_log_level level, const char * format, ...){
     }
 }
 
-
-
 struct ggml_metal_context * ggml_metal_init(int n_cb) {
     GGML_METAL_LOG_INFO("%s: allocating\n", __func__);
 
-    id <MTLDevice> device;
+    id<MTLDevice> device;
     NSString * s;
 
 #if TARGET_OS_OSX
@@ -215,6 +267,9 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
 
             NSString * sourcePath;
             NSString * ggmlMetalPathResources = [[NSProcessInfo processInfo].environment objectForKey:@"GGML_METAL_PATH_RESOURCES"];
+
+            GGML_METAL_LOG_INFO("%s: GGML_METAL_PATH_RESOURCES = %s\n", __func__, ggmlMetalPathResources ? [ggmlMetalPathResources UTF8String] : "nil");
+
             if (ggmlMetalPathResources) {
                 sourcePath = [ggmlMetalPathResources stringByAppendingPathComponent:@"ggml-metal.metal"];
             } else {
@@ -245,6 +300,29 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
         }
     }
 
+#if TARGET_OS_OSX
+    // print MTL GPU family:
+    GGML_METAL_LOG_INFO("%s: GPU name:   %s\n", __func__, [[ctx->device name] UTF8String]);
+
+    // determine max supported GPU family
+    // https://developer.apple.com/metal/Metal-Shading-Language-Specification.pdf
+    // https://developer.apple.com/metal/Metal-Feature-Set-Tables.pdf
+    for (int i = MTLGPUFamilyApple1 + 20; i >= MTLGPUFamilyApple1; --i) {
+        if ([ctx->device supportsFamily:i]) {
+            GGML_METAL_LOG_INFO("%s: GPU family: MTLGPUFamilyApple%d (%d)\n", __func__, i - (int) MTLGPUFamilyApple1 + 1, i);
+            break;
+        }
+    }
+
+    GGML_METAL_LOG_INFO("%s: hasUnifiedMemory              = %s\n",       __func__, ctx->device.hasUnifiedMemory ? "true" : "false");
+    GGML_METAL_LOG_INFO("%s: recommendedMaxWorkingSetSize  = %8.2f MB\n", __func__, ctx->device.recommendedMaxWorkingSetSize / 1e6);
+    if (ctx->device.maxTransferRate != 0) {
+        GGML_METAL_LOG_INFO("%s: maxTransferRate               = %8.2f MB/s\n", __func__, ctx->device.maxTransferRate / 1e6);
+    } else {
+        GGML_METAL_LOG_INFO("%s: maxTransferRate               = built-in GPU\n", __func__);
+    }
+#endif
+
     // load kernels
     {
         NSError * error = nil;
@@ -266,11 +344,15 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
         GGML_METAL_ADD_KERNEL(add_row);
         GGML_METAL_ADD_KERNEL(mul);
         GGML_METAL_ADD_KERNEL(mul_row);
+        GGML_METAL_ADD_KERNEL(div);
+        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);
@@ -288,6 +370,7 @@ 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);
@@ -304,6 +387,21 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
         GGML_METAL_ADD_KERNEL(mul_mv_q4_K_f32);
         GGML_METAL_ADD_KERNEL(mul_mv_q5_K_f32);
         GGML_METAL_ADD_KERNEL(mul_mv_q6_K_f32);
+        GGML_METAL_ADD_KERNEL(mul_mv_id_f32_f32);
+        //GGML_METAL_ADD_KERNEL(mul_mv_id_f16_f16);
+        GGML_METAL_ADD_KERNEL(mul_mv_id_f16_f32);
+        //GGML_METAL_ADD_KERNEL(mul_mv_id_f16_f32_1row);
+        //GGML_METAL_ADD_KERNEL(mul_mv_id_f16_f32_l4);
+        GGML_METAL_ADD_KERNEL(mul_mv_id_q4_0_f32);
+        GGML_METAL_ADD_KERNEL(mul_mv_id_q4_1_f32);
+        GGML_METAL_ADD_KERNEL(mul_mv_id_q5_0_f32);
+        GGML_METAL_ADD_KERNEL(mul_mv_id_q5_1_f32);
+        GGML_METAL_ADD_KERNEL(mul_mv_id_q8_0_f32);
+        GGML_METAL_ADD_KERNEL(mul_mv_id_q2_K_f32);
+        GGML_METAL_ADD_KERNEL(mul_mv_id_q3_K_f32);
+        GGML_METAL_ADD_KERNEL(mul_mv_id_q4_K_f32);
+        GGML_METAL_ADD_KERNEL(mul_mv_id_q5_K_f32);
+        GGML_METAL_ADD_KERNEL(mul_mv_id_q6_K_f32);
         if ([ctx->device supportsFamily:MTLGPUFamilyApple7]) {
             GGML_METAL_ADD_KERNEL(mul_mm_f32_f32);
             GGML_METAL_ADD_KERNEL(mul_mm_f16_f32);
@@ -317,43 +415,44 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
             GGML_METAL_ADD_KERNEL(mul_mm_q4_K_f32);
             GGML_METAL_ADD_KERNEL(mul_mm_q5_K_f32);
             GGML_METAL_ADD_KERNEL(mul_mm_q6_K_f32);
+            GGML_METAL_ADD_KERNEL(mul_mm_id_f32_f32);
+            GGML_METAL_ADD_KERNEL(mul_mm_id_f16_f32);
+            GGML_METAL_ADD_KERNEL(mul_mm_id_q4_0_f32);
+            GGML_METAL_ADD_KERNEL(mul_mm_id_q4_1_f32);
+            GGML_METAL_ADD_KERNEL(mul_mm_id_q5_0_f32);
+            GGML_METAL_ADD_KERNEL(mul_mm_id_q5_1_f32);
+            GGML_METAL_ADD_KERNEL(mul_mm_id_q8_0_f32);
+            GGML_METAL_ADD_KERNEL(mul_mm_id_q2_K_f32);
+            GGML_METAL_ADD_KERNEL(mul_mm_id_q3_K_f32);
+            GGML_METAL_ADD_KERNEL(mul_mm_id_q4_K_f32);
+            GGML_METAL_ADD_KERNEL(mul_mm_id_q5_K_f32);
+            GGML_METAL_ADD_KERNEL(mul_mm_id_q6_K_f32);
         }
         GGML_METAL_ADD_KERNEL(rope_f32);
         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);
+        GGML_METAL_ADD_KERNEL(cpy_f32_q4_0);
+        GGML_METAL_ADD_KERNEL(cpy_f32_q4_1);
+        //GGML_METAL_ADD_KERNEL(cpy_f32_q5_0);
+        //GGML_METAL_ADD_KERNEL(cpy_f32_q5_1);
         GGML_METAL_ADD_KERNEL(cpy_f16_f16);
+        GGML_METAL_ADD_KERNEL(cpy_f16_f32);
         GGML_METAL_ADD_KERNEL(concat);
         GGML_METAL_ADD_KERNEL(sqr);
+        GGML_METAL_ADD_KERNEL(sum_rows);
 
 #undef GGML_METAL_ADD_KERNEL
     }
 
-#if TARGET_OS_OSX
-    // print MTL GPU family:
-    GGML_METAL_LOG_INFO("%s: GPU name:   %s\n", __func__, [[ctx->device name] UTF8String]);
-
-    // determine max supported GPU family
-    // https://developer.apple.com/metal/Metal-Shading-Language-Specification.pdf
-    // https://developer.apple.com/metal/Metal-Feature-Set-Tables.pdf
-    for (int i = MTLGPUFamilyApple1 + 20; i >= MTLGPUFamilyApple1; --i) {
-        if ([ctx->device supportsFamily:i]) {
-            GGML_METAL_LOG_INFO("%s: GPU family: MTLGPUFamilyApple%d (%d)\n", __func__, i - (int) MTLGPUFamilyApple1 + 1, i);
-            break;
-        }
-    }
-
-    GGML_METAL_LOG_INFO("%s: hasUnifiedMemory              = %s\n",        __func__, ctx->device.hasUnifiedMemory ? "true" : "false");
-    GGML_METAL_LOG_INFO("%s: recommendedMaxWorkingSetSize  = %8.2f MiB\n", __func__, ctx->device.recommendedMaxWorkingSetSize / 1024.0 / 1024.0);
-    if (ctx->device.maxTransferRate != 0) {
-        GGML_METAL_LOG_INFO("%s: maxTransferRate               = %8.2f MiB/s\n", __func__, ctx->device.maxTransferRate / 1024.0 / 1024.0);
-    } else {
-        GGML_METAL_LOG_INFO("%s: maxTransferRate               = built-in GPU\n", __func__);
-    }
-#endif
-
     return ctx;
 }
 
@@ -367,11 +466,15 @@ void ggml_metal_free(struct ggml_metal_context * ctx) {
     GGML_METAL_DEL_KERNEL(add_row);
     GGML_METAL_DEL_KERNEL(mul);
     GGML_METAL_DEL_KERNEL(mul_row);
+    GGML_METAL_DEL_KERNEL(div);
+    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);
@@ -389,6 +492,7 @@ 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);
@@ -405,6 +509,21 @@ void ggml_metal_free(struct ggml_metal_context * ctx) {
     GGML_METAL_DEL_KERNEL(mul_mv_q4_K_f32);
     GGML_METAL_DEL_KERNEL(mul_mv_q5_K_f32);
     GGML_METAL_DEL_KERNEL(mul_mv_q6_K_f32);
+    GGML_METAL_DEL_KERNEL(mul_mv_id_f32_f32);
+    //GGML_METAL_DEL_KERNEL(mul_mv_id_f16_f16);
+    GGML_METAL_DEL_KERNEL(mul_mv_id_f16_f32);
+    //GGML_METAL_DEL_KERNEL(mul_mv_id_f16_f32_1row);
+    //GGML_METAL_DEL_KERNEL(mul_mv_id_f16_f32_l4);
+    GGML_METAL_DEL_KERNEL(mul_mv_id_q4_0_f32);
+    GGML_METAL_DEL_KERNEL(mul_mv_id_q4_1_f32);
+    GGML_METAL_DEL_KERNEL(mul_mv_id_q5_0_f32);
+    GGML_METAL_DEL_KERNEL(mul_mv_id_q5_1_f32);
+    GGML_METAL_DEL_KERNEL(mul_mv_id_q8_0_f32);
+    GGML_METAL_DEL_KERNEL(mul_mv_id_q2_K_f32);
+    GGML_METAL_DEL_KERNEL(mul_mv_id_q3_K_f32);
+    GGML_METAL_DEL_KERNEL(mul_mv_id_q4_K_f32);
+    GGML_METAL_DEL_KERNEL(mul_mv_id_q5_K_f32);
+    GGML_METAL_DEL_KERNEL(mul_mv_id_q6_K_f32);
     if ([ctx->device supportsFamily:MTLGPUFamilyApple7]) {
         GGML_METAL_DEL_KERNEL(mul_mm_f32_f32);
         GGML_METAL_DEL_KERNEL(mul_mm_f16_f32);
@@ -418,16 +537,40 @@ void ggml_metal_free(struct ggml_metal_context * ctx) {
         GGML_METAL_DEL_KERNEL(mul_mm_q4_K_f32);
         GGML_METAL_DEL_KERNEL(mul_mm_q5_K_f32);
         GGML_METAL_DEL_KERNEL(mul_mm_q6_K_f32);
+        GGML_METAL_DEL_KERNEL(mul_mm_id_f32_f32);
+        GGML_METAL_DEL_KERNEL(mul_mm_id_f16_f32);
+        GGML_METAL_DEL_KERNEL(mul_mm_id_q4_0_f32);
+        GGML_METAL_DEL_KERNEL(mul_mm_id_q4_1_f32);
+        GGML_METAL_DEL_KERNEL(mul_mm_id_q5_0_f32);
+        GGML_METAL_DEL_KERNEL(mul_mm_id_q5_1_f32);
+        GGML_METAL_DEL_KERNEL(mul_mm_id_q8_0_f32);
+        GGML_METAL_DEL_KERNEL(mul_mm_id_q2_K_f32);
+        GGML_METAL_DEL_KERNEL(mul_mm_id_q3_K_f32);
+        GGML_METAL_DEL_KERNEL(mul_mm_id_q4_K_f32);
+        GGML_METAL_DEL_KERNEL(mul_mm_id_q5_K_f32);
+        GGML_METAL_DEL_KERNEL(mul_mm_id_q6_K_f32);
     }
     GGML_METAL_DEL_KERNEL(rope_f32);
     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);
+    GGML_METAL_DEL_KERNEL(cpy_f32_q4_0);
+    GGML_METAL_DEL_KERNEL(cpy_f32_q4_1);
+    //GGML_METAL_DEL_KERNEL(cpy_f32_q5_0);
+    //GGML_METAL_DEL_KERNEL(cpy_f32_q5_1);
     GGML_METAL_DEL_KERNEL(cpy_f16_f16);
+    GGML_METAL_DEL_KERNEL(cpy_f16_f32);
     GGML_METAL_DEL_KERNEL(concat);
     GGML_METAL_DEL_KERNEL(sqr);
+    GGML_METAL_DEL_KERNEL(sum_rows);
 
 #undef GGML_METAL_DEL_KERNEL
 
@@ -471,6 +614,25 @@ int * ggml_metal_get_concur_list(struct ggml_metal_context * ctx) {
     return ctx->concur_list;
 }
 
+// temporarily defined here for compatibility between ggml-backend and the old API
+
+struct ggml_backend_metal_buffer {
+    void   * data;
+    size_t   size;
+
+    id<MTLBuffer> metal;
+};
+
+struct ggml_backend_metal_buffer_context {
+    void * all_data;
+    size_t all_size;
+    bool owned;
+
+    // multiple buffers are used only to avoid the maximum buffer size limitation when using mmap
+    int n_buffers;
+    struct ggml_backend_metal_buffer buffers[GGML_METAL_MAX_BUFFERS];
+};
+
 // finds the Metal buffer that contains the tensor data on the GPU device
 // the assumption is that there is 1-to-1 mapping between the host and device memory buffers, so we can find the
 // Metal buffer based on the host memory pointer
@@ -480,8 +642,29 @@ static id<MTLBuffer> ggml_metal_get_buffer(struct ggml_metal_context * ctx, stru
 
     const int64_t tsize = ggml_nbytes(t);
 
-    if (t->buffer && t->buffer->backend && t->buffer->backend->context) {
-        ctx = t->buffer->backend->context;
+    ggml_backend_buffer_t buffer = t->view_src ? t->view_src->buffer : t->buffer;
+
+    // compatibility with ggml-backend
+    if (buffer && buffer->buft == ggml_backend_metal_buffer_type()) {
+        struct ggml_backend_metal_buffer_context * buf_ctx = (struct ggml_backend_metal_buffer_context *) buffer->context;
+
+        // find the view that contains the tensor fully
+        for (int i = 0; i < buf_ctx->n_buffers; ++i) {
+            const int64_t ioffs = (int64_t) t->data - (int64_t) buf_ctx->buffers[i].data;
+
+            //GGML_METAL_LOG_INFO("ioffs = %10ld, tsize = %10ld, sum = %10ld, buf_ctx->buffers[%d].size = %10ld\n", ioffs, tsize, ioffs + tsize, i, buf_ctx->buffers[i].size);
+            if (ioffs >= 0 && ioffs + tsize <= (int64_t) buf_ctx->buffers[i].size) {
+                *offs = (size_t) ioffs;
+
+                //GGML_METAL_LOG_INFO("%s: tensor '%16s', offs = %8ld\n", __func__, t->name, *offs);
+
+                return buf_ctx->buffers[i].metal;
+            }
+        }
+
+        GGML_METAL_LOG_ERROR("%s: error: tensor '%s' buffer is nil\n", __func__, t->name);
+
+        return nil;
     }
 
     // find the view that contains the tensor fully
@@ -706,6 +889,83 @@ void ggml_metal_graph_find_concurrency(
     }
 }
 
+static bool ggml_metal_supports_op(const struct ggml_tensor * op) {
+    switch (op->op) {
+        case GGML_OP_UNARY:
+            switch (ggml_get_unary_op(op)) {
+                case GGML_UNARY_OP_TANH:
+                case GGML_UNARY_OP_RELU:
+                case GGML_UNARY_OP_GELU:
+                case GGML_UNARY_OP_GELU_QUICK:
+                case GGML_UNARY_OP_SILU:
+                    return true;
+                default:
+                    return false;
+            }
+        case GGML_OP_NONE:
+        case GGML_OP_RESHAPE:
+        case GGML_OP_VIEW:
+        case GGML_OP_TRANSPOSE:
+        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:
+        case GGML_OP_SQR:
+        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;
+        case GGML_OP_CPY:
+        case GGML_OP_DUP:
+        case GGML_OP_CONT:
+            {
+                switch (op->src[0]->type) {
+                    case GGML_TYPE_F32:
+                        switch (op->type) {
+                           case GGML_TYPE_F16:
+                           case GGML_TYPE_F32:
+                           case GGML_TYPE_Q8_0:
+                           case GGML_TYPE_Q4_0:
+                           case GGML_TYPE_Q4_1:
+                                return true;
+                           default:
+                                return false;
+                        }
+                    case GGML_TYPE_F16:
+                        switch (op->type) {
+                           case GGML_TYPE_F16:
+                           case GGML_TYPE_F32:
+                                return true;
+                           default:
+                                return false;
+                        }
+                    default:
+                        return false;
+                };
+            }
+        case GGML_OP_DIAG_MASK_INF:
+        case GGML_OP_GET_ROWS:
+            {
+                return op->ne[3] == 1;
+            }
+        default:
+            return false;
+    }
+}
 void ggml_metal_graph_compute(
         struct ggml_metal_context * ctx,
                struct ggml_cgraph * gf) {
@@ -776,6 +1036,11 @@ void ggml_metal_graph_compute(
                         } break;
                 }
 
+                if (!ggml_metal_supports_op(dst)) {
+                    GGML_METAL_LOG_ERROR("%s: error: unsupported op '%s'\n", __func__, ggml_op_desc(dst));
+                    GGML_ASSERT(!"unsupported op");
+                }
+
                 const int64_t  ne00 = src0 ? src0->ne[0] : 0;
                 const int64_t  ne01 = src0 ? src0->ne[1] : 0;
                 const int64_t  ne02 = src0 ? src0->ne[2] : 0;
@@ -868,25 +1133,42 @@ void ggml_metal_graph_compute(
                             [encoder dispatchThreadgroups:MTLSizeMake(ne1, ne2, ne3) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
                         } break;
                     case GGML_OP_ADD:
+                    case GGML_OP_MUL:
+                    case GGML_OP_DIV:
                         {
-                            GGML_ASSERT(ggml_is_contiguous(src0));
-                            GGML_ASSERT(ggml_is_contiguous(src1));
+                            const size_t offs = 0;
 
                             bool bcast_row = false;
 
                             int64_t nb = ne00;
 
-                            if (ggml_nelements(src1) == ne10 && ne00 % 4 == 0) {
+                            id<MTLComputePipelineState> pipeline = nil;
+
+                            if (ggml_nelements(src1) == ne10 && ggml_is_contiguous(src1) && ne00 % 4 == 0 && ne10 % 4 == 0) {
+                                GGML_ASSERT(ggml_is_contiguous(src0));
+
                                 // src1 is a row
                                 GGML_ASSERT(ne11 == 1);
 
                                 nb = ne00 / 4;
-                                [encoder setComputePipelineState:ctx->pipeline_add_row];
+                                switch (dst->op) {
+                                    case GGML_OP_ADD: pipeline = ctx->pipeline_add_row; break;
+                                    case GGML_OP_MUL: pipeline = ctx->pipeline_mul_row; break;
+                                    case GGML_OP_DIV: pipeline = ctx->pipeline_div_row; break;
+                                    default: GGML_ASSERT(false);
+                                }
 
                                 bcast_row = true;
                             } else {
-                                [encoder setComputePipelineState:ctx->pipeline_add];
+                                switch (dst->op) {
+                                    case GGML_OP_ADD: pipeline = ctx->pipeline_add; break;
+                                    case GGML_OP_MUL: pipeline = ctx->pipeline_mul; break;
+                                    case GGML_OP_DIV: pipeline = ctx->pipeline_div; break;
+                                    default: GGML_ASSERT(false);
+                                }
                             }
+
+                            [encoder setComputePipelineState:pipeline];
                             [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];
@@ -914,48 +1196,104 @@ 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;
 
                                 [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
                             } else {
-                                const int nth = MIN(1024, ne0);
+                                const int nth = MIN((int) pipeline.maxTotalThreadsPerThreadgroup, ne0);
 
                                 [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
                             }
                         } break;
-                    case GGML_OP_MUL:
+                    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));
 
-                            // utilize float4
-                            GGML_ASSERT(ne00 % 4 == 0);
-                            const int64_t nb = ne00/4;
+                            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];
 
-                            if (ggml_nelements(src1) == ne10) {
-                                // src1 is a row
-                                GGML_ASSERT(ne11 == 1);
-                                [encoder setComputePipelineState:ctx->pipeline_mul_row];
-                            } else {
-                                [encoder setComputePipelineState:ctx->pipeline_mul];
+                            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:&nb     length:sizeof(nb) atIndex:3];
+                            [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 int64_t n = ggml_nelements(dst)/4;
+                            const int nth = MIN(1024, ne0);
 
-                            [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                            [encoder dispatchThreadgroups:MTLSizeMake(ne11, ne12, ne13) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
                         } break;
                     case GGML_OP_SCALE:
                         {
                             GGML_ASSERT(ggml_is_contiguous(src0));
 
-                            const float scale = *(const float *) src1->data;
+                            const float scale = *(const float *) dst->op_params;
 
                             int64_t n = ggml_nelements(dst);
 
@@ -966,24 +1304,23 @@ void ggml_metal_graph_compute(
                                 [encoder setComputePipelineState:ctx->pipeline_scale];
                             }
 
-                            [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
-                            [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
+                            [encoder setBuffer:id_src0   offset:offs_src0 atIndex:0];
+                            [encoder setBuffer:id_dst    offset:offs_dst  atIndex:1];
                             [encoder setBytes:&scale length:sizeof(scale) atIndex:2];
 
                             [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
                         } break;
                     case GGML_OP_UNARY:
                         switch (ggml_get_unary_op(gf->nodes[i])) {
-                            case GGML_UNARY_OP_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:
                                 {
@@ -1004,6 +1341,28 @@ 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:
@@ -1023,6 +1382,40 @@ void ggml_metal_graph_compute(
                             const int64_t n = ggml_nelements(dst);
                             [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
                         } break;
+                    case GGML_OP_SUM_ROWS:
+                        {
+                            GGML_ASSERT(src0->nb[0] == ggml_type_size(src0->type));
+
+                            [encoder setComputePipelineState:ctx->pipeline_sum_rows];
+                            [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:&ne10 length:sizeof(ne10) atIndex:10];
+                            [encoder setBytes:&ne11 length:sizeof(ne11) atIndex:11];
+                            [encoder setBytes:&ne12 length:sizeof(ne12) atIndex:12];
+                            [encoder setBytes:&ne13 length:sizeof(ne13) atIndex:13];
+                            [encoder setBytes:&nb10 length:sizeof(nb10) atIndex:14];
+                            [encoder setBytes:&nb11 length:sizeof(nb11) atIndex:15];
+                            [encoder setBytes:&nb12 length:sizeof(nb12) atIndex:16];
+                            [encoder setBytes:&nb13 length:sizeof(nb13) atIndex:17];
+                            [encoder setBytes:&ne0  length:sizeof(ne0)  atIndex:18];
+                            [encoder setBytes:&ne1  length:sizeof(ne1)  atIndex:19];
+                            [encoder setBytes:&ne2  length:sizeof(ne2)  atIndex:20];
+                            [encoder setBytes:&ne3  length:sizeof(ne3)  atIndex:21];
+                            [encoder setBytes:&nb0  length:sizeof(nb0)  atIndex:22];
+                            [encoder setBytes:&nb1  length:sizeof(nb1)  atIndex:23];
+                            [encoder setBytes:&nb2  length:sizeof(nb2)  atIndex:24];
+                            [encoder setBytes:&nb3  length:sizeof(nb3)  atIndex:25];
+
+                            [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                        } break;
                     case GGML_OP_SOFT_MAX:
                         {
                             int nth = 32; // SIMD width
@@ -1042,7 +1435,11 @@ void ggml_metal_graph_compute(
                             const float scale = ((float *) dst->op_params)[0];
 
                             [encoder setBuffer:id_src0 offset:offs_src0   atIndex:0];
-                            [encoder setBuffer:id_src1 offset:offs_src1   atIndex:1];
+                            if (id_src1) {
+                                [encoder setBuffer:id_src1 offset:offs_src1   atIndex:1];
+                            } else {
+                                [encoder setBuffer:id_src0 offset:offs_src0   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];
@@ -1077,13 +1474,17 @@ void ggml_metal_graph_compute(
                     case GGML_OP_MUL_MAT:
                         {
                             GGML_ASSERT(ne00 == ne10);
-                            GGML_ASSERT(ne03 == ne13);
 
-                            const uint gqa = ne12/ne02;
+                            // TODO: assert that dim2 and dim3 are contiguous
+                            GGML_ASSERT(ne12 % ne02 == 0);
+                            GGML_ASSERT(ne13 % ne03 == 0);
+
+                            const uint r2 = ne12/ne02;
+                            const uint r3 = ne13/ne03;
 
                             // find the break-even point where the matrix-matrix kernel becomes more efficient compared
                             // to the matrix-vector kernel
-                            int ne11_mm_min = src0t == GGML_TYPE_F16 ? 1 : 16;
+                            int ne11_mm_min = 1;
 
 #if 0
                             // the numbers below are measured on M2 Ultra for 7B and 13B models
@@ -1114,7 +1515,7 @@ void ggml_metal_graph_compute(
                                 !ggml_is_transposed(src1) &&
                                 src1t == GGML_TYPE_F32 &&
                                 ne00 % 32 == 0 && ne00 >= 64 &&
-                                ne11 > ne11_mm_min) {
+                                (ne11 > ne11_mm_min || (ggml_is_quantized(src0t) && ne12 > 1))) {
                                 //printf("matrix: ne00 = %6d, ne01 = %6d, ne02 = %6d, ne11 = %6d, ne12 = %6d\n", ne00, ne01, ne02, ne11, ne12);
                                 switch (src0->type) {
                                     case GGML_TYPE_F32:  [encoder setComputePipelineState:ctx->pipeline_mul_mm_f32_f32];  break;
@@ -1144,9 +1545,10 @@ void ggml_metal_graph_compute(
                                 [encoder setBytes:&nb12    length:sizeof(nb12) atIndex:10];
                                 [encoder setBytes:&ne0     length:sizeof(ne0)  atIndex:11];
                                 [encoder setBytes:&ne1     length:sizeof(ne1)  atIndex:12];
-                                [encoder setBytes:&gqa     length:sizeof(gqa)  atIndex:13];
+                                [encoder setBytes:&r2      length:sizeof(r2)   atIndex:13];
+                                [encoder setBytes:&r3      length:sizeof(r3)   atIndex:14];
                                 [encoder setThreadgroupMemoryLength:8192 atIndex:0];
-                                [encoder dispatchThreadgroups:MTLSizeMake( (ne11 + 31)/32, (ne01 + 63)/64, ne12) threadsPerThreadgroup:MTLSizeMake(128, 1, 1)];
+                                [encoder dispatchThreadgroups:MTLSizeMake( (ne11 + 31)/32, (ne01 + 63)/64, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(128, 1, 1)];
                             } else {
                                 int nth0 = 32;
                                 int nth1 = 1;
@@ -1182,90 +1584,60 @@ void ggml_metal_graph_compute(
                                         } break;
                                     case GGML_TYPE_Q4_0:
                                         {
-                                            GGML_ASSERT(ne02 == 1);
-                                            GGML_ASSERT(ne12 == 1);
-
                                             nth0 = 8;
                                             nth1 = 8;
                                             [encoder setComputePipelineState:ctx->pipeline_mul_mv_q4_0_f32];
                                         } break;
                                     case GGML_TYPE_Q4_1:
                                         {
-                                            GGML_ASSERT(ne02 == 1);
-                                            GGML_ASSERT(ne12 == 1);
-
                                             nth0 = 8;
                                             nth1 = 8;
                                             [encoder setComputePipelineState:ctx->pipeline_mul_mv_q4_1_f32];
                                         } break;
                                     case GGML_TYPE_Q5_0:
                                         {
-                                            GGML_ASSERT(ne02 == 1);
-                                            GGML_ASSERT(ne12 == 1);
-
                                             nth0 = 8;
                                             nth1 = 8;
                                             [encoder setComputePipelineState:ctx->pipeline_mul_mv_q5_0_f32];
                                         } break;
                                     case GGML_TYPE_Q5_1:
                                         {
-                                            GGML_ASSERT(ne02 == 1);
-                                            GGML_ASSERT(ne12 == 1);
-
                                             nth0 = 8;
                                             nth1 = 8;
                                             [encoder setComputePipelineState:ctx->pipeline_mul_mv_q5_1_f32];
                                         } break;
                                     case GGML_TYPE_Q8_0:
                                         {
-                                            GGML_ASSERT(ne02 == 1);
-                                            GGML_ASSERT(ne12 == 1);
-
                                             nth0 = 8;
                                             nth1 = 8;
                                             [encoder setComputePipelineState:ctx->pipeline_mul_mv_q8_0_f32];
                                         } break;
                                     case GGML_TYPE_Q2_K:
                                         {
-                                            GGML_ASSERT(ne02 == 1);
-                                            GGML_ASSERT(ne12 == 1);
-
                                             nth0 = 2;
                                             nth1 = 32;
                                             [encoder setComputePipelineState:ctx->pipeline_mul_mv_q2_K_f32];
                                         } break;
                                     case GGML_TYPE_Q3_K:
                                         {
-                                            GGML_ASSERT(ne02 == 1);
-                                            GGML_ASSERT(ne12 == 1);
-
                                             nth0 = 2;
                                             nth1 = 32;
                                             [encoder setComputePipelineState:ctx->pipeline_mul_mv_q3_K_f32];
                                         } break;
                                     case GGML_TYPE_Q4_K:
                                         {
-                                            GGML_ASSERT(ne02 == 1);
-                                            GGML_ASSERT(ne12 == 1);
-
                                             nth0 = 4; //1;
                                             nth1 = 8; //32;
                                             [encoder setComputePipelineState:ctx->pipeline_mul_mv_q4_K_f32];
                                         } break;
                                     case GGML_TYPE_Q5_K:
                                         {
-                                            GGML_ASSERT(ne02 == 1);
-                                            GGML_ASSERT(ne12 == 1);
-
                                             nth0 = 2;
                                             nth1 = 32;
                                             [encoder setComputePipelineState:ctx->pipeline_mul_mv_q5_K_f32];
                                         } break;
                                     case GGML_TYPE_Q6_K:
                                         {
-                                            GGML_ASSERT(ne02 == 1);
-                                            GGML_ASSERT(ne12 == 1);
-
                                             nth0 = 2;
                                             nth1 = 32;
                                             [encoder setComputePipelineState:ctx->pipeline_mul_mv_q6_K_f32];
@@ -1294,31 +1666,281 @@ void ggml_metal_graph_compute(
                                 [encoder setBytes:&nb12 length:sizeof(nb12) atIndex:14];
                                 [encoder setBytes:&ne0  length:sizeof(ne0)  atIndex:15];
                                 [encoder setBytes:&ne1  length:sizeof(ne1)  atIndex:16];
-                                [encoder setBytes:&gqa  length:sizeof(gqa)  atIndex:17];
+                                [encoder setBytes:&r2   length:sizeof(r2)   atIndex:17];
+                                [encoder setBytes:&r3   length:sizeof(r3)   atIndex:18];
 
                                 if (src0t == GGML_TYPE_Q4_0 || src0t == GGML_TYPE_Q4_1 ||
                                     src0t == GGML_TYPE_Q5_0 || src0t == GGML_TYPE_Q5_1 || src0t == GGML_TYPE_Q8_0 ||
                                     src0t == GGML_TYPE_Q2_K) { // || src0t == GGML_TYPE_Q4_K) {
-                                    [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                                    [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
                                 }
                                 else if (src0t == GGML_TYPE_Q4_K) {
-                                    [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                                    [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
                                 }
                                 else if (src0t == GGML_TYPE_Q3_K) {
 #ifdef GGML_QKK_64
-                                    [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 1)/2, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                                    [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 1)/2, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
 #else
-                                    [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                                    [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
 #endif
                                 }
                                 else if (src0t == GGML_TYPE_Q5_K) {
-                                    [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                                    [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
                                 }
                                 else if (src0t == GGML_TYPE_Q6_K) {
-                                    [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 1)/2, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                                    [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 1)/2, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
                                 } else {
-                                    int64_t ny = (ne11 + nrows - 1)/nrows;
-                                    [encoder dispatchThreadgroups:MTLSizeMake(ne01, ny, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                                    const int64_t ny = (ne11 + nrows - 1)/nrows;
+                                    [encoder dispatchThreadgroups:MTLSizeMake(ne01, ny, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                                }
+                            }
+                        } break;
+                    case GGML_OP_MUL_MAT_ID:
+                        {
+                            //GGML_ASSERT(ne00 == ne10);
+                            //GGML_ASSERT(ne03 == ne13);
+
+                            GGML_ASSERT(src0t == GGML_TYPE_I32);
+
+                            const int n_as = ((int32_t *) dst->op_params)[1];
+
+                            // TODO: make this more general
+                            GGML_ASSERT(n_as <= 8);
+
+                            struct ggml_tensor * src2 = gf->nodes[i]->src[2];
+
+                            const int64_t  ne20 = src2 ? src2->ne[0] : 0;
+                            const int64_t  ne21 = src2 ? src2->ne[1] : 0;
+                            const int64_t  ne22 = src2 ? src2->ne[2] : 0;
+                            const int64_t  ne23 = src2 ? src2->ne[3] : 0; GGML_UNUSED(ne23);
+
+                            const uint64_t nb20 = src2 ? src2->nb[0] : 0; GGML_UNUSED(nb20);
+                            const uint64_t nb21 = src2 ? src2->nb[1] : 0;
+                            const uint64_t nb22 = src2 ? src2->nb[2] : 0;
+                            const uint64_t nb23 = src2 ? src2->nb[3] : 0; GGML_UNUSED(nb23);
+
+                            const enum ggml_type src2t = src2 ? src2->type : GGML_TYPE_COUNT; GGML_UNUSED(src2t);
+
+                            GGML_ASSERT(!ggml_is_transposed(src2));
+                            GGML_ASSERT(!ggml_is_transposed(src1));
+
+                            GGML_ASSERT(ne20 % 32 == 0);
+                            // !!!!!!!!! TODO: this assert is probably required but not sure!
+                            //GGML_ASSERT(ne20 >= 64);
+                            GGML_ASSERT(src1t == GGML_TYPE_F32);
+
+                            const uint r2 = ne12/ne22;
+                            const uint r3 = ne13/ne23;
+
+                            // find the break-even point where the matrix-matrix kernel becomes more efficient compared
+                            // to the matrix-vector kernel
+                            int ne11_mm_min = 1;
+
+                            const int idx = ((int32_t *) dst->op_params)[0];
+
+                            // batch size
+                            GGML_ASSERT(ne01 == ne11);
+
+                            const int64_t _ne1 = 1; // kernel_mul_mm_impl needs a reference in constant memory
+
+                            // for now the matrix-matrix multiplication kernel only works on A14+/M1+ SoCs
+                            // AMD GPU and older A-chips will reuse matrix-vector multiplication kernel
+                            // !!!
+                            // TODO: for now, always use mat-vec kernels until we figure out how to improve the
+                            //       indirect matrix multiplication
+                            // !!!
+                            if ([ctx->device supportsFamily:MTLGPUFamilyApple7] && _ne1 > ne11_mm_min) {
+                                switch (src2->type) {
+                                    case GGML_TYPE_F32:  [encoder setComputePipelineState:ctx->pipeline_mul_mm_id_f32_f32];  break;
+                                    case GGML_TYPE_F16:  [encoder setComputePipelineState:ctx->pipeline_mul_mm_id_f16_f32];  break;
+                                    case GGML_TYPE_Q4_0: [encoder setComputePipelineState:ctx->pipeline_mul_mm_id_q4_0_f32]; break;
+                                    case GGML_TYPE_Q4_1: [encoder setComputePipelineState:ctx->pipeline_mul_mm_id_q4_1_f32]; break;
+                                    case GGML_TYPE_Q5_0: [encoder setComputePipelineState:ctx->pipeline_mul_mm_id_q5_0_f32]; break;
+                                    case GGML_TYPE_Q5_1: [encoder setComputePipelineState:ctx->pipeline_mul_mm_id_q5_1_f32]; break;
+                                    case GGML_TYPE_Q8_0: [encoder setComputePipelineState:ctx->pipeline_mul_mm_id_q8_0_f32]; break;
+                                    case GGML_TYPE_Q2_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_id_q2_K_f32]; break;
+                                    case GGML_TYPE_Q3_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_id_q3_K_f32]; break;
+                                    case GGML_TYPE_Q4_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_id_q4_K_f32]; break;
+                                    case GGML_TYPE_Q5_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_id_q5_K_f32]; break;
+                                    case GGML_TYPE_Q6_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_id_q6_K_f32]; break;
+                                    default: GGML_ASSERT(false && "MUL_MAT_ID not implemented");
+                                }
+                                [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:&nb01    length:sizeof(nb01) atIndex:3];
+                                [encoder setBytes:&ne20    length:sizeof(ne20) atIndex:4];
+                                [encoder setBytes:&ne22    length:sizeof(ne22) atIndex:5];
+                                [encoder setBytes:&nb21    length:sizeof(nb21) atIndex:6];
+                                [encoder setBytes:&nb22    length:sizeof(nb22) atIndex:7];
+                                [encoder setBytes:&ne12    length:sizeof(ne12) atIndex:8];
+                                [encoder setBytes:&ne13    length:sizeof(ne13) atIndex:9];
+                                [encoder setBytes:&nb10    length:sizeof(nb10) atIndex:10];
+                                [encoder setBytes:&nb11    length:sizeof(nb11) atIndex:11];
+                                [encoder setBytes:&nb12    length:sizeof(nb12) atIndex:12];
+                                [encoder setBytes:&ne0     length:sizeof(ne0)  atIndex:13];
+                                [encoder setBytes:&_ne1    length:sizeof(_ne1) atIndex:14];
+                                [encoder setBytes:&nb1     length:sizeof(nb1)  atIndex:15];
+                                [encoder setBytes:&r2      length:sizeof(r2)   atIndex:16];
+                                [encoder setBytes:&r3      length:sizeof(r3)   atIndex:17];
+                                [encoder setBytes:&idx     length:sizeof(idx)  atIndex:18];
+                                // TODO: how to make this an array? read Metal docs
+                                for (int j = 0; j < n_as; ++j) {
+                                    struct ggml_tensor * src_cur = dst->src[2 + j];
+
+                                    size_t offs_src_cur = 0;
+                                    id<MTLBuffer> id_src_cur = ggml_metal_get_buffer(ctx, src_cur, &offs_src_cur);
+
+                                    [encoder setBuffer:id_src_cur offset:offs_src_cur atIndex:19 + j];
+                                }
+
+                                [encoder setThreadgroupMemoryLength:8192 atIndex:0];
+
+                                // TODO: processing one row at a time (ne11 -> 1) is not efficient
+                                [encoder dispatchThreadgroups:MTLSizeMake( (_ne1 + 31)/32, (ne21 + 63)/64, ne01*ne12*ne13) threadsPerThreadgroup:MTLSizeMake(128, 1, 1)];
+                            } else {
+                                int nth0 = 32;
+                                int nth1 = 1;
+                                int nrows = 1;
+                                //printf("vector: ne00 = %6d, ne01 = %6d, ne02 = %6d, ne11 = %6d, ne12 = %6d\n", ne00, ne01, ne02, ne11, ne12);
+
+                                // use custom matrix x vector kernel
+                                switch (src2t) {
+                                    case GGML_TYPE_F32:
+                                        {
+                                            GGML_ASSERT(src1t == GGML_TYPE_F32);
+                                            [encoder setComputePipelineState:ctx->pipeline_mul_mv_id_f32_f32];
+                                        } break;
+                                    case GGML_TYPE_F16:
+                                        {
+                                            GGML_ASSERT(src1t == GGML_TYPE_F32);
+                                            nth0 = 32;
+                                            nth1 = 1;
+                                            [encoder setComputePipelineState:ctx->pipeline_mul_mv_id_f16_f32];
+                                        } break;
+                                    case GGML_TYPE_Q4_0:
+                                        {
+                                            nth0 = 8;
+                                            nth1 = 8;
+                                            [encoder setComputePipelineState:ctx->pipeline_mul_mv_id_q4_0_f32];
+                                        } break;
+                                    case GGML_TYPE_Q4_1:
+                                        {
+                                            nth0 = 8;
+                                            nth1 = 8;
+                                            [encoder setComputePipelineState:ctx->pipeline_mul_mv_id_q4_1_f32];
+                                        } break;
+                                    case GGML_TYPE_Q5_0:
+                                        {
+                                            nth0 = 8;
+                                            nth1 = 8;
+                                            [encoder setComputePipelineState:ctx->pipeline_mul_mv_id_q5_0_f32];
+                                        } break;
+                                    case GGML_TYPE_Q5_1:
+                                        {
+                                            nth0 = 8;
+                                            nth1 = 8;
+                                            [encoder setComputePipelineState:ctx->pipeline_mul_mv_id_q5_1_f32];
+                                        } break;
+                                    case GGML_TYPE_Q8_0:
+                                        {
+                                            nth0 = 8;
+                                            nth1 = 8;
+                                            [encoder setComputePipelineState:ctx->pipeline_mul_mv_id_q8_0_f32];
+                                        } break;
+                                    case GGML_TYPE_Q2_K:
+                                        {
+                                            nth0 = 2;
+                                            nth1 = 32;
+                                            [encoder setComputePipelineState:ctx->pipeline_mul_mv_id_q2_K_f32];
+                                        } break;
+                                    case GGML_TYPE_Q3_K:
+                                        {
+                                            nth0 = 2;
+                                            nth1 = 32;
+                                            [encoder setComputePipelineState:ctx->pipeline_mul_mv_id_q3_K_f32];
+                                        } break;
+                                    case GGML_TYPE_Q4_K:
+                                        {
+                                            nth0 = 4; //1;
+                                            nth1 = 8; //32;
+                                            [encoder setComputePipelineState:ctx->pipeline_mul_mv_id_q4_K_f32];
+                                        } break;
+                                    case GGML_TYPE_Q5_K:
+                                        {
+                                            nth0 = 2;
+                                            nth1 = 32;
+                                            [encoder setComputePipelineState:ctx->pipeline_mul_mv_id_q5_K_f32];
+                                        } break;
+                                    case GGML_TYPE_Q6_K:
+                                        {
+                                            nth0 = 2;
+                                            nth1 = 32;
+                                            [encoder setComputePipelineState:ctx->pipeline_mul_mv_id_q6_K_f32];
+                                        } break;
+                                    default:
+                                        {
+                                            GGML_METAL_LOG_ERROR("Asserting on type %d\n", (int)src0t);
+                                            GGML_ASSERT(false && "not implemented");
+                                        }
+                                };
+
+                                [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:&nb01 length:sizeof(nb01) atIndex:3];
+                                [encoder setBytes:&ne20 length:sizeof(ne20) atIndex:4];
+                                [encoder setBytes:&ne21 length:sizeof(ne21) atIndex:5];
+                                [encoder setBytes:&ne22 length:sizeof(ne22) atIndex:6];
+                                [encoder setBytes:&nb20 length:sizeof(nb20) atIndex:7];
+                                [encoder setBytes:&nb21 length:sizeof(nb21) atIndex:8];
+                                [encoder setBytes:&nb22 length:sizeof(nb22) atIndex:9];
+                                [encoder setBytes:&ne10 length:sizeof(ne10) atIndex:10];
+                                [encoder setBytes:&_ne1 length:sizeof(_ne1) atIndex:11];
+                                [encoder setBytes:&ne12 length:sizeof(ne12) atIndex:12];
+                                [encoder setBytes:&ne13 length:sizeof(ne13) atIndex:13];
+                                [encoder setBytes:&nb10 length:sizeof(nb10) atIndex:14];
+                                [encoder setBytes:&nb11 length:sizeof(nb11) atIndex:15];
+                                [encoder setBytes:&nb12 length:sizeof(nb12) atIndex:16];
+                                [encoder setBytes:&ne0  length:sizeof(ne0)  atIndex:17];
+                                [encoder setBytes:&_ne1 length:sizeof(_ne1) atIndex:18];
+                                [encoder setBytes:&nb1  length:sizeof(nb1)  atIndex:19];
+                                [encoder setBytes:&r2   length:sizeof(r2)   atIndex:20];
+                                [encoder setBytes:&r3   length:sizeof(r3)   atIndex:21];
+                                [encoder setBytes:&idx  length:sizeof(idx)  atIndex:22];
+                                // TODO: how to make this an array? read Metal docs
+                                for (int j = 0; j < n_as; ++j) {
+                                    struct ggml_tensor * src_cur = dst->src[2 + j];
+
+                                    size_t offs_src_cur = 0;
+                                    id<MTLBuffer> id_src_cur = ggml_metal_get_buffer(ctx, src_cur, &offs_src_cur);
+
+                                    [encoder setBuffer:id_src_cur offset:offs_src_cur atIndex:23 + j];
+                                }
+
+                                if (src2t == GGML_TYPE_Q4_0 || src2t == GGML_TYPE_Q4_1 ||
+                                    src2t == GGML_TYPE_Q5_0 || src2t == GGML_TYPE_Q5_1 || src2t == GGML_TYPE_Q8_0 ||
+                                    src2t == GGML_TYPE_Q2_K) { // || src2t == GGML_TYPE_Q4_K) {
+                                    [encoder dispatchThreadgroups:MTLSizeMake((ne21 + 7)/8, _ne1, ne01*ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                                }
+                                else if (src2t == GGML_TYPE_Q4_K) {
+                                    [encoder dispatchThreadgroups:MTLSizeMake((ne21 + 3)/4, _ne1, ne01*ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                                }
+                                else if (src2t == GGML_TYPE_Q3_K) {
+#ifdef GGML_QKK_64
+                                    [encoder dispatchThreadgroups:MTLSizeMake((ne21 + 1)/2, _ne1, ne01*ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+#else
+                                    [encoder dispatchThreadgroups:MTLSizeMake((ne21 + 3)/4, _ne1, ne01*ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+#endif
+                                }
+                                else if (src2t == GGML_TYPE_Q5_K) {
+                                    [encoder dispatchThreadgroups:MTLSizeMake((ne21 + 3)/4, _ne1, ne01*ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                                }
+                                else if (src2t == GGML_TYPE_Q6_K) {
+                                    [encoder dispatchThreadgroups:MTLSizeMake((ne21 + 1)/2, _ne1, ne01*ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                                } else {
+                                    const int64_t ny = (_ne1 + nrows - 1)/nrows;
+                                    [encoder dispatchThreadgroups:MTLSizeMake(ne21, ny, ne01*ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
                                 }
                             }
                         } break;
@@ -1340,16 +1962,19 @@ void ggml_metal_graph_compute(
                                 default: GGML_ASSERT(false && "not implemented");
                             }
 
-                            [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 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( int64_t) atIndex:3];
                             [encoder setBytes:&nb01 length:sizeof(uint64_t) atIndex:4];
-                            [encoder setBytes:&nb1  length:sizeof(uint64_t) atIndex:5];
+                            [encoder setBytes:&nb02 length:sizeof(uint64_t) atIndex:5];
+                            [encoder setBytes:&ne10 length:sizeof( int64_t) atIndex:6];
+                            [encoder setBytes:&nb10 length:sizeof( int64_t) atIndex:7];
+                            [encoder setBytes:&nb11 length:sizeof( int64_t) atIndex:8];
+                            [encoder setBytes:&nb1  length:sizeof(uint64_t) atIndex:9];
+                            [encoder setBytes:&nb2  length:sizeof(uint64_t) atIndex:10];
 
-                            const int64_t n = ggml_nelements(src1);
-
-                            [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                            [encoder dispatchThreadgroups:MTLSizeMake(ne10, ne11, 1) threadsPerThreadgroup:MTLSizeMake(32, 1, 1)];
                         } break;
                     case GGML_OP_RMS_NORM:
                         {
@@ -1376,6 +2001,38 @@ 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;
@@ -1545,18 +2202,123 @@ 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);
+                            GGML_ASSERT( dst->type == GGML_TYPE_I32);
+
+                            const int nrows = ggml_nrows(src0);
+
+                            enum ggml_sort_order order = (enum ggml_sort_order) dst->op_params[0];
+
+                            switch (order) {
+                                case GGML_SORT_ASC:  [encoder setComputePipelineState:ctx->pipeline_argsort_f32_i32_asc];  break;
+                                case GGML_SORT_DESC: [encoder setComputePipelineState:ctx->pipeline_argsort_f32_i32_desc]; break;
+                                default: GGML_ASSERT(false);
+                            };
+
+                            [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 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:
                         {
-                            const int nth = MIN(1024, ne00);
+                            GGML_ASSERT(ne00 % ggml_blck_size(src0->type) == 0);
+
+                            int nth = MIN(1024, ne00/ggml_blck_size(src0->type));
 
                             switch (src0t) {
                                 case GGML_TYPE_F32:
                                     {
+                                        GGML_ASSERT(ne0 % ggml_blck_size(dst->type) == 0);
+
                                         switch (dstt) {
-                                            case GGML_TYPE_F16: [encoder setComputePipelineState:ctx->pipeline_cpy_f32_f16]; break;
-                                            case GGML_TYPE_F32: [encoder setComputePipelineState:ctx->pipeline_cpy_f32_f32]; break;
+                                            case GGML_TYPE_F16:  [encoder setComputePipelineState:ctx->pipeline_cpy_f32_f16];  break;
+                                            case GGML_TYPE_F32:  [encoder setComputePipelineState:ctx->pipeline_cpy_f32_f32];  break;
+                                            case GGML_TYPE_Q8_0: [encoder setComputePipelineState:ctx->pipeline_cpy_f32_q8_0]; break;
+                                            case GGML_TYPE_Q4_0: [encoder setComputePipelineState:ctx->pipeline_cpy_f32_q4_0]; break;
+                                            case GGML_TYPE_Q4_1: [encoder setComputePipelineState:ctx->pipeline_cpy_f32_q4_1]; break;
+                                            //case GGML_TYPE_Q5_0: [encoder setComputePipelineState:ctx->pipeline_cpy_f32_q5_0]; break;
+                                            //case GGML_TYPE_Q5_1: [encoder setComputePipelineState:ctx->pipeline_cpy_f32_q5_1]; break;
                                             default: GGML_ASSERT(false && "not implemented");
                                         };
                                     } break;
@@ -1564,7 +2326,7 @@ void ggml_metal_graph_compute(
                                     {
                                         switch (dstt) {
                                             case GGML_TYPE_F16: [encoder setComputePipelineState:ctx->pipeline_cpy_f16_f16]; break;
-                                            case GGML_TYPE_F32: GGML_ASSERT(false && "cpy_f16_f32 not implemented"); break;
+                                            case GGML_TYPE_F32: [encoder setComputePipelineState:ctx->pipeline_cpy_f16_f32]; break;
                                             default: GGML_ASSERT(false && "not implemented");
                                         };
                                     } break;
@@ -1631,6 +2393,260 @@ void ggml_metal_graph_compute(
 
 // backend interface
 
+// default buffer
+static id<MTLDevice> g_backend_device = nil;
+static int g_backend_device_ref_count = 0;
+
+static id<MTLDevice> ggml_backend_metal_get_device(void) {
+    if (g_backend_device == nil) {
+        g_backend_device = MTLCreateSystemDefaultDevice();
+    }
+
+    g_backend_device_ref_count++;
+
+    return g_backend_device;
+}
+
+static void ggml_backend_metal_free_device(void) {
+    assert(g_backend_device_ref_count > 0);
+
+    g_backend_device_ref_count--;
+
+    if (g_backend_device_ref_count == 0) {
+        [g_backend_device release];
+        g_backend_device = nil;
+    }
+}
+
+static void * ggml_backend_metal_buffer_get_base(ggml_backend_buffer_t buffer) {
+    struct ggml_backend_metal_buffer_context * ctx = (struct ggml_backend_metal_buffer_context *)buffer->context;
+
+    return ctx->all_data;
+}
+
+static void ggml_backend_metal_buffer_free_buffer(ggml_backend_buffer_t buffer) {
+    struct ggml_backend_metal_buffer_context * ctx = (struct ggml_backend_metal_buffer_context *)buffer->context;
+
+    for (int i = 0; i < ctx->n_buffers; i++) {
+        [ctx->buffers[i].metal release];
+    }
+    ggml_backend_metal_free_device();
+
+    if (ctx->owned) {
+        free(ctx->all_data);
+    }
+
+    free(ctx);
+}
+
+static void ggml_backend_metal_buffer_set_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
+    memcpy((char *)tensor->data + offset, data, size);
+
+    UNUSED(buffer);
+}
+
+static void ggml_backend_metal_buffer_get_tensor(ggml_backend_buffer_t buffer, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) {
+    memcpy(data, (const char *)tensor->data + offset, size);
+
+    UNUSED(buffer);
+}
+
+static void ggml_backend_metal_buffer_cpy_tensor_from(ggml_backend_buffer_t buffer, struct ggml_tensor * src, struct ggml_tensor * dst) {
+    ggml_backend_tensor_get(src, dst->data, 0, ggml_nbytes(src));
+
+    UNUSED(buffer);
+}
+
+static void ggml_backend_metal_buffer_cpy_tensor_to(ggml_backend_buffer_t buffer, struct ggml_tensor * src, struct ggml_tensor * dst) {
+    ggml_backend_tensor_set(dst, src->data, 0, ggml_nbytes(src));
+
+    UNUSED(buffer);
+}
+
+static void ggml_backend_metal_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
+    struct ggml_backend_metal_buffer_context * ctx = (struct ggml_backend_metal_buffer_context *)buffer->context;
+
+    memset(ctx->all_data, value, ctx->all_size);
+}
+
+static struct ggml_backend_buffer_i ggml_backend_metal_buffer_i = {
+    /* .free_buffer     = */ ggml_backend_metal_buffer_free_buffer,
+    /* .get_base        = */ ggml_backend_metal_buffer_get_base,
+    /* .init_tensor     = */ NULL,
+    /* .set_tensor      = */ ggml_backend_metal_buffer_set_tensor,
+    /* .get_tensor      = */ ggml_backend_metal_buffer_get_tensor,
+    /* .cpy_tensor_from = */ ggml_backend_metal_buffer_cpy_tensor_from,
+    /* .cpy_tensor_to   = */ ggml_backend_metal_buffer_cpy_tensor_to,
+    /* .clear           = */ ggml_backend_metal_buffer_clear,
+};
+
+// default buffer type
+
+static ggml_backend_buffer_t ggml_backend_metal_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
+    struct ggml_backend_metal_buffer_context * ctx = malloc(sizeof(struct ggml_backend_metal_buffer_context));
+
+    const size_t size_page = sysconf(_SC_PAGESIZE);
+
+    size_t size_aligned = size;
+    if ((size_aligned % size_page) != 0) {
+        size_aligned += (size_page - (size_aligned % size_page));
+    }
+
+    id<MTLDevice> device = ggml_backend_metal_get_device();
+
+    ctx->all_data = ggml_metal_host_malloc(size_aligned);
+    ctx->all_size = size_aligned;
+    ctx->owned = true;
+    ctx->n_buffers = 1;
+
+    ctx->buffers[0].data = ctx->all_data;
+    ctx->buffers[0].size = size;
+    ctx->buffers[0].metal = [device newBufferWithBytesNoCopy:ctx->all_data
+                    length:size_aligned
+                    options:MTLResourceStorageModeShared
+                    deallocator:nil];
+
+    if (ctx->buffers[0].metal == nil) {
+        GGML_METAL_LOG_ERROR("%s: error: failed to allocate buffer, size = %8.2f MiB\n", __func__, size_aligned / 1024.0 / 1024.0);
+        free(ctx);
+        ggml_backend_metal_free_device();
+        return NULL;
+    }
+
+    GGML_METAL_LOG_INFO("%s: allocated buffer, size = %8.2f MiB", __func__, size_aligned / 1024.0 / 1024.0);
+
+
+#if TARGET_OS_OSX
+    GGML_METAL_LOG_INFO(", (%8.2f / %8.2f)",
+            device.currentAllocatedSize / 1024.0 / 1024.0,
+            device.recommendedMaxWorkingSetSize / 1024.0 / 1024.0);
+
+    if (device.currentAllocatedSize > device.recommendedMaxWorkingSetSize) {
+        GGML_METAL_LOG_WARN("%s: warning: current allocated size is greater than the recommended max working set size\n", __func__);
+    } else {
+        GGML_METAL_LOG_INFO("\n");
+    }
+#else
+    GGML_METAL_LOG_INFO(", (%8.2f)\n", device.currentAllocatedSize / 1024.0 / 1024.0);
+#endif
+
+
+    return ggml_backend_buffer_init(buft, ggml_backend_metal_buffer_i, ctx, size);
+}
+
+static size_t ggml_backend_metal_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
+    return 32;
+    UNUSED(buft);
+}
+
+static bool ggml_backend_metal_buffer_type_supports_backend(ggml_backend_buffer_type_t buft, ggml_backend_t backend) {
+    return ggml_backend_is_metal(backend) || ggml_backend_is_cpu(backend);
+
+    UNUSED(buft);
+}
+
+static bool ggml_backend_metal_buffer_type_is_host(ggml_backend_buffer_type_t buft) {
+    return true;
+
+    UNUSED(buft);
+}
+
+ggml_backend_buffer_type_t ggml_backend_metal_buffer_type(void) {
+    static struct ggml_backend_buffer_type ggml_backend_buffer_type_metal = {
+        /* .iface = */ {
+            /* .alloc_buffer     = */ ggml_backend_metal_buffer_type_alloc_buffer,
+            /* .get_alignment    = */ ggml_backend_metal_buffer_type_get_alignment,
+            /* .get_alloc_size   = */ NULL, // defaults to ggml_nbytes
+            /* .supports_backend = */ ggml_backend_metal_buffer_type_supports_backend,
+            /* .is_host          = */ ggml_backend_metal_buffer_type_is_host,
+        },
+        /* .context = */ NULL,
+    };
+
+    return &ggml_backend_buffer_type_metal;
+}
+
+// buffer from ptr
+
+ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void * data, size_t size, size_t max_size) {
+    struct ggml_backend_metal_buffer_context * ctx = malloc(sizeof(struct ggml_backend_metal_buffer_context));
+
+    ctx->all_data = data;
+    ctx->all_size = size;
+    ctx->owned = false;
+    ctx->n_buffers = 0;
+
+    const size_t size_page = sysconf(_SC_PAGESIZE);
+    size_t size_aligned = size;
+    if ((size_aligned % size_page) != 0) {
+        size_aligned += (size_page - (size_aligned % size_page));
+    }
+
+    id<MTLDevice> device = ggml_backend_metal_get_device();
+
+    // the buffer fits into the max buffer size allowed by the device
+    if (size_aligned <= device.maxBufferLength) {
+        ctx->buffers[ctx->n_buffers].data = data;
+        ctx->buffers[ctx->n_buffers].size = size;
+
+        ctx->buffers[ctx->n_buffers].metal = [device newBufferWithBytesNoCopy:data length:size_aligned options:MTLResourceStorageModeShared deallocator:nil];
+
+        if (ctx->buffers[ctx->n_buffers].metal == nil) {
+            GGML_METAL_LOG_ERROR("%s: error: failed to allocate buffer, size = %8.2f MiB\n", __func__, size_aligned / 1024.0 / 1024.0);
+            return false;
+        }
+
+        GGML_METAL_LOG_INFO("%s: allocated buffer, size = %8.2f MiB", __func__, size_aligned / 1024.0 / 1024.0);
+
+        ++ctx->n_buffers;
+    } else {
+        // this overlap between the views will guarantee that the tensor with the maximum size will fully fit into
+        // one of the views
+        const size_t size_ovlp = ((max_size + size_page - 1) / size_page + 1) * size_page; // round-up 2 pages just in case
+        const size_t size_step = device.maxBufferLength - size_ovlp;
+        const size_t size_view = device.maxBufferLength;
+
+        for (size_t i = 0; i < size; i += size_step) {
+            const size_t size_step_aligned = (i + size_view <= size) ? size_view : (size_aligned - i);
+
+            ctx->buffers[ctx->n_buffers].data = (void *) ((uint8_t *) data + i);
+            ctx->buffers[ctx->n_buffers].size = size_step_aligned;
+
+            ctx->buffers[ctx->n_buffers].metal = [device newBufferWithBytesNoCopy:(void *) ((uint8_t *) data + i) length:size_step_aligned options:MTLResourceStorageModeShared deallocator:nil];
+
+            if (ctx->buffers[ctx->n_buffers].metal == nil) {
+                GGML_METAL_LOG_ERROR("%s: error: failed to allocate buffer, size = %8.2f MiB\n", __func__, size_step_aligned / 1024.0 / 1024.0);
+                return false;
+            }
+
+            GGML_METAL_LOG_INFO("%s: allocated buffer, size = %8.2f MiB, offs = %12ld", __func__, size_step_aligned / 1024.0 / 1024.0, i);
+            if (i + size_step < size) {
+                GGML_METAL_LOG_INFO("\n");
+            }
+
+            ++ctx->n_buffers;
+        }
+    }
+
+#if TARGET_OS_OSX
+    GGML_METAL_LOG_INFO(", (%8.2f / %8.2f)",
+            device.currentAllocatedSize / 1024.0 / 1024.0,
+            device.recommendedMaxWorkingSetSize / 1024.0 / 1024.0);
+
+    if (device.currentAllocatedSize > device.recommendedMaxWorkingSetSize) {
+        GGML_METAL_LOG_WARN("%s: warning: current allocated size is greater than the recommended max working set size\n", __func__);
+    } else {
+        GGML_METAL_LOG_INFO("\n");
+    }
+#else
+    GGML_METAL_LOG_INFO(", (%8.2f)\n", device.currentAllocatedSize / 1024.0 / 1024.0);
+#endif
+
+    return ggml_backend_buffer_init(ggml_backend_metal_buffer_type(), ggml_backend_metal_buffer_i, ctx, size);
+}
+
+// backend
+
 static const char * ggml_backend_metal_name(ggml_backend_t backend) {
     return "Metal";
 
@@ -1643,69 +2659,8 @@ static void ggml_backend_metal_free(ggml_backend_t backend) {
     free(backend);
 }
 
-static void * ggml_backend_metal_buffer_get_base(ggml_backend_buffer_t buffer) {
-    return (void *)buffer->context;
-}
-
-static void ggml_backend_metal_buffer_free_buffer(ggml_backend_buffer_t buffer) {
-    free(buffer->context);
-    UNUSED(buffer);
-}
-
-static struct ggml_backend_buffer_i metal_backend_buffer_i = {
-    /* .free_buffer    = */ ggml_backend_metal_buffer_free_buffer,
-    /* .get_base       = */ ggml_backend_metal_buffer_get_base,
-    /* .get_alloc_size = */ NULL, // defaults to ggml_nbytes
-    /* .init_tensor    = */ NULL, // no initialization required
-    /* .free_tensor    = */ NULL, // no cleanup required
-};
-
-static ggml_backend_buffer_t ggml_backend_metal_alloc_buffer(ggml_backend_t backend, size_t size) {
-    struct ggml_metal_context * ctx = (struct ggml_metal_context *)backend->context;
-
-    void * data = ggml_metal_host_malloc(size);
-
-    // TODO: set proper name of the buffers
-    ggml_metal_add_buffer(ctx, "backend", data, size, 0);
-
-    return ggml_backend_buffer_init(backend, metal_backend_buffer_i, data, size);
-}
-
-static size_t ggml_backend_metal_get_alignment(ggml_backend_t backend) {
-    return 32;
-    UNUSED(backend);
-}
-
-static void ggml_backend_metal_set_tensor_async(ggml_backend_t backend, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
-    GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor write out of bounds");
-    GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
-
-    memcpy((char *)tensor->data + offset, data, size);
-
-    UNUSED(backend);
-}
-
-static void ggml_backend_metal_get_tensor_async(ggml_backend_t backend, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) {
-    GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor read out of bounds");
-    GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
-
-    memcpy(data, (const char *)tensor->data + offset, size);
-
-    UNUSED(backend);
-}
-
-static void ggml_backend_metal_synchronize(ggml_backend_t backend) {
-    UNUSED(backend);
-}
-
-static void ggml_backend_metal_cpy_tensor_from(ggml_backend_t backend, struct ggml_tensor * src, struct ggml_tensor * dst) {
-    ggml_backend_tensor_get(src, dst->data, 0, ggml_nbytes(src));
-
-    UNUSED(backend);
-}
-
-static void ggml_backend_metal_cpy_tensor_to(ggml_backend_t backend, struct ggml_tensor * src, struct ggml_tensor * dst) {
-    ggml_backend_tensor_set_async(dst, src->data, 0, ggml_nbytes(src));
+static ggml_backend_buffer_type_t ggml_backend_metal_get_default_buffer_type(ggml_backend_t backend) {
+    return ggml_backend_metal_buffer_type();
 
     UNUSED(backend);
 }
@@ -1717,32 +2672,33 @@ static void ggml_backend_metal_graph_compute(ggml_backend_t backend, struct ggml
 }
 
 static bool ggml_backend_metal_supports_op(ggml_backend_t backend, const struct ggml_tensor * op) {
-    return true;
+    return ggml_metal_supports_op(op);
+
     UNUSED(backend);
-    UNUSED(op);
 }
 
 static struct ggml_backend_i metal_backend_i = {
-    /* .get_name            = */ ggml_backend_metal_name,
-    /* .free                = */ ggml_backend_metal_free,
-    /* .alloc_buffer        = */ ggml_backend_metal_alloc_buffer,
-    /* .get_alignment       = */ ggml_backend_metal_get_alignment,
-    /* .set_tensor_async    = */ ggml_backend_metal_set_tensor_async,
-    /* .get_tensor_async    = */ ggml_backend_metal_get_tensor_async,
-    /* .synchronize         = */ ggml_backend_metal_synchronize,
-    /* .cpy_tensor_from     = */ ggml_backend_metal_cpy_tensor_from,
-    /* .cpy_tensor_to       = */ ggml_backend_metal_cpy_tensor_to,
-    /* .graph_plan_create   = */ NULL, // the metal implementation does not require creating graph plans atm
-    /* .graph_plan_free     = */ NULL,
-    /* .graph_plan_compute  = */ NULL,
-    /* .graph_compute       = */ ggml_backend_metal_graph_compute,
-    /* .supports_op         = */ ggml_backend_metal_supports_op,
+    /* .get_name                = */ ggml_backend_metal_name,
+    /* .free                    = */ ggml_backend_metal_free,
+    /* .get_default_buffer_type = */ ggml_backend_metal_get_default_buffer_type,
+    /* .set_tensor_async        = */ NULL,
+    /* .get_tensor_async        = */ NULL,
+    /* .cpy_tensor_from_async   = */ NULL,
+    /* .cpy_tensor_to_async     = */ NULL,
+    /* .synchronize             = */ NULL,
+    /* .graph_plan_create       = */ NULL,
+    /* .graph_plan_free         = */ NULL,
+    /* .graph_plan_compute      = */ NULL,
+    /* .graph_compute           = */ ggml_backend_metal_graph_compute,
+    /* .supports_op             = */ ggml_backend_metal_supports_op,
 };
 
 ggml_backend_t ggml_backend_metal_init(void) {
-    struct ggml_metal_context * ctx = malloc(sizeof(struct ggml_metal_context));
+    struct ggml_metal_context * ctx = ggml_metal_init(GGML_DEFAULT_N_THREADS);
 
-    ctx = ggml_metal_init(GGML_DEFAULT_N_THREADS);
+    if (ctx == NULL) {
+        return NULL;
+    }
 
     ggml_backend_t metal_backend = malloc(sizeof(struct ggml_backend));
 
