llama : adapt to F16 KQ_pos

Thanks Johannes again for the tip

.devops/nix/docker.nix

@@ -1,37 +0,0 @@
-{
-  lib,
-  dockerTools,
-  buildEnv,
-  llama-cpp,
-  interactive ? true,
-  coreutils,
-}:
-
-# A tar that can be fed into `docker load`:
-#
-# $ nix build .#llamaPackages.docker
-# $ docker load < result
-
-# For details and variations cf.
-# - https://nixos.org/manual/nixpkgs/unstable/#ssec-pkgs-dockerTools-buildLayeredImage
-# - https://discourse.nixos.org/t/a-faster-dockertools-buildimage-prototype/16922
-# - https://nixery.dev/
-
-# Approximate (compressed) sizes, at the time of writing, are:
-#
-# .#llamaPackages.docker: 125M;
-# .#llamaPackagesCuda.docker: 537M;
-# .#legacyPackages.aarch64-linux.llamaPackagesXavier.docker: 415M.
-
-dockerTools.buildLayeredImage {
-  name = llama-cpp.pname;
-  tag = "latest";
-
-  contents =
-    [ llama-cpp ]
-    ++ lib.optionals interactive [
-      coreutils
-      dockerTools.binSh
-      dockerTools.caCertificates
-    ];
-}

.devops/nix/package.nix

@@ -255,11 +255,11 @@ effectiveStdenv.mkDerivation (
       # Configurations we don't want even the CI to evaluate. Results in the
       # "unsupported platform" messages. This is mostly a no-op, because
       # cudaPackages would've refused to evaluate anyway.
-      badPlatforms = optionals (useCuda || useOpenCL) lib.platforms.darwin;
+      badPlatforms = optionals (useCuda || useOpenCL || useVulkan) lib.platforms.darwin;
 
       # Configurations that are known to result in build failures. Can be
       # overridden by importing Nixpkgs with `allowBroken = true`.
-      broken = (useMetalKit && !effectiveStdenv.isDarwin);
+      broken = (useMetalKit && !effectiveStdenv.isDarwin) || (useVulkan && effectiveStdenv.isDarwin);
 
       description = "Inference of LLaMA model in pure C/C++${descriptionSuffix}";
       homepage = "https://github.com/ggerganov/llama.cpp/";

.devops/nix/scope.nix

@@ -12,8 +12,5 @@ lib.makeScope newScope (
   self: {
     inherit llamaVersion;
     llama-cpp = self.callPackage ./package.nix { };
-    docker = self.callPackage ./docker.nix { };
-    docker-min = self.callPackage ./docker.nix { interactive = false; };
-    sif = self.callPackage ./sif.nix { };
   }
 )

.devops/nix/sif.nix

@@ -1,27 +0,0 @@
-{
-  lib,
-  singularity-tools,
-  llama-cpp,
-  bashInteractive,
-  interactive ? false,
-}:
-
-let
-    optionalInt = cond: x: if cond then x else 0;
-in
-singularity-tools.buildImage rec {
-  inherit (llama-cpp) name;
-  contents = [ llama-cpp ] ++ lib.optionals interactive [ bashInteractive ];
-
-  # These are excessive (but safe) for most variants. Building singularity
-  # images requires superuser privileges, so we build them inside a VM in a
-  # writable image of pre-determined size.
-  #
-  # ROCm is currently affected by https://github.com/NixOS/nixpkgs/issues/276846
-  #
-  # Expected image sizes:
-  # - cpu/blas: 150M,
-  # - cuda, all gencodes: 560M,
-  diskSize = 4096 + optionalInt llama-cpp.useRocm 16384;
-  memSize = diskSize;
-}

.github/ISSUE_TEMPLATE/bug.md

@@ -7,5 +7,3 @@ assignees: ''
 ---
 
 Please include information about your system, the steps to reproduce the bug, and the version of llama.cpp that you are using. If possible, please provide a minimal code example that reproduces the bug.
-
-If the bug concerns the server, please try to reproduce it first using the [server test scenario framework](https://github.com/ggerganov/llama.cpp/tree/master/examples/server/tests).

.github/workflows/build.yml

@@ -669,7 +669,8 @@ jobs:
         run: |
           cd examples/llama.android
 
-          ./gradlew build --no-daemon
+          # Skip armeabi-v7a for now (https://github.com/llvm/llvm-project/issues/65820).
+          ./gradlew build --no-daemon -Pskip-armeabi-v7a
 
 #  freeBSD-latest:
 #    runs-on: macos-12

.github/workflows/nix-ci-aarch64.yml

@@ -19,6 +19,7 @@ on:
 
 jobs:
   nix-build-aarch64:
+    if: ${{ vars.CACHIX_NAME != '' }}
     runs-on: ubuntu-latest
     steps:
     - name: Checkout repository
@@ -36,8 +37,8 @@ jobs:
         extra-conf: |
           extra-platforms = aarch64-linux
           extra-system-features = nixos-test kvm
-          extra-substituters = https://llama-cpp.cachix.org https://cuda-maintainers.cachix.org
-          extra-trusted-public-keys = llama-cpp.cachix.org-1:H75X+w83wUKTIPSO1KWy9ADUrzThyGs8P5tmAbkWhQc= cuda-maintainers.cachix.org-1:0dq3bujKpuEPMCX6U4WylrUDZ9JyUG0VpVZa7CNfq5E=
+          extra-substituters = https://${{ vars.CACHIX_NAME }}.cachix.org https://cuda-maintainers.cachix.org
+          extra-trusted-public-keys = ${{ vars.CACHIX_PUBLIC_KEY }} cuda-maintainers.cachix.org-1:0dq3bujKpuEPMCX6U4WylrUDZ9JyUG0VpVZa7CNfq5E=
     - uses: DeterminateSystems/magic-nix-cache-action@v2
       with:
         upstream-cache: https://${{ matrix.cachixName }}.cachix.org
@@ -45,7 +46,7 @@ jobs:
       uses: cachix/cachix-action@v13
       with:
         authToken: '${{ secrets.CACHIX_AUTH_TOKEN }}'
-        name: llama-cpp
+        name: ${{ vars.CACHIX_NAME }}
     - name: Show all output paths
       run: >
           nix run github:nix-community/nix-eval-jobs

.github/workflows/nix-ci.yml

@@ -23,8 +23,8 @@ jobs:
       with:
         github-token: ${{ secrets.GITHUB_TOKEN }}
         extra-conf: |
-          extra-substituters = https://llama-cpp.cachix.org https://cuda-maintainers.cachix.org
-          extra-trusted-public-keys = llama-cpp.cachix.org-1:H75X+w83wUKTIPSO1KWy9ADUrzThyGs8P5tmAbkWhQc= cuda-maintainers.cachix.org-1:0dq3bujKpuEPMCX6U4WylrUDZ9JyUG0VpVZa7CNfq5E=
+          extra-substituters = https://${{ vars.CACHIX_NAME }}.cachix.org https://cuda-maintainers.cachix.org
+          extra-trusted-public-keys = ${{ vars.CACHIX_PUBLIC_KEY }} cuda-maintainers.cachix.org-1:0dq3bujKpuEPMCX6U4WylrUDZ9JyUG0VpVZa7CNfq5E=
     - uses: DeterminateSystems/magic-nix-cache-action@v2
       with:
         upstream-cache: https://${{ matrix.cachixName }}.cachix.org
@@ -37,6 +37,7 @@ jobs:
           --flake
           ".#packages.$(nix eval --raw --impure --expr builtins.currentSystem)"
   nix-build:
+    if: ${{ vars.CACHIX_NAME != '' }}
     strategy:
       fail-fast: false
       matrix:
@@ -50,8 +51,8 @@ jobs:
       with:
         github-token: ${{ secrets.GITHUB_TOKEN }}
         extra-conf: |
-          extra-substituters = https://llama-cpp.cachix.org https://cuda-maintainers.cachix.org
-          extra-trusted-public-keys = llama-cpp.cachix.org-1:H75X+w83wUKTIPSO1KWy9ADUrzThyGs8P5tmAbkWhQc= cuda-maintainers.cachix.org-1:0dq3bujKpuEPMCX6U4WylrUDZ9JyUG0VpVZa7CNfq5E=
+          extra-substituters = https://${{ vars.CACHIX_NAME }}.cachix.org https://cuda-maintainers.cachix.org
+          extra-trusted-public-keys = ${{ vars.CACHIX_PUBLIC_KEY }} cuda-maintainers.cachix.org-1:0dq3bujKpuEPMCX6U4WylrUDZ9JyUG0VpVZa7CNfq5E=
     - uses: DeterminateSystems/magic-nix-cache-action@v2
       with:
         upstream-cache: https://${{ matrix.cachixName }}.cachix.org
@@ -59,7 +60,7 @@ jobs:
       uses: cachix/cachix-action@v13
       with:
         authToken: '${{ secrets.CACHIX_AUTH_TOKEN }}'
-        name: llama-cpp
+        name: ${{ vars.CACHIX_NAME }}
     - name: Build
       run: >
           nix run github:Mic92/nix-fast-build