@@ -1759,7 +2715,26 @@ bool ggml_backend_is_metal(ggml_backend_t backend) {
 }
 
 void ggml_backend_metal_set_n_cb(ggml_backend_t backend, int n_cb) {
+    GGML_ASSERT(ggml_backend_is_metal(backend));
+
     struct ggml_metal_context * ctx = (struct ggml_metal_context *)backend->context;
 
     ggml_metal_set_n_cb(ctx, n_cb);
 }
+
+bool ggml_backend_metal_supports_family(ggml_backend_t backend, int family) {
+    GGML_ASSERT(ggml_backend_is_metal(backend));
+
+    struct ggml_metal_context * ctx = (struct ggml_metal_context *)backend->context;
+
+    return [ctx->device supportsFamily:(MTLGPUFamilyApple1 + family - 1)];
+}
+
+ggml_backend_t ggml_backend_reg_metal_init(const char * params, void * user_data); // silence warning
+
+ggml_backend_t ggml_backend_reg_metal_init(const char * params, void * user_data) {
+    return ggml_backend_metal_init();
+
+    GGML_UNUSED(params);
+    GGML_UNUSED(user_data);
+}
diff --git a/ggml-metal.metal b/ggml-metal.metal
index 9a79f815f..d5b54e112 100644
--- a/ggml-metal.metal
+++ b/ggml-metal.metal
@@ -3,6 +3,8 @@
 using namespace metal;
 
 #define MAX(x, y) ((x) > (y) ? (x) : (y))
+#define MIN(x, y) ((x) < (y) ? (x) : (y))
+#define SWAP(x, y) { auto tmp = (x); (x) = (y); (y) = tmp; }
 
 #define QK4_0 32
 #define QR4_0 2
@@ -41,10 +43,65 @@ typedef struct {
 
 #define N_SIMDWIDTH 32 // assuming SIMD group size is 32
 
-// general-purpose kernel for addition of two tensors
-// pros: works for non-contiguous tensors, supports broadcast across dims 1, 2 and 3
+enum ggml_sort_order {
+    GGML_SORT_ASC,
+    GGML_SORT_DESC,
+};
+
+// general-purpose kernel for addition, multiplication and division of two tensors
+// pros: works for non-contiguous tensors, supports broadcast across all dims
 // cons: not very efficient
 kernel void kernel_add(
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        constant  int64_t & ne00,
+        constant  int64_t & ne01,
+        constant  int64_t & ne02,
+        constant  int64_t & ne03,
+        constant  int64_t & nb00,
+        constant  int64_t & nb01,
+        constant  int64_t & nb02,
+        constant  int64_t & nb03,
+        constant  int64_t & ne10,
+        constant  int64_t & ne11,
+        constant  int64_t & ne12,
+        constant  int64_t & ne13,
+        constant  int64_t & nb10,
+        constant  int64_t & nb11,
+        constant  int64_t & nb12,
+        constant  int64_t & nb13,
+        constant  int64_t & ne0,
+        constant  int64_t & ne1,
+        constant  int64_t & ne2,
+        constant  int64_t & ne3,
+        constant  int64_t & nb0,
+        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]]) {
+    const int64_t i03 = tgpig.z;
+    const int64_t i02 = tgpig.y;
+    const int64_t i01 = tgpig.x;
+
+    const int64_t i13 = i03 % ne13;
+    const int64_t i12 = i02 % ne12;
+    const int64_t i11 = i01 % ne11;
+
+    device const char * src0_ptr = src0 + i03*nb03 + i02*nb02 + i01*nb01 + offs;
+    device const char * src1_ptr = src1 + i13*nb13 + i12*nb12 + i11*nb11;
+    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;
+        *((device float *)(dst_ptr + i0*nb0)) = *((device float *)(src0_ptr + i0*nb00)) + *((device float *)(src1_ptr + i10*nb10));
+    }
+}
+
+kernel void kernel_mul(
         device const char * src0,
         device const char * src1,
         device       char * dst,
@@ -83,16 +140,62 @@ 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 + 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;
+    device const char * src0_ptr = src0 + i03*nb03 + i02*nb02 + i01*nb01;
+    device const char * src1_ptr = src1 + i13*nb13 + i12*nb12 + i11*nb11;
+    device       char * dst_ptr  = dst  + i03*nb3  + i02*nb2  + i01*nb1;
 
     for (int i0 = tpitg.x; i0 < ne0; i0 += ntg.x) {
-        ((device float *)dst_ptr)[0] = ((device float *)src0_ptr)[0] + ((device float *)src1_ptr)[0];
+        const int i10 = i0 % ne10;
+        *((device float *)(dst_ptr + i0*nb0)) = *((device float *)(src0_ptr + i0*nb00)) * *((device float *)(src1_ptr + i10*nb10));
+    }
+}
 
-        src0_ptr += ntg.x*nb00;
-        src1_ptr += ntg.x*nb10;
-        dst_ptr  += ntg.x*nb0;
+kernel void kernel_div(
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        constant  int64_t & ne00,
+        constant  int64_t & ne01,
+        constant  int64_t & ne02,
+        constant  int64_t & ne03,
+        constant  int64_t & nb00,
+        constant  int64_t & nb01,
+        constant  int64_t & nb02,
+        constant  int64_t & nb03,
+        constant  int64_t & ne10,
+        constant  int64_t & ne11,
+        constant  int64_t & ne12,
+        constant  int64_t & ne13,
+        constant  int64_t & nb10,
+        constant  int64_t & nb11,
+        constant  int64_t & nb12,
+        constant  int64_t & nb13,
+        constant  int64_t & ne0,
+        constant  int64_t & ne1,
+        constant  int64_t & ne2,
+        constant  int64_t & ne3,
+        constant  int64_t & nb0,
+        constant  int64_t & nb1,
+        constant  int64_t & nb2,
+        constant  int64_t & nb3,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3   ntg[[threads_per_threadgroup]]) {
+    const int64_t i03 = tgpig.z;
+    const int64_t i02 = tgpig.y;
+    const int64_t i01 = tgpig.x;
+
+    const int64_t i13 = i03 % ne13;
+    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 * src1_ptr = src1 + i13*nb13 + i12*nb12 + i11*nb11;
+    device       char * dst_ptr  = dst  + i03*nb3  + i02*nb2  + i01*nb1;
+
+    for (int i0 = tpitg.x; i0 < ne0; i0 += ntg.x) {
+        const int i10 = i0 % ne10;
+        *((device float *)(dst_ptr + i0*nb0)) = *((device float *)(src0_ptr + i0*nb00)) / *((device float *)(src1_ptr + i10*nb10));
     }
 }
 
@@ -102,30 +205,29 @@ 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];
 }
 
-kernel void kernel_mul(
-        device const float4 * src0,
-        device const float4 * src1,
-        device       float4 * dst,
-        uint tpig[[thread_position_in_grid]]) {
-    dst[tpig] = src0[tpig] * src1[tpig];
-}
-
-// assumption: src1 is a row
-// broadcast src1 into src0
 kernel void kernel_mul_row(
         device const float4 * src0,
         device const float4 * src1,
         device       float4 * dst,
-        constant    int64_t & nb,
+        constant    int64_t & nb  [[buffer(28)]],
         uint tpig[[thread_position_in_grid]]) {
     dst[tpig] = src0[tpig] * src1[tpig % nb];
 }
 
+kernel void kernel_div_row(
+        device const float4 * src0,
+        device const float4 * src1,
+        device       float4 * dst,
+        constant    int64_t & nb  [[buffer(28)]],
+        uint tpig[[thread_position_in_grid]]) {
+    dst[tpig] = src0[tpig] / src1[tpig % nb];
+}
+
 kernel void kernel_scale(
         device const float * src0,
         device       float * dst,
@@ -142,14 +244,6 @@ kernel void kernel_scale_4(
     dst[tpig] = src0[tpig] * scale;
 }
 
-kernel void kernel_silu(
-        device const float4 * src0,
-        device       float4 * dst,
-        uint tpig[[thread_position_in_grid]]) {
-    device const float4 & x = src0[tpig];
-    dst[tpig] = x / (1.0f + exp(-x));
-}
-
 kernel void kernel_relu(
         device const float * src0,
         device       float * dst,
@@ -157,15 +251,17 @@ kernel void kernel_relu(
     dst[tpig] = max(0.0f, src0[tpig]);
 }
 
-kernel void kernel_sqr(
+kernel void kernel_tanh(
         device const float * src0,
         device       float * dst,
         uint tpig[[thread_position_in_grid]]) {
-    dst[tpig] = src0[tpig] * src0[tpig];
+    device const float & x = src0[tpig];
+    dst[tpig] = precise::tanh(x);
 }
 
-constant float GELU_COEF_A    = 0.044715f;
-constant float SQRT_2_OVER_PI = 0.79788456080286535587989211986876f;
+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,
@@ -180,6 +276,78 @@ kernel void kernel_gelu(
     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,
+        uint tpig[[thread_position_in_grid]]) {
+    device const float4 & x = src0[tpig];
+    dst[tpig] = x / (1.0f + exp(-x));
+}
+
+kernel void kernel_sqr(
+        device const float * src0,
+        device       float * dst,
+        uint tpig[[thread_position_in_grid]]) {
+    dst[tpig] = src0[tpig] * src0[tpig];
+}
+
+kernel void kernel_sum_rows(
+        device const float * src0,
+        device       float * dst,
+        constant  int64_t & ne00,
+        constant  int64_t & ne01,
+        constant  int64_t & ne02,
+        constant  int64_t & ne03,
+        constant  int64_t & nb00,
+        constant  int64_t & nb01,
+        constant  int64_t & nb02,
+        constant  int64_t & nb03,
+        constant  int64_t & ne10,
+        constant  int64_t & ne11,
+        constant  int64_t & ne12,
+        constant  int64_t & ne13,
+        constant  int64_t & nb10,
+        constant  int64_t & nb11,
+        constant  int64_t & nb12,
+        constant  int64_t & nb13,
+        constant  int64_t & ne0,
+        constant  int64_t & ne1,
+        constant  int64_t & ne2,
+        constant  int64_t & ne3,
+        constant  int64_t & nb0,
+        constant  int64_t & nb1,
+        constant  int64_t & nb2,
+        constant  int64_t & nb3,
+        uint3 tpig[[thread_position_in_grid]]) {
+    int64_t i3 = tpig.z;
+    int64_t i2 = tpig.y;
+    int64_t i1 = tpig.x;
+
+    if (i3 >= ne03 || i2 >= ne02 || i1 >= ne01) {
+        return;
+    }
+
+    device const float * src_row = (device const float *) ((device const char *) src0 + i1*nb01 + i2*nb02 + i3*nb03);
+    device       float * dst_row = (device       float *) ((device       char *) dst  + i1*nb1  + i2*nb2  + i3*nb3);
+
+    float row_sum = 0;
+
+    for (int64_t i0 = 0; i0 < ne00; i0++) {
+        row_sum += src_row[i0];
+    }
+
+    dst_row[0] = row_sum;
+}
+
 kernel void kernel_soft_max(
         device const float * src0,
         device const float * src1,
@@ -198,9 +366,9 @@ kernel void kernel_soft_max(
     const int64_t i02 = (tgpig - i03*ne02*ne01) / ne01;
     const int64_t i01 = (tgpig - i03*ne02*ne01 - i02*ne01);
 
-    device const float * psrc0 =        src0 + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
-    device const float * pmask = src1 ? src1                                      + i01*ne00 : nullptr;
-    device       float * pdst  =        dst  + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
+    device const float * psrc0 =         src0 + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
+    device const float * pmask = src1 != src0 ? src1                               + i01*ne00 : nullptr;
+    device       float * pdst  =         dst  + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
 
     // parallel max
     float lmax = -INFINITY;
@@ -236,7 +404,12 @@ kernel void kernel_soft_max(
         pdst[i00] = exp_psrc0;
     }
 
+    // This barrier fixes a failing test
+    // ref: https://github.com/ggerganov/ggml/pull/621#discussion_r1425156335
+    threadgroup_barrier(mem_flags::mem_none);
+
     float sum = simd_sum(lsum);
+
     if (ntg > N_SIMDWIDTH) {
         if (sgitg == 0) {
             buf[tiisg] = 0.0f;
@@ -279,9 +452,9 @@ kernel void kernel_soft_max_4(
     const int64_t i02 = (tgpig - i03*ne02*ne01) / ne01;
     const int64_t i01 = (tgpig - i03*ne02*ne01 - i02*ne01);
 
-    device const float4 * psrc4 =        (device const float4 *)(src0 + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00);
-    device const float4 * pmask = src1 ? (device const float4 *)(src1 +                                      i01*ne00) : nullptr;
-    device       float4 * pdst4 =        (device       float4 *)(dst  + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00);
+    device const float4 * psrc4 =                (device const float4 *)(src0 + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00);
+    device const float4 * pmask = src1 != src0 ? (device const float4 *)(src1 +                                      i01*ne00) : nullptr;
+    device       float4 * pdst4 =                (device       float4 *)(dst  + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00);
 
     // parallel max
     float4 lmax4 = -INFINITY;
@@ -319,7 +492,13 @@ kernel void kernel_soft_max_4(
     }
 
     const float lsum = lsum4[0] + lsum4[1] + lsum4[2] + lsum4[3];
+
+    // This barrier fixes a failing test
+    // ref: https://github.com/ggerganov/ggml/pull/621#discussion_r1425156335
+    threadgroup_barrier(mem_flags::mem_none);
+
     float sum = simd_sum(lsum);
+
     if (ntg > N_SIMDWIDTH) {
         if (sgitg == 0) {
             buf[tiisg] = 0.0f;
@@ -490,6 +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
@@ -582,9 +849,20 @@ inline float block_q_n_dot_y(device const block_q5_1 * qb_curr, float sumy, thre
 //      giard against the number of rows not being divisible by
 //      N_DST, so this is another explicit assumption of the implementation.
 template<typename block_q_type, int nr, int nsg, int nw>
-void mul_vec_q_n_f32(device const void * src0, device const float * src1, device float * dst,
-                    int64_t ne00, int64_t ne01, int64_t ne02, int64_t ne10, int64_t ne12, int64_t ne0, int64_t ne1, uint gqa,
-                    uint3 tgpig, uint tiisg, uint sgitg) {
+void mul_vec_q_n_f32_impl(
+        device const void  * src0,
+        device const float * src1,
+        device       float * dst,
+                   int64_t   ne00,
+                   int64_t   ne01,
+                   int64_t   ne02,
+                   int64_t   ne10,
+                   int64_t   ne12,
+                   int64_t   ne0,
+                   int64_t   ne1,
+                   uint      r2,
+                   uint      r3,
+                   uint3 tgpig, uint tiisg, uint sgitg) {
     const int nb = ne00/QK4_0;
 
     const int r0 = tgpig.x;
@@ -593,7 +871,10 @@ void mul_vec_q_n_f32(device const void * src0, device const float * src1, device
 
     const int first_row = (r0 * nsg + sgitg) * nr;
 
-    const uint offset0 = first_row * nb + im/gqa*(nb*ne0);
+    const uint i12 = im%ne12;
+    const uint i13 = im/ne12;
+
+    const uint offset0 = first_row * nb + (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
 
     device const block_q_type * x = (device const block_q_type *) src0 + offset0;
     device const float        * y = (device const float        *) src1 + r1*ne10 + im*ne00*ne1;
@@ -643,13 +924,14 @@ kernel void kernel_mul_mv_q4_0_f32(
         constant   int64_t & ne02[[buffer(5)]],
         constant   int64_t & ne10[[buffer(9)]],
         constant   int64_t & ne12[[buffer(11)]],
-        constant   int64_t & ne0[[buffer(15)]],
-        constant   int64_t & ne1[[buffer(16)]],
-        constant   uint    & gqa[[buffer(17)]],
+        constant   int64_t & ne0 [[buffer(15)]],
+        constant   int64_t & ne1 [[buffer(16)]],
+        constant   uint    & r2  [[buffer(17)]],
+        constant   uint    & r3  [[buffer(18)]],
         uint3 tgpig[[threadgroup_position_in_grid]],
         uint  tiisg[[thread_index_in_simdgroup]],
         uint  sgitg[[simdgroup_index_in_threadgroup]]) {
-    mul_vec_q_n_f32<block_q4_0, N_DST, N_SIMDGROUP, N_SIMDWIDTH>(src0,src1,dst,ne00,ne01,ne02,ne10,ne12,ne0,ne1,gqa,tgpig,tiisg,sgitg);
+    mul_vec_q_n_f32_impl<block_q4_0, N_DST, N_SIMDGROUP, N_SIMDWIDTH>(src0,src1,dst,ne00,ne01,ne02,ne10,ne12,ne0,ne1,r2,r3,tgpig,tiisg,sgitg);
 }
 
 kernel void kernel_mul_mv_q4_1_f32(
@@ -661,13 +943,14 @@ kernel void kernel_mul_mv_q4_1_f32(
         constant   int64_t & ne02[[buffer(5)]],
         constant   int64_t & ne10[[buffer(9)]],
         constant   int64_t & ne12[[buffer(11)]],
-        constant   int64_t & ne0[[buffer(15)]],
-        constant   int64_t & ne1[[buffer(16)]],
-        constant   uint    & gqa[[buffer(17)]],
+        constant   int64_t & ne0 [[buffer(15)]],
+        constant   int64_t & ne1 [[buffer(16)]],
+        constant   uint    & r2  [[buffer(17)]],
+        constant   uint    & r3  [[buffer(18)]],
         uint3 tgpig[[threadgroup_position_in_grid]],
         uint tiisg[[thread_index_in_simdgroup]],
         uint sgitg[[simdgroup_index_in_threadgroup]]) {
-     mul_vec_q_n_f32<block_q4_1, N_DST, N_SIMDGROUP, N_SIMDWIDTH>(src0,src1,dst,ne00,ne01,ne02,ne10,ne12,ne0,ne1,gqa,tgpig,tiisg,sgitg);
+     mul_vec_q_n_f32_impl<block_q4_1, N_DST, N_SIMDGROUP, N_SIMDWIDTH>(src0,src1,dst,ne00,ne01,ne02,ne10,ne12,ne0,ne1,r2,r3,tgpig,tiisg,sgitg);
 }
 
 kernel void kernel_mul_mv_q5_0_f32(
@@ -679,13 +962,14 @@ kernel void kernel_mul_mv_q5_0_f32(
         constant   int64_t & ne02[[buffer(5)]],
         constant   int64_t & ne10[[buffer(9)]],
         constant   int64_t & ne12[[buffer(11)]],
-        constant   int64_t & ne0[[buffer(15)]],
-        constant   int64_t & ne1[[buffer(16)]],
-        constant   uint    & gqa[[buffer(17)]],
+        constant   int64_t & ne0 [[buffer(15)]],
+        constant   int64_t & ne1 [[buffer(16)]],
+        constant   uint    & r2  [[buffer(17)]],
+        constant   uint    & r3  [[buffer(18)]],
         uint3 tgpig[[threadgroup_position_in_grid]],
         uint  tiisg[[thread_index_in_simdgroup]],
         uint  sgitg[[simdgroup_index_in_threadgroup]]) {
-    mul_vec_q_n_f32<block_q5_0, N_DST, N_SIMDGROUP, N_SIMDWIDTH>(src0,src1,dst,ne00,ne01,ne02,ne10,ne12,ne0,ne1,gqa,tgpig,tiisg,sgitg);
+    mul_vec_q_n_f32_impl<block_q5_0, N_DST, N_SIMDGROUP, N_SIMDWIDTH>(src0,src1,dst,ne00,ne01,ne02,ne10,ne12,ne0,ne1,r2,r3,tgpig,tiisg,sgitg);
 }
 
 kernel void kernel_mul_mv_q5_1_f32(
@@ -697,33 +981,35 @@ kernel void kernel_mul_mv_q5_1_f32(
         constant   int64_t & ne02[[buffer(5)]],
         constant   int64_t & ne10[[buffer(9)]],
         constant   int64_t & ne12[[buffer(11)]],
-        constant   int64_t & ne0[[buffer(15)]],
-        constant   int64_t & ne1[[buffer(16)]],
-        constant   uint    & gqa[[buffer(17)]],
+        constant   int64_t & ne0 [[buffer(15)]],
+        constant   int64_t & ne1 [[buffer(16)]],
+        constant   uint    & r2  [[buffer(17)]],
+        constant   uint    & r3  [[buffer(18)]],
         uint3 tgpig[[threadgroup_position_in_grid]],
         uint  tiisg[[thread_index_in_simdgroup]],
         uint  sgitg[[simdgroup_index_in_threadgroup]]) {
-    mul_vec_q_n_f32<block_q5_1, N_DST, N_SIMDGROUP, N_SIMDWIDTH>(src0,src1,dst,ne00,ne01,ne02,ne10,ne12,ne0,ne1,gqa,tgpig,tiisg,sgitg);
+    mul_vec_q_n_f32_impl<block_q5_1, N_DST, N_SIMDGROUP, N_SIMDWIDTH>(src0,src1,dst,ne00,ne01,ne02,ne10,ne12,ne0,ne1,r2,r3,tgpig,tiisg,sgitg);
 }
 
 
 #define NB_Q8_0 8
 
-kernel void kernel_mul_mv_q8_0_f32(
+void kernel_mul_mv_q8_0_f32_impl(
         device const  void * src0,
         device const float * src1,
         device       float * dst,
         constant   int64_t & ne00,
-        constant   int64_t & ne01[[buffer(4)]],
-        constant   int64_t & ne02[[buffer(5)]],
-        constant   int64_t & ne10[[buffer(9)]],
-        constant   int64_t & ne12[[buffer(11)]],
-        constant   int64_t & ne0[[buffer(15)]],
-        constant   int64_t & ne1[[buffer(16)]],
-        constant   uint    & gqa[[buffer(17)]],
+        constant   int64_t & ne01,
+        constant   int64_t & ne02,
+        constant   int64_t & ne10,
+        constant   int64_t & ne12,
+        constant   int64_t & ne0,
+        constant   int64_t & ne1,
+        constant   uint    & r2,
+        constant   uint    & r3,
         uint3 tgpig[[threadgroup_position_in_grid]],
-        uint tiisg[[thread_index_in_simdgroup]],
-        uint sgitg[[simdgroup_index_in_threadgroup]]) {
+        uint  tiisg[[thread_index_in_simdgroup]],
+        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
     const int nr  = N_DST;
     const int nsg = N_SIMDGROUP;
     const int nw  = N_SIMDWIDTH;
@@ -732,8 +1018,14 @@ kernel void kernel_mul_mv_q8_0_f32(
     const int r0 = tgpig.x;
     const int r1 = tgpig.y;
     const int im = tgpig.z;
+
     const int first_row = (r0 * nsg + sgitg) * nr;
-    const uint offset0 = first_row * nb + im/gqa*(nb*ne0);
+
+    const uint i12 = im%ne12;
+    const uint i13 = im/ne12;
+
+    const uint offset0 = first_row * nb + (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
+
     device const block_q8_0 * x = (device const block_q8_0 *) src0 + offset0;
     device const float      * y = (device const float      *) src1 + r1*ne10 + im*ne00*ne1;
 
@@ -771,9 +1063,29 @@ kernel void kernel_mul_mv_q8_0_f32(
     }
 }
 
+[[host_name("kernel_mul_mv_q8_0_f32")]]
+kernel void kernel_mul_mv_q8_0_f32(
+        device const  void * src0,
+        device const float * src1,
+        device       float * dst,
+        constant   int64_t & ne00,
+        constant   int64_t & ne01,
+        constant   int64_t & ne02,
+        constant   int64_t & ne10,
+        constant   int64_t & ne12,
+        constant   int64_t & ne0,
+        constant   int64_t & ne1,
+        constant   uint    & r2   [[buffer(17)]],
+        constant   uint    & r3   [[buffer(18)]],
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint  tiisg[[thread_index_in_simdgroup]],
+        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
+    kernel_mul_mv_q8_0_f32_impl(src0,src1,dst,ne00,ne01,ne02,ne10,ne12,ne0,ne1,r2,r3,tgpig,tiisg,sgitg);
+}
+
 #define N_F32_F32 4
 
-kernel void kernel_mul_mv_f32_f32(
+void kernel_mul_mv_f32_f32_impl(
         device const  char * src0,
         device const  char * src1,
         device       float * dst,
@@ -791,6 +1103,8 @@ kernel void kernel_mul_mv_f32_f32(
         constant  uint64_t & nb12,
         constant   int64_t & ne0,
         constant   int64_t & ne1,
+        constant   uint    & r2,
+        constant   uint    & r3,
         uint3 tgpig[[threadgroup_position_in_grid]],
         uint  tiisg[[thread_index_in_simdgroup]]) {
 
@@ -798,7 +1112,12 @@ kernel void kernel_mul_mv_f32_f32(
     const int64_t rb = tgpig.y*N_F32_F32;
     const int64_t im = tgpig.z;
 
-    device const float * x = (device const float *) (src0 + r0*nb01 + im/(ne12/ne02)*nb02);
+    const uint i12 = im%ne12;
+    const uint i13 = im/ne12;
+
+    const uint offset0 = r0*nb01 + (i12/r2)*nb02 + (i13/r3)*nb02*ne02;
+
+    device const float * x = (device const float *) (src0 + offset0);
 
     if (ne00 < 128) {
         for (int row = 0; row < N_F32_F32; ++row) {
@@ -844,6 +1163,32 @@ kernel void kernel_mul_mv_f32_f32(
     }
 }
 
+[[host_name("kernel_mul_mv_f32_f32")]]
+kernel void kernel_mul_mv_f32_f32(
+        device const  char * src0,
+        device const  char * src1,
+        device       float * dst,
+        constant   int64_t & ne00,
+        constant   int64_t & ne01,
+        constant   int64_t & ne02,
+        constant  uint64_t & nb00,
+        constant  uint64_t & nb01,
+        constant  uint64_t & nb02,
+        constant   int64_t & ne10,
+        constant   int64_t & ne11,
+        constant   int64_t & ne12,
+        constant  uint64_t & nb10,
+        constant  uint64_t & nb11,
+        constant  uint64_t & nb12,
+        constant   int64_t & ne0,
+        constant   int64_t & ne1,
+        constant   uint    & r2   [[buffer(17)]],
+        constant   uint    & r3   [[buffer(18)]],
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint  tiisg[[thread_index_in_simdgroup]]) {
+    kernel_mul_mv_f32_f32_impl(src0, src1, dst, ne00, ne01, ne02, nb00, nb01, nb02, ne10, ne11, ne12, nb10, nb11, nb12, ne0, ne1, r2, r3, tgpig, tiisg);
+}
+
 #define N_F16_F16 4
 
 kernel void kernel_mul_mv_f16_f16(
@@ -864,6 +1209,8 @@ kernel void kernel_mul_mv_f16_f16(
         constant  uint64_t & nb12,
         constant   int64_t & ne0,
         constant   int64_t & ne1,
+        constant   uint    & r2   [[buffer(17)]],
+        constant   uint    & r3   [[buffer(18)]],
         uint3 tgpig[[threadgroup_position_in_grid]],
         uint  tiisg[[thread_index_in_simdgroup]]) {
 
@@ -871,7 +1218,12 @@ kernel void kernel_mul_mv_f16_f16(
     const int64_t rb = tgpig.y*N_F16_F16;
     const int64_t im = tgpig.z;
 
-    device const half * x = (device const half *) (src0 + r0*nb01 + im/(ne12/ne02)*nb02);
+    const uint i12 = im%ne12;
+    const uint i13 = im/ne12;
+
+    const uint offset0 = r0*nb01 + (i12/r2)*nb02 + (i13/r3)*nb02*ne02;
+
+    device const half * x = (device const half *) (src0 + offset0);
 
     if (ne00 < 128) {
         for (int row = 0; row < N_F16_F16; ++row) {
@@ -917,7 +1269,7 @@ kernel void kernel_mul_mv_f16_f16(
     }
 }
 
-kernel void kernel_mul_mv_f16_f32_1row(
+void kernel_mul_mv_f16_f32_1row_impl(
         device const  char * src0,
         device const  char * src1,
         device       float * dst,
@@ -935,6 +1287,8 @@ kernel void kernel_mul_mv_f16_f32_1row(
         constant  uint64_t & nb12,
         constant   int64_t & ne0,
         constant   int64_t & ne1,
+        constant   uint    & r2,
+        constant   uint    & r3,
         uint3 tgpig[[threadgroup_position_in_grid]],
         uint  tiisg[[thread_index_in_simdgroup]]) {
 
@@ -942,7 +1296,12 @@ kernel void kernel_mul_mv_f16_f32_1row(
     const int64_t r1 = tgpig.y;
     const int64_t im = tgpig.z;
 
-    device const half  * x = (device const half  *) (src0 + r0*nb01 + im/(ne12/ne02)*nb02);
+    const uint i12 = im%ne12;
+    const uint i13 = im/ne12;
+
+    const uint offset0 = r0*nb01 + (i12/r2)*nb02 + (i13/r3)*nb02*ne02;
+
+    device const half  * x = (device const half  *) (src0 + offset0);
     device const float * y = (device const float *) (src1 + r1*nb11 + im*nb12);
 
     float sumf = 0;
@@ -966,12 +1325,10 @@ kernel void kernel_mul_mv_f16_f32_1row(
             dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum;
         }
     }
-
 }
 
-#define N_F16_F32 4
-
-kernel void kernel_mul_mv_f16_f32(
+[[host_name("kernel_mul_mv_f16_f32_1row")]]
+kernel void kernel_mul_mv_f16_f32_1row(
         device const  char * src0,
         device const  char * src1,
         device       float * dst,
@@ -989,6 +1346,35 @@ kernel void kernel_mul_mv_f16_f32(
         constant  uint64_t & nb12,
         constant   int64_t & ne0,
         constant   int64_t & ne1,
+        constant   uint    & r2   [[buffer(17)]],
+        constant   uint    & r3   [[buffer(18)]],
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint  tiisg[[thread_index_in_simdgroup]]) {
+    kernel_mul_mv_f16_f32_1row_impl(src0, src1, dst, ne00, ne01, ne02, nb00, nb01, nb02, ne10, ne11, ne12, nb10, nb11, nb12, ne0, ne1, r2, r3, tgpig, tiisg);
+}
+
+#define N_F16_F32 4
+
+void kernel_mul_mv_f16_f32_impl(
+        device const  char * src0,
+        device const  char * src1,
+        device       float * dst,
+        constant   int64_t & ne00,
+        constant   int64_t & ne01,
+        constant   int64_t & ne02,
+        constant  uint64_t & nb00,
+        constant  uint64_t & nb01,
+        constant  uint64_t & nb02,
+        constant   int64_t & ne10,
+        constant   int64_t & ne11,
+        constant   int64_t & ne12,
+        constant  uint64_t & nb10,
+        constant  uint64_t & nb11,
+        constant  uint64_t & nb12,
+        constant   int64_t & ne0,
+        constant   int64_t & ne1,
+        constant   uint    & r2,
+        constant   uint    & r3,
         uint3 tgpig[[threadgroup_position_in_grid]],
         uint tiisg[[thread_index_in_simdgroup]]) {
 
@@ -996,7 +1382,12 @@ kernel void kernel_mul_mv_f16_f32(
     const int64_t rb = tgpig.y*N_F16_F32;
     const int64_t im = tgpig.z;
 
-    device const half * x = (device const half *) (src0 + r0*nb01 + im/(ne12/ne02)*nb02);
+    const uint i12 = im%ne12;
+    const uint i13 = im/ne12;
+
+    const uint offset0 = r0*nb01 + (i12/r2)*nb02 + (i13/r3)*nb02*ne02;
+
+    device const half * x = (device const half *) (src0 + offset0);
 
     if (ne00 < 128) {
         for (int row = 0; row < N_F16_F32; ++row) {
@@ -1042,6 +1433,32 @@ kernel void kernel_mul_mv_f16_f32(
     }
 }
 
+[[host_name("kernel_mul_mv_f16_f32")]]
+kernel void kernel_mul_mv_f16_f32(
+        device const  char * src0,
+        device const  char * src1,
+        device       float * dst,
+        constant   int64_t & ne00,
+        constant   int64_t & ne01,
+        constant   int64_t & ne02,
+        constant  uint64_t & nb00,
+        constant  uint64_t & nb01,
+        constant  uint64_t & nb02,
+        constant   int64_t & ne10,
+        constant   int64_t & ne11,
+        constant   int64_t & ne12,
+        constant  uint64_t & nb10,
+        constant  uint64_t & nb11,
+        constant  uint64_t & nb12,
+        constant   int64_t & ne0,
+        constant   int64_t & ne1,
+        constant   uint    & r2   [[buffer(17)]],
+        constant   uint    & r3   [[buffer(18)]],
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint tiisg[[thread_index_in_simdgroup]]) {
+    kernel_mul_mv_f16_f32_impl(src0, src1, dst, ne00, ne01, ne02, nb00, nb01, nb02, ne10, ne11, ne12, nb10, nb11, nb12, ne0, ne1, r2, r3, tgpig, tiisg);
+}
+
 // Assumes row size (ne00) is a multiple of 4
 kernel void kernel_mul_mv_f16_f32_l4(
         device const  char * src0,
@@ -1061,6 +1478,8 @@ kernel void kernel_mul_mv_f16_f32_l4(
         constant  uint64_t & nb12,
         constant   int64_t & ne0,
         constant   int64_t & ne1,
+        constant   uint    & r2   [[buffer(17)]],
+        constant   uint    & r3   [[buffer(18)]],
         uint3 tgpig[[threadgroup_position_in_grid]],
         uint tiisg[[thread_index_in_simdgroup]]) {
 
@@ -1068,7 +1487,12 @@ kernel void kernel_mul_mv_f16_f32_l4(
     const int64_t r0 = tgpig.x;
     const int64_t im = tgpig.z;
 
-    device const half4 * x4 = (device const half4 *) (src0 + r0*nb01 + im/(ne12/ne02)*nb02);
+    const uint i12 = im%ne12;
+    const uint i13 = im/ne12;
+
+    const uint offset0 = r0*nb01 + (i12/r2)*nb02 + (i13/r3)*nb02*ne02;
+
+    device const half4 * x4 = (device const half4 *) (src0 + offset0);
 
     for (int r1 = 0; r1 < nrows; ++r1) {
         device const float4 * y4 = (device const float4 *) (src1 + r1*nb11 + im*nb12);
@@ -1120,17 +1544,21 @@ kernel void kernel_alibi_f32(
     const int64_t i2 = (n - i3*ne2*ne1*ne0) / (ne1*ne0);
     const int64_t i1 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0) / ne0;
     const int64_t i0 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0 - i1*ne0);
+    const int64_t k = i3*ne3 + i2;
 
-    device float * dst_data = (device float *) ((device char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
     float m_k;
-    if (i2 < n_heads_log2_floor) {
-        m_k = pow(m0, i2 + 1);
+    if (k < n_heads_log2_floor) {
+        m_k = pow(m0, k + 1);
     } else {
-        m_k = pow(m1, 2 * (i2 - n_heads_log2_floor) + 1);
+        m_k = pow(m1, 2 * (k - n_heads_log2_floor) + 1);
     }
+
+    device       char * dst_row = (device char *) dst + i3*nb3 + i2*nb2 + i1*nb1;
+    device const char * src_row = (device char *) src0 + i03*nb03 + i02*nb02 + i01*nb01;
     for (int64_t i00 = tpitg.x; i00 < ne00; i00 += ntg.x) {
-        device const float * src = (device float *)((device char *) src0 + i03*nb03 + i02*nb02 + i01*nb01 + i00*nb00);
-        dst_data[i00] = src[0] + m_k * (i00 - ne00 + 1);
+        const  float   src_v = *(device float *)(src_row + i00*nb00);
+        device float * dst_v =  (device float *)(dst_row + i00*nb0);
+        *dst_v = i00 * m_k + src_v;
     }
 }
 
@@ -1274,8 +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;
@@ -1294,6 +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];
             }
         }
     }
@@ -1335,9 +1772,160 @@ 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,
+        device     int32_t * dst,
+        constant   int64_t & ncols,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]]);
+
+template<ggml_sort_order order>
+kernel void kernel_argsort_f32_i32(
+        device const float   * x,
+        device       int32_t * dst,
+        constant     int64_t & ncols,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]]) {
+    // bitonic sort
+    int col = tpitg[0];
+    int row = tgpig[1];
+
+    if (col >= ncols) return;
+
+    device const float   * x_row   = x   + row * ncols;
+    device       int32_t * dst_row = dst + row * ncols;
+
+    // initialize indices
+    if (col < ncols) {
+        dst_row[col] = col;
+    }
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    for (int k = 2; k <= ncols; k *= 2) {
+        for (int j = k / 2; j > 0; j /= 2) {
+            int ixj = col ^ j;
+            if (ixj > col) {
+                if ((col & k) == 0) {
+                    if (order == GGML_SORT_ASC ? x_row[dst_row[col]] > x_row[dst_row[ixj]] : x_row[dst_row[col]] < x_row[dst_row[ixj]]) {
+                        SWAP(dst_row[col], dst_row[ixj]);
+                    }
+                } else {
+                    if (order == GGML_SORT_ASC ? x_row[dst_row[col]] < x_row[dst_row[ixj]] : x_row[dst_row[col]] > x_row[dst_row[ixj]]) {
+                        SWAP(dst_row[col], dst_row[ixj]);
+                    }
+                }
+            }
+            threadgroup_barrier(mem_flags::mem_threadgroup);
+        }
+    }
+}
+
+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,
+        device  const half * src0,
+        device        half * dst,
         constant   int64_t & ne00,
         constant   int64_t & ne01,
         constant   int64_t & ne02,
@@ -1376,6 +1964,47 @@ kernel void kernel_cpy_f16_f16(
     }
 }
 
+kernel void kernel_cpy_f16_f32(
+        device  const half * src0,
+        device       float * 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 i03 = tgpig[2];
+    const int64_t i02 = tgpig[1];
+    const int64_t i01 = tgpig[0];
+
+    const int64_t n = i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
+
+    const int64_t i3 = n / (ne2*ne1*ne0);
+    const int64_t i2 = (n - i3*ne2*ne1*ne0) / (ne1*ne0);
+    const int64_t i1 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0) / ne0;
+    const int64_t i0 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0 - i1*ne0);
+
+    device float * dst_data = (device float *) ((device char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
+
+    for (int64_t i00 = tpitg.x; i00 < ne00; i00 += ntg.x) {
+        device const half * src = (device half *)((device char *) src0 + i03*nb03 + i02*nb02 + i01*nb01 + i00*nb00);
+        dst_data[i00] = src[0];
+    }
+}
+
 kernel void kernel_cpy_f32_f16(
         device const float * src0,
         device        half * dst,
@@ -1460,10 +2089,201 @@ kernel void kernel_cpy_f32_f32(
     }
 }
 
+kernel void kernel_cpy_f32_q8_0(
+        device const float * src0,
+        device        void * 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 i03 = tgpig[2];
+    const int64_t i02 = tgpig[1];
+    const int64_t i01 = tgpig[0];
+
+    const int64_t n = i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
+
+    const int64_t i3 = n / (ne2*ne1*ne0);
+    const int64_t i2 = (n - i3*ne2*ne1*ne0) / (ne1*ne0);
+    const int64_t i1 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0) / ne0;
+    const int64_t i0 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0 - i1*ne0)/QK8_0;
+
+    device block_q8_0 * dst_data = (device block_q8_0 *) ((device char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
+
+    for (int64_t i00 = tpitg.x*QK8_0; i00 < ne00; i00 += ntg.x*QK8_0) {
+        device const float * src = (device float *)((device char *) src0 + i03*nb03 + i02*nb02 + i01*nb01 + i00*nb00);
+
+        float amax = 0.0f; // absolute max
+
+        for (int j = 0; j < QK8_0; j++) {
+            const float v = src[j];
+            amax = MAX(amax, fabs(v));
+        }
+
+        const float d = amax / ((1 << 7) - 1);
+        const float id = d ? 1.0f/d : 0.0f;
+
+        dst_data[i00/QK8_0].d = d;
+
+        for (int j = 0; j < QK8_0; ++j) {
+            const float x0 = src[j]*id;
+
+            dst_data[i00/QK8_0].qs[j] = round(x0);
+        }
+    }
+}
+
+kernel void kernel_cpy_f32_q4_0(
+        device const float * src0,
+        device        void * 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 i03 = tgpig[2];
+    const int64_t i02 = tgpig[1];
+    const int64_t i01 = tgpig[0];
+
+    const int64_t n = i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
+
+    const int64_t i3 = n / (ne2*ne1*ne0);
+    const int64_t i2 = (n - i3*ne2*ne1*ne0) / (ne1*ne0);
+    const int64_t i1 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0) / ne0;
+    const int64_t i0 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0 - i1*ne0)/QK4_0;
+
+    device block_q4_0 * dst_data = (device block_q4_0 *) ((device char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
+
+    for (int64_t i00 = tpitg.x*QK4_0; i00 < ne00; i00 += ntg.x*QK4_0) {
+        device const float * src = (device float *)((device char *) src0 + i03*nb03 + i02*nb02 + i01*nb01 + i00*nb00);
+
+        float amax = 0.0f; // absolute max
+        float max  = 0.0f;
+
+        for (int j = 0; j < QK4_0; j++) {
+            const float v = src[j];
+            if (amax < fabs(v)) {
+                amax = fabs(v);
+                max  = v;
+            }
+        }
+
+        const float d = max / -8;
+        const float id = d ? 1.0f/d : 0.0f;
+
+        dst_data[i00/QK4_0].d = d;
+
+        for (int j = 0; j < QK4_0/2; ++j) {
+            const float x0 = src[0       + j]*id;
+            const float x1 = src[QK4_0/2 + j]*id;
+
+            const uint8_t xi0 = MIN(15, (int8_t)(x0 + 8.5f));
+            const uint8_t xi1 = MIN(15, (int8_t)(x1 + 8.5f));
+
+            dst_data[i00/QK4_0].qs[j]  = xi0;
+            dst_data[i00/QK4_0].qs[j] |= xi1 << 4;
+        }
+    }
+}
+
+kernel void kernel_cpy_f32_q4_1(
+        device const float * src0,
+        device        void * 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 i03 = tgpig[2];
+    const int64_t i02 = tgpig[1];
+    const int64_t i01 = tgpig[0];
+
+    const int64_t n = i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
+
+    const int64_t i3 = n / (ne2*ne1*ne0);
+    const int64_t i2 = (n - i3*ne2*ne1*ne0) / (ne1*ne0);
+    const int64_t i1 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0) / ne0;
+    const int64_t i0 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0 - i1*ne0)/QK4_1;
+
+    device block_q4_1 * dst_data = (device block_q4_1 *) ((device char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
+
+    for (int64_t i00 = tpitg.x*QK4_1; i00 < ne00; i00 += ntg.x*QK4_1) {
+        device const float * src = (device float *)((device char *) src0 + i03*nb03 + i02*nb02 + i01*nb01 + i00*nb00);
+
+        float min = FLT_MAX;
+        float max = -FLT_MAX;
+
+        for (int j = 0; j < QK4_1; j++) {
+            const float v = src[j];
+            if (min > v) min = v;
+            if (max < v) max = v;
+        }
+
+        const float d = (max - min) / ((1 << 4) - 1);
+        const float id = d ? 1.0f/d : 0.0f;
+
+        dst_data[i00/QK4_1].d = d;
+        dst_data[i00/QK4_1].m = min;
+
+        for (int j = 0; j < QK4_1/2; ++j) {
+            const float x0 = (src[0       + j] - min)*id;
+            const float x1 = (src[QK4_1/2 + j] - min)*id;
+
+            const uint8_t xi0 = MIN(15, (int8_t)(x0 + 0.5f));
+            const uint8_t xi1 = MIN(15, (int8_t)(x1 + 0.5f));
+
+            dst_data[i00/QK4_1].qs[j]  = xi0;
+            dst_data[i00/QK4_1].qs[j] |= xi1 << 4;
+        }
+    }
+}
+
 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,
@@ -1500,7 +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;
 
@@ -1608,32 +2428,39 @@ static inline uchar4 get_scale_min_k4(int j, device const uint8_t * q) {
 
 //====================================== dot products =========================
 
-kernel void kernel_mul_mv_q2_K_f32(
+void kernel_mul_mv_q2_K_f32_impl(
         device const  void * src0,
         device const float * src1,
         device       float * dst,
         constant   int64_t & ne00,
-        constant   int64_t & ne01[[buffer(4)]],
-        constant   int64_t & ne02[[buffer(5)]],
-        constant   int64_t & ne10[[buffer(9)]],
-        constant   int64_t & ne12[[buffer(11)]],
-        constant   int64_t & ne0[[buffer(15)]],
-        constant   int64_t & ne1[[buffer(16)]],
-        constant   uint    & gqa[[buffer(17)]],
+        constant   int64_t & ne01,
+        constant   int64_t & ne02,
+        constant   int64_t & ne10,
+        constant   int64_t & ne12,
+        constant   int64_t & ne0,
+        constant   int64_t & ne1,
+        constant   uint    & r2,
+        constant   uint    & r3,
         uint3 tgpig[[threadgroup_position_in_grid]],
-        uint tiisg[[thread_index_in_simdgroup]],
-        uint sgitg[[simdgroup_index_in_threadgroup]]) {
+        uint  tiisg[[thread_index_in_simdgroup]],
+        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
 
     const int nb = ne00/QK_K;
     const int r0 = tgpig.x;
     const int r1 = tgpig.y;
-    const int r2 = tgpig.z;
+    const int im = tgpig.z;
 
     const int first_row = (r0 * N_SIMDGROUP + sgitg) * N_DST;
     const int ib_row = first_row * nb;
-    const uint offset0 = r2/gqa*(nb*ne0);
+
+    const uint i12 = im%ne12;
+    const uint i13 = im/ne12;
+
+    const uint offset0 = (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
+
     device const block_q2_K * x = (device const block_q2_K *) src0 + ib_row + offset0;
-    device const float      * y = (device const float      *) src1 + r1*ne10 + r2*ne00*ne1;
+    device const float      * y = (device const float      *) src1 + r1*ne10 + im*ne00*ne1;
+
     float yl[32];
     float sumf[N_DST]={0.f}, all_sum;
 
@@ -1642,11 +2469,11 @@ kernel void kernel_mul_mv_q2_K_f32(
 #if QK_K == 256
     const int ix = tiisg/8;  // 0...3
     const int it = tiisg%8;  // 0...7
-    const int im = it/4;     // 0 or 1
+    const int iq = it/4;     // 0 or 1
     const int ir = it%4;     // 0...3
     const int is = (8*ir)/16;// 0 or 1
 
-    device const float * y4 = y + ix * QK_K + 128 * im + 8 * ir;
+    device const float * y4 = y + ix * QK_K + 128 * iq + 8 * ir;
 
     for (int ib = ix; ib < nb; ib += 4) {
 
@@ -1658,8 +2485,8 @@ kernel void kernel_mul_mv_q2_K_f32(
             yl[i+24] = y4[i+96]; sumy[3] += yl[i+24];
         }
 
-        device const uint8_t  * sc = (device const uint8_t  *)x[ib].scales + 8*im + is;
-        device const uint16_t * qs = (device const uint16_t *)x[ib].qs + 16 * im + 4 * ir;
+        device const uint8_t  * sc = (device const uint8_t  *)x[ib].scales + 8*iq + is;
+        device const uint16_t * qs = (device const uint16_t *)x[ib].qs + 16 * iq + 4 * ir;
         device const half     * dh = &x[ib].d;
 
         for (int row = 0; row < N_DST; row++) {
@@ -1746,13 +2573,13 @@ kernel void kernel_mul_mv_q2_K_f32(
     for (int row = 0; row < N_DST; ++row) {
         all_sum = simd_sum(sumf[row]);
         if (tiisg == 0) {
-            dst[r1*ne0 + r2*ne0*ne1 + first_row + row] = all_sum;
+            dst[r1*ne0 + im*ne0*ne1 + first_row + row] = all_sum;
         }
     }
 }
 
-#if QK_K == 256
-kernel void kernel_mul_mv_q3_K_f32(
+[[host_name("kernel_mul_mv_q2_K_f32")]]
+kernel void kernel_mul_mv_q2_K_f32(
         device const  void * src0,
         device const float * src1,
         device       float * dst,
@@ -1761,23 +2588,50 @@ kernel void kernel_mul_mv_q3_K_f32(
         constant   int64_t & ne02[[buffer(5)]],
         constant   int64_t & ne10[[buffer(9)]],
         constant   int64_t & ne12[[buffer(11)]],
-        constant   int64_t & ne0[[buffer(15)]],
-        constant   int64_t & ne1[[buffer(16)]],
-        constant   uint    & gqa[[buffer(17)]],
+        constant   int64_t & ne0 [[buffer(15)]],
+        constant   int64_t & ne1 [[buffer(16)]],
+        constant   uint    & r2  [[buffer(17)]],
+        constant   uint    & r3  [[buffer(18)]],
         uint3 tgpig[[threadgroup_position_in_grid]],
-        uint tiisg[[thread_index_in_simdgroup]],
-        uint sgitg[[simdgroup_index_in_threadgroup]]) {
+        uint  tiisg[[thread_index_in_simdgroup]],
+        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
+
+    kernel_mul_mv_q2_K_f32_impl(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3, tgpig, tiisg, sgitg);
+}
+
+#if QK_K == 256
+void kernel_mul_mv_q3_K_f32_impl(
+        device const  void * src0,
+        device const float * src1,
+        device       float * dst,
+        constant   int64_t & ne00,
+        constant   int64_t & ne01,
+        constant   int64_t & ne02,
+        constant   int64_t & ne10,
+        constant   int64_t & ne12,
+        constant   int64_t & ne0,
+        constant   int64_t & ne1,
+        constant   uint    & r2,
+        constant   uint    & r3,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint  tiisg[[thread_index_in_simdgroup]],
+        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
 
     const int nb = ne00/QK_K;
 
     const int64_t r0 = tgpig.x;
     const int64_t r1 = tgpig.y;
-    const int64_t r2 = tgpig.z;
+    const int64_t im = tgpig.z;
 
     const int first_row = (r0 * N_SIMDGROUP + sgitg) * 2;
-    const uint offset0 = r2/gqa*(nb*ne0);
+
+    const uint i12 = im%ne12;
+    const uint i13 = im/ne12;
+
+    const uint offset0 = (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
+
     device const block_q3_K * x = (device const block_q3_K *) src0 + first_row*nb + offset0;
-    device const float     * yy = (device const float      *) src1 + r1*ne10 + r2*ne00*ne1;
+    device const float     * yy = (device const float      *) src1 + r1*ne10 + im*ne00*ne1;
 
     float yl[32];
 
@@ -1899,40 +2753,47 @@ kernel void kernel_mul_mv_q3_K_f32(
     }
     if (tiisg == 0) {
         for (int row = 0; row < 2; ++row) {
-            dst[r1*ne0 + r2*ne0*ne1 + first_row + row] = sumf1[row];
+            dst[r1*ne0 + im*ne0*ne1 + first_row + row] = sumf1[row];
         }
     }
 }
 #else
-kernel void kernel_mul_mv_q3_K_f32(
+void kernel_mul_mv_q3_K_f32_impl(
         device const  void * src0,
         device const float * src1,
         device       float * dst,
         constant   int64_t & ne00,
-        constant   int64_t & ne01[[buffer(4)]],
-        constant   int64_t & ne02[[buffer(5)]],
-        constant   int64_t & ne10[[buffer(9)]],
-        constant   int64_t & ne12[[buffer(11)]],
-        constant   int64_t & ne0[[buffer(15)]],
-        constant   int64_t & ne1[[buffer(16)]],
-        constant   uint    & gqa[[buffer(17)]],
+        constant   int64_t & ne01,
+        constant   int64_t & ne02,
+        constant   int64_t & ne10,
+        constant   int64_t & ne12,
+        constant   int64_t & ne0,
+        constant   int64_t & ne1,
+        constant   uint    & r2,
+        constant   uint    & r3,
         uint3 tgpig[[threadgroup_position_in_grid]],
-        uint tiisg[[thread_index_in_simdgroup]],
-        uint sgitg[[simdgroup_index_in_threadgroup]]) {
+        uint  tiisg[[thread_index_in_simdgroup]],
+        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
 
     const int nb = ne00/QK_K;
 
     const int64_t r0 = tgpig.x;
     const int64_t r1 = tgpig.y;
-    const int64_t r2 = tgpig.z;
+    const int64_t im = tgpig.z;
 
     const int row = 2 * r0 + sgitg;
-    const uint offset0 = r2/gqa*(nb*ne0);
+
+    const uint i12 = im%ne12;
+    const uint i13 = im/ne12;
+
+    const uint offset0 = (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
+
     device const block_q3_K * x = (device const block_q3_K *) src0 + row*nb + offset0;
-    device const float     * yy = (device const float      *) src1 + r1*ne10 + r2*ne00*ne1;
+    device const float     * yy = (device const float      *) src1 + r1*ne10 + im*ne00*ne1;
+
     const int ix = tiisg/4;
     const int il = 4 * (tiisg%4);// 0, 4, 8, 12
-    const int im = il/8;         // 0, 0, 1, 1
+    const int iq = il/8;         // 0, 0, 1, 1
     const int in = il%8;         // 0, 4, 0, 4
 
     float2 sum = {0.f, 0.f};
@@ -1952,7 +2813,7 @@ kernel void kernel_mul_mv_q3_K_f32(
         const float d4 = d_all * ((int32_t)(s[0] & 0xF000) - 32768) * 1.f/262144.f;
 
         for (int l = 0; l < 4; l += 2) {
-            const uint16_t hm = h[l/2] >> im;
+            const uint16_t hm = h[l/2] >> iq;
             sum[0] += y[l+ 0] * d1 * ((int32_t)(q[l/2] & 0x0003) - ((hm & 0x0001) ? 0 :  4))
                     + y[l+16] * d2 * ((int32_t)(q[l/2] & 0x000c) - ((hm & 0x0004) ? 0 : 16))
                     + y[l+32] * d3 * ((int32_t)(q[l/2] & 0x0030) - ((hm & 0x0010) ? 0 : 64))
@@ -1968,28 +2829,50 @@ kernel void kernel_mul_mv_q3_K_f32(
 
     const float tot = simd_sum(sumf);
     if (tiisg == 0) {
-        dst[r1*ne0 + r2*ne0*ne1 + row] = tot;
+        dst[r1*ne0 + im*ne0*ne1 + row] = tot;
     }
 
 }
 #endif
 
-#if QK_K == 256
-kernel void kernel_mul_mv_q4_K_f32(
+[[host_name("kernel_mul_mv_q3_K_f32")]]
+kernel void kernel_mul_mv_q3_K_f32(
         device const  void * src0,
         device const float * src1,
         device       float * dst,
         constant   int64_t & ne00,
-        constant   int64_t & ne01 [[buffer(4)]],
-        constant   int64_t & ne02 [[buffer(5)]],
-        constant   int64_t & ne10 [[buffer(9)]],
-        constant   int64_t & ne12 [[buffer(11)]],
-        constant   int64_t & ne0  [[buffer(15)]],
-        constant   int64_t & ne1  [[buffer(16)]],
-        constant   uint    & gqa  [[buffer(17)]],
+        constant   int64_t & ne01[[buffer(4)]],
+        constant   int64_t & ne02[[buffer(5)]],
+        constant   int64_t & ne10[[buffer(9)]],
+        constant   int64_t & ne12[[buffer(11)]],
+        constant   int64_t & ne0 [[buffer(15)]],
+        constant   int64_t & ne1 [[buffer(16)]],
+        constant   uint    & r2  [[buffer(17)]],
+        constant   uint    & r3  [[buffer(18)]],
         uint3 tgpig[[threadgroup_position_in_grid]],
-        uint tiisg[[thread_index_in_simdgroup]],
-        uint sgitg[[simdgroup_index_in_threadgroup]]) {
+        uint  tiisg[[thread_index_in_simdgroup]],
+        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
+
+    kernel_mul_mv_q3_K_f32_impl(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3, tgpig, tiisg, sgitg);
+}
+
+#if QK_K == 256
+void kernel_mul_mv_q4_K_f32_impl(
+        device const  void * src0,
+        device const float * src1,
+        device       float * dst,
+        constant   int64_t & ne00,
+        constant   int64_t & ne01,
+        constant   int64_t & ne02,
+        constant   int64_t & ne10,
+        constant   int64_t & ne12,
+        constant   int64_t & ne0,
+        constant   int64_t & ne1,
+        constant   uint    & r2,
+        constant   uint    & r3,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint  tiisg[[thread_index_in_simdgroup]],
+        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
 
     const uint16_t kmask1 = 0x3f3f;
     const uint16_t kmask2 = 0x0f0f;
@@ -1997,26 +2880,32 @@ kernel void kernel_mul_mv_q4_K_f32(
 
     const int ix = tiisg/8;  // 0...3
     const int it = tiisg%8;  // 0...7
-    const int im = it/4;     // 0 or 1
+    const int iq = it/4;     // 0 or 1
     const int ir = it%4;     // 0...3
 
     const int nb = ne00/QK_K;
     const int r0 = tgpig.x;
     const int r1 = tgpig.y;
-    const int r2 = tgpig.z;
+    const int im = tgpig.z;
     //const int first_row = (r0 * N_SIMDGROUP + sgitg) * N_DST;
     const int first_row = r0 * N_DST;
     const int ib_row = first_row * nb;
-    const uint offset0 = r2/gqa*(nb*ne0);
+
+    const uint i12 = im%ne12;
+    const uint i13 = im/ne12;
+
+    const uint offset0 = (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
+
     device const block_q4_K * x = (device const block_q4_K *) src0 + ib_row + offset0;
-    device const float      * y = (device const float      *) src1 + r1*ne10 + r2*ne00*ne1;
+    device const float      * y = (device const float      *) src1 + r1*ne10 + im*ne00*ne1;
+
     float yl[16];
     float yh[16];
     float sumf[N_DST]={0.f}, all_sum;
 
     const int step = sizeof(block_q4_K) * nb / 2;
 
-    device const float * y4 = y + ix * QK_K + 64 * im + 8 * ir;
+    device const float * y4 = y + ix * QK_K + 64 * iq + 8 * ir;
 
     uint16_t sc16[4];
     thread const uint8_t * sc8 = (thread const uint8_t *)sc16;
@@ -2031,8 +2920,8 @@ kernel void kernel_mul_mv_q4_K_f32(
             yh[i+8] = y4[i+160]; sumy[3] += yh[i+8];
         }
 
-        device const uint16_t * sc = (device const uint16_t *)x[ib].scales + im;
-        device const uint16_t * q1 = (device const uint16_t *)x[ib].qs + 16 * im + 4 * ir;
+        device const uint16_t * sc = (device const uint16_t *)x[ib].scales + iq;
+        device const uint16_t * q1 = (device const uint16_t *)x[ib].qs + 16 * iq + 4 * ir;
         device const half     * dh = &x[ib].d;
 
         for (int row = 0; row < N_DST; row++) {
@@ -2076,23 +2965,24 @@ kernel void kernel_mul_mv_q4_K_f32(
     for (int row = 0; row < N_DST; ++row) {
         all_sum = simd_sum(sumf[row]);
         if (tiisg == 0) {
-            dst[r1*ne0 + r2*ne0*ne1 + first_row + row] = all_sum;
+            dst[r1*ne0 + im*ne0*ne1 + first_row + row] = all_sum;
         }
     }
 }
 #else
-kernel void kernel_mul_mv_q4_K_f32(
+void kernel_mul_mv_q4_K_f32_impl(
         device const  void * src0,
         device const float * src1,
         device       float * dst,
         constant   int64_t & ne00,
-        constant   int64_t & ne01[[buffer(4)]],
-        constant   int64_t & ne02[[buffer(5)]],
-        constant   int64_t & ne10[[buffer(9)]],
-        constant   int64_t & ne12[[buffer(11)]],
-        constant   int64_t & ne0[[buffer(15)]],
-        constant   int64_t & ne1[[buffer(16)]],
-        constant   uint    & gqa[[buffer(17)]],
+        constant   int64_t & ne01,
+        constant   int64_t & ne02,
+        constant   int64_t & ne10,
+        constant   int64_t & ne12,
+        constant   int64_t & ne0,
+        constant   int64_t & ne1,
+        constant   uint    & r2,
+        constant   uint    & r3,
         uint3 tgpig[[threadgroup_position_in_grid]],
         uint tiisg[[thread_index_in_simdgroup]],
         uint sgitg[[simdgroup_index_in_threadgroup]]) {
@@ -2103,12 +2993,18 @@ kernel void kernel_mul_mv_q4_K_f32(
     const int nb = ne00/QK_K;
     const int r0 = tgpig.x;
     const int r1 = tgpig.y;
-    const int r2 = tgpig.z;
+    const int im = tgpig.z;
     const int first_row = (r0 * N_SIMDGROUP + sgitg) * N_DST;
     const int ib_row = first_row * nb;
-    const uint offset0 = r2/gqa*(nb*ne0);
+
+    const uint i12 = im%ne12;
+    const uint i13 = im/ne12;
+
+    const uint offset0 = (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
+
     device const block_q4_K * x = (device const block_q4_K *) src0 + ib_row + offset0;
-    device const float      * y = (device const float      *) src1 + r1*ne10 + r2*ne00*ne1;
+    device const float      * y = (device const float      *) src1 + r1*ne10 + im*ne00*ne1;
+
     float yl[8];
     float yh[8];
     float sumf[N_DST]={0.f}, all_sum;
@@ -2164,13 +3060,14 @@ kernel void kernel_mul_mv_q4_K_f32(
     for (int row = 0; row < N_DST; ++row) {
         all_sum = simd_sum(sumf[row]);
         if (tiisg == 0) {
-            dst[r1*ne0+ r2*ne0*ne1 + first_row + row] = all_sum;
+            dst[r1*ne0+ im*ne0*ne1 + first_row + row] = all_sum;
         }
     }
 }
 #endif
 
-kernel void kernel_mul_mv_q5_K_f32(
+[[host_name("kernel_mul_mv_q4_K_f32")]]
+kernel void kernel_mul_mv_q4_K_f32(
         device const  void * src0,
         device const float * src1,
         device       float * dst,
@@ -2179,23 +3076,49 @@ kernel void kernel_mul_mv_q5_K_f32(
         constant   int64_t & ne02[[buffer(5)]],
         constant   int64_t & ne10[[buffer(9)]],
         constant   int64_t & ne12[[buffer(11)]],
-        constant   int64_t & ne0[[buffer(15)]],
-        constant   int64_t & ne1[[buffer(16)]],
-        constant   uint    & gqa[[buffer(17)]],
+        constant   int64_t & ne0 [[buffer(15)]],
+        constant   int64_t & ne1 [[buffer(16)]],
+        constant   uint    & r2  [[buffer(17)]],
+        constant   uint    & r3  [[buffer(18)]],
         uint3 tgpig[[threadgroup_position_in_grid]],
         uint tiisg[[thread_index_in_simdgroup]],
         uint sgitg[[simdgroup_index_in_threadgroup]]) {
 
+    kernel_mul_mv_q4_K_f32_impl(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3, tgpig, tiisg, sgitg);
+}
+
+void kernel_mul_mv_q5_K_f32_impl(
+        device const  void * src0,
+        device const float * src1,
+        device       float * dst,
+        constant   int64_t & ne00,
+        constant   int64_t & ne01,
+        constant   int64_t & ne02,
+        constant   int64_t & ne10,
+        constant   int64_t & ne12,
+        constant   int64_t & ne0,
+        constant   int64_t & ne1,
+        constant   uint    & r2,
+        constant   uint    & r3,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint  tiisg[[thread_index_in_simdgroup]],
+        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
+
     const int nb = ne00/QK_K;
 
     const int64_t r0 = tgpig.x;
     const int64_t r1 = tgpig.y;
-    const int r2 = tgpig.z;
+    const int im = tgpig.z;
 
     const int first_row = (r0 * N_SIMDGROUP + sgitg) * 2;
-    const uint offset0 = r2/gqa*(nb*ne0);
+
+    const uint i12 = im%ne12;
+    const uint i13 = im/ne12;
+
+    const uint offset0 = (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
+
     device const block_q5_K * x = (device const block_q5_K *) src0 + first_row*nb + offset0;
-    device const float     * yy = (device const float      *) src1 + r1*ne10 + r2*ne00*ne1;
+    device const float     * yy = (device const float      *) src1 + r1*ne10 + im*ne00*ne1;
 
     float sumf[2]={0.f};
 
@@ -2211,15 +3134,15 @@ kernel void kernel_mul_mv_q5_K_f32(
 
     const int tid = tiisg/4;
     const int ix  = tiisg%4;
-    const int im  = tid/4;
+    const int iq  = tid/4;
     const int ir  = tid%4;
     const int n   = 8;
 
     const int l0 = n*ir;
-    const int q_offset = 32*im + l0;
-    const int y_offset = 64*im + l0;
+    const int q_offset = 32*iq + l0;
+    const int y_offset = 64*iq + l0;
 
-    const uint8_t hm1 = 1u << (2*im);
+    const uint8_t hm1 = 1u << (2*iq);
     const uint8_t hm2 = hm1 << 1;
     const uint8_t hm3 = hm1 << 4;
     const uint8_t hm4 = hm2 << 4;
@@ -2234,7 +3157,7 @@ kernel void kernel_mul_mv_q5_K_f32(
         device const uint8_t * q1 = x[i].qs + q_offset;
         device const uint8_t * qh = x[i].qh + l0;
         device const half * dh = &x[i].d;
-        device const uint16_t * a = (device const uint16_t *)x[i].scales + im;
+        device const uint16_t * a = (device const uint16_t *)x[i].scales + iq;
 
         device const float * y2 = y1 + 128;
         float4 sumy = {0.f, 0.f, 0.f, 0.f};
@@ -2290,7 +3213,7 @@ kernel void kernel_mul_mv_q5_K_f32(
 
     const int il = 4 * (tiisg/8);  // 0, 4, 8, 12
     const int ix = tiisg%8;
-    const int im = il/8;         // 0, 0, 1, 1
+    const int iq = il/8;         // 0, 0, 1, 1
     const int in = il%8;         // 0, 4, 0, 4
 
     device const float * y = yy + ix*QK_K + il;
@@ -2315,7 +3238,7 @@ kernel void kernel_mul_mv_q5_K_f32(
 
             float2 acc = {0.f, 0.f};
             for (int l = 0; l < 4; ++l) {
-                const uint8_t hl = h[l] >> im;
+                const uint8_t hl = h[l] >> iq;
                 acc[0] += yl[l+0] * s[0] * ((int16_t)(q[l+ 0] & 0x0F) - (hl & 0x01 ? 0 : 16))
                         + yl[l+4] * s[1] * ((int16_t)(q[l+16] & 0x0F) - (hl & 0x04 ? 0 : 16));
                 acc[1] += yh[l+0] * s[2] * ((int16_t)(q[l+ 0] & 0xF0) - (hl & 0x10 ? 0 : 256))
@@ -2337,13 +3260,13 @@ kernel void kernel_mul_mv_q5_K_f32(
     for (int row = 0; row < 2; ++row) {
         const float tot = simd_sum(sumf[row]);
         if (tiisg == 0) {
-            dst[r1*ne0 + r2*ne0*ne1 + first_row + row] = tot;
+            dst[r1*ne0 + im*ne0*ne1 + first_row + row] = tot;
         }
     }
-
 }
 
-kernel void kernel_mul_mv_q6_K_f32(
+[[host_name("kernel_mul_mv_q5_K_f32")]]
+kernel void kernel_mul_mv_q5_K_f32(
         device const  void * src0,
         device const float * src1,
         device       float * dst,
@@ -2352,12 +3275,33 @@ kernel void kernel_mul_mv_q6_K_f32(
         constant   int64_t & ne02[[buffer(5)]],
         constant   int64_t & ne10[[buffer(9)]],
         constant   int64_t & ne12[[buffer(11)]],
-        constant   int64_t & ne0[[buffer(15)]],
-        constant   int64_t & ne1[[buffer(16)]],
-        constant   uint    & gqa[[buffer(17)]],
+        constant   int64_t & ne0 [[buffer(15)]],
+        constant   int64_t & ne1 [[buffer(16)]],
+        constant   uint    & r2  [[buffer(17)]],
+        constant   uint    & r3  [[buffer(18)]],
         uint3 tgpig[[threadgroup_position_in_grid]],
-        uint tiisg[[thread_index_in_simdgroup]],
-        uint sgitg[[simdgroup_index_in_threadgroup]]) {
+        uint  tiisg[[thread_index_in_simdgroup]],
+        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
+
+    kernel_mul_mv_q5_K_f32_impl(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3, tgpig, tiisg, sgitg);
+}
+
+void kernel_mul_mv_q6_K_f32_impl(
+        device const  void * src0,
+        device const float * src1,
+        device       float * dst,
+        constant   int64_t & ne00,
+        constant   int64_t & ne01,
+        constant   int64_t & ne02,
+        constant   int64_t & ne10,
+        constant   int64_t & ne12,
+        constant   int64_t & ne0,
+        constant   int64_t & ne1,
+        constant   uint    & r2,
+        constant   uint    & r3,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint  tiisg[[thread_index_in_simdgroup]],
+        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
 
     const uint8_t kmask1 = 0x03;
     const uint8_t kmask2 = 0x0C;
@@ -2368,12 +3312,17 @@ kernel void kernel_mul_mv_q6_K_f32(
 
     const int64_t r0 = tgpig.x;
     const int64_t r1 = tgpig.y;
-    const int r2 = tgpig.z;
+    const int     im = tgpig.z;
 
     const int row = 2 * r0 + sgitg;
-    const uint offset0 = r2/gqa*(nb*ne0);
+
+    const uint i12 = im%ne12;
+    const uint i13 = im/ne12;
+
+    const uint offset0 = (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
+
     device const block_q6_K * x = (device const block_q6_K *) src0 + row * nb + offset0;
-    device const float     * yy = (device const float      *) src1 + r1*ne10 + r2*ne00*ne1;
+    device const float     * yy = (device const float      *) src1 + r1*ne10 + im*ne00*ne1;
 
     float sumf = 0;
 
@@ -2439,10 +3388,31 @@ kernel void kernel_mul_mv_q6_K_f32(
 
     const float tot = simd_sum(sumf);
     if (tiisg == 0) {
-        dst[r1*ne0 + r2*ne0*ne1 + row] = tot;
+        dst[r1*ne0 + im*ne0*ne1 + row] = tot;
     }
 }
 
+[[host_name("kernel_mul_mv_q6_K_f32")]]
+kernel void kernel_mul_mv_q6_K_f32(
+        device const  void * src0,
+        device const float * src1,
+        device       float * dst,
+        constant   int64_t & ne00,
+        constant   int64_t & ne01[[buffer(4)]],
+        constant   int64_t & ne02[[buffer(5)]],
+        constant   int64_t & ne10[[buffer(9)]],
+        constant   int64_t & ne12[[buffer(11)]],
+        constant   int64_t & ne0 [[buffer(15)]],
+        constant   int64_t & ne1 [[buffer(16)]],
+        constant   uint    & r2  [[buffer(17)]],
+        constant   uint    & r3  [[buffer(18)]],
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint  tiisg[[thread_index_in_simdgroup]],
+        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
+
+    kernel_mul_mv_q6_K_f32_impl(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3, tgpig, tiisg, sgitg);
+}
+
 //============================= templates and their specializations =============================
 
 // NOTE: this is not dequantizing - we are simply fitting the template
@@ -2560,10 +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
@@ -2633,10 +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;
@@ -2663,13 +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;
     }
@@ -2717,22 +3687,90 @@ void dequantize_q6_K(device const block_q6_K *xb, short il, thread type4x4 & reg
 template<typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread float4x4 &)>
 kernel void kernel_get_rows(
         device const  void * src0,
-        device const   int * src1,
+        device const  char * src1,
         device       float * dst,
         constant   int64_t & ne00,
         constant  uint64_t & nb01,
+        constant  uint64_t & nb02,
+        constant   int64_t & ne10,
+        constant  uint64_t & nb10,
+        constant  uint64_t & nb11,
         constant  uint64_t & nb1,
-        uint                 tgpig[[threadgroup_position_in_grid]],
+        constant  uint64_t & nb2,
+        uint3                tgpig[[threadgroup_position_in_grid]],
         uint                 tiitg[[thread_index_in_threadgroup]],
-        uint                 tptg[[threads_per_threadgroup]]) {
-    const int i = tgpig;
-    const int r = ((device int32_t *) src1)[i];
+        uint3                tptg [[threads_per_threadgroup]]) {
+    //const int64_t i = tgpig;
+    //const int64_t r = ((device int32_t *) src1)[i];
 
-    for (int ind = tiitg; ind < ne00/16; ind += tptg) {
+    const int64_t i10 = tgpig.x;
+    const int64_t i11 = tgpig.y;
+
+    const int64_t r = ((device int32_t *) ((device char *) src1 + i11*nb11 + i10*nb10))[0];
+
+    const int64_t i02 = i11;
+
+    for (int64_t ind = tiitg; ind < ne00/16; ind += tptg.x) {
         float4x4 temp;
         dequantize_func(
-            ((device const block_q *) ((device char *) src0 + r*nb01)) + ind/nl, ind%nl, temp);
-        *(((device float4x4 *) ((device char *) dst + i*nb1)) + ind) = temp;
+            ((device const block_q *) ((device char *) src0 + r*nb01 + i02*nb02)) + ind/nl, ind%nl, temp);
+        *(((device float4x4 *) ((device char *) dst + i11*nb2 + i10*nb1)) + ind) = temp;
+    }
+}
+
+kernel void kernel_get_rows_f32(
+        device const  void * src0,
+        device const  char * src1,
+        device       float * dst,
+        constant   int64_t & ne00,
+        constant  uint64_t & nb01,
+        constant  uint64_t & nb02,
+        constant   int64_t & ne10,
+        constant  uint64_t & nb10,
+        constant  uint64_t & nb11,
+        constant  uint64_t & nb1,
+        constant  uint64_t & nb2,
+        uint3                tgpig[[threadgroup_position_in_grid]],
+        uint                 tiitg[[thread_index_in_threadgroup]],
+        uint3                tptg [[threads_per_threadgroup]]) {
+    const int64_t i10 = tgpig.x;
+    const int64_t i11 = tgpig.y;
+
+    const int64_t r = ((device int32_t *) ((device char *) src1 + i11*nb11 + i10*nb10))[0];
+
+    const int64_t i02 = i11;
+
+    for (int ind = tiitg; ind < ne00; ind += tptg.x) {
+        ((device float *) ((device char *) dst + i11*nb2 + i10*nb1))[ind] =
+            ((device float *) ((device char *) src0 + r*nb01 + i02*nb02))[ind];
+    }
+}
+
+kernel void kernel_get_rows_f16(
+        device const  void * src0,
+        device const  char * src1,
+        device       float * dst,
+        constant   int64_t & ne00,
+        constant  uint64_t & nb01,
+        constant  uint64_t & nb02,
+        constant   int64_t & ne10,
+        constant  uint64_t & nb10,
+        constant  uint64_t & nb11,
+        constant  uint64_t & nb1,
+        constant  uint64_t & nb2,
+        uint3                tgpig[[threadgroup_position_in_grid]],
+        uint                 tiitg[[thread_index_in_threadgroup]],
+        uint3                tptg [[threads_per_threadgroup]]) {
+    const int64_t i10 = tgpig.x;
+    const int64_t i11 = tgpig.y;
+
+    const int64_t r = ((device int32_t *) ((device char *) src1 + i11*nb11 + i10*nb10))[0];
+
+    const int64_t i02 = i11;
+
+    for (int ind = tiitg; ind < ne00; ind += tptg.x) {
+        ((device float *) ((device char *) dst + i11*nb2 + i10*nb1))[ind] =
+            ((device half *) ((device char *) src0 + r*nb01 + i02*nb02))[ind];
     }
 }
 
@@ -2749,24 +3787,25 @@ kernel void kernel_get_rows(
 
 // each block_q contains 16*nl weights
 template<typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread half4x4 &)>
-kernel void kernel_mul_mm(device const  uchar * src0,
-                          device const  uchar * src1,
-                          device        float * dst,
-                          constant    int64_t & ne00,
-                          constant    int64_t & ne02,
-                          constant    int64_t & nb01,
-                          constant    int64_t & nb02,
-                          constant    int64_t & ne12,
-                          constant    int64_t & nb10,
-                          constant    int64_t & nb11,
-                          constant    int64_t & nb12,
-                          constant    int64_t & ne0,
-                          constant    int64_t & ne1,
-                          constant       uint & gqa,
-                          threadgroup   uchar * shared_memory [[threadgroup(0)]],
-                          uint3                 tgpig[[threadgroup_position_in_grid]],
-                          uint                  tiitg[[thread_index_in_threadgroup]],
-                          uint                  sgitg[[simdgroup_index_in_threadgroup]]) {
+void kernel_mul_mm_impl(device const  uchar * src0,
+                        device const  uchar * src1,
+                        device        float * dst,
+                        constant    int64_t & ne00,
+                        constant    int64_t & ne02,
+                        constant    int64_t & nb01,
+                        constant    int64_t & nb02,
+                        constant    int64_t & ne12,
+                        constant    int64_t & nb10,
+                        constant    int64_t & nb11,
+                        constant    int64_t & nb12,
+                        constant    int64_t & ne0,
+                        constant    int64_t & ne1,
+                        constant       uint & r2,
+                        constant       uint & r3,
+                        threadgroup   uchar * shared_memory [[threadgroup(0)]],
+                        uint3                 tgpig[[threadgroup_position_in_grid]],
+                        uint                  tiitg[[thread_index_in_threadgroup]],
+                        uint                  sgitg[[simdgroup_index_in_threadgroup]]) {
 
     threadgroup half  * sa = (threadgroup half  *)(shared_memory);
     threadgroup float * sb = (threadgroup float *)(shared_memory + 4096);
@@ -2792,7 +3831,10 @@ kernel void kernel_mul_mm(device const  uchar * src0,
 
     short il = (tiitg % THREAD_PER_ROW);
 
-    uint   offset0 = im/gqa*nb02;
+    const uint i12 = im%ne12;
+    const uint i13 = im/ne12;
+
+    uint   offset0 = (i12/r2)*nb02 + (i13/r3)*(nb02*ne02);
     ushort offset1 = il/nl;
 
     device const block_q * x = (device const block_q *)(src0 + (r0 * BLOCK_SIZE_M + thread_row) * nb01 + offset0) + offset1;
@@ -2876,17 +3918,137 @@ kernel void kernel_mul_mm(device const  uchar * src0,
     }
 }
 
+template<typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread half4x4 &)>
+kernel void kernel_mul_mm(device const  uchar * src0,
+                          device const  uchar * src1,
+                          device        float * dst,
+                          constant    int64_t & ne00,
+                          constant    int64_t & ne02,
+                          constant    int64_t & nb01,
+                          constant    int64_t & nb02,
+                          constant    int64_t & ne12,
+                          constant    int64_t & nb10,
+                          constant    int64_t & nb11,
+                          constant    int64_t & nb12,
+                          constant    int64_t & ne0,
+                          constant    int64_t & ne1,
+                          constant       uint & r2,
+                          constant       uint & r3,
+                          threadgroup   uchar * shared_memory [[threadgroup(0)]],
+                          uint3                 tgpig[[threadgroup_position_in_grid]],
+                          uint                  tiitg[[thread_index_in_threadgroup]],
+                          uint                  sgitg[[simdgroup_index_in_threadgroup]]) {
+    kernel_mul_mm_impl<block_q, nl, dequantize_func>(
+        src0,
+        src1,
+        dst,
+        ne00,
+        ne02,
+        nb01,
+        nb02,
+        ne12,
+        nb10,
+        nb11,
+        nb12,
+        ne0,
+        ne1,
+        r2,
+        r3,
+        shared_memory,
+        tgpig,
+        tiitg,
+        sgitg);
+}
+
+template<typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread half4x4 &)>
+kernel void kernel_mul_mm_id(
+        device const   uchar * ids,
+        device const   uchar * src1,
+        device         uchar * dst,
+        constant     int64_t & nbi1,
+        constant     int64_t & ne00,
+        constant     int64_t & ne02,
+        constant     int64_t & nb01,
+        constant     int64_t & nb02,
+        constant     int64_t & ne12,
+        constant     int64_t & ne13,
+        constant     int64_t & nb10,
+        constant     int64_t & nb11,
+        constant     int64_t & nb12,
+        constant     int64_t & ne0,
+        constant     int64_t & ne1,
+        constant     int64_t & nb1,
+        constant        uint & r2,
+        constant        uint & r3,
+        constant         int & idx,
+        device const   uchar * src00,
+        device const   uchar * src01,
+        device const   uchar * src02,
+        device const   uchar * src03,
+        device const   uchar * src04,
+        device const   uchar * src05,
+        device const   uchar * src06,
+        device const   uchar * src07,
+        threadgroup    uchar * shared_memory [[threadgroup(0)]],
+        uint3                  tgpig[[threadgroup_position_in_grid]],
+        uint                   tiitg[[thread_index_in_threadgroup]],
+        uint                   sgitg[[simdgroup_index_in_threadgroup]]) {
+    device const uchar * src0[8] = {src00, src01, src02, src03, src04, src05, src06, src07};
+
+    const int64_t bid = tgpig.z/(ne12*ne13);
+
+    tgpig.z = tgpig.z%(ne12*ne13);
+
+    const int32_t id = ((device int32_t *) (ids + bid*nbi1))[idx];
+
+    kernel_mul_mm_impl<block_q, nl, dequantize_func>(
+        src0[id],
+        src1 + bid*nb11,
+        (device float *) (dst + bid*nb1),
+        ne00,
+        ne02,
+        nb01,
+        nb02,
+        ne12,
+        nb10,
+        nb11,
+        nb12,
+        ne0,
+        ne1,
+        r2,
+        r3,
+        shared_memory,
+        tgpig,
+        tiitg,
+        sgitg);
+}
+
 #if QK_K == 256
 #define QK_NL 16
 #else
 #define QK_NL 4
 #endif
 
-typedef void (get_rows_t)(device const void *, device const int *, device float *, constant int64_t &, \
-                          constant uint64_t &, constant uint64_t &, uint, uint, uint);
+//
+// get rows
+//
 
-template [[host_name("kernel_get_rows_f32")]]  kernel get_rows_t kernel_get_rows<float4x4,   1, dequantize_f32>;
-template [[host_name("kernel_get_rows_f16")]]  kernel get_rows_t kernel_get_rows<half4x4,    1, dequantize_f16>;
+typedef void (get_rows_t)(
+        device const void * src0,
+        device const char * src1,
+        device      float * dst,
+        constant  int64_t & ne00,
+        constant uint64_t & nb01,
+        constant uint64_t & nb02,
+        constant  int64_t & ne10,
+        constant uint64_t & nb10,
+        constant uint64_t & nb11,
+        constant uint64_t & nb1,
+        constant uint64_t & nb2,
+        uint3, uint, uint3);
+
+//template [[host_name("kernel_get_rows_f32")]]  kernel get_rows_t kernel_get_rows<float4x4,   1, dequantize_f32>;
+//template [[host_name("kernel_get_rows_f16")]]  kernel get_rows_t kernel_get_rows<half4x4,    1, dequantize_f16>;
 template [[host_name("kernel_get_rows_q4_0")]] kernel get_rows_t kernel_get_rows<block_q4_0, 2, dequantize_q4_0>;
 template [[host_name("kernel_get_rows_q4_1")]] kernel get_rows_t kernel_get_rows<block_q4_1, 2, dequantize_q4_1>;
 template [[host_name("kernel_get_rows_q5_0")]] kernel get_rows_t kernel_get_rows<block_q5_0, 2, dequantize_q5_0>;
@@ -2898,6 +4060,10 @@ template [[host_name("kernel_get_rows_q4_K")]] kernel get_rows_t kernel_get_rows
 template [[host_name("kernel_get_rows_q5_K")]] kernel get_rows_t kernel_get_rows<block_q5_K, QK_NL, dequantize_q5_K>;
 template [[host_name("kernel_get_rows_q6_K")]] kernel get_rows_t kernel_get_rows<block_q6_K, QK_NL, dequantize_q6_K>;
 
+//
+// matrix-matrix multiplication
+//
+
 typedef void (mat_mm_t)(
         device const  uchar * src0,
         device const  uchar * src1,
@@ -2912,8 +4078,10 @@ typedef void (mat_mm_t)(
         constant    int64_t & nb12,
         constant    int64_t & ne0,
         constant    int64_t & ne1,
-        constant       uint & gqa,
-        threadgroup uchar *, uint3, uint, uint);
+        constant       uint & r2,
+        constant       uint & r3,
+        threadgroup   uchar *,
+        uint3, uint, uint);
 
 template [[host_name("kernel_mul_mm_f32_f32")]]  kernel mat_mm_t kernel_mul_mm<float4x4,   1,     dequantize_f32>;
 template [[host_name("kernel_mul_mm_f16_f32")]]  kernel mat_mm_t kernel_mul_mm<half4x4,    1,     dequantize_f16>;
@@ -2927,3 +4095,823 @@ template [[host_name("kernel_mul_mm_q3_K_f32")]] kernel mat_mm_t kernel_mul_mm<b
 template [[host_name("kernel_mul_mm_q4_K_f32")]] kernel mat_mm_t kernel_mul_mm<block_q4_K, QK_NL, dequantize_q4_K>;
 template [[host_name("kernel_mul_mm_q5_K_f32")]] kernel mat_mm_t kernel_mul_mm<block_q5_K, QK_NL, dequantize_q5_K>;
 template [[host_name("kernel_mul_mm_q6_K_f32")]] kernel mat_mm_t kernel_mul_mm<block_q6_K, QK_NL, dequantize_q6_K>;
+
+//
+// indirect matrix-matrix multiplication
+//
+
+typedef void (mat_mm_id_t)(
+        device const   uchar * ids,
+        device const   uchar * src1,
+        device         uchar * dst,
+        constant     int64_t & nbi1,
+        constant     int64_t & ne00,
+        constant     int64_t & ne02,
+        constant     int64_t & nb01,
+        constant     int64_t & nb02,
+        constant     int64_t & ne12,
+        constant     int64_t & ne13,
+        constant     int64_t & nb10,
+        constant     int64_t & nb11,
+        constant     int64_t & nb12,
+        constant     int64_t & ne0,
+        constant     int64_t & ne1,
+        constant     int64_t & nb1,
+        constant        uint & r2,
+        constant        uint & r3,
+        constant         int & idx,
+        device const   uchar * src00,
+        device const   uchar * src01,
+        device const   uchar * src02,
+        device const   uchar * src03,
+        device const   uchar * src04,
+        device const   uchar * src05,
+        device const   uchar * src06,
+        device const   uchar * src07,
+        threadgroup    uchar *,
+        uint3, uint, uint);
+
+template [[host_name("kernel_mul_mm_id_f32_f32")]]  kernel mat_mm_id_t kernel_mul_mm_id<float4x4,   1,     dequantize_f32>;
+template [[host_name("kernel_mul_mm_id_f16_f32")]]  kernel mat_mm_id_t kernel_mul_mm_id<half4x4,    1,     dequantize_f16>;
+template [[host_name("kernel_mul_mm_id_q4_0_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_q4_0, 2,     dequantize_q4_0>;
+template [[host_name("kernel_mul_mm_id_q4_1_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_q4_1, 2,     dequantize_q4_1>;
+template [[host_name("kernel_mul_mm_id_q5_0_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_q5_0, 2,     dequantize_q5_0>;
+template [[host_name("kernel_mul_mm_id_q5_1_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_q5_1, 2,     dequantize_q5_1>;
+template [[host_name("kernel_mul_mm_id_q8_0_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_q8_0, 2,     dequantize_q8_0>;
+template [[host_name("kernel_mul_mm_id_q2_K_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_q2_K, QK_NL, dequantize_q2_K>;
+template [[host_name("kernel_mul_mm_id_q3_K_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_q3_K, QK_NL, dequantize_q3_K>;
+template [[host_name("kernel_mul_mm_id_q4_K_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_q4_K, QK_NL, dequantize_q4_K>;
+template [[host_name("kernel_mul_mm_id_q5_K_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_q5_K, QK_NL, dequantize_q5_K>;
+template [[host_name("kernel_mul_mm_id_q6_K_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_q6_K, QK_NL, dequantize_q6_K>;
+
+//
+// matrix-vector multiplication
+//
+
+[[host_name("kernel_mul_mv_id_f32_f32")]]
+kernel void kernel_mul_mv_id_f32_f32(
+        device const    char * ids,
+        device const    char * src1,
+        device         uchar * dst,
+        constant     int64_t & nbi1,
+        constant     int64_t & ne00,
+        constant     int64_t & ne01,
+        constant     int64_t & ne02,
+        constant    uint64_t & nb00,
+        constant    uint64_t & nb01,
+        constant    uint64_t & nb02,
+        constant     int64_t & ne10,
+        constant     int64_t & ne11,
+        constant     int64_t & ne12,
+        constant     int64_t & ne13,
+        constant    uint64_t & nb10,
+        constant    uint64_t & nb11,
+        constant    uint64_t & nb12,
+        constant     int64_t & ne0,
+        constant     int64_t & ne1,
+        constant     int64_t & nb1,
+        constant        uint & r2,
+        constant        uint & r3,
+        constant         int & idx,
+        device const    char * src00,
+        device const    char * src01,
+        device const    char * src02,
+        device const    char * src03,
+        device const    char * src04,
+        device const    char * src05,
+        device const    char * src06,
+        device const    char * src07,
+        uint3                  tgpig[[threadgroup_position_in_grid]],
+        uint                   tiitg[[thread_index_in_threadgroup]],
+        uint                   tiisg[[thread_index_in_simdgroup]],
+        uint                   sgitg[[simdgroup_index_in_threadgroup]]) {
+    device const char * src0[8] = {src00, src01, src02, src03, src04, src05, src06, src07};
+
+    const int64_t bid = tgpig.z/(ne12*ne13);
+
+    tgpig.z = tgpig.z%(ne12*ne13);
+
+    const int32_t id = ((device int32_t *) (ids + bid*nbi1))[idx];
+
+    kernel_mul_mv_f32_f32_impl(
+        src0[id],
+        src1 + bid*nb11,
+        (device float *) (dst + bid*nb1),
+        ne00,
+        ne01,
+        ne02,
+        nb00,
+        nb01,
+        nb02,
+        ne10,
+        ne11,
+        ne12,
+        nb10,
+        nb11,
+        nb12,
+        ne0,
+        ne1,
+        r2,
+        r3,
+        tgpig,
+        tiisg);
+}
+
+[[host_name("kernel_mul_mv_id_f16_f32")]]
+kernel void kernel_mul_mv_id_f16_f32(
+        device const    char * ids,
+        device const    char * src1,
+        device         uchar * dst,
+        constant     int64_t & nbi1,
+        constant     int64_t & ne00,
+        constant     int64_t & ne01,
+        constant     int64_t & ne02,
+        constant    uint64_t & nb00,
+        constant    uint64_t & nb01,
+        constant    uint64_t & nb02,
+        constant     int64_t & ne10,
+        constant     int64_t & ne11,
+        constant     int64_t & ne12,
+        constant     int64_t & ne13,
+        constant    uint64_t & nb10,
+        constant    uint64_t & nb11,
+        constant    uint64_t & nb12,
+        constant     int64_t & ne0,
+        constant     int64_t & ne1,
+        constant     int64_t & nb1,
+        constant        uint & r2,
+        constant        uint & r3,
+        constant         int & idx,
+        device const    char * src00,
+        device const    char * src01,
+        device const    char * src02,
+        device const    char * src03,
+        device const    char * src04,
+        device const    char * src05,
+        device const    char * src06,
+        device const    char * src07,
+        uint3                  tgpig[[threadgroup_position_in_grid]],
+        uint                   tiitg[[thread_index_in_threadgroup]],
+        uint                   tiisg[[thread_index_in_simdgroup]],
+        uint                   sgitg[[simdgroup_index_in_threadgroup]]) {
+    device const char * src0[8] = {src00, src01, src02, src03, src04, src05, src06, src07};
+
+    const int64_t bid = tgpig.z/(ne12*ne13);
+
+    tgpig.z = tgpig.z%(ne12*ne13);
+
+    const int32_t id = ((device int32_t *) (ids + bid*nbi1))[idx];
+
+    kernel_mul_mv_f16_f32_impl(
+        src0[id],
+        src1 + bid*nb11,
+        (device float *) (dst + bid*nb1),
+        ne00,
+        ne01,
+        ne02,
+        nb00,
+        nb01,
+        nb02,
+        ne10,
+        ne11,
+        ne12,
+        nb10,
+        nb11,
+        nb12,
+        ne0,
+        ne1,
+        r2,
+        r3,
+        tgpig,
+        tiisg);
+}
+
+[[host_name("kernel_mul_mv_id_q8_0_f32")]]
+kernel void kernel_mul_mv_id_q8_0_f32(
+        device const    char * ids,
+        device const    char * src1,
+        device         uchar * dst,
+        constant     int64_t & nbi1,
+        constant     int64_t & ne00,
+        constant     int64_t & ne01,
+        constant     int64_t & ne02,
+        constant    uint64_t & nb00,
+        constant    uint64_t & nb01,
+        constant    uint64_t & nb02,
+        constant     int64_t & ne10,
+        constant     int64_t & ne11,
+        constant     int64_t & ne12,
+        constant     int64_t & ne13,
+        constant    uint64_t & nb10,
+        constant    uint64_t & nb11,
+        constant    uint64_t & nb12,
+        constant     int64_t & ne0,
+        constant     int64_t & ne1,
+        constant     int64_t & nb1,
+        constant        uint & r2,
+        constant        uint & r3,
+        constant         int & idx,
+        device const    char * src00,
+        device const    char * src01,
+        device const    char * src02,
+        device const    char * src03,
+        device const    char * src04,
+        device const    char * src05,
+        device const    char * src06,
+        device const    char * src07,
+        uint3                  tgpig[[threadgroup_position_in_grid]],
+        uint                   tiitg[[thread_index_in_threadgroup]],
+        uint                   tiisg[[thread_index_in_simdgroup]],
+        uint                   sgitg[[simdgroup_index_in_threadgroup]]) {
+    device const char * src0[8] = {src00, src01, src02, src03, src04, src05, src06, src07};
+
+    const int64_t bid = tgpig.z/(ne12*ne13);
+
+    tgpig.z = tgpig.z%(ne12*ne13);
+
+    const int32_t id = ((device int32_t *) (ids + bid*nbi1))[idx];
+
+    kernel_mul_mv_q8_0_f32_impl(
+        src0[id],
+        (device const float *) (src1 + bid*nb11),
+        (device       float *) ( dst + bid*nb1),
+        ne00,
+        ne01,
+        ne02,
+        ne10,
+        ne12,
+        ne0,
+        ne1,
+        r2,
+        r3,
+        tgpig,
+        tiisg,
+        sgitg);
+}
+
+[[host_name("kernel_mul_mv_id_q4_0_f32")]]
+kernel void kernel_mul_mv_id_q4_0_f32(
+        device const    char * ids,
+        device const    char * src1,
+        device         uchar * dst,
+        constant     int64_t & nbi1,
+        constant     int64_t & ne00,
+        constant     int64_t & ne01,
+        constant     int64_t & ne02,
+        constant    uint64_t & nb00,
+        constant    uint64_t & nb01,
+        constant    uint64_t & nb02,
+        constant     int64_t & ne10,
+        constant     int64_t & ne11,
+        constant     int64_t & ne12,
+        constant     int64_t & ne13,
+        constant    uint64_t & nb10,
+        constant    uint64_t & nb11,
+        constant    uint64_t & nb12,
+        constant     int64_t & ne0,
+        constant     int64_t & ne1,
+        constant     int64_t & nb1,
+        constant        uint & r2,
+        constant        uint & r3,
+        constant         int & idx,
+        device const    char * src00,
+        device const    char * src01,
+        device const    char * src02,
+        device const    char * src03,
+        device const    char * src04,
+        device const    char * src05,
+        device const    char * src06,
+        device const    char * src07,
+        uint3                  tgpig[[threadgroup_position_in_grid]],
+        uint                   tiitg[[thread_index_in_threadgroup]],
+        uint                   tiisg[[thread_index_in_simdgroup]],
+        uint                   sgitg[[simdgroup_index_in_threadgroup]]) {
+    device const char * src0[8] = {src00, src01, src02, src03, src04, src05, src06, src07};
+
+    const int64_t bid = tgpig.z/(ne12*ne13);
+
+    tgpig.z = tgpig.z%(ne12*ne13);
+
+    const int32_t id = ((device int32_t *) (ids + bid*nbi1))[idx];
+
+    mul_vec_q_n_f32_impl<block_q4_0, N_DST, N_SIMDGROUP, N_SIMDWIDTH>(
+        src0[id],
+        (device const float *) (src1 + bid*nb11),
+        (device       float *) ( dst + bid*nb1),
+        ne00,
+        ne01,
+        ne02,
+        ne10,
+        ne12,
+        ne0,
+        ne1,
+        r2,
+        r3,
+        tgpig,
+        tiisg,
+        sgitg);
+}
+
+[[host_name("kernel_mul_mv_id_q4_1_f32")]]
+kernel void kernel_mul_mv_id_q4_1_f32(
+        device const    char * ids,
+        device const    char * src1,
+        device         uchar * dst,
+        constant     int64_t & nbi1,
+        constant     int64_t & ne00,
+        constant     int64_t & ne01,
+        constant     int64_t & ne02,
+        constant    uint64_t & nb00,
+        constant    uint64_t & nb01,
+        constant    uint64_t & nb02,
+        constant     int64_t & ne10,
+        constant     int64_t & ne11,
+        constant     int64_t & ne12,
+        constant     int64_t & ne13,
+        constant    uint64_t & nb10,
+        constant    uint64_t & nb11,
+        constant    uint64_t & nb12,
+        constant     int64_t & ne0,
+        constant     int64_t & ne1,
+        constant     int64_t & nb1,
+        constant        uint & r2,
+        constant        uint & r3,
+        constant         int & idx,
+        device const    char * src00,
+        device const    char * src01,
+        device const    char * src02,
+        device const    char * src03,
+        device const    char * src04,
+        device const    char * src05,
+        device const    char * src06,
+        device const    char * src07,
+        uint3                  tgpig[[threadgroup_position_in_grid]],
+        uint                   tiitg[[thread_index_in_threadgroup]],
+        uint                   tiisg[[thread_index_in_simdgroup]],
+        uint                   sgitg[[simdgroup_index_in_threadgroup]]) {
+    device const char * src0[8] = {src00, src01, src02, src03, src04, src05, src06, src07};
+
+    const int64_t bid = tgpig.z/(ne12*ne13);
+
+    tgpig.z = tgpig.z%(ne12*ne13);
+
+    const int32_t id = ((device int32_t *) (ids + bid*nbi1))[idx];
+
+    mul_vec_q_n_f32_impl<block_q4_1, N_DST, N_SIMDGROUP, N_SIMDWIDTH>(
+        src0[id],
+        (device const float *) (src1 + bid*nb11),
+        (device       float *) ( dst + bid*nb1),
+        ne00,
+        ne01,
+        ne02,
+        ne10,
+        ne12,
+        ne0,
+        ne1,
+        r2,
+        r3,
+        tgpig,
+        tiisg,
+        sgitg);
+}
+
+[[host_name("kernel_mul_mv_id_q5_0_f32")]]
+kernel void kernel_mul_mv_id_q5_0_f32(
+        device const    char * ids,
+        device const    char * src1,
+        device         uchar * dst,
+        constant     int64_t & nbi1,
+        constant     int64_t & ne00,
+        constant     int64_t & ne01,
+        constant     int64_t & ne02,
+        constant    uint64_t & nb00,
+        constant    uint64_t & nb01,
+        constant    uint64_t & nb02,
+        constant     int64_t & ne10,
+        constant     int64_t & ne11,
+        constant     int64_t & ne12,
+        constant     int64_t & ne13,
+        constant    uint64_t & nb10,
+        constant    uint64_t & nb11,
+        constant    uint64_t & nb12,
+        constant     int64_t & ne0,
+        constant     int64_t & ne1,
+        constant     int64_t & nb1,
+        constant        uint & r2,
+        constant        uint & r3,
+        constant         int & idx,
+        device const    char * src00,
+        device const    char * src01,
+        device const    char * src02,
+        device const    char * src03,
+        device const    char * src04,
+        device const    char * src05,
+        device const    char * src06,
+        device const    char * src07,
+        uint3                  tgpig[[threadgroup_position_in_grid]],
+        uint                   tiitg[[thread_index_in_threadgroup]],
+        uint                   tiisg[[thread_index_in_simdgroup]],
+        uint                   sgitg[[simdgroup_index_in_threadgroup]]) {
+    device const char * src0[8] = {src00, src01, src02, src03, src04, src05, src06, src07};
+
+    const int64_t bid = tgpig.z/(ne12*ne13);
+
+    tgpig.z = tgpig.z%(ne12*ne13);
+
+    const int32_t id = ((device int32_t *) (ids + bid*nbi1))[idx];
+
+    mul_vec_q_n_f32_impl<block_q5_0, N_DST, N_SIMDGROUP, N_SIMDWIDTH>(
+        src0[id],
+        (device const float *) (src1 + bid*nb11),
+        (device       float *) ( dst + bid*nb1),
+        ne00,
+        ne01,
+        ne02,
+        ne10,
+        ne12,
+        ne0,
+        ne1,
+        r2,
+        r3,
+        tgpig,
+        tiisg,
+        sgitg);
+}
+
+[[host_name("kernel_mul_mv_id_q5_1_f32")]]
+kernel void kernel_mul_mv_id_q5_1_f32(
+        device const    char * ids,
+        device const    char * src1,
+        device         uchar * dst,
+        constant     int64_t & nbi1,
+        constant     int64_t & ne00,
+        constant     int64_t & ne01,
+        constant     int64_t & ne02,
+        constant    uint64_t & nb00,
+        constant    uint64_t & nb01,
+        constant    uint64_t & nb02,
+        constant     int64_t & ne10,
+        constant     int64_t & ne11,
+        constant     int64_t & ne12,
+        constant     int64_t & ne13,
+        constant    uint64_t & nb10,
+        constant    uint64_t & nb11,
+        constant    uint64_t & nb12,
+        constant     int64_t & ne0,
+        constant     int64_t & ne1,
+        constant     int64_t & nb1,
+        constant        uint & r2,
+        constant        uint & r3,
+        constant         int & idx,
+        device const    char * src00,
+        device const    char * src01,
+        device const    char * src02,
+        device const    char * src03,
+        device const    char * src04,
+        device const    char * src05,
+        device const    char * src06,
+        device const    char * src07,
+        uint3                  tgpig[[threadgroup_position_in_grid]],
+        uint                   tiitg[[thread_index_in_threadgroup]],
+        uint                   tiisg[[thread_index_in_simdgroup]],
+        uint                   sgitg[[simdgroup_index_in_threadgroup]]) {
+    device const char * src0[8] = {src00, src01, src02, src03, src04, src05, src06, src07};
+
+    const int64_t bid = tgpig.z/(ne12*ne13);
+
+    tgpig.z = tgpig.z%(ne12*ne13);
+
+    const int32_t id = ((device int32_t *) (ids + bid*nbi1))[idx];
+
+    mul_vec_q_n_f32_impl<block_q5_1, N_DST, N_SIMDGROUP, N_SIMDWIDTH>(
+        src0[id],
+        (device const float *) (src1 + bid*nb11),
+        (device       float *) ( dst + bid*nb1),
+        ne00,
+        ne01,
+        ne02,
+        ne10,
+        ne12,
+        ne0,
+        ne1,
+        r2,
+        r3,
+        tgpig,
+        tiisg,
+        sgitg);
+}
+
+[[host_name("kernel_mul_mv_id_q2_K_f32")]]
+kernel void kernel_mul_mv_id_q2_K_f32(
+        device const    char * ids,
+        device const    char * src1,
+        device         uchar * dst,
+        constant     int64_t & nbi1,
+        constant     int64_t & ne00,
+        constant     int64_t & ne01,
+        constant     int64_t & ne02,
+        constant    uint64_t & nb00,
+        constant    uint64_t & nb01,
+        constant    uint64_t & nb02,
+        constant     int64_t & ne10,
+        constant     int64_t & ne11,
+        constant     int64_t & ne12,
+        constant     int64_t & ne13,
+        constant    uint64_t & nb10,
+        constant    uint64_t & nb11,
+        constant    uint64_t & nb12,
+        constant     int64_t & ne0,
+        constant     int64_t & ne1,
+        constant     int64_t & nb1,
+        constant        uint & r2,
+        constant        uint & r3,
+        constant         int & idx,
+        device const    char * src00,
+        device const    char * src01,
+        device const    char * src02,
+        device const    char * src03,
+        device const    char * src04,
+        device const    char * src05,
+        device const    char * src06,
+        device const    char * src07,
+        uint3                  tgpig[[threadgroup_position_in_grid]],
+        uint                   tiitg[[thread_index_in_threadgroup]],
+        uint                   tiisg[[thread_index_in_simdgroup]],
+        uint                   sgitg[[simdgroup_index_in_threadgroup]]) {
+    device const char * src0[8] = {src00, src01, src02, src03, src04, src05, src06, src07};
+
+    const int64_t bid = tgpig.z/(ne12*ne13);
+
+    tgpig.z = tgpig.z%(ne12*ne13);
+
+    const int32_t id = ((device int32_t *) (ids + bid*nbi1))[idx];
+
+    kernel_mul_mv_q2_K_f32_impl(
+        src0[id],
+        (device const float *) (src1 + bid*nb11),
+        (device       float *) ( dst + bid*nb1),
+        ne00,
+        ne01,
+        ne02,
+        ne10,
+        ne12,
+        ne0,
+        ne1,
+        r2,
+        r3,
+        tgpig,
+        tiisg,
+        sgitg);
+}
+
+[[host_name("kernel_mul_mv_id_q3_K_f32")]]
+kernel void kernel_mul_mv_id_q3_K_f32(
+        device const    char * ids,
+        device const    char * src1,
+        device         uchar * dst,
+        constant     int64_t & nbi1,
+        constant     int64_t & ne00,
+        constant     int64_t & ne01,
+        constant     int64_t & ne02,
+        constant    uint64_t & nb00,
+        constant    uint64_t & nb01,
+        constant    uint64_t & nb02,
+        constant     int64_t & ne10,
+        constant     int64_t & ne11,
+        constant     int64_t & ne12,
+        constant     int64_t & ne13,
+        constant    uint64_t & nb10,
+        constant    uint64_t & nb11,
+        constant    uint64_t & nb12,
+        constant     int64_t & ne0,
+        constant     int64_t & ne1,
+        constant     int64_t & nb1,
+        constant        uint & r2,
+        constant        uint & r3,
+        constant         int & idx,
+        device const    char * src00,
+        device const    char * src01,
+        device const    char * src02,
+        device const    char * src03,
+        device const    char * src04,
+        device const    char * src05,
+        device const    char * src06,
+        device const    char * src07,
+        uint3                  tgpig[[threadgroup_position_in_grid]],
+        uint                   tiitg[[thread_index_in_threadgroup]],
+        uint                   tiisg[[thread_index_in_simdgroup]],
+        uint                   sgitg[[simdgroup_index_in_threadgroup]]) {
+    device const char * src0[8] = {src00, src01, src02, src03, src04, src05, src06, src07};
+
+    const int64_t bid = tgpig.z/(ne12*ne13);
+
+    tgpig.z = tgpig.z%(ne12*ne13);
+
+    const int32_t id = ((device int32_t *) (ids + bid*nbi1))[idx];
+
+    kernel_mul_mv_q3_K_f32_impl(
+        src0[id],
+        (device const float *) (src1 + bid*nb11),
+        (device       float *) ( dst + bid*nb1),
+        ne00,
+        ne01,
+        ne02,
+        ne10,
+        ne12,
+        ne0,
+        ne1,
+        r2,
+        r3,
+        tgpig,
+        tiisg,
+        sgitg);
+}
+
+[[host_name("kernel_mul_mv_id_q4_K_f32")]]
+kernel void kernel_mul_mv_id_q4_K_f32(
+        device const    char * ids,
+        device const    char * src1,
+        device         uchar * dst,
+        constant     int64_t & nbi1,
+        constant     int64_t & ne00,
+        constant     int64_t & ne01,
+        constant     int64_t & ne02,
+        constant    uint64_t & nb00,
+        constant    uint64_t & nb01,
+        constant    uint64_t & nb02,
+        constant     int64_t & ne10,
+        constant     int64_t & ne11,
+        constant     int64_t & ne12,
+        constant     int64_t & ne13,
+        constant    uint64_t & nb10,
+        constant    uint64_t & nb11,
+        constant    uint64_t & nb12,
+        constant     int64_t & ne0,
+        constant     int64_t & ne1,
+        constant     int64_t & nb1,
+        constant        uint & r2,
+        constant        uint & r3,
+        constant         int & idx,
+        device const    char * src00,
+        device const    char * src01,
+        device const    char * src02,
+        device const    char * src03,
+        device const    char * src04,
+        device const    char * src05,
+        device const    char * src06,
+        device const    char * src07,
+        uint3                  tgpig[[threadgroup_position_in_grid]],
+        uint                   tiitg[[thread_index_in_threadgroup]],
+        uint                   tiisg[[thread_index_in_simdgroup]],
+        uint                   sgitg[[simdgroup_index_in_threadgroup]]) {
+    device const char * src0[8] = {src00, src01, src02, src03, src04, src05, src06, src07};
+
+    const int64_t bid = tgpig.z/(ne12*ne13);
+
+    tgpig.z = tgpig.z%(ne12*ne13);
+
+    const int32_t id = ((device int32_t *) (ids + bid*nbi1))[idx];
+
+    kernel_mul_mv_q4_K_f32_impl(
+        src0[id],
+        (device const float *) (src1 + bid*nb11),
+        (device       float *) ( dst + bid*nb1),
+        ne00,
+        ne01,
+        ne02,
+        ne10,
+        ne12,
+        ne0,
+        ne1,
+        r2,
+        r3,
+        tgpig,
+        tiisg,
+        sgitg);
+}
+
+[[host_name("kernel_mul_mv_id_q5_K_f32")]]
+kernel void kernel_mul_mv_id_q5_K_f32(
+        device const    char * ids,
+        device const    char * src1,
+        device         uchar * dst,
+        constant     int64_t & nbi1,
+        constant     int64_t & ne00,
+        constant     int64_t & ne01,
+        constant     int64_t & ne02,
+        constant    uint64_t & nb00,
+        constant    uint64_t & nb01,
+        constant    uint64_t & nb02,
+        constant     int64_t & ne10,
+        constant     int64_t & ne11,
+        constant     int64_t & ne12,
+        constant     int64_t & ne13,
+        constant    uint64_t & nb10,
+        constant    uint64_t & nb11,
+        constant    uint64_t & nb12,
+        constant     int64_t & ne0,
+        constant     int64_t & ne1,
+        constant     int64_t & nb1,
+        constant        uint & r2,
+        constant        uint & r3,
+        constant         int & idx,
+        device const    char * src00,
+        device const    char * src01,
+        device const    char * src02,
+        device const    char * src03,
+        device const    char * src04,
+        device const    char * src05,
+        device const    char * src06,
+        device const    char * src07,
+        uint3                  tgpig[[threadgroup_position_in_grid]],
+        uint                   tiitg[[thread_index_in_threadgroup]],
+        uint                   tiisg[[thread_index_in_simdgroup]],
+        uint                   sgitg[[simdgroup_index_in_threadgroup]]) {
+    device const char * src0[8] = {src00, src01, src02, src03, src04, src05, src06, src07};
+
+    const int64_t bid = tgpig.z/(ne12*ne13);
+
+    tgpig.z = tgpig.z%(ne12*ne13);
+
+    const int32_t id = ((device int32_t *) (ids + bid*nbi1))[idx];
+
+    kernel_mul_mv_q5_K_f32_impl(
+        src0[id],
+        (device const float *) (src1 + bid*nb11),
+        (device       float *) ( dst + bid*nb1),
+        ne00,
+        ne01,
+        ne02,
+        ne10,
+        ne12,
+        ne0,
+        ne1,
+        r2,
+        r3,
+        tgpig,
+        tiisg,
+        sgitg);
+}
+
+[[host_name("kernel_mul_mv_id_q6_K_f32")]]
+kernel void kernel_mul_mv_id_q6_K_f32(
+        device const    char * ids,
+        device const    char * src1,
+        device         uchar * dst,
+        constant     int64_t & nbi1,
+        constant     int64_t & ne00,
+        constant     int64_t & ne01,
+        constant     int64_t & ne02,
+        constant    uint64_t & nb00,
+        constant    uint64_t & nb01,
+        constant    uint64_t & nb02,
+        constant     int64_t & ne10,
+        constant     int64_t & ne11,
+        constant     int64_t & ne12,
+        constant     int64_t & ne13,
+        constant    uint64_t & nb10,
+        constant    uint64_t & nb11,
+        constant    uint64_t & nb12,
+        constant     int64_t & ne0,
+        constant     int64_t & ne1,
+        constant     int64_t & nb1,
+        constant        uint & r2,
+        constant        uint & r3,
+        constant         int & idx,
+        device const    char * src00,
+        device const    char * src01,
+        device const    char * src02,
+        device const    char * src03,
+        device const    char * src04,
+        device const    char * src05,
+        device const    char * src06,
+        device const    char * src07,
+        uint3                  tgpig[[threadgroup_position_in_grid]],
+        uint                   tiitg[[thread_index_in_threadgroup]],
+        uint                   tiisg[[thread_index_in_simdgroup]],
+        uint                   sgitg[[simdgroup_index_in_threadgroup]]) {
+    device const char * src0[8] = {src00, src01, src02, src03, src04, src05, src06, src07};
+
+    const int64_t bid = tgpig.z/(ne12*ne13);
+
+    tgpig.z = tgpig.z%(ne12*ne13);
+
+    const int32_t id = ((device int32_t *) (ids + bid*nbi1))[idx];
+
+    kernel_mul_mv_q6_K_f32_impl(
+        src0[id],
+        (device const float *) (src1 + bid*nb11),
+        (device       float *) ( dst + bid*nb1),
+        ne00,
+        ne01,
+        ne02,
+        ne10,
+        ne12,
+        ne0,
+        ne1,
+        r2,
+        r3,
+        tgpig,
+        tiisg,
+        sgitg);
+}
diff --git a/ggml-quants.c b/ggml-quants.c
index 7285d5f7f..a15a24048 100644
--- a/ggml-quants.c
+++ b/ggml-quants.c
@@ -3114,7 +3114,7 @@ void ggml_vec_dot_q5_0_q8_0(const int n, float * restrict s, const void * restri
 
     size_t vl = __riscv_vsetvl_e8m1(qk/2);
 
-    // These tempory registers are for masking and shift operations
+    // These temporary registers are for masking and shift operations
     vuint32m2_t vt_1 = __riscv_vid_v_u32m2(vl);
     vuint32m2_t vt_2 = __riscv_vsll_vv_u32m2(__riscv_vmv_v_x_u32m2(1, vl), vt_1, vl);
 
@@ -3677,7 +3677,7 @@ void ggml_vec_dot_q2_K_q8_K(const int n, float * restrict s, const void * restri
 
         const uint8x16_t mins = vshrq_n_u8(mins_and_scales, 4);
         const ggml_int16x8x2_t q8sums = ggml_vld1q_s16_x2(y[i].bsums);
-        const ggml_int16x8x2_t mins16 = {vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(mins))), vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(mins)))};
+        const ggml_int16x8x2_t mins16 = {{vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(mins))), vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(mins)))}};
         const int32x4_t s0 = vaddq_s32(vmull_s16(vget_low_s16 (mins16.val[0]), vget_low_s16 (q8sums.val[0])),
                                        vmull_s16(vget_high_s16(mins16.val[0]), vget_high_s16(q8sums.val[0])));
         const int32x4_t s1 = vaddq_s32(vmull_s16(vget_low_s16 (mins16.val[1]), vget_low_s16 (q8sums.val[1])),
@@ -4757,7 +4757,7 @@ void ggml_vec_dot_q3_K_q8_K(const int n, float * restrict s, const void * restri
 
             vl = 16;
 
-            // retreive lane to multiply with scale
+            // retrieve lane to multiply with scale
             vint32m2_t aux0_0 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(a0, 0), (scale[0]), vl);
             vint32m2_t aux0_1 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(a0, 1), (scale[1]), vl);
             vint32m2_t aux1_0 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(a1, 0), (scale[2]), vl);
@@ -6626,7 +6626,7 @@ void ggml_vec_dot_q6_K_q8_K(const int n, float * restrict s, const void * restri
 
         const ggml_int16x8x2_t q8sums = ggml_vld1q_s16_x2(y[i].bsums);
         const int8x16_t scales = vld1q_s8(scale);
-        const ggml_int16x8x2_t q6scales = {vmovl_s8(vget_low_s8(scales)), vmovl_s8(vget_high_s8(scales))};
+        const ggml_int16x8x2_t q6scales = {{vmovl_s8(vget_low_s8(scales)), vmovl_s8(vget_high_s8(scales))}};
 
         const int32x4_t prod = vaddq_s32(vaddq_s32(vmull_s16(vget_low_s16 (q8sums.val[0]), vget_low_s16 (q6scales.val[0])),
                                                    vmull_s16(vget_high_s16(q8sums.val[0]), vget_high_s16(q6scales.val[0]))),
diff --git a/ggml.c b/ggml.c
index e2687ef4f..3656422d7 100644
--- a/ggml.c
+++ b/ggml.c
@@ -1,4 +1,4 @@
-#define _CRT_SECURE_NO_DEPRECATE // Disables ridiculous "unsafe" warnigns on Windows
+#define _CRT_SECURE_NO_DEPRECATE // Disables ridiculous "unsafe" warnings on Windows
 #define _USE_MATH_DEFINES // For M_PI on MSVC
 
 #include "ggml-impl.h"
@@ -33,7 +33,7 @@
 // we should just be careful :)
 #pragma warning(disable: 4244 4267)
 
-// disable POSIX deprecation warnigns
+// disable POSIX deprecation warnings
 // these functions are never going away, anyway
 #pragma warning(disable: 4996)
 #endif
@@ -233,24 +233,6 @@ inline static void * ggml_aligned_malloc(size_t size) {
 #define UNUSED GGML_UNUSED
 #define SWAP(x, y, T) do { T SWAP = x; x = y; y = SWAP; } while (0)
 
-//
-// tensor access macros
-//
-
-#define GGML_TENSOR_UNARY_OP_LOCALS \
-    GGML_TENSOR_LOCALS(int64_t, ne0, src0, ne) \
-    GGML_TENSOR_LOCALS(size_t,  nb0, src0, nb) \
-    GGML_TENSOR_LOCALS(int64_t, ne,  dst,  ne) \
-    GGML_TENSOR_LOCALS(size_t,  nb,  dst,  nb)
-
-#define GGML_TENSOR_BINARY_OP_LOCALS \
-    GGML_TENSOR_LOCALS(int64_t, ne0, src0, ne) \
-    GGML_TENSOR_LOCALS(size_t,  nb0, src0, nb) \
-    GGML_TENSOR_LOCALS(int64_t, ne1, src1, ne) \
-    GGML_TENSOR_LOCALS(size_t,  nb1, src1, nb) \
-    GGML_TENSOR_LOCALS(int64_t, ne,  dst,  ne) \
-    GGML_TENSOR_LOCALS(size_t,  nb,  dst,  nb)
-
 #if defined(GGML_USE_ACCELERATE)
 #include <Accelerate/Accelerate.h>
 #if defined(GGML_USE_CLBLAST) // allow usage of CLBlast alongside Accelerate functions
@@ -1413,7 +1395,7 @@ inline static void ggml_vec_step_f32 (const int n, float * y, const float * x) {
 inline static void ggml_vec_tanh_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = tanhf(x[i]);  }
 inline static void ggml_vec_elu_f32  (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = (x[i] > 0.f) ? x[i] : expf(x[i])-1; }
 inline static void ggml_vec_relu_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = (x[i] > 0.f) ? x[i] : 0.f; }
-inline static void ggml_vec_leaky_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = (x[i] > 0.f) ? x[i] : 0.1f*x[i]; }
+inline static void ggml_vec_leaky_relu_f32 (const int n, float * y, const float * x, const float ns) { for (int i = 0; i < n; ++i) y[i] = ((x[i] > 0.f) ? x[i] : 0.f) + ns * ((x[i] < 0.0f) ? x[i] : 0.f); }
 
 static const float GELU_COEF_A     = 0.044715f;
 static const float GELU_QUICK_COEF = -1.702f;
@@ -1613,6 +1595,7 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
     "GROUP_NORM",
 
     "MUL_MAT",
+    "MUL_MAT_ID",
     "OUT_PROD",
 
     "SCALE",
@@ -1640,6 +1623,9 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
     "POOL_1D",
     "POOL_2D",
     "UPSCALE",
+    "PAD",
+    "ARGSORT",
+    "LEAKY_RELU",
 
     "FLASH_ATTN",
     "FLASH_FF",
@@ -1666,7 +1652,7 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
     "CROSS_ENTROPY_LOSS_BACK",
 };
 
-static_assert(GGML_OP_COUNT == 68, "GGML_OP_COUNT != 68");
+static_assert(GGML_OP_COUNT == 72, "GGML_OP_COUNT != 72");
 
 static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
     "none",
@@ -1695,6 +1681,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
     "group_norm(x)",
 
     "X*Y",
+    "X[i]*Y",
     "X*Y",
 
     "x*v",
@@ -1722,6 +1709,9 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
     "pool_1d(x)",
     "pool_2d(x)",
     "upscale(x)",
+    "pad(x)",
+    "argsort(x)",
+    "leaky_relu(x)",
 
     "flash_attn(x)",
     "flash_ff(x)",
@@ -1748,15 +1738,32 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
     "cross_entropy_loss_back(x,y)",
 };
 
-static_assert(GGML_OP_COUNT == 68, "GGML_OP_COUNT != 68");
+static_assert(GGML_OP_COUNT == 72, "GGML_OP_COUNT != 72");
 
 static_assert(GGML_OP_POOL_COUNT == 2, "GGML_OP_POOL_COUNT != 2");
 
+
+static const char * GGML_UNARY_OP_NAME[GGML_UNARY_OP_COUNT] = {
+    "ABS",
+    "SGN",
+    "NEG",
+    "STEP",
+    "TANH",
+    "ELU",
+    "RELU",
+    "GELU",
+    "GELU_QUICK",
+    "SILU",
+};
+
+static_assert(GGML_UNARY_OP_COUNT == 10, "GGML_UNARY_OP_COUNT != 10");
+
+
 static_assert(sizeof(struct ggml_object)%GGML_MEM_ALIGN == 0, "ggml_object size must be a multiple of GGML_MEM_ALIGN");
 static_assert(sizeof(struct ggml_tensor)%GGML_MEM_ALIGN == 0, "ggml_tensor size must be a multiple of GGML_MEM_ALIGN");
 
 // WARN:
-// Mis-confguration can lead to problem that's hard to reason about:
+// Mis-configuration can lead to problem that's hard to reason about:
 // * At best  it crash or talks nosense.
 // * At worst it talks slightly difference but hard to perceive.
 //
@@ -1771,6 +1778,7 @@ static void ggml_setup_op_has_task_pass(void) {
 
         p[GGML_OP_ACC                    ] = true;
         p[GGML_OP_MUL_MAT                ] = true;
+        p[GGML_OP_MUL_MAT_ID             ] = true;
         p[GGML_OP_OUT_PROD               ] = true;
         p[GGML_OP_SET                    ] = true;
         p[GGML_OP_GET_ROWS_BACK          ] = true;
@@ -1989,12 +1997,6 @@ size_t ggml_nbytes_pad(const struct ggml_tensor * tensor) {
     return GGML_PAD(ggml_nbytes(tensor), GGML_MEM_ALIGN);
 }
 
-size_t ggml_nbytes_split(const struct ggml_tensor * tensor, int nrows_split) {
-    static_assert(GGML_MAX_DIMS == 4, "GGML_MAX_DIMS is not 4 - update this function");
-
-    return (nrows_split*tensor->ne[0]*ggml_type_size(tensor->type))/ggml_blck_size(tensor->type);
-}
-
 int ggml_blck_size(enum ggml_type type) {
     return type_traits[type].blck_size;
 }
@@ -2003,8 +2005,13 @@ size_t ggml_type_size(enum ggml_type type) {
     return type_traits[type].type_size;
 }
 
-float ggml_type_sizef(enum ggml_type type) {
-    return ((float)(type_traits[type].type_size))/type_traits[type].blck_size;
+size_t ggml_row_size(enum ggml_type type, int64_t ne) {
+    assert(ne % ggml_blck_size(type) == 0);
+    return ggml_type_size(type)*ne/ggml_blck_size(type);
+}
+
+double ggml_type_sizef(enum ggml_type type) {
+    return ((double)(type_traits[type].type_size))/type_traits[type].blck_size;
 }
 
 const char * ggml_type_name(enum ggml_type type) {
@@ -2023,28 +2030,55 @@ const char * ggml_op_symbol(enum ggml_op op) {
     return GGML_OP_SYMBOL[op];
 }
 
+const char * ggml_unary_op_name(enum ggml_unary_op op) {
+    return GGML_UNARY_OP_NAME[op];
+}
+
+const char * ggml_op_desc(const struct ggml_tensor * t) {
+    if (t->op == GGML_OP_UNARY) {
+        enum ggml_unary_op uop = ggml_get_unary_op(t);
+        return ggml_unary_op_name(uop);
+    }
+    else {
+        return ggml_op_name(t->op);
+    }
+}
+
 size_t ggml_element_size(const struct ggml_tensor * tensor) {
     return ggml_type_size(tensor->type);
 }
 
-static inline bool ggml_is_scalar(const struct ggml_tensor * tensor) {
+bool ggml_is_scalar(const struct ggml_tensor * tensor) {
     static_assert(GGML_MAX_DIMS == 4, "GGML_MAX_DIMS is not 4 - update this function");
 
     return tensor->ne[0] == 1 && tensor->ne[1] == 1 && tensor->ne[2] == 1 && tensor->ne[3] == 1;
 }
 
-static inline bool ggml_is_vector(const struct ggml_tensor * tensor) {
+bool ggml_is_vector(const struct ggml_tensor * tensor) {
     static_assert(GGML_MAX_DIMS == 4, "GGML_MAX_DIMS is not 4 - update this function");
 
     return tensor->ne[1] == 1 && tensor->ne[2] == 1 && tensor->ne[3] == 1;
 }
 
-static inline bool ggml_is_matrix(const struct ggml_tensor * tensor) {
+bool ggml_is_matrix(const struct ggml_tensor * tensor) {
     static_assert(GGML_MAX_DIMS == 4, "GGML_MAX_DIMS is not 4 - update this function");
 
     return tensor->ne[2] == 1 && tensor->ne[3] == 1;
 }
 
+bool ggml_is_3d(const struct ggml_tensor * tensor) {
+    return tensor->ne[3] == 1;
+}
+
+int ggml_n_dims(const struct ggml_tensor * tensor) {
+    for (int i = GGML_MAX_DIMS - 1; i >= 1; --i) {
+        if (tensor->ne[i] > 1) {
+            return i + 1;
+        }
+    }
+    return 1;
+}
+
 static inline bool ggml_can_mul_mat(const struct ggml_tensor * t0, const struct ggml_tensor * t1) {
     static_assert(GGML_MAX_DIMS == 4, "GGML_MAX_DIMS is not 4 - update this function");
 
@@ -2349,20 +2383,8 @@ size_t ggml_get_mem_size(const struct ggml_context * ctx) {
 size_t ggml_get_max_tensor_size(const struct ggml_context * ctx) {
     size_t max_size = 0;
 
-    struct ggml_object * obj = ctx->objects_begin;
-
-    while (obj != NULL) {
-        if (obj->type == GGML_OBJECT_TENSOR) {
-            struct ggml_tensor * tensor = (struct ggml_tensor *) ((char *) ctx->mem_buffer + obj->offs);
-
-            const size_t size = ggml_nbytes(tensor);
-
-            if (max_size < size) {
-                max_size = size;
-            }
-        }
-
-        obj = obj->next;
+    for (struct ggml_tensor * tensor = ggml_get_first_tensor(ctx); tensor != NULL; tensor = ggml_get_next_tensor(ctx, tensor)) {
+        max_size = MAX(max_size, ggml_nbytes(tensor));
     }
 
     return max_size;
@@ -2451,7 +2473,7 @@ static struct ggml_tensor * ggml_new_tensor_impl(
         view_src   = view_src->view_src;
     }
 
-    size_t data_size = ggml_type_size(type)*(ne[0]/ggml_blck_size(type));
+    size_t data_size = ggml_row_size(type, ne[0]);
     for (int i = 1; i < n_dims; i++) {
         data_size *= ne[i];
     }
@@ -2494,7 +2516,6 @@ static struct ggml_tensor * ggml_new_tensor_impl(
         /*.type         =*/ type,
         /*.backend      =*/ GGML_BACKEND_CPU,
         /*.buffer       =*/ NULL,
-        /*.n_dims       =*/ n_dims,
         /*.ne           =*/ { 1, 1, 1, 1 },
         /*.nb           =*/ { 0, 0, 0, 0 },
         /*.op           =*/ GGML_OP_NONE,
@@ -2601,7 +2622,7 @@ struct ggml_tensor * ggml_new_f32(struct ggml_context * ctx, float value) {
 }
 
 struct ggml_tensor * ggml_dup_tensor(struct ggml_context * ctx, const struct ggml_tensor * src) {
-    return ggml_new_tensor(ctx, src->type, src->n_dims, src->ne);
+    return ggml_new_tensor(ctx, src->type, GGML_MAX_DIMS, src->ne);
 }
 
 static void ggml_set_op_params(struct ggml_tensor * tensor, const void * params, size_t params_size) {
@@ -3050,7 +3071,7 @@ struct ggml_tensor * ggml_format_name(struct ggml_tensor * tensor, const char *
 struct ggml_tensor * ggml_view_tensor(
         struct ggml_context * ctx,
         struct ggml_tensor  * src) {
-    struct ggml_tensor * result = ggml_new_tensor_impl(ctx, src->type, src->n_dims, src->ne, src, 0);
+    struct ggml_tensor * result = ggml_new_tensor_impl(ctx, src->type, GGML_MAX_DIMS, src->ne, src, 0);
     ggml_format_name(result, "%s (view)", src->name);
 
     for (int i = 0; i < GGML_MAX_DIMS; i++) {
@@ -3060,7 +3081,7 @@ struct ggml_tensor * ggml_view_tensor(
     return result;
 }
 
-struct ggml_tensor * ggml_get_first_tensor(struct ggml_context * ctx) {
+struct ggml_tensor * ggml_get_first_tensor(const struct ggml_context * ctx) {
     struct ggml_object * obj = ctx->objects_begin;
 
     char * const mem_buffer = ctx->mem_buffer;
@@ -3076,7 +3097,7 @@ struct ggml_tensor * ggml_get_first_tensor(struct ggml_context * ctx) {
     return NULL;
 }
 
-struct ggml_tensor * ggml_get_next_tensor(struct ggml_context * ctx, struct ggml_tensor * tensor) {
+struct ggml_tensor * ggml_get_next_tensor(const struct ggml_context * ctx, struct ggml_tensor * tensor) {
     struct ggml_object * obj = (struct ggml_object *) ((char *)tensor - GGML_OBJECT_SIZE);
     obj = obj->next;
 
@@ -3154,9 +3175,7 @@ static struct ggml_tensor * ggml_add_impl(
         struct ggml_tensor * a,
         struct ggml_tensor * b,
         bool inplace) {
-    // TODO: support less-strict constraint
-    //       GGML_ASSERT(ggml_can_repeat(b, a));
-    GGML_ASSERT(ggml_can_repeat_rows(b, a));
+    GGML_ASSERT(ggml_can_repeat(b, a));
 
     bool is_node = false;
 
@@ -3210,10 +3229,10 @@ static struct ggml_tensor * ggml_add_cast_impl(
         is_node = true;
     }
 
-    struct ggml_tensor * result = ggml_new_tensor(ctx, type, a->n_dims, a->ne);
+    struct ggml_tensor * result = ggml_new_tensor(ctx, type, GGML_MAX_DIMS, a->ne);
 
     result->op   = GGML_OP_ADD;
-    result->grad = is_node ? ggml_new_tensor(ctx, GGML_TYPE_F32, a->n_dims, a->ne) : NULL;
+    result->grad = is_node ? ggml_new_tensor(ctx, GGML_TYPE_F32, GGML_MAX_DIMS, a->ne) : NULL;
     result->src[0] = a;
     result->src[1] = b;
 
@@ -3371,9 +3390,7 @@ static struct ggml_tensor * ggml_mul_impl(
         struct ggml_tensor * a,
         struct ggml_tensor * b,
         bool inplace) {
-    // TODO: support less-strict constraint
-    //       GGML_ASSERT(ggml_can_repeat(b, a));
-    GGML_ASSERT(ggml_can_repeat_rows(b, a));
+    GGML_ASSERT(ggml_can_repeat(b, a));
 
     bool is_node = false;
 
@@ -3418,7 +3435,7 @@ static struct ggml_tensor * ggml_div_impl(
         struct ggml_tensor * a,
         struct ggml_tensor * b,
         bool inplace) {
-    GGML_ASSERT(ggml_are_same_shape(a, b));
+    GGML_ASSERT(ggml_can_repeat(b, a));
 
     bool is_node = false;
 
@@ -3584,12 +3601,12 @@ struct ggml_tensor * ggml_sum_rows(
         is_node = true;
     }
 
-    int64_t ne[4] = {1,1,1,1};
-    for (int i=1; i<a->n_dims; ++i) {
+    int64_t ne[GGML_MAX_DIMS] = { 1 };
+    for (int i = 1; i < GGML_MAX_DIMS; ++i) {
         ne[i] = a->ne[i];
     }
 
-    struct ggml_tensor * result = ggml_new_tensor(ctx, a->type, a->n_dims, ne);
+    struct ggml_tensor * result = ggml_new_tensor(ctx, a->type, GGML_MAX_DIMS, ne);
 
     result->op   = GGML_OP_SUM_ROWS;
     result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
@@ -3610,8 +3627,8 @@ struct ggml_tensor * ggml_mean(
         is_node = true;
     }
 
-    int64_t ne[GGML_MAX_DIMS] = { 1, a->ne[1], a->ne[2], a->ne[3] };
-    struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, a->n_dims, ne);
+    int64_t ne[4] = { 1, a->ne[1], a->ne[2], a->ne[3] };
+    struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne);
 
     result->op   = GGML_OP_MEAN;
     result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
@@ -3633,8 +3650,7 @@ struct ggml_tensor * ggml_argmax(
         is_node = true;
     }
 
-    int64_t ne[GGML_MAX_DIMS] = { a->ne[1], 1, 1, 1 };
-    struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_I32, a->n_dims, ne);
+    struct ggml_tensor * result = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, a->ne[1]);
 
     result->op   = GGML_OP_ARGMAX;
     result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
@@ -3657,7 +3673,7 @@ struct ggml_tensor * ggml_repeat(
         is_node = true;
     }
 
-    struct ggml_tensor * result = ggml_new_tensor(ctx, a->type, b->n_dims, b->ne);
+    struct ggml_tensor * result = ggml_new_tensor(ctx, a->type, GGML_MAX_DIMS, b->ne);
 
     result->op   = GGML_OP_REPEAT;
     result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
@@ -3684,7 +3700,7 @@ struct ggml_tensor * ggml_repeat_back(
         return a;
     }
 
-    struct ggml_tensor * result = ggml_new_tensor(ctx, a->type, b->n_dims, b->ne);
+    struct ggml_tensor * result = ggml_new_tensor(ctx, a->type, GGML_MAX_DIMS, b->ne);
 
     result->op   = GGML_OP_REPEAT_BACK;
     result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
@@ -3815,12 +3831,25 @@ struct ggml_tensor * ggml_relu_inplace(
     return ggml_unary_inplace(ctx, a, GGML_UNARY_OP_RELU);
 }
 
-// ggml_leaky
+// ggml_leaky_relu
 
-struct ggml_tensor * ggml_leaky(
+struct ggml_tensor * ggml_leaky_relu(
         struct ggml_context * ctx,
-        struct ggml_tensor  * a) {
-    return ggml_unary(ctx, a, GGML_UNARY_OP_LEAKY);
+        struct ggml_tensor  * a, float negative_slope, bool inplace) {
+    bool is_node = false;
+
+    if (!inplace && (a->grad)) {
+        is_node = true;
+    }
+
+    struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
+    ggml_set_op_params(result, &negative_slope, sizeof(negative_slope));
+
+    result->op   = GGML_OP_LEAKY_RELU;
+    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
+    result->src[0] = a;
+
+    return result;
 }
 