.github/workflows/server.yml

@@ -1,127 +0,0 @@
-# Server build and tests
-name: Server
-
-on:
-  workflow_dispatch: # allows manual triggering
-  push:
-    branches:
-      - master
-      - test/server-add-ci-test # FIXME remove
-    paths: ['.github/workflows/**', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.swift', '**/*.m', 'examples/server/**.*']
-  pull_request:
-    types: [opened, synchronize, reopened]
-    paths: ['**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.swift', '**/*.m', 'examples/server/**.*']
-
-jobs:
-  server:
-    runs-on: ubuntu-latest
-
-    strategy:
-      matrix:
-        build: [noavx, avx2, avx, avx512, cublas, clblast, openblas, kompute, vulkan]
-        sanitizer: [ADDRESS, THREAD, UNDEFINED]
-        build_type: [Debug, Release]
-        include:
-          - build: 'noavx'
-            defines: '-DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DLLAMA_AVX=OFF -DLLAMA_AVX2=OFF -DLLAMA_FMA=OFF'
-            image: ubuntu:latest
-          - build: 'avx2'
-            defines: '-DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON'
-            image: ubuntu:latest
-          - build: 'avx'
-            defines: '-DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DLLAMA_AVX2=OFF'
-            image: ubuntu:latest
-          - build: 'avx512'
-            defines: '-DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DLLAMA_AVX512=ON'
-            image: ubuntu:latest
-            experimental: true
-          - build: 'cublas'
-            defines: '-DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DLLAMA_CUBLAS=ON'
-            image: nvidia/cuda:12.3.1-devel-ubuntu22.04
-            arch_not_available: true # require nvidia docker engine
-          - build: 'clblast'
-            defines: '-DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DLLAMA_CLBLAST=ON'
-            image: ubuntu:latest
-            arch_not_available: true
-          - build: 'openblas'
-            defines: '-DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DLLAMA_BLAS=ON -DLLAMA_BLAS_VENDOR=OpenBLAS'
-            image: ubuntu:latest
-          - build: 'kompute'
-            defines: '-DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DLLAMA_KOMPUTE=ON -DKOMPUTE_OPT_DISABLE_VULKAN_VERSION_CHECK=ON'
-            image: ubuntu:latest
-            arch_not_available: true
-          - build: 'vulkan'
-            defines: '-DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DLLAMA_VULKAN=ON'
-            image: ubuntu:latest
-            arch_not_available: true
-
-    container:
-      image: ${{ matrix.image }}
-      ports:
-        - 8888
-      options: --cpus 4
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v3
-
-      - name: Dependencies
-        id: depends
-        run: |
-          apt-get update
-          apt-get -y install \
-            build-essential \
-            pkg-config \
-            git \
-            cmake \
-            python3-pip \
-            wget \
-            psmisc
-
-      - name: Download CLBlast
-        id: get_clblast
-        if: ${{ matrix.build == 'clblast' }}
-        run: |
-          apt install -y libclblast-dev
-
-      - name: Download OpenBLAS
-        id: get_openblas
-        if: ${{ matrix.build == 'openblas' }}
-        run: |
-          apt-get -y install libopenblas-dev
-
-      - name: Install Vulkan SDK
-        id: get_vulkan
-        if: ${{ matrix.build == 'kompute' || matrix.build == 'vulkan' }}
-        run: |
-          wget -qO- https://packages.lunarg.com/lunarg-signing-key-pub.asc | tee /etc/apt/trusted.gpg.d/lunarg.asc
-          wget -qO /etc/apt/sources.list.d/lunarg-vulkan-jammy.list http://packages.lunarg.com/vulkan/lunarg-vulkan-jammy.list
-          apt-get update
-          apt-get -y install vulkan-sdk
-
-      - name: Build
-        id: cmake_build
-        run: |
-          mkdir build
-          cd build
-          cmake .. -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON -DCMAKE_BUILD_TYPE=${{ matrix.build_type }} ${{ matrix.defines }}
-          cmake --build . --config ${{ matrix.build_type }} -j $(nproc) --target server
-
-      - name: Tests dependencies
-        id: test_dependencies
-        run: |
-          pip install -r examples/server/tests/requirements.txt
-
-      - name: Download models
-        id: download_models
-        run: |
-          cd examples/server/tests
-          ../../../scripts/hf.sh --repo ggml-org/models --file tinyllamas/stories260K.gguf
-
-      - name: Tests
-        id: server_integration_test
-        continue-on-error: ${{ matrix.experimental || matrix.arch_not_available }}
-        run: |
-          cd examples/server/tests
-          PORT=8888 ./tests.sh

CMakeLists.txt

@@ -110,7 +110,6 @@ option(LLAMA_VULKAN_RUN_TESTS                "llama: run Vulkan tests"
 option(LLAMA_METAL                           "llama: use Metal"                                 ${LLAMA_METAL_DEFAULT})
 option(LLAMA_METAL_NDEBUG                    "llama: disable Metal debugging"                   OFF)
 option(LLAMA_METAL_SHADER_DEBUG              "llama: compile Metal with -fno-fast-math"         OFF)
-option(LLAMA_METAL_EMBED_LIBRARY             "llama: embed Metal library"                       OFF)
 option(LLAMA_KOMPUTE                         "llama: use Kompute"                               OFF)
 option(LLAMA_MPI                             "llama: use MPI"                                   OFF)
 option(LLAMA_QKK_64                          "llama: use super-block size of 64 for k-quants"   OFF)
@@ -146,6 +145,14 @@ set(THREADS_PREFER_PTHREAD_FLAG ON)
 find_package(Threads REQUIRED)
 include(CheckCXXCompilerFlag)
 
+if (LLAMA_FATAL_WARNINGS)
+    if (CMAKE_CXX_COMPILER_ID MATCHES "GNU" OR CMAKE_CXX_COMPILER_ID MATCHES "Clang")
+        add_compile_options(-Werror)
+    elseif (CMAKE_CXX_COMPILER_ID STREQUAL "MSVC")
+        add_compile_options(/WX)
+    endif()
+endif()
+
 # enable libstdc++ assertions for debug builds
 if (CMAKE_SYSTEM_NAME MATCHES "Linux")
     add_compile_definitions($<$<CONFIG:Debug>:_GLIBCXX_ASSERTIONS>)
@@ -202,29 +209,6 @@ if (LLAMA_METAL)
     # copy ggml-metal.metal to bin directory
     configure_file(ggml-metal.metal ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.metal COPYONLY)
 
-    if (LLAMA_METAL_EMBED_LIBRARY)
-        enable_language(ASM)
-        add_compile_definitions(GGML_METAL_EMBED_LIBRARY)
-
-        set(METALLIB_SOURCE "${CMAKE_SOURCE_DIR}/ggml-metal.metal")
-        file(MAKE_DIRECTORY "${CMAKE_BINARY_DIR}/autogenerated")
-        set(EMBED_METALLIB_ASSEMBLY "${CMAKE_BINARY_DIR}/autogenerated/ggml-embed-metallib.s")
-
-        add_custom_command(
-            OUTPUT ${EMBED_METALLIB_ASSEMBLY}
-            COMMAND echo ".section __DATA,__ggml_metallib" > ${EMBED_METALLIB_ASSEMBLY}
-            COMMAND echo ".globl _ggml_metallib_start" >> ${EMBED_METALLIB_ASSEMBLY}
-            COMMAND echo "_ggml_metallib_start:" >> ${EMBED_METALLIB_ASSEMBLY}
-            COMMAND echo ".incbin \\\"${METALLIB_SOURCE}\\\"" >> ${EMBED_METALLIB_ASSEMBLY}
-            COMMAND echo ".globl _ggml_metallib_end" >> ${EMBED_METALLIB_ASSEMBLY}
-            COMMAND echo "_ggml_metallib_end:" >> ${EMBED_METALLIB_ASSEMBLY}
-            DEPENDS ${METALLIB_SOURCE}
-            COMMENT "Generate assembly for embedded Metal library"
-        )
-
-        set(GGML_SOURCES_METAL ${GGML_SOURCES_METAL} ${EMBED_METALLIB_ASSEMBLY})
-    endif()
-
     if (LLAMA_METAL_SHADER_DEBUG)
         # custom command to do the following:
         #   xcrun -sdk macosx metal    -fno-fast-math -c ggml-metal.metal -o ggml-metal.air
@@ -763,24 +747,15 @@ function(get_flags CCID CCVER)
     set(GF_CXX_FLAGS ${CXX_FLAGS} PARENT_SCOPE)
 endfunction()
 
-if (LLAMA_FATAL_WARNINGS)
-    if (CMAKE_CXX_COMPILER_ID MATCHES "GNU" OR CMAKE_CXX_COMPILER_ID MATCHES "Clang")
-        list(APPEND C_FLAGS   -Werror)
-        list(APPEND CXX_FLAGS -Werror)
-    elseif (CMAKE_CXX_COMPILER_ID STREQUAL "MSVC")
-        add_compile_options(/WX)
-    endif()
-endif()
-
 if (LLAMA_ALL_WARNINGS)
     if (NOT MSVC)
-        list(APPEND WARNING_FLAGS -Wall -Wextra -Wpedantic -Wcast-qual -Wno-unused-function)
-        list(APPEND C_FLAGS       -Wshadow -Wstrict-prototypes -Wpointer-arith -Wmissing-prototypes
-                                  -Werror=implicit-int -Werror=implicit-function-declaration)
-        list(APPEND CXX_FLAGS     -Wmissing-declarations -Wmissing-noreturn)
+        set(WARNING_FLAGS -Wall -Wextra -Wpedantic -Wcast-qual -Wno-unused-function)
+        set(C_FLAGS       -Wshadow -Wstrict-prototypes -Wpointer-arith -Wmissing-prototypes
+                          -Werror=implicit-int -Werror=implicit-function-declaration)
+        set(CXX_FLAGS     -Wmissing-declarations -Wmissing-noreturn)
 
-        list(APPEND C_FLAGS   ${WARNING_FLAGS})
-        list(APPEND CXX_FLAGS ${WARNING_FLAGS})
+        set(C_FLAGS   ${WARNING_FLAGS} ${C_FLAGS})
+        set(CXX_FLAGS ${WARNING_FLAGS} ${CXX_FLAGS})
 
         get_flags(${CMAKE_CXX_COMPILER_ID} ${CMAKE_CXX_COMPILER_VERSION})
 
@@ -798,10 +773,6 @@ set(CUDA_CXX_FLAGS "")
 if (LLAMA_CUBLAS)
     set(CUDA_FLAGS -use_fast_math)
 
-    if (LLAMA_FATAL_WARNINGS)
-        list(APPEND CUDA_FLAGS -Werror all-warnings)
-    endif()
-
     if (LLAMA_ALL_WARNINGS AND NOT MSVC)
         set(NVCC_CMD ${CMAKE_CUDA_COMPILER} .c)
         if (NOT CMAKE_CUDA_HOST_COMPILER STREQUAL "")
@@ -936,16 +907,10 @@ if (CMAKE_OSX_ARCHITECTURES STREQUAL "arm64" OR CMAKE_GENERATOR_PLATFORM_LWR STR
             list(APPEND ARCH_FLAGS -mfpu=neon-fp-armv8 -mno-unaligned-access)
         endif()
         if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "armv7")
-            if ("${CMAKE_SYSTEM_NAME}" STREQUAL "Android")
-                # Android armeabi-v7a
-                list(APPEND ARCH_FLAGS -mfpu=neon-vfpv4 -mno-unaligned-access -funsafe-math-optimizations)
-            else()
-                # Raspberry Pi 2
-                list(APPEND ARCH_FLAGS -mfpu=neon-fp-armv8 -mno-unaligned-access -funsafe-math-optimizations)
-            endif()
+            # Raspberry Pi 2
+            list(APPEND ARCH_FLAGS -mfpu=neon-fp-armv8 -mno-unaligned-access -funsafe-math-optimizations)
         endif()
         if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "armv8")
-            # Android arm64-v8a
             # Raspberry Pi 3, 4, Zero 2 (32-bit)
             list(APPEND ARCH_FLAGS -mno-unaligned-access)
         endif()

Makefile

@@ -97,10 +97,9 @@ endif
 #
 
 # keep standard at C11 and C++11
-MK_CPPFLAGS  = -I. -Icommon
-MK_CFLAGS    = -std=c11   -fPIC
-MK_CXXFLAGS  = -std=c++11 -fPIC
-MK_NVCCFLAGS = -std=c++11
+MK_CPPFLAGS = -I. -Icommon
+MK_CFLAGS   = -std=c11   -fPIC
+MK_CXXFLAGS = -std=c++11 -fPIC
 