 // ggml_gelu
@@ -4007,8 +4036,9 @@ static struct ggml_tensor * ggml_group_norm_impl(
 
     struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
 
-    result->op = GGML_OP_GROUP_NORM;
     result->op_params[0] = n_groups;
+
+    result->op = GGML_OP_GROUP_NORM;
     result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
     result->src[0] = a;
     result->src[1] = NULL; // TODO: maybe store epsilon here?
@@ -4046,7 +4076,7 @@ struct ggml_tensor * ggml_mul_mat(
     }
 
     const int64_t ne[4] = { a->ne[1], b->ne[1], b->ne[2], b->ne[3] };
-    struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, MAX(a->n_dims, b->n_dims), ne);
+    struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne);
 
     result->op   = GGML_OP_MUL_MAT;
     result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
@@ -4056,6 +4086,59 @@ struct ggml_tensor * ggml_mul_mat(
     return result;
 }
 
+void ggml_mul_mat_set_prec(
+        struct ggml_tensor * a,
+        enum ggml_prec       prec) {
+    const int32_t prec_i32 = (int32_t) prec;
+
+    ggml_set_op_params_i32(a, 0, prec_i32);
+}
+
+// ggml_mul_mat_id
+
+struct ggml_tensor * ggml_mul_mat_id(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * const as[],
+        int                   n_as,
+        struct ggml_tensor  * ids,
+        int                   id,
+        struct ggml_tensor  * b) {
+
+    GGML_ASSERT(ids->type == GGML_TYPE_I32);
+    GGML_ASSERT(ids->ne[2] == 1 && ids->ne[3] == 1);
+    GGML_ASSERT(ids->ne[1] == b->ne[1]);
+    GGML_ASSERT(ids->ne[2] == b->ne[2] && ids->ne[3] == b->ne[3]);
+    GGML_ASSERT(n_as > 0 && n_as <= GGML_MAX_SRC - 2);
+    GGML_ASSERT(id >= 0 && id < ids->ne[0]);
+
+    bool is_node = false;
+
+    if (as[0]->grad || b->grad) {
+        is_node = true;
+    }
+
+    const int64_t ne[4] = { as[0]->ne[1], b->ne[1], b->ne[2], b->ne[3] };
+    struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne);
+
+    ggml_set_op_params_i32(result, 0, id);
+    ggml_set_op_params_i32(result, 1, n_as);
+
+    result->op   = GGML_OP_MUL_MAT_ID;
+    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
+    result->src[0] = ids;
+    result->src[1] = b;
+
+    for (int i = 0; i < n_as; i++) {
+        struct ggml_tensor * a = as[i];
+        GGML_ASSERT(ggml_are_same_shape(as[0], a));
+        GGML_ASSERT(ggml_can_mul_mat(a, b));
+        GGML_ASSERT(!ggml_is_transposed(a));
+        result->src[i + 2] = a;
+    }
+
+    return result;
+}
+
 // ggml_out_prod
 
 struct ggml_tensor * ggml_out_prod(
@@ -4073,7 +4156,7 @@ struct ggml_tensor * ggml_out_prod(
 
     // a is broadcastable to b for ne[2] and ne[3] -> use b->ne[2] and b->ne[3]
     const int64_t ne[4] = { a->ne[0], b->ne[0], b->ne[2], b->ne[3] };
-    struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, MAX(a->n_dims, b->n_dims), ne);
+    struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne);
 
     result->op   = GGML_OP_OUT_PROD;
     result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
@@ -4088,23 +4171,23 @@ struct ggml_tensor * ggml_out_prod(
 static struct ggml_tensor * ggml_scale_impl(
         struct ggml_context * ctx,
         struct ggml_tensor  * a,
-        struct ggml_tensor  * b,
+        float                 s,
         bool inplace) {
-    GGML_ASSERT(ggml_is_scalar(b));
     GGML_ASSERT(ggml_is_padded_1d(a));
 
     bool is_node = false;
 
-    if (a->grad || b->grad) {
+    if (a->grad) {
         is_node = true;
     }
 
     struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
 
+    ggml_set_op_params(result, &s, sizeof(s));
+
     result->op   = GGML_OP_SCALE;
     result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
     result->src[0] = a;
-    result->src[1] = b;
 
     return result;
 }
@@ -4112,15 +4195,15 @@ static struct ggml_tensor * ggml_scale_impl(
 struct ggml_tensor * ggml_scale(
         struct ggml_context * ctx,
         struct ggml_tensor * a,
-        struct ggml_tensor * b) {
-    return ggml_scale_impl(ctx, a, b, false);
+        float                s) {
+    return ggml_scale_impl(ctx, a, s, false);
 }
 
 struct ggml_tensor * ggml_scale_inplace(
         struct ggml_context * ctx,
         struct ggml_tensor * a,
-        struct ggml_tensor * b) {
-    return ggml_scale_impl(ctx, a, b, true);
+        float                s) {
+    return ggml_scale_impl(ctx, a, s, true);
 }
 
 // ggml_set
@@ -4209,7 +4292,7 @@ struct ggml_tensor * ggml_set_2d_inplace(
         struct ggml_tensor *  b,
         size_t                nb1,
         size_t                offset) {
-    return ggml_set_impl(ctx, a, b, nb1, a->nb[2], a->nb[3], offset, false);
+    return ggml_set_impl(ctx, a, b, nb1, a->nb[2], a->nb[3], offset, true);
 }
 
 // ggml_cpy
@@ -4358,7 +4441,7 @@ struct ggml_tensor * ggml_reshape(
         //GGML_ASSERT(false);
     }
 
-    struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, b->n_dims, b->ne, a, 0);
+    struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, GGML_MAX_DIMS, b->ne, a, 0);
     ggml_format_name(result, "%s (reshaped)", a->name);
 
     result->op   = GGML_OP_RESHAPE;
@@ -4673,7 +4756,9 @@ struct ggml_tensor * ggml_get_rows(
         struct ggml_context * ctx,
         struct ggml_tensor  * a,
         struct ggml_tensor  * b) {
-    GGML_ASSERT(ggml_is_matrix(a) && ggml_is_vector(b) && b->type == GGML_TYPE_I32);
+    GGML_ASSERT(a->ne[2] == b->ne[1]);
+    GGML_ASSERT(b->ne[3] == 1);
+    GGML_ASSERT(b->type == GGML_TYPE_I32);
 
     bool is_node = false;
 
@@ -4683,7 +4768,7 @@ struct ggml_tensor * ggml_get_rows(
 
     // TODO: implement non F32 return
     //struct ggml_tensor * result = ggml_new_tensor_2d(ctx, a->type, a->ne[0], b->ne[0]);
-    struct ggml_tensor * result = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, a->ne[0], b->ne[0]);
+    struct ggml_tensor * result = ggml_new_tensor_4d(ctx, GGML_TYPE_F32, a->ne[0], b->ne[0], b->ne[1], b->ne[2]);
 
     result->op   = GGML_OP_GET_ROWS;
     result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
@@ -4734,7 +4819,7 @@ struct ggml_tensor * ggml_diag(
     }
 
     const int64_t ne[4] = { a->ne[0], a->ne[0], a->ne[2], a->ne[3] };
-    struct ggml_tensor * result = ggml_new_tensor(ctx, a->type, MAX(a->n_dims, 2), ne);
+    struct ggml_tensor * result = ggml_new_tensor(ctx, a->type, 4, ne);
 
     result->op   = GGML_OP_DIAG;
     result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
@@ -5381,7 +5466,7 @@ struct ggml_tensor * ggml_pool_1d(
         is_node = true;
     }
 
-    const int64_t ne[3] = {
+    const int64_t ne[2] = {
         ggml_calc_pool_output_size(a->ne[0], k0, s0, p0),
         a->ne[1],
     };
@@ -5461,6 +5546,30 @@ static struct ggml_tensor * ggml_upscale_impl(
     return result;
 }
 
+struct ggml_tensor * ggml_pad(
+    struct ggml_context * ctx,
+    struct ggml_tensor  * a,
+    int p0, int p1, int p2, int p3) {
+    bool is_node = false;
+
+    if (a->grad) {
+        GGML_ASSERT(false); // TODO: implement backward
+        is_node = true;
+    }
+
+    struct ggml_tensor * result = ggml_new_tensor_4d(ctx, a->type,
+            a->ne[0] + p0,
+            a->ne[1] + p1,
+            a->ne[2] + p2,
+            a->ne[3] + p3);
+
+    result->op = GGML_OP_PAD;
+    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
+    result->src[0] = a;
+
+    return result;
+}
+
 struct ggml_tensor * ggml_upscale(
     struct ggml_context * ctx,
     struct ggml_tensor * a,
@@ -5468,6 +5577,43 @@ struct ggml_tensor * ggml_upscale(
     return ggml_upscale_impl(ctx, a, scale_factor);
 }
 
+// ggml_argsort
+
+struct ggml_tensor * ggml_argsort(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        enum ggml_sort_order  order) {
+    bool is_node = false;
+
+    struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_I32, GGML_MAX_DIMS, a->ne);
+
+    ggml_set_op_params_i32(result, 0, (int32_t) order);
+
+    result->op   = GGML_OP_ARGSORT;
+    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
+    result->src[0] = a;
+
+    return result;
+}
+
+// ggml_top_k
+
+struct ggml_tensor * ggml_top_k(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        int                   k) {
+    GGML_ASSERT(a->ne[0] >= k);
+
+    struct ggml_tensor * result = ggml_argsort(ctx, a, GGML_SORT_DESC);
+
+    result = ggml_view_4d(ctx, result,
+                k, result->ne[1], result->ne[2], result->ne[3],
+                   result->nb[1], result->nb[2], result->nb[3],
+                0);
+
+    return result;
+}
+
 // ggml_flash_attn
 
 struct ggml_tensor * ggml_flash_attn(
@@ -5486,7 +5632,7 @@ struct ggml_tensor * ggml_flash_attn(
     }
 
     //struct ggml_tensor * result = ggml_dup_tensor(ctx, q);
-    struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, q->n_dims, q->ne);
+    struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, GGML_MAX_DIMS, q->ne);
 
     int32_t t = masked ? 1 : 0;
     ggml_set_op_params(result, &t, sizeof(t));
@@ -5519,7 +5665,7 @@ struct ggml_tensor * ggml_flash_ff(
     }
 
     //struct ggml_tensor * result = ggml_dup_tensor(ctx, a);
-    struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, a->n_dims, a->ne);
+    struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, GGML_MAX_DIMS, a->ne);
 
     result->op   = GGML_OP_FLASH_FF;
     result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
@@ -5635,7 +5781,6 @@ struct ggml_tensor * ggml_win_part(
     const int np  = npx*npy;
 
     const int64_t ne[4] = { a->ne[0], w, w, np, };
-
     struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne);
 
     int32_t params[] = { npx, npy, w };
@@ -6827,7 +6972,7 @@ static void ggml_compute_forward_add_f32(
         const struct ggml_tensor * src0,
         const struct ggml_tensor * src1,
         struct ggml_tensor * dst) {
-    GGML_ASSERT(ggml_can_repeat_rows(src1, src0) && ggml_are_same_shape(src0, dst));
+    GGML_ASSERT(ggml_can_repeat(src1, src0) && ggml_are_same_shape(src0, dst));
 
     if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
         return;
@@ -6860,16 +7005,19 @@ static void ggml_compute_forward_add_f32(
             const int64_t i13 = i03 % ne13;
             const int64_t i12 = i02 % ne12;
             const int64_t i11 = i01 % ne11;
+            const int64_t nr0 = ne00 / ne10;
 
             float * dst_ptr  = (float *) ((char *) dst->data  + i03*nb3  + i02*nb2  + i01*nb1 );
             float * src0_ptr = (float *) ((char *) src0->data + i03*nb03 + i02*nb02 + i01*nb01);
             float * src1_ptr = (float *) ((char *) src1->data + i13*nb13 + i12*nb12 + i11*nb11);
 
+            for (int64_t r = 0; r < nr0; ++r) {
 #ifdef GGML_USE_ACCELERATE
-            vDSP_vadd(src0_ptr, 1, src1_ptr, 1, dst_ptr, 1, ne00);
+                vDSP_vadd(src0_ptr + r*ne10, 1, src1_ptr, 1, dst_ptr + r*ne10, 1, ne10);
 #else
-            ggml_vec_add_f32(ne00, dst_ptr, src0_ptr, src1_ptr);
+                ggml_vec_add_f32(ne10, dst_ptr + r*ne10, src0_ptr + r*ne10, src1_ptr);
 #endif
+            }
         }
     } else {
         // src1 is not contiguous
@@ -6886,8 +7034,9 @@ static void ggml_compute_forward_add_f32(
             float * dst_ptr  = (float *) ((char *) dst->data  + i03*nb3  + i02*nb2  + i01*nb1 );
             float * src0_ptr = (float *) ((char *) src0->data + i03*nb03 + i02*nb02 + i01*nb01);
 
-            for (int i0 = 0; i0 < ne0; i0++) {
-                float * src1_ptr = (float *) ((char *) src1->data + i13*nb13 + i12*nb12 + i11*nb11 + i0*nb10);
+            for (int64_t i0 = 0; i0 < ne0; ++i0) {
+                const int64_t i10 = i0 % ne10;
+                float * src1_ptr = (float *) ((char *) src1->data + i13*nb13 + i12*nb12 + i11*nb11 + i10*nb10);
 
                 dst_ptr[i0] = src0_ptr[i0] + *src1_ptr;
             }
@@ -7421,7 +7570,7 @@ static void ggml_compute_forward_acc_f32(
     GGML_ASSERT(ggml_is_contiguous(dst) && ggml_is_contiguous(src0));
 
     // view src0 and dst with these strides and data offset inbytes during acc
-    // nb0 is implicitely element_size because src0 and dst are contiguous
+    // nb0 is implicitly element_size because src0 and dst are contiguous
     size_t nb1     = ((int32_t *) dst->op_params)[0];
     size_t nb2     = ((int32_t *) dst->op_params)[1];
     size_t nb3     = ((int32_t *) dst->op_params)[2];
@@ -7607,7 +7756,7 @@ static void ggml_compute_forward_mul_f32(
         const struct ggml_tensor * src0,
         const struct ggml_tensor * src1,
         struct ggml_tensor * dst) {
-    GGML_ASSERT(ggml_can_repeat_rows(src1, src0) && ggml_are_same_shape(src0, dst));
+    GGML_ASSERT(ggml_can_repeat(src1, src0) && ggml_are_same_shape(src0, dst));
 
     if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
         return;
@@ -7617,6 +7766,8 @@ static void ggml_compute_forward_mul_f32(
 
 #ifdef GGML_USE_CLBLAST
     if (src1->backend == GGML_BACKEND_GPU) {
+        // TODO: OpenCL kernel support full broadcast
+        GGML_ASSERT(ggml_can_repeat_rows(src1, src0));
         if (ith == 0) {
             ggml_cl_mul(src0, src1, dst);
         }
@@ -7630,7 +7781,6 @@ static void ggml_compute_forward_mul_f32(
 
     GGML_ASSERT( nb0 == sizeof(float));
     GGML_ASSERT(nb00 == sizeof(float));
-    GGML_ASSERT(ne00 == ne10);
 
     if (nb10 == sizeof(float)) {
         for (int64_t ir = ith; ir < nr; ir += nth) {
@@ -7642,20 +7792,21 @@ static void ggml_compute_forward_mul_f32(
             const int64_t i13 = i03 % ne13;
             const int64_t i12 = i02 % ne12;
             const int64_t i11 = i01 % ne11;
+            const int64_t nr0 = ne00 / ne10;
 
             float * dst_ptr  = (float *) ((char *) dst->data  + i03*nb3  + i02*nb2  + i01*nb1 );
             float * src0_ptr = (float *) ((char *) src0->data + i03*nb03 + i02*nb02 + i01*nb01);
             float * src1_ptr = (float *) ((char *) src1->data + i13*nb13 + i12*nb12 + i11*nb11);
 
+            for (int64_t r = 0 ; r < nr0; ++r) {
 #ifdef GGML_USE_ACCELERATE
-            UNUSED(ggml_vec_mul_f32);
+                UNUSED(ggml_vec_mul_f32);
 
-            vDSP_vmul( src0_ptr, 1, src1_ptr, 1, dst_ptr,  1, ne00);
+                vDSP_vmul(src0_ptr + r*ne10, 1, src1_ptr, 1, dst_ptr + r*ne10, 1, ne10);
 #else
-            ggml_vec_mul_f32(ne00, dst_ptr, src0_ptr, src1_ptr);
+                ggml_vec_mul_f32(ne10, dst_ptr + r*ne10, src0_ptr + r*ne10, src1_ptr);
 #endif
-                // }
-            // }
+            }
         }
     } else {
         // src1 is not contiguous
@@ -7673,8 +7824,9 @@ static void ggml_compute_forward_mul_f32(
             float * dst_ptr  = (float *) ((char *) dst->data  + i03*nb3  + i02*nb2  + i01*nb1 );
             float * src0_ptr = (float *) ((char *) src0->data + i03*nb03 + i02*nb02 + i01*nb01);
 
-            for (int64_t i0 = 0; i0 < ne00; i0++) {
-                float * src1_ptr = (float *) ((char *) src1->data + i13*nb13 + i12*nb12 + i11*nb11 + i0*nb10);
+            for (int64_t i0 = 0; i0 < ne00; ++i0) {
+                const int64_t i10 = i0 % ne10;
+                float * src1_ptr = (float *) ((char *) src1->data + i13*nb13 + i12*nb12 + i11*nb11 + i10*nb10);
 
                 dst_ptr[i0] = src0_ptr[i0] * (*src1_ptr);
             }
@@ -7708,14 +7860,16 @@ static void ggml_compute_forward_div_f32(
         const struct ggml_tensor * src0,
         const struct ggml_tensor * src1,
         struct ggml_tensor * dst) {
-    assert(params->ith == 0);
-    assert(ggml_are_same_shape(src0, src1) && ggml_are_same_shape(src0, dst));
+    GGML_ASSERT(ggml_can_repeat(src1, src0) && ggml_are_same_shape(src0, dst));
 
     if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
         return;
     }
 
-    const int nr  = ggml_nrows(src0);
+    const int ith = params->ith;
+    const int nth = params->nth;
+
+    const int64_t nr = ggml_nrows(src0);
 
     GGML_TENSOR_BINARY_OP_LOCALS
 
@@ -7723,41 +7877,50 @@ static void ggml_compute_forward_div_f32(
     GGML_ASSERT(nb00 == sizeof(float));
 
     if (nb10 == sizeof(float)) {
-        for (int ir = 0; ir < nr; ++ir) {
-            // src0, src1 and dst are same shape => same indices
-            const int i3 = ir/(ne2*ne1);
-            const int i2 = (ir - i3*ne2*ne1)/ne1;
-            const int i1 = (ir - i3*ne2*ne1 - i2*ne1);
+        for (int64_t ir = ith; ir < nr; ir += nth) {
+            // src0 and dst are same shape => same indices
+            const int64_t i03 = ir/(ne02*ne01);
+            const int64_t i02 = (ir - i03*ne02*ne01)/ne01;
+            const int64_t i01 = (ir - i03*ne02*ne01 - i02*ne01);
 
+            const int64_t i13 = i03 % ne13;
+            const int64_t i12 = i02 % ne12;
+            const int64_t i11 = i01 % ne11;
+            const int64_t nr0 = ne00 / ne10;
+
+            float * dst_ptr  = (float *) ((char *) dst->data  + i03*nb3  + i02*nb2  + i01*nb1 );
+            float * src0_ptr = (float *) ((char *) src0->data + i03*nb03 + i02*nb02 + i01*nb01);
+            float * src1_ptr = (float *) ((char *) src1->data + i13*nb13 + i12*nb12 + i11*nb11);
+
+            for (int64_t r = 0; r < nr0; ++r) {
 #ifdef GGML_USE_ACCELERATE
-            UNUSED(ggml_vec_div_f32);
+                UNUSED(ggml_vec_div_f32);
 
-            vDSP_vdiv(
-                    (float *) ((char *) src1->data + i3*nb13 + i2*nb12 + i1*nb11), 1,
-                    (float *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01), 1,
-                    (float *) ((char *) dst->data  + i3*nb3  + i2*nb2  + i1*nb1 ), 1,
-                    ne0);
+                vDSP_vdiv(src1_ptr, 1, src0_ptr + r*ne10, 1, dst_ptr + r*ne10, 1, ne10);
 #else
-            ggml_vec_div_f32(ne0,
-                    (float *) ((char *) dst->data  + i3*nb3  + i2*nb2  + i1*nb1 ),
-                    (float *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01),
-                    (float *) ((char *) src1->data + i3*nb13 + i2*nb12 + i1*nb11));
+                ggml_vec_div_f32(ne10, dst_ptr + r*ne10, src0_ptr + r*ne10, src1_ptr);
 #endif
-                // }
-            // }
+            }
         }
     } else {
         // src1 is not contiguous
-        for (int ir = 0; ir < nr; ++ir) {
-            // src0, src1 and dst are same shape => same indices
-            const int i3 = ir/(ne2*ne1);
-            const int i2 = (ir - i3*ne2*ne1)/ne1;
-            const int i1 = (ir - i3*ne2*ne1 - i2*ne1);
+        for (int64_t ir = ith; ir < nr; ir += nth) {
+            // src0 and dst are same shape => same indices
+            // src1 is broadcastable across src0 and dst in i1, i2, i3
+            const int64_t i03 = ir/(ne02*ne01);
+            const int64_t i02 = (ir - i03*ne02*ne01)/ne01;
+            const int64_t i01 = (ir - i03*ne02*ne01 - i02*ne01);
 
-            float * dst_ptr  = (float *) ((char *) dst->data  + i3*nb3  + i2*nb2  + i1*nb1 );
-            float * src0_ptr = (float *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01);
-            for (int i0 = 0; i0 < ne0; i0++) {
-                float * src1_ptr = (float *) ((char *) src1->data + i3*nb13 + i2*nb12 + i1*nb11 + i0*nb10);
+            const int64_t i13 = i03 % ne13;
+            const int64_t i12 = i02 % ne12;
+            const int64_t i11 = i01 % ne11;
+
+            float * dst_ptr  = (float *) ((char *) dst->data  + i03*nb3  + i02*nb2  + i01*nb1 );
+            float * src0_ptr = (float *) ((char *) src0->data + i03*nb03 + i02*nb02 + i01*nb01);
+
+            for (int64_t i0 = 0; i0 < ne00; ++i0) {
+                const int64_t i10 = i0 % ne10;
+                float * src1_ptr = (float *) ((char *) src1->data + i13*nb13 + i12*nb12 + i11*nb11 + i10*nb10);
 
                 dst_ptr[i0] = src0_ptr[i0] / (*src1_ptr);
             }
@@ -8203,7 +8366,7 @@ static void ggml_compute_forward_repeat_f16(
         return;
     }
 
-    GGML_TENSOR_UNARY_OP_LOCALS;
+    GGML_TENSOR_UNARY_OP_LOCALS
 
     // guaranteed to be an integer due to the check in ggml_can_repeat
     const int nr0 = (int)(ne0/ne00);
@@ -8348,6 +8511,7 @@ static void ggml_compute_forward_concat_f32(
     GGML_ASSERT(src0->nb[0] == sizeof(float));
 
     const int ith = params->ith;
+    const int nth = params->nth;
 
     GGML_TENSOR_BINARY_OP_LOCALS
 
@@ -8357,7 +8521,7 @@ static void ggml_compute_forward_concat_f32(
     GGML_ASSERT(nb10 == sizeof(float));
 
     for (int i3 = 0; i3 < ne3; i3++) {
-        for (int i2 = ith; i2 < ne2; i2++) {
+        for (int i2 = ith; i2 < ne2; i2 += nth) {
             if (i2 < ne02) { // src0
                 for (int i1 = 0; i1 < ne1; i1++) {
                     for (int i0 = 0; i0 < ne0; i0++) {
@@ -8869,10 +9033,9 @@ static void ggml_compute_forward_silu(
             } break;
     }
 }
+// ggml_compute_forward_leaky_relu
 
-// ggml_compute_forward_leaky
-
-static void ggml_compute_forward_leaky_f32(
+static void ggml_compute_forward_leaky_relu_f32(
         const struct ggml_compute_params * params,
         const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
@@ -8886,24 +9049,27 @@ static void ggml_compute_forward_leaky_f32(
     const int n  = ggml_nrows(src0);
     const int nc = src0->ne[0];
 
+    float negative_slope;
+    memcpy(&negative_slope, dst->op_params, sizeof(float));
+
     assert(dst->nb[0]  == sizeof(float));
     assert(src0->nb[0] == sizeof(float));
 
     for (int i = 0; i < n; i++) {
-        ggml_vec_leaky_f32(nc,
+        ggml_vec_leaky_relu_f32(nc,
                 (float *) ((char *) dst->data  + i*( dst->nb[1])),
-                (float *) ((char *) src0->data + i*(src0->nb[1])));
+                (float *) ((char *) src0->data + i*(src0->nb[1])), negative_slope);
     }
 }
 
-static void ggml_compute_forward_leaky(
+static void ggml_compute_forward_leaky_relu(
         const struct ggml_compute_params * params,
         const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
     switch (src0->type) {
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_leaky_f32(params, src0, dst);
+                ggml_compute_forward_leaky_relu_f32(params, src0, dst);
             } break;
         default:
             {
@@ -8998,6 +9164,8 @@ static void ggml_compute_forward_norm_f32(
     float eps;
     memcpy(&eps, dst->op_params, sizeof(float));
 
+    GGML_ASSERT(eps > 0.0f);
+
     // TODO: optimize
     for (int64_t i03 = 0; i03 < ne03; i03++) {
         for (int64_t i02 = 0; i02 < ne02; i02++) {
@@ -9067,6 +9235,8 @@ static void ggml_compute_forward_rms_norm_f32(
     float eps;
     memcpy(&eps, dst->op_params, sizeof(float));
 
+    GGML_ASSERT(eps > 0.0f);
+
     // TODO: optimize
     for (int64_t i03 = 0; i03 < ne03; i03++) {
         for (int64_t i02 = 0; i02 < ne02; i02++) {
@@ -9392,8 +9562,11 @@ static bool ggml_compute_forward_mul_mat_use_blas(
     const int64_t ne0 = dst->ne[0];
     const int64_t ne1 = dst->ne[1];
 
+    // NOTE: with GGML_OP_MUL_MAT_ID we don't want to go through the BLAS branch because it will dequantize (to_float)
+    //       all the experts for each batch element and the processing would become incredibly slow
     // TODO: find the optimal values for these
-    if (ggml_is_contiguous(src0) &&
+    if (dst->op != GGML_OP_MUL_MAT_ID &&
+        ggml_is_contiguous(src0) &&
         ggml_is_contiguous(src1) &&
       //src0->type == GGML_TYPE_F32 &&
         src1->type == GGML_TYPE_F32 &&
@@ -9481,8 +9654,7 @@ static void ggml_compute_forward_mul_mat(
 
                 const void  * x = (char *)            src0->data + i02*nb02 + i03*nb03;
                 const float * y = (float *) ((char *) src1->data + i12*nb12 + i13*nb13);
-
-                float * d = (float *) ((char *) dst->data + i12*nb2 + i13*nb3);
+                      float * d = (float *) ((char *)  dst->data + i12*nb2  + i13*nb3);
 
                 if (type != GGML_TYPE_F32) {
                             float * const wdata    = params->wdata;
@@ -9499,10 +9671,10 @@ static void ggml_compute_forward_mul_mat(
                 }
 
                 cblas_sgemm(CblasRowMajor, CblasNoTrans, CblasTrans,
-                        ne11, ne01, ne10,
-                        1.0f,    y, ne10,
-                                 x, ne00,
-                        0.0f,    d, ne01);
+                          ne1, ne01, ne10,
+                         1.0f,    y, ne10,
+                                  x, ne00,
+                         0.0f,    d, ne01);
             }
         }
 
@@ -9515,7 +9687,10 @@ static void ggml_compute_forward_mul_mat(
     if (params->type == GGML_TASK_INIT) {
         if (src1->type != vec_dot_type) {
             char * wdata = params->wdata;
-            const size_t row_size = ne10*ggml_type_size(vec_dot_type)/ggml_blck_size(vec_dot_type);
+            const size_t row_size = ggml_row_size(vec_dot_type, ne10);
+
+            assert(params->wsize >= ne11*ne12*ne13*row_size);
+            assert(src1->type == GGML_TYPE_F32);
 
             for (int64_t i13 = 0; i13 < ne13; ++i13) {
                 for (int64_t i12 = 0; i12 < ne12; ++i12) {
@@ -9535,10 +9710,10 @@ static void ggml_compute_forward_mul_mat(
     }
 
     const void * wdata    = (src1->type == vec_dot_type) ? src1->data : params->wdata;
-    const size_t row_size = ne10*ggml_type_size(vec_dot_type)/ggml_blck_size(vec_dot_type);
+    const size_t row_size = ggml_row_size(vec_dot_type, ne10);
 
-    const int64_t nr0 = ne01;           // src0 rows
-    const int64_t nr1 = ne11*ne12*ne13; // src1 rows
+    const int64_t nr0 = ne01;          // src0 rows
+    const int64_t nr1 = ne1*ne12*ne13; // src1 rows
 
     //printf("nr0 = %lld, nr1 = %lld\n", nr0, nr1);
 
@@ -9580,9 +9755,9 @@ static void ggml_compute_forward_mul_mat(
     for (int64_t iir1 = ir110; iir1 < ir111; iir1 += blck_1) {
         for (int64_t iir0 = ir010; iir0 < ir011; iir0 += blck_0) {
             for (int64_t ir1 = iir1; ir1 < iir1 + blck_1 && ir1 < ir111; ++ir1) {
-                const int64_t i13 = (ir1/(ne12*ne11));
-                const int64_t i12 = (ir1 - i13*ne12*ne11)/ne11;
-                const int64_t i11 = (ir1 - i13*ne12*ne11 - i12*ne11);
+                const int64_t i13 = (ir1/(ne12*ne1));
+                const int64_t i12 = (ir1 - i13*ne12*ne1)/ne1;
+                const int64_t i11 = (ir1 - i13*ne12*ne1 - i12*ne1);
 
                 // broadcast src0 into src1
                 const int64_t i03 = i13/r3;
@@ -9618,6 +9793,197 @@ static void ggml_compute_forward_mul_mat(
     }
 }
 
+// ggml_compute_forward_mul_mat_id
+
+static void ggml_compute_forward_mul_mat_id(
+        const struct ggml_compute_params * params,
+        const struct ggml_tensor * ids,
+        const struct ggml_tensor * src1,
+              struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[2]; // only for GGML_TENSOR_BINARY_OP_LOCALS
+
+    GGML_TENSOR_BINARY_OP_LOCALS
+
+    const int ith = params->ith;
+    const int nth = params->nth;
+
+    const enum ggml_type type = src0->type;
+
+    const bool src1_cont = ggml_is_contiguous(src1);
+
+    ggml_vec_dot_t    const vec_dot               = type_traits[type].vec_dot;
+    enum ggml_type    const vec_dot_type          = type_traits[type].vec_dot_type;
+    ggml_from_float_t const from_float_to_vec_dot = type_traits[vec_dot_type].from_float;
+
+    GGML_ASSERT(ne0 == ne01);
+    GGML_ASSERT(ne1 == ne11);
+    GGML_ASSERT(ne2 == ne12);
+    GGML_ASSERT(ne3 == ne13);
+
+    // we don't support permuted src0 or src1
+    GGML_ASSERT(nb00 == ggml_type_size(type));
+    GGML_ASSERT(nb10 == ggml_type_size(src1->type));
+
+    // dst cannot be transposed or permuted
+    GGML_ASSERT(nb0 == sizeof(float));
+    GGML_ASSERT(nb0 <= nb1);
+    GGML_ASSERT(nb1 <= nb2);
+    GGML_ASSERT(nb2 <= nb3);
+
+    // broadcast factors
+    const int64_t r2 = ne12/ne02;
+    const int64_t r3 = ne13/ne03;
+
+    // row groups
+    const int id   = ggml_get_op_params_i32(dst, 0);
+    const int n_as = ggml_get_op_params_i32(dst, 1);
+
+    char * wdata_src1_end = (src1->type == vec_dot_type) ?
+            (char *) params->wdata :
+            (char *) params->wdata + GGML_PAD(ggml_row_size(vec_dot_type, ggml_nelements(src1)), sizeof(int64_t));
+
+    int64_t * matrix_row_counts = (int64_t *) (wdata_src1_end); // [n_as]
+    int64_t * matrix_rows       = matrix_row_counts + n_as;     // [n_as][ne11]
+
+    #define MMID_MATRIX_ROW(row_id, i1) matrix_rows[(row_id)*ne11 + (i1)]
+
+   if (params->type == GGML_TASK_INIT) {
+        char * wdata = params->wdata;
+        if (src1->type != vec_dot_type) {
+            const size_t row_size = ggml_row_size(vec_dot_type, ne10);
+
+            assert(params->wsize >= ne11*ne12*ne13*row_size);
+            assert(src1->type == GGML_TYPE_F32);
+
+            for (int64_t i13 = 0; i13 < ne13; ++i13) {
+                for (int64_t i12 = 0; i12 < ne12; ++i12) {
+                    for (int64_t i11 = 0; i11 < ne11; ++i11) {
+                        from_float_to_vec_dot((float *)((char *) src1->data + i13*nb13 + i12*nb12 + i11*nb11), (void *) wdata, ne10);
+                        wdata += row_size;
+                    }
+                }
+            }
+        }
+
+        // initialize matrix_row_counts
+        GGML_ASSERT(wdata == wdata_src1_end);
+        memset(matrix_row_counts, 0, n_as*sizeof(int64_t));
+
+        // group rows by src0 matrix
+        for (int64_t i01 = 0; i01 < ids->ne[1]; i01++) {
+            const int32_t row_id = *(const int32_t *) ((const char *) ids->data + i01*ids->nb[1] + id*ids->nb[0]);
+
+            GGML_ASSERT(row_id >= 0 && row_id < n_as);
+            MMID_MATRIX_ROW(row_id, matrix_row_counts[row_id]) = i01;
+            matrix_row_counts[row_id] += 1;
+        }
+
+        return;
+    }
+
+    if (params->type == GGML_TASK_FINALIZE) {
+        return;
+    }
+
+    // compute each matrix multiplication in sequence
+    for (int cur_a = 0; cur_a < n_as; ++cur_a) {
+        const int64_t cne1 = matrix_row_counts[cur_a];
+
+        if (cne1 == 0) {
+            continue;
+        }
+
+        const struct ggml_tensor * src0_cur = dst->src[cur_a + 2];
+
+        const void * wdata    = (src1->type == vec_dot_type) ? src1->data : params->wdata;
+        const size_t row_size = ggml_row_size(vec_dot_type, ne10);
+
+        const int64_t nr0 = ne01;           // src0 rows
+        const int64_t nr1 = cne1*ne12*ne13; // src1 rows
+
+        //printf("nr0 = %lld, nr1 = %lld\n", nr0, nr1);
+
+        // distribute the thread work across the inner or outer loop based on which one is larger
+
+        const int64_t nth0 = nr0 > nr1 ? nth : 1; // parallelize by src0 rows
+        const int64_t nth1 = nr0 > nr1 ? 1 : nth; // parallelize by src1 rows
+
+        const int64_t ith0 = ith % nth0;
+        const int64_t ith1 = ith / nth0;
+
+        const int64_t dr0 = (nr0 + nth0 - 1)/nth0;
+        const int64_t dr1 = (nr1 + nth1 - 1)/nth1;
+
+        const int64_t ir010 = dr0*ith0;
+        const int64_t ir011 = MIN(ir010 + dr0, nr0);
+
+        const int64_t ir110 = dr1*ith1;
+        const int64_t ir111 = MIN(ir110 + dr1, nr1);
+
+        //printf("ir010 = %6lld, ir011 = %6lld, ir110 = %6lld, ir111 = %6lld\n", ir010, ir011, ir110, ir111);
+
+        // threads with no work simply yield (not sure if it helps)
+        if (ir010 >= ir011 || ir110 >= ir111) {
+            sched_yield();
+            continue;
+        }
+
+        assert(ne12 % ne02 == 0);
+        assert(ne13 % ne03 == 0);
+
+        // block-tiling attempt
+        const int64_t blck_0 = 16;
+        const int64_t blck_1 = 16;
+
+        // attempt to reduce false-sharing (does not seem to make a difference)
+        float tmp[16];
+
+        for (int64_t iir1 = ir110; iir1 < ir111; iir1 += blck_1) {
+            for (int64_t iir0 = ir010; iir0 < ir011; iir0 += blck_0) {
+                for (int64_t ir1 = iir1; ir1 < iir1 + blck_1 && ir1 < ir111; ++ir1) {
+                    const int64_t  i13 = (ir1/(ne12*cne1)); // Note: currently, src1 is always a matrix
+                    const int64_t  i12 = (ir1 - i13*ne12*cne1)/cne1;
+                    const int64_t _i11 = (ir1 - i13*ne12*cne1 - i12*cne1);
+                    const int64_t  i11 = MMID_MATRIX_ROW(cur_a, _i11);
+
+                    // broadcast src0 into src1
+                    const int64_t i03 = i13/r3;
+                    const int64_t i02 = i12/r2;
+
+                    const int64_t i1 = i11;
+                    const int64_t i2 = i12;
+                    const int64_t i3 = i13;
+
+                    const char * src0_row = (const char *) src0_cur->data + (0 + i02*nb02 + i03*nb03);
+
+                    // desc: when src1 is not a contiguous memory block we have to calculate the offset using the strides
+                    //       if it is, then we have either copied the data to params->wdata and made it contiguous or we are using
+                    //       the original src1 data pointer, so we should index using the indices directly
+                    // TODO: this is a bit of a hack, we should probably have a better way to handle this
+                    const char * src1_col = (const char *) wdata +
+                        (src1_cont || src1->type != vec_dot_type
+                        ? (i11      + i12*ne11 + i13*ne12*ne11)*row_size
+                        : (i11*nb11 + i12*nb12 + i13*nb13));
+
+                    float * dst_col = (float *) ((char *) dst->data + (i1*nb1 + i2*nb2 + i3*nb3));
+
+                    //for (int64_t ir0 = iir0; ir0 < iir0 + blck_0 && ir0 < ir011; ++ir0) {
+                    //    vec_dot(ne00, &dst_col[ir0], src0_row + ir0*nb01, src1_col);
+                    //}
+
+                    for (int64_t ir0 = iir0; ir0 < iir0 + blck_0 && ir0 < ir011; ++ir0) {
+                        vec_dot(ne00, &tmp[ir0 - iir0], src0_row + ir0*nb01, src1_col);
+                    }
+                    memcpy(&dst_col[iir0], tmp, (MIN(iir0 + blck_0, ir011) - iir0)*sizeof(float));
+                }
+            }
+        }
+    }
+
+    #undef MMID_MATRIX_ROW
+}
+
 // ggml_compute_forward_out_prod
 
 static void ggml_compute_forward_out_prod_f32(
@@ -9959,19 +10325,18 @@ static void ggml_compute_forward_out_prod(
 static void ggml_compute_forward_scale_f32(
         const struct ggml_compute_params * params,
         const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
         struct ggml_tensor * dst) {
     GGML_ASSERT(ggml_is_contiguous(src0));
     GGML_ASSERT(ggml_is_contiguous(dst));
     GGML_ASSERT(ggml_are_same_shape(src0, dst));
-    GGML_ASSERT(ggml_is_scalar(src1));
 
     if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
         return;
     }
 
     // scale factor
-    const float v = *(float *) src1->data;
+    float v;
+    memcpy(&v, dst->op_params, sizeof(float));
 
     const int ith = params->ith;
     const int nth = params->nth;
@@ -10002,12 +10367,11 @@ static void ggml_compute_forward_scale_f32(
 static void ggml_compute_forward_scale(
         const struct ggml_compute_params * params,
         const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
         struct ggml_tensor * dst) {
     switch (src0->type) {
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_scale_f32(params, src0, src1, dst);
+                ggml_compute_forward_scale_f32(params, src0, dst);
             } break;
         default:
             {
@@ -10027,7 +10391,7 @@ static void ggml_compute_forward_set_f32(
     GGML_ASSERT(ggml_is_contiguous(dst) && ggml_is_contiguous(src0));
 
     // view src0 and dst with these strides and data offset inbytes during set
-    // nb0 is implicitely element_size because src0 and dst are contiguous
+    // nb0 is implicitly element_size because src0 and dst are contiguous
     size_t nb1     = ((int32_t *) dst->op_params)[0];
     size_t nb2     = ((int32_t *) dst->op_params)[1];
     size_t nb3     = ((int32_t *) dst->op_params)[2];
@@ -10191,21 +10555,30 @@ static void ggml_compute_forward_get_rows_q(
         return;
     }
 
-    const int nc = src0->ne[0];
-    const int nr = ggml_nelements(src1);
+    GGML_TENSOR_BINARY_OP_LOCALS
+
+    const int64_t nc = ne00;
+    const int64_t nr = ggml_nelements(src1); GGML_UNUSED(nr);
+
     const enum ggml_type type = src0->type;
     ggml_to_float_t const dequantize_row_q = type_traits[type].to_float;
 
-    assert( dst->ne[0] == nc);
-    assert( dst->ne[1] == nr);
-    assert(src0->nb[0] == ggml_type_size(type));
+    assert(ne0  == nc);
+    assert(ne02 == ne11);
+    assert(nb00 == ggml_type_size(type));
+    assert(ggml_nrows(dst) == nr);
 
-    for (int i = 0; i < nr; ++i) {
-        const int r = ((int32_t *) src1->data)[i];
+    // TODO: multi-thread
+    for (int64_t i12 = 0; i12 < ne12; ++i12) {
+        for (int64_t i11 = 0; i11 < ne11; ++i11) {
+            for (int64_t i10 = 0; i10 < ne10; ++i10) {
+                const int64_t i01 = *(int32_t *) ((char *) src1->data + i10*nb10 + i11*nb11 + i12*nb12);
 
-        dequantize_row_q(
-                (const void *) ((char *) src0->data + r*src0->nb[1]),
-                     (float *) ((char *)  dst->data + i*dst->nb[1]), nc);
+                dequantize_row_q(
+                        (const void *) ((char *) src0->data + i01*nb01 + i11*nb02 + i12*nb03),
+                             (float *) ((char *)  dst->data + i10*nb1  + i11*nb2  + i12*nb3), nc);
+            }
+        }
     }
 }
 
@@ -10220,19 +10593,26 @@ static void ggml_compute_forward_get_rows_f16(
         return;
     }
 
-    const int nc = src0->ne[0];
-    const int nr = ggml_nelements(src1);
+    GGML_TENSOR_BINARY_OP_LOCALS
 
-    assert( dst->ne[0] == nc);
-    assert( dst->ne[1] == nr);
-    assert(src0->nb[0] == sizeof(ggml_fp16_t));
+    const int64_t nc = ne00;
+    const int64_t nr = ggml_nelements(src1); GGML_UNUSED(nr);
 
-    for (int i = 0; i < nr; ++i) {
-        const int r = ((int32_t *) src1->data)[i];
+    assert(ne0  == nc);
+    assert(ne02 == ne11);
+    assert(nb00 == sizeof(ggml_fp16_t));
+    assert(ggml_nrows(dst) == nr);
 
-        for (int j = 0; j < nc; ++j) {
-            ggml_fp16_t v = ((ggml_fp16_t *) ((char *) src0->data + r*src0->nb[1]))[j];
-            ((float *) ((char *)  dst->data + i*dst->nb[1]))[j] = GGML_FP16_TO_FP32(v);
+    // TODO: multi-thread
+    for (int64_t i12 = 0; i12 < ne12; ++i12) {
+        for (int64_t i11 = 0; i11 < ne11; ++i11) {
+            for (int64_t i10 = 0; i10 < ne10; ++i10) {
+                const int64_t i01 = *(int32_t *) ((char *) src1->data + i10*nb10 + i11*nb11 + i12*nb12);
+
+                ggml_fp16_to_fp32_row(
+                        (const void *) ((char *) src0->data + i01*nb01 + i11*nb02 + i12*nb03),
+                             (float *) ((char *)  dst->data + i10*nb1  + i11*nb2  + i12*nb3), nc);
+            }
         }
     }
 }
@@ -10248,19 +10628,27 @@ static void ggml_compute_forward_get_rows_f32(
         return;
     }
 
-    const int nc = src0->ne[0];
-    const int nr = ggml_nelements(src1);
+    GGML_TENSOR_BINARY_OP_LOCALS
 
-    assert( dst->ne[0] == nc);
-    assert( dst->ne[1] == nr);
-    assert(src0->nb[0] == sizeof(float));
+    const int64_t nc = ne00;
+    const int64_t nr = ggml_nelements(src1); GGML_UNUSED(nr);
 
-    for (int i = 0; i < nr; ++i) {
-        const int r = ((int32_t *) src1->data)[i];
+    assert(ne0  == nc);
+    assert(ne02 == ne11);
+    assert(nb00 == sizeof(float));
+    assert(ggml_nrows(dst) == nr);
 
-        ggml_vec_cpy_f32(nc,
-                (float *) ((char *)  dst->data + i*dst->nb[1]),
-                (float *) ((char *) src0->data + r*src0->nb[1]));
+    // TODO: multi-thread
+    for (int64_t i12 = 0; i12 < ne12; ++i12) {
+        for (int64_t i11 = 0; i11 < ne11; ++i11) {
+            for (int64_t i10 = 0; i10 < ne10; ++i10) {
+                const int64_t i01 = *(int32_t *) ((char *) src1->data + i10*nb10 + i11*nb11 + i12*nb12);
+
+                ggml_vec_cpy_f32(nc,
+                        (float *) ((char *)  dst->data + i10*nb1  + i11*nb2  + i12*nb3),
+                        (float *) ((char *) src0->data + i01*nb01 + i11*nb02 + i12*nb03));
+            }
+        }
     }
 }
 
@@ -11172,10 +11560,13 @@ static void ggml_compute_forward_rope_f32(
                     }
                 } else {
                     // TODO: this might be wrong for ne0 != n_dims - need double check
-                    // ref:  https://github.com/huggingface/transformers/blob/main/src/transformers/models/gpt_neox/modeling_gpt_neox.py#LL251C1-L294C28
+                    //       it seems we have to rope just the first n_dims elements and do nothing with the rest
+                    // ref:  https://github.com/ml-explore/mlx/blob/dc2edc762c797e3b8de50b1dad4dc0a131691033/benchmarks/python/llama_jax_bench.py#L11-L26
                     theta_base *= freq_scale;
-                    for (int64_t ib = 0; ib < ne0/n_dims; ++ib) {
-                        for (int64_t ic = 0; ic < n_dims; ic += 2) {
+                    for (int64_t ic = 0; ic < ne0; ic += 2) {
+                        if (ic < n_dims) {
+                            const int64_t ib = 0;
+
                             // simplified from `(ib * n_dims + ic) * inv_ndims`
                             float cur_rot = inv_ndims * ic - ib;
 
@@ -11198,6 +11589,14 @@ static void ggml_compute_forward_rope_f32(
 
                             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;
+
+                            const float * const src = (float *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
+                                  float * dst_data  = (float *)((char *)  dst->data + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
+
+                            dst_data[0] = src[0];
+                            dst_data[1] = src[1];
                         }
                     }
                 }
@@ -11325,10 +11724,13 @@ static void ggml_compute_forward_rope_f16(
                     }
                 } else {
                     // TODO: this might be wrong for ne0 != n_dims - need double check
-                    // ref:  https://github.com/huggingface/transformers/blob/main/src/transformers/models/gpt_neox/modeling_gpt_neox.py#LL251C1-L294C28
+                    //       it seems we have to rope just the first n_dims elements and do nothing with the rest
+                    // ref:  https://github.com/ml-explore/mlx/blob/dc2edc762c797e3b8de50b1dad4dc0a131691033/benchmarks/python/llama_jax_bench.py#L11-L26
                     theta_base *= freq_scale;
-                    for (int64_t ib = 0; ib < ne0/n_dims; ++ib) {
-                        for (int64_t ic = 0; ic < n_dims; ic += 2) {
+                    for (int64_t ic = 0; ic < ne0; ic += 2) {
+                        if (ic < n_dims) {
+                            const int64_t ib = 0;
+
                             // simplified from `(ib * n_dims + ic) * inv_ndims`
                             float cur_rot = inv_ndims * ic - ib;
 
@@ -11351,6 +11753,14 @@ static void ggml_compute_forward_rope_f16(
 
                             dst_data[0]        = GGML_FP32_TO_FP16(x0*cos_theta - x1*sin_theta);
                             dst_data[n_dims/2] = GGML_FP32_TO_FP16(x0*sin_theta + x1*cos_theta);
+                        } else {
+                            const int64_t i0 = ic;
+
+                            const ggml_fp16_t * const src = (ggml_fp16_t *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
+                                  ggml_fp16_t * dst_data  = (ggml_fp16_t *)((char *)  dst->data + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
+
+                            dst_data[0] = src[0];
+                            dst_data[1] = src[1];
                         }
                     }
                 }
@@ -11980,6 +12390,7 @@ static void ggml_compute_forward_upscale_f32(
     GGML_ASSERT(src0->nb[0] == sizeof(float));
 
     const int ith = params->ith;
+    const int nth = params->nth;
 
     GGML_TENSOR_UNARY_OP_LOCALS
 
@@ -11987,16 +12398,17 @@ static void ggml_compute_forward_upscale_f32(
 
     // TODO: optimize
 
-    for (int i03 = 0; i03 < ne03; i03++) {
-        for (int i02 = ith; i02 < ne02; i02++) {
-            for (int m = 0; m < dst->ne[1]; m++) {
-                int i01 = m / scale_factor;
-                for (int n = 0; n < dst->ne[0]; n++) {
-                    int i00 = n / scale_factor;
+    for (int64_t i3 = 0; i3 < ne3; i3++) {
+        const int64_t i03 = i3;
+        for (int64_t i2 = ith; i2 < ne2; i2 += nth) {
+            const int64_t i02 = i2;
+            for (int64_t i1 = 0; i1 < ne1; i1++) {
+                const int64_t i01 = i1 / scale_factor;
+                for (int64_t i0 = 0; i0 < ne0; i0++) {
+                    const int64_t i00 = i0 / scale_factor;
 
-                    const float * x = (float *)((char *) src0->data + i00 * nb00 +i01 * nb01 + i02 * nb02 + i03 * nb03);
-
-                    float * y = (float *)((char *) dst->data + n * dst->nb[0] + m * dst->nb[1] + i02 * dst->nb[2] + i03 * dst->nb[3]);
+                    const float * x = (float *)((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03);
+                          float * y = (float *)((char *)  dst->data +  i0*nb0  +  i1*nb1  +  i2*nb2  +  i3*nb3);
 
                     *y = *x;
                 }
@@ -12021,6 +12433,125 @@ static void ggml_compute_forward_upscale(
     }
 }
 
+// ggml_compute_forward_pad
+
+static void ggml_compute_forward_pad_f32(
+    const struct ggml_compute_params * params,
+    const struct ggml_tensor * src0,
+          struct ggml_tensor * dst) {
+
+    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
+        return;
+    }
+
+    GGML_ASSERT(src0->nb[0] == sizeof(float));
+    GGML_ASSERT( dst->nb[0] == sizeof(float));
+
+    const int ith = params->ith;
+    const int nth = params->nth;
+
+    GGML_TENSOR_UNARY_OP_LOCALS
+
+    float * dst_ptr = (float *) dst->data;
+
+    // TODO: optimize
+
+    for (int64_t i2 = 0; i2 < ne2; ++i2) {
+        for (int64_t i1 = ith; i1 < ne1; i1 += nth) {
+            for (int64_t i0 = 0; i0 < ne0; ++i0) {
+                for (int64_t i3 = 0; i3 < ne3; ++i3) {
+                    const int64_t dst_idx = i3*(ne0*ne1*ne2) + i2*(ne0*ne1) + i1*ne0 + i0;
+
+                    const float * src_ptr = (const float *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
+
+                    if (i0 < ne00 && i1 < ne01 && i2 < ne02 && i3 < ne03) {
+                        dst_ptr[dst_idx] = *src_ptr;
+                    } else {
+                        dst_ptr[dst_idx] = 0;
+                    }
+                }
+            }
+        }
+    }
+}
+
+static void ggml_compute_forward_pad(
+    const struct ggml_compute_params * params,
+    const struct ggml_tensor * src0,
+    struct ggml_tensor * dst) {
+    switch (src0->type) {
+        case GGML_TYPE_F32:
+            {
+                ggml_compute_forward_pad_f32(params, src0, dst);
+            } break;
+        default:
+            {
+                GGML_ASSERT(false);
+            } break;
+    }
+}
+
+// ggml_compute_forward_argsort
+
+static void ggml_compute_forward_argsort_f32(
+    const struct ggml_compute_params * params,
+    const struct ggml_tensor * src0,
+    struct ggml_tensor * dst) {
+
+    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
+        return;
+    }
+
+    GGML_TENSOR_UNARY_OP_LOCALS
+
+    GGML_ASSERT(nb0 == sizeof(float));
+
+    const int ith = params->ith;
+    const int nth = params->nth;
+
+    const int64_t nr = ggml_nrows(src0);
+
+    enum ggml_sort_order order = (enum ggml_sort_order) ggml_get_op_params_i32(dst, 0);
+
+    for (int64_t i = ith; i < nr; i += nth) {
+        int32_t * dst_data = (int32_t *)((char *) dst->data + i*nb1);
+        const float * src_data = (float *)((char *) src0->data + i*nb01);
+
+        for (int64_t j = 0; j < ne0; j++) {
+            dst_data[j] = j;
+        }
+
+        // C doesn't have a functional sort, so we do a bubble sort instead
+        for (int64_t j = 0; j < ne0; j++) {
+            for (int64_t k = j + 1; k < ne0; k++) {
+                if ((order == GGML_SORT_ASC && src_data[dst_data[j]] > src_data[dst_data[k]]) ||
+                    (order == GGML_SORT_DESC && src_data[dst_data[j]] < src_data[dst_data[k]])) {
+                    int32_t tmp = dst_data[j];
+                    dst_data[j] = dst_data[k];
+                    dst_data[k] = tmp;
+                }
+            }
+        }
+    }
+}
+
+static void ggml_compute_forward_argsort(
+    const struct ggml_compute_params * params,
+    const struct ggml_tensor * src0,
+    struct ggml_tensor * dst) {
+
+    switch (src0->type) {
+        case GGML_TYPE_F32:
+            {
+                ggml_compute_forward_argsort_f32(params, src0, dst);
+            } break;
+        default:
+            {
+                GGML_ASSERT(false);
+            } break;
+    }
+}
+
 // ggml_compute_forward_flash_attn
 
 static void ggml_compute_forward_flash_attn_f32(
@@ -13167,10 +13698,6 @@ static void ggml_compute_forward_unary(
             {
                 ggml_compute_forward_silu(params, src0, dst);
             } break;
-        case GGML_UNARY_OP_LEAKY:
-            {
-                ggml_compute_forward_leaky(params, src0, dst);
-            } break;
         default:
             {
                 GGML_ASSERT(false);
@@ -13844,13 +14371,17 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
             {
                 ggml_compute_forward_mul_mat(params, tensor->src[0], tensor->src[1], tensor);
             } break;
+        case GGML_OP_MUL_MAT_ID:
+            {
+                ggml_compute_forward_mul_mat_id(params, tensor->src[0], tensor->src[1], tensor);
+            } break;
         case GGML_OP_OUT_PROD:
             {
                 ggml_compute_forward_out_prod(params, tensor->src[0], tensor->src[1], tensor);
             } break;
         case GGML_OP_SCALE:
             {
-                ggml_compute_forward_scale(params, tensor->src[0], tensor->src[1], tensor);
+                ggml_compute_forward_scale(params, tensor->src[0], tensor);
             } break;
         case GGML_OP_SET:
             {
@@ -13948,6 +14479,18 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
             {
                 ggml_compute_forward_upscale(params, tensor->src[0], tensor);
             } break;
+        case GGML_OP_PAD:
+            {
+                ggml_compute_forward_pad(params, tensor->src[0], tensor);
+            } break;
+        case GGML_OP_ARGSORT:
+            {
+                ggml_compute_forward_argsort(params, tensor->src[0], tensor);
+            } break;
+        case GGML_OP_LEAKY_RELU:
+            {
+                ggml_compute_forward_leaky_relu(params, tensor->src[0], tensor);
+            } break;
         case GGML_OP_FLASH_ATTN:
             {
                 const int32_t t = ggml_get_op_params_i32(tensor, 0);
@@ -14202,7 +14745,7 @@ static struct ggml_tensor * ggml_recompute_graph_node(
         return replacements->vals[i];
     }
 
-    struct ggml_tensor * clone = ggml_new_tensor(ctx, node->type, node->n_dims, node->ne);
+    struct ggml_tensor * clone = ggml_new_tensor(ctx, node->type, GGML_MAX_DIMS, node->ne);
 
     // insert clone into replacements
     GGML_ASSERT(replacements->set.keys[i] == NULL); // assert that we don't overwrite
@@ -14272,7 +14815,7 @@ void ggml_build_backward_gradient_checkpointing(
             // insert new tensors recomputing src, reusing already made replacements,
             // remember replacements: remember new tensors with mapping from corresponding gf nodes
             // recurse for input tensors,
-            // unless (i.e. terminating when) input tensors are replacments (like checkpoints)
+            // unless (i.e. terminating when) input tensors are replacements (like checkpoints)
             node->src[k] = ggml_recompute_graph_node(ctx, gf, replacements, node->src[k]);
         }
         // insert rewritten backward node with replacements made into resulting backward graph gb
@@ -14294,7 +14837,7 @@ static struct ggml_tensor * ggml_add_or_set(struct ggml_context * ctx, struct gg
 
 static struct ggml_tensor * ggml_acc_or_set(struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * b, size_t nb1, size_t nb2, size_t nb3, size_t offset, struct ggml_hash_set zero_table) {
     if (ggml_hash_contains(zero_table, a)) {
-        struct ggml_tensor * a_zero = ggml_scale(ctx, a, ggml_new_f32(ctx, 0));
+        struct ggml_tensor * a_zero = ggml_scale(ctx, a, 0.0f);
         return ggml_acc_impl(ctx, a_zero, b, nb1, nb2, nb3, offset, false);
     } else {
         return ggml_acc_impl(ctx, a, b, nb1, nb2, nb3, offset, false);
@@ -14430,7 +14973,7 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
                                 src0->grad,
                                 ggml_scale(ctx,
                                     ggml_mul(ctx, src0, tensor->grad),
-                                    ggml_new_f32(ctx, 2.0f)),
+                                    2.0f),
                                 zero_table);
                 }
             } break;
@@ -14444,7 +14987,7 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
                                     ggml_div(ctx,
                                         tensor->grad,
                                         tensor),
-                                    ggml_new_f32(ctx, 0.5f)),
+                                    0.5f),
                                 zero_table);
                 }
             } break;
@@ -14598,6 +15141,10 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
                                 zero_table);
                 }
             } break;
+        case GGML_OP_MUL_MAT_ID:
+            {
+                GGML_ASSERT(false); // TODO: not implemented
+            } break;
         case GGML_OP_OUT_PROD:
             {
                 GGML_ASSERT(false); // TODO: not implemented
@@ -14606,17 +15153,13 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
             {
                 // necessary for llama
                 if (src0->grad) {
+                    float s;
+                    memcpy(&s, tensor->op_params, sizeof(float));
+
                     src0->grad =
                         ggml_add_or_set(ctx,
                             src0->grad,
-                            ggml_scale_impl(ctx, tensor->grad, src1, false),
-                            zero_table);
-                }
-                if (src1->grad) {
-                    src1->grad =
-                        ggml_add_or_set(ctx,
-                            src1->grad,
-                            ggml_sum(ctx, ggml_mul_impl(ctx, tensor->grad, src0, false)),
+                            ggml_scale_impl(ctx, tensor->grad, s, false),
                             zero_table);
                 }
             } break;
@@ -14794,6 +15337,8 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
                     const int n_past = ((int32_t *) tensor->op_params)[0];
                     src0->grad =
                         ggml_add_or_set(ctx, src0->grad,
+                            /* ggml_diag_mask_inf_impl() shouldn't be here */
+                            /* ref:  https://github.com/ggerganov/llama.cpp/pull/4203#discussion_r1412377992 */
                             ggml_diag_mask_zero_impl(ctx, tensor->grad, n_past, false),
                         zero_table);
                 }
@@ -14936,6 +15481,18 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
             {
                 GGML_ASSERT(false); // TODO: not implemented
             } break;
+        case GGML_OP_PAD:
+            {
+                GGML_ASSERT(false); // TODO: not implemented
+            } break;
+        case GGML_OP_ARGSORT:
+            {
+                GGML_ASSERT(false); // TODO: not implemented
+            } break;
+        case GGML_OP_LEAKY_RELU:
+            {
+                GGML_ASSERT(false); // TODO: not implemented
+            } break;
         case GGML_OP_FLASH_ATTN:
             {
                 struct ggml_tensor * flash_grad = NULL;
@@ -15296,12 +15853,8 @@ struct ggml_cgraph * ggml_new_graph(struct ggml_context * ctx) {
     return ggml_new_graph_custom(ctx, GGML_DEFAULT_GRAPH_SIZE, false);
 }
 
-struct ggml_cgraph * ggml_graph_view(struct ggml_context * ctx, struct ggml_cgraph * cgraph0, int i0, int i1) {
-    const size_t obj_size = sizeof(struct ggml_cgraph);
-    struct ggml_object * obj = ggml_new_object(ctx, GGML_OBJECT_GRAPH, obj_size);
-    struct ggml_cgraph * cgraph = (struct ggml_cgraph *) ((char *) ctx->mem_buffer + obj->offs);
-
-    *cgraph = (struct ggml_cgraph) {
+struct ggml_cgraph ggml_graph_view(struct ggml_cgraph * cgraph0, int i0, int i1) {
+    struct ggml_cgraph cgraph = {
         /*.size         =*/ 0,
         /*.n_nodes      =*/ i1 - i0,
         /*.n_leafs      =*/ 0,
@@ -15536,7 +16089,6 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
                 n_tasks = n_threads;
             } break;
         case GGML_OP_SUB:
-        case GGML_OP_DIV:
         case GGML_OP_SQR:
         case GGML_OP_SQRT:
         case GGML_OP_LOG:
@@ -15546,6 +16098,7 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
         case GGML_OP_ARGMAX:
         case GGML_OP_REPEAT:
         case GGML_OP_REPEAT_BACK:
+        case GGML_OP_LEAKY_RELU:
             {
                 n_tasks = 1;
             } break;
@@ -15558,7 +16111,6 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
                 case GGML_UNARY_OP_TANH:
                 case GGML_UNARY_OP_ELU:
                 case GGML_UNARY_OP_RELU:
-                case GGML_UNARY_OP_LEAKY:
                     {
                         n_tasks = 1;
                     } break;
@@ -15569,10 +16121,13 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
                     {
                         n_tasks = n_threads;
                     } break;
+                default:
+                    GGML_ASSERT(false);
             }
             break;
         case GGML_OP_SILU_BACK:
         case GGML_OP_MUL:
+        case GGML_OP_DIV:
         case GGML_OP_NORM:
         case GGML_OP_RMS_NORM:
         case GGML_OP_RMS_NORM_BACK:
@@ -15610,6 +16165,10 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
                 }
 #endif
             } break;
+        case GGML_OP_MUL_MAT_ID:
+            {
+                n_tasks = n_threads;
+            } break;
         case GGML_OP_OUT_PROD:
             {
                 n_tasks = n_threads;
@@ -15629,7 +16188,6 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
             } break;
         case GGML_OP_DIAG_MASK_ZERO:
         case GGML_OP_DIAG_MASK_INF:
-        case GGML_OP_SOFT_MAX:
         case GGML_OP_SOFT_MAX_BACK:
         case GGML_OP_ROPE:
         case GGML_OP_ROPE_BACK:
@@ -15645,6 +16203,10 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
             {
                 n_tasks = 1; //TODO
             } break;
+        case GGML_OP_SOFT_MAX:
+            {
+                n_tasks = MIN(MIN(4, n_threads), ggml_nrows(node->src[0]));
+            } break;
         case GGML_OP_CONV_TRANSPOSE_1D:
             {
                 n_tasks = n_threads;
@@ -15666,6 +16228,14 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
             {
                 n_tasks = n_threads;
             } break;
+        case GGML_OP_PAD:
+            {
+                n_tasks = n_threads;
+            } break;
+        case GGML_OP_ARGSORT:
+            {
+                n_tasks = n_threads;
+            } break;
         case GGML_OP_FLASH_ATTN:
             {
                 n_tasks = n_threads;
@@ -15728,6 +16298,10 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
             {
                 n_tasks = 1;
             } break;
+        case GGML_OP_COUNT:
+            {
+                GGML_ASSERT(false);
+            } break;
         default:
             {
                 fprintf(stderr, "%s: op not implemented: ", __func__);
@@ -15876,18 +16450,16 @@ struct ggml_cplan ggml_graph_plan(struct ggml_cgraph * cgraph, int n_threads) {
 
     // thread scheduling for the different operations + work buffer size estimation
     for (int i = 0; i < cgraph->n_nodes; i++) {
-        int n_tasks = 1;
-
         struct ggml_tensor * node = cgraph->nodes[i];
 
+        const int n_tasks = ggml_get_n_tasks(node, n_threads);
+
         size_t cur = 0;
 
         switch (node->op) {
             case GGML_OP_CPY:
             case GGML_OP_DUP:
                 {
-                    n_tasks = n_threads;
-
                     if (ggml_is_quantized(node->type)) {
                         cur = ggml_type_size(GGML_TYPE_F32) * node->ne[0] * n_tasks;
                     }
@@ -15895,16 +16467,12 @@ struct ggml_cplan ggml_graph_plan(struct ggml_cgraph * cgraph, int n_threads) {
             case GGML_OP_ADD:
             case GGML_OP_ADD1:
                 {
-                    n_tasks = n_threads;
-
                     if (ggml_is_quantized(node->src[0]->type)) {
                         cur = ggml_type_size(GGML_TYPE_F32) * node->src[0]->ne[0] * n_tasks;
                     }
                 } break;
             case GGML_OP_ACC:
                 {
-                    n_tasks = n_threads;
-
                     if (ggml_is_quantized(node->src[0]->type)) {
                         cur = ggml_type_size(GGML_TYPE_F32) * node->src[1]->ne[0] * n_tasks;
                     }
@@ -15927,21 +16495,30 @@ struct ggml_cplan ggml_graph_plan(struct ggml_cgraph * cgraph, int n_threads) {
                     } else
 #endif
                     if (node->src[1]->type != vec_dot_type) {
-                        cur = ggml_type_size(vec_dot_type)*ggml_nelements(node->src[1])/ggml_blck_size(vec_dot_type);
+                        cur = ggml_row_size(vec_dot_type, ggml_nelements(node->src[1]));
                     }
                 } break;
+            case GGML_OP_MUL_MAT_ID:
+                {
+                    const struct ggml_tensor * src0 = node->src[2];
+                    const struct ggml_tensor * src1 = node->src[1];
+                    const enum ggml_type vec_dot_type = type_traits[src0->type].vec_dot_type;
+                    if (src1->type != vec_dot_type) {
+                        cur = ggml_row_size(vec_dot_type, ggml_nelements(src1));
+                    }
+                    const int n_as = ggml_get_op_params_i32(node, 1);
+                    cur = GGML_PAD(cur, sizeof(int64_t));        // align
+                    cur += n_as * sizeof(int64_t);               // matrix_row_counts
+                    cur += n_as * src1->ne[1] * sizeof(int64_t); // matrix_rows
+                } break;
             case GGML_OP_OUT_PROD:
                 {
-                    n_tasks = n_threads;
-
                     if (ggml_is_quantized(node->src[0]->type)) {
                         cur = ggml_type_size(GGML_TYPE_F32) * node->src[0]->ne[0] * n_tasks;
                     }
                 } break;
             case GGML_OP_SOFT_MAX:
                 {
-                    n_tasks = MIN(MIN(4, n_threads), ggml_nrows(node->src[0]));
-
                     cur = ggml_type_size(GGML_TYPE_F32) * node->ne[0] * n_tasks;
                 } break;
             case GGML_OP_CONV_TRANSPOSE_1D:
@@ -15969,10 +16546,6 @@ struct ggml_cplan ggml_graph_plan(struct ggml_cgraph * cgraph, int n_threads) {
                         GGML_ASSERT(false);
                     }
                 } break;
-            case GGML_OP_IM2COL:
-                {
-                    n_tasks = n_threads;
-                } break;
             case GGML_OP_CONV_TRANSPOSE_2D:
                 {
                     const int64_t ne00 = node->src[0]->ne[0]; // W
@@ -15989,8 +16562,6 @@ struct ggml_cplan ggml_graph_plan(struct ggml_cgraph * cgraph, int n_threads) {
                 } break;
             case GGML_OP_FLASH_ATTN:
                 {
-                    n_tasks = n_threads;
-
                     const int64_t ne11 = ggml_up(node->src[1]->ne[1], GGML_SOFT_MAX_UNROLL);
 
                     if (node->src[1]->type == GGML_TYPE_F32) {
@@ -16003,8 +16574,6 @@ struct ggml_cplan ggml_graph_plan(struct ggml_cgraph * cgraph, int n_threads) {
                 } break;
             case GGML_OP_FLASH_FF:
                 {
-                    n_tasks = n_threads;
-
                     if (node->src[1]->type == GGML_TYPE_F32) {
                         cur  = sizeof(float)*node->src[1]->ne[1]*n_tasks; // TODO: this can become (n_tasks-1)
                         cur += sizeof(float)*node->src[1]->ne[1]*n_tasks; // this is overestimated by x2
@@ -16015,8 +16584,6 @@ struct ggml_cplan ggml_graph_plan(struct ggml_cgraph * cgraph, int n_threads) {
                 } break;
             case GGML_OP_FLASH_ATTN_BACK:
                 {
-                    n_tasks = n_threads;
-
                     const int64_t    D = node->src[0]->ne[0];
                     const int64_t ne11 = ggml_up(node->src[1]->ne[1], GGML_SOFT_MAX_UNROLL);
                     const int64_t mxDn = MAX(D, ne11) * 2; // *2 because of S and SM in ggml_compute_forward_flash_attn_back
@@ -16031,8 +16598,6 @@ struct ggml_cplan ggml_graph_plan(struct ggml_cgraph * cgraph, int n_threads) {
 
             case GGML_OP_CROSS_ENTROPY_LOSS:
                 {
-                    n_tasks = n_threads;
-
                     cur = ggml_type_size(node->type)*(n_tasks + node->src[0]->ne[0]*n_tasks);
                 } break;
             case GGML_OP_COUNT:
@@ -16174,7 +16739,7 @@ static void ggml_graph_export_leaf(const struct ggml_tensor * tensor, FILE * fou
     fprintf(fout, "%-6s %-12s %8d %" PRId64 " %" PRId64 " %" PRId64 " %" PRId64 " %16zu %16zu %16zu %16zu %16p %32s\n",
             ggml_type_name(tensor->type),
             ggml_op_name  (tensor->op),
-            tensor->n_dims,
+            ggml_n_dims(tensor),
             ne[0], ne[1], ne[2], ne[3],
             nb[0], nb[1], nb[2], nb[3],
             tensor->data,
@@ -16189,7 +16754,7 @@ static void ggml_graph_export_node(const struct ggml_tensor * tensor, const char
             arg,
             ggml_type_name(tensor->type),
             ggml_op_name  (tensor->op),
-            tensor->n_dims,
+            ggml_n_dims(tensor),
             ne[0], ne[1], ne[2], ne[3],
             nb[0], nb[1], nb[2], nb[3],
             tensor->data,
@@ -16279,11 +16844,9 @@ void ggml_graph_export(const struct ggml_cgraph * cgraph, const char * fname) {
 
                 const uint32_t type   = tensor->type;
                 const uint32_t op     = tensor->op;
-                const uint32_t n_dims = tensor->n_dims;
 
                 fwrite(&type,   sizeof(uint32_t), 1, fout);
                 fwrite(&op,     sizeof(uint32_t), 1, fout);
-                fwrite(&n_dims, sizeof(uint32_t), 1, fout);
 
                 for (int j = 0; j < GGML_MAX_DIMS; ++j) {
                     const uint64_t ne = tensor->ne[j];
@@ -16313,11 +16876,9 @@ void ggml_graph_export(const struct ggml_cgraph * cgraph, const char * fname) {
 
                 const uint32_t type   = tensor->type;
                 const uint32_t op     = tensor->op;
-                const uint32_t n_dims = tensor->n_dims;
 
                 fwrite(&type,   sizeof(uint32_t), 1, fout);
                 fwrite(&op,     sizeof(uint32_t), 1, fout);
-                fwrite(&n_dims, sizeof(uint32_t), 1, fout);
 
                 for (int j = 0; j < GGML_MAX_DIMS; ++j) {
                     const uint64_t ne = tensor->ne[j];
@@ -16489,12 +17050,10 @@ struct ggml_cgraph * ggml_graph_import(const char * fname, struct ggml_context *
         {
             uint32_t type;
             uint32_t op;
-            uint32_t n_dims;
 
             for (uint32_t i = 0; i < n_leafs; ++i) {
                 type   = *(const uint32_t *) ptr; ptr += sizeof(type);
                 op     = *(const uint32_t *) ptr; ptr += sizeof(op);
-                n_dims = *(const uint32_t *) ptr; ptr += sizeof(n_dims);
 
                 int64_t ne[GGML_MAX_DIMS];
                 size_t  nb[GGML_MAX_DIMS];
@@ -16510,7 +17069,7 @@ struct ggml_cgraph * ggml_graph_import(const char * fname, struct ggml_context *
                     nb[j] = nb_cur;
                 }
 
-                struct ggml_tensor * tensor = ggml_new_tensor(*ctx_eval, (enum ggml_type) type, n_dims, ne);
+                struct ggml_tensor * tensor = ggml_new_tensor(*ctx_eval, (enum ggml_type) type, GGML_MAX_DIMS, ne);
 
                 tensor->op = (enum ggml_op) op;
 
@@ -16527,7 +17086,7 @@ struct ggml_cgraph * ggml_graph_import(const char * fname, struct ggml_context *
 
                 ptr += ggml_nbytes(tensor);
 
-                fprintf(stderr, "%s: loaded leaf %d: '%16s', %3d dims, %9zu bytes\n", __func__, i, tensor->name, n_dims, ggml_nbytes(tensor));
+                fprintf(stderr, "%s: loaded leaf %d: '%16s', %9zu bytes\n", __func__, i, tensor->name, ggml_nbytes(tensor));
             }
         }
 
@@ -16537,12 +17096,10 @@ struct ggml_cgraph * ggml_graph_import(const char * fname, struct ggml_context *
         {
             uint32_t type;
             uint32_t op;
-            uint32_t n_dims;
 
             for (uint32_t i = 0; i < n_nodes; ++i) {
                 type   = *(const uint32_t *) ptr; ptr += sizeof(type);
                 op     = *(const uint32_t *) ptr; ptr += sizeof(op);
-                n_dims = *(const uint32_t *) ptr; ptr += sizeof(n_dims);
 
                 enum ggml_op eop = (enum ggml_op) op;
 
@@ -16613,7 +17170,7 @@ struct ggml_cgraph * ggml_graph_import(const char * fname, struct ggml_context *
                         } break;
                     default:
                         {
-                            tensor = ggml_new_tensor(*ctx_eval, (enum ggml_type) type, n_dims, ne);
+                            tensor = ggml_new_tensor(*ctx_eval, (enum ggml_type) type, GGML_MAX_DIMS, ne);
 
                             tensor->op = eop;
                         } break;
@@ -16632,7 +17189,7 @@ struct ggml_cgraph * ggml_graph_import(const char * fname, struct ggml_context *
 
                 result->nodes[i] = tensor;
 
-                fprintf(stderr, "%s: loaded node %d: '%16s', %3d dims, %9zu bytes\n", __func__, i, tensor->name, n_dims, ggml_nbytes(tensor));
+                fprintf(stderr, "%s: loaded node %d: '%16s', %9zu bytes\n", __func__, i, tensor->name, ggml_nbytes(tensor));
             }
         }
     }
@@ -16770,7 +17327,7 @@ void ggml_graph_dump_dot(const struct ggml_cgraph * gb, const struct ggml_cgraph
             fprintf(fp, "(%s)|", ggml_type_name(node->type));
         }
 
-        if (node->n_dims == 2) {
+        if (ggml_is_matrix(node)) {
             fprintf(fp, "%d [%" PRId64 ", %" PRId64 "] | <x>%s", i, node->ne[0], node->ne[1], ggml_op_symbol(node->op));
         } else {
             fprintf(fp, "%d [%" PRId64 ", %" PRId64 ", %" PRId64 "] | <x>%s", i, node->ne[0], node->ne[1], node->ne[2], ggml_op_symbol(node->op));
@@ -17037,7 +17594,7 @@ static enum ggml_opt_result ggml_opt_adam(
             int64_t i = 0;
             for (int p = 0; p < np; ++p) {
                 const int64_t ne = ggml_nelements(ps[p]);
-                const float p_decay = ((ps[p]->n_dims >= decay_min_ndim) ? decay : 0.0f) * sched;
+                const float p_decay = ((ggml_n_dims(ps[p]) >= decay_min_ndim) ? decay : 0.0f) * sched;
                 for (int64_t j = 0; j < ne; ++j) {
                     float x  = ggml_get_f32_1d(ps[p], j);
                     float g_ = g[i]*gnorm;
@@ -17819,8 +18376,8 @@ size_t ggml_quantize_q5_0(const float * src, void * dst, int n, int k, int64_t *
             memcpy(&qh, &y[i].qh, sizeof(qh));
 
             for (int j = 0; j < QK5_0; j += 2) {
-                const uint8_t vh0 = ((qh & (1u << (j + 0 ))) >> (j + 0 )) << 4;
-                const uint8_t vh1 = ((qh & (1u << (j + 16))) >> (j + 12));
+                const uint8_t vh0 = ((qh & (1u << (j/2 + 0 ))) >> (j/2 + 0 )) << 4;
+                const uint8_t vh1 = ((qh & (1u << (j/2 + 16))) >> (j/2 + 12));
 
                 // cast to 16 bins
                 const uint8_t vi0 = ((y[i].qs[j/2] & 0x0F) | vh0) / 2;
@@ -17849,8 +18406,8 @@ size_t ggml_quantize_q5_1(const float * src, void * dst, int n, int k, int64_t *
             memcpy(&qh, &y[i].qh, sizeof(qh));
 
             for (int j = 0; j < QK5_1; j += 2) {
-                const uint8_t vh0 = ((qh & (1u << (j + 0 ))) >> (j + 0 )) << 4;
-                const uint8_t vh1 = ((qh & (1u << (j + 16))) >> (j + 12));
+                const uint8_t vh0 = ((qh & (1u << (j/2 + 0 ))) >> (j/2 + 0 )) << 4;
+                const uint8_t vh1 = ((qh & (1u << (j/2 + 16))) >> (j/2 + 12));
 
                 // cast to 16 bins
                 const uint8_t vi0 = ((y[i].qs[j/2] & 0x0F) | vh0) / 2;
@@ -18040,6 +18597,7 @@ struct gguf_kv {
 
 struct gguf_header {
     char magic[4];
+
     uint32_t version;
     uint64_t n_tensors; // GGUFv2
     uint64_t n_kv;      // GGUFv2
@@ -18129,7 +18687,7 @@ struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_p
 
         for (uint32_t i = 0; i < sizeof(magic); i++) {
             if (magic[i] != GGUF_MAGIC[i]) {
-                fprintf(stderr, "%s: invalid magic characters %s.\n", __func__, magic);
+                fprintf(stderr, "%s: invalid magic characters '%c%c%c%c'\n", __func__, magic[0], magic[1], magic[2], magic[3]);
                 fclose(file);
                 return NULL;
             }
@@ -18144,7 +18702,6 @@ struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_p
     {
         strncpy(ctx->header.magic, magic, 4);
 
-
         ctx->kv    = NULL;
         ctx->infos = NULL;
         ctx->data  = NULL;
@@ -18311,7 +18868,7 @@ struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_p
                 return NULL;
             }
 
-            const size_t size_cur = (ne*ggml_type_size(info->type))/ggml_blck_size(info->type);
+            const size_t size_cur = ggml_row_size(info->type, ne);
 
             ctx->size += GGML_PAD(size_cur, ctx->alignment);
         }
@@ -18640,6 +19197,10 @@ char * gguf_get_tensor_name(const struct gguf_context * ctx, int i) {
     return ctx->infos[i].name.data;
 }
 
+enum ggml_type gguf_get_tensor_type(const struct gguf_context * ctx, int i) {
+    return ctx->infos[i].type;
+}
+
 // returns the index
 static int gguf_get_or_add_key(struct gguf_context * ctx, const char * key) {
     const int idx = gguf_find_key(ctx, key);
@@ -18815,8 +19376,8 @@ void gguf_add_tensor(
         ctx->infos[idx].ne[i] = 1;
     }
 
-    ctx->infos[idx].n_dims = tensor->n_dims;
-    for (int i = 0; i < tensor->n_dims; i++) {
+    ctx->infos[idx].n_dims = ggml_n_dims(tensor);
+    for (uint32_t i = 0; i < ctx->infos[idx].n_dims; i++) {
         ctx->infos[idx].ne[i] = tensor->ne[i];
     }
 
diff --git a/ggml.h b/ggml.h
index 2f6787d4e..338f355a4 100644
--- a/ggml.h
+++ b/ggml.h
@@ -215,9 +215,9 @@
 #define GGML_QNT_VERSION_FACTOR 1000 // do not change this
 
 #define GGML_MAX_DIMS           4
-#define GGML_MAX_PARAMS         1024
+#define GGML_MAX_PARAMS         2048
 #define GGML_MAX_CONTEXTS       64
-#define GGML_MAX_SRC            6
+#define GGML_MAX_SRC            10
 #define GGML_MAX_NAME           64
 #define GGML_MAX_OP_PARAMS      64
 #define GGML_DEFAULT_N_THREADS  4
@@ -283,13 +283,27 @@
     const type prefix##3 = (pointer)->array[3]; \
     GGML_UNUSED(prefix##3);
 
+#define GGML_TENSOR_UNARY_OP_LOCALS \
+    GGML_TENSOR_LOCALS(int64_t, ne0, src0, ne) \
+    GGML_TENSOR_LOCALS(size_t,  nb0, src0, nb) \
+    GGML_TENSOR_LOCALS(int64_t, ne,  dst,  ne) \
+    GGML_TENSOR_LOCALS(size_t,  nb,  dst,  nb)
+
+#define GGML_TENSOR_BINARY_OP_LOCALS \
+    GGML_TENSOR_LOCALS(int64_t, ne0, src0, ne) \
+    GGML_TENSOR_LOCALS(size_t,  nb0, src0, nb) \
+    GGML_TENSOR_LOCALS(int64_t, ne1, src1, ne) \
+    GGML_TENSOR_LOCALS(size_t,  nb1, src1, nb) \
+    GGML_TENSOR_LOCALS(int64_t, ne,  dst,  ne) \
+    GGML_TENSOR_LOCALS(size_t,  nb,  dst,  nb)
+
 #ifdef  __cplusplus
 extern "C" {
 #endif
 
 #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;
@@ -329,6 +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,
@@ -381,6 +401,7 @@ extern "C" {
         GGML_OP_GROUP_NORM,
 
         GGML_OP_MUL_MAT,
+        GGML_OP_MUL_MAT_ID,
         GGML_OP_OUT_PROD,
 
         GGML_OP_SCALE,
@@ -407,8 +428,10 @@ extern "C" {
         GGML_OP_CONV_TRANSPOSE_2D,
         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,
@@ -448,7 +471,8 @@ extern "C" {
         GGML_UNARY_OP_GELU,
         GGML_UNARY_OP_GELU_QUICK,
         GGML_UNARY_OP_SILU,
-        GGML_UNARY_OP_LEAKY
+
+        GGML_UNARY_OP_COUNT,
     };
 
     enum ggml_object_type {
@@ -460,7 +484,8 @@ extern "C" {
     enum ggml_log_level {
         GGML_LOG_LEVEL_ERROR = 2,
         GGML_LOG_LEVEL_WARN = 3,
-        GGML_LOG_LEVEL_INFO = 4
+        GGML_LOG_LEVEL_INFO = 4,
+        GGML_LOG_LEVEL_DEBUG = 5
     };
 
     // ggml object
@@ -484,7 +509,6 @@ 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)
@@ -516,7 +540,7 @@ 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);
@@ -621,16 +645,22 @@ 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);
     GGML_API const char * ggml_op_symbol(enum ggml_op   op);
 
+    GGML_API const char * ggml_unary_op_name(enum ggml_unary_op op);
+    GGML_API const char * ggml_op_desc(const struct ggml_tensor * t); // unary or op name
+
     GGML_API size_t  ggml_element_size(const struct ggml_tensor * tensor);
 
     GGML_API bool    ggml_is_quantized(enum ggml_type type);
@@ -641,6 +671,11 @@ 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);
 
@@ -701,8 +736,8 @@ extern "C" {
     GGML_API struct ggml_tensor * ggml_view_tensor(struct ggml_context * ctx, struct ggml_tensor * src);
 
     // Context tensor enumeration and lookup
-    GGML_API struct ggml_tensor * ggml_get_first_tensor(struct ggml_context * ctx);
-    GGML_API struct ggml_tensor * ggml_get_next_tensor (struct ggml_context * ctx, struct ggml_tensor * tensor);
+    GGML_API struct ggml_tensor * ggml_get_first_tensor(const struct ggml_context * ctx);
+    GGML_API struct ggml_tensor * ggml_get_next_tensor (const struct ggml_context * ctx, struct ggml_tensor * tensor);
     GGML_API struct ggml_tensor * ggml_get_tensor(struct ggml_context * ctx, const char * name);
 
     GGML_API struct ggml_tensor * ggml_set_zero(struct ggml_tensor * tensor);
@@ -773,6 +808,9 @@ 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,
@@ -937,15 +975,14 @@ 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);
@@ -1027,6 +1064,22 @@ 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(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * const as[],
+            int                   n_as,
+            struct ggml_tensor  * ids,
+            int                   id,
+            struct ggml_tensor  * b);
+
     // A: m columns, n rows,
     // B: p columns, n rows,
     // result is m columns, p rows
@@ -1042,13 +1095,13 @@ extern "C" {
     GGML_API struct ggml_tensor * ggml_scale(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
-            struct ggml_tensor  * b);
+            float                 s);
 
     // in-place, returns view(a)
     GGML_API struct ggml_tensor * ggml_scale_inplace(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
-            struct ggml_tensor  * b);
+            float                 s);
 
     // b -> view(a,offset,nb1,nb2,3), return modified a
     GGML_API struct ggml_tensor * ggml_set(
@@ -1234,6 +1287,7 @@ extern "C" {
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
 
+    // supports 3D: a->ne[2] == b->ne[1]
     GGML_API struct ggml_tensor * ggml_get_rows(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
@@ -1520,6 +1574,32 @@ 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,
+        GGML_SORT_DESC,
+    };
+
+    GGML_API struct ggml_tensor * ggml_argsort(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            enum ggml_sort_order  order);
+
+    // top k elements per row
+    GGML_API struct ggml_tensor * ggml_top_k(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            int                   k);
+
     GGML_API struct ggml_tensor * ggml_flash_attn(
             struct ggml_context * ctx,
             struct ggml_tensor  * q,
@@ -1581,7 +1661,6 @@ extern "C" {
             int                   kh);
 
     // used in sam
-
     GGML_API struct ggml_tensor * ggml_add_rel_pos(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
@@ -1756,7 +1835,7 @@ extern "C" {
     GGML_API struct ggml_cgraph * ggml_new_graph         (struct ggml_context * ctx); // size = GGML_DEFAULT_GRAPH_SIZE, grads = false
     GGML_API struct ggml_cgraph * ggml_new_graph_custom  (struct ggml_context * ctx, size_t size, bool grads);
     GGML_API struct ggml_cgraph * ggml_graph_dup         (struct ggml_context * ctx, struct ggml_cgraph * cgraph);
-    GGML_API struct ggml_cgraph * ggml_graph_view        (struct ggml_context * ctx, struct ggml_cgraph * cgraph, int i0, int i1);
+    GGML_API struct ggml_cgraph   ggml_graph_view        (struct ggml_cgraph * cgraph, int i0, int i1);
     GGML_API void                 ggml_graph_cpy         (struct ggml_cgraph * src, struct ggml_cgraph * dst);
     GGML_API void                 ggml_graph_reset       (struct ggml_cgraph * cgraph);  // zero grads
     GGML_API void                 ggml_graph_clear       (struct ggml_cgraph * cgraph);
@@ -2057,10 +2136,11 @@ extern "C" {
     GGML_API const void * gguf_get_arr_data(const struct gguf_context * ctx, int key_id);
     GGML_API const char * gguf_get_arr_str (const struct gguf_context * ctx, int key_id, int i);
 
-    GGML_API int    gguf_get_n_tensors    (const struct gguf_context * ctx);
-    GGML_API int    gguf_find_tensor      (const struct gguf_context * ctx, const char * name);
-    GGML_API size_t gguf_get_tensor_offset(const struct gguf_context * ctx, int i);
-    GGML_API char * gguf_get_tensor_name  (const struct gguf_context * ctx, int i);
+    GGML_API int            gguf_get_n_tensors    (const struct gguf_context * ctx);
+    GGML_API int            gguf_find_tensor      (const struct gguf_context * ctx, const char * name);
+    GGML_API size_t         gguf_get_tensor_offset(const struct gguf_context * ctx, int i);
+    GGML_API char *         gguf_get_tensor_name  (const struct gguf_context * ctx, int i);
+    GGML_API enum ggml_type gguf_get_tensor_type  (const struct gguf_context * ctx, int i);
 
     // overrides existing values or adds a new one
     GGML_API void gguf_set_val_u8  (struct gguf_context * ctx, const char * key, uint8_t  val);
diff --git a/gguf-py/README.md b/gguf-py/README.md
index 502b6a510..22d7ffa52 100644
--- a/gguf-py/README.md
+++ b/gguf-py/README.md
@@ -3,7 +3,7 @@
 This is a Python package for writing binary files in the [GGUF](https://github.com/ggerganov/ggml/pull/302)
 (GGML Universal File) format.
 
-See [convert-llama-hf-to-gguf.py](https://github.com/ggerganov/llama.cpp/blob/master/convert-llama-hf-to-gguf.py)
+See [convert-llama-hf-to-gguf.py](https://github.com/ggerganov/llama.cpp/blob/master/convert-hf-to-gguf.py)
 as an example for its usage.
 