 # -Ofast tends to produce faster code, but may not be available for some compilers.
 ifdef LLAMA_FAST
@@ -173,7 +172,7 @@ ifdef LLAMA_DEBUG
 	MK_LDFLAGS  += -g
 
 	ifeq ($(UNAME_S),Linux)
-		MK_CPPFLAGS += -D_GLIBCXX_ASSERTIONS
+		MK_CXXFLAGS += -Wp,-D_GLIBCXX_ASSERTIONS
 	endif
 else
 	MK_CPPFLAGS += -DNDEBUG
@@ -217,7 +216,7 @@ MK_CFLAGS    += $(WARN_FLAGS) -Wshadow -Wstrict-prototypes -Wpointer-arith -Wmis
 MK_CXXFLAGS  += $(WARN_FLAGS) -Wmissing-declarations -Wmissing-noreturn
 
 ifeq ($(LLAMA_FATAL_WARNINGS),1)
-	MK_CFLAGS   += -Werror
+	MK_CFLAGS += -Werror
 	MK_CXXFLAGS += -Werror
 endif
 
@@ -385,9 +384,6 @@ ifdef LLAMA_CUBLAS
 	MK_LDFLAGS   += -lcuda -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 -L/usr/lib/wsl/lib
 	OBJS         += ggml-cuda.o
 	MK_NVCCFLAGS += -use_fast_math
-ifdef LLAMA_FATAL_WARNINGS
-	MK_NVCCFLAGS += -Werror all-warnings
-endif # LLAMA_FATAL_WARNINGS
 ifndef JETSON_EOL_MODULE_DETECT
 	MK_NVCCFLAGS += --forward-unknown-to-host-compiler
 endif # JETSON_EOL_MODULE_DETECT
@@ -446,9 +442,9 @@ ifdef LLAMA_CUDA_CCBIN
 endif
 ggml-cuda.o: ggml-cuda.cu ggml-cuda.h
 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) $(CPPFLAGS) -Xcompiler "$(CUDA_CXXFLAGS)" -c $< -o $@
+	$(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) $(NVCCFLAGS) $(CPPFLAGS) -Xcompiler "$(CUDA_CXXFLAGS)" -c $< -o $@
+	$(NVCC) $(NVCCFLAGS) -Xcompiler "$(CUDA_CXXFLAGS)" -c $< -o $@
 endif # JETSON_EOL_MODULE_DETECT
 endif # LLAMA_CUBLAS
 
@@ -533,29 +529,11 @@ ifdef LLAMA_METAL
 ifdef LLAMA_METAL_NDEBUG
 	MK_CPPFLAGS += -DGGML_METAL_NDEBUG
 endif
-ifdef LLAMA_METAL_EMBED_LIBRARY
-	MK_CPPFLAGS += -DGGML_METAL_EMBED_LIBRARY
-	OBJS        += ggml-metal-embed.o
-endif
 endif # LLAMA_METAL
 
 ifdef LLAMA_METAL
 ggml-metal.o: ggml-metal.m ggml-metal.h
 	$(CC) $(CFLAGS) -c $< -o $@
-
-ifdef LLAMA_METAL_EMBED_LIBRARY
-ggml-metal-embed.o: ggml-metal.metal
-	@echo "Embedding Metal library"
-	$(eval TEMP_ASSEMBLY=$(shell mktemp))
-	@echo ".section __DATA, __ggml_metallib" > $(TEMP_ASSEMBLY)
-	@echo ".globl _ggml_metallib_start" >> $(TEMP_ASSEMBLY)
-	@echo "_ggml_metallib_start:" >> $(TEMP_ASSEMBLY)
-	@echo ".incbin \"$<\"" >> $(TEMP_ASSEMBLY)
-	@echo ".globl _ggml_metallib_end" >> $(TEMP_ASSEMBLY)
-	@echo "_ggml_metallib_end:" >> $(TEMP_ASSEMBLY)
-	@$(AS) $(TEMP_ASSEMBLY) -o $@
-	@rm -f ${TEMP_ASSEMBLY}
-endif
 endif # LLAMA_METAL
 
 ifdef LLAMA_MPI
@@ -567,10 +545,9 @@ GF_CC := $(CC)
 include scripts/get-flags.mk
 
 # combine build flags with cmdline overrides
-override CPPFLAGS  := $(MK_CPPFLAGS) $(CPPFLAGS)
-override CFLAGS    := $(CPPFLAGS) $(MK_CFLAGS) $(GF_CFLAGS) $(CFLAGS)
-BASE_CXXFLAGS      := $(MK_CXXFLAGS) $(CXXFLAGS)
-override CXXFLAGS  := $(BASE_CXXFLAGS) $(HOST_CXXFLAGS) $(GF_CXXFLAGS) $(CPPFLAGS)
+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)
 
@@ -597,7 +574,7 @@ $(info I CC:        $(shell $(CC)   --version | head -n 1))
 $(info I CXX:       $(shell $(CXX)  --version | head -n 1))
 ifdef LLAMA_CUBLAS
 $(info I NVCC:      $(shell $(NVCC) --version | tail -n 1))
-CUDA_VERSION := $(shell $(NVCC) --version | grep -oP 'release (\K[0-9]+\.[0-9])')
+CUDA_VERSION := $(shell nvcc --version | grep -oP 'release (\K[0-9]+\.[0-9])')
 ifeq ($(shell awk -v "v=$(CUDA_VERSION)" 'BEGIN { print (v < 11.7) }'),1)
 ifndef CUDA_DOCKER_ARCH
 ifndef CUDA_POWER_ARCH
@@ -719,7 +696,7 @@ save-load-state: examples/save-load-state/save-load-state.cpp ggml.o llama.o $(C
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-server: examples/server/server.cpp examples/server/oai.hpp examples/server/utils.hpp 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 examples/llava/llava.h examples/llava/llava.cpp common/stb_image.h ggml.o llama.o $(COMMON_DEPS) grammar-parser.o $(OBJS)
+server: examples/server/server.cpp examples/server/oai.hpp examples/server/utils.hpp 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) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) -c examples/llava/clip.cpp -o $(call GET_OBJ_FILE, examples/llava/clip.cpp) -Wno-cast-qual
 	$(CXX) $(CXXFLAGS) -Iexamples/server $(filter-out %.h %.hpp $< examples/llava/clip.cpp,$^) $(call GET_OBJ_FILE, $<) $(call GET_OBJ_FILE, examples/llava/clip.cpp) -o $@ $(LDFLAGS) $(LWINSOCK2)

README-sycl.md

@@ -272,7 +272,7 @@ Please install [Visual Studio](https://visualstudio.microsoft.com/) which impact
 
 a. Please follow the procedure in [Get the Intel® oneAPI Base Toolkit ](https://www.intel.com/content/www/us/en/developer/tools/oneapi/base-toolkit.html).
 
-Recommend to install to default folder: **C:\Program Files (x86)\Intel\oneAPI**.
+Recommend to install to default folder: **/opt/intel/oneapi**.
 
 Following guide uses the default folder as example. If you use other folder, please modify the following guide info with your folder.
 

README.md

@@ -10,9 +10,13 @@ Inference of Meta's [LLaMA](https://arxiv.org/abs/2302.13971) model (and others)
 
 ### Hot topics
 
-- Support for chat templates: [Wiki (contributions welcome)](https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template)
-- Support for Gemma models: https://github.com/ggerganov/llama.cpp/pull/5631
-- Non-linear quantization IQ4_NL: https://github.com/ggerganov/llama.cpp/pull/5590
+- Remove LLAMA_MAX_DEVICES and LLAMA_SUPPORTS_GPU_OFFLOAD: https://github.com/ggerganov/llama.cpp/pull/5240
+- Incoming backends: https://github.com/ggerganov/llama.cpp/discussions/5138
+  - [SYCL backend](README-sycl.md) is ready (1/28/2024), support Linux/Windows in Intel GPUs (iGPU, Arc/Flex/Max series)
+- New SOTA quantized models, including pure 2-bits: https://huggingface.co/ikawrakow
+- 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
 - Looking for contributions to improve and maintain the `server` example: https://github.com/ggerganov/llama.cpp/issues/4216
 
 ----
@@ -103,7 +107,6 @@ Typically finetunes of the base models below are supported as well.
 - [x] [Orion 14B](https://github.com/ggerganov/llama.cpp/pull/5118)
 - [x] [InternLM2](https://huggingface.co/models?search=internlm2)
 - [x] [CodeShell](https://github.com/WisdomShell/codeshell)
-- [x] [Gemma](https://ai.google.dev/gemma)
 
 **Multimodal models:**
 
@@ -114,9 +117,6 @@ Typically finetunes of the base models below are supported as well.
 - [x] [MobileVLM 1.7B/3B models](https://huggingface.co/models?search=mobileVLM)
 - [x] [Yi-VL](https://huggingface.co/models?search=Yi-VL)
 
-**HTTP server**
-
-[llama.cpp web server](./examples/server) is a lightweight [OpenAI API](https://github.com/openai/openai-openapi) compatible HTTP server that can be used to serve local models and easily connect them to existing clients.
 
 **Bindings:**
 
@@ -145,7 +145,6 @@ Unless otherwise noted these projects are open-source with permissive licensing:
 - [nat/openplayground](https://github.com/nat/openplayground)
 - [Faraday](https://faraday.dev/) (proprietary)
 - [LMStudio](https://lmstudio.ai/) (proprietary)
-- [LocalAI](https://github.com/mudler/LocalAI) (MIT)
 - [LostRuins/koboldcpp](https://github.com/LostRuins/koboldcpp) (AGPL)
 - [Mozilla-Ocho/llamafile](https://github.com/Mozilla-Ocho/llamafile)
 - [nomic-ai/gpt4all](https://github.com/nomic-ai/gpt4all)
@@ -157,8 +156,6 @@ Unless otherwise noted these projects are open-source with permissive licensing:
 - [pythops/tenere](https://github.com/pythops/tenere) (AGPL)
 - [semperai/amica](https://github.com/semperai/amica)
 - [withcatai/catai](https://github.com/withcatai/catai)
-- [Mobile-Artificial-Intelligence/maid](https://github.com/Mobile-Artificial-Intelligence/maid) (MIT)
-- [Msty](https://msty.app) (proprietary)
 
 ---
 

build.zig

@@ -123,7 +123,6 @@ pub fn build(b: *std.build.Builder) !void {
     const grammar_parser = make.obj("grammar-parser", "common/grammar-parser.cpp");
     const train = make.obj("train", "common/train.cpp");
     const clip = make.obj("clip", "examples/llava/clip.cpp");
-    const llava = make.obj("llava", "examples/llava/llava.cpp");
 
     _ = make.exe("main", "examples/main/main.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common, buildinfo, sampling, console, grammar_parser });
     _ = make.exe("quantize", "examples/quantize/quantize.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common, buildinfo });
@@ -132,7 +131,7 @@ pub fn build(b: *std.build.Builder) !void {
     _ = make.exe("finetune", "examples/finetune/finetune.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common, buildinfo, train });
     _ = make.exe("train-text-from-scratch", "examples/train-text-from-scratch/train-text-from-scratch.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common, buildinfo, train });
 