 ## Installation
@@ -61,7 +61,7 @@ If you want to publish the package manually for any reason, you need to have `tw
 pip install build twine
 ```
 
-Then, folow these steps to release a new version:
+Then, follow these steps to release a new version:
 
 1. Bump the version in `pyproject.toml`.
 2. Build the package:
diff --git a/gguf-py/gguf/constants.py b/gguf-py/gguf/constants.py
index 685c88f1a..390dca049 100644
--- a/gguf-py/gguf/constants.py
+++ b/gguf-py/gguf/constants.py
@@ -38,6 +38,8 @@ class Keys:
         FEED_FORWARD_LENGTH   = "{arch}.feed_forward_length"
         USE_PARALLEL_RESIDUAL = "{arch}.use_parallel_residual"
         TENSOR_DATA_LAYOUT    = "{arch}.tensor_data_layout"
+        EXPERT_COUNT          = "{arch}.expert_count"
+        EXPERT_USED_COUNT     = "{arch}.expert_used_count"
 
     class Attention:
         HEAD_COUNT        = "{arch}.attention.head_count"
@@ -93,6 +95,7 @@ class MODEL_ARCH(IntEnum):
     BLOOM     = auto()
     STABLELM  = auto()
     QWEN      = auto()
+    PHI2      = auto()
 
 
 class MODEL_TENSOR(IntEnum):
@@ -111,10 +114,14 @@ class MODEL_TENSOR(IntEnum):
     ATTN_NORM       = auto()
     ATTN_NORM_2     = auto()
     ATTN_ROT_EMBD   = auto()
+    FFN_GATE_INP    = auto()
+    FFN_NORM        = auto()
     FFN_GATE        = auto()
     FFN_DOWN        = auto()
     FFN_UP          = auto()
-    FFN_NORM        = auto()
+    FFN_GATE_EXP    = auto()
+    FFN_DOWN_EXP    = auto()
+    FFN_UP_EXP      = auto()
     ATTN_Q_NORM     = auto()
     ATTN_K_NORM     = auto()
 
@@ -134,6 +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] = {
@@ -154,10 +162,14 @@ TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
     MODEL_TENSOR.ATTN_ROT_EMBD:   "blk.{bid}.attn_rot_embd",
     MODEL_TENSOR.ATTN_Q_NORM:     "blk.{bid}.attn_q_norm",
     MODEL_TENSOR.ATTN_K_NORM:     "blk.{bid}.attn_k_norm",
+    MODEL_TENSOR.FFN_GATE_INP:    "blk.{bid}.ffn_gate_inp",
     MODEL_TENSOR.FFN_NORM:        "blk.{bid}.ffn_norm",
     MODEL_TENSOR.FFN_GATE:        "blk.{bid}.ffn_gate",
     MODEL_TENSOR.FFN_DOWN:        "blk.{bid}.ffn_down",
     MODEL_TENSOR.FFN_UP:          "blk.{bid}.ffn_up",
+    MODEL_TENSOR.FFN_GATE_EXP:    "blk.{bid}.ffn_gate.{xid}",
+    MODEL_TENSOR.FFN_DOWN_EXP:    "blk.{bid}.ffn_down.{xid}",
+    MODEL_TENSOR.FFN_UP_EXP:      "blk.{bid}.ffn_up.{xid}",
 }
 
 MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
@@ -172,10 +184,14 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.ATTN_V,
         MODEL_TENSOR.ATTN_OUT,
         MODEL_TENSOR.ATTN_ROT_EMBD,
+        MODEL_TENSOR.FFN_GATE_INP,
         MODEL_TENSOR.FFN_NORM,
         MODEL_TENSOR.FFN_GATE,
         MODEL_TENSOR.FFN_DOWN,
         MODEL_TENSOR.FFN_UP,
+        MODEL_TENSOR.FFN_GATE_EXP,
+        MODEL_TENSOR.FFN_DOWN_EXP,
+        MODEL_TENSOR.FFN_UP_EXP,
     ],
     MODEL_ARCH.GPTNEOX: [
         MODEL_TENSOR.TOKEN_EMBD,
@@ -336,6 +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
 }
 
diff --git a/gguf-py/gguf/gguf_writer.py b/gguf-py/gguf/gguf_writer.py
index b8ec977c8..73e021607 100644
--- a/gguf-py/gguf/gguf_writer.py
+++ b/gguf-py/gguf/gguf_writer.py
@@ -339,6 +339,12 @@ class GGUFWriter:
     def add_clamp_kqv(self, value: float) -> None:
         self.add_float32(Keys.Attention.CLAMP_KQV.format(arch=self.arch), value)
 
+    def add_expert_count(self, count: int) -> None:
+        self.add_uint32(Keys.LLM.EXPERT_COUNT.format(arch=self.arch), count)
+
+    def add_expert_used_count(self, count: int) -> None:
+        self.add_uint32(Keys.LLM.EXPERT_USED_COUNT.format(arch=self.arch), count)
+
     def add_layer_norm_eps(self, value: float) -> None:
         self.add_float32(Keys.Attention.LAYERNORM_EPS.format(arch=self.arch), value)
 
diff --git a/gguf-py/gguf/tensor_mapping.py b/gguf-py/gguf/tensor_mapping.py
index cc6236014..6fcbdbc1c 100644
--- a/gguf-py/gguf/tensor_mapping.py
+++ b/gguf-py/gguf/tensor_mapping.py
@@ -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
@@ -149,6 +155,11 @@ class TensorNameMap:
             "model.layers.{bid}.ln2",                                        # yi
         ),
 
+        MODEL_TENSOR.FFN_GATE_INP: (
+            "layers.{bid}.feed_forward.gate",           # mixtral
+            "model.layers.{bid}.block_sparse_moe.gate", # mixtral
+        ),
+
         # Feed-forward up
         MODEL_TENSOR.FFN_UP: (
             "gpt_neox.layers.{bid}.mlp.dense_h_to_4h",                # gptneox
@@ -162,13 +173,24 @@ 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: (
+            "layers.{bid}.feed_forward.experts.{xid}.w3",           # mixtral
+            "model.layers.{bid}.block_sparse_moe.experts.{xid}.w3", # mixtral
         ),
 
         # Feed-forward gate
         MODEL_TENSOR.FFN_GATE: (
-            "model.layers.{bid}.mlp.gate_proj",  # llama-hf refact
-            "layers.{bid}.feed_forward.w1",      # llama-pth
-            "transformer.h.{bid}.mlp.w2",        # qwen
+            "model.layers.{bid}.mlp.gate_proj",           # llama-hf refact
+            "layers.{bid}.feed_forward.w1",               # llama-pth
+            "transformer.h.{bid}.mlp.w2",                 # qwen
+        ),
+
+        MODEL_TENSOR.FFN_GATE_EXP: (
+            "layers.{bid}.feed_forward.experts.{xid}.w1",           # mixtral
+            "model.layers.{bid}.block_sparse_moe.experts.{xid}.w1", # mixtral
         ),
 
         # Feed-forward down
@@ -183,6 +205,12 @@ 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: (
+            "layers.{bid}.feed_forward.experts.{xid}.w2",           # mixtral
+            "model.layers.{bid}.block_sparse_moe.experts.{xid}.w2", # mixtral
         ),
 
         MODEL_TENSOR.ATTN_Q_NORM: (
@@ -213,11 +241,14 @@ class TensorNameMap:
             for tensor, keys in self.block_mappings_cfg.items():
                 if tensor not in MODEL_TENSORS[arch]:
                     continue
-                tensor_name = TENSOR_NAMES[tensor].format(bid = bid)
-                self.mapping[tensor_name] = (tensor, tensor_name)
-                for key in keys:
-                    key = key.format(bid = bid)
-                    self.mapping[key] = (tensor, tensor_name)
+                # TODO: make this configurable
+                n_experts = 8
+                for xid in range(n_experts):
+                    tensor_name = TENSOR_NAMES[tensor].format(bid = bid, xid = xid)
+                    self.mapping[tensor_name] = (tensor, tensor_name)
+                    for key in keys:
+                        key = key.format(bid = bid, xid = xid)
+                        self.mapping[key] = (tensor, tensor_name)
 
     def get_type_and_name(self, key: str, try_suffixes: Sequence[str] = ()) -> tuple[MODEL_TENSOR, str] | None:
         result = self.mapping.get(key)
diff --git a/gguf-py/gguf/vocab.py b/gguf-py/gguf/vocab.py
index de3e5edb5..cd1942975 100644
--- a/gguf-py/gguf/vocab.py
+++ b/gguf-py/gguf/vocab.py
@@ -84,7 +84,7 @@ class SpecialVocab:
         merges_file = path / 'merges.txt'
         if not merges_file.is_file():
             return False
-        with open(merges_file, 'r') 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
diff --git a/gguf-py/pyproject.toml b/gguf-py/pyproject.toml
index e6374bfe8..9789c2c87 100644
--- a/gguf-py/pyproject.toml
+++ b/gguf-py/pyproject.toml
@@ -1,6 +1,6 @@
 [tool.poetry]
 name = "gguf"
-version = "0.6.0"
+version = "0.7.0"
 description = "Read and write ML models in GGUF for GGML"
 authors = ["GGML <ggml@ggml.ai>"]
 packages = [
diff --git a/llama.cpp b/llama.cpp
index fd905ade7..4e4495739 100644
--- a/llama.cpp
+++ b/llama.cpp
@@ -1,11 +1,12 @@
 #define LLAMA_API_INTERNAL
+//#define LLAMA_GGML_BACKEND_CUDA_TEST // for testing only - enables ggml-cuda through ggml-backend, disables partial offloading
 #include "llama.h"
 
 #include "unicode.h"
 
 #include "ggml.h"
-
 #include "ggml-alloc.h"
+#include "ggml-backend.h"
 
 #ifdef GGML_USE_CUBLAS
 #  include "ggml-cuda.h"
@@ -32,6 +33,7 @@
         #include <unistd.h>
         #if defined(_POSIX_MAPPED_FILES)
             #include <sys/mman.h>
+            #include <fcntl.h>
         #endif
         #if defined(_POSIX_MEMLOCK_RANGE)
             #include <sys/resource.h>
@@ -74,6 +76,7 @@
 #include <set>
 #include <sstream>
 #include <thread>
+#include <type_traits>
 #include <unordered_map>
 
 #if defined(_MSC_VER)
@@ -90,7 +93,8 @@
 #define LLAMA_ATTRIBUTE_FORMAT(...)
 #endif
 
-#define LLAMA_MAX_NODES 8192
+#define LLAMA_MAX_NODES   8192
+#define LLAMA_MAX_EXPERTS 8
 
 //
 // logging
@@ -193,6 +197,7 @@ enum llm_arch {
     LLM_ARCH_BLOOM,
     LLM_ARCH_STABLELM,
     LLM_ARCH_QWEN,
+    LLM_ARCH_PHI2,
     LLM_ARCH_UNKNOWN,
 };
 
@@ -210,6 +215,7 @@ 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 {
@@ -230,6 +236,8 @@ enum llm_kv {
     LLM_KV_FEED_FORWARD_LENGTH,
     LLM_KV_USE_PARALLEL_RESIDUAL,
     LLM_KV_TENSOR_DATA_LAYOUT,
+    LLM_KV_EXPERT_COUNT,
+    LLM_KV_EXPERT_USED_COUNT,
 
     LLM_KV_ATTENTION_HEAD_COUNT,
     LLM_KV_ATTENTION_HEAD_COUNT_KV,
@@ -280,6 +288,8 @@ static std::map<llm_kv, std::string> LLM_KV_NAMES = {
     { LLM_KV_FEED_FORWARD_LENGTH,           "%s.feed_forward_length"   },
     { LLM_KV_USE_PARALLEL_RESIDUAL,         "%s.use_parallel_residual" },
     { LLM_KV_TENSOR_DATA_LAYOUT,            "%s.tensor_data_layout"    },
+    { LLM_KV_EXPERT_COUNT,                  "%s.expert_count"          },
+    { LLM_KV_EXPERT_USED_COUNT,             "%s.expert_used_count"     },
 
     { LLM_KV_ATTENTION_HEAD_COUNT,          "%s.attention.head_count"             },
     { LLM_KV_ATTENTION_HEAD_COUNT_KV,       "%s.attention.head_count_kv"          },
@@ -337,10 +347,14 @@ enum llm_tensor {
     LLM_TENSOR_ATTN_NORM,
     LLM_TENSOR_ATTN_NORM_2,
     LLM_TENSOR_ATTN_ROT_EMBD,
+    LLM_TENSOR_FFN_GATE_INP,
+    LLM_TENSOR_FFN_NORM,
     LLM_TENSOR_FFN_GATE,
     LLM_TENSOR_FFN_DOWN,
     LLM_TENSOR_FFN_UP,
-    LLM_TENSOR_FFN_NORM,
+    LLM_TENSOR_FFN_DOWN_EXP,
+    LLM_TENSOR_FFN_GATE_EXP,
+    LLM_TENSOR_FFN_UP_EXP,
     LLM_TENSOR_ATTN_Q_NORM,
     LLM_TENSOR_ATTN_K_NORM,
 };
@@ -359,10 +373,14 @@ static std::map<llm_arch, std::map<llm_tensor, std::string>> LLM_TENSOR_NAMES =
             { LLM_TENSOR_ATTN_V,          "blk.%d.attn_v" },
             { LLM_TENSOR_ATTN_OUT,        "blk.%d.attn_output" },
             { LLM_TENSOR_ATTN_ROT_EMBD,   "blk.%d.attn_rot_embd" },
+            { LLM_TENSOR_FFN_GATE_INP,    "blk.%d.ffn_gate_inp" },
             { LLM_TENSOR_FFN_NORM,        "blk.%d.ffn_norm" },
             { LLM_TENSOR_FFN_GATE,        "blk.%d.ffn_gate" },
             { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down" },
             { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
+            { LLM_TENSOR_FFN_GATE_EXP,    "blk.%d.ffn_gate.%d" },
+            { LLM_TENSOR_FFN_DOWN_EXP,    "blk.%d.ffn_down.%d" },
+            { LLM_TENSOR_FFN_UP_EXP,      "blk.%d.ffn_up.%d" },
         },
     },
     {
@@ -536,6 +554,19 @@ 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,
@@ -584,27 +615,16 @@ struct LLM_TN {
     std::string operator()(llm_tensor tensor, const std::string & suffix, int bid) const {
         return ::format(LLM_TENSOR_NAMES[arch].at(tensor).c_str(), bid) + "." + suffix;
     }
+
+    std::string operator()(llm_tensor tensor, const std::string & suffix, int bid, int xid) const {
+        return ::format(LLM_TENSOR_NAMES[arch].at(tensor).c_str(), bid, xid) + "." + suffix;
+    }
 };
 
 //
 // gguf helpers
 //
 
-#define GGUF_GET_KEY(ctx, dst, func, type, req, key) \
-do { \
-    const std::string skey(key); \
-    const int kid = gguf_find_key(ctx, skey.c_str()); \
-    if (kid >= 0) { \
-        enum gguf_type ktype = gguf_get_kv_type(ctx, kid); \
-        if (ktype != (type)) { \
-            throw std::runtime_error(format("key %s has wrong type: %s", skey.c_str(), gguf_type_name(ktype))); \
-        } \
-        (dst) = func(ctx, kid); \
-    } else if (req) { \
-        throw std::runtime_error(format("key not found in model: %s", skey.c_str())); \
-    } \
-} while (0)
-
 static std::map<int8_t, std::string> LLAMA_ROPE_SCALING_TYPES = {
     { LLAMA_ROPE_SCALING_NONE,   "none"   },
     { LLAMA_ROPE_SCALING_LINEAR, "linear" },
@@ -638,7 +658,7 @@ static std::string gguf_data_to_str(enum gguf_type type, const void * data, int
     }
 }
 
-static std::string gguf_kv_to_str(struct gguf_context * ctx_gguf, int i) {
+static std::string gguf_kv_to_str(const struct gguf_context * ctx_gguf, int i) {
     const enum gguf_type type = gguf_get_kv_type(ctx_gguf, i);
 
     switch (type) {
@@ -694,38 +714,6 @@ static void ggml_graph_compute_helper(std::vector<uint8_t> & buf, ggml_cgraph *
 // llama helpers
 //
 
-inline void * llama_host_malloc(size_t n) {
-#ifdef GGML_USE_CUBLAS
-    if (ggml_cublas_loaded()) {
-        return ggml_cuda_host_malloc(n);
-    } else {
-        return malloc(n);
-    }
-#elif GGML_USE_METAL
-    return ggml_metal_host_malloc(n);
-#elif GGML_USE_CPU_HBM
-    return hbw_malloc(n);
-#else
-    return malloc(n);
-#endif
-}
-
-inline void llama_host_free(void * ptr) {
-#ifdef GGML_USE_CUBLAS
-    if (ggml_cublas_loaded()) {
-        return ggml_cuda_host_free(ptr);
-    } else {
-        return free(ptr);
-    }
-#elif GGML_USE_METAL
-    return ggml_metal_host_free(ptr);
-#elif GGML_USE_CPU_HBM
-    return hbw_free(ptr);
-#else
-    return free(ptr);
-#endif
-}
-
 #if defined(_WIN32)
 static std::string llama_format_win_err(DWORD err) {
     LPSTR buf;
@@ -740,40 +728,10 @@ static std::string llama_format_win_err(DWORD err) {
 }
 #endif
 
-struct llama_buffer {
-    void * data = NULL;
-    size_t size = 0;
-
-    // fallback to malloc / free
-    // useful in cases where CUDA can try to allocate PINNED memory
-    bool fallback = false;
-
-    void resize(size_t n) {
-        llama_host_free(data);
-
-        data = llama_host_malloc(n);
-        if (!data) {
-            fallback = true;
-            data = malloc(n);
-        } else {
-            fallback = false;
-        }
-
-        GGML_ASSERT(data);
-        size = n;
-    }
-
-    ~llama_buffer() {
-        if (data) {
-            if (fallback) { // NOLINT
-                free(data);
-            } else {
-                llama_host_free(data);
-            }
-        }
-
-        data = NULL;
-    }
+template <typename T>
+struct no_init {
+    T value;
+    no_init() { /* do nothing */ }
 };
 
 struct llama_file {
@@ -820,7 +778,7 @@ struct llama_file {
             throw std::runtime_error(format("read error: %s", strerror(errno)));
         }
         if (ret != 1) {
-            throw std::runtime_error(std::string("unexpectedly reached end of file"));
+            throw std::runtime_error("unexpectedly reached end of file");
         }
     }
 
@@ -861,6 +819,9 @@ struct llama_mmap {
 #ifdef _POSIX_MAPPED_FILES
     static constexpr bool SUPPORTED = true;
 
+    // list of mapped fragments (first_offset, last_offset)
+    std::vector<std::pair<size_t, size_t>> mapped_fragments;
+
     llama_mmap(struct llama_file * file, size_t prefetch = (size_t) -1 /* -1 = max value */, bool numa = false) {
         size = file->size;
         int fd = fileno(file->fp);
@@ -868,17 +829,22 @@ struct llama_mmap {
         // prefetch/readahead impairs performance on NUMA systems
         if (numa) { prefetch = 0; }
 #ifdef __linux__
+        // advise the kernel to read the file sequentially (increases readahead)
+        if (posix_fadvise(fd, 0, 0, POSIX_FADV_SEQUENTIAL)) {
+            LLAMA_LOG_WARN("warning: posix_fadvise(.., POSIX_FADV_SEQUENTIAL) failed: %s\n",
+                    strerror(errno));
+        }
         if (prefetch) { flags |= MAP_POPULATE; }
 #endif
         addr = mmap(NULL, file->size, PROT_READ, flags, fd, 0);
-        if (addr == MAP_FAILED) {
+        if (addr == MAP_FAILED) { // NOLINT
             throw std::runtime_error(format("mmap failed: %s", strerror(errno)));
         }
 
         if (prefetch > 0) {
-            // Advise the kernel to preload the mapped memory
+            // advise the kernel to preload the mapped memory
             if (posix_madvise(addr, std::min(file->size, prefetch), POSIX_MADV_WILLNEED)) {
-                fprintf(stderr, "warning: posix_madvise(.., POSIX_MADV_WILLNEED) failed: %s\n",
+                LLAMA_LOG_WARN("warning: posix_madvise(.., POSIX_MADV_WILLNEED) failed: %s\n",
                         strerror(errno));
             }
         }
@@ -886,41 +852,108 @@ struct llama_mmap {
             // advise the kernel not to use readahead
             // (because the next page might not belong on the same node)
             if (posix_madvise(addr, file->size, POSIX_MADV_RANDOM)) {
-                fprintf(stderr, "warning: posix_madvise(.., POSIX_MADV_RANDOM) failed: %s\n",
+                LLAMA_LOG_WARN("warning: posix_madvise(.., POSIX_MADV_RANDOM) failed: %s\n",
                         strerror(errno));
             }
         }
+
+        // initialize list of mapped_fragments
+        mapped_fragments.emplace_back(0, file->size);
+    }
+
+    static void align_range(size_t * first, size_t * last, size_t page_size) {
+        // align first to the next page
+        size_t offset_in_page = *first & (page_size - 1);
+        size_t offset_to_page = offset_in_page == 0 ? 0 : page_size - offset_in_page;
+        *first += offset_to_page;
+
+        // align last to the previous page
+        *last = *last & ~(page_size - 1);
+
+        if (*last <= *first) {
+            *last = *first;
+        }
+    }
+
+    // partially unmap the file in the range [first, last)
+    void unmap_fragment(size_t first, size_t last) {
+        // note: this function must not be called multiple times with overlapping ranges
+        // otherwise, there is a risk of invalidating addresses that have been repurposed for other mappings
+        int page_size = sysconf(_SC_PAGESIZE);
+        align_range(&first, &last, page_size);
+        size_t len = last - first;
+
+        if (len == 0) {
+            return;
+        }
+
+        GGML_ASSERT(first % page_size == 0);
+        GGML_ASSERT(last % page_size == 0);
+        GGML_ASSERT(last > first);
+
+        void * next_page_start = (uint8_t *) addr + first;
+
+        // unmap the range
+        if (munmap(next_page_start, len)) {
+            LLAMA_LOG_WARN("warning: munmap failed: %s\n", strerror(errno));
+        }
+
+        // update the list of mapped fragments to avoid unmapping the same range again in the destructor
+        std::vector<std::pair<size_t, size_t>> new_mapped_fragments;
+        for (const auto & frag : mapped_fragments) {
+            if (frag.first < first && frag.second > last) {
+                // the range is in the middle of the fragment, split it
+                new_mapped_fragments.emplace_back(frag.first, first);
+                new_mapped_fragments.emplace_back(last, frag.second);
+            } else if (frag.first < first && frag.second > first) {
+                // the range starts in the middle of the fragment
+                new_mapped_fragments.emplace_back(frag.first, first);
+            } else if (frag.first < last && frag.second > last) {
+                // the range ends in the middle of the fragment
+                new_mapped_fragments.emplace_back(last, frag.second);
+            } else if (frag.first >= first && frag.second <= last) {
+                // the range covers the entire fragment
+            } else {
+                // the range is outside the fragment
+                new_mapped_fragments.push_back(frag);
+            }
+        }
+        mapped_fragments = std::move(new_mapped_fragments);
     }
 
     ~llama_mmap() {
-        munmap(addr, size);
+        for (const auto & frag : mapped_fragments) {
+            if (munmap((char *) addr + frag.first, frag.second - frag.first)) {
+                LLAMA_LOG_WARN("warning: munmap failed: %s\n", strerror(errno));
+            }
+        }
     }
 #elif defined(_WIN32)
     static constexpr bool SUPPORTED = true;
 
-    llama_mmap(struct llama_file * file, bool prefetch = true, bool numa = false) {
-        (void) numa;
+    llama_mmap(struct llama_file * file, size_t prefetch = (size_t) -1, bool numa = false) {
+        GGML_UNUSED(numa);
 
         size = file->size;
 
         HANDLE hFile = (HANDLE) _get_osfhandle(_fileno(file->fp));
 
         HANDLE hMapping = CreateFileMappingA(hFile, NULL, PAGE_READONLY, 0, 0, NULL);
-        DWORD error = GetLastError();
 
         if (hMapping == NULL) {
+            DWORD error = GetLastError();
             throw std::runtime_error(format("CreateFileMappingA failed: %s", llama_format_win_err(error).c_str()));
         }
 
         addr = MapViewOfFile(hMapping, FILE_MAP_READ, 0, 0, 0);
-        error = GetLastError();
+        DWORD error = GetLastError();
         CloseHandle(hMapping);
 
         if (addr == NULL) {
             throw std::runtime_error(format("MapViewOfFile failed: %s", llama_format_win_err(error).c_str()));
         }
 
-        if (prefetch) {
+        if (prefetch > 0) {
             // PrefetchVirtualMemory is only present on Windows 8 and above, so we dynamically load it
             BOOL (WINAPI *pPrefetchVirtualMemory) (HANDLE, ULONG_PTR, PWIN32_MEMORY_RANGE_ENTRY, ULONG);
             HMODULE hKernel32 = GetModuleHandleW(L"kernel32.dll");
@@ -932,30 +965,43 @@ struct llama_mmap {
                 // advise the kernel to preload the mapped memory
                 WIN32_MEMORY_RANGE_ENTRY range;
                 range.VirtualAddress = addr;
-                range.NumberOfBytes = (SIZE_T)size;
+                range.NumberOfBytes = (SIZE_T) std::min(size, prefetch);
                 if (!pPrefetchVirtualMemory(GetCurrentProcess(), 1, &range, 0)) {
-                    fprintf(stderr, "warning: PrefetchVirtualMemory failed: %s\n",
+                    LLAMA_LOG_WARN("warning: PrefetchVirtualMemory failed: %s\n",
                             llama_format_win_err(GetLastError()).c_str());
                 }
             }
         }
     }
 
+    void unmap_fragment(size_t first, size_t last) {
+        // not supported
+        GGML_UNUSED(first);
+        GGML_UNUSED(last);
+    }
+
     ~llama_mmap() {
         if (!UnmapViewOfFile(addr)) {
-            fprintf(stderr, "warning: UnmapViewOfFile failed: %s\n",
+            LLAMA_LOG_WARN("warning: UnmapViewOfFile failed: %s\n",
                     llama_format_win_err(GetLastError()).c_str());
         }
     }
 #else
     static constexpr bool SUPPORTED = false;
 
-    llama_mmap(struct llama_file * file, bool prefetch = true, bool numa = false) {
-        (void) file;
-        (void) prefetch;
-        (void) numa;
+    llama_mmap(struct llama_file * file, size_t prefetch = -1, bool numa = false) {
+        GGML_UNUSED(file);
+        GGML_UNUSED(prefetch);
+        GGML_UNUSED(numa);
 
-        throw std::runtime_error(std::string("mmap not supported"));
+        throw std::runtime_error("mmap not supported");
+    }
+
+    void unmap_fragment(size_t first, size_t last) {
+        GGML_UNUSED(first);
+        GGML_UNUSED(last);
+
+        throw std::runtime_error("mmap not supported");
     }
 #endif
 };
@@ -1130,6 +1176,26 @@ static std::string llama_token_to_piece(const struct llama_context * ctx, llama_
     return std::string(result.data(), result.size());
 }
 
+static ggml_backend_buffer_type_t llama_default_buffer_type(int n_gpu_layers) {
+#ifdef GGML_USE_METAL
+    if (n_gpu_layers > 0) {
+        return ggml_backend_metal_buffer_type();
+    }
+#elif defined(GGML_USE_CUBLAS) && defined(LLAMA_GGML_BACKEND_CUDA_TEST)
+    if (n_gpu_layers > 0) {
+        return ggml_backend_cuda_buffer_type(0);
+    }
+#elif defined(GGML_USE_CUBLAS)
+    return ggml_backend_cuda_host_buffer_type();
+#elif defined(GGML_USE_CPU_HBM)
+    return ggml_backend_cpu_hbm_buffer_type();
+#endif
+
+    return ggml_backend_cpu_buffer_type();
+
+    GGML_UNUSED(n_gpu_layers);
+}
+
 //
 // globals
 //
@@ -1178,6 +1244,8 @@ struct llama_hparams {
     uint32_t n_layer;
     uint32_t n_rot;
     uint32_t n_ff;
+    uint32_t n_expert = 0;
+    uint32_t n_expert_used = 0;
 
     float f_norm_eps;
     float f_norm_rms_eps;
@@ -1192,15 +1260,18 @@ struct llama_hparams {
     float f_max_alibi_bias;
 
     bool operator!=(const llama_hparams & other) const {
-        if (this->vocab_only  != other.vocab_only)  return true;
-        if (this->n_vocab     != other.n_vocab)     return true;
-        if (this->n_ctx_train != other.n_ctx_train) return true;
-        if (this->n_embd      != other.n_embd)      return true;
-        if (this->n_head      != other.n_head)      return true;
-        if (this->n_head_kv   != other.n_head_kv)   return true;
-        if (this->n_layer     != other.n_layer)     return true;
-        if (this->n_rot       != other.n_rot)       return true;
-        if (this->n_ff        != other.n_ff)        return true;
+        if (this->vocab_only    != other.vocab_only)    return true;
+        if (this->n_vocab       != other.n_vocab)       return true;
+        if (this->n_ctx_train   != other.n_ctx_train)   return true;
+        if (this->n_embd        != other.n_embd)        return true;
+        if (this->n_head        != other.n_head)        return true;
+        if (this->n_head_kv     != other.n_head_kv)     return true;
+        if (this->n_layer       != other.n_layer)       return true;
+        if (this->n_rot         != other.n_rot)         return true;
+        if (this->n_ff          != other.n_ff)          return true;
+        if (this->n_expert      != other.n_expert)      return true;
+        if (this->n_expert_used != other.n_expert_used) return true;
+
         if (this->rope_finetuned  != other.rope_finetuned)  return true;
         if (this->n_yarn_orig_ctx != other.n_yarn_orig_ctx) return true;
 
@@ -1245,6 +1316,7 @@ struct llama_cparams {
     float yarn_beta_slow;
 
     bool mul_mat_q;
+    bool offload_kqv;
 };
 
 struct llama_layer {
@@ -1281,6 +1353,12 @@ struct llama_layer {
     struct ggml_tensor * ffn_down; // w2
     struct ggml_tensor * ffn_up;   // w3
 
+    // ff MoE
+    struct ggml_tensor * ffn_gate_inp;
+    struct ggml_tensor * ffn_gate_exp[LLAMA_MAX_EXPERTS];
+    struct ggml_tensor * ffn_down_exp[LLAMA_MAX_EXPERTS];
+    struct ggml_tensor * ffn_up_exp  [LLAMA_MAX_EXPERTS];
+
     // ff bias
     struct ggml_tensor * ffn_down_b; // b2
     struct ggml_tensor * ffn_up_b;   // b3
@@ -1313,24 +1391,27 @@ struct llama_kv_cache {
 
     std::vector<llama_kv_cell> cells;
 
-    struct ggml_tensor * k = NULL;
-    struct ggml_tensor * v = NULL;
+    std::vector<struct ggml_tensor *> k_l; // per layer
+    std::vector<struct ggml_tensor *> v_l;
 
     struct ggml_context * ctx = NULL;
 
-    llama_buffer buf;
+    ggml_backend_buffer_t buf = NULL;
 
     ~llama_kv_cache() {
+#if defined(GGML_USE_CUBLAS) && !defined(LLAMA_GGML_BACKEND_CUDA_TEST)
+        if (ggml_cublas_loaded()) {
+            for (size_t i = 0; i < k_l.size(); ++i) {
+                ggml_cuda_free_data(k_l[i]);
+                ggml_cuda_free_data(v_l[i]);
+            }
+        }
+#endif
         if (ctx) {
             ggml_free(ctx);
         }
 
-#ifdef GGML_USE_CUBLAS
-        if (ggml_cublas_loaded()) {
-            ggml_cuda_free_data(k);
-            ggml_cuda_free_data(v);
-        }
-#endif
+        ggml_backend_buffer_free(buf);
     }
 };
 
@@ -1370,11 +1451,11 @@ struct llama_vocab {
     id special_suffix_id = 32008;
     id special_eot_id    = 32010;
 
-    int find_bpe_rank(std::string token_left, std::string token_right) const {
-        GGML_ASSERT(token_left.find(" ") == std::string::npos);
-        GGML_ASSERT(token_left.find("\n") == std::string::npos);
-        GGML_ASSERT(token_right.find(" ") == std::string::npos);
-        GGML_ASSERT(token_right.find("\n") == std::string::npos);
+    int find_bpe_rank(const std::string & token_left, const std::string & token_right) const {
+        GGML_ASSERT(token_left.find(' ') == std::string::npos);
+        GGML_ASSERT(token_left.find('\n') == std::string::npos);
+        GGML_ASSERT(token_right.find(' ') == std::string::npos);
+        GGML_ASSERT(token_right.find('\n') == std::string::npos);
 
         auto it = bpe_ranks.find(std::make_pair(token_left, token_right));
         if (it == bpe_ranks.end()) {
@@ -1403,6 +1484,7 @@ 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;
 
@@ -1415,7 +1497,7 @@ struct llama_model {
     struct ggml_context * ctx = NULL;
 
     // the model memory buffer
-    llama_buffer buf;
+    ggml_backend_buffer_t buf = NULL;
 
     // model memory mapped file
     std::unique_ptr<llama_mmap> mapping;
@@ -1431,11 +1513,7 @@ struct llama_model {
     int64_t t_start_us = 0;
 
     ~llama_model() {
-        if (ctx) {
-            ggml_free(ctx);
-        }
-
-#ifdef GGML_USE_CUBLAS
+#if defined(GGML_USE_CUBLAS) && !defined(LLAMA_GGML_BACKEND_CUDA_TEST)
         if (ggml_cublas_loaded()) {
             for (size_t i = 0; i < tensors_by_name.size(); ++i) {
                 ggml_cuda_free_data(tensors_by_name[i].second);
@@ -1449,24 +1527,26 @@ struct llama_model {
             ggml_cl_free_data(tensors_by_name[i].second);
         }
 #endif
+        if (ctx) {
+            ggml_free(ctx);
+        }
+
+        ggml_backend_buffer_free(buf);
     }
 };
 
 struct llama_context {
     llama_context(const llama_model & model) : model(model), t_start_us(model.t_start_us), t_load_us(model.t_load_us) {}
     ~llama_context() {
-#ifdef GGML_USE_METAL
-        if (ctx_metal) {
-            ggml_metal_free(ctx_metal);
-        }
-#endif
-        if (alloc) {
-            ggml_allocr_free(alloc);
-        }
+        ggml_allocr_free(alloc);
+        ggml_backend_buffer_free(buf_alloc);
+        ggml_backend_free(backend);
     }
 
     llama_cparams cparams;
 
+    ggml_backend_t backend = nullptr;
+
     const llama_model & model;
 
     // key + value cache for the self attention
@@ -1488,23 +1568,22 @@ 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])
     std::vector<float> embedding;
 
-    // reusable buffer for `struct ggml_graph_plan.work_data`
-    std::vector<uint8_t> work_buffer;
-
     // memory buffers used to evaluate the model
-    llama_buffer buf_compute;
-
-    llama_buffer buf_alloc;
+    std::vector<uint8_t> buf_compute_meta;
+    ggml_backend_buffer_t buf_alloc = NULL;
     ggml_allocr * alloc = NULL;
 
-#ifdef GGML_USE_METAL
-    ggml_metal_context * ctx_metal = NULL;
-#endif
+    // temporary buffer for copying data to/from the backend
+    std::vector<no_init<uint8_t>> buf_copy;
 
 #ifdef GGML_USE_MPI
     ggml_mpi_context * ctx_mpi = NULL;
@@ -1518,15 +1597,14 @@ struct llama_context {
 static bool llama_kv_cache_init(
         const struct llama_hparams & hparams,
              struct llama_kv_cache & cache,
-                         ggml_type   wtype,
+                         ggml_type   ktype,
+                         ggml_type   vtype,
                           uint32_t   n_ctx,
-                               int   n_gpu_layers) {
+                               int   n_gpu_layers,
+                              bool   offload) {
     const uint32_t n_embd  = hparams.n_embd_gqa();
     const uint32_t n_layer = hparams.n_layer;
 
-    const int64_t n_mem      = n_layer*n_ctx;
-    const int64_t n_elements = n_embd*n_mem;
-
     cache.has_shift = false;
 
     cache.head = 0;
@@ -1536,47 +1614,61 @@ static bool llama_kv_cache_init(
     cache.cells.clear();
     cache.cells.resize(n_ctx);
 
-    cache.buf.resize(2u*n_elements*ggml_type_size(wtype) + 2u*ggml_tensor_overhead());
-    memset(cache.buf.data, 0, cache.buf.size);
-
     struct ggml_init_params params;
-    params.mem_size   = cache.buf.size;
-    params.mem_buffer = cache.buf.data;
-    params.no_alloc   = false;
+    params.mem_size   = 2u*n_layer*ggml_tensor_overhead();
+    params.mem_buffer = NULL;
+    params.no_alloc   = true;
 
     cache.ctx = ggml_init(params);
 
+    size_t vram_kv_cache = 0;
+
     if (!cache.ctx) {
         LLAMA_LOG_ERROR("%s: failed to allocate memory for kv cache\n", __func__);
         return false;
     }
 
-    cache.k = ggml_new_tensor_1d(cache.ctx, wtype, n_elements);
-    cache.v = ggml_new_tensor_1d(cache.ctx, wtype, n_elements);
-    ggml_set_name(cache.k, "cache_k");
-    ggml_set_name(cache.v, "cache_v");
+    cache.k_l.reserve(n_layer);
+    cache.v_l.reserve(n_layer);
 
-    (void) n_gpu_layers;
+    const int i_gpu_start = (int) n_layer - n_gpu_layers;
 
-#ifdef GGML_USE_CUBLAS
-    if (ggml_cublas_loaded()) {
-        size_t vram_kv_cache = 0;
-
-        if (n_gpu_layers > (int)n_layer + 1) {
-            ggml_cuda_assign_buffers_no_scratch(cache.v);
-            LLAMA_LOG_INFO("%s: offloading v cache to GPU\n", __func__);
-            vram_kv_cache += ggml_nbytes(cache.v);
-        }
-        if (n_gpu_layers > (int)n_layer + 2) {
-            ggml_cuda_assign_buffers_no_scratch(cache.k);
-            LLAMA_LOG_INFO("%s: offloading k cache to GPU\n", __func__);
-            vram_kv_cache += ggml_nbytes(cache.k);
-        }
-        if (vram_kv_cache > 0) {
-            LLAMA_LOG_INFO("%s: VRAM kv self = %.2f MiB\n", __func__, vram_kv_cache / 1024.0 / 1024.0);
+    for (int i = 0; i < (int) n_layer; i++) {
+        ggml_tensor * k = ggml_new_tensor_1d(cache.ctx, ktype, n_embd*n_ctx);
+        ggml_tensor * v = ggml_new_tensor_1d(cache.ctx, vtype, n_embd*n_ctx);
+        ggml_format_name(k, "cache_k_l%d", i);
+        ggml_format_name(v, "cache_v_l%d", i);
+        cache.k_l.push_back(k);
+        cache.v_l.push_back(v);
+#if defined(GGML_USE_CUBLAS) && !defined(LLAMA_GGML_BACKEND_CUDA_TEST)
+        if (i >= i_gpu_start) {
+            if (offload) {
+                ggml_cuda_assign_buffers_no_scratch(k);
+                ggml_cuda_assign_buffers_no_scratch(v);
+                vram_kv_cache += ggml_nbytes(k);
+                vram_kv_cache += ggml_nbytes(v);
+                // HACK: mark tensor as allocated
+                k->data = v->data = (void *)(uintptr_t)1;
+            }
         }
+#endif // GGML_USE_CUBLAS
     }
-#endif
+
+    // allocate tensors
+    cache.buf = ggml_backend_alloc_ctx_tensors_from_buft(cache.ctx, llama_default_buffer_type(n_gpu_layers));
+
+    // buf may be NULL with full offload
+    if (cache.buf) {
+        // initialize the buffer to avoid NaNs in the padding
+        ggml_backend_buffer_clear(cache.buf, 0);
+    }
+
+    if (vram_kv_cache > 0) {
+        LLAMA_LOG_INFO("%s: VRAM kv self = %.2f MB\n", __func__, vram_kv_cache / 1024.0 / 1024.0);
+    }
+
+    GGML_UNUSED(i_gpu_start);
+    GGML_UNUSED(offload);
 
     return true;
 }
@@ -1797,6 +1889,169 @@ static std::string llama_format_tensor_shape(const struct ggml_tensor * t) {
     return buf;
 }
 
+namespace GGUFMeta {
+    template <typename T, gguf_type gt_, T (*gfun)(const gguf_context *, const int)>
+    struct GKV_Base_Type {
+        static constexpr gguf_type gt = gt_;
+
+        static T getter(const gguf_context * ctx, const int kid) {
+            return gfun(ctx, kid);
+        }
+    };
+
+    template<typename T> struct GKV_Base;
+
+    template<> struct GKV_Base<bool        >: GKV_Base_Type<bool,         GGUF_TYPE_BOOL,    gguf_get_val_bool> {};
+    template<> struct GKV_Base<uint8_t     >: GKV_Base_Type<uint8_t,      GGUF_TYPE_UINT8,   gguf_get_val_u8  > {};
+    template<> struct GKV_Base<uint16_t    >: GKV_Base_Type<uint16_t,     GGUF_TYPE_UINT16,  gguf_get_val_u16 > {};
+    template<> struct GKV_Base<uint32_t    >: GKV_Base_Type<uint32_t,     GGUF_TYPE_UINT32,  gguf_get_val_u32 > {};
+    template<> struct GKV_Base<uint64_t    >: GKV_Base_Type<uint64_t,     GGUF_TYPE_UINT64,  gguf_get_val_u64 > {};
+    template<> struct GKV_Base<int8_t      >: GKV_Base_Type<int8_t,       GGUF_TYPE_INT8,    gguf_get_val_i8  > {};
+    template<> struct GKV_Base<int16_t     >: GKV_Base_Type<int16_t,      GGUF_TYPE_INT16,   gguf_get_val_i16 > {};
+    template<> struct GKV_Base<int32_t     >: GKV_Base_Type<int32_t,      GGUF_TYPE_INT32,   gguf_get_val_i32 > {};
+    template<> struct GKV_Base<int64_t     >: GKV_Base_Type<int64_t,      GGUF_TYPE_INT64,   gguf_get_val_i64 > {};
+    template<> struct GKV_Base<float       >: GKV_Base_Type<float,        GGUF_TYPE_FLOAT32, gguf_get_val_f32 > {};
+    template<> struct GKV_Base<double      >: GKV_Base_Type<double,       GGUF_TYPE_FLOAT64, gguf_get_val_f64 > {};
+    template<> struct GKV_Base<const char *>: GKV_Base_Type<const char *, GGUF_TYPE_STRING,  gguf_get_val_str > {};
+
+    template<> struct GKV_Base<std::string> {
+        static constexpr gguf_type gt = GGUF_TYPE_STRING;
+
+        static std::string getter(const gguf_context * ctx, const int kid) {
+            return gguf_get_val_str(ctx, kid);
+        }
+    };
+
+    struct ArrayInfo{
+        const gguf_type gt;
+        const size_t length;
+        const void * data;
+    };
+
+    template<> struct GKV_Base<ArrayInfo> {
+        public:
+        static constexpr gguf_type gt = GGUF_TYPE_ARRAY;
+        static ArrayInfo getter(const gguf_context *ctx, const int k) {
+            return ArrayInfo {
+                gguf_get_arr_type(ctx, k),
+                size_t(gguf_get_arr_n(ctx, k)),
+                gguf_get_arr_data(ctx, k),
+            };
+        }
+    };
+
+    template<typename T>
+    class GKV: public GKV_Base<T> {
+        GKV() = delete;
+
+        public:
+        static T get_kv(const gguf_context * ctx, const int k) {
+            const enum gguf_type kt = gguf_get_kv_type(ctx, k);
+
+            if (kt != GKV::gt) {
+                throw std::runtime_error(format("key %s has wrong type %s but expected type %s",
+                    gguf_get_key(ctx, k), gguf_type_name(kt), gguf_type_name(GKV::gt)));
+            }
+            return GKV::getter(ctx, k);
+        }
+
+        static const char * override_type_to_str(const llama_model_kv_override_type ty) {
+            switch (ty) {
+                case LLAMA_KV_OVERRIDE_BOOL:  return "bool";
+                case LLAMA_KV_OVERRIDE_INT:   return "int";
+                case LLAMA_KV_OVERRIDE_FLOAT: return "float";
+            }
+            return "unknown";
+        }
+
+        static bool validate_override(const llama_model_kv_override_type expected_type, const struct llama_model_kv_override *override) {
+            if (!override) { return false; }
+            if (override->tag == expected_type) {
+                LLAMA_LOG_INFO("%s: Using metadata override (%5s) '%s' = ",
+                    __func__, override_type_to_str(override->tag), override->key);
+                switch (override->tag) {
+                    case LLAMA_KV_OVERRIDE_BOOL:  {
+                        printf("%s\n", override->bool_value ? "true" : "false");
+                    } break;
+                    case LLAMA_KV_OVERRIDE_INT:   {
+                        printf("%" PRId64 "\n", override->int_value);
+                    } break;
+                    case LLAMA_KV_OVERRIDE_FLOAT: {
+                        printf("%.6f\n", override->float_value);
+                    } break;
+                    default:
+                        // Shouldn't be possible to end up here, but just in case...
+                        throw std::runtime_error(
+                            format("Unsupported attempt to override %s type for metadata key %s\n",
+                                override_type_to_str(override->tag), override->key));
+                }
+                return true;
+            }
+            LLAMA_LOG_WARN("%s: Warning: Bad metadata override type for key '%s', expected %s but got %s\n",
+                __func__, override->key, override_type_to_str(expected_type), override_type_to_str(override->tag));
+            return false;
+        }
+
+        template<typename OT>
+        static typename std::enable_if<std::is_same<OT, bool>::value, bool>::type
+        try_override(OT & target, const struct llama_model_kv_override *override) {
+            if (validate_override(LLAMA_KV_OVERRIDE_BOOL, override)) {
+                target = override->bool_value;
+                return true;
+            }
+            return false;
+        }
+
+        template<typename OT>
+        static typename std::enable_if<!std::is_same<OT, bool>::value && std::is_integral<OT>::value, bool>::type
+        try_override(OT & target, const struct llama_model_kv_override *override) {
+            if (validate_override(LLAMA_KV_OVERRIDE_INT, override)) {
+                target = override->int_value;
+                return true;
+            }
+            return false;
+        }
+
+        template<typename OT>
+        static typename std::enable_if<std::is_floating_point<OT>::value, bool>::type
+        try_override(T & target, const struct llama_model_kv_override *override) {
+            if (validate_override(LLAMA_KV_OVERRIDE_FLOAT, override)) {
+                target = override->float_value;
+                return true;
+            }
+            return false;
+        }
+
+        template<typename OT>
+        static typename std::enable_if<std::is_same<OT, std::string>::value, bool>::type
+        try_override(T & target, const struct llama_model_kv_override *override) {
+            (void)target;
+            (void)override;
+            if (!override) { return false; }
+            // Currently, we should never end up here so it would be a bug if we do.
+            throw std::runtime_error(format("Unsupported attempt to override string type for metadata key %s\n",
+                override ? override->key : "NULL"));
+        }
+
+        static bool set(const gguf_context * ctx, const int k, T & target, const struct llama_model_kv_override *override = nullptr) {
+            if (try_override<T>(target, override)) {
+                return true;
+            }
+            if (k < 0) { return false; }
+            target = get_kv(ctx, k);
+            return true;
+        }
+
+        static bool set(const gguf_context * ctx, const char * key, T & target, const struct llama_model_kv_override *override = nullptr) {
+            return set(ctx, gguf_find_key(ctx, key), target, override);
+        }
+
+        static bool set(const gguf_context * ctx, const std::string & key, T & target, const struct llama_model_kv_override *override = nullptr) {
+            return set(ctx, key.c_str(), target, override);
+        }
+    };
+}
+
 struct llama_model_loader {
     int n_kv      = 0;
     int n_tensors = 0;
@@ -1812,21 +2067,34 @@ struct llama_model_loader {
     llama_fver  fver;
 
     std::unique_ptr<llama_mmap> mapping;
+    std::unordered_map<std::string, struct llama_model_kv_override> kv_overrides;
 
     struct gguf_context * ctx_gguf = NULL;
     struct ggml_context * ctx_meta = NULL;
 
-    llama_model_loader(const std::string & fname, bool use_mmap) : file(fname.c_str(), "rb") {
+    std::string arch_name;
+    LLM_KV      llm_kv    = LLM_KV(LLM_ARCH_UNKNOWN);
+
+    llama_model_loader(const std::string & fname, bool use_mmap, const struct llama_model_kv_override * param_overrides_p) : file(fname.c_str(), "rb") {
         struct gguf_init_params params = {
             /*.no_alloc = */ true,
             /*.ctx      = */ &ctx_meta,
         };
 
+        if (param_overrides_p != nullptr) {
+            for (const struct llama_model_kv_override *p = param_overrides_p; p->key[0] != 0; p++) {
+                kv_overrides.insert({std::string(p->key), *p});
+            }
+        }
+
         ctx_gguf = gguf_init_from_file(fname.c_str(), params);
         if (!ctx_gguf) {
             throw std::runtime_error(format("%s: failed to load model from %s\n", __func__, fname.c_str()));
         }
 
+        get_key(llm_kv(LLM_KV_GENERAL_ARCHITECTURE), arch_name, false);
+        llm_kv = LLM_KV(llm_arch_from_string(arch_name));
+
         n_kv      = gguf_get_n_kv(ctx_gguf);
         n_tensors = gguf_get_n_tensors(ctx_gguf);
 
@@ -1851,17 +2119,16 @@ struct llama_model_loader {
             enum ggml_type type_max = GGML_TYPE_F32;
 
             for (int i = 0; i < n_tensors; i++) {
-                const char * name = gguf_get_tensor_name(ctx_gguf, i);
-                struct ggml_tensor * meta = ggml_get_tensor(ctx_meta, name);
+                enum ggml_type type = gguf_get_tensor_type(ctx_gguf, i);
 
-                n_type[meta->type]++;
+                n_type[type]++;
 
-                if (n_type_max < n_type[meta->type]) {
-                    n_type_max = n_type[meta->type];
-                    type_max   = meta->type;
+                if (n_type_max < n_type[type]) {
+                    n_type_max = n_type[type];
+                    type_max   = 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) {
@@ -1894,6 +2161,7 @@ struct llama_model_loader {
                 }
             }
 
+            LLAMA_LOG_INFO("%s: Dumping metadata keys/values. Note: KV overrides do not apply in this output.\n", __func__);
             for (int i = 0; i < n_kv; i++) {
                 const char * name           = gguf_get_key(ctx_gguf, i);
                 const enum gguf_type type   = gguf_get_kv_type(ctx_gguf, i);
@@ -1939,53 +2207,78 @@ struct llama_model_loader {
         }
     }
 
+    template<typename T>
+    typename std::enable_if<std::is_integral<T>::value, bool>::type
+    get_arr_n(const std::string & key, T & result, const bool required = true) {
+        const int kid = gguf_find_key(ctx_gguf, key.c_str());
+
+        if (kid < 0) {
+            if (required) {
+                throw std::runtime_error(format("key not found in model: %s", key.c_str()));
+            }
+            return false;
+        }
+
+        struct GGUFMeta::ArrayInfo arr_info =
+            GGUFMeta::GKV<GGUFMeta::ArrayInfo>::get_kv(ctx_gguf, kid);
+
+
+        result = arr_info.length;
+        return true;
+    }
+
+    template<typename T>
+    typename std::enable_if<std::is_integral<T>::value, bool>::type
+    get_arr_n(const enum llm_kv kid, T & result, const bool required = true) {
+        return get_arr_n(llm_kv(kid), result, required);
+    }
+
+    template<typename T>
+    bool get_key(const std::string & key, T & result, const bool required = true) {
+        auto it = kv_overrides.find(key);
+
+        const struct llama_model_kv_override * override =
+            it != kv_overrides.end() ? &it->second : nullptr;
+
+        const bool found = GGUFMeta::GKV<T>::set(ctx_gguf, key, result, override);
+
+        if (required && !found) {
+            throw std::runtime_error(format("key not found in model: %s", key.c_str()));
+        }
+
+        return found;
+    }
+
+    template<typename T>
+    bool get_key(const enum llm_kv kid, T & result, const bool required = true) {
+        return get_key(llm_kv(kid), result, required);
+    }
+
     std::string get_arch_name() const {
-        const auto kv = LLM_KV(LLM_ARCH_UNKNOWN);
-
-        std::string arch_name;
-        GGUF_GET_KEY(ctx_gguf, arch_name, gguf_get_val_str, GGUF_TYPE_STRING, false, kv(LLM_KV_GENERAL_ARCHITECTURE));
-
         return arch_name;
     }
 
     enum llm_arch get_arch() const {
-        const std::string arch_name = get_arch_name();
-
-        return llm_arch_from_string(arch_name);
+        return llm_kv.arch;
     }
 
     const char * get_tensor_name(int i) const {
         return gguf_get_tensor_name(ctx_gguf, i);
     }
 
-    struct ggml_tensor * get_tensor_meta(int i) const {
-        return ggml_get_tensor(ctx_meta, get_tensor_name(i));
+    struct ggml_tensor * get_tensor_meta(const char * name) const {
+        return ggml_get_tensor(ctx_meta, name);
     }
 
-    void calc_sizes(size_t & ctx_size_p, size_t & mmapped_size_p) const {
-        ctx_size_p     = 0;
-        mmapped_size_p = 0;
-
-        for (int i = 0; i < n_tensors; i++) {
-            struct ggml_tensor * meta = get_tensor_meta(i);
-            ctx_size_p += sizeof(struct ggml_tensor) + GGML_OBJECT_SIZE;
-            (use_mmap ? mmapped_size_p : ctx_size_p) += ggml_nbytes_pad(meta);
-        }
+    struct ggml_tensor * get_tensor_meta(int i) const {
+        return get_tensor_meta(get_tensor_name(i));
     }
 
     struct ggml_tensor * create_tensor_for(struct ggml_context * ctx, struct ggml_tensor * meta, ggml_backend_type backend) {
-        if (backend != GGML_BACKEND_CPU) {
-            ggml_set_no_alloc(ctx, true);
-        }
-
         struct ggml_tensor * tensor = ggml_dup_tensor(ctx, meta);
         tensor->backend = backend; // TODO: ggml_set_backend
         ggml_set_name(tensor, ggml_get_name(meta));
 
-        if (backend != GGML_BACKEND_CPU) {
-            ggml_set_no_alloc(ctx, use_mmap);
-        }
-
         n_created++;
 
         return tensor;
@@ -2043,91 +2336,144 @@ struct llama_model_loader {
         return gguf_get_data_offset(ctx_gguf) + gguf_get_tensor_offset(ctx_gguf, idx);
     }
 
+    void init_mapping(bool prefetch = true) {
+        /*
+        // prefetch only CPU tensors
+        if (use_mmap) {
+            size_t size_pref = 0; // prefetch
+
+            for (int i = 0; i < gguf_get_n_tensors(ctx_gguf); i++) {
+                struct ggml_tensor * cur = ggml_get_tensor(ctx, gguf_get_tensor_name(ctx_gguf, i));
+                if (cur->backend == GGML_BACKEND_CPU) {
+                    size_t tensor_end = gguf_get_tensor_offset(ctx_gguf, i) + ggml_nbytes(cur);
+                    size_pref = std::max(size_pref, tensor_end);
+                }
+            }
+            mapping.reset(new llama_mmap(&file, gguf_get_data_offset(ctx_gguf) + size_pref, ggml_is_numa()));
+        }
+        */
+        // prefetch the whole file - all the data is needed anyway
+        if (use_mmap) {
+            mapping.reset(new llama_mmap(&file, prefetch ? -1 : 0, ggml_is_numa()));
+        }
+    }
+
+    // for backwards compatibility, does not support ggml-backend
     void load_data_for(struct ggml_tensor * cur) const {
         const size_t offs = file_offset(ggml_get_name(cur));
 
-        if (use_mmap) {
-            cur->data = (uint8_t *) mapping->addr + offs;
+        if (use_mmap && mapping) {
+            GGML_ASSERT(cur->data == nullptr);
+            cur->data = (uint8_t *)mapping->addr + offs;
         } else {
+            GGML_ASSERT(cur->data != nullptr);
             file.seek(offs, SEEK_SET);
             file.read_raw(cur->data, ggml_nbytes(cur));
         }
     }
 
-    void load_all_data(struct ggml_context * ctx, llama_progress_callback progress_callback, void * progress_callback_user_data, llama_mlock * lmlock) {
+    // Returns false if cancelled by progress_callback
+    bool load_all_data(struct ggml_context * ctx, llama_progress_callback progress_callback, void * progress_callback_user_data, ggml_backend_buffer_t buf_mmap, llama_mlock * lmlock) const {
         size_t size_data = 0;
-        size_t size_lock = 0;
-        size_t size_pref = 0; // prefetch
 
         for (int i = 0; i < gguf_get_n_tensors(ctx_gguf); i++) {
             struct ggml_tensor * cur = ggml_get_tensor(ctx, gguf_get_tensor_name(ctx_gguf, i));
             size_data += ggml_nbytes(cur);
-            if (cur->backend == GGML_BACKEND_CPU) {
-                size_pref += ggml_nbytes(cur);
-            }
         }
 
-        if (use_mmap) {
-            mapping.reset(new llama_mmap(&file, size_pref, ggml_is_numa()));
+        if (use_mmap && buf_mmap) {
             if (lmlock) {
                 lmlock->init(mapping->addr);
             }
         }
 
-        size_t done_size = 0;
+#if (defined(GGML_USE_CUBLAS) && !defined(LLAMA_GGML_BACKEND_CUDA_TEST)) || defined(GGML_USE_CLBLAST)
+        const bool legacy_offload = true;
+#else
+        const bool legacy_offload = false;
+#endif
+
+        std::vector<no_init<uint8_t>> read_buf;
+
+        size_t size_done = 0;
+
+        size_t mmap_first = -1;
+        size_t mmap_last  = 0;
+
         for (int i = 0; i < gguf_get_n_tensors(ctx_gguf); i++) {
             struct ggml_tensor * cur = ggml_get_tensor(ctx, gguf_get_tensor_name(ctx_gguf, i));
             GGML_ASSERT(cur); // unused tensors should have been caught by load_data already
 
             if (progress_callback) {
-                progress_callback((float) done_size / size_data, progress_callback_user_data);
+                if (!progress_callback((float) size_done / size_data, progress_callback_user_data)) {
+                    return false;
+                }
             }
 
-            // allocate temp buffer if not using mmap
-            if (!use_mmap && cur->data == NULL) {
-                GGML_ASSERT(cur->backend != GGML_BACKEND_CPU);
-                #ifdef GGML_USE_CPU_HBM
-                cur->data = (uint8_t*)hbw_malloc(ggml_nbytes(cur));
-                #else
-                cur->data = (uint8_t*)malloc(ggml_nbytes(cur));
-                #endif
-            }
+            const size_t offs = file_offset(ggml_get_name(cur));
 
-            load_data_for(cur);
-
-            switch (cur->backend) {
-                case GGML_BACKEND_CPU:
-                    if (use_mmap && lmlock) {
-                        size_lock += ggml_nbytes(cur);
-                        lmlock->grow_to(size_lock);
+            if (!legacy_offload || cur->backend == GGML_BACKEND_CPU) {
+                if (use_mmap && mapping) {
+                    if (buf_mmap) {
+                        ggml_backend_tensor_alloc(buf_mmap, cur, (uint8_t *) mapping->addr + offs);
+                        if (lmlock) {
+                            lmlock->grow_to(offs + ggml_nbytes(cur));
+                        }
+                        mmap_first = std::min(mmap_first, offs);
+                        mmap_last  = std::max(mmap_last,  offs + ggml_nbytes(cur));
+                    } else {
+                        ggml_backend_tensor_set(cur, (uint8_t *) mapping->addr + offs, 0, ggml_nbytes(cur));
                     }
-                    break;
-#ifdef GGML_USE_CUBLAS
-                case GGML_BACKEND_GPU:
-                case GGML_BACKEND_GPU_SPLIT:
-                    // old code:
-                    //ggml_cuda_transform_tensor(lt.data, lt.ggml_tensor);
-
-                    // TODO: test if this works !!
-                    ggml_cuda_transform_tensor(cur->data, cur);
-                    if (!use_mmap) {
-                        free(cur->data);
+                } else {
+                    if (ggml_backend_buffer_is_host(cur->buffer)) {
+                        file.seek(offs, SEEK_SET);
+                        file.read_raw(cur->data, ggml_nbytes(cur));
+                    } else {
+                        read_buf.resize(ggml_nbytes(cur));
+                        file.seek(offs, SEEK_SET);
+                        file.read_raw(read_buf.data(), ggml_nbytes(cur));
+                        ggml_backend_tensor_set(cur, read_buf.data(), 0, ggml_nbytes(cur));
                     }
-                    break;
+                }
+            } else {
+                // HACK: mark tensor as allocated
+                cur->data = (void *)(uintptr_t)1;
+                void * data;
+                if (use_mmap && mapping) {
+                    data = (uint8_t *) mapping->addr + offs;
+                } else {
+                    read_buf.resize(ggml_nbytes(cur));
+                    file.seek(offs, SEEK_SET);
+                    file.read_raw(read_buf.data(), ggml_nbytes(cur));
+                    data = read_buf.data();
+                }
+
+#if defined(GGML_USE_CUBLAS) && !defined(LLAMA_GGML_BACKEND_CUDA_TEST)
+                ggml_cuda_transform_tensor(data, cur);
 #elif defined(GGML_USE_CLBLAST)
-                case GGML_BACKEND_GPU:
-                    ggml_cl_transform_tensor(cur->data, cur);
-                    if (!use_mmap) {
-                        free(cur->data);
-                    }
-                    break;
+                GGML_ASSERT(cur->backend == GGML_BACKEND_GPU);
+                ggml_cl_transform_tensor(data, cur);
+#else
+                GGML_ASSERT(!"GPU tensor without a GPU backend");
+                GGML_UNUSED(data);
 #endif
-                default:
-                    continue;
             }
 
-            done_size += ggml_nbytes(cur);
+            size_done += ggml_nbytes(cur);
         }
+
+        // unmap offloaded tensors and metadata
+        if (use_mmap && mapping) {
+            mapping->unmap_fragment(0, mmap_first);
+            mapping->unmap_fragment(mmap_last, mapping->size);
+        }
+
+        if (progress_callback) {
+            // Even though the model is done loading, we still honor
+            // cancellation since we need to free allocations.
+            return progress_callback(1.0f, progress_callback_user_data);
+        }
+        return true;
     }
 };
 
@@ -2150,25 +2496,25 @@ 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";
     }
@@ -2201,11 +2547,8 @@ static void llm_load_arch(llama_model_loader & ml, llama_model & model) {
 static void llm_load_hparams(
         llama_model_loader & ml,
         llama_model & model) {
-    struct gguf_context * ctx = ml.ctx_gguf;
-
-    const auto kv = LLM_KV(model.arch);
-
     auto & hparams = model.hparams;
+    const gguf_context * ctx = ml.ctx_gguf;
 
     // get metadata as string
     for (int i = 0; i < gguf_get_n_kv(ctx); i++) {
@@ -2219,42 +2562,51 @@ static void llm_load_hparams(
     }
 
     // get general kv
-    GGUF_GET_KEY(ctx, model.name, gguf_get_val_str, GGUF_TYPE_STRING, false, kv(LLM_KV_GENERAL_NAME));
+    ml.get_key(LLM_KV_GENERAL_NAME, model.name, false);
 
     // get hparams kv
-    GGUF_GET_KEY(ctx, hparams.n_vocab,        gguf_get_arr_n,   GGUF_TYPE_ARRAY,  true, kv(LLM_KV_TOKENIZER_LIST));
-    GGUF_GET_KEY(ctx, hparams.n_ctx_train,    gguf_get_val_u32, GGUF_TYPE_UINT32, true, kv(LLM_KV_CONTEXT_LENGTH));
-    GGUF_GET_KEY(ctx, hparams.n_embd,         gguf_get_val_u32, GGUF_TYPE_UINT32, true, kv(LLM_KV_EMBEDDING_LENGTH));
-    GGUF_GET_KEY(ctx, hparams.n_ff,           gguf_get_val_u32, GGUF_TYPE_UINT32, true, kv(LLM_KV_FEED_FORWARD_LENGTH));
-    GGUF_GET_KEY(ctx, hparams.n_head,         gguf_get_val_u32, GGUF_TYPE_UINT32, true, kv(LLM_KV_ATTENTION_HEAD_COUNT));
-    GGUF_GET_KEY(ctx, hparams.n_layer,        gguf_get_val_u32, GGUF_TYPE_UINT32, true, kv(LLM_KV_BLOCK_COUNT));
+    ml.get_arr_n(LLM_KV_TOKENIZER_LIST,       hparams.n_vocab);
+    ml.get_key  (LLM_KV_CONTEXT_LENGTH,       hparams.n_ctx_train);
+    ml.get_key  (LLM_KV_EMBEDDING_LENGTH,     hparams.n_embd);
+    ml.get_key  (LLM_KV_FEED_FORWARD_LENGTH,  hparams.n_ff);
+    ml.get_key  (LLM_KV_ATTENTION_HEAD_COUNT, hparams.n_head);
+    ml.get_key  (LLM_KV_BLOCK_COUNT,          hparams.n_layer);
+    ml.get_key  (LLM_KV_EXPERT_COUNT,         hparams.n_expert,      false);
+    ml.get_key  (LLM_KV_EXPERT_USED_COUNT,    hparams.n_expert_used, false);
+
+    GGML_ASSERT(hparams.n_expert <= LLAMA_MAX_EXPERTS);
+    GGML_ASSERT(hparams.n_expert_used <= hparams.n_expert);
+    if (hparams.n_expert > 0) {
+        GGML_ASSERT(hparams.n_expert_used > 0);
+    } else {
+        GGML_ASSERT(hparams.n_expert_used == 0);
+    }
 
     // n_head_kv is optional, default to n_head
     hparams.n_head_kv = hparams.n_head;
-    GGUF_GET_KEY(ctx, hparams.n_head_kv, gguf_get_val_u32, GGUF_TYPE_UINT32, false, kv(LLM_KV_ATTENTION_HEAD_COUNT_KV));
+    ml.get_key(LLM_KV_ATTENTION_HEAD_COUNT_KV, hparams.n_head_kv, false);
 
-    hparams.rope_finetuned = false;
-    GGUF_GET_KEY(ctx, hparams.rope_finetuned, gguf_get_val_bool, GGUF_TYPE_BOOL, false,
-                 kv(LLM_KV_ROPE_SCALING_FINETUNED));
+    bool rope_finetuned = false;
+    ml.get_key(LLM_KV_ROPE_SCALING_FINETUNED, rope_finetuned, false);
+    hparams.rope_finetuned = rope_finetuned;
 
     hparams.n_yarn_orig_ctx = hparams.n_ctx_train;
-    GGUF_GET_KEY(ctx, hparams.n_yarn_orig_ctx, gguf_get_val_u32, GGUF_TYPE_UINT32, false,
-                 kv(LLM_KV_ROPE_SCALING_ORIG_CTX_LEN));
+    ml.get_key(LLM_KV_ROPE_SCALING_ORIG_CTX_LEN, hparams.n_yarn_orig_ctx, false);
 
     // rope_freq_base (optional)
     hparams.rope_freq_base_train = 10000.0f;
-    GGUF_GET_KEY(ctx, hparams.rope_freq_base_train, gguf_get_val_f32, GGUF_TYPE_FLOAT32, false, kv(LLM_KV_ROPE_FREQ_BASE));
+    ml.get_key(LLM_KV_ROPE_FREQ_BASE, hparams.rope_freq_base_train, false);
 
     std::string rope_scaling("linear");
-    GGUF_GET_KEY(ctx, rope_scaling, gguf_get_val_str, GGUF_TYPE_STRING, false, kv(LLM_KV_ROPE_SCALING_TYPE));
+    ml.get_key(LLM_KV_ROPE_SCALING_TYPE, rope_scaling, false);
     hparams.rope_scaling_type_train = llama_rope_scaling_type_from_string(rope_scaling);
     GGML_ASSERT(hparams.rope_scaling_type_train != LLAMA_ROPE_SCALING_UNSPECIFIED);
 
     // rope_freq_scale (inverse of the kv) is optional
     float ropescale = 0.0f;
-    GGUF_GET_KEY(ctx, ropescale, gguf_get_val_f32, GGUF_TYPE_FLOAT32, false, kv(LLM_KV_ROPE_SCALING_FACTOR));
-    if (ropescale == 0.0f) { // try the old key name
-        GGUF_GET_KEY(ctx, ropescale, gguf_get_val_f32, GGUF_TYPE_FLOAT32, false, kv(LLM_KV_ROPE_SCALE_LINEAR));
+    if (!ml.get_key(LLM_KV_ROPE_SCALING_FACTOR, ropescale, false)) {
+        // try the old key name
+        ml.get_key(LLM_KV_ROPE_SCALE_LINEAR, ropescale, false);
     }
     hparams.rope_freq_scale_train = ropescale == 0.0f ? 1.0f : 1.0f/ropescale;
 
@@ -2262,7 +2614,7 @@ static void llm_load_hparams(
     {
         hparams.n_rot = hparams.n_embd / hparams.n_head;
 
-        GGUF_GET_KEY(ctx, hparams.n_rot, gguf_get_val_u32, GGUF_TYPE_UINT32, false, kv(LLM_KV_ROPE_DIMENSION_COUNT));
+        ml.get_key(LLM_KV_ROPE_DIMENSION_COUNT, hparams.n_rot, false);
 
         if (model.arch == LLM_ARCH_LLAMA || model.arch == LLM_ARCH_FALCON) {
             if (hparams.n_rot != hparams.n_embd / hparams.n_head) {
@@ -2277,9 +2629,10 @@ static void llm_load_hparams(
     switch (model.arch) {
         case LLM_ARCH_LLAMA:
             {
-                GGUF_GET_KEY(ctx, hparams.f_norm_rms_eps, gguf_get_val_f32, GGUF_TYPE_FLOAT32, true, kv(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS));
+                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;
@@ -2291,7 +2644,7 @@ static void llm_load_hparams(
             } break;
         case LLM_ARCH_FALCON:
             {
-                GGUF_GET_KEY(ctx, hparams.f_norm_eps, gguf_get_val_f32, GGUF_TYPE_FLOAT32, true, kv(LLM_KV_ATTENTION_LAYERNORM_EPS));
+                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps);
 
                 switch (hparams.n_layer) {
                     case 32: model.type = e_model::MODEL_7B; break;
@@ -2301,7 +2654,7 @@ static void llm_load_hparams(
             } break;
         case LLM_ARCH_BAICHUAN:
             {
-                GGUF_GET_KEY(ctx, hparams.f_norm_rms_eps, gguf_get_val_f32, GGUF_TYPE_FLOAT32, true, kv(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS));
+                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
                 switch (hparams.n_layer) {
                     case 32: model.type = e_model::MODEL_7B; break;
                     case 40: model.type = e_model::MODEL_13B; break;
@@ -2310,7 +2663,7 @@ static void llm_load_hparams(
             } break;
         case LLM_ARCH_STARCODER:
             {
-                GGUF_GET_KEY(ctx, hparams.f_norm_eps, gguf_get_val_f32, GGUF_TYPE_FLOAT32, true, kv(LLM_KV_ATTENTION_LAYERNORM_EPS));
+                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps);
                 switch (hparams.n_layer) {
                     case 24: model.type = e_model::MODEL_1B; break;
                     case 36: model.type = e_model::MODEL_3B; break;
@@ -2321,7 +2674,7 @@ static void llm_load_hparams(
             } break;
         case LLM_ARCH_PERSIMMON:
             {
-                GGUF_GET_KEY(ctx, hparams.f_norm_eps, gguf_get_val_f32, GGUF_TYPE_FLOAT32, true, kv(LLM_KV_ATTENTION_LAYERNORM_EPS));
+                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps);
                 switch (hparams.n_layer) {
                     case 36: model.type = e_model::MODEL_8B; break;
                     default: model.type = e_model::MODEL_UNKNOWN;
@@ -2329,7 +2682,7 @@ static void llm_load_hparams(
             } break;
         case LLM_ARCH_REFACT:
             {
-                GGUF_GET_KEY(ctx, hparams.f_norm_rms_eps, gguf_get_val_f32, GGUF_TYPE_FLOAT32, true, kv(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS));
+                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
                 switch (hparams.n_layer) {
                     case 32: model.type = e_model::MODEL_1B; break;
                     default: model.type = e_model::MODEL_UNKNOWN;
@@ -2337,7 +2690,7 @@ static void llm_load_hparams(
             } break;
         case LLM_ARCH_BLOOM:
             {
-                GGUF_GET_KEY(ctx, hparams.f_norm_eps, gguf_get_val_f32, GGUF_TYPE_FLOAT32, true, kv(LLM_KV_ATTENTION_LAYERNORM_EPS));
+                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps);
 
                 switch (hparams.n_layer) {
                     case 24: model.type = e_model::MODEL_1B; break;
@@ -2352,9 +2705,9 @@ static void llm_load_hparams(
             {
                 hparams.f_clamp_kqv = 0.0f;
 
-                GGUF_GET_KEY(ctx, hparams.f_norm_eps, gguf_get_val_f32, GGUF_TYPE_FLOAT32, true, kv(LLM_KV_ATTENTION_LAYERNORM_EPS));
-                GGUF_GET_KEY(ctx, hparams.f_clamp_kqv, gguf_get_val_f32, GGUF_TYPE_FLOAT32, false, kv(LLM_KV_ATTENTION_CLAMP_KQV));
-                GGUF_GET_KEY(ctx, hparams.f_max_alibi_bias, gguf_get_val_f32, GGUF_TYPE_FLOAT32, true, kv(LLM_KV_ATTENTION_MAX_ALIBI_BIAS));
+                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS,  hparams.f_norm_eps);
+                ml.get_key(LLM_KV_ATTENTION_CLAMP_KQV,      hparams.f_clamp_kqv, false);
+                ml.get_key(LLM_KV_ATTENTION_MAX_ALIBI_BIAS, hparams.f_max_alibi_bias);
 
                 switch (hparams.n_layer) {
                     case 32: model.type = e_model::MODEL_7B; break;
@@ -2364,7 +2717,7 @@ static void llm_load_hparams(
             } break;
         case LLM_ARCH_STABLELM:
             {
-                GGUF_GET_KEY(ctx, hparams.f_norm_eps, gguf_get_val_f32, GGUF_TYPE_FLOAT32, true, kv(LLM_KV_ATTENTION_LAYERNORM_EPS));
+                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;
@@ -2373,13 +2726,23 @@ static void llm_load_hparams(
             } break;
         case LLM_ARCH_QWEN:
             {
-                GGUF_GET_KEY(ctx, hparams.f_norm_rms_eps, gguf_get_val_f32, GGUF_TYPE_FLOAT32, true, kv(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS));
+                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+
                 switch (hparams.n_layer) {
                     case 32: model.type = e_model::MODEL_7B; break;
                     case 40: model.type = e_model::MODEL_13B; break;
                     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;
     }
@@ -2421,7 +2784,7 @@ static void llm_load_vocab(
     {
         std::string tokenizer_name;
 
-        GGUF_GET_KEY(ctx, tokenizer_name, gguf_get_val_str, GGUF_TYPE_STRING, true, kv(LLM_KV_TOKENIZER_MODEL));
+        ml.get_key(LLM_KV_TOKENIZER_MODEL, tokenizer_name);
 
         if (tokenizer_name == "llama") {
             vocab.type = LLAMA_VOCAB_TYPE_SPM;
@@ -2511,34 +2874,31 @@ static void llm_load_vocab(
         };
         for (const auto & it : special_token_types) {
             const std::string & key = kv(std::get<0>(it));
-            int32_t & id = std::get<1>(it), old_id = id;
+            int32_t & id = std::get<1>(it);
 
-            GGUF_GET_KEY(ctx, id, gguf_get_val_u32, GGUF_TYPE_UINT32, false, key);
-            // Must be >= -1 and < vocab size. Since the key is unsigned, -1
-            // can only come from the default value, so there's no point in
-            // validating that.
-            if (size_t(id + 1) > vocab.id_to_token.size()) {
-                LLAMA_LOG_WARN("%s: bad special token: '%s' = %d, using default id %d\n",
-                    __func__, key.c_str(), id, old_id);
-                id = old_id;
+            uint32_t new_id;
+            if (!ml.get_key(std::get<0>(it), new_id, false)) {
+                continue;
+            }
+            if (new_id >= vocab.id_to_token.size()) {
+                LLAMA_LOG_WARN("%s: bad special token: '%s' = %ud, using default id %d\n",
+                    __func__, key.c_str(), new_id, id);
+            } else {
+                id = new_id;
             }
 
         }
 
         // Handle add_bos_token and add_eos_token
-        std::string key = kv(LLM_KV_TOKENIZER_ADD_BOS);
-        int kid = gguf_find_key(ctx, key.c_str());
-        enum gguf_type ktype = kid < 0 ? GGUF_TYPE_COUNT : gguf_get_kv_type(ctx, kid);
-        vocab.special_add_bos = ktype == GGUF_TYPE_BOOL ? gguf_get_val_bool(ctx, kid) : -1;
-        if (ktype != GGUF_TYPE_BOOL && ktype != GGUF_TYPE_COUNT) {
-            LLAMA_LOG_WARN("%s: bad field type %d for '%s' - ignoring\n", __func__, ktype, key.c_str());
-        }
-        key = kv(LLM_KV_TOKENIZER_ADD_EOS);
-        kid = gguf_find_key(ctx, key.c_str());
-        ktype = kid < 0 ? GGUF_TYPE_COUNT : gguf_get_kv_type(ctx, kid);
-        vocab.special_add_eos = ktype == GGUF_TYPE_BOOL ? gguf_get_val_bool(ctx, kid) : -1;
-        if (ktype != GGUF_TYPE_BOOL && ktype != GGUF_TYPE_COUNT) {
-            LLAMA_LOG_WARN("%s: bad field type %d for '%s' - ignoring\n", __func__, ktype, key.c_str());
+        {
+            bool temp = true;
+
+            if (ml.get_key(LLM_KV_TOKENIZER_ADD_BOS, temp, false)) {
+                vocab.special_add_bos = int(temp);
+            }
+            if (ml.get_key(LLM_KV_TOKENIZER_ADD_EOS, temp, false)) {
+                vocab.special_add_eos = int(temp);
+            }
         }
     }
 
@@ -2549,7 +2909,7 @@ static void llm_load_vocab(
         // The assumption is, since special tokens aren't meant to be exposed to end user, they are designed
         //  to be unmatchable by the tokenizer, therefore tokens from the vocab, which are unmatchable by the tokenizer
         //  are special tokens.
-        // From testing, this appears to corelate 1:1 with special tokens.
+        // From testing, this appears to correlate 1:1 with special tokens.
         //
 
         // Counting special tokens and verifying in only one direction
@@ -2662,6 +3022,8 @@ static void llm_load_print_meta(llama_model_loader & ml, llama_model & model) {
     LLAMA_LOG_INFO("%s: f_clamp_kqv      = %.1e\n",   __func__, hparams.f_clamp_kqv);
     LLAMA_LOG_INFO("%s: f_max_alibi_bias = %.1e\n",   __func__, hparams.f_max_alibi_bias);
     LLAMA_LOG_INFO("%s: n_ff             = %u\n",     __func__, hparams.n_ff);
+    LLAMA_LOG_INFO("%s: n_expert         = %u\n",     __func__, hparams.n_expert);
+    LLAMA_LOG_INFO("%s: n_expert_used    = %u\n",     __func__, hparams.n_expert_used);
     LLAMA_LOG_INFO("%s: rope scaling     = %s\n",     __func__, rope_scaling_type.c_str());
     LLAMA_LOG_INFO("%s: freq_base_train  = %.1f\n",   __func__, hparams.rope_freq_base_train);
     LLAMA_LOG_INFO("%s: freq_scale_train = %g\n",     __func__, hparams.rope_freq_scale_train);
@@ -2688,7 +3050,8 @@ static void llm_load_print_meta(llama_model_loader & ml, llama_model & model) {
     if (vocab.linefeed_id    != -1) { LLAMA_LOG_INFO( "%s: LF token         = %d '%s'\n", __func__, vocab.linefeed_id,    vocab.id_to_token[vocab.linefeed_id].text.c_str() );    }
 }
 
-static void llm_load_tensors(
+// Returns false if cancelled by progress_callback
+static bool llm_load_tensors(
         llama_model_loader & ml,
         llama_model & model,
         int n_gpu_layers,
@@ -2704,25 +3067,16 @@ static void llm_load_tensors(
 
     model.n_gpu_layers = n_gpu_layers;
 
-    size_t ctx_size;
-    size_t mmapped_size;
+    size_t ctx_size = ggml_tensor_overhead() * ml.n_tensors;
 
-    ml.calc_sizes(ctx_size, mmapped_size);
-
-    LLAMA_LOG_INFO("%s: ggml ctx size = %7.2f MiB\n", __func__, ctx_size/1024.0/1024.0);
+    LLAMA_LOG_INFO("%s: ggml ctx size       = %7.2f MiB\n", __func__, ctx_size/1024.0/1024.0);
 
     // create the ggml context
     {
-        model.buf.resize(ctx_size);
-        if (use_mlock) {
-            model.mlock_buf.init   (model.buf.data);
-            model.mlock_buf.grow_to(model.buf.size);
-        }
-
         struct ggml_init_params params = {
-            /*.mem_size   =*/ model.buf.size,
-            /*.mem_buffer =*/ model.buf.data,
-            /*.no_alloc   =*/ ml.use_mmap,
+            /*.mem_size   =*/ ctx_size,
+            /*.mem_buffer =*/ NULL,
+            /*.no_alloc   =*/ true,
         };
 
         model.ctx = ggml_init(params);
@@ -2733,25 +3087,24 @@ 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
+#if defined(GGML_USE_CUBLAS) && !defined(LLAMA_GGML_BACKEND_CUDA_TEST)
     if (ggml_cublas_loaded()) {
         LLAMA_LOG_INFO("%s: using " GGML_CUDA_NAME " for GPU acceleration\n", __func__);
         ggml_cuda_set_main_device(main_gpu);
 
-        llama_backend_offload = GGML_BACKEND_GPU;
+        llama_backend_offload       = GGML_BACKEND_GPU;
         llama_backend_offload_split = GGML_BACKEND_GPU_SPLIT;
     }
 #elif defined(GGML_USE_CLBLAST)
         LLAMA_LOG_INFO("%s: using OpenCL for GPU acceleration\n", __func__);
-        llama_backend_offload = GGML_BACKEND_GPU;
+        llama_backend_offload       = GGML_BACKEND_GPU;
         llama_backend_offload_split = GGML_BACKEND_GPU;
 #endif
 
-    // prepare memory for the weights
-    size_t vram_weights = 0;
+    // create tensors for the weights
     {
         const int64_t n_embd     = hparams.n_embd;
         const int64_t n_embd_gqa = hparams.n_embd_gqa();
@@ -2771,14 +3124,7 @@ static void llm_load_tensors(
                         ggml_backend_type backend_output;
 
                         if (n_gpu_layers > int(n_layer)) {
-                            // norm is not performance relevant on its own but keeping it in VRAM reduces data copying
-                            // on Windows however this is detrimental unless everything is on the GPU
-#ifndef _WIN32
-                            backend_norm = llama_backend_offload;
-#else
-                            backend_norm = n_gpu_layers <= (int) n_layer + 2 ? GGML_BACKEND_CPU : llama_backend_offload;
-#endif // _WIN32
-
+                            backend_norm   = llama_backend_offload;
                             backend_output = llama_backend_offload_split;
                         } else {
                             backend_norm   = GGML_BACKEND_CPU;
@@ -2787,13 +3133,6 @@ static void llm_load_tensors(
 
                         model.output_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd},          backend_norm);
                         model.output      = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, backend_output);
-
-                        if (backend_norm == GGML_BACKEND_GPU) {
-                            vram_weights += ggml_nbytes(model.output_norm);
-                        }
-                        if (backend_output == GGML_BACKEND_GPU_SPLIT) {
-                            vram_weights += ggml_nbytes(model.output);
-                        }
                     }
 
                     const uint32_t n_ff = hparams.n_ff;
@@ -2823,20 +3162,25 @@ static void llm_load_tensors(
 
                         layer.ffn_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, backend);
 
-                        layer.ffn_gate = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, backend_split);
-                        layer.ffn_down = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, backend_split);
-                        layer.ffn_up   = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, backend_split);
+                        layer.ffn_gate_inp = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd}, backend, false);
 
-                        if (backend == GGML_BACKEND_GPU) {
-                            vram_weights +=
-                                ggml_nbytes(layer.attn_norm) + ggml_nbytes(layer.wq) + ggml_nbytes(layer.wk) +
-                                ggml_nbytes(layer.wv) + ggml_nbytes(layer.wo) +
-                                (layer.bq ? ggml_nbytes(layer.bq) : 0) +
-                                (layer.bk ? ggml_nbytes(layer.bk) : 0) +
-                                (layer.bv ? ggml_nbytes(layer.bv) : 0) +
-                                (layer.bo ? ggml_nbytes(layer.bo) : 0) +
-                                ggml_nbytes(layer.ffn_norm) + ggml_nbytes(layer.ffn_gate) +
-                                ggml_nbytes(layer.ffn_down) + ggml_nbytes(layer.ffn_up);
+                        if (layer.ffn_gate_inp == nullptr) {
+                            GGML_ASSERT(hparams.n_expert      == 0);
+                            GGML_ASSERT(hparams.n_expert_used == 0);
+
+                            layer.ffn_gate = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, backend_split);
+                            layer.ffn_down = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, backend_split);
+                            layer.ffn_up   = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, backend_split);
+                        } else {
+                            GGML_ASSERT(hparams.n_expert      > 0);
+                            GGML_ASSERT(hparams.n_expert_used > 0);
+
+                            // MoE branch
+                            for (uint32_t x = 0; x < hparams.n_expert; ++x) {
+                                layer.ffn_gate_exp[x] = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_GATE_EXP, "weight", i, x), {n_embd,   n_ff}, backend_split);