-    const server = make.exe("server", "examples/server/server.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common, buildinfo, sampling, grammar_parser, clip, llava });
+    const server = make.exe("server", "examples/server/server.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common, buildinfo, sampling, grammar_parser, clip });
     if (server.target.isWindows()) {
         server.linkSystemLibrary("ws2_32");
     }

common/common.cpp

@@ -295,9 +295,9 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
                 break;
             }
             std::string value(argv[i]);
-            /**/ if (value == "none")   { params.rope_scaling_type = LLAMA_ROPE_SCALING_TYPE_NONE; }
-            else if (value == "linear") { params.rope_scaling_type = LLAMA_ROPE_SCALING_TYPE_LINEAR; }
-            else if (value == "yarn")   { params.rope_scaling_type = LLAMA_ROPE_SCALING_TYPE_YARN; }
+            /**/ if (value == "none")   { params.rope_scaling_type = LLAMA_ROPE_SCALING_NONE; }
+            else if (value == "linear") { params.rope_scaling_type = LLAMA_ROPE_SCALING_LINEAR; }
+            else if (value == "yarn")   { params.rope_scaling_type = LLAMA_ROPE_SCALING_YARN; }
             else { invalid_param = true; break; }
         } else if (arg == "--rope-scale") {
             if (++i >= argc) {
@@ -630,11 +630,11 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
             }
             std::string arg_next = argv[i];
             if (arg_next == "none") {
-                params.split_mode = LLAMA_SPLIT_MODE_NONE;
+                params.split_mode = LLAMA_SPLIT_NONE;
             } else if (arg_next == "layer") {
-                params.split_mode = LLAMA_SPLIT_MODE_LAYER;
+                params.split_mode = LLAMA_SPLIT_LAYER;
             } else if (arg_next == "row") {
-                params.split_mode = LLAMA_SPLIT_MODE_ROW;
+                params.split_mode = LLAMA_SPLIT_ROW;
             } else {
                 invalid_param = true;
                 break;
@@ -837,15 +837,15 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
             sep++;
             if (strncmp(sep, "int:", 4) == 0) {
                 sep += 4;
-                kvo.tag = LLAMA_KV_OVERRIDE_TYPE_INT;
+                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_TYPE_FLOAT;
+                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_TYPE_BOOL;
+                kvo.tag = LLAMA_KV_OVERRIDE_BOOL;
                 if (std::strcmp(sep, "true") == 0) {
                     kvo.bool_value = true;
                 } else if (std::strcmp(sep, "false") == 0) {

common/common.h

@@ -61,7 +61,7 @@ struct gpt_params {
     float   p_split               = 0.1f;  // speculative decoding split probability
     int32_t n_gpu_layers          = -1;    // number of layers to store in VRAM (-1 - use default)
     int32_t n_gpu_layers_draft    = -1;    // number of layers to store in VRAM for the draft model (-1 - use default)
-    llama_split_mode split_mode   = LLAMA_SPLIT_MODE_LAYER; // how to split the model across GPUs
+    llama_split_mode split_mode   = LLAMA_SPLIT_LAYER; // how to split the model across GPUs
     int32_t main_gpu              = 0;     // the GPU that is used for scratch and small tensors
     float   tensor_split[128]     = {0};   // how split tensors should be distributed across GPUs
     int32_t n_beams               = 0;     // if non-zero then use beam search of given width.
@@ -75,7 +75,7 @@ struct gpt_params {
     float   yarn_beta_fast        = 32.0f; // YaRN low correction dim
     float   yarn_beta_slow        = 1.0f;  // YaRN high correction dim
     int32_t yarn_orig_ctx         = 0;     // YaRN original context length
-    int32_t rope_scaling_type     = LLAMA_ROPE_SCALING_TYPE_UNSPECIFIED;
+    int32_t rope_scaling_type     = LLAMA_ROPE_SCALING_UNSPECIFIED;
     ggml_numa_strategy numa       = GGML_NUMA_STRATEGY_DISABLED;
 
     // // sampling parameters

common/sampling.cpp

@@ -266,7 +266,7 @@ static llama_token llama_sampling_sample_impl(
             //    }
             //}
 
-            //LOG("sampled token: %5d: '%s'\n", id, llama_token_to_piece(ctx_main, id).c_str());
+            LOG("sampled token: %5d: '%s'\n", id, llama_token_to_piece(ctx_main, id).c_str());
         }
     }
 

common/train.cpp

@@ -31,7 +31,7 @@ struct train_state  * init_train_state() {
 
     state->opt = new struct ggml_opt_context;
     state->opt->ctx = NULL;
-    state->opt->params = ggml_opt_default_params(GGML_OPT_TYPE_ADAM);
+    state->opt->params = ggml_opt_default_params(GGML_OPT_ADAM);
     state->opt->params.graph_size = LLAMA_TRAIN_MAX_NODES;
     state->opt->loss_after = 0.0f;
 
@@ -556,7 +556,7 @@ void load_opt_context_gguf(struct gguf_context * fctx, struct ggml_context * f_g
     std::string opt_type;
     GGUF_GET_KEY(fctx, opt_type, gguf_get_val_str, GGUF_TYPE_STRING, true, LLM_KV_OPTIMIZER_TYPE);
     if (opt_type == LLM_KV_OPTIMIZER_TYPE_ADAM) {
-        opt->params.type = GGML_OPT_TYPE_ADAM;
+        opt->params.type = GGML_OPT_ADAM;
 
         GGUF_GET_KEY(fctx, opt->adam.fx_best,          gguf_get_val_f32, GGUF_TYPE_FLOAT32, true, LLM_KV_OPTIMIZER_ADAM_BEST_LOSS);
         GGUF_GET_KEY(fctx, opt->adam.fx_prev,          gguf_get_val_f32, GGUF_TYPE_FLOAT32, true, LLM_KV_OPTIMIZER_ADAM_PREVIOUS_LOSS);
@@ -568,7 +568,7 @@ void load_opt_context_gguf(struct gguf_context * fctx, struct ggml_context * f_g
         copy_tensor_by_name(opt->adam.v,  f_ggml_ctx, LLM_TENSOR_OPTIMIZER_ADAM_SECOND_MOMENTS);
         copy_tensor_by_name(opt->adam.pf, f_ggml_ctx, LLM_TENSOR_OPTIMIZER_ADAM_PAST_LOSS_VALUES);
     } else if (opt_type == LLM_KV_OPTIMIZER_TYPE_LBFGS) {
-        opt->params.type = GGML_OPT_TYPE_LBFGS;
+        opt->params.type = GGML_OPT_LBFGS;
 
         GGUF_GET_KEY(fctx, opt->params.lbfgs.m,         gguf_get_val_u32, GGUF_TYPE_UINT32,  true, LLM_KV_OPTIMIZER_LBFGS_APPROX_HESSIAN_COUNT);
         GGUF_GET_KEY(fctx, opt->lbfgs.fx_best,          gguf_get_val_f32, GGUF_TYPE_FLOAT32, true, LLM_KV_OPTIMIZER_LBFGS_BEST_LOSS);
@@ -603,7 +603,7 @@ void save_opt_context_gguf(struct gguf_context * fctx, struct ggml_opt_context *
     gguf_set_val_bool(fctx, LLM_KV_OPTIMIZER_JUST_INITIALIZED, opt->just_initialized);
 
     switch (opt->params.type) {
-        case GGML_OPT_TYPE_ADAM:
+        case GGML_OPT_ADAM:
             {
                 gguf_set_val_str(fctx, LLM_KV_OPTIMIZER_TYPE, LLM_KV_OPTIMIZER_TYPE_ADAM);
                 gguf_set_val_f32(fctx, LLM_KV_OPTIMIZER_ADAM_BEST_LOSS,            opt->adam.fx_best);
@@ -622,7 +622,7 @@ void save_opt_context_gguf(struct gguf_context * fctx, struct ggml_opt_context *
                     gguf_add_tensor(fctx, opt->adam.pf);
                 }
             } break;
-        case GGML_OPT_TYPE_LBFGS:
+        case GGML_OPT_LBFGS:
             {
                 gguf_set_val_str(fctx, LLM_KV_OPTIMIZER_TYPE, LLM_KV_OPTIMIZER_TYPE_LBFGS);
                 gguf_set_val_u32(fctx, LLM_KV_OPTIMIZER_LBFGS_APPROX_HESSIAN_COUNT, opt->params.lbfgs.m);

convert-hf-to-gguf.py

@@ -192,7 +192,7 @@ class Model:
             return RefactModel
         if model_architecture == "PersimmonForCausalLM":
             return PersimmonModel
-        if model_architecture in ("StableLmForCausalLM", "StableLMEpochForCausalLM", "LlavaStableLMEpochForCausalLM"):
+        if model_architecture in ("StableLMEpochForCausalLM", "LlavaStableLMEpochForCausalLM"):
             return StableLMModel
         if model_architecture == "QWenLMHeadModel":
             return QwenModel
@@ -218,8 +218,6 @@ class Model:
             return BertModel
         if model_architecture == "NomicBertModel":
             return NomicBertModel
-        if model_architecture == "GemmaForCausalLM":
-            return GemmaModel
         return Model
 
     def _is_model_safetensors(self) -> bool:
@@ -253,7 +251,7 @@ class Model:
             return gguf.MODEL_ARCH.REFACT
         if arch == "PersimmonForCausalLM":
             return gguf.MODEL_ARCH.PERSIMMON
-        if arch in ("StableLmForCausalLM", "StableLMEpochForCausalLM", "LlavaStableLMEpochForCausalLM"):
+        if arch in ("StableLMEpochForCausalLM", "LlavaStableLMEpochForCausalLM"):
             return gguf.MODEL_ARCH.STABLELM
         if arch == "QWenLMHeadModel":
             return gguf.MODEL_ARCH.QWEN
@@ -279,8 +277,6 @@ class Model:
             return gguf.MODEL_ARCH.BERT
         if arch == "NomicBertModel":
             return gguf.MODEL_ARCH.NOMIC_BERT
-        if arch == "GemmaForCausalLM":
-            return gguf.MODEL_ARCH.GEMMA
 
         raise NotImplementedError(f'Architecture "{arch}" not supported!')
 