+                                layer.ffn_down_exp[x] = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN_EXP, "weight", i, x), {  n_ff, n_embd}, backend_split);
+                                layer.ffn_up_exp[x]   = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP_EXP,   "weight", i, x), {n_embd,   n_ff}, backend_split);
+                            }
                         }
                     }
                 } break;
@@ -2848,14 +3192,7 @@ static void llm_load_tensors(
                         ggml_backend_type backend_output;
 
                         if (n_gpu_layers > int(n_layer)) {
-                            // norm is not performance relevant on its own but keeping it in VRAM reduces data copying
-                            // on Windows however this is detrimental unless everything is on the GPU
-#ifndef _WIN32
-                            backend_norm = llama_backend_offload;
-#else
-                            backend_norm = n_gpu_layers <= (int) n_layer + 2 ? GGML_BACKEND_CPU : llama_backend_offload;
-#endif // _WIN32
-
+                            backend_norm   = llama_backend_offload;
                             backend_output = llama_backend_offload_split;
                         } else {
                             backend_norm   = GGML_BACKEND_CPU;
@@ -2864,13 +3201,6 @@ static void llm_load_tensors(
 
                         model.output_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd},          backend_norm);
                         model.output      = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, backend_output);
-
-                        if (backend_norm == GGML_BACKEND_GPU) {
-                            vram_weights += ggml_nbytes(model.output_norm);
-                        }
-                        if (backend_output == GGML_BACKEND_GPU_SPLIT) {
-                            vram_weights += ggml_nbytes(model.output);
-                        }
                     }
 
                     const uint32_t n_ff = hparams.n_ff;
@@ -2897,19 +3227,10 @@ static void llm_load_tensors(
                         layer.ffn_gate = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, backend_split);
                         layer.ffn_down = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, backend_split);
                         layer.ffn_up   = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, backend_split);
-
-                        if (backend == GGML_BACKEND_GPU) {
-                            vram_weights +=
-                                ggml_nbytes(layer.attn_norm) + ggml_nbytes(layer.wq)       + ggml_nbytes(layer.wk)       +
-                                ggml_nbytes(layer.wv)        + ggml_nbytes(layer.wo)       + ggml_nbytes(layer.ffn_norm) +
-                                ggml_nbytes(layer.ffn_gate)  + ggml_nbytes(layer.ffn_down) + ggml_nbytes(layer.ffn_up);
-                        }
                     }
                 } break;
             case LLM_ARCH_FALCON:
                 {
-                    // TODO: CPU-only for now
-
                     model.tok_embd = ml.create_tensor(ctx, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, GGML_BACKEND_CPU);
 
                     // output
@@ -2918,14 +3239,7 @@ static void llm_load_tensors(
                         ggml_backend_type backend_output;
 
                         if (n_gpu_layers > int(n_layer)) {
-                            // norm is not performance relevant on its own but keeping it in VRAM reduces data copying
-                            // on Windows however this is detrimental unless everything is on the GPU
-#ifndef _WIN32
-                            backend_norm = llama_backend_offload;
-#else
-                            backend_norm = n_gpu_layers <= (int) n_layer + 2 ? GGML_BACKEND_CPU : llama_backend_offload;
-#endif // _WIN32
-
+                            backend_norm   = llama_backend_offload;
                             backend_output = llama_backend_offload_split;
                         } else {
                             backend_norm   = GGML_BACKEND_CPU;
@@ -2935,14 +3249,6 @@ static void llm_load_tensors(
                         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);
-
-                        if (backend_norm == GGML_BACKEND_GPU) {
-                            vram_weights += ggml_nbytes(model.output_norm);
-                            vram_weights += ggml_nbytes(model.output_norm_b);
-                        }
-                        if (backend_output == GGML_BACKEND_GPU_SPLIT) {
-                            vram_weights += ggml_nbytes(model.output);
-                        }
                     }
 
                     const uint32_t n_ff = hparams.n_ff;
@@ -2963,11 +3269,6 @@ static void llm_load_tensors(
                         if (gguf_find_tensor(ml.ctx_gguf, tn(LLM_TENSOR_ATTN_NORM_2, "weight", i).c_str()) >= 0) {
                             layer.attn_norm_2   = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_NORM_2, "weight", i), {n_embd}, backend);
                             layer.attn_norm_2_b = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_NORM_2, "bias", i),   {n_embd}, backend);
-
-                            if (backend == GGML_BACKEND_GPU) {
-                                vram_weights += ggml_nbytes(layer.attn_norm_2);
-                                vram_weights += ggml_nbytes(layer.attn_norm_2_b);
-                            }
                         }
 
                         layer.wqkv = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}, backend_split);
@@ -2975,13 +3276,6 @@ static void llm_load_tensors(
 
                         layer.ffn_down = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, backend_split);
                         layer.ffn_up   = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, backend_split);
-
-                        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.wo)          +
-                                ggml_nbytes(layer.ffn_down)  + ggml_nbytes(layer.ffn_up);
-                        }
                     }
                 } break;
             case LLM_ARCH_STARCODER:
@@ -2995,14 +3289,7 @@ static void llm_load_tensors(
                         ggml_backend_type backend_output;
 
                         if (n_gpu_layers > int(n_layer)) {
-                            // norm is not performance relevant on its own but keeping it in VRAM reduces data copying
-                            // on Windows however this is detrimental unless everything is on the GPU
-#ifndef _WIN32
-                            backend_norm = llama_backend_offload;
-#else
-                            backend_norm = n_gpu_layers <= (int) n_layer + 2 ? GGML_BACKEND_CPU : llama_backend_offload;
-#endif // _WIN32
-
+                            backend_norm   = llama_backend_offload;
                             backend_output = llama_backend_offload_split;
                         } else {
                             backend_norm   = GGML_BACKEND_CPU;
@@ -3012,14 +3299,6 @@ static void llm_load_tensors(
                         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);
-
-                        if (backend_norm == GGML_BACKEND_GPU) {
-                            vram_weights += ggml_nbytes(model.output_norm);
-                            vram_weights += ggml_nbytes(model.output_norm_b);
-                        }
-                        if (backend_output == GGML_BACKEND_GPU_SPLIT) {
-                            vram_weights += ggml_nbytes(model.output);
-                        }
                     }
 
                     const uint32_t n_ff = hparams.n_ff;
@@ -3051,16 +3330,6 @@ static void llm_load_tensors(
 
                         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_norm)  + ggml_nbytes(layer.ffn_norm_b)  +
-                                ggml_nbytes(layer.ffn_down)  + ggml_nbytes(layer.ffn_down_b)  +
-                                ggml_nbytes(layer.ffn_up)    + ggml_nbytes(layer.ffn_up_b);
-                        }
                     }
                 } break;
             case LLM_ARCH_PERSIMMON:
@@ -3072,21 +3341,7 @@ static void llm_load_tensors(
                         ggml_backend_type backend_output;
 
                         if (n_gpu_layers > int(n_layer)) {
-#ifdef GGML_USE_CUBLAS
-                            if (n_gpu_layers > int(n_layer + 1)) {
-                                LLAMA_LOG_ERROR("%s: CUDA backend missing Persimmon CUDA ops, can offload at most %ld layers. See: https://github.com/ggerganov/llama.cpp/issues/4038\n",
-                                    __func__, n_layer + 1);
-                                throw std::runtime_error("Persimmon CUDA offload failed");
-                            }
-#endif
-                            // norm is not performance relevant on its own but keeping it in VRAM reduces data copying
-                            // on Windows however this is detrimental unless everything is on the GPU
-#ifndef _WIN32
-                            backend_norm = llama_backend_offload;
-#else
-                            backend_norm = n_gpu_layers <= (int) n_layer + 2 ? GGML_BACKEND_CPU : llama_backend_offload;
-#endif // _WIN32
-
+                            backend_norm   = llama_backend_offload;
                             backend_output = llama_backend_offload_split;
                         } else {
                             backend_norm   = GGML_BACKEND_CPU;
@@ -3096,14 +3351,6 @@ static void llm_load_tensors(
                         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);
-
-                        if (backend_norm == GGML_BACKEND_GPU) {
-                            vram_weights += ggml_nbytes(model.output_norm);
-                            vram_weights += ggml_nbytes(model.output_norm_b);
-                        }
-                        if (backend_output == GGML_BACKEND_GPU_SPLIT) {
-                            vram_weights += ggml_nbytes(model.output);
-                        }
                     }
 
                     const uint32_t n_ff = hparams.n_ff;
@@ -3133,8 +3380,6 @@ static void llm_load_tensors(
                 } break;
             case LLM_ARCH_BLOOM:
                 {
-                    // TODO: CPU-only for now
-
                     model.tok_embd   = ml.create_tensor(ctx, tn(LLM_TENSOR_TOKEN_EMBD,      "weight"), {n_embd, n_vocab}, GGML_BACKEND_CPU);
                     model.tok_norm   = ml.create_tensor(ctx, tn(LLM_TENSOR_TOKEN_EMBD_NORM, "weight"), {n_embd},          GGML_BACKEND_CPU);
                     model.tok_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_TOKEN_EMBD_NORM, "bias"),   {n_embd},          GGML_BACKEND_CPU);
@@ -3145,14 +3390,7 @@ static void llm_load_tensors(
                         ggml_backend_type backend_output;
 
                         if (n_gpu_layers > int(n_layer)) {
-                            // norm is not performance relevant on its own but keeping it in VRAM reduces data copying
-                            // on Windows however this is detrimental unless everything is on the GPU
-#ifndef _WIN32
-                            backend_norm = llama_backend_offload;
-#else
-                            backend_norm = n_gpu_layers <= (int) n_layer + 2 ? GGML_BACKEND_CPU : llama_backend_offload;
-#endif // _WIN32
-
+                            backend_norm   = llama_backend_offload;
                             backend_output = llama_backend_offload_split;
                         } else {
                             backend_norm   = GGML_BACKEND_CPU;
@@ -3162,14 +3400,6 @@ static void llm_load_tensors(
                         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);
-
-                        if (backend_norm == GGML_BACKEND_GPU) {
-                            vram_weights += ggml_nbytes(model.output_norm);
-                            vram_weights += ggml_nbytes(model.output_norm_b);
-                        }
-                        if (backend_output == GGML_BACKEND_GPU_SPLIT) {
-                            vram_weights += ggml_nbytes(model.output);
-                        }
                     }
 
                     const uint32_t n_ff = hparams.n_ff;
@@ -3201,16 +3431,6 @@ static void llm_load_tensors(
 
                         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_norm)  + ggml_nbytes(layer.ffn_norm_b)  +
-                                ggml_nbytes(layer.ffn_up)    + ggml_nbytes(layer.ffn_up_b)    +
-                                ggml_nbytes(layer.ffn_down)  + ggml_nbytes(layer.ffn_down_b);
-                        }
                     }
                 } break;
             case LLM_ARCH_MPT:
@@ -3223,14 +3443,7 @@ static void llm_load_tensors(
                         ggml_backend_type backend_output;
 
                         if (n_gpu_layers > int(n_layer)) {
-                            // norm is not performance relevant on its own but keeping it in VRAM reduces data copying
-                            // on Windows however this is detrimental unless everything is on the GPU
-#ifndef _WIN32
-                            backend_norm = llama_backend_offload;
-#else
-                            backend_norm = n_gpu_layers <= (int) n_layer + 2 ? GGML_BACKEND_CPU : llama_backend_offload;
-#endif // _WIN32
-
+                            backend_norm   = llama_backend_offload;
                             backend_output = llama_backend_offload_split;
                         } else {
                             backend_norm   = GGML_BACKEND_CPU;
@@ -3239,13 +3452,6 @@ static void llm_load_tensors(
 
                         model.output_norm   = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd},          backend_norm);
                         model.output        = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, backend_output);
-
-                        if (backend_norm == GGML_BACKEND_GPU) {
-                            vram_weights += ggml_nbytes(model.output_norm);
-                        }
-                        if (backend_output == GGML_BACKEND_GPU_SPLIT) {
-                            vram_weights += ggml_nbytes(model.output);
-                        }
                     }
 
                     const uint32_t n_ff = hparams.n_ff;
@@ -3268,16 +3474,6 @@ static void llm_load_tensors(
 
                         layer.ffn_down = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, backend_split);
                         layer.ffn_up   = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, backend_split);
-
-                        if (backend == GGML_BACKEND_GPU) {
-                            vram_weights +=
-                                ggml_nbytes(layer.attn_norm) +
-                                ggml_nbytes(layer.wqkv)      +
-                                ggml_nbytes(layer.wo)        +
-                                ggml_nbytes(layer.ffn_norm)  +
-                                ggml_nbytes(layer.ffn_down)  +
-                                ggml_nbytes(layer.ffn_up);
-                        }
                     }
                 } break;
             case LLM_ARCH_STABLELM:
@@ -3290,14 +3486,7 @@ static void llm_load_tensors(
                         ggml_backend_type backend_output;
 
                         if (n_gpu_layers > int(n_layer)) {
-                            // norm is not performance relevant on its own but keeping it in VRAM reduces data copying
-                            // on Windows however this is detrimental unless everything is on the GPU
-#ifndef _WIN32
-                            backend_norm = llama_backend_offload;
-#else
-                            backend_norm = n_gpu_layers <= (int) n_layer + 2 ? GGML_BACKEND_CPU : llama_backend_offload;
-#endif // _WIN32
-
+                            backend_norm   = llama_backend_offload;
                             backend_output = llama_backend_offload_split;
                         } else {
                             backend_norm   = GGML_BACKEND_CPU;
@@ -3307,13 +3496,6 @@ static void llm_load_tensors(
                         model.output_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT_NORM, "bias"), {n_embd},          backend_norm);
                         model.output_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd},          backend_norm);
                         model.output      = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, backend_output);
-
-                        if (backend_norm == GGML_BACKEND_GPU) {
-                            vram_weights += ggml_nbytes(model.output_norm);
-                        }
-                        if (backend_output == GGML_BACKEND_GPU_SPLIT) {
-                            vram_weights += ggml_nbytes(model.output);
-                        }
                     }
 
                     const uint32_t n_ff = hparams.n_ff;
@@ -3345,13 +3527,6 @@ static void llm_load_tensors(
                         layer.ffn_gate = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, backend_split);
                         layer.ffn_down = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, backend_split);
                         layer.ffn_up = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, backend_split);
-
-                        if (backend == GGML_BACKEND_GPU) {
-                            vram_weights +=
-                                ggml_nbytes(layer.attn_norm) + ggml_nbytes(layer.wq)       + ggml_nbytes(layer.wk)       +
-                                ggml_nbytes(layer.wv)        + ggml_nbytes(layer.wo)       + ggml_nbytes(layer.ffn_norm) +
-                                ggml_nbytes(layer.ffn_gate)  + ggml_nbytes(layer.ffn_down) + ggml_nbytes(layer.ffn_up);
-                        }
                     }
                 } break;
             case LLM_ARCH_QWEN:
@@ -3362,14 +3537,7 @@ static void llm_load_tensors(
                         ggml_backend_type backend_output;
 
                         if (n_gpu_layers > int(n_layer)) {
-                            // norm is not performance relevant on its own but keeping it in VRAM reduces data copying
-                            // on Windows however this is detrimental unless everything is on the GPU
-#ifndef _WIN32
-                            backend_norm = llama_backend_offload;
-#else
-                            backend_norm = n_gpu_layers <= (int) n_layer + 2 ? GGML_BACKEND_CPU : llama_backend_offload;
-#endif // _WIN32
-
+                            backend_norm   = llama_backend_offload;
                             backend_output = llama_backend_offload_split;
                         } else {
                             backend_norm   = GGML_BACKEND_CPU;
@@ -3378,14 +3546,7 @@ static void llm_load_tensors(
 
                         model.output_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd},          backend_norm);
                         model.output      = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, backend_output);
-
-                        if (backend_norm == GGML_BACKEND_GPU) {
-                            vram_weights += ggml_nbytes(model.output_norm);
-                        }
-                        if (backend_output == GGML_BACKEND_GPU_SPLIT) {
-                            vram_weights += ggml_nbytes(model.output);
-                        }
-                    }
+                   }
 
                     const uint32_t n_ff = hparams.n_ff / 2;
 
@@ -3410,16 +3571,59 @@ static void llm_load_tensors(
                         layer.ffn_gate = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, backend_split);
                         layer.ffn_down = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, backend_split);
                         layer.ffn_up   = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, backend_split);
-
-                        if (backend == GGML_BACKEND_GPU) {
-                            vram_weights +=
-                                ggml_nbytes(layer.attn_norm) + ggml_nbytes(layer.wqkv)     + ggml_nbytes(layer.bqkv)     +
-                                ggml_nbytes(layer.wo)        + ggml_nbytes(layer.ffn_norm) + ggml_nbytes(layer.ffn_gate) +
-                                ggml_nbytes(layer.ffn_down)  + ggml_nbytes(layer.ffn_up);
-                        }
                     }
                 } 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);
+                    }
+
+                    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);
+                    }
+                } break;
             default:
                 throw std::runtime_error("unknown architecture");
         }
@@ -3427,16 +3631,78 @@ static void llm_load_tensors(
 
     ml.done_getting_tensors();
 
+    ml.init_mapping();
+
+    // allocate tensors
+    size_t vram_weights = 0;
+    size_t buf_size = 0;
+
+    ggml_backend_buffer_type_t buft = llama_default_buffer_type(n_gpu_layers);
+
+    for (struct ggml_tensor * t = ggml_get_first_tensor(ctx); t != nullptr; t = ggml_get_next_tensor(ctx, t)) {
+        // GGML_BACKEND_GPU tensors are for CUDA and OpenCL only, which are handled separately without ggml-backend
+        if (t->backend == GGML_BACKEND_CPU) {
+            buf_size += GGML_PAD(ggml_backend_buft_get_alloc_size(buft, t), ggml_backend_buft_get_alignment(buft));
+        } else {
+            vram_weights += ggml_nbytes(t);
+        }
+    }
+
+    // create backend buffer
+    ggml_backend_buffer_t buf_mmap = nullptr;
+
+#ifdef GGML_USE_METAL
+    if (n_gpu_layers > 0) {
+        if (ml.use_mmap) {
+            const size_t max_size = ggml_get_max_tensor_size(ctx);
+            model.buf = ggml_backend_metal_buffer_from_ptr(ml.mapping->addr, ml.mapping->size, max_size);
+            buf_mmap = model.buf;
+        } else {
+            model.buf = ggml_backend_alloc_ctx_tensors_from_buft(ctx, ggml_backend_metal_buffer_type());
+        }
+    }
+#elif defined(GGML_USE_CUBLAS) && defined(LLAMA_GGML_BACKEND_CUDA_TEST)
+    // for testing only
+    if (n_gpu_layers > 0) {
+        model.buf = ggml_backend_alloc_ctx_tensors_from_buft(ctx, ggml_backend_cuda_buffer_type(0));
+    }
+#endif
+
+    if (model.buf == nullptr) {
+        // CPU backend, and indirectly CUDA and OpenCL
+        if (ml.use_mmap) {
+            model.buf = ggml_backend_cpu_buffer_from_ptr(ml.mapping->addr, ml.mapping->size);
+            buf_mmap = model.buf;
+        } else {
+            // allocate only CPU tensors
+            model.buf = ggml_backend_buft_alloc_buffer(buft, buf_size);
+            ggml_tallocr_t alloc = ggml_tallocr_new_from_buffer(model.buf);
+            for (struct ggml_tensor * t = ggml_get_first_tensor(ctx); t != nullptr; t = ggml_get_next_tensor(ctx, t)) {
+                if (t->backend == GGML_BACKEND_CPU) {
+                    ggml_tallocr_alloc(alloc, t);
+                }
+            }
+            ggml_tallocr_free(alloc);
+        }
+    }
+
+    if (use_mlock && ggml_backend_buffer_is_host(model.buf)) {
+        model.mlock_buf.init   (ggml_backend_buffer_get_base(model.buf));
+        model.mlock_buf.grow_to(ggml_backend_buffer_get_size(model.buf));
+    }
+
     // print memory requirements
     {
-        // this is the total memory required to run the inference
-        size_t mem_required =
-            ctx_size +
-            mmapped_size - vram_weights; // weights in VRAM not in memory
+        size_t sys_mem_required = ctx_size + buf_size;
 
-        LLAMA_LOG_INFO("%s: mem required  = %7.2f MiB\n", __func__, mem_required / 1024.0 / 1024.0);
+        if (sys_mem_required > 0) {
+            LLAMA_LOG_INFO("%s: system memory used  = %7.2f MiB\n", __func__, sys_mem_required / 1024.0 / 1024.0);
+        }
+        if (vram_weights > 0) {
+            LLAMA_LOG_INFO("%s: VRAM used           = %7.2f MiB\n", __func__, vram_weights / 1024.0 / 1024.0);
+        }
 
-#if defined(GGML_USE_CUBLAS) || defined(GGML_USE_CLBLAST)
+#if (defined(GGML_USE_CUBLAS) && !defined(LLAMA_GGML_BACKEND_CUDA_TEST)) || defined(GGML_USE_CLBLAST)
         const int n_gpu = std::min(n_gpu_layers, int(hparams.n_layer));
 
         LLAMA_LOG_INFO("%s: offloading %d repeating layers to GPU\n", __func__, n_gpu);
@@ -3444,38 +3710,27 @@ static void llm_load_tensors(
             LLAMA_LOG_INFO("%s: offloading non-repeating layers to GPU\n", __func__);
         }
 
-#ifdef GGML_USE_CUBLAS
-        const int max_backend_supported_layers = hparams.n_layer + 3;
-        const int max_offloadable_layers       = hparams.n_layer + 3;
-#elif GGML_USE_CLBLAST
         const int max_backend_supported_layers = hparams.n_layer + 1;
         const int max_offloadable_layers       = hparams.n_layer + 1;
-#endif // GGML_USE_CUBLAS
 
         LLAMA_LOG_INFO("%s: offloaded %d/%d layers to GPU\n", __func__, std::min(n_gpu_layers, max_offloadable_layers), max_backend_supported_layers);
-        LLAMA_LOG_INFO("%s: VRAM used: %.2f MiB\n", __func__, vram_weights / 1024.0 / 1024.0);
-#else
-        (void) n_gpu_layers;
 #endif // defined(GGML_USE_CUBLAS) || defined(GGML_USE_CLBLAST)
     }
 
-    // populate `tensors_by_name`
+#if defined(GGML_USE_CUBLAS) && !defined(LLAMA_GGML_BACKEND_CUDA_TEST)
+    ggml_cuda_set_tensor_split(tensor_split);
+#else
+    GGML_UNUSED(tensor_split);
+#endif // GGML_USE_CUBLAS
+
+    // populate tensors_by_name
     for (int i = 0; i < ml.n_tensors; ++i) {
         struct ggml_tensor * cur = ggml_get_tensor(ctx, ml.get_tensor_name(i));
         model.tensors_by_name.emplace_back(ggml_get_name(cur), cur);
     }
 
-    (void) tensor_split;
-#ifdef GGML_USE_CUBLAS
-    {
-        ggml_cuda_set_tensor_split(tensor_split);
-    }
-#endif
-
-    ml.load_all_data(ctx, progress_callback, progress_callback_user_data, use_mlock ? &model.mlock_mmap : NULL);
-
-    if (progress_callback) {
-        progress_callback(1.0f, progress_callback_user_data);
+    if (!ml.load_all_data(ctx, progress_callback, progress_callback_user_data, buf_mmap, use_mlock ? &model.mlock_mmap : NULL)) {
+        return false;
     }
 
     model.mapping = std::move(ml.mapping);
@@ -3483,11 +3738,13 @@ static void llm_load_tensors(
     // loading time will be recalculate after the first eval, so
     // we take page faults deferred by mmap() into consideration
     model.t_load_us = ggml_time_us() - model.t_start_us;
+    return true;
 }
 
-static bool llama_model_load(const std::string & fname, llama_model & model, const llama_model_params & params) {
+// Returns 0 on success, -1 on error, and -2 on cancellation via llama_progress_callback
+static int llama_model_load(const std::string & fname, llama_model & model, const llama_model_params & params) {
     try {
-        llama_model_loader ml(fname, params.use_mmap);
+        llama_model_loader ml(fname, params.use_mmap, params.kv_overrides);
 
         model.hparams.vocab_only = params.vocab_only;
 
@@ -3503,19 +3760,21 @@ static bool llama_model_load(const std::string & fname, llama_model & model, con
 
         if (params.vocab_only) {
             LLAMA_LOG_INFO("%s: vocab only - skipping tensors\n", __func__);
-            return true;
+            return 0;
         }
 
-        llm_load_tensors(
+        if (!llm_load_tensors(
             ml, model, params.n_gpu_layers, params.main_gpu, params.tensor_split, params.use_mlock,
             params.progress_callback, params.progress_callback_user_data
-        );
+        )) {
+            return -2;
+        }
     } catch (const std::exception & err) {
         LLAMA_LOG_ERROR("error loading model: %s\n", err.what());
-        return false;
+        return -1;
     }
 
-    return true;
+    return 0;
 }
 
 //
@@ -3614,11 +3873,11 @@ static void llm_build_k_shift(
         struct ggml_tensor * tmp =
             // we rotate only the first n_rot dimensions
             ggml_rope_custom_inplace(ctx,
-                    ggml_view_3d(ctx, kv.k,
+                    ggml_view_3d(ctx, kv.k_l[il],
                         n_embd_head, n_head_kv, n_ctx,
-                        ggml_element_size(kv.k)*n_embd_head,
-                        ggml_element_size(kv.k)*n_embd_gqa,
-                        ggml_element_size(kv.k)*n_embd_gqa*n_ctx*il),
+                        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);
         cb(tmp, "K_shifted", il);
@@ -3645,13 +3904,13 @@ static void llm_build_kv_store(
     //struct ggml_tensor * v_cur_t = ggml_transpose(ctx, v_cur); // TODO: reshape above is likely not needed
     cb(v_cur_t, "v_cur_t", il);
 
-    struct ggml_tensor * k_cache_view = ggml_view_1d(ctx, kv.k, n_tokens*n_embd_gqa,
-            (ggml_element_size(kv.k)*n_embd_gqa)*(il*n_ctx + kv_head));
+    struct ggml_tensor * k_cache_view = ggml_view_1d(ctx, kv.k_l[il], n_tokens*n_embd_gqa,
+            (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, n_tokens, n_embd_gqa,
-            (   n_ctx)*ggml_element_size(kv.v),
-            (il*n_ctx)*ggml_element_size(kv.v)*n_embd_gqa + kv_head*ggml_element_size(kv.v));
+    struct ggml_tensor * v_cache_view = ggml_view_2d(ctx, kv.v_l[il], n_tokens, n_embd_gqa,
+            (  n_ctx)*ggml_element_size(kv.v_l[il]),
+            (kv_head)*ggml_element_size(kv.v_l[il]));
     cb(v_cache_view, "v_cache_view", il);
 
     // important: storing RoPE-ed version of K in the KV cache!
@@ -3780,17 +4039,18 @@ 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,
          struct ggml_tensor * wo_b,
          struct ggml_tensor * q_cur,
-         struct ggml_tensor * kq_scale,
          struct ggml_tensor * kq_mask,
                     int64_t   n_ctx,
                     int32_t   n_tokens,
                     int32_t   n_kv,
                     float     max_alibi_bias,
+                    float     kq_scale,
          const llm_build_cb & cb,
                     int       il) {
     const int64_t n_embd      = hparams.n_embd;
@@ -3803,16 +4063,22 @@ static struct ggml_tensor * llm_build_kqv(
     cb(q, "q", il);
 
     struct ggml_tensor * k =
-        ggml_view_3d(ctx, kv.k,
+        ggml_view_3d(ctx, kv.k_l[il],
                 n_embd_head, n_kv, n_head_kv,
-                ggml_element_size(kv.k)*n_embd_gqa,
-                ggml_element_size(kv.k)*n_embd_head,
-                ggml_element_size(kv.k)*n_embd_gqa*n_ctx*il);
+                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);
@@ -3832,17 +4098,17 @@ 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, kq_scale);
         cb(kq, "kq_soft_max_ext", il);
     }
 
     // split cached v into n_head heads
     struct ggml_tensor * v =
-        ggml_view_3d(ctx, kv.v,
+        ggml_view_3d(ctx, kv.v_l[il],
                 n_kv, n_embd_head, n_head_kv,
-                ggml_element_size(kv.v)*n_ctx,
-                ggml_element_size(kv.v)*n_ctx*n_embd_head,
-                ggml_element_size(kv.v)*n_ctx*n_embd_gqa*il);
+                ggml_element_size(kv.v_l[il])*n_ctx,
+                ggml_element_size(kv.v_l[il])*n_ctx*n_embd_head,
+                0);
     cb(v, "v", il);
 
     struct ggml_tensor * kqv = ggml_mul_mat(ctx, v, kq);
@@ -3880,6 +4146,8 @@ struct llm_build_context {
     const int64_t n_head_kv;
     const int64_t n_embd_head;
     const int64_t n_embd_gqa;
+    const int64_t n_expert;
+    const int64_t n_expert_used;
 
     const float freq_base;
     const float freq_scale;
@@ -3899,7 +4167,7 @@ struct llm_build_context {
 
     const llm_build_cb & cb;
 
-    llama_buffer & buf_compute;
+    std::vector<uint8_t> & buf_compute_meta;
 
     struct ggml_context * ctx0 = nullptr;
 
@@ -3909,33 +4177,35 @@ struct llm_build_context {
     const llama_batch  & batch,
     const llm_build_cb & cb,
                   bool   worst_case) :
-        model         (lctx.model),
-        hparams       (model.hparams),
-        cparams       (lctx.cparams),
-        batch         (batch),
-        kv_self       (lctx.kv_self),
-        n_embd        (hparams.n_embd),
-        n_layer       (hparams.n_layer),
-        n_ctx         (cparams.n_ctx),
-        n_head        (hparams.n_head),
-        n_head_kv     (hparams.n_head_kv),
-        n_embd_head   (hparams.n_embd_head()),
-        n_embd_gqa    (hparams.n_embd_gqa()),
-        freq_base     (cparams.rope_freq_base),
-        freq_scale    (cparams.rope_freq_scale),
-        ext_factor    (cparams.yarn_ext_factor),
-        attn_factor   (cparams.yarn_attn_factor),
-        beta_fast     (cparams.yarn_beta_fast),
-        beta_slow     (cparams.yarn_beta_slow),
-        norm_eps      (hparams.f_norm_eps),
-        norm_rms_eps  (hparams.f_norm_rms_eps),
-        n_tokens      (batch.n_tokens),
-        n_kv          (worst_case ? n_ctx            : kv_self.n),
-        kv_head       (worst_case ? n_ctx - n_tokens : kv_self.head),
-        n_orig_ctx    (cparams.n_yarn_orig_ctx),
-        do_rope_shift (worst_case || kv_self.has_shift),
-        cb            (cb),
-        buf_compute   (lctx.buf_compute) {
+        model            (lctx.model),
+        hparams          (model.hparams),
+        cparams          (lctx.cparams),
+        batch            (batch),
+        kv_self          (lctx.kv_self),
+        n_embd           (hparams.n_embd),
+        n_layer          (hparams.n_layer),
+        n_ctx            (cparams.n_ctx),
+        n_head           (hparams.n_head),
+        n_head_kv        (hparams.n_head_kv),
+        n_embd_head      (hparams.n_embd_head()),
+        n_embd_gqa       (hparams.n_embd_gqa()),
+        n_expert         (hparams.n_expert),
+        n_expert_used    (hparams.n_expert_used),
+        freq_base        (cparams.rope_freq_base),
+        freq_scale       (cparams.rope_freq_scale),
+        ext_factor       (cparams.yarn_ext_factor),
+        attn_factor      (cparams.yarn_attn_factor),
+        beta_fast        (cparams.yarn_beta_fast),
+        beta_slow        (cparams.yarn_beta_slow),
+        norm_eps         (hparams.f_norm_eps),
+        norm_rms_eps     (hparams.f_norm_rms_eps),
+        n_tokens         (batch.n_tokens),
+        n_kv             (worst_case ? n_ctx            : kv_self.n),
+        kv_head          (worst_case ? n_ctx - n_tokens : kv_self.head),
+        n_orig_ctx       (cparams.n_yarn_orig_ctx),
+        do_rope_shift    (worst_case || kv_self.has_shift),
+        cb               (cb),
+        buf_compute_meta (lctx.buf_compute_meta) {
             GGML_ASSERT(!!kv_self.ctx);
 
             // all initializations should be done in init()
@@ -3943,8 +4213,8 @@ struct llm_build_context {
 
     void init() {
         struct ggml_init_params params = {
-            /*.mem_size   =*/ buf_compute.size,
-            /*.mem_buffer =*/ buf_compute.data,
+            /*.mem_size   =*/ buf_compute_meta.size(),
+            /*.mem_buffer =*/ buf_compute_meta.data(),
             /*.no_alloc   =*/ true,
         };
 
@@ -3973,10 +4243,6 @@ struct llm_build_context {
         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);
-        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);
@@ -4035,9 +4301,9 @@ 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_mask, n_ctx, n_tokens, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il);
                 cb(cur, "kqv_out", il);
             }
 
@@ -4045,7 +4311,7 @@ struct llm_build_context {
             cb(ffn_inp, "ffn_inp", il);
 
             // feed-forward network
-            {
+            if (model.layers[il].ffn_gate_inp == nullptr) {
                 cur = llm_build_norm(ctx0, ffn_inp, hparams,
                         model.layers[il].ffn_norm, NULL,
                         LLM_NORM_RMS, cb, il);
@@ -4057,6 +4323,69 @@ struct llm_build_context {
                         model.layers[il].ffn_down, NULL,
                         LLM_FFN_SILU, LLM_FFN_PAR, cb, il);
                 cb(cur, "ffn_out", il);
+            } else {
+                // MoE branch
+                cur = llm_build_norm(ctx0, ffn_inp, hparams,
+                        model.layers[il].ffn_norm, NULL,
+                        LLM_NORM_RMS, cb, il);
+                cb(cur, "ffn_norm", il);
+
+                ggml_tensor * logits = ggml_mul_mat(ctx0, model.layers[il].ffn_gate_inp, cur); // [n_tokens, num_experts]
+                cb(logits, "ffn_moe_logits", il);
+
+                ggml_tensor * probs = ggml_soft_max(ctx0, logits); // [n_tokens, num_experts]
+                cb(probs, "ffn_moe_probs", il);
+
+                // select experts
+                ggml_tensor * selected_experts = ggml_top_k(ctx0, probs, n_expert_used); // [n_tokens, num_experts_per_tok]
+                cb(selected_experts->src[0], "ffn_moe_argsort", il);
+
+                ggml_tensor * weights = ggml_get_rows(ctx0,
+                        ggml_reshape_3d(ctx0, probs, 1, n_expert, n_tokens), selected_experts);
+                cb(weights, "ffn_moe_weights", il);
+
+                weights = ggml_reshape_2d(ctx0, weights, n_expert_used, n_tokens); // [n_tokens, num_experts_per_tok]
+
+                ggml_tensor * weights_sum = ggml_sum_rows(ctx0, weights);
+                cb(weights_sum, "ffn_moe_weights_sum", il);
+
+                weights = ggml_div(ctx0, weights, weights_sum); // [n_tokens, num_experts_per_tok]
+                cb(weights, "ffn_moe_weights_norm", il);
+
+                // compute expert outputs
+                ggml_tensor * moe_out = nullptr;
+
+                for (int i = 0; i < n_expert_used; ++i) {
+                    ggml_tensor * cur_expert;
+
+                    ggml_tensor * cur_up = ggml_mul_mat_id(ctx0, model.layers[il].ffn_up_exp, n_expert, selected_experts, i, cur);
+                    cb(cur_up, "ffn_moe_up", il);
+
+                    ggml_tensor * cur_gate = ggml_mul_mat_id(ctx0, model.layers[il].ffn_gate_exp, n_expert, selected_experts, i, cur);
+                    cb(cur_gate, "ffn_moe_gate", il);
+
+                    cur_gate = ggml_silu(ctx0, cur_gate);
+                    cb(cur_gate, "ffn_moe_silu", il);
+
+                    cur_expert = ggml_mul(ctx0, cur_up, cur_gate); // [n_tokens, n_embd]
+                    cb(cur_expert, "ffn_moe_gate_par", il);
+
+                    cur_expert = ggml_mul_mat_id(ctx0, model.layers[il].ffn_down_exp, n_expert, selected_experts, i, cur_expert); // [n_tokens, n_embd]
+                    cb(cur_expert, "ffn_moe_down", il);
+
+                    cur_expert = ggml_mul(ctx0, cur_expert,
+                            ggml_view_2d(ctx0, weights, 1, n_tokens, weights->nb[1], i*weights->nb[0]));
+                    cb(cur_expert, "ffn_moe_weighted", il);
+
+                    if (i == 0) {
+                        moe_out = cur_expert;
+                    } else {
+                        moe_out = ggml_add(ctx0, moe_out, cur_expert);
+                        cb(moe_out, "ffn_moe_out", il);
+                    }
+                }
+
+                cur = moe_out;
             }
 
             cur = ggml_add(ctx0, cur, ffn_inp);
@@ -4095,10 +4424,6 @@ struct llm_build_context {
         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);
-        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);
@@ -4155,9 +4480,9 @@ 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_mask, n_ctx, n_tokens, n_kv, max_alibi_bias, 1.0f/sqrtf(float(n_embd_head)), cb, il);
                 cb(cur, "kqv_out", il);
             }
 
@@ -4215,10 +4540,6 @@ struct llm_build_context {
         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);
-        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);
@@ -4279,9 +4600,9 @@ 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_mask, n_ctx, n_tokens, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il);
                 cb(cur, "kqv_out", il);
             }
 
@@ -4338,10 +4659,6 @@ struct llm_build_context {
         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);
-        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);
@@ -4379,9 +4696,9 @@ 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_mask, n_ctx, n_tokens, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il);
                 cb(cur, "kqv_out", il);
             }
 
@@ -4434,13 +4751,11 @@ struct llm_build_context {
         inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, cb);
         cb(inpL, "imp_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);
 
-        // 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);
 
@@ -4586,9 +4901,9 @@ 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_mask, n_ctx, n_tokens, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il);
                 cb(cur, "kqv_out", il);
             }
 
@@ -4642,10 +4957,6 @@ struct llm_build_context {
         inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, cb);
         cb(inpL, "inp_embd", -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);
@@ -4677,9 +4988,9 @@ 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_mask, n_ctx, n_tokens, n_kv, 8.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il);
                 cb(cur, "kqv_out", il);
             }
 
@@ -4733,10 +5044,6 @@ struct llm_build_context {
         inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, cb);
         cb(inpL, "inp_embd", -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);
@@ -4774,9 +5081,9 @@ 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_mask, n_ctx, n_tokens, n_kv, 8.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il);
                 cb(cur, "kqv_out", il);
             }
 
@@ -4827,10 +5134,6 @@ struct llm_build_context {
         inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, cb);
         cb(inpL, "inp_embd", -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);
@@ -4868,9 +5171,9 @@ 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_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);
             }
 
@@ -4930,10 +5233,6 @@ struct llm_build_context {
         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);
-        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);
@@ -4981,9 +5280,9 @@ 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_mask, n_ctx, n_tokens, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il);
                 cb(cur, "kqv_out", il);
             }
 
@@ -5043,11 +5342,7 @@ struct llm_build_context {
         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);
-        cb(KQ_scale, "KQ_scale", -1);
-
-        // KQ_mask (mask for 1 head, it wil be broadcasted to all heads)
+        // 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);
 
@@ -5098,9 +5393,9 @@ 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_mask, n_ctx, n_tokens, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il);
                 cb(cur, "kqv_out", il);
             }
 
@@ -5142,6 +5437,116 @@ 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);
+
+        // 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);
+
+                // 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
+                Qcur = ggml_scale(ctx0, Qcur, 1.0f/sqrtf(float(n_embd_head)));
+                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_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;
     }
 };
@@ -5154,10 +5559,10 @@ struct llm_build_context {
 enum llm_offload_func_e {
     OFFLOAD_FUNC_NOP,
     OFFLOAD_FUNC,
-    OFFLOAD_FUNC_KQ,
-    OFFLOAD_FUNC_V,
+    OFFLOAD_FUNC_FRC, // force offload
+    OFFLOAD_FUNC_KQV,
     OFFLOAD_FUNC_NR,
-    OFFLOAD_FUNC_EMB,
+    OFFLOAD_FUNC_EMB, // embeddings
     OFFLOAD_FUNC_OUT,
 };
 
@@ -5241,11 +5646,11 @@ static const std::unordered_map<const char *, llm_offload_func_e> k_offload_map
   //{ "inp_embd",                   OFFLOAD_FUNC_NR  }, // TODO: missing K-quants get_rows kernel
     { "pos_embd",                   OFFLOAD_FUNC_NR  },
 
-    { "inp_pos",                    OFFLOAD_FUNC_KQ  }, // this is often used for KQ ops (e.g. rope)
-    { "KQ_scale",                   OFFLOAD_FUNC_KQ  },
-    { "KQ_mask",                    OFFLOAD_FUNC_KQ  },
-    { "K_shift",                    OFFLOAD_FUNC_KQ  },
-    { "K_shifted",                  OFFLOAD_FUNC_KQ  },
+    { "inp_pos",                    OFFLOAD_FUNC_FRC }, // this is often used for KQ ops (e.g. rope)
+    { "KQ_mask",                    OFFLOAD_FUNC_FRC },
+    { "K_shift",                    OFFLOAD_FUNC_FRC },
+
+    { "K_shifted",                  OFFLOAD_FUNC     },
 
     { "inp_norm",                   OFFLOAD_FUNC_NR  },
     { "inp_norm_w",                 OFFLOAD_FUNC_NR  },
@@ -5258,38 +5663,38 @@ static const std::unordered_map<const char *, llm_offload_func_e> k_offload_map
     { "attn_norm",                  OFFLOAD_FUNC     },
     { "attn_norm_2",                OFFLOAD_FUNC     },
 
-    { "wqkv",                       OFFLOAD_FUNC_KQ  },
-    { "bqkv",                       OFFLOAD_FUNC_KQ  },
-    { "wqkv_clamped",               OFFLOAD_FUNC_KQ  },
+    { "wqkv",                       OFFLOAD_FUNC_KQV },
+    { "bqkv",                       OFFLOAD_FUNC_KQV },
+    { "wqkv_clamped",               OFFLOAD_FUNC_KQV },
 
-    { "tmpk",                       OFFLOAD_FUNC_KQ  },
-    { "tmpq",                       OFFLOAD_FUNC_KQ  },
-    { "tmpv",                       OFFLOAD_FUNC_V   },
-    { "Kcur",                       OFFLOAD_FUNC_KQ  },
-    { "Qcur",                       OFFLOAD_FUNC_KQ  },
-    { "Vcur",                       OFFLOAD_FUNC_V   },
+    { "tmpk",                       OFFLOAD_FUNC_KQV },
+    { "tmpq",                       OFFLOAD_FUNC_KQV },
+    { "tmpv",                       OFFLOAD_FUNC_KQV },
+    { "Kcur",                       OFFLOAD_FUNC_KQV },
+    { "Qcur",                       OFFLOAD_FUNC_KQV },
+    { "Vcur",                       OFFLOAD_FUNC_KQV },
 
-    { "krot",                       OFFLOAD_FUNC_KQ  },
-    { "qrot",                       OFFLOAD_FUNC_KQ  },
-    { "kpass",                      OFFLOAD_FUNC_KQ  },
-    { "qpass",                      OFFLOAD_FUNC_KQ  },
-    { "krotated",                   OFFLOAD_FUNC_KQ  },
-    { "qrotated",                   OFFLOAD_FUNC_KQ  },
+    { "krot",                       OFFLOAD_FUNC_KQV },
+    { "qrot",                       OFFLOAD_FUNC_KQV },
+    { "kpass",                      OFFLOAD_FUNC_KQV },
+    { "qpass",                      OFFLOAD_FUNC_KQV },
+    { "krotated",                   OFFLOAD_FUNC_KQV },
+    { "qrotated",                   OFFLOAD_FUNC_KQV },
 
-    { "q",                          OFFLOAD_FUNC_KQ  },
-    { "k",                          OFFLOAD_FUNC_KQ  },
-    { "kq",                         OFFLOAD_FUNC_KQ  },
-    { "kq_scaled",                  OFFLOAD_FUNC_KQ  },
-    { "kq_scaled_alibi",            OFFLOAD_FUNC_KQ  },
-    { "kq_masked",                  OFFLOAD_FUNC_KQ  },
-    { "kq_soft_max",                OFFLOAD_FUNC_V   },
-    { "kq_soft_max_ext",            OFFLOAD_FUNC_V   },
-    { "v",                          OFFLOAD_FUNC_V   },
-    { "kqv",                        OFFLOAD_FUNC_V   },
-    { "kqv_merged",                 OFFLOAD_FUNC_V   },
-    { "kqv_merged_cont",            OFFLOAD_FUNC_V   },
-    { "kqv_wo",                     OFFLOAD_FUNC_V   },
-    { "kqv_out",                    OFFLOAD_FUNC_V   },
+    { "q",                          OFFLOAD_FUNC_KQV },
+    { "k",                          OFFLOAD_FUNC_KQV },
+    { "kq",                         OFFLOAD_FUNC_KQV },
+    { "kq_scaled",                  OFFLOAD_FUNC_KQV },
+    { "kq_scaled_alibi",            OFFLOAD_FUNC_KQV },
+    { "kq_masked",                  OFFLOAD_FUNC_KQV },
+    { "kq_soft_max",                OFFLOAD_FUNC_KQV },
+    { "kq_soft_max_ext",            OFFLOAD_FUNC_KQV },
+    { "v",                          OFFLOAD_FUNC_KQV },
+    { "kqv",                        OFFLOAD_FUNC_KQV },
+    { "kqv_merged",                 OFFLOAD_FUNC_KQV },
+    { "kqv_merged_cont",            OFFLOAD_FUNC_KQV },
+    { "kqv_wo",                     OFFLOAD_FUNC_KQV },
+    { "kqv_out",                    OFFLOAD_FUNC_KQV },
 
     { "ffn_inp",                    OFFLOAD_FUNC     },
     { "ffn_norm",                   OFFLOAD_FUNC     },
@@ -5308,9 +5713,24 @@ static const std::unordered_map<const char *, llm_offload_func_e> k_offload_map
     { "ffn_relu",                   OFFLOAD_FUNC     },
     { "ffn_sqr(relu)",              OFFLOAD_FUNC     },
 
+    { "ffn_moe_logits",             OFFLOAD_FUNC     },
+    { "ffn_moe_probs",              OFFLOAD_FUNC     },
+    { "ffn_moe_argsort",            OFFLOAD_FUNC     },
+    { "ffn_moe_weights",            OFFLOAD_FUNC     },
+    { "ffn_moe_weights_sum",        OFFLOAD_FUNC     },
+    { "ffn_moe_weights_norm",       OFFLOAD_FUNC     },
+    { "ffn_moe_weighted",           OFFLOAD_FUNC     },
+    { "ffn_moe_up",                 OFFLOAD_FUNC     },
+    { "ffn_moe_gate",               OFFLOAD_FUNC     },
+    { "ffn_moe_silu",               OFFLOAD_FUNC     },
+    { "ffn_moe_gate_par",           OFFLOAD_FUNC     },
+    { "ffn_moe_down",               OFFLOAD_FUNC     },
+    { "ffn_moe_out",                OFFLOAD_FUNC     },
+
     { "l_out",                      OFFLOAD_FUNC     },
 
     { "result_norm",                OFFLOAD_FUNC_EMB },
+    { "result_output_no_bias",      OFFLOAD_FUNC_EMB },
     { "result_output",              OFFLOAD_FUNC_OUT },
 };
 
@@ -5328,11 +5748,10 @@ 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_KQ_scale = false;
     bool alloc_inp_KQ_mask  = false;
     bool alloc_inp_K_shift  = false;
 
-#ifdef GGML_USE_CUBLAS
+#if defined(GGML_USE_CUBLAS) && !defined(LLAMA_GGML_BACKEND_CUDA_TEST)
     const bool do_offload = true;
 #else
     const bool do_offload = true; // TODO: set to false after finishing refactoring
@@ -5360,7 +5779,7 @@ static struct ggml_cgraph * llama_build_graph(
             if (!ggml_allocr_is_measure(lctx.alloc) && batch.token) {
                 const int64_t n_tokens = cur->ne[0];
 
-                memcpy(cur->data, batch.token, n_tokens*ggml_element_size(cur));
+                ggml_backend_tensor_set(cur, batch.token, 0, n_tokens*ggml_element_size(cur));
             }
 
             alloc_inp_tokens = true;
@@ -5373,7 +5792,7 @@ static struct ggml_cgraph * llama_build_graph(
                 const int64_t n_embd   = cur->ne[0];
                 const int64_t n_tokens = cur->ne[1];
 
-                memcpy(cur->data, batch.embd, n_tokens*n_embd*ggml_element_size(cur));
+                ggml_backend_tensor_set(cur, batch.embd, 0, n_tokens*n_embd*ggml_element_size(cur));
             }
 
             alloc_inp_embd = true;
@@ -5385,27 +5804,13 @@ static struct ggml_cgraph * llama_build_graph(
             if (!ggml_allocr_is_measure(lctx.alloc) && batch.pos) {
                 const int64_t n_tokens = cur->ne[0];
 
-                int32_t * data = (int32_t *) cur->data;
-
-                for (int i = 0; i < n_tokens; ++i) {
-                    data[i] = batch.pos[i];
-                }
+                static_assert(std::is_same<llama_pos, int32_t>::value, "llama_pos must be int32_t");
+                ggml_backend_tensor_set(cur, batch.pos, 0, n_tokens*ggml_element_size(cur));
             }
 
             alloc_inp_pos = 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();
-                ggml_set_f32(cur, 1.0f/sqrtf(float(n_embd_head)));
-            }
-
-            alloc_inp_KQ_scale = true;
-        }
-
         if (!alloc_inp_KQ_mask && strcmp(name, "KQ_mask") == 0) {
             ggml_allocr_alloc(lctx.alloc, cur);
 
@@ -5413,8 +5818,13 @@ static struct ggml_cgraph * llama_build_graph(
                 const int64_t n_kv     = cur->ne[0];
                 const int64_t n_tokens = cur->ne[1];
 
-                float * data = (float *) cur->data;
-                memset(data, 0, ggml_nbytes(cur));
+                float * data;
+                if (ggml_backend_buffer_is_host(cur->buffer)) {
+                    data = (float *) cur->data;
+                } else {
+                    lctx.buf_copy.resize(ggml_nbytes(cur));
+                    data = (float *) lctx.buf_copy.data();
+                }
 
                 for (int h = 0; h < 1; ++h) {
                     for (int j = 0; j < n_tokens; ++j) {
@@ -5422,12 +5832,20 @@ static struct ggml_cgraph * llama_build_graph(
                         const llama_seq_id seq_id = batch.seq_id[j][0];
 
                         for (int i = 0; i < n_kv; ++i) {
+                            float f;
                             if (!lctx.kv_self.cells[i].has_seq_id(seq_id) || lctx.kv_self.cells[i].pos > pos) {
-                                data[h*(n_kv*n_tokens) + j*n_kv + i] = -INFINITY;
+                                f = -INFINITY;
+                            } else {
+                                f = 0;
                             }
+                            data[h*(n_kv*n_tokens) + j*n_kv + i] = f;
                         }
                     }
                 }
+
+                if (data != cur->data) {
+                    ggml_backend_tensor_set(cur, data, 0, ggml_nbytes(cur));
+                }
             }
 
             alloc_inp_KQ_mask = true;
@@ -5439,11 +5857,21 @@ static struct ggml_cgraph * llama_build_graph(
             if (!ggml_allocr_is_measure(lctx.alloc)) {
                 const int64_t n_ctx = cur->ne[0];
 
-                int32_t * data = (int32_t *) cur->data;
+                int32_t * data;
+                if (ggml_backend_buffer_is_host(cur->buffer)) {
+                    data = (int32_t *) cur->data;
+                } else {
+                    lctx.buf_copy.resize(ggml_nbytes(cur));
+                    data = (int32_t *) lctx.buf_copy.data();
+                }
 
                 for (int i = 0; i < n_ctx; ++i) {
                     data[i] = lctx.kv_self.cells[i].delta;
                 }
+
+                if (data != cur->data) {
+                    ggml_backend_tensor_set(cur, data, 0, ggml_nbytes(cur));
+                }
             }
 
             alloc_inp_K_shift = true;
@@ -5480,16 +5908,16 @@ static struct ggml_cgraph * llama_build_graph(
         static const std::unordered_map<llm_offload_func_e, std::string, std::hash<int>> k_offload_func_name = {
             { OFFLOAD_FUNC_NOP, "CPU" },
             { OFFLOAD_FUNC_OUT, "CPU" },
-#ifdef GGML_USE_CUBLAS
-            { OFFLOAD_FUNC,     "GPU (CUDA)" },
-            { OFFLOAD_FUNC_KQ,  "GPU (CUDA) KQ" },
-            { OFFLOAD_FUNC_V,   "GPU (CUDA) V" },
-            { OFFLOAD_FUNC_NR,  "GPU (CUDA) NR" },
+#if defined(GGML_USE_CUBLAS) && !defined(LLAMA_GGML_BACKEND_CUDA_TEST)
+            { OFFLOAD_FUNC,     "GPU (CUDA)"     },
+            { OFFLOAD_FUNC_FRC, "GPU (CUDA) FRC" },
+            { OFFLOAD_FUNC_KQV, "GPU (CUDA) KQV" },
+            { OFFLOAD_FUNC_NR,  "GPU (CUDA) NR"  },
             { OFFLOAD_FUNC_EMB, "GPU (CUDA) EMB" },
 #else
             { OFFLOAD_FUNC,     "CPU" },
-            { OFFLOAD_FUNC_KQ,  "CPU" },
-            { OFFLOAD_FUNC_V,   "CPU" },
+            { OFFLOAD_FUNC_FRC, "CPU" },
+            { OFFLOAD_FUNC_KQV, "CPU" },
             { OFFLOAD_FUNC_NR,  "CPU" },
             { OFFLOAD_FUNC_EMB, "CPU" },
 #endif // GGML_USE_CUBLAS
@@ -5522,21 +5950,26 @@ static struct ggml_cgraph * llama_build_graph(
                     }
                 }
                 break;
+            case OFFLOAD_FUNC_FRC:
+                if (!lctx.cparams.offload_kqv) {
+                    func_e = OFFLOAD_FUNC_NOP;
+                } break;
+            case OFFLOAD_FUNC_KQV:
+                if (!lctx.cparams.offload_kqv) {
+                    func_e = OFFLOAD_FUNC_NOP;
+                } else {
+                    if (n_gpu_layers < n_layer) {
+                        if (il < i_gpu_start) {
+                            func_e = OFFLOAD_FUNC_NOP;
+                        }
+                    }
+                }
+                break;
             case OFFLOAD_FUNC_NR:
                 if (n_gpu_layers <= n_layer + 0) {
                     func_e = OFFLOAD_FUNC_NOP;
                 }
                 break;
-            case OFFLOAD_FUNC_V:
-                if (n_gpu_layers <= n_layer + 1) {
-                    func_e = OFFLOAD_FUNC_NOP;
-                }
-                break;
-            case OFFLOAD_FUNC_KQ:
-                if (n_gpu_layers <= n_layer + 2) {
-                    func_e = OFFLOAD_FUNC_NOP;
-                }
-                break;
             case OFFLOAD_FUNC_EMB:
                 if (!offload_emb || n_gpu_layers < n_layer) {
                     func_e = OFFLOAD_FUNC_NOP;
@@ -5548,7 +5981,7 @@ static struct ggml_cgraph * llama_build_graph(
         offload_func_t func = ggml_offload_nop;
 
         // this is needed for compatibility with Metal for example
-#ifdef GGML_USE_CUBLAS
+#if defined(GGML_USE_CUBLAS) && !defined(LLAMA_GGML_BACKEND_CUDA_TEST)
         static offload_func_t ggml_offload_gpu = ggml_cuda_assign_buffers_no_alloc;
 #else
         static offload_func_t ggml_offload_gpu = ggml_offload_nop;
@@ -5558,8 +5991,8 @@ static struct ggml_cgraph * llama_build_graph(
             case OFFLOAD_FUNC_NOP:
             case OFFLOAD_FUNC_OUT: func = ggml_offload_nop; break;
             case OFFLOAD_FUNC:
-            case OFFLOAD_FUNC_KQ:
-            case OFFLOAD_FUNC_V:
+            case OFFLOAD_FUNC_KQV:
+            case OFFLOAD_FUNC_FRC:
             case OFFLOAD_FUNC_NR:
             case OFFLOAD_FUNC_EMB: func = ggml_offload_gpu; break;
             default: GGML_ASSERT(false);
@@ -5622,6 +6055,10 @@ 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);
     }
@@ -5699,7 +6136,7 @@ static int llama_decode_internal(
     const int64_t n_embd  = hparams.n_embd;
     const int64_t n_vocab = hparams.n_vocab;
 
-    // helpers for smoother batch API transistion
+    // helpers for smoother batch API transition
     // after deprecating the llama_eval calls, these will be removed
     std::vector<llama_pos> pos;
 
@@ -5745,6 +6182,7 @@ static int llama_decode_internal(
     // after enough generations, the benefit from this heuristic disappears
     // if we start defragmenting the cache, the benefit from this will be more important
     kv_self.n = std::min((int32_t) cparams.n_ctx, std::max(32, GGML_PAD(llama_kv_cache_cell_max(kv_self), 32)));
+    //kv_self.n = llama_kv_cache_cell_max(kv_self);
 
     //printf("kv_self.n = %5d, kv_self.used = %5d, kv_self.head = %5d\n", kv_self.n, kv_self.used, kv_self.head);
 
@@ -5754,18 +6192,23 @@ 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];
+    if (strcmp(embeddings->name, "result_norm") != 0) {
+        embeddings = gf->nodes[gf->n_nodes - 3];
+        GGML_ASSERT(strcmp(embeddings->name, "result_norm") == 0);
+    }
 
-    GGML_ASSERT(strcmp(res->name,        "result_output") == 0);
-    GGML_ASSERT(strcmp(embeddings->name, "result_norm")   == 0);
-
-
-#ifdef GGML_USE_CUBLAS
+#if defined(GGML_USE_CUBLAS) && !defined(LLAMA_GGML_BACKEND_CUDA_TEST)
+    char * buf_alloc_base = (char *)ggml_backend_buffer_get_base(lctx.buf_alloc);
     for (int i = 0; i < gf->n_leafs; i++) {
         ggml_tensor * node = gf->leafs[i];
         if (node->backend == GGML_BACKEND_GPU && node->extra == NULL) {
-            ggml_cuda_assign_scratch_offset(node, (char*)node->data - (char *) lctx.buf_alloc.data);
+            ggml_cuda_assign_scratch_offset(node, (char *)node->data - buf_alloc_base);
             ggml_cuda_copy_to_device(node);
         }
     }
@@ -5773,7 +6216,7 @@ static int llama_decode_internal(
     for (int i = 0; i < gf->n_nodes; i++) {
         ggml_tensor * node = gf->nodes[i];
         if (node->backend == GGML_BACKEND_GPU && node->extra == NULL) {
-            ggml_cuda_assign_scratch_offset(node, (char*)node->data - (char *) lctx.buf_alloc.data);
+            ggml_cuda_assign_scratch_offset(node, (char *)node->data - buf_alloc_base);
         }
     }
 
@@ -5795,28 +6238,28 @@ static int llama_decode_internal(
         n_threads = std::min(4, n_threads);
     }
 
-    const bool fully_offloaded = model.n_gpu_layers >= (int) hparams.n_layer + 3;
+    const bool fully_offloaded = model.n_gpu_layers >= (int) hparams.n_layer + 1;
     if (ggml_cpu_has_cublas() && fully_offloaded) {
         n_threads = 1;
     }
 
-#if GGML_USE_MPI
+#ifdef GGML_USE_MPI
     const int64_t n_layer = hparams.n_layer;
     ggml_mpi_graph_compute_pre(lctx.ctx_mpi, gf, n_layer);
 #endif
 
 #ifdef GGML_USE_METAL
-    if (lctx.ctx_metal) {
-        ggml_metal_set_n_cb     (lctx.ctx_metal, n_threads);
-        ggml_metal_graph_compute(lctx.ctx_metal, gf);
-    } else {
-        ggml_graph_compute_helper(lctx.work_buffer, gf, n_threads);
+    if (ggml_backend_is_metal(lctx.backend)) {
+        ggml_backend_metal_set_n_cb(lctx.backend, n_threads);
     }
-#else
-    ggml_graph_compute_helper(lctx.work_buffer, gf, n_threads);
 #endif
 
-#if GGML_USE_MPI
+    if (ggml_backend_is_cpu(lctx.backend)) {
+        ggml_backend_cpu_set_n_threads(lctx.backend, n_threads);
+    }
+    ggml_backend_graph_compute(lctx.backend, gf);
+
+#ifdef GGML_USE_MPI
     ggml_mpi_graph_compute_post(lctx.ctx_mpi, gf, n_layer);
 #endif
 
@@ -5854,20 +6297,37 @@ 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++) {
                 if (batch.logits[i] == 0) {
                     continue;
                 }
-                memcpy(logits_out.data() + (n_vocab*i), (float *) ggml_get_data(res) + (n_vocab*i), sizeof(float)*n_vocab);
+                ggml_backend_tensor_get(res, logits_out.data() + (n_vocab*i), (n_vocab*i)*sizeof(float), n_vocab*sizeof(float));
+#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);
+            ggml_backend_tensor_get(res, logits_out.data(), 0, n_vocab*n_tokens*sizeof(float));
+#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);
+            ggml_backend_tensor_get(res, logits_out.data(), (n_vocab*(n_tokens - 1))*sizeof(float), n_vocab*sizeof(float));
+#ifndef NDEBUG
+            logits_valid[0] = true;
+#endif
         }
     }
 
@@ -5876,7 +6336,7 @@ static int llama_decode_internal(
         auto & embedding_out = lctx.embedding;
 
         embedding_out.resize(n_embd);
-        memcpy(embedding_out.data(), (float *) ggml_get_data(embeddings) + (n_embd*(n_tokens - 1)), sizeof(float)*n_embd);
+        ggml_backend_tensor_get(embeddings, embedding_out.data(), (n_embd*(n_tokens - 1))*sizeof(float), n_embd*sizeof(float));
     }
 
     // measure the performance only for the single-token evals
@@ -6477,12 +6937,12 @@ static void tokenizer_st_partition(const llama_vocab & vocab, std::forward_list<
 
                 // loop over the text
                 while (true) {
-                    // find the first occurence of a given special token in this fragment
+                    // find the first occurrence of a given special token in this fragment
                     //  passing offset argument only limit the "search area" but match coordinates
                     //  are still relative to the source full raw_text
                     auto match = raw_text->find(special_token, raw_text_base_offset);
 
-                    // no occurences found, stop processing this fragment for a given special token
+                    // no occurrences found, stop processing this fragment for a given special token
                     if (match == std::string::npos) break;
 
                     // check if match is within bounds of offset <-> length
@@ -6654,14 +7114,13 @@ struct llama_grammar_candidate {
 // Decodes a UTF-8 string which may end in an incomplete sequence. Adds a terminating 0 for use as
 // pointer. If an invalid sequence is encountered, returns `llama_partial_utf8.n_remain == -1`.
 static std::pair<std::vector<uint32_t>, llama_partial_utf8> decode_utf8(
-        const char         * src,
-        size_t               n_src,
+        const std::string & src,
         llama_partial_utf8   partial_start) {
     static const int      lookup[] = { 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 2, 2, 3, 4 };
-    const char          * pos      = src;
+    const char          * pos      = src.c_str();
     std::vector<uint32_t> code_points;
     // common english strings have the same number of codepoints and bytes. `+ 1` for the terminating 0.
-    code_points.reserve(n_src + 1);
+    code_points.reserve(src.size() + 1);
     uint32_t              value    = partial_start.value;
     int                   n_remain = partial_start.n_remain;
 
@@ -6712,13 +7171,6 @@ static std::pair<std::vector<uint32_t>, llama_partial_utf8> decode_utf8(
     return std::make_pair(std::move(code_points), llama_partial_utf8{ value, n_remain });
 }
 
-static std::pair<std::vector<uint32_t>, llama_partial_utf8> decode_utf8(
-        std::string src,
-        llama_partial_utf8 partial_start
-) {
-    return decode_utf8(src.c_str(), src.size(), partial_start);
-}
-
 // returns true iff pos points to the end of one of the definitions of a rule
 static bool llama_grammar_is_end_of_sequence(const llama_grammar_element * pos) {
     switch (pos->type) {
@@ -7357,7 +7809,9 @@ void llama_sample_grammar(struct llama_context * ctx, llama_token_data_array * c
     const llama_token eos = llama_token_eos(&ctx->model);
 
     std::vector<std::pair<std::vector<uint32_t>, llama_partial_utf8>> candidates_decoded;
+    candidates_decoded.reserve(candidates->size);
     std::vector<llama_grammar_candidate>                              candidates_grammar;
+    candidates_grammar.reserve(candidates->size);
 
     for (size_t i = 0; i < candidates->size; ++i) {
         const llama_token id    = candidates->data[i].id;
@@ -7687,7 +8141,7 @@ struct llama_beam_search_data {
     }
 
     // Min-heaps are used to efficiently collect the top-k elements (k=n_beams).
-    // The repetative patterns below reflect the 2 stages of heaps:
+    // The repetitive patterns below reflect the 2 stages of heaps:
     //  * Gather elements until the vector is full, then call std::make_heap() on it.
     //  * If the heap is full and a new element is found that should be included, pop the
     //    least element to the back(), replace it with the new, then push it into the heap.
@@ -7840,12 +8294,6 @@ void llama_beam_search(llama_context * ctx,
 // quantization
 //
 
-template <typename T>
-struct no_init {
-    T value;
-    no_init() { /* do nothing */ }
-};
-
 struct quantize_state_internal {
     const llama_model                 & model;
     const llama_model_quantize_params * params;
@@ -7925,11 +8373,9 @@ static void llama_convert_tensor_internal(
     workers.clear();
 }
 
-static ggml_type get_k_quant_type(
-    quantize_state_internal & qs,
-    ggml_type new_type, const ggml_tensor * tensor, llama_ftype ftype
-) {
+static ggml_type get_k_quant_type(quantize_state_internal & qs, ggml_type new_type, const ggml_tensor * tensor, llama_ftype ftype) {
     const std::string name = ggml_get_name(tensor);
+
     // TODO: avoid hardcoded tensor names - use the TN_* constants
     const llm_arch arch = qs.model.arch;
     const auto       tn = LLM_TN(arch);
@@ -7963,7 +8409,18 @@ static ggml_type get_k_quant_type(
             // nearly negligible increase in model size by quantizing this tensor with more bits:
             if (new_type == GGML_TYPE_Q3_K || new_type == GGML_TYPE_Q4_K) new_type = GGML_TYPE_Q5_K;
         }
+        if (qs.model.hparams.n_expert == 8) {
+            // for the 8-expert model, bumping this to Q8_0 trades just ~128MB
+            // TODO: explore better strategies
+            new_type = GGML_TYPE_Q8_0;
+        }
         ++qs.i_attention_wv;
+    } else if (name.find("attn_k.weight") != std::string::npos) {
+        if (qs.model.hparams.n_expert == 8) {
+            // for the 8-expert model, bumping this to Q8_0 trades just ~128MB
+            // TODO: explore better strategies
+            new_type = GGML_TYPE_Q8_0;
+        }
     } else if (name.find("ffn_down.weight") != std::string::npos) {
         if      (ftype == LLAMA_FTYPE_MOSTLY_Q2_K) new_type = GGML_TYPE_Q3_K;
         else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M) {
@@ -8078,10 +8535,8 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
     constexpr bool use_mmap = false;
 #endif
 
-    llama_model_loader ml(fname_inp, use_mmap);
-    if (ml.use_mmap) {
-        ml.mapping.reset(new llama_mmap(&ml.file, /* prefetch */ 0, ggml_is_numa()));
-    }
+    llama_model_loader ml(fname_inp, use_mmap, NULL);
+    ml.init_mapping(false); // no prefetching?
 
     llama_model model;
     llm_load_arch(ml, model);
@@ -8172,10 +8627,13 @@ 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;
 
+        // do not quantize expert gating tensors
+        quantize &= name.find("ffn_gate_inp.weight") == std::string::npos;
+
         enum ggml_type new_type;
         void * new_data;
         size_t new_size;
@@ -8324,74 +8782,63 @@ 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;
-    std::vector<uint8_t> base_buf;
-    if (path_base_model) {
+
+   if (path_base_model) {
         LLAMA_LOG_INFO("%s: loading base model from '%s'\n", __func__, path_base_model);
-        ml.reset(new llama_model_loader(path_base_model, /*use_mmap*/ true));
-
-        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;
-        base_params.mem_size   = base_buf.size();
-        base_params.mem_buffer = base_buf.data();
-        base_params.no_alloc   = ml->use_mmap;
-
-        base_ctx = ggml_init(base_params);
-
-        // maybe this should in llama_model_loader
-        if (ml->use_mmap) {
-            ml->mapping.reset(new llama_mmap(&ml->file, /* prefetch */ 0, ggml_is_numa()));
-        }
+        ml.reset(new llama_model_loader(path_base_model, /*use_mmap*/ true, /*kv_overrides*/ nullptr));
+        ml->init_mapping(false); // no prefetching
     }
 
     // read tensors and apply
@@ -8401,27 +8848,35 @@ 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
@@ -8435,7 +8890,7 @@ 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());
@@ -8454,22 +8909,15 @@ 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;
 
@@ -8482,7 +8930,7 @@ static int llama_apply_lora_from_file_internal(
             offload_func_t offload_func               = ggml_offload_nop;
             offload_func_t offload_func_force_inplace = ggml_offload_nop;
 
-#ifdef GGML_USE_CUBLAS
+#if defined(GGML_USE_CUBLAS) && !defined(LLAMA_GGML_BACKEND_CUDA_TEST)
             if (dest_t->backend == GGML_BACKEND_GPU || dest_t->backend == GGML_BACKEND_GPU_SPLIT) {
                 if (dest_t->type != GGML_TYPE_F16) {
                     throw std::runtime_error(format(
@@ -8499,13 +8947,11 @@ 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->get_tensor_meta(base_name.c_str());
                 ml->load_data_for(base_t);
             } else {
                 base_t = dest_t;
@@ -8534,43 +8980,42 @@ 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_set_name(scale_tensor, "scale_tensor");
-
-                BA = ggml_scale_inplace(lora_ctx, BA, scale_tensor);
+                BA = ggml_scale_inplace(lora_ctx.get(), BA, scaling);
                 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++;
@@ -8580,12 +9025,6 @@ 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);
 
@@ -8602,6 +9041,7 @@ struct llama_model_params llama_model_default_params() {
         /*.tensor_split                =*/ nullptr,
         /*.progress_callback           =*/ nullptr,
         /*.progress_callback_user_data =*/ nullptr,
+        /*.kv_overrides                =*/ nullptr,
         /*.vocab_only                  =*/ false,
         /*.use_mmap                    =*/ true,
         /*.use_mlock                   =*/ false,
@@ -8629,10 +9069,12 @@ struct llama_context_params llama_context_default_params() {
         /*.yarn_beta_fast              =*/ 32.0f,
         /*.yarn_beta_slow              =*/ 1.0f,
         /*.yarn_orig_ctx               =*/ 0,
+        /*.type_k                      =*/ GGML_TYPE_F16,
+        /*.type_v                      =*/ GGML_TYPE_F16,
         /*.mul_mat_q                   =*/ true,
-        /*.f16_kv                      =*/ true,
         /*.logits_all                  =*/ false,
         /*.embedding                   =*/ false,
+        /*.offload_kqv                 =*/ true,
     };
 
     return result;
@@ -8712,11 +9154,18 @@ struct llama_model * llama_load_model_from_file(
                     LLAMA_LOG_INFO("\n");
                 }
             }
+            return true;
         };
     }
 
-    if (!llama_model_load(path_model, *model, params)) {
-        LLAMA_LOG_ERROR("%s: failed to load model\n", __func__);
+    int status = llama_model_load(path_model, *model, params);
+    GGML_ASSERT(status <= 0);
+    if (status < 0) {
+        if (status == -1) {
+            LLAMA_LOG_ERROR("%s: failed to load model\n", __func__);
+        } else if (status == -2) {
+            LLAMA_LOG_INFO("%s: cancelled model load\n", __func__);
+        }
         delete model;
         return nullptr;
     }
@@ -8749,6 +9198,7 @@ struct llama_context * llama_new_context_with_model(
     cparams.yarn_beta_fast   = params.yarn_beta_fast;
     cparams.yarn_beta_slow   = params.yarn_beta_slow;
     cparams.mul_mat_q        = params.mul_mat_q;
+    cparams.offload_kqv      = params.offload_kqv;
 
     cparams.n_ctx            = params.n_ctx           == 0    ? hparams.n_ctx_train           : params.n_ctx;
     cparams.rope_freq_base   = params.rope_freq_base  == 0.0f ? hparams.rope_freq_base_train  : params.rope_freq_base;
@@ -8782,19 +9232,68 @@ struct llama_context * llama_new_context_with_model(
     ctx->rng = std::mt19937(params.seed);
     ctx->logits_all = params.logits_all;
 
-    ggml_type memory_type = params.f16_kv ? GGML_TYPE_F16 : GGML_TYPE_F32;
+    const ggml_type type_k = params.type_k;
+    const ggml_type type_v = params.type_v;
+
+    GGML_ASSERT(hparams.n_embd_head() % ggml_blck_size(type_k) == 0);
+    GGML_ASSERT(hparams.n_embd_head() % ggml_blck_size(type_v) == 0);
 
     // reserve memory for context buffers
     if (!hparams.vocab_only) {
-        if (!llama_kv_cache_init(ctx->model.hparams, ctx->kv_self, memory_type, cparams.n_ctx, model->n_gpu_layers)) {
+        // initialize backend
+#ifdef GGML_USE_METAL
+        if (model->n_gpu_layers > 0) {
+            ctx->backend = ggml_backend_metal_init();
+            if (ctx->backend == nullptr) {
+                LLAMA_LOG_ERROR("%s: failed to initialize Metal backend\n", __func__);
+            }
+        }
+#elif defined(GGML_USE_CUBLAS) && defined(LLAMA_GGML_BACKEND_CUDA_TEST)
+        // for testing only
+        if (model->n_gpu_layers > 0) {
+            ctx->backend = ggml_backend_cuda_init(0);
+            if (ctx->backend == nullptr) {
+                LLAMA_LOG_ERROR("%s: failed to initialize CUDA backend\n", __func__);
+            }
+        }
+#endif
+
+        if (ctx->backend == nullptr && ggml_backend_buffer_is_host(model->buf)) {
+            ctx->backend = ggml_backend_cpu_init();
+            if (ctx->backend == nullptr) {
+                LLAMA_LOG_ERROR("%s: failed to initialize CPU backend\n", __func__);
+            }
+        }
+
+        if (ctx->backend == nullptr) {
+            LLAMA_LOG_ERROR("%s: failed to initialize a backend\n", __func__);
+            delete ctx;
+            return nullptr;
+        }
+
+        if (!llama_kv_cache_init(ctx->model.hparams, ctx->kv_self, type_k, type_v,
+                cparams.n_ctx, model->n_gpu_layers, cparams.offload_kqv)) {
             LLAMA_LOG_ERROR("%s: llama_kv_cache_init() failed for self-attention cache\n", __func__);
             llama_free(ctx);
             return nullptr;
         }
 
         {
-            const size_t memory_size = ggml_nbytes(ctx->kv_self.k) + ggml_nbytes(ctx->kv_self.v);
-            LLAMA_LOG_INFO("%s: kv self size  = %7.2f MiB\n", __func__, memory_size / 1024.0 / 1024.0);
+            size_t memory_size_k = 0;
+            size_t memory_size_v = 0;
+
+            for (auto & k : ctx->kv_self.k_l) {
+                memory_size_k += ggml_nbytes(k);
+            }
+
+            for (auto & v : ctx->kv_self.v_l) {
+                memory_size_v += ggml_nbytes(v);
+            }
+
+            LLAMA_LOG_INFO("%s: KV self size  = %7.2f MiB, K (%s): %7.2f MiB, V (%s): %7.2f MiB\n", __func__,
+                (float)(memory_size_k + memory_size_v) / (1024.0f * 1024.0f),
+                ggml_type_name(type_k), (float)memory_size_k / (1024.0f * 1024.0f),
+                ggml_type_name(type_v), (float)memory_size_v / (1024.0f * 1024.0f));
         }
 
         // resized during inference
@@ -8809,12 +9308,11 @@ struct llama_context * llama_new_context_with_model(
         }
 
         {
-            static const size_t tensor_alignment = 32;
             // the compute buffer is used to store the tensor and graph structs, while the allocator buffer is used for the tensor data
-            ctx->buf_compute.resize(ggml_tensor_overhead()*LLAMA_MAX_NODES + ggml_graph_overhead());
+            ctx->buf_compute_meta.resize(ggml_tensor_overhead()*LLAMA_MAX_NODES + ggml_graph_overhead());
 
             // create measure allocator
-            ctx->alloc = ggml_allocr_new_measure(tensor_alignment);
+            ctx->alloc = ggml_allocr_new_measure_from_backend(ctx->backend);
 
             // build worst-case graph
             int n_tokens = (int)std::min(cparams.n_ctx, cparams.n_batch);
@@ -8822,94 +9320,50 @@ struct llama_context * llama_new_context_with_model(
             llama_token token = llama_token_bos(&ctx->model); // not actually used by llama_build_graph, but required to choose between token and embedding inputs graph