@@ -622,6 +618,11 @@ class MPTModel(Model):
 
             self.gguf_writer.add_tensor(new_name, data)
 
+            # note: MPT output is tied to (same as) wte in original model;
+            # for easier implementation in llama.cpp it's duplicated in GGUF, though :/
+            if new_name == "token_embd.weight":
+                self.gguf_writer.add_tensor("output.weight", data)
+
 
 class OrionModel(Model):
     def set_vocab(self):
@@ -654,8 +655,6 @@ class OrionModel(Model):
         self.gguf_writer.add_feed_forward_length(self.hparams["intermediate_size"])
         self.gguf_writer.add_head_count(head_count)
         self.gguf_writer.add_head_count_kv(head_count_kv)
-        # note: config provides rms norm but it is actually layer norm
-        # ref:  https://huggingface.co/OrionStarAI/Orion-14B-Chat/blob/276a17221ce42beb45f66fac657a41540e71f4f5/modeling_orion.py#L570-L571
         self.gguf_writer.add_layer_norm_eps(self.hparams["rms_norm_eps"])
 
     def write_tensors(self):
@@ -1032,6 +1031,7 @@ class PersimmonModel(Model):
         self.gguf_writer.add_head_count_kv(head_count_kv)
         self.gguf_writer.add_rope_freq_base(self.hparams["rope_theta"])
         self.gguf_writer.add_layer_norm_eps(self.hparams["layer_norm_eps"])
+        self.gguf_writer.add_layer_norm_rms_eps(self.hparams["rms_norm_eps"])
 
     def set_vocab(self):
         self._set_vocab_sentencepiece()
@@ -1074,11 +1074,10 @@ class StableLMModel(Model):
         self.gguf_writer.add_embedding_length(hparams["hidden_size"])
         self.gguf_writer.add_block_count(block_count)
         self.gguf_writer.add_feed_forward_length(hparams["intermediate_size"])
-        rotary_factor = self.find_hparam(["partial_rotary_factor", "rope_pct"])
-        self.gguf_writer.add_rope_dimension_count(int(rotary_factor * (hparams["hidden_size"] // hparams["num_attention_heads"])))
+        self.gguf_writer.add_rope_dimension_count(int(hparams["rope_pct"] * (hparams["hidden_size"] // hparams["num_attention_heads"])))
         self.gguf_writer.add_head_count(hparams["num_attention_heads"])
         self.gguf_writer.add_parallel_residual(hparams["use_parallel_residual"] if "use_parallel_residual" in hparams else True)
-        self.gguf_writer.add_layer_norm_eps(self.find_hparam(["layer_norm_eps", "norm_eps"]))
+        self.gguf_writer.add_layer_norm_eps(1e-5)
 
 
 class MixtralModel(Model):
@@ -1786,63 +1785,6 @@ class NomicBertModel(BertModel):
             yield name, data
 
 
-class GemmaModel(Model):
-    def set_vocab(self):
-        self._set_vocab_sentencepiece()
-
-    def set_gguf_parameters(self):
-        hparams = self.hparams
-        block_count = hparams["num_hidden_layers"]
-
-        self.gguf_writer.add_name(self.dir_model.name)
-        self.gguf_writer.add_context_length(hparams["max_position_embeddings"])
-        self.gguf_writer.add_embedding_length(hparams["hidden_size"])
-        self.gguf_writer.add_block_count(block_count)
-        self.gguf_writer.add_feed_forward_length(hparams["intermediate_size"])
-        self.gguf_writer.add_head_count(hparams["num_attention_heads"])
-        self.gguf_writer.add_head_count_kv(self.hparams["num_key_value_heads"] if "num_key_value_heads" in hparams else hparams["num_attention_heads"])
-        self.gguf_writer.add_layer_norm_rms_eps(self.hparams["rms_norm_eps"])
-        self.gguf_writer.add_key_length(hparams["head_dim"])
-        self.gguf_writer.add_value_length(hparams["head_dim"])
-        self.gguf_writer.add_file_type(self.ftype)
-
-    def write_tensors(self):
-        block_count = self.hparams.get("n_layers", self.hparams.get("num_hidden_layers", self.hparams.get("n_layer")))
-        tensor_map = gguf.get_tensor_name_map(self.model_arch, block_count)
-
-        for name, data_torch in self.get_tensors():
-            # ref: https://github.com/huggingface/transformers/blob/fc37f38915372c15992b540dfcbbe00a916d4fc6/src/transformers/models/gemma/modeling_gemma.py#L89
-            if name.endswith("norm.weight"):
-                data_torch = data_torch + 1
-
-            old_dtype = data_torch.dtype
-
-            # convert any unsupported data types to float32
-            if data_torch.dtype not in (torch.float16, torch.float32):
-                data_torch = data_torch.to(torch.float32)
-
-            data = data_torch.squeeze().numpy()
-
-            # map tensor names
-            new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
-            if new_name is None:
-                print(f"Can not map tensor {name!r}")
-                sys.exit()
-
-            n_dims = len(data.shape)
-            data_dtype = data.dtype
-
-            data = data.astype(np.float32)
-
-            # if f16 desired, convert any float32 2-dim weight tensors to float16
-            if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
-                data = data.astype(np.float16)
-
-            print(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
-
-            self.gguf_writer.add_tensor(new_name, data)
-
-
 ###### CONVERSION LOGIC ######
 
 

examples/baby-llama/baby-llama.cpp

@@ -1547,7 +1547,7 @@ int main(int argc, char ** argv) {
 
         float error_before_opt = ggml_get_f32_1d(e, 0);
 
-        struct ggml_opt_params opt_params_lbfgs = ggml_opt_default_params(GGML_OPT_TYPE_LBFGS);
+        struct ggml_opt_params opt_params_lbfgs = ggml_opt_default_params(GGML_OPT_LBFGS);
         opt_params_lbfgs.print_forward_graph = false;
         opt_params_lbfgs.print_backward_graph = false;
         opt_params_lbfgs.lbfgs.n_iter = 16;

examples/batched-bench/batched-bench.cpp

@@ -105,7 +105,7 @@ int main(int argc, char ** argv) {
 
     ctx_params.seed      = 1234;
     ctx_params.n_ctx     = n_kv_max;
-    ctx_params.n_batch   = 512;
+    ctx_params.n_batch   = 2048;
     ctx_params.mul_mat_q = mmq;
 
     ctx_params.n_threads       = params.n_threads;

examples/finetune/finetune.cpp

@@ -1531,7 +1531,7 @@ int main(int argc, char ** argv) {
     lora.hparams.n_rank_output         = n_rank_output;
 
     // set opt params from command line
-    opt->params = ggml_opt_default_params(GGML_OPT_TYPE_ADAM);
+    opt->params = ggml_opt_default_params(GGML_OPT_ADAM);
     opt->params.print_forward_graph     = false;
     opt->params.print_backward_graph    = false;
     opt->params.graph_size              = LLAMA_TRAIN_MAX_NODES;

examples/infill/infill.cpp

@@ -447,8 +447,8 @@ int main(int argc, char ** argv) {
                 LOG("context full, swapping: n_past = %d, n_left = %d, n_ctx = %d, n_keep = %d, n_discard = %d\n",
                     n_past, n_left, n_ctx, params.n_keep, n_discard);
 
-                llama_kv_cache_seq_rm (ctx, 0, params.n_keep + 1            , params.n_keep + n_discard + 1);
-                llama_kv_cache_seq_add(ctx, 0, params.n_keep + 1 + n_discard, n_past, -n_discard);
+                llama_kv_cache_seq_rm   (ctx, 0, params.n_keep + 1            , params.n_keep + n_discard + 1);
+                llama_kv_cache_seq_shift(ctx, 0, params.n_keep + 1 + n_discard, n_past, -n_discard);
 
                 n_past -= n_discard;
 

examples/json-schema-to-grammar.py

@@ -87,21 +87,7 @@ class SchemaConverter:
         elif schema_type == 'array' and 'items' in schema:
             # TODO `prefixItems` keyword
             item_rule_name = self.visit(schema['items'], f'{name}{"-" if name else ""}item')
-            list_item_operator = f'("," space {item_rule_name})'
-            successive_items = ""
-            min_items = schema.get("minItems", 0)
-            if min_items > 0:
-               first_item = f"({item_rule_name})"
-               successive_items = list_item_operator * (min_items - 1)
-               min_items -= 1
-            else:
-               first_item = f"({item_rule_name})?"
-            max_items = schema.get("maxItems")
-            if max_items is not None and max_items > min_items:
-                successive_items += (list_item_operator + "?") * (max_items - min_items - 1)
-            else:
-                successive_items += list_item_operator + "*"
-            rule = f'"[" space {first_item} {successive_items} "]" space'
+            rule = f'"[" space ({item_rule_name} ("," space {item_rule_name})*)? "]" space'
             return self._add_rule(rule_name, rule)
 
         else:

examples/llama-bench/llama-bench.cpp

@@ -157,9 +157,9 @@ static const char * output_format_str(output_formats format) {
 
 static const char * split_mode_str(llama_split_mode mode) {
     switch (mode) {
-        case LLAMA_SPLIT_MODE_NONE:  return "none";
-        case LLAMA_SPLIT_MODE_LAYER: return "layer";
-        case LLAMA_SPLIT_MODE_ROW:   return "row";
+        case LLAMA_SPLIT_NONE:  return "none";
+        case LLAMA_SPLIT_LAYER: return "layer";
+        case LLAMA_SPLIT_ROW:   return "row";
         default: GGML_ASSERT(!"invalid split mode");
     }
 }
@@ -193,7 +193,7 @@ static const cmd_params cmd_params_defaults = {
     /* type_v        */ {GGML_TYPE_F16},
     /* n_threads     */ {get_num_physical_cores()},
     /* n_gpu_layers  */ {99},
-    /* split_mode    */ {LLAMA_SPLIT_MODE_LAYER},
+    /* split_mode    */ {LLAMA_SPLIT_LAYER},
     /* main_gpu      */ {0},
     /* no_kv_offload */ {false},
     /* mul_mat_q     */ {true},
@@ -358,11 +358,11 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
             for (const auto & m : p) {
                 llama_split_mode mode;
                 if (m == "none") {
-                    mode = LLAMA_SPLIT_MODE_NONE;
+                    mode = LLAMA_SPLIT_NONE;
                 } else if (m == "layer") {
-                    mode = LLAMA_SPLIT_MODE_LAYER;
+                    mode = LLAMA_SPLIT_LAYER;
                 } else if (m == "row") {
-                    mode = LLAMA_SPLIT_MODE_ROW;
+                    mode = LLAMA_SPLIT_ROW;
                 } else {
                     invalid_param = true;
                     break;

examples/llama.android/app/build.gradle.kts

@@ -21,8 +21,12 @@ android {
             useSupportLibrary = true
         }
         ndk {
-            // Add NDK properties if wanted, e.g.
-            // abiFilters += listOf("arm64-v8a")
+            // Workaround for https://github.com/llvm/llvm-project/issues/65820
+            // affecting armeabi-v7a. Skip armeabi-v7a when invoked with
+            // -Pskip-armeabi-v7a (e.g., ./gradlew build -Pskip-armeabi-v7a).
+            if (project.hasProperty("skip-armeabi-v7a")) {
+                abiFilters += listOf("arm64-v8a", "x86_64", "x86")
+            }
         }
         externalNativeBuild {
             cmake {

examples/llava/README.md

@@ -59,39 +59,14 @@ python ./convert.py ../llava-v1.5-7b --skip-unknown
 Now both the LLaMA part and the image encoder is in the `llava-v1.5-7b` directory.
 
 ## LLaVA 1.6 gguf conversion
-1) First clone a LLaVA 1.6 model:
-```console
-git clone https://huggingface.co/liuhaotian/llava-v1.6-vicuna-7b
-```
-2) Use `llava-surgery-v2.py` which also supports llava-1.5 variants pytorch as well as safetensor models:
-```console
-python examples/llava/llava-surgery-v2.py -C -m ../llava-v1.6-vicuna-7b/
-```
+
+1) Backup your pth/safetensor model files as llava-surgery modifies them
+2) Use `python llava-surgery-v2.py -C -m /path/to/hf-model` which also supports llava-1.5 variants pytorch as well as safetensor models:
 - you will find a llava.projector and a llava.clip file in your model directory
-3) Copy the llava.clip file into a subdirectory (like vit), rename it to pytorch_model.bin and add a fitting vit configuration to the directory:
-```console
-mkdir vit
-cp ../llava-v1.6-vicuna-7b/llava.clip vit/pytorch_model.bin
-cp ../llava-v1.6-vicuna-7b/llava.projector vit/
-curl -s -q https://huggingface.co/cmp-nct/llava-1.6-gguf/raw/main/config_vit.json -o vit/config.json
-```
-
-4) Create the visual gguf model:
-```console
-python ./examples/llava/convert-image-encoder-to-gguf.py -m vit --llava-projector vit/llava.projector --output-dir vit --clip-model-is-vision
-```
+3) Copy the llava.clip file into a subdirectory (like vit), rename it to pytorch_model.bin and add a fitting vit configuration to the directory (https://huggingface.co/cmp-nct/llava-1.6-gguf/blob/main/config_vit.json) and rename it to config.json.
+4) Create the visual gguf model: `python ./examples/llava/convert-image-encoder-to-gguf.py -m ../path/to/vit --llava-projector ../path/to/llava.projector --output-dir ../path/to/output --clip-model-is-vision`
 - This is similar to llava-1.5, the difference is that we tell the encoder that we are working with the pure vision model part of CLIP
-
-5) Then convert the model to gguf format:
-```console
-python ./convert.py ../llava-v1.6-vicuna-7b/ --skip-unknown
-```
-
-6) And finally we can run the llava-cli using the 1.6 model version:
-```console
-./llava-cli -m ../llava-v1.6-vicuna-7b/ggml-model-f16.gguf --mmproj vit/mmproj-model-f16.gguf --image some-image.jpg -c 4096
-```
-
+5) Everything else as usual: convert.py the hf model, quantize as needed
 **note** llava-1.6 needs more context than llava-1.5, at least 3000 is needed (just run it at -c 4096)
 **note** llava-1.6 greatly benefits from batched prompt processing (defaults work)
 

examples/llava/clip.cpp

@@ -616,9 +616,9 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
             KQ = ggml_soft_max_inplace(ctx0, KQ);
             struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V, KQ);
             KQV = ggml_reshape_4d(ctx0, KQV, d_head, num_positions, n_head, batch_size);
-            KQV = ggml_permute(ctx0, KQV, 0, 2, 1, 3);
+            KQV = ggml_cont(ctx0, ggml_permute(ctx0, KQV, 0, 2, 1, 3));
 
-            cur = ggml_cont_3d(ctx0, KQV, hidden_size, num_positions, batch_size);
+            cur = ggml_cpy(ctx0, KQV, ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, hidden_size, num_positions, batch_size));
         }
 
         // attention output

examples/llava/llava-surgery-v2.py

@@ -65,7 +65,9 @@ def clean_vision_tower_from_checkpoint(checkpoint_path):
         for name in clip_tensors:
             del checkpoint[name]
 
+        # Save the updated checkpoint
         checkpoint_path = checkpoint_path
+        save_model(checkpoint, checkpoint_path, file_type)
         return True
     return False
 
@@ -150,6 +152,16 @@ for name in first_mm_tensors:
 if len(projector) > 0:
     save_model(projector, f"{args.model}/llava.projector", 'pytorch')
 
+for name in mm_tensors:
+    del last_checkpoint[name]
+for name in first_mm_tensors:
+    del first_checkpoint[name]
+
+if len(mm_tensors) > 0:
+    save_model(last_checkpoint, projector_checkpoint_path, file_type)
+if len(first_mm_tensors) > 0:
+    save_model(first_checkpoint, newline_checkpoint_path, file_type)
+
 print("Done!")
 print(f"Now you can convert {args.model} to a a regular LLaMA GGUF file.")
 print(f"Also, use {args.model}/llava.projector to prepare a llava-encoder.gguf file.")

examples/llava/llava.cpp

@@ -152,7 +152,7 @@ static bool clip_llava_handle_patches(clip_ctx * ctx_clip, std::vector<float *>
 
     ggml_tensor * newline_tmp = clip_get_newline_tensor(ctx_clip);
     model.newline = ggml_new_tensor_1d(model.ctx, GGML_TYPE_F32, newline_tmp->ne[0]);
-    if (newline_tmp->backend != GGML_BACKEND_TYPE_CPU) {
+    if (newline_tmp->backend != GGML_BACKEND_CPU) {
         if (newline_tmp->buffer == NULL) {
             printf("newline_tmp tensor buffer is NULL\n");
         }
@@ -311,7 +311,7 @@ bool llava_validate_embed_size(const llama_context * ctx_llama, const clip_ctx *
     return true;
 }
 
-bool llava_image_embed_make_with_clip_img(clip_ctx * ctx_clip, int n_threads, const clip_image_u8 * img, float ** image_embd_out, int * n_img_pos_out) {
+static bool llava_image_embed_make_with_clip_img(clip_ctx * ctx_clip, int n_threads, const clip_image_u8 * img, float ** image_embd_out, int * n_img_pos_out) {
     float * image_embd = (float *)malloc(clip_embd_nbytes(ctx_clip)*6); // TODO: base on gridsize/llava model
     if (!image_embd) {
         fprintf(stderr, "Unable to allocate memory for image embeddings\n");

examples/llava/llava.h

@@ -31,8 +31,6 @@ struct llava_image_embed {
 /** sanity check for clip <-> llava embed size match */
 LLAVA_API bool llava_validate_embed_size(const llama_context * ctx_llama, const clip_ctx * ctx_clip);
 
-LLAVA_API bool llava_image_embed_make_with_clip_img(clip_ctx * ctx_clip, int n_threads, const clip_image_u8 * img, float ** image_embd_out, int * n_img_pos_out);
-
 /** build an image embed from image file bytes */
 LLAVA_API struct llava_image_embed * llava_image_embed_make_with_bytes(struct clip_ctx * ctx_clip, int n_threads, const unsigned char * image_bytes, int image_bytes_length);
 /** build an image embed from a path to an image filename */

examples/main/main.cpp

@@ -334,8 +334,6 @@ int main(int argc, char ** argv) {
     // number of tokens to keep when resetting context
     if (params.n_keep < 0 || params.n_keep > (int) embd_inp.size() || params.instruct || params.chatml) {
         params.n_keep = (int)embd_inp.size();
-    } else {
-        params.n_keep += add_bos; // always keep the BOS token
     }
 
     // prefix & suffix for instruct mode
@@ -385,8 +383,8 @@ int main(int argc, char ** argv) {
             }
         }
 
-        if (params.n_keep > add_bos) {
-            LOG_TEE("%s: static prompt based on n_keep: '", __func__);
+        if (params.n_keep > 0) {
+        LOG_TEE("%s: static prompt based on n_keep: '", __func__);
             for (int i = 0; i < params.n_keep; i++) {
                 LOG_TEE("%s", llama_token_to_piece(ctx, embd_inp[i]).c_str());
             }
@@ -542,14 +540,14 @@ int main(int argc, char ** argv) {
                         break;
                     }
 
-                    const int n_left    = n_past - params.n_keep;
+                    const int n_left    = n_past - params.n_keep - 1;
                     const int n_discard = n_left/2;
 
                     LOG("context full, swapping: n_past = %d, n_left = %d, n_ctx = %d, n_keep = %d, n_discard = %d\n",
                             n_past, n_left, n_ctx, params.n_keep, n_discard);
 
-                    llama_kv_cache_seq_rm (ctx, 0, params.n_keep            , params.n_keep + n_discard);
-                    llama_kv_cache_seq_add(ctx, 0, params.n_keep + n_discard, n_past, -n_discard);
+                    llama_kv_cache_seq_rm   (ctx, 0, params.n_keep + 1            , params.n_keep + n_discard + 1);
+                    llama_kv_cache_seq_shift(ctx, 0, params.n_keep + 1 + n_discard, n_past, -n_discard);
 
                     n_past -= n_discard;
 
@@ -576,9 +574,9 @@ int main(int argc, char ** argv) {
                     LOG("div:   [%6d, %6d] / %6d -> [%6d, %6d]\n", ga_i + ib*bd, ga_i + ib*bd + ga_w, ga_n, (ga_i + ib*bd)/ga_n, (ga_i + ib*bd + ga_w)/ga_n);
                     LOG("shift: [%6d, %6d] + %6d -> [%6d, %6d]\n", ga_i + ib*bd + ga_w, n_past + ib*bd, dd, ga_i + ib*bd + ga_w + dd, n_past + ib*bd + dd);
 
-                    llama_kv_cache_seq_add(ctx, 0, ga_i,                n_past,              ib*bd);
-                    llama_kv_cache_seq_div(ctx, 0, ga_i + ib*bd,        ga_i + ib*bd + ga_w, ga_n);
-                    llama_kv_cache_seq_add(ctx, 0, ga_i + ib*bd + ga_w, n_past + ib*bd,      dd);
+                    llama_kv_cache_seq_shift(ctx, 0, ga_i,                n_past,              ib*bd);
+                    llama_kv_cache_seq_div  (ctx, 0, ga_i + ib*bd,        ga_i + ib*bd + ga_w, ga_n);
+                    llama_kv_cache_seq_shift(ctx, 0, ga_i + ib*bd + ga_w, n_past + ib*bd,      dd);
 
                     n_past -= bd;
 

examples/passkey/passkey.cpp

@@ -126,7 +126,7 @@ int main(int argc, char ** argv) {
     const int n_batch     = ctx_params.n_batch;
     const int n_batch_grp = ctx_params.n_batch/n_grp;
 
-    LOG_TEE("\n%s: n_len = %d, n_ctx = %d, n_kv_req = %d, n_grp = %d, n_batch = %d, n_junk = %d, i_pos = %d\n", __func__, n_len, n_ctx, n_kv_req, n_grp, n_batch, n_junk, i_pos);
+    LOG_TEE("\n%s: n_len = %d, n_ctx = %d, n_kv_req = %d, n_grp = %d, n_batch = %d\n", __func__, n_len, n_ctx, n_kv_req, n_grp, n_batch);
 
     // print the prompt token-by-token
 
@@ -146,11 +146,10 @@ int main(int argc, char ** argv) {
             const int ib = i/n_batch - 1;
             const int bd = n_batch_grp*(n_grp - 1);
 
-            llama_kv_cache_seq_add (ctx, 0, n_past - n_batch,         n_past,         ib*bd);
-            llama_kv_cache_seq_div (ctx, 0, n_past - n_batch + ib*bd, n_past + ib*bd, n_grp);
-            llama_kv_cache_update  (ctx);
+            llama_kv_cache_seq_shift(ctx, 0, n_past - n_batch,         n_past,         ib*bd);
+            llama_kv_cache_seq_div  (ctx, 0, n_past - n_batch + ib*bd, n_past + ib*bd, n_grp);
 
-            n_past = llama_kv_cache_seq_pos_max(ctx, 0) + 1;
+            n_past -= bd;
         }
 
         llama_batch_clear(batch);
@@ -180,12 +179,10 @@ int main(int argc, char ** argv) {
 
         LOG_TEE("%s: shifting KV cache with %d\n", __func__, n_discard);
 
-        llama_kv_cache_seq_rm (ctx, 0, n_keep            , n_keep + n_discard);
-        llama_kv_cache_seq_add(ctx, 0, n_keep + n_discard, n_ctx,  -n_discard);
-        llama_kv_cache_defrag (ctx);
-        llama_kv_cache_update (ctx);
+        llama_kv_cache_seq_rm   (ctx, 0, n_keep            , n_keep + n_discard);
+        llama_kv_cache_seq_shift(ctx, 0, n_keep + n_discard, n_ctx,  -n_discard);
 
-        n_past = llama_kv_cache_seq_pos_max(ctx, 0) + 1;
+        n_past -= n_discard;
 
         llama_batch_clear(batch);
 
@@ -211,12 +208,10 @@ int main(int argc, char ** argv) {
         if (n_discard > 0) {
             LOG_TEE("%s: shifting KV cache with %d to free space for the answer\n", __func__, n_discard);
 
-            llama_kv_cache_seq_rm (ctx, 0, n_keep            , n_keep + n_discard);
-            llama_kv_cache_seq_add(ctx, 0, n_keep + n_discard, n_ctx,  -n_discard);
-            llama_kv_cache_defrag (ctx);
-            llama_kv_cache_update (ctx);
+            llama_kv_cache_seq_rm   (ctx, 0, n_keep            , n_keep + n_discard);
+            llama_kv_cache_seq_shift(ctx, 0, n_keep + n_discard, n_ctx,  -n_discard);
 
-            n_past = llama_kv_cache_seq_pos_max(ctx, 0) + 1;
+            n_past -= n_discard;
         }
     }
 

examples/quantize/quantize.cpp

@@ -27,14 +27,11 @@ static const std::vector<struct quant_option> QUANT_OPTIONS = {
     { "Q2_K",   LLAMA_FTYPE_MOSTLY_Q2_K,   " 2.63G, +0.6717 ppl @ LLaMA-v1-7B", },
     { "Q2_K_S", LLAMA_FTYPE_MOSTLY_Q2_K_S, " 2.16G, +9.0634 ppl @ LLaMA-v1-7B", },
     { "IQ3_XXS",LLAMA_FTYPE_MOSTLY_IQ3_XXS," 3.06 bpw quantization",            },
-    { "IQ3_S",  LLAMA_FTYPE_MOSTLY_IQ3_S,  " 3.44 bpw quantization",            },
-    { "IQ3_M",  LLAMA_FTYPE_MOSTLY_IQ3_M,  " 3.66 bpw quantization mix",         },
     { "Q3_K",   LLAMA_FTYPE_MOSTLY_Q3_K_M, "alias for Q3_K_M" },
     { "Q3_K_XS",LLAMA_FTYPE_MOSTLY_Q3_K_XS,"3-bit extra small quantization"   , },
     { "Q3_K_S", LLAMA_FTYPE_MOSTLY_Q3_K_S, " 2.75G, +0.5551 ppl @ LLaMA-v1-7B", },
     { "Q3_K_M", LLAMA_FTYPE_MOSTLY_Q3_K_M, " 3.07G, +0.2496 ppl @ LLaMA-v1-7B", },
     { "Q3_K_L", LLAMA_FTYPE_MOSTLY_Q3_K_L, " 3.35G, +0.1764 ppl @ LLaMA-v1-7B", },
-    { "IQ4_NL", LLAMA_FTYPE_MOSTLY_IQ4_NL, " 4.25 bpw non-linear quantization", },
     { "Q4_K",   LLAMA_FTYPE_MOSTLY_Q4_K_M, "alias for Q4_K_M", },
     { "Q4_K_S", LLAMA_FTYPE_MOSTLY_Q4_K_S, " 3.59G, +0.0992 ppl @ LLaMA-v1-7B", },
     { "Q4_K_M", LLAMA_FTYPE_MOSTLY_Q4_K_M, " 3.80G, +0.0532 ppl @ LLaMA-v1-7B", },

examples/server/README.md

@@ -1,20 +1,8 @@
-# LLaMA.cpp HTTP Server
+# llama.cpp/example/server
 
-Fast, lightweight, pure C/C++ HTTP server based on [httplib](https://github.com/yhirose/cpp-httplib), [nlohmann::json](https://github.com/nlohmann/json) and **llama.cpp**.
+This example demonstrates a simple HTTP API server and a simple web front end to interact with llama.cpp.
 
-Set of LLM REST APIs and a simple web front end to interact with llama.cpp.
-
-**Features:**
- * LLM inference of F16 and quantum models on GPU and CPU
- * [OpenAI API](https://github.com/openai/openai-openapi) compatible chat completions and embeddings routes
- * Parallel decoding with multi-user support
- * Continuous batching
- * Multimodal (wip)
- * Monitoring endpoints
-
-The project is under active development, and we are [looking for feedback and contributors](https://github.com/ggerganov/llama.cpp/issues/4216).
-
-**Command line options:**
+Command line options:
 
 - `--threads N`, `-t N`: Set the number of threads to use during generation.
 - `-tb N, --threads-batch N`: Set the number of threads to use during batch and prompt processing. If not specified, the number of threads will be set to the number of threads used for generation.
@@ -51,12 +39,8 @@ see https://github.com/ggerganov/llama.cpp/issues/1437
 - `--mmproj MMPROJ_FILE`: Path to a multimodal projector file for LLaVA.
 - `--grp-attn-n`: Set the group attention factor to extend context size through self-extend(default: 1=disabled), used together with group attention width `--grp-attn-w`
 - `--grp-attn-w`: Set the group attention width to extend context size through self-extend(default: 512), used together with group attention factor `--grp-attn-n`
-- `-n N, --n-predict N`: Set the maximum tokens to predict (default: -1)
+- `-n, --n-predict`: Set the maximum tokens to predict (default: -1)
 - `--slots-endpoint-disable`: To disable slots state monitoring endpoint. Slots state may contain user data, prompts included.
-- `--metrics`: enable prometheus `/metrics` compatible endpoint (default: disabled)
-- `--chat-template JINJA_TEMPLATE`: Set custom jinja chat template. This parameter accepts a string, not a file name (default: template taken from model's metadata). We only support [some pre-defined templates](https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template)
-- `--log-disable`: Output logs to stdout only, default: enabled.
-- `--log-format FORMAT`: Define the log output to FORMAT: json or text (default: json)
 
 ## Build
 
@@ -113,12 +97,6 @@ curl --request POST \
     --data '{"prompt": "Building a website can be done in 10 simple steps:","n_predict": 128}'
 ```
 
-## Advanced testing
-
-We implemented a [server test framework](./tests/README.md) using human-readable scenario.
-
-*Before submitting an issue, please try to reproduce it with this format.*
-
 ## Node JS Test
 
 You need to have [Node.js](https://nodejs.org/en) installed.
@@ -156,13 +134,10 @@ node index.js
 ## API Endpoints
 
 - **GET** `/health`: Returns the current state of the server:
-  - 503 -> `{"status": "loading model"}` if the model is still being loaded.
-  - 500 -> `{"status": "error"}` if the model failed to load.
-  - 200 -> `{"status": "ok", "slots_idle": 1, "slots_processing": 2 }` if the model is successfully loaded and the server is ready for further requests mentioned below.
-  - 200 -> `{"status": "no slot available", "slots_idle": 0, "slots_processing": 32}` if no slot are currently available.
-  - 503 -> `{"status": "no slot available", "slots_idle": 0, "slots_processing": 32}` if the query parameter `fail_on_no_slot` is provided and no slot are currently available.
-
-  If the query parameter `include_slots` is passed, `slots` field will contain internal slots data except if `--slots-endpoint-disable` is set.
+  - `{"status": "loading model"}` if the model is still being loaded.
+  - `{"status": "error"}` if the model failed to load.
+  - `{"status": "ok"}` if the model is successfully loaded and the server is ready for further requests mentioned below.
+  - `{"status": "no slot available", "slots_idle": 0, "slots_processing": 32}` if no slot are currently available
 
 - **POST** `/completion`: Given a `prompt`, it returns the predicted completion.
 
@@ -172,7 +147,7 @@ node index.js
 
     `temperature`: Adjust the randomness of the generated text (default: 0.8).
 
-    `dynatemp_range`: Dynamic temperature range. The final temperature will be in the range of `[temperature - dynatemp_range; temperature + dynatemp_range]` (default: 0.0, 0.0 = disabled).
+    `dynatemp_range`: Dynamic temperature range (default: 0.0, 0.0 = disabled).
 
     `dynatemp_exponent`: Dynamic temperature exponent (default: 1.0).
 
@@ -230,7 +205,7 @@ node index.js
 
     `slot_id`: Assign the completion task to an specific slot. If is -1 the task will be assigned to a Idle slot (default: -1)
 
-    `cache_prompt`: Re-use previously cached prompt from the last request if possible. This may prevent re-caching the prompt from scratch. (default: false)
+    `cache_prompt`: Save the prompt and generation for avoid reprocess entire prompt if a part of this isn't change (default: false)
 
     `system_prompt`: Change the system prompt (initial prompt of all slots), this is useful for chat applications. [See more](#change-system-prompt-on-runtime)
 
@@ -263,7 +238,7 @@ Notice that each `probs` is an array of length `n_probs`.
 
 - `content`: Completion result as a string (excluding `stopping_word` if any). In case of streaming mode, will contain the next token as a string.
 - `stop`: Boolean for use with `stream` to check whether the generation has stopped (Note: This is not related to stopping words array `stop` from input options)
-- `generation_settings`: The provided options above excluding `prompt` but including `n_ctx`, `model`. These options may differ from the original ones in some way (e.g. bad values filtered out, strings converted to tokens, etc.).
+- `generation_settings`: The provided options above excluding `prompt` but including `n_ctx`, `model`
 - `model`: The path to the model loaded with `-m`
 - `prompt`: The provided `prompt`
 - `stopped_eos`: Indicating whether the completion has stopped because it encountered the EOS token
@@ -472,18 +447,6 @@ Notice that each `probs` is an array of length `n_probs`.
 ]
 ```
 
-- **GET** `/metrics`: [Prometheus](https://prometheus.io/) compatible metrics exporter endpoint if `--metrics` is enabled:
-
-Available metrics:
-- `llamacpp:prompt_tokens_total`: Number of prompt tokens processed.
-- `llamacpp:tokens_predicted_total`: Number of generation tokens processed.
-- `llamacpp:prompt_tokens_seconds`: Average prompt throughput in tokens/s.
-- `llamacpp:predicted_tokens_seconds`: Average generation throughput in tokens/s.
-- `llamacpp:kv_cache_usage_ratio`: KV-cache usage. 1 means 100 percent usage.
-- `llamacpp:kv_cache_tokens`: KV-cache tokens.
-- `llamacpp:requests_processing`: Number of request processing.
-- `llamacpp:requests_deferred`: Number of request deferred.
-
 ## More examples
 
 ### Change system prompt on runtime

examples/server/oai.hpp

@@ -15,11 +15,13 @@
 using json = nlohmann::json;
 
 inline static json oaicompat_completion_params_parse(
-    const struct llama_model * model,
     const json &body, /* openai api json semantics */
     const std::string &chat_template)
 {
     json llama_params;
+    std::string formatted_prompt = chat_template == "chatml"
+        ? format_chatml(body["messages"])  // OpenAI 'messages' to chatml (with <|im_start|>,...)
+        : format_llama2(body["messages"]); // OpenAI 'messages' to llama2 (with [INST],...)
 
     llama_params["__oaicompat"] = true;
 
@@ -32,7 +34,7 @@ inline static json oaicompat_completion_params_parse(
     // https://platform.openai.com/docs/api-reference/chat/create
     llama_sampling_params default_sparams;
     llama_params["model"]             = json_value(body, "model", std::string("unknown"));
-    llama_params["prompt"]            = format_chat(model, chat_template, body["messages"]);
+    llama_params["prompt"]            = formatted_prompt;
     llama_params["cache_prompt"]      = json_value(body, "cache_prompt", false);
     llama_params["temperature"]       = json_value(body, "temperature", 0.0);
     llama_params["top_k"]             = json_value(body, "top_k", default_sparams.top_k);

examples/server/server.cpp

@@ -5,7 +5,6 @@
 #include "oai.hpp"
 
 #include "../llava/clip.h"
-#include "../llava/llava.h"
 
 #include "stb_image.h"
 
@@ -38,16 +37,14 @@ struct server_params
     std::string hostname = "127.0.0.1";
     std::vector<std::string> api_keys;
     std::string public_path = "examples/server/public";
-    std::string chat_template = "";
+    std::string chat_template = "chatml";
     int32_t port = 8080;
     int32_t read_timeout = 600;
     int32_t write_timeout = 600;
     bool slots_endpoint = true;
-    bool metrics_endpoint = false;
 };
 
 bool server_verbose = false;
-bool server_log_json = true;
 
 static size_t common_part(const std::vector<llama_token> &a, const std::vector<llama_token> &b)
 {
@@ -303,76 +300,12 @@ struct llama_client_slot
     }
 
     void print_timings() const {
-       char buffer[512];
-        double t_token = t_prompt_processing / num_prompt_tokens_processed;
-        double n_tokens_second = 1e3 / t_prompt_processing * num_prompt_tokens_processed;
-        sprintf(buffer, "prompt eval time     = %10.2f ms / %5d tokens (%8.2f ms per token, %8.2f tokens per second)",
-                t_prompt_processing, num_prompt_tokens_processed,
-                t_token, n_tokens_second);
-        LOG_INFO(buffer, {
-            {"slot_id",                     id},
-            {"task_id",                     task_id},
-            {"t_prompt_processing",         t_prompt_processing},
-            {"num_prompt_tokens_processed", num_prompt_tokens_processed},
-            {"t_token",                     t_token},
-            {"n_tokens_second",             n_tokens_second},
-        });
-
-        t_token = t_token_generation / n_decoded;
-        n_tokens_second = 1e3 / t_token_generation * n_decoded;
-        sprintf(buffer, "generation eval time = %10.2f ms / %5d runs   (%8.2f ms per token, %8.2f tokens per second)",
-                t_token_generation, n_decoded,
-                t_token, n_tokens_second);
-        LOG_INFO(buffer, {
-            {"slot_id",            id},
-            {"task_id",            task_id},
-            {"t_token_generation", t_token_generation},
-            {"n_decoded",          n_decoded},
-            {"t_token",            t_token},
-            {"n_tokens_second",    n_tokens_second},
-        });
-
-        sprintf(buffer, "          total time = %10.2f ms", t_prompt_processing + t_token_generation);
-        LOG_INFO(buffer, {
-            {"slot_id",             id},
-            {"task_id",             task_id},
-            {"t_prompt_processing", t_prompt_processing},
-            {"t_token_generation",  t_token_generation},
-            {"t_total",             t_prompt_processing + t_token_generation},
-        });
-    }
-};
-
-struct llama_metrics {
-    uint64_t n_prompt_tokens_processed_total = 0;
-    uint64_t n_tokens_predicted_total        = 0;
-
-    uint64_t n_prompt_tokens_processed = 0;
-    uint64_t t_prompt_processing       = 0;
-
-    uint64_t n_tokens_predicted       = 0;
-    uint64_t t_tokens_generation      = 0;
-
-
-    void on_prompt_eval(const llama_client_slot &slot) {
-        n_prompt_tokens_processed_total += slot.num_prompt_tokens_processed;
-
-        n_prompt_tokens_processed += slot.num_prompt_tokens_processed;
-        t_prompt_processing       += slot.t_prompt_processing;
-    }
-
-    void on_prediction(const llama_client_slot &slot) {
-        n_tokens_predicted_total += slot.n_decoded;
-
-        n_tokens_predicted  += slot.n_decoded;
-        t_tokens_generation += slot.t_token_generation;
-    }
-
-    void reset_bucket() {
-        n_prompt_tokens_processed = 0;
-        t_prompt_processing       = 0;
-        n_tokens_predicted        = 0;
-        t_tokens_generation       = 0;
+        LOG_TEE("\n");
+        LOG_TEE("%s: prompt eval time = %10.2f ms / %5d tokens (%8.2f ms per token, %8.2f tokens per second)\n",
+            __func__, t_prompt_processing, num_prompt_tokens_processed, t_prompt_processing / num_prompt_tokens_processed, 1e3 / t_prompt_processing * num_prompt_tokens_processed);
+        LOG_TEE("%s:        eval time = %10.2f ms / %5d runs   (%8.2f ms per token, %8.2f tokens per second)\n",
+            __func__, t_token_generation, n_decoded,t_token_generation / n_decoded, 1e3 / t_token_generation * n_decoded);
+        LOG_TEE("%s:       total time = %10.2f ms\n", __func__, t_prompt_processing + t_token_generation);
     }
 };
 
@@ -410,8 +343,6 @@ struct llama_server_context
     llama_server_queue queue_tasks;
     llama_server_response queue_results;
 
-    llama_metrics metrics;
-
     ~llama_server_context()
     {
         if (ctx)
@@ -431,7 +362,7 @@ struct llama_server_context
         params = params_;
         if (!params.mmproj.empty()) {
             multimodal = true;
-            LOG_INFO("Multi Modal Mode Enabled", {});
+            LOG_TEE("Multi Modal Mode Enabled");
             clp_ctx = clip_model_load(params.mmproj.c_str(), /*verbosity=*/ 1);
             if(clp_ctx == nullptr) {
                 LOG_ERROR("unable to load clip model", {{"model", params.mmproj}});
@@ -468,23 +399,13 @@ struct llama_server_context
         return true;
     }
 
-    void validate_model_chat_template(server_params & sparams) {
-        llama_chat_message chat[] = {{"user", "test"}};
-        std::vector<char> buf(1);
-        int res = llama_chat_apply_template(model, nullptr, chat, 1, true, buf.data(), buf.size());
-        if (res < 0) {
-            LOG_ERROR("The chat template comes with this model is not yet supported, falling back to chatml. This may cause the model to output suboptimal responses", {});
-            sparams.chat_template = "<|im_start|>"; // llama_chat_apply_template only checks if <|im_start|> exist in the template
-        }
-    }
-
     void initialize() {
         // create slots
         all_slots_are_idle = true;
 
         const int32_t n_ctx_slot = n_ctx / params.n_parallel;
 
-        LOG_INFO("initializing slots", {{"n_slots", params.n_parallel}});
+        LOG_TEE("Available slots:\n");
         for (int i = 0; i < params.n_parallel; i++)
         {
             llama_client_slot slot;
@@ -493,10 +414,7 @@ struct llama_server_context
             slot.n_ctx = n_ctx_slot;
             slot.n_predict = params.n_predict;
 
-            LOG_INFO("new slot", {
-                {"slot_id",    slot.id},
-                {"n_ctx_slot", slot.n_ctx}
-            });
+            LOG_TEE(" -> Slot %i - max context: %i\n", slot.id, n_ctx_slot);
 
             const int ga_n = params.grp_attn_n;
             const int ga_w = params.grp_attn_w;
@@ -506,12 +424,7 @@ struct llama_server_context
                 GGML_ASSERT(ga_w % ga_n == 0            && "ga_w must be a multiple of ga_n");     // NOLINT
                 //GGML_ASSERT(n_ctx_train % ga_w == 0     && "n_ctx_train must be a multiple of ga_w");    // NOLINT
                 //GGML_ASSERT(n_ctx >= n_ctx_train * ga_n && "n_ctx must be at least n_ctx_train * ga_n"); // NOLINT
-
-                LOG_INFO("slot self-extend", {
-                    {"slot_id",   slot.id},
-                    {"ga_n",      ga_n},
-                    {"ga_w",      ga_w}
-                });
+                LOG_TEE(" -> Slot %i - self-extend: ga_n = %d, ga_w = %d\n", slot.id, ga_n, ga_w);
             }
 
             slot.ga_i = 0;
@@ -805,16 +718,10 @@ struct llama_server_context
                     img_sl.img_data = clip_image_u8_init();
                     if (!clip_image_load_from_bytes(image_buffer.data(), image_buffer.size(), img_sl.img_data))
                     {
-                        LOG_ERROR("failed to load image", {
-                            {"slot_id",   slot->id},
-                            {"img_sl_id", img_sl.id}
-                        });
+                        LOG_TEE("slot %i - failed to load image [id: %i]\n", slot->id, img_sl.id);
                         return false;
                     }
-                    LOG_VERBOSE("image loaded", {
-                        {"slot_id",   slot->id},
-                        {"img_sl_id", img_sl.id}
-                    });
+                    LOG_TEE("slot %i - loaded