             ggml_cgraph * gf = llama_build_graph(*ctx, llama_batch_get_one(&token, n_tokens, n_past, 0));
 
-#ifdef GGML_USE_METAL
-            if (model->n_gpu_layers > 0) {
-                ctx->ctx_metal = ggml_metal_init(1);
-                if (!ctx->ctx_metal) {
-                    LLAMA_LOG_ERROR("%s: ggml_metal_init() failed\n", __func__);
-                    llama_free(ctx);
-                    return NULL;
-                }
-                //ggml_metal_graph_find_concurrency(ctx->ctx_metal, gf, false);
-                //ggml_allocr_set_parse_seq(ctx->alloc, ggml_metal_get_concur_list(ctx->ctx_metal), ggml_metal_if_optimized(ctx->ctx_metal));
-            }
-#endif
             // measure memory requirements for the graph
-            size_t alloc_size = ggml_allocr_alloc_graph(ctx->alloc, gf) + tensor_alignment;
+            size_t alloc_size = ggml_allocr_alloc_graph(ctx->alloc, gf);
 
-            LLAMA_LOG_INFO("%s: compute buffer total size = %.2f MiB\n", __func__, (ctx->buf_compute.size + alloc_size) / 1024.0 / 1024.0);
+            LLAMA_LOG_INFO("%s: compute buffer total size = %.2f MiB\n", __func__, (ctx->buf_compute_meta.size() + alloc_size) / 1024.0 / 1024.0);
 
-            // recreate allocator with exact memory requirements
+            // create allocator again with exact memory requirements
             ggml_allocr_free(ctx->alloc);
 
-            ctx->buf_alloc.resize(alloc_size);
-            ctx->alloc = ggml_allocr_new(ctx->buf_alloc.data, ctx->buf_alloc.size, tensor_alignment);
-#ifdef GGML_USE_METAL
-            if (ctx->ctx_metal) {
-                //ggml_allocr_set_parse_seq(ctx->alloc, ggml_metal_get_concur_list(ctx->ctx_metal), ggml_metal_if_optimized(ctx->ctx_metal));
-            }
-#endif
-#ifdef GGML_USE_CUBLAS
-            ggml_cuda_set_scratch_size(alloc_size);
-            LLAMA_LOG_INFO("%s: VRAM scratch buffer: %.2f MiB\n", __func__, alloc_size / 1024.0 / 1024.0);
+            ctx->buf_alloc = ggml_backend_alloc_buffer(ctx->backend, alloc_size);
+            ctx->alloc = ggml_allocr_new_from_buffer(ctx->buf_alloc);
+#if defined(GGML_USE_CUBLAS) && !defined(LLAMA_GGML_BACKEND_CUDA_TEST)
+            if (model->n_gpu_layers > 0) {
+                ggml_cuda_set_scratch_size(alloc_size);
+                LLAMA_LOG_INFO("%s: VRAM scratch buffer: %.2f MiB\n", __func__, alloc_size / 1024.0 / 1024.0);
 
-            // calculate total VRAM usage
-            auto add_tensor = [](const ggml_tensor * t, size_t & size) {
-                if (t->backend == GGML_BACKEND_GPU || t->backend == GGML_BACKEND_GPU_SPLIT) {
-                    size += ggml_nbytes(t);
+                // calculate total VRAM usage
+                auto add_tensor = [](const ggml_tensor * t, size_t & size) {
+                    if (t->backend == GGML_BACKEND_GPU || t->backend == GGML_BACKEND_GPU_SPLIT) {
+                        size += ggml_nbytes(t);
+                    }
+                };
+                size_t model_vram_size = 0;
+                for (const auto & kv : model->tensors_by_name) {
+                    add_tensor(kv.second, model_vram_size);
                 }
-            };
-            size_t model_vram_size = 0;
-            for (const auto & kv : model->tensors_by_name) {
-                add_tensor(kv.second, model_vram_size);
+
+                size_t kv_vram_size = 0;
+                for (auto & k : ctx->kv_self.k_l) {
+                    add_tensor(k, kv_vram_size);
+                }
+                for (auto & v : ctx->kv_self.v_l) {
+                    add_tensor(v, kv_vram_size);
+                }
+
+                size_t ctx_vram_size = alloc_size + kv_vram_size;
+                size_t total_vram_size = model_vram_size + ctx_vram_size;
+
+                LLAMA_LOG_INFO("%s: total VRAM used: %.2f MiB (model: %.2f MiB, context: %.2f MiB)\n", __func__,
+                        total_vram_size / 1024.0 / 1024.0,
+                        model_vram_size / 1024.0 / 1024.0,
+                        ctx_vram_size   / 1024.0 / 1024.0);
             }
-
-            size_t kv_vram_size = 0;
-            add_tensor(ctx->kv_self.k, kv_vram_size);
-            add_tensor(ctx->kv_self.v, kv_vram_size);
-
-            size_t ctx_vram_size = alloc_size + kv_vram_size;
-            size_t total_vram_size = model_vram_size + ctx_vram_size;
-
-            LLAMA_LOG_INFO("%s: total VRAM used: %.2f MiB (model: %.2f MiB, context: %.2f MiB)\n", __func__,
-                    total_vram_size / 1024.0 / 1024.0,
-                    model_vram_size / 1024.0 / 1024.0,
-                    ctx_vram_size   / 1024.0 / 1024.0);
 #endif
         }
-
-#ifdef GGML_USE_METAL
-        if (model->n_gpu_layers > 0) {
-            // this allocates all Metal resources and memory buffers
-
-            void * data_ptr  = NULL;
-            size_t data_size = 0;
-
-            if (ctx->model.mapping) {
-                data_ptr  = ctx->model.mapping->addr;
-                data_size = ctx->model.mapping->size;
-            } else {
-                data_ptr  = ggml_get_mem_buffer(ctx->model.ctx);
-                data_size = ggml_get_mem_size  (ctx->model.ctx);
-            }
-
-            const size_t max_size = ggml_get_max_tensor_size(ctx->model.ctx);
-
-            LLAMA_LOG_INFO("%s: max tensor size = %8.2f MiB\n", __func__, max_size/1024.0/1024.0);
-
-#define LLAMA_METAL_CHECK_BUF(result)                            \
-            if (!(result)) {                                             \
-                LLAMA_LOG_ERROR("%s: failed to add buffer\n", __func__); \
-                llama_free(ctx);                                         \
-                return NULL;                                             \
-            }
-
-            LLAMA_METAL_CHECK_BUF(ggml_metal_add_buffer(ctx->ctx_metal, "data",  data_ptr, data_size, max_size));
-            LLAMA_METAL_CHECK_BUF(ggml_metal_add_buffer(ctx->ctx_metal, "kv",    ctx->kv_self.buf.data, ctx->kv_self.buf.size, 0));
-            LLAMA_METAL_CHECK_BUF(ggml_metal_add_buffer(ctx->ctx_metal, "alloc", ctx->buf_alloc.data, ctx->buf_alloc.size, 0));
-#undef LLAMA_METAL_CHECK_BUF
-        }
-#endif
     }
 
 #ifdef GGML_USE_MPI
@@ -8937,10 +9391,14 @@ 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;
 }
@@ -9197,7 +9655,7 @@ size_t llama_get_state_size(const struct llama_context * ctx) {
     const size_t s_embedding       = ctx->embedding.size() * sizeof(float);
     const size_t s_kv_size         = sizeof(size_t);
     const size_t s_kv_ntok         = sizeof(int);
-    const size_t s_kv              = ctx->kv_self.buf.size;
+    const size_t s_kv              = ggml_backend_buffer_get_size(ctx->kv_self.buf);
 
     const size_t s_total = (
         + s_rng_size
@@ -9325,7 +9783,7 @@ static void llama_copy_state_data_internal(struct llama_context * ctx, llama_dat
         const auto   n_embd  = hparams.n_embd_gqa();
         const auto   n_ctx   = cparams.n_ctx;
 
-        const size_t   kv_buf_size = kv_self.buf.size;
+        const size_t   kv_buf_size = ggml_backend_buffer_get_size(kv_self.buf);
         const uint32_t kv_head     = kv_self.head;
         const uint32_t kv_size     = kv_self.size;
         const uint32_t kv_used     = kv_self.used;
@@ -9336,37 +9794,48 @@ static void llama_copy_state_data_internal(struct llama_context * ctx, llama_dat
         data_ctx->write(&kv_used,     sizeof(kv_used));
 
         if (kv_buf_size) {
-            const size_t elt_size = ggml_element_size(kv_self.k);
+            const size_t elt_size = ggml_element_size(kv_self.k_l[0]);
 
-            ggml_context * cpy_ctx = ggml_init({ 6*ggml_tensor_overhead() + ggml_graph_overhead(), NULL, /* no_alloc */ true });
+            ggml_context * cpy_ctx = ggml_init({ 6*n_layer*ggml_tensor_overhead() + ggml_graph_overhead(), NULL, /* no_alloc */ true });
             ggml_cgraph * gf = ggml_new_graph(cpy_ctx);
 
-            ggml_tensor * kout3d = ggml_new_tensor_3d(cpy_ctx, kv_self.k->type, n_embd, kv_head, n_layer);
-            std::vector<uint8_t> kout3d_data(ggml_nbytes(kout3d), 0);
-            kout3d->data = kout3d_data.data();
+            std::vector<struct ggml_tensor *> kout2d(n_layer);
+            std::vector<struct ggml_tensor *> vout2d(n_layer);
 
-            ggml_tensor * vout3d = ggml_new_tensor_3d(cpy_ctx, kv_self.v->type, kv_head, n_embd, n_layer);
-            std::vector<uint8_t> vout3d_data(ggml_nbytes(vout3d), 0);
-            vout3d->data = vout3d_data.data();
+            for (int il = 0; il < (int) n_layer; ++il) {
+                kout2d[il] = ggml_new_tensor_2d(cpy_ctx, kv_self.k_l[il]->type, n_embd, kv_head);
+                vout2d[il] = ggml_new_tensor_2d(cpy_ctx, kv_self.v_l[il]->type, kv_head, n_embd);
 
-            ggml_tensor * k3d = ggml_view_3d(cpy_ctx, kv_self.k,
-                n_embd, kv_head, n_layer,
-                elt_size*n_embd, elt_size*n_embd*n_ctx, 0);
+                ggml_tensor * k2d = ggml_view_2d(cpy_ctx, kv_self.k_l[il],
+                        n_embd, kv_head,
+                        elt_size*n_embd, 0);
 
-            ggml_tensor * v3d = ggml_view_3d(cpy_ctx, kv_self.v,
-                kv_head, n_embd, n_layer,
-                elt_size*n_ctx, elt_size*n_ctx*n_embd, 0);
+                ggml_tensor * v2d = ggml_view_2d(cpy_ctx, kv_self.v_l[il],
+                        kv_head, n_embd,
+                        elt_size*n_ctx, 0);
 
-            ggml_build_forward_expand(gf, ggml_cpy(cpy_ctx, k3d, kout3d));
-            ggml_build_forward_expand(gf, ggml_cpy(cpy_ctx, v3d, vout3d));
-            ggml_graph_compute_helper(ctx->work_buffer, gf, /*n_threads*/ 1);
+                ggml_build_forward_expand(gf, ggml_cpy(cpy_ctx, k2d, kout2d[il]));
+                ggml_build_forward_expand(gf, ggml_cpy(cpy_ctx, v2d, vout2d[il]));
+            }
+
+            ggml_backend_buffer_t buf = ggml_backend_alloc_ctx_tensors(cpy_ctx, ctx->backend);
+
+            ggml_backend_graph_compute(ctx->backend, gf);
+
+            std::vector<uint8_t> tmp_buf;
+            for (int il = 0; il < (int) n_layer; ++il) {
+                tmp_buf.resize(ggml_nbytes(kout2d[il]));
+                ggml_backend_tensor_get(kout2d[il], tmp_buf.data(), 0, tmp_buf.size());
+                data_ctx->write(tmp_buf.data(), tmp_buf.size());
+
+                tmp_buf.resize(ggml_nbytes(vout2d[il]));
+                ggml_backend_tensor_get(vout2d[il], tmp_buf.data(), 0, tmp_buf.size());
+                data_ctx->write(tmp_buf.data(), tmp_buf.size());
+            }
 
             ggml_free(cpy_ctx);
 
-            // our data is now in the kout3d_data and vout3d_data buffers
-            // write them to file
-            data_ctx->write(kout3d_data.data(), kout3d_data.size());
-            data_ctx->write(vout3d_data.data(), vout3d_data.size());
+            ggml_backend_buffer_free(buf);
         }
 
         for (uint32_t i = 0; i < kv_size; ++i) {
@@ -9464,34 +9933,48 @@ size_t llama_set_state_data(struct llama_context * ctx, uint8_t * src) {
         memcpy(&kv_used,     inp, sizeof(kv_used));     inp += sizeof(kv_used);
 
         if (kv_buf_size) {
-            GGML_ASSERT(kv_self.buf.size == kv_buf_size);
+            GGML_ASSERT(ggml_backend_buffer_get_size(kv_self.buf) == kv_buf_size);
 
-            const size_t elt_size = ggml_element_size(kv_self.k);
+            const size_t elt_size = ggml_element_size(kv_self.k_l[0]);
 
-            ggml_context * cpy_ctx = ggml_init({ 6*ggml_tensor_overhead() + ggml_graph_overhead(), NULL, /* no_alloc */ true });
+            ggml_context * cpy_ctx = ggml_init({ 6*n_layer*ggml_tensor_overhead() + ggml_graph_overhead(), NULL, /* no_alloc */ true });
             ggml_cgraph * gf = ggml_new_graph(cpy_ctx);
 
-            ggml_tensor * kin3d = ggml_new_tensor_3d(cpy_ctx, kv_self.k->type, n_embd, kv_head, n_layer);
-            kin3d->data = (void *) inp;
-            inp += ggml_nbytes(kin3d);
+            std::vector<struct ggml_tensor *> kin2d(n_layer);
+            std::vector<struct ggml_tensor *> vin2d(n_layer);
 
-            ggml_tensor * vin3d = ggml_new_tensor_3d(cpy_ctx, kv_self.v->type, kv_head, n_embd, n_layer);
-            vin3d->data = (void *) inp;
-            inp += ggml_nbytes(vin3d);
+            for (int il = 0; il < n_layer; ++il) {
+                kin2d[il] = ggml_new_tensor_2d(cpy_ctx, kv_self.k_l[il]->type, n_embd, kv_head);
+                vin2d[il] = ggml_new_tensor_2d(cpy_ctx, kv_self.v_l[il]->type, kv_head, n_embd);
 
-            ggml_tensor * k3d = ggml_view_3d(cpy_ctx, kv_self.k,
-                n_embd, kv_head, n_layer,
-                elt_size*n_embd, elt_size*n_embd*n_ctx, 0);
+                ggml_tensor * k2d = ggml_view_2d(cpy_ctx, kv_self.k_l[il],
+                    n_embd, kv_head,
+                    elt_size*n_embd, 0);
 
-            ggml_tensor * v3d = ggml_view_3d(cpy_ctx, kv_self.v,
-                kv_head, n_embd, n_layer,
-                elt_size*n_ctx, elt_size*n_ctx*n_embd, 0);
+                ggml_tensor * v2d = ggml_view_2d(cpy_ctx, kv_self.v_l[il],
+                    kv_head, n_embd,
+                    elt_size*n_ctx, 0);
 
-            ggml_build_forward_expand(gf, ggml_cpy(cpy_ctx, kin3d, k3d));
-            ggml_build_forward_expand(gf, ggml_cpy(cpy_ctx, vin3d, v3d));
-            ggml_graph_compute_helper(ctx->work_buffer, gf, /*n_threads*/ 1);
+                ggml_build_forward_expand(gf, ggml_cpy(cpy_ctx, kin2d[il], k2d));
+                ggml_build_forward_expand(gf, ggml_cpy(cpy_ctx, vin2d[il], v2d));
+            }
+
+            ggml_backend_buffer_t buf = ggml_backend_alloc_ctx_tensors(cpy_ctx, ctx->backend);
+
+            // load data into the tensors
+            for (int il = 0; il < n_layer; ++il) {
+                ggml_backend_tensor_set(kin2d[il], inp, 0, ggml_nbytes(kin2d[il]));
+                inp += ggml_nbytes(kin2d[il]);
+
+                ggml_backend_tensor_set(vin2d[il], inp, 0, ggml_nbytes(vin2d[il]));
+                inp += ggml_nbytes(vin2d[il]);
+            }
+
+            ggml_backend_graph_compute(ctx->backend, gf);
 
             ggml_free(cpy_ctx);
+
+            ggml_backend_buffer_free(buf);
         }
 
         ctx->kv_self.head = kv_head;
@@ -9714,6 +10197,7 @@ 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;
 }
 
diff --git a/llama.h b/llama.h
index 89cb6198e..af76bae2d 100644
--- a/llama.h
+++ b/llama.h
@@ -39,10 +39,11 @@
 
 #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
-#define LLAMA_SESSION_VERSION 2
+#define LLAMA_SESSION_VERSION 3
 
 #if defined(GGML_USE_CUBLAS) || defined(GGML_USE_CLBLAST) || defined(GGML_USE_METAL)
 // Defined when llama.cpp is compiled with support for offloading model layers to GPU.
@@ -126,7 +127,7 @@ extern "C" {
         bool sorted;
     } llama_token_data_array;
 
-    typedef void (*llama_progress_callback)(float progress, void *ctx);
+    typedef bool (*llama_progress_callback)(float progress, void *ctx);
 
     // Input data for llama_decode
     // A llama_batch object can contain input about one or many sequences
@@ -158,16 +159,38 @@ extern "C" {
         llama_seq_id all_seq_id; // used if seq_id == NULL
     } llama_batch;
 
+    enum llama_model_kv_override_type {
+        LLAMA_KV_OVERRIDE_INT,
+        LLAMA_KV_OVERRIDE_FLOAT,
+        LLAMA_KV_OVERRIDE_BOOL,
+    };
+
+    struct llama_model_kv_override {
+        char key[128];
+        enum llama_model_kv_override_type tag;
+        union {
+            int64_t int_value;
+            double float_value;
+            bool bool_value;
+        };
+    };
+
     struct llama_model_params {
         int32_t n_gpu_layers; // number of layers to store in VRAM
         int32_t main_gpu;     // the GPU that is used for scratch and small tensors
         const float * tensor_split; // how to split layers across multiple GPUs (size: LLAMA_MAX_DEVICES)
 
-        // called with a progress value between 0 and 1, pass NULL to disable
+        // Called with a progress value between 0.0 and 1.0. Pass NULL to disable.
+        // If the provided progress_callback returns true, model loading continues.
+        // If it returns false, model loading is immediately aborted.
         llama_progress_callback progress_callback;
+
         // context pointer passed to the progress callback
         void * progress_callback_user_data;
 
+        // override key-value pairs of the model meta data
+        const struct llama_model_kv_override * kv_overrides;
+
         // Keep the booleans together to avoid misalignment during copy-by-value.
         bool vocab_only; // only load the vocabulary, no weights
         bool use_mmap;   // use mmap if possible
@@ -191,11 +214,14 @@ extern "C" {
         float    yarn_beta_slow;   // YaRN high correction dim
         uint32_t yarn_orig_ctx;    // YaRN original context size
 
+        enum ggml_type type_k; // data type for K cache
+        enum ggml_type type_v; // data type for V cache
+
         // Keep the booleans together to avoid misalignment during copy-by-value.
-        bool mul_mat_q;  // if true, use experimental mul_mat_q kernels (DEPRECATED - always true)
-        bool f16_kv;     // use fp16 for KV cache, fp32 otherwise
-        bool logits_all; // the llama_eval() call computes all logits, not just the last one
-        bool embedding;  // embedding mode only
+        bool mul_mat_q;   // if true, use experimental mul_mat_q kernels (DEPRECATED - always true)
+        bool logits_all;  // the llama_eval() call computes all logits, not just the last one (DEPRECATED - set llama_batch.logits instead)
+        bool embedding;   // embedding mode only
+        bool offload_kqv; // whether to offload the KQV ops (including the KV cache) to GPU
     };
 
     // model quantization parameters
@@ -290,7 +316,9 @@ 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);
 
diff --git a/requirements.txt b/requirements.txt
index 81c909d0b..1a1162566 100644
--- a/requirements.txt
+++ b/requirements.txt
@@ -1,3 +1,5 @@
 numpy==1.24.4
 sentencepiece==0.1.98
+transformers>=4.34.0
 gguf>=0.1.0
+protobuf>=4.21.0
diff --git a/scripts/get-flags.mk b/scripts/get-flags.mk
new file mode 100644
index 000000000..596d7ead1
--- /dev/null
+++ b/scripts/get-flags.mk
@@ -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
diff --git a/scripts/sync-ggml.sh b/scripts/sync-ggml.sh
index 4024531b1..0097db435 100755
--- a/scripts/sync-ggml.sh
+++ b/scripts/sync-ggml.sh
@@ -20,5 +20,6 @@ cp -rpv ../ggml/include/ggml/ggml.h         ./ggml.h
 cp -rpv ../ggml/include/ggml/ggml-alloc.h   ./ggml-alloc.h
 cp -rpv ../ggml/include/ggml/ggml-backend.h ./ggml-backend.h
 
-cp -rpv ../ggml/tests/test-opt.cpp    ./tests/test-opt.cpp
-cp -rpv ../ggml/tests/test-grad0.cpp  ./tests/test-grad0.cpp
+cp -rpv ../ggml/tests/test-opt.cpp         ./tests/test-opt.cpp
+cp -rpv ../ggml/tests/test-grad0.cpp       ./tests/test-grad0.cpp
+cp -rpv ../ggml/tests/test-backend-ops.cpp ./tests/test-backend-ops.cpp
diff --git a/tests/CMakeLists.txt b/tests/CMakeLists.txt
index c8b4bc254..e42237c7a 100644
--- a/tests/CMakeLists.txt
+++ b/tests/CMakeLists.txt
@@ -22,26 +22,32 @@ endfunction()
 llama_build_and_test_executable(test-quantize-fns.cpp)
 llama_build_and_test_executable(test-quantize-perf.cpp)
 llama_build_and_test_executable(test-sampling.cpp)
+
 llama_build_executable(test-tokenizer-0-llama.cpp)
 llama_test_executable (test-tokenizer-0-llama test-tokenizer-0-llama.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-llama.gguf)
+
 llama_build_executable(test-tokenizer-0-falcon.cpp)
 llama_test_executable (test-tokenizer-0-falcon test-tokenizer-0-falcon.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-falcon.gguf)
+
 llama_build_executable(test-tokenizer-1-llama.cpp)
-llama_test_executable (test-tokenizer-1-llama test-tokenizer-1-llama.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-llama.gguf)
-llama_test_executable(test-tokenizer-1-baichuan test-tokenizer-1-llama.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-baichuan.gguf)
+llama_test_executable (test-tokenizer-1-llama    test-tokenizer-1-llama.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-llama.gguf)
+llama_test_executable (test-tokenizer-1-baichuan test-tokenizer-1-llama.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-baichuan.gguf)
+
 llama_build_executable(test-tokenizer-1-bpe.cpp)
-llama_test_executable (test-tokenizer-1-falcon test-tokenizer-1-bpe.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-falcon.gguf)
-llama_test_executable(test-tokenizer-1-aquila test-tokenizer-1-bpe.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-aquila.gguf)
-llama_test_executable(test-tokenizer-1-mpt test-tokenizer-1-bpe.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-mpt.gguf)
-llama_test_executable(test-tokenizer-1-stablelm-3b-4e1t test-tokenizer-1-bpe.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-stablelm-3b-4e1t.gguf)
-llama_test_executable(test-tokenizer-1-gpt-neox test-tokenizer-1-bpe.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-gpt-neox.gguf)
-llama_test_executable(test-tokenizer-1-refact test-tokenizer-1-bpe.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-refact.gguf)
-llama_test_executable(test-tokenizer-1-starcoder test-tokenizer-1-bpe.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-starcoder.gguf)
-# llama_test_executable(test-tokenizer-1-bloom test-tokenizer-1-bpe.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-bloom.gguf) # BIG
+llama_test_executable (test-tokenizer-1-falcon           test-tokenizer-1-bpe.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-falcon.gguf)
+llama_test_executable (test-tokenizer-1-aquila           test-tokenizer-1-bpe.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-aquila.gguf)
+llama_test_executable (test-tokenizer-1-mpt              test-tokenizer-1-bpe.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-mpt.gguf)
+llama_test_executable (test-tokenizer-1-stablelm-3b-4e1t test-tokenizer-1-bpe.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-stablelm-3b-4e1t.gguf)
+llama_test_executable (test-tokenizer-1-gpt-neox         test-tokenizer-1-bpe.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-gpt-neox.gguf)
+llama_test_executable (test-tokenizer-1-refact           test-tokenizer-1-bpe.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-refact.gguf)
+llama_test_executable (test-tokenizer-1-starcoder        test-tokenizer-1-bpe.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-starcoder.gguf)
+# llama_test_executable (test-tokenizer-1-bloom test-tokenizer-1-bpe.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-bloom.gguf) # BIG
+
 llama_build_and_test_executable(test-grammar-parser.cpp)
 llama_build_and_test_executable(test-llama-grammar.cpp)
-llama_build_and_test_executable(test-grad0.cpp) # SLOW
+llama_build_and_test_executable(test-grad0.cpp)
 # llama_build_and_test_executable(test-opt.cpp) # SLOW
+llama_build_and_test_executable(test-backend-ops.cpp)
 
 llama_build_and_test_executable(test-rope.cpp)
 
diff --git a/tests/test-backend-ops.cpp b/tests/test-backend-ops.cpp
new file mode 100644
index 000000000..f3df8a8c6
--- /dev/null
+++ b/tests/test-backend-ops.cpp
@@ -0,0 +1,1690 @@
+#include <ggml.h>
+#include <ggml-alloc.h>
+#include <ggml-backend.h>
+#include <ggml-backend-impl.h>
+#include <algorithm>
+#include <array>
+#include <cfloat>
+#include <cstring>
+#include <functional>
+#include <memory>
+#include <random>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string>
+#include <thread>
+#include <vector>
+
+
+static void init_tensor_uniform(ggml_tensor * tensor, float min = -1.0f, float max = 1.0f) {
+    size_t size = ggml_nelements(tensor);
+    std::vector<float> data(size);
+
+#if 0
+    std::default_random_engine generator(rd());
+    std::uniform_real_distribution<float> distribution(min, max);
+
+    for (size_t i = 0; i < size; i++) {
+        data[i] = distribution(generator);
+    }
+#endif
+    auto init_thread = [&](size_t start, size_t end) {
+        std::random_device rd;
+        std::default_random_engine generator(rd());
+        std::uniform_real_distribution<float> distribution(min, max);
+
+        for (size_t i = start; i < end; i++) {
+            data[i] = distribution(generator);
+        }
+    };
+
+    size_t n_threads = std::thread::hardware_concurrency();
+    std::vector<std::thread> threads;
+    threads.reserve(n_threads);
+    for (size_t i = 0; i < n_threads; i++) {
+        size_t start =     i*size/n_threads;
+        size_t end   = (i+1)*size/n_threads;
+        threads.emplace_back(init_thread, start, end);
+    }
+    for (auto & t : threads) {
+        t.join();
+    }
+
+    if (tensor->type == GGML_TYPE_F32 || tensor->type == GGML_TYPE_I32) {
+        ggml_backend_tensor_set(tensor, data.data(), 0, size * sizeof(float));
+    } else if (ggml_is_quantized(tensor->type) || tensor->type == GGML_TYPE_F16) {
+        GGML_ASSERT(size % ggml_blck_size(tensor->type) == 0);
+        std::vector<uint8_t> dataq(ggml_row_size(tensor->type, size));
+        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());
+    } else {
+        GGML_ASSERT(false);
+    }
+}
+
+static std::vector<float> tensor_to_float(const ggml_tensor * t) {
+    std::vector<float> tv;
+    tv.reserve(ggml_nelements(t));
+
+    std::vector<uint8_t> buf(ggml_nbytes(t));
+    ggml_backend_tensor_get(t, buf.data(), 0, ggml_nbytes(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++) {
+        for (int64_t i2 = 0; i2 < t->ne[2]; i2++) {
+            for (int64_t i1 = 0; i1 < t->ne[1]; i1++) {
+                for (int64_t i0 = 0; i0 < t->ne[0]; i0 += bs) {
+                    size_t i = i3*t->nb[3] + i2*t->nb[2] + i1*t->nb[1] + i0/bs*t->nb[0];
+                    if (t->type == GGML_TYPE_F16) {
+                        tv.push_back(ggml_fp16_to_fp32(*(ggml_fp16_t*)&buf[i]));
+                    } else if (t->type == GGML_TYPE_F32) {
+                        tv.push_back(*(float *) &buf[i]);
+                    } else if (t->type == GGML_TYPE_I32) {
+                        tv.push_back((float)*(int32_t *) &buf[i]);
+                    } else if (quantized) {
+                        tt.to_float(&buf[i], vq.data(), bs);
+                        tv.insert(tv.end(), vq.begin(), vq.end());
+                    } else {
+                        GGML_ASSERT(false);
+                    }
+                }
+            }
+        }
+    }
+
+    return tv;
+}
+
+/*
+static double cosine_similarity(const float * v1, const float * v2, size_t n) {
+    double dot = 0.0;
+    double mag1 = 0.0;
+    double mag2 = 0.0;
+
+    for (size_t i = 0; i < n; i++) {
+        if (std::isnan(v1[i]) || std::isnan(v2[i])) {
+            return -1.0f;
+        }
+        if (std::isinf(v1[i]) && std::isinf(v2[i])) {
+            continue;
+        }
+        dot  += v1[i]*v2[i];
+        mag1 += v1[i]*v1[i];
+        mag2 += v2[i]*v2[i];
+    }
+
+    return dot/sqrt(mag1*mag2);
+}
+
+static float distance(const float * v1, const float * v2, size_t n) {
+    double d = 0.0;
+
+    for (size_t i = 0; i < n; i++) {
+        if (std::isnan(v1[i]) || std::isnan(v2[i])) {
+            return INFINITY;
+        }
+        if (std::isinf(v1[i]) && std::isinf(v2[i])) {
+            continue;
+        }
+        d += (v1[i] - v2[i])*(v1[i] - v2[i]);
+    }
+
+    return sqrt(d);
+}
+
+static float vec_len(const float * v, size_t n) {
+    double d = 0.0;
+
+    for (size_t i = 0; i < n; i++) {
+        if (std::isnan(v[i])) {
+            return INFINITY;
+        }
+        if (std::isinf(v[i])) {
+            continue;
+        }
+        d += v[i]*v[i];
+    }
+
+    return sqrt(d);
+}
+*/
+
+// normalized mean squared error = mse(a, b) / mse(a, 0)
+static double nmse(const float * a, const float * b, size_t n) {
+    double mse_a_b = 0.0;
+    double mse_a_0 = 0.0;
+
+    for (size_t i = 0; i < n; i++) {
+        float a_i = a[i];
+        float b_i = b[i];
+
+        mse_a_b += (a_i - b_i) * (a_i - b_i);
+        mse_a_0 += a_i * a_i;
+    }
+
+    return mse_a_b / mse_a_0;
+}
+
+// utils for printing the variables of the test cases
+#define VAR_TO_STR(x) (#x "=" + var_to_str(x))
+
+template<typename T>
+static std::string var_to_str(const T & x) {
+    return std::to_string(x);
+}
+
+template<typename T, size_t N>
+static std::string var_to_str(const T (&x)[N]) {
+    std::string s = "[";
+    for (size_t i = 0; i < N; i++) {
+        if (i > 0) {
+            s += ",";
+        }
+        s += var_to_str(x[i]);
+    }
+    s += "]";
+    return s;
+}
+
+template<typename T, size_t N>
+static std::string var_to_str(const std::array<T, N> & x) {
+    std::string s = "[";
+    for (size_t i = 0; i < N; i++) {
+        if (i > 0) {
+            s += ",";
+        }
+        s += var_to_str(x[i]);
+    }
+    s += "]";
+    return s;
+}
+
+//static std::string var_to_str(ggml_unary_op unary_op) {
+//    return ggml_unary_op_name(unary_op);
+//}
+
+static std::string var_to_str(ggml_type type) {
+    return ggml_type_name(type);
+}
+
+#define VARS_TO_STR1(a) VAR_TO_STR(a)
+#define VARS_TO_STR2(a, b) VAR_TO_STR(a) + "," + VAR_TO_STR(b)
+#define VARS_TO_STR3(a, b, c) VAR_TO_STR(a) + "," + VARS_TO_STR2(b, c)
+#define VARS_TO_STR4(a, b, c, d) VAR_TO_STR(a) + "," + VARS_TO_STR3(b, c, d)
+#define VARS_TO_STR5(a, b, c, d, e) VAR_TO_STR(a) + "," + VARS_TO_STR4(b, c, d, e)
+#define VARS_TO_STR6(a, b, c, d, e, f) VAR_TO_STR(a) + "," + VARS_TO_STR5(b, c, d, e, f)
+#define VARS_TO_STR7(a, b, c, d, e, f, g) VAR_TO_STR(a) + "," + VARS_TO_STR6(b, c, d, e, f, g)
+#define VARS_TO_STR8(a, b, c, d, e, f, g, h) VAR_TO_STR(a) + "," + VARS_TO_STR7(b, c, d, e, f, g, h)
+#define VARS_TO_STR9(a, b, c, d, e, f, g, h, i) VAR_TO_STR(a) + "," + VARS_TO_STR8(b, c, d, e, f, g, h, i)
+#define VARS_TO_STR10(a, b, c, d, e, f, g, h, i, j) VAR_TO_STR(a) + "," + VARS_TO_STR9(b, c, d, e, f, g, h, i, j)
+#define VARS_TO_STR11(a, b, c, d, e, f, g, h, i, j, k) VAR_TO_STR(a) + "," + VARS_TO_STR10(b, c, d, e, f, g, h, i, j, k)
+
+
+// accept FLT_MAX as infinity
+static bool isinf_or_max(float f) {
+    return std::isinf(f) || f == FLT_MAX || f == -FLT_MAX;
+}
+
+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() {}
+
+    virtual std::string op_desc(ggml_tensor * t) {
+        return ggml_op_desc(t);
+    }
+
+    virtual std::string vars() {
+        return "";
+    }
+
+    virtual ggml_tensor * build_graph(ggml_context * ctx) = 0;
+
+    virtual double max_nmse_err() {
+        return 1e-7;
+    }
+
+    virtual void initialize_tensors(ggml_context * ctx) {
+        for (ggml_tensor * t = ggml_get_first_tensor(ctx); t != nullptr; t = ggml_get_next_tensor(ctx, t)) {
+            init_tensor_uniform(t);
+        }
+    }
+
+    virtual size_t op_size(ggml_tensor * t) {
+        size_t size = ggml_nbytes(t);
+        // add source tensors
+        for (int i = 0; i < GGML_MAX_SRC; i++) {
+            if (t->src[i] != NULL) {
+                size += ggml_nbytes(t->src[i]);
+            }
+        }
+        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,
+            /* .no_alloc = */ true,
+        };
+        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) {
+            //printf("  %s: skipping\n", op_desc(out).c_str());
+            ggml_free(ctx);
+            return true;
+        }
+
+        printf("  %s(%s): ", op_desc(out).c_str(), vars().c_str());
+        fflush(stdout);
+
+        // check if backends support op
+        for (ggml_backend_t backend : {backend1, backend2}) {
+            if (!ggml_backend_supports_op(backend, out)) {
+                printf("not supported\n");
+                ggml_free(ctx);
+                return true;
+            }
+        }
+
+        // post-graph sentinel
+        add_sentinel(ctx);
+
+        // allocate
+        ggml_backend_buffer_t buf = ggml_backend_alloc_ctx_tensors(ctx, backend1);
+
+        // build graph
+        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);
+
+        // compare
+        struct callback_userdata {
+            bool   ok;
+            double max_err;
+        };
+
+        callback_userdata ud {
+            true,
+            max_nmse_err(),
+        };
+
+        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);
+
+            for (size_t i = 0; i < f1.size(); i++) {
+                // check for nans
+                if (std::isnan(f1[i]) || std::isnan(f2[i])) {
+                    printf("[%s] NaN at index %zu (%f %f) ", ggml_op_desc(t1), i, f1[i], f2[i]);
+                    ud->ok = false;
+                    return true;
+                }
+                // check for infs: both must be inf of the same sign, or both must be finite
+                if (isinf_or_max(f1[i]) || isinf_or_max(f2[i])) {
+                    if (isinf_or_max(f1[i]) && isinf_or_max(f2[i])) {
+                        if (std::signbit(f1[i]) != std::signbit(f2[i])) {
+                            printf("[%s] inf sign mismatch: %f %f ", ggml_op_desc(t1), f1[i], f2[i]);
+                            ud->ok = false;
+                            return true;
+                        }
+                    } else {
+                        printf("[%s] inf mismatch: %f %f ", ggml_op_desc(t1), f1[i], f2[i]);
+                        ud->ok = false;
+                        return true;
+                    }
+                }
+            }
+
+            double err = nmse(f1.data(), f2.data(), f1.size());
+            if (err > ud->max_err) {
+                printf("[%s] NMSE = %f ", ggml_op_desc(t1), err);
+                //for (int i = 0; i < f1.size(); i++) {
+                //    printf("%5d %9.6f %9.6f, diff = %9.6f\n", i, f1[i], f2[i], f1[i] - f2[i]);
+                //}
+                //printf("\n");
+                //exit(1);
+                ud->ok = false;
+            }
+            return true;
+
+            GGML_UNUSED(index);
+        };
+
+        ggml_backend_compare_graph_backend(backend1, backend2, gf, callback, &ud);
+
+        if (ud.ok) {
+            printf("\033[1;32mOK\033[0m\n");
+        } else {
+            printf("\033[1;31mFAIL\033[0m\n");
+        }
+
+        ggml_backend_buffer_free(buf);
+
+        ggml_free(ctx);
+
+        return ud.ok;
+    }
+
+    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 = {
+            /* .mem_size = */ ggml_tensor_overhead()*128 + ggml_graph_overhead_custom(graph_nodes, false),
+            /* .mem_base = */ NULL,
+            /* .no_alloc = */ true,
+        };
+        ggml_context * ctx = ggml_init(params);
+
+        ggml_tensor * out = build_graph(ctx);
+
+        if (op_name != nullptr && op_desc(out) != op_name) {
+            //printf("  %s: skipping\n", op_desc(out).c_str());
+            ggml_free(ctx);
+            return true;
+        }
+
+        int len = printf("  %s(%s): ", op_desc(out).c_str(), vars().c_str());
+        fflush(stdout);
+
+        // check if backends support op
+        if (!ggml_backend_supports_op(backend, out)) {
+            printf("not supported\n");
+            ggml_free(ctx);
+            return true;
+        }
+
+        // align while also leaving some margin for variations in parameters
+        int align = 20;
+        int last = (len + align - 1) / align * align;
+        if (last - len < 5) {
+            last += align;
+        }
+        last = std::max(last, 60);
+        printf("%*s", last - len, "");
+
+        // allocate
+        ggml_backend_buffer_t buf = ggml_backend_alloc_ctx_tensors(ctx, backend);
+
+        // randomize tensors
+        initialize_tensors(ctx);
+
+        // build graph
+        ggml_cgraph * gf = ggml_new_graph_custom(ctx, graph_nodes, false);
+        ggml_build_forward_expand(gf, out);
+
+        // warmup run
+        ggml_backend_graph_compute(backend, gf);
+
+        // duplicate the op
+        size_t target_size = ggml_backend_is_cpu(backend) ? 1ULL << 33 : 1ULL << 35; // 8 GB CPU, 32 GB GPU
+        int n_runs = std::min((size_t)gf->size - gf->n_nodes, target_size / op_size(out)) + 1;
+        for (int i = 1; i < n_runs; i++) {
+            gf->nodes[gf->n_nodes++] = out;
+        }
+
+        // calculate memory
+        size_t mem = n_runs * op_size(out);
+        auto tensor_op_size = [](ggml_tensor * t) {
+            size_t size = ggml_nbytes(t);
+            // add source tensors
+            for (int i = 0; i < GGML_MAX_SRC; i++) {
+                if (t->src[i] != NULL) {
+                    size += ggml_nbytes(t->src[i]);
+                }
+            }
+            return size;
+        };
+        for (int i = 0; i < gf->n_nodes; i++) {
+            if (ggml_is_view_op(gf->nodes[i]->op) || gf->nodes[i] == out) {
+                continue;
+            }
+            mem += tensor_op_size(gf->nodes[i]);
+        }
+
+        // run
+        ggml_backend_synchronize(backend);
+
+        int64_t start_time = ggml_time_us();
+        ggml_backend_graph_compute(backend, gf);
+        ggml_backend_synchronize(backend);
+        int64_t end_time = ggml_time_us();
+        double time_us = end_time - start_time;
+
+        printf("    %5d runs - %8.2f us/run - %8zu kB/run - \033[1;34m%7.2f GB/s\033[0m\n",
+            n_runs,
+            time_us / n_runs,
+            op_size(out) / 1024,
+            mem / (time_us/1e6) / 1024.0 / 1024.0 / 1024.0);
+
+        ggml_backend_buffer_free(buf);
+
+        ggml_free(ctx);
+
+        return true;
+    }
+};
+
+// GGML_OP_UNARY
+struct test_unary : public test_case {
+    const ggml_unary_op op;
+    const ggml_type type;
+    const std::array<int64_t, 4> ne;
+
+    std::string vars() override {
+        return VARS_TO_STR2(type, ne);
+    }
+
+    test_unary(ggml_unary_op op,
+            ggml_type type = GGML_TYPE_F32,
+            std::array<int64_t, 4> ne = {128, 10, 10, 10})
+        : op(op), type(type), ne(ne) {}
+
+    ggml_tensor * build_graph(ggml_context * ctx) override {
+        ggml_tensor * in = ggml_new_tensor(ctx, type, 4, ne.data());
+        ggml_tensor * out = ggml_unary(ctx, in, op);
+        return out;
+    }
+};
+
+// GGML_OP_GET_ROWS
+struct test_get_rows : public test_case {
+    const ggml_type type;
+    const int n; // cols
+    const int m; // rows
+    const int r; // rows to get
+    const int b; // batch size
+    const bool v; // view (non-contiguous src1)
+
+    std::string vars() override {
+        return VARS_TO_STR6(type, n, m, r, b, v);
+    }
+
+    test_get_rows(ggml_type type = GGML_TYPE_F32, int n = 10, int m = 5, int r = 3, int b = 1, bool v = false)
+        : type(type), n(n), m(m), r(r), b(b), v(v) {}
+
+    ggml_tensor * build_graph(ggml_context * ctx) override {
+        ggml_tensor * in = ggml_new_tensor_3d(ctx, type, n, m, b);
+        ggml_tensor * rows = ggml_new_tensor_2d(ctx, GGML_TYPE_I32, r, b);
+        if (v) {
+            rows = ggml_view_2d(ctx, rows, r/2, b, rows->nb[1], 0);
+        }
+        ggml_tensor * out = ggml_get_rows(ctx, in, rows);
+        return out;
+    }
+
+    void initialize_tensors(ggml_context * ctx) override {
+        for (ggml_tensor * t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) {
+            if (t->type == GGML_TYPE_I32) {
+                if (ggml_is_view_op(t->op)) { continue; }
+                // rows
+                std::vector<int> data(r*b);
+                for (int i = 0; i < r*b; i++) {
+                    data[i] = rand() % m;
+                }
+                ggml_backend_tensor_set(t, data.data(), 0, r * b * sizeof(int));
+            } else {
+                init_tensor_uniform(t);
+            }
+        }
+    }
+};
+
+// GGML_OP_REPEAT
+struct test_repeat : public test_case {
+    const ggml_type type;
+    const std::array<int64_t, 4> ne;
+    const std::array<int, 4> nr;
+
+    std::string vars() override {
+        return VARS_TO_STR3(type, ne, nr);
+    }
+
+    size_t op_size(ggml_tensor * t) override {
+        return ggml_nbytes(t) * 2;
+    }
+
+    test_repeat(ggml_type type = GGML_TYPE_F32,
+            std::array<int64_t, 4> ne = {10, 10, 10, 10},
+            std::array<int, 4> nr = {2, 2, 2, 2})
+        : type(type), ne(ne), nr(nr) {}
+
+    ggml_tensor * build_graph(ggml_context * ctx) override {
+        ggml_tensor * target = ggml_new_tensor_4d(ctx, type, ne[0]*nr[0], ne[1]*nr[1], ne[2]*nr[2], ne[3]*nr[3]);
+        ggml_tensor * src = ggml_new_tensor(ctx, type, 4, ne.data());
+        ggml_tensor * out = ggml_repeat(ctx, src, target);
+        return out;
+    }
+};
+
+// GGML_OP_DUP
+struct test_dup : public test_case {
+    const ggml_type type;
+    const std::array<int64_t, 4> ne;
+
+    std::string vars() override {
+        return VARS_TO_STR2(type, ne);
+    }
+
+    test_dup(ggml_type type = GGML_TYPE_F32,
+            std::array<int64_t, 4> ne = {10, 10, 10, 1})
+        : type(type), ne(ne) {}
+
+    ggml_tensor * build_graph(ggml_context * ctx) override {
+        ggml_tensor * src = ggml_new_tensor(ctx, type, 4, ne.data());
+        ggml_tensor * out = ggml_dup(ctx, src);
+        return out;
+    }
+};
+
+// GGML_OP_CPY
+struct test_cpy : public test_case {
+    const ggml_type type_src;
+    const ggml_type type_dst;
+    const std::array<int64_t, 4> ne;
+
+    std::string vars() override {
+        return VARS_TO_STR3(type_src, type_dst, ne);
+    }
+
+    size_t op_size(ggml_tensor * t) override {
+        return ggml_nbytes(t) + ggml_nbytes(t->src[0]);
+    }
+
+    test_cpy(ggml_type type_src = GGML_TYPE_F32, ggml_type type_dst = GGML_TYPE_F32,
+            std::array<int64_t, 4> ne = {10, 10, 10, 1})
+        : type_src(type_src), type_dst(type_dst), ne(ne) {}
+
+    ggml_tensor * build_graph(ggml_context * ctx) override {
+        ggml_tensor * src = ggml_new_tensor(ctx, type_src, 4, ne.data());
+        ggml_tensor * dst = ggml_new_tensor(ctx, type_dst, 4, ne.data());
+        ggml_tensor * out = ggml_cpy(ctx, src, dst);
+        return out;
+    }
+};
+
+// GGML_OP_CONT
+struct test_cont : public test_case {
+    const ggml_type type;
+    const std::array<int64_t, 4> ne;
+
+    std::string vars() override {
+        return VARS_TO_STR2(type, ne);
+    }
+
+    test_cont(ggml_type type = GGML_TYPE_F32,
+            std::array<int64_t, 4> ne = {10, 10, 10, 1})
+        : type(type), ne(ne) {}
+
+    ggml_tensor * build_graph(ggml_context * ctx) override {
+        ggml_tensor * src = ggml_new_tensor(ctx, type, 4, ne.data());
+        src = ggml_transpose(ctx, src);
+        ggml_tensor * out = ggml_cont(ctx, src);
+
+        return out;
+    }
+};
+
+// GGML_OP_ADD
+// GGML_OP_MUL
+// GGML_OP_DIV
+struct test_bin_bcast : public test_case {
+    using op_t = ggml_tensor * (*) (ggml_context *, ggml_tensor *, ggml_tensor *);
+    op_t op;
+    const ggml_type type;
+    const std::array<int64_t, 4> ne;
+    const std::array<int, 4> nr;
+
+    std::string vars() override {
+        return VARS_TO_STR3(type, ne, nr);
+    }
+
+    size_t op_size(ggml_tensor * t) override {
+        return ggml_nbytes(t) * 3;
+    }
+
+    test_bin_bcast(op_t op, ggml_type type = GGML_TYPE_F32,
+            std::array<int64_t, 4> ne = {10, 10, 1, 1},
+            std::array<int, 4> nr = {1, 2, 1, 1})
+        : op(op), type(type), ne(ne), nr(nr) {}
+
+    ggml_tensor * build_graph(ggml_context * ctx) override {
+        ggml_tensor * a = ggml_new_tensor_4d(ctx, type, ne[0]*nr[0], ne[1]*nr[1], ne[2]*nr[2], ne[3]*nr[3]);
+        ggml_tensor * b = ggml_new_tensor(ctx, type, 4, ne.data());
+        ggml_tensor * out = op(ctx, a, b);
+        return out;
+    }
+
+    void initialize_tensors(ggml_context * ctx) override {
+        for (ggml_tensor * t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) {
+            if (op == ggml_div) {
+                // avoid division by zero
+                init_tensor_uniform(t, 1.0f, 2.0f);
+            } else {
+                init_tensor_uniform(t);
+            }
+        }
+    }
+};
+
+// GGML_OP_SCALE
+struct test_scale : public test_case {
+    const ggml_type type;
+    const std::array<int64_t, 4> ne;
+    float scale;
+
+    std::string vars() override {
+        return VARS_TO_STR3(type, ne, scale);
+    }
+
+    test_scale(ggml_type type = GGML_TYPE_F32,
+            std::array<int64_t, 4> ne = {10, 10, 10, 10},
+            float scale = 2.0f)
+        : type(type), ne(ne), scale(scale) {}
+
+    ggml_tensor * build_graph(ggml_context * ctx) override {
+        ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data());
+        ggml_tensor * out = ggml_scale(ctx, a, scale);
+        return out;
+    }
+};
+
+// GGML_OP_NORM
+struct test_norm : public test_case {
+    const ggml_type type;
+    const std::array<int64_t, 4> ne;
+    float eps;
+
+    std::string vars() override {
+        return VARS_TO_STR3(type, ne, eps);
+    }
+
+    test_norm(ggml_type type = GGML_TYPE_F32,
+            std::array<int64_t, 4> ne = {64, 10, 10, 10},
+            float eps = 1e-6f)
+        : type(type), ne(ne), eps(eps) {}
+
+    ggml_tensor * build_graph(ggml_context * ctx) override {
+        ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data());
+        ggml_tensor * out = ggml_norm(ctx, a, eps);
+        return out;
+    }
+};
+
+// GGML_OP_RMS_NORM
+struct test_rms_norm : public test_case {
+    const ggml_type type;
+    const std::array<int64_t, 4> ne;
+    float eps;
+
+    std::string vars() override {
+        return VARS_TO_STR3(type, ne, eps);
+    }
+
+    test_rms_norm(ggml_type type = GGML_TYPE_F32,
+            std::array<int64_t, 4> ne = {64, 10, 10, 10},
+            float eps = 1e-6f)
+        : type(type), ne(ne), eps(eps) {}
+
+    ggml_tensor * build_graph(ggml_context * ctx) override {
+        ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data());
+        ggml_tensor * out = ggml_rms_norm(ctx, a, eps);
+        return out;
+    }
+};
+
+// GGML_OP_MUL_MAT
+struct test_mul_mat : public test_case {
+    const ggml_type type_a;
+    const ggml_type type_b;
+    const int64_t m;
+    const int64_t n;
+    const int64_t k;
+    const std::array<int64_t, 2> bs; // dims 3 and 4
+    const std::array<int64_t, 2> nr; // repeat in dims 3 and 4
+
+    std::string vars() override {
+        return VARS_TO_STR7(type_a, type_b, m, n, k, bs, nr);
+    }
+
+    double max_nmse_err() override {
+        return 5e-4;
+    }
+
+    size_t op_size(ggml_tensor * t) override {
+        size_t a = ggml_nbytes(t->src[0]) * n * nr[0] * nr[1];
+        size_t b = ggml_nbytes(t->src[1]) * m;
+        size_t c  = ggml_nbytes(t);
+        return a + b + c;
+
+        GGML_UNUSED(t);
+    }
+
+    test_mul_mat(ggml_type type_a = GGML_TYPE_F32, ggml_type type_b = GGML_TYPE_F32,
+            int64_t m = 32, int64_t n = 32, int64_t k = 32,
+            std::array<int64_t, 2> bs = {10, 10},
+            std::array<int64_t, 2> nr = {2, 2})
+        : type_a(type_a), type_b(type_b), m(m), n(n), k(k), bs(bs), nr(nr) {}
+
+    ggml_tensor * build_graph(ggml_context * ctx) override {
+        // C^T = A * B^T: (k, m) * (k, n) => (m, n)
+        ggml_tensor * a = ggml_new_tensor_4d(ctx, type_a, k, m, bs[0]      , bs[1]);
+        ggml_tensor * b = ggml_new_tensor_4d(ctx, type_b, k, n, bs[0]*nr[0], bs[1]*nr[1]);
+        ggml_tensor * out = ggml_mul_mat(ctx, a, b);
+        return out;
+    }
+};
+
+// GGML_OP_MUL_MAT_ID
+struct test_mul_mat_id : public test_case {
+    const ggml_type type_a;
+    const ggml_type type_b;
+    const int n_mats;
+    const int id;
+    const int64_t m;
+    const int64_t n;
+    const int64_t k;
+    const bool v; // view (non-contiguous ids)
+
+    std::string vars() override {
+        return VARS_TO_STR8(type_a, type_b, n_mats, id, m, n, k, v);
+    }
+
+    double max_nmse_err() override {
+        return 5e-4;
+    }
+
+    size_t op_size(ggml_tensor * t) override {
+        size_t a = ggml_nbytes(t->src[2]) * n;
+        size_t b = ggml_nbytes(t->src[1]) * m;
+        size_t c  = ggml_nbytes(t);
+        return a + b + c;
+
+        GGML_UNUSED(t);
+    }
+
+    test_mul_mat_id(ggml_type type_a = GGML_TYPE_F32, ggml_type type_b = GGML_TYPE_F32,
+            int n_mats = 2, int id = 0,
+            int64_t m = 32, int64_t n = 32, int64_t k = 32, bool v = false)
+        : type_a(type_a), type_b(type_b), n_mats(n_mats), id(id),
+            m(m), n(n), k(k), v(v) {}
+
+    ggml_tensor * build_graph(ggml_context * ctx) override {
+        // C^T = A * B^T: (k, m) * (k, n) => (m, n)
+        std::vector<ggml_tensor *> mats;
+        for (int i = 0; i < n_mats; i++) {
+            ggml_tensor * a = ggml_new_tensor_2d(ctx, type_a, k, m);
+            mats.push_back(a);
+        }
+        ggml_tensor * ids = ggml_new_tensor_2d(ctx, GGML_TYPE_I32, n_mats, n);
+        if (v) {
+            ids = ggml_view_2d(ctx, ids, n_mats/2, ids->ne[1], ids->nb[1], 0);
+        }
+        ggml_tensor * b = ggml_new_tensor_2d(ctx, type_b, k, n);
+        ggml_tensor * out = ggml_mul_mat_id(ctx, mats.data(), n_mats, ids, v ? id/2 : id, b);
+        return out;
+    }
+
+    void initialize_tensors(ggml_context * ctx) override {
+        std::random_device rd;
+        std::default_random_engine rng(rd());
+        for (ggml_tensor * t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) {
+            if (t->type == GGML_TYPE_I32) {
+                if (ggml_is_view_op(t->op)) { continue; }
+                // ids
+                for (int64_t r = 0; r < ggml_nrows(t); r++) {
+                    std::vector<int32_t> data(t->ne[0]);
+                    for (int i = 0; i < t->ne[0]; i++) {
+                        data[i] = i % n_mats;
+                    }
+                    std::shuffle(data.begin(), data.end(), rng);
+                    ggml_backend_tensor_set(t, data.data(), r * t->nb[1], t->ne[0] * sizeof(int32_t));
+                }
+            } else {
+                init_tensor_uniform(t);
+            }
+        }
+    }
+};
+
+// GGML_OP_SQR
+struct test_sqr : public test_case {
+    const ggml_type type;
+    const std::array<int64_t, 4> ne;
+
+    std::string vars() override {
+        return VARS_TO_STR2(type, ne);
+    }
+
+    test_sqr(ggml_type type = GGML_TYPE_F32,
+            std::array<int64_t, 4> ne = {10, 10, 10, 10})
+        : type(type), ne(ne) {}
+
+    ggml_tensor * build_graph(ggml_context * ctx) override {
+        ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data());
+        ggml_tensor * out = ggml_sqr(ctx, a);
+        return out;
+    }
+};
+
+// GGML_OP_CLAMP
+struct test_clamp : public test_case {
+    const ggml_type type;
+    const std::array<int64_t, 4> ne;
+    float min;
+    float max;
+
+    std::string vars() override {
+        return VARS_TO_STR4(type, ne, min, max);
+    }
+
+    test_clamp(ggml_type type = GGML_TYPE_F32,
+            std::array<int64_t, 4> ne = {10, 10, 10, 10},
+            float min = -0.5f, float max = 0.5f)
+        : type(type), ne(ne), min(min), max(max) {}
+
+    ggml_tensor * build_graph(ggml_context * ctx) override {
+        ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data());
+        ggml_tensor * out = ggml_clamp(ctx, a, min, max);
+        return out;
+    }
+};
+
+// GGML_OP_DIAG_MASK_INF
+struct test_diag_mask_inf : public test_case {
+    const ggml_type type;
+    const std::array<int64_t, 4> ne;
+    const int n_past;
+
+    std::string vars() override {
+        return VARS_TO_STR3(type, ne, n_past);
+    }
+
+    test_diag_mask_inf(ggml_type type = GGML_TYPE_F32,
+            std::array<int64_t, 4> ne = {10, 10, 10, 10},
+            int n_past = 5)
+        : type(type), ne(ne), n_past(n_past) {}
+
+    ggml_tensor * build_graph(ggml_context * ctx) override {
+        ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data());
+        ggml_tensor * out = ggml_diag_mask_inf(ctx, a, n_past);
+        return out;
+    }
+};
+
+// GGML_OP_SOFT_MAX
+struct test_soft_max : public test_case {
+    const ggml_type type;
+    const std::array<int64_t, 4> ne;
+
+    std::string vars() override {
+        return VARS_TO_STR2(type, ne);
+    }
+
+    test_soft_max(ggml_type type = GGML_TYPE_F32,
+            std::array<int64_t, 4> ne = {10, 10, 10, 10})
+        : type(type), ne(ne) {}
+
+    ggml_tensor * build_graph(ggml_context * ctx) override {
+        ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data());
+        ggml_tensor * out = ggml_soft_max(ctx, a);
+        return out;
+    }
+};
+
+// GGML_OP_ROPE
+struct test_rope : public test_case {
+    const ggml_type type;
+    const std::array<int64_t, 4> ne;
+    int n_dims;
+    int mode;
+    int n_ctx;
+
+    std::string vars() override {
+        return VARS_TO_STR5(type, ne, n_dims, mode, n_ctx);
+    }
+
+    test_rope(ggml_type type = GGML_TYPE_F32,
+            std::array<int64_t, 4> ne = {10, 10, 10, 1},
+            int n_dims = 10, int mode = 0, int n_ctx = 512)
+        : type(type), ne(ne), n_dims(n_dims), mode(mode), n_ctx(n_ctx) {}
+
+    ggml_tensor * build_graph(ggml_context * ctx) override {
+        ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data());
+        ggml_tensor * pos = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, ne[2]);
+        ggml_tensor * out = ggml_rope(ctx, a, pos, n_dims, mode, n_ctx);
+        return out;
+    }
+
+    void initialize_tensors(ggml_context * ctx) override {
+        for (ggml_tensor * t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) {
+            if (t->type == GGML_TYPE_I32) {
+                // pos
+                std::vector<int> data(ne[2]);
+                for (int i = 0; i < ne[2]; i++) {
+                    data[i] = rand() % n_ctx;
+                }
+                ggml_backend_tensor_set(t, data.data(), 0, ne[2] * sizeof(int));
+            } else {
+                init_tensor_uniform(t);
+            }
+        }
+    }
+};
+
+// GGML_OP_ALIBI
+struct test_alibi : public test_case {
+    const ggml_type type;
+    const std::array<int64_t, 4> ne;
+    int n_past;
+    int n_head;
+    float bias_max;
+
+    std::string vars() override {
+        return VARS_TO_STR5(type, ne, n_past, n_head, bias_max);
+    }
+
+    test_alibi(ggml_type type = GGML_TYPE_F32,
+            std::array<int64_t, 4> ne = {10, 10, 10, 10},
+            int n_past = 512, int n_head = 10, float bias_max = 0.5f)
+        : type(type), ne(ne), n_past(n_past), n_head(n_head), bias_max(bias_max) {}
+
+    ggml_tensor * build_graph(ggml_context * ctx) override {
+        ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data());
+        ggml_tensor * out = ggml_alibi(ctx, a, n_past, n_head, bias_max);
+        return out;
+    }
+};
+
+// GGML_OP_IM2COL
+struct test_im2col : public test_case {
+    const ggml_type type_input;
+    const ggml_type type_kernel;
+    const std::array<int64_t, 4> ne_input;
+    const std::array<int64_t, 4> ne_kernel;
+    // stride
+    const int s0;
+    const int s1;
+    // padding
+    const int p0;
+    const int p1;
+    // dilatation
+    const int d0;
+    const int d1;
+    // mode
+    const bool is_2D;
+
+    std::string vars() override {
+        return VARS_TO_STR11(type_input, type_kernel, ne_input, ne_kernel, s0, s1, p0, p1, d0, d1, is_2D);
+    }
+
+    test_im2col(ggml_type type_input = GGML_TYPE_F32, ggml_type type_kernel = GGML_TYPE_F16,
+            std::array<int64_t, 4> ne_input = {10, 10, 3, 1}, // [input_width, input_height, input_channels, 1]
+            std::array<int64_t, 4> ne_kernel = {3, 3, 3, 1}, // [kernel_width, kernel_height, input_channels, 1]
+            int s0 = 1, int s1 = 1,
+            int p0 = 1, int p1 = 1,
+            int d0 = 1, int d1 = 1,
+            bool is_2D = true)
+        : type_input(type_input), type_kernel(type_kernel), ne_input(ne_input), ne_kernel(ne_kernel), s0(s0), s1(s1), p0(p0), p1(p1), d0(d0), d1(d1), is_2D(is_2D) {}
+
+    ggml_tensor * build_graph(ggml_context * ctx) override {
+        ggml_tensor * input = ggml_new_tensor(ctx, type_input, 4, ne_input.data());
+        ggml_tensor * kernel = ggml_new_tensor(ctx, type_kernel, 4, ne_kernel.data());
+        ggml_tensor * out = ggml_im2col(ctx, kernel, input, s0, s1, p0, p1, d0, d1, is_2D);
+        return out;
+    }
+};
+
+// GGML_OP_CONCAT
+struct test_concat : public test_case {
+    const ggml_type type;
+    const std::array<int64_t, 4> ne;
+    const int64_t b_ne2;
+
+    std::string vars() override {
+        return VARS_TO_STR3(type, ne, b_ne2);
+    }
+
+    test_concat(ggml_type type = GGML_TYPE_F32,
+            std::array<int64_t, 4> ne = {10, 10, 10, 10},
+            int64_t b_ne2 = 10)
+        : type(type), ne(ne), b_ne2(b_ne2) {}
+
+    ggml_tensor * build_graph(ggml_context * ctx) override {
+        ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data());
+        ggml_tensor * b = ggml_new_tensor_4d(ctx, type, ne[0], ne[1], b_ne2, ne[3]);
+        ggml_tensor * out = ggml_concat(ctx, a, b);
+        return out;
+    }
+};
+
+// GGML_OP_ARGSORT
+struct test_argsort : public test_case {
+    const ggml_type type;
+    const std::array<int64_t, 4> ne;
+    ggml_sort_order order;
+
+    std::string vars() override {
+        return VARS_TO_STR3(type, ne, order);
+    }
+
+    test_argsort(ggml_type type = GGML_TYPE_F32,
+            std::array<int64_t, 4> ne = {16, 10, 10, 10},
+            ggml_sort_order order = GGML_SORT_ASC)
+        : type(type), ne(ne), order(order) {}
+
+    ggml_tensor * build_graph(ggml_context * ctx) override {
+        ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data());
+        ggml_tensor * out = ggml_argsort(ctx, a, order);
+        return out;
+    }
+
+    void initialize_tensors(ggml_context * ctx) override {
+        std::random_device rd;
+        std::default_random_engine rng(rd());
+        for (ggml_tensor * t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) {
+            if (t->type == GGML_TYPE_I32) {
+                // indices
+                std::vector<int> data(ggml_nelements(t));
+                for (int i = 0; i < ggml_nelements(t); i++) {
+                    data[i] = rand();
+                }
+                std::shuffle(data.begin(), data.end(), rng);
+                ggml_backend_tensor_set(t, data.data(), 0, ne[0]*ne[1]*ne[2]*ne[3] * sizeof(int));
+            } else if (t->type == GGML_TYPE_F32) {
+                // initialize with unique values to avoid ties
+                for (int64_t r = 0; r < ggml_nrows(t); r++) {
+                    std::vector<float> data(t->ne[0]);
+                    for (int i = 0; i < t->ne[0]; i++) {
+                        data[i] = i;
+                    }
+                    std::shuffle(data.begin(), data.end(), rng);
+                    ggml_backend_tensor_set(t, data.data(), r * t->nb[1], t->ne[0] * sizeof(float));
+                }
+            } else {
+                GGML_ASSERT(false);
+            }
+        }
+    }
+};
+
+// GGML_OP_SUM_ROWS
+struct test_sum_rows : public test_case {
+    const ggml_type type;
+    const std::array<int64_t, 4> ne;
+
+    std::string vars() override {
+        return VARS_TO_STR2(type, ne);
+    }
+
+    test_sum_rows(ggml_type type = GGML_TYPE_F32,
+            std::array<int64_t, 4> ne = {10, 10, 10, 10})
+        : type(type), ne(ne) {}
+
+    ggml_tensor * build_graph(ggml_context * ctx) override {
+        ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data());
+        ggml_tensor * out = ggml_sum_rows(ctx, a);
+        return out;
+    }
+};
+
+// 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;
+    const int n_experts_per_tok;
+    const int n_tokens;
+    const int n_embd;
+    const int n_ff;
+
+    std::string op_desc(ggml_tensor * t) override {
+        return "MOE";
+
+        GGML_UNUSED(t);
+    }
+
+    std::string vars() override {
+        return VARS_TO_STR5(n_experts, n_experts_per_tok, n_tokens, n_embd, n_ff);
+    }
+
+    test_moe(int n_experts = 8, int n_experts_per_tok = 2, int n_tokens = 1, int n_embd = 4096, int n_ff = 14336)
+        : n_experts(n_experts), n_experts_per_tok(n_experts_per_tok), n_tokens(n_tokens), n_embd(n_embd), n_ff(n_ff) {
+    }
+
+    ggml_tensor * build_graph(ggml_context * ctx) override {
+        ggml_tensor * ffn_gate_inp = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_experts);
+
+        std::vector<ggml_tensor *> ffn_up_exp(n_experts);
+        std::vector<ggml_tensor *> ffn_gate_exp(n_experts);
+        std::vector<ggml_tensor *> ffn_down_exp(n_experts);
+
+        for (int i = 0; i < n_experts; ++i) {
+            ffn_up_exp[i] = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_ff);
+            ffn_gate_exp[i] = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_ff);
+            ffn_down_exp[i] = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_ff, n_embd);
+        }
+
+        ggml_tensor * cur = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_tokens);
+
+        ggml_tensor * logits = ggml_mul_mat(ctx, ffn_gate_inp, cur);
+        ggml_tensor * probs = ggml_soft_max_ext(ctx, logits, nullptr, 1.0f/sqrtf(n_embd));
+
+        // select experts
+        ggml_tensor * selected_experts = ggml_top_k(ctx, probs, n_experts_per_tok);
+
+        ggml_tensor * weights = ggml_get_rows(ctx,
+                ggml_reshape_3d(ctx, probs, 1, n_experts, n_tokens), selected_experts);
+
+        weights = ggml_reshape_2d(ctx, weights, n_experts_per_tok, n_tokens);
+
+        ggml_tensor * weights_sum = ggml_sum_rows(ctx, weights);
+
+        weights = ggml_div(ctx, weights, weights_sum);
+
+        // compute expert outputs
+        ggml_tensor * moe_out = nullptr;
+
+        for (int i = 0; i < n_experts_per_tok; ++i) {
+            ggml_tensor * cur_expert;
+
+            ggml_tensor * cur_up = ggml_mul_mat_id(ctx, ffn_up_exp.data(), n_experts, selected_experts, i, cur);
+
+            ggml_tensor * cur_gate = ggml_mul_mat_id(ctx, ffn_gate_exp.data(), n_experts, selected_experts, i, cur);
+
+            cur_gate = ggml_silu(ctx, cur_gate);
+
+            cur_expert = ggml_mul(ctx, cur_up, cur_gate);
+
+            cur_expert = ggml_mul_mat_id(ctx, ffn_down_exp.data(), n_experts, selected_experts, i, cur_expert);
+
+            cur_expert = ggml_mul(ctx, cur_expert,
+                    ggml_view_2d(ctx, weights, 1, n_tokens, weights->nb[1], i*weights->nb[0]));
+
+            if (i == 0) {
+                moe_out = cur_expert;
+            } else {
+                moe_out = ggml_add(ctx, moe_out, cur_expert);
+            }
+        }
+
+        cur = moe_out;
+
+        return cur;
+    }
+};
+
+static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op_name) {
+    std::vector<std::unique_ptr<test_case>> test_cases;
+
+    const ggml_type all_types[] = {
+        GGML_TYPE_F32, GGML_TYPE_F16,
+        GGML_TYPE_Q4_0, GGML_TYPE_Q4_1,
+        GGML_TYPE_Q5_0, GGML_TYPE_Q5_1,
+        GGML_TYPE_Q8_0,
+        GGML_TYPE_Q2_K, GGML_TYPE_Q3_K,
+        GGML_TYPE_Q4_K, GGML_TYPE_Q5_K,
+        GGML_TYPE_Q6_K
+    };
+
+    // unary ops
+    for (int op = 0; op < GGML_UNARY_OP_COUNT; op++) {
+        test_cases.emplace_back(new test_unary((ggml_unary_op) op));
+    }
+
+    test_cases.emplace_back(new test_get_rows(GGML_TYPE_F32, 1, 8, 2, 1, false));
+    for (ggml_type type : all_types) {
+        for (int b : {1, 7}) {
+            for (bool v : {false, true}) {
+                test_cases.emplace_back(new test_get_rows(type, 256, 5, 4, b, v));
+            }
+        }
+    }
+
+    test_cases.emplace_back(new test_repeat(GGML_TYPE_F32, {10, 10, 10, 10}, {1, 1, 1, 1}));
+    test_cases.emplace_back(new test_repeat(GGML_TYPE_F32, {10, 10, 10, 10}, {2, 1, 1, 1}));
+    test_cases.emplace_back(new test_repeat(GGML_TYPE_F32, {10, 10, 10, 10}, {1, 2, 1, 1}));
+    test_cases.emplace_back(new test_repeat(GGML_TYPE_F32, {10, 10, 10, 10}, {1, 1, 2, 1}));
+    test_cases.emplace_back(new test_repeat(GGML_TYPE_F32, {10, 10, 10, 10}, {1, 1, 1, 2}));
+
+    test_cases.emplace_back(new test_dup());
+
+    for (ggml_type type : all_types) {
+       test_cases.emplace_back(new test_cpy(GGML_TYPE_F32, type, {256, 10, 10, 1}));
+    }
+
+    test_cases.emplace_back(new test_cont());
+
+    auto add_test_bin_bcast = [&](ggml_type type, std::array<int64_t, 4> ne, std::array<int, 4> nr) {
+        for (auto op : {ggml_add, ggml_mul, ggml_div}) {
+            test_cases.emplace_back(new test_bin_bcast(op, type, ne, nr));
+        }
+    };
+
+    add_test_bin_bcast(GGML_TYPE_F32, {1, 1, 8, 1}, {1, 1, 1, 1});
+    add_test_bin_bcast(GGML_TYPE_F32, {1, 1, 1, 1}, {32, 1, 1, 1});
+    add_test_bin_bcast(GGML_TYPE_F32, {1, 1, 320, 320}, {1, 1, 1, 1});
+    add_test_bin_bcast(GGML_TYPE_F32, {16, 10, 1, 1}, {1, 1, 1, 1});
+    add_test_bin_bcast(GGML_TYPE_F32, {16, 10, 10, 1}, {1, 1, 1, 1});
+    add_test_bin_bcast(GGML_TYPE_F32, {16, 10, 10, 10}, {1, 1, 1, 1});
+    add_test_bin_bcast(GGML_TYPE_F32, {16, 10, 10, 10}, {2, 1, 1, 1});
+    add_test_bin_bcast(GGML_TYPE_F32, {16, 10, 10, 10}, {1, 2, 1, 1});
+    add_test_bin_bcast(GGML_TYPE_F32, {16, 10, 10, 10}, {1, 1, 2, 1});
+    add_test_bin_bcast(GGML_TYPE_F32, {16, 10, 10, 10}, {1, 1, 1, 2});
+    add_test_bin_bcast(GGML_TYPE_F32, {16, 10, 10, 10}, {1, 1, 2, 2});
+    add_test_bin_bcast(GGML_TYPE_F32, {16, 10, 10, 10}, {1, 2, 2, 2});
+    add_test_bin_bcast(GGML_TYPE_F32, {16, 10, 10, 10}, {2, 2, 2, 2});
+
+    // stable diffusion
+    add_test_bin_bcast(GGML_TYPE_F32, {1280, 1, 1, 1}, {1, 1, 1, 1});
+    add_test_bin_bcast(GGML_TYPE_F32, {1280, 1, 1, 1}, {1, 16, 16, 1});
+    add_test_bin_bcast(GGML_TYPE_F32, {1280, 16, 16, 1}, {1, 1, 1, 1});
+    add_test_bin_bcast(GGML_TYPE_F32, {1280, 1, 1, 1}, {1, 256, 1, 1});
+    add_test_bin_bcast(GGML_TYPE_F32, {1, 1, 1280, 1}, {16, 16, 1, 1});
+    add_test_bin_bcast(GGML_TYPE_F32, {16, 16, 1280, 1}, {1, 1, 1, 1});
+    add_test_bin_bcast(GGML_TYPE_F32, {1, 1, 1920, 1}, {16, 16, 1, 1});
+    add_test_bin_bcast(GGML_TYPE_F32, {1, 1, 2560, 1}, {16, 16, 1, 1});
+    add_test_bin_bcast(GGML_TYPE_F32, {1, 1, 1280, 1}, {32, 32, 1, 1});
+    add_test_bin_bcast(GGML_TYPE_F32, {1, 1, 1920, 1}, {32, 32, 1, 1});
+    add_test_bin_bcast(GGML_TYPE_F32, {1, 1, 640, 1}, {32, 32, 1, 1});
+    add_test_bin_bcast(GGML_TYPE_F32, {5120, 1, 1, 1}, {1, 256, 1, 1});
+    add_test_bin_bcast(GGML_TYPE_F32, {640, 1, 1, 1}, {1, 1, 1, 1});
+    //add_test_bin_bcast(GGML_TYPE_F32, {3, 3, 2560, 1280}, {1, 1, 1, 1});
+    //add_test_bin_bcast(GGML_TYPE_F32, {3, 3, 2560, 1280}, {2, 1, 1, 1});
+
+    test_cases.emplace_back(new test_scale());
+
+    for (float eps : {1e-6f, 1e-5f, 1e-3f, 1e-1f}) {
+        test_cases.emplace_back(new test_norm(GGML_TYPE_F32, {64, 10, 10, 10}, eps));
+        test_cases.emplace_back(new test_rms_norm(GGML_TYPE_F32, {64, 10, 10, 10}, eps));
+    }
+
+    for (ggml_type type_a : all_types) {
+        for (ggml_type type_b : {GGML_TYPE_F32 /*, GGML_TYPE_F16 */}) {
+            // FIXME: CPU crashes on f16xf16
+            test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 1, 256, { 1,  1}, {1, 1}));
+            test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 1, 256, {10,  1}, {1, 1}));
+            test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 1, 256, {10,  1}, {2, 1}));
+            test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 1, 256, {10, 10}, {1, 1}));
+            test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 1, 256, {10, 10}, {2, 1}));
+            test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 1, 256, {10, 10}, {1, 2}));
+            test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 1, 256, {10, 10}, {2, 2}));
+
+            test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 16, 256, { 1,  1}, {1, 1}));
+            test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 16, 256, {10,  1}, {1, 1}));
+            test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 16, 256, {10,  1}, {2, 1}));
+            test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 16, 256, {10, 10}, {1, 1}));
+            test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 16, 256, {10, 10}, {2, 1}));
+            test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 16, 256, {10, 10}, {1, 2}));
+            test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 16, 256, {10, 10}, {2, 2}));
+        }
+    }
+
+    for (ggml_type type_a : all_types) {
+        for (ggml_type type_b : {GGML_TYPE_F32 /*, GGML_TYPE_F16 */}) {
+            for (int n_mats : {2, 4, 8}) {
+                for (int id = 0; id < n_mats; id++) {
+                    for (bool v : {false, true}) {
+                        test_cases.emplace_back(new test_mul_mat_id(type_a, type_b, n_mats, id, 16, 16, 256, v));
+                    }
+                }
+            }
+        }
+    }
+
+    test_cases.emplace_back(new test_sqr());
+    test_cases.emplace_back(new test_clamp());
+
+    test_cases.emplace_back(new test_diag_mask_inf(GGML_TYPE_F32, {10, 10,  1,  1}, 5));
+    test_cases.emplace_back(new test_diag_mask_inf(GGML_TYPE_F32, {10, 10, 10,  1}, 5));
+    test_cases.emplace_back(new test_diag_mask_inf(GGML_TYPE_F32, {10, 10, 10, 10}, 5));
+
+    test_cases.emplace_back(new test_soft_max());
+
+    for (ggml_type type : {GGML_TYPE_F32, GGML_TYPE_F16}) {
+        test_cases.emplace_back(new test_rope(type, {128,  32, 10, 1}, 128, 0, 512)); // llama 7B
+        test_cases.emplace_back(new test_rope(type, {128,  40, 10, 1}, 128, 0, 512)); // llama 13B
+        test_cases.emplace_back(new test_rope(type, {128,  52, 10, 1}, 128, 0, 512)); // llama 30B
+        test_cases.emplace_back(new test_rope(type, {128,  64, 10, 1}, 128, 0, 512)); // llama 65B
+        test_cases.emplace_back(new test_rope(type, { 64,   1, 10, 1},  64, 2, 512)); // neox (falcon 7B)
+        test_cases.emplace_back(new test_rope(type, { 64,  71, 10, 1},  64, 2, 512)); // neox (falcon 7B)
+        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());
+    test_cases.emplace_back(new test_im2col());
+    test_cases.emplace_back(new test_concat());
+
+    for (ggml_sort_order order : {GGML_SORT_ASC, GGML_SORT_DESC}) {
+        test_cases.emplace_back(new test_argsort(GGML_TYPE_F32, {8, 1, 1, 1}, order));
+        test_cases.emplace_back(new test_argsort(GGML_TYPE_F32, {16, 10, 10, 10}, order));
+    }
+
+    test_cases.emplace_back(new test_sum_rows());
+    test_cases.emplace_back(new test_upscale());
+    test_cases.emplace_back(new test_group_norm());
+    test_cases.emplace_back(new test_acc());
+    test_cases.emplace_back(new test_pad());
+    test_cases.emplace_back(new test_leaky_relu());
+
+#if !defined(__SANITIZE_THREAD__)
+    // FIXME: these tests use too much memory with thread sanitizer
+    test_cases.emplace_back(new test_moe(8, 2, 1, 4096, 8*1024));
+    //test_cases.emplace_back(new test_moe(8, 2, 8, 4096, 14336));
+#endif
+
+    // run tests
+    if (mode == MODE_TEST) {
+        ggml_backend_t backend_cpu = ggml_backend_cpu_init();
+
+        size_t n_ok = 0;
+        for (auto & test : test_cases) {
+            if (test->eval(backend, backend_cpu, op_name)) {
+                n_ok++;
+            }
+        }
+        printf("  %zu/%zu tests passed\n", n_ok, test_cases.size());
+
+        ggml_backend_free(backend_cpu);
+
+        return n_ok == test_cases.size();
+    }
+
+    if (mode == MODE_PERF) {
+        for (auto & test : test_cases) {
+            test->eval_perf(backend, op_name);
+        }
+        return true;
+    }
+
+    GGML_ASSERT(false);
+    return false;
+}
+
+static void usage(char ** argv) {
+    printf("Usage: %s [mode] [-o op] [-b backend]\n", argv[0]);
+    printf("  valid modes are: test (compare with CPU backend for correctness) or perf (performance evaluation)\n");
+    printf("  op names are as given by ggml_op_desc()\n");
+}
+
+int main(int argc, char ** argv) {
+    test_mode mode = MODE_TEST;
+    const char * op_name = NULL;
+    const char * backend = NULL;
+
+    for (int i = 1; i < argc; i++) {
+        if (strcmp(argv[i], "test") == 0) {
+            mode = MODE_TEST;
+        } else if (strcmp(argv[i], "perf") == 0) {
+            mode = MODE_PERF;
+        } else if (strcmp(argv[i], "-o") == 0) {
+            if (i + 1 < argc) {
+                op_name = argv[++i];
+            } else {
+                usage(argv);
+                return 1;
+            }
+        } else if (strcmp(argv[i], "-b") == 0) {
+            if (i + 1 < argc) {
+                backend = argv[++i];
+            } else {
+                usage(argv);
+                return 1;
+            }
+        } else {
+            usage(argv);
+            return 1;
+        }
+    }
+
+    // enumerate backends
+    printf("Testing %zu backends\n\n", ggml_backend_reg_get_count());
+
+    size_t n_ok = 0;
+
+    for (size_t i = 0; i < ggml_backend_reg_get_count(); i++) {
+        printf("Backend %zu/%zu (%s)\n", i + 1, ggml_backend_reg_get_count(), ggml_backend_reg_get_name(i));
+
+        if (backend != NULL && strcmp(backend, ggml_backend_reg_get_name(i)) != 0) {
+            printf("  Skipping\n");
+            n_ok++;
+            continue;
+        }
+
+        ggml_backend_t backend = ggml_backend_reg_init_backend(i, NULL);
+        GGML_ASSERT(backend != NULL);
+        printf("  Backend name: %s\n", ggml_backend_name(backend));
+
+        bool ok = test_backend(backend, mode, op_name);
+
+        printf("  Backend %s: ", ggml_backend_name(backend));
+        if (ok) {
+            printf("\033[1;32mOK\033[0m\n");
+            n_ok++;
+        } else {
+            printf("\033[1;31mFAIL\033[0m\n");
+        }
+
+        printf("\n");
+
+        ggml_backend_free(backend);
+    }
+
+    printf("%zu/%zu backends passed\n", n_ok, ggml_backend_reg_get_count());
+
+    if (n_ok != ggml_backend_reg_get_count()) {
+        printf("\033[1;31mFAIL\033[0m\n");
+        return 1;
+    }
+
+    printf("\033[1;32mOK\033[0m\n");
+    return 0;
+}
diff --git a/tests/test-grad0.cpp b/tests/test-grad0.cpp
index 7fe9154dd..14914def5 100644
--- a/tests/test-grad0.cpp
+++ b/tests/test-grad0.cpp
@@ -1,4 +1,4 @@
-#define _CRT_SECURE_NO_DEPRECATE // Disables ridiculous "unsafe" warnigns on Windows
+#define _CRT_SECURE_NO_DEPRECATE // Disables ridiculous "unsafe" warnings on Windows
 #include "ggml.h"
 
 #include <cmath>
@@ -881,19 +881,19 @@ int main(int argc, const char ** argv) {
         // scale
         {
             srand(seed);
-            const int nargs = 2;
+            const int nargs = 1;
 
             int64_t ne2[4];
             ne2[0] = 1;
 
             for (int ndims = 1; ndims <= 2; ++ndims) {
-                x[1] = get_random_tensor_f32(ctx0, 1, ne2, -1.0f, 1.0f);
                 x[0] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
 
-                ggml_set_param(ctx0, x[0]);
-                ggml_set_param(ctx0, x[1]);
+                const float s = -1.0f + 2.0f*frand();
 
-                struct ggml_tensor * f = ggml_sum(ctx0, ggml_scale(ctx0, x[0], x[1]));
+                ggml_set_param(ctx0, x[0]);
+
+                struct ggml_tensor * f = ggml_sum(ctx0, ggml_scale(ctx0, x[0], s));
 
                 check_gradient("scale", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, INFINITY);
             }
@@ -1395,7 +1395,7 @@ int main(int argc, const char ** argv) {
                                                 ggml_add1(ctx0,
                                                     ggml_scale(ctx0,
                                                         ggml_soft_max(ctx0, x[0]),
-                                                        ggml_new_f32(ctx0, 1.0f - eps)),
+                                                        1.0f - eps),
                                                     ggml_new_f32(ctx0, eps))));
 
                 check_gradient("softmax", ctx0, x, f, ndims, nargs, 1e-3f, 2e-1f, INFINITY);
diff --git a/tests/test-quantize-perf.cpp b/tests/test-quantize-perf.cpp
index 88fac0e23..09d410b7f 100644
--- a/tests/test-quantize-perf.cpp
+++ b/tests/test-quantize-perf.cpp
@@ -117,7 +117,7 @@ static void usage(char * argv[]) {
     printf("  --size SIZE           set test size, divisible by 32 (L1_SIZE:%d)\n", L1_SIZE);
     printf("  -3                    use size as L1, L2, L3 sizes (L1:%d L2:%d L3:%d)\n", L1_SIZE, L2_SIZE, L3_SIZE);
     printf("  -4                    use size as L1, L2, L3, MEM sizes (L1:%d L2:%d L3:%d MEM:%d)\n", L1_SIZE, L2_SIZE, L3_SIZE, MEM_SIZE);
-    printf("  --op OP               set test opration as quantize_row_q_reference, quantize_row_q, dequantize_row_q,\n");
+    printf("  --op OP               set test operation as quantize_row_q_reference, quantize_row_q, dequantize_row_q,\n");
     printf("                        quantize_row_q_dot, vec_dot_q (all)\n");
     printf("  --type TYPE           set test type as");
     for (int i = 0; i < GGML_TYPE_COUNT; i++) {
@@ -202,7 +202,7 @@ int main(int argc, char * argv[]) {
             }
             int alignment = std::stoi(argv[i]);
             if (alignment < 0 || alignment > MAX_ALIGNMENT) {
-            fprintf(stderr, "error: aligment-offset must be less than %d\n", MAX_ALIGNMENT);
+            fprintf(stderr, "error: alignment-offset must be less than %d\n", MAX_ALIGNMENT);
                 invalid_param = true;
                 break;
             }
@@ -286,7 +286,7 @@ int main(int argc, char * argv[]) {
                         qfns.from_float_reference(test_data1, test_q1, size);
                         return test_q1[0];
                     };
-                    size_t quantized_size = 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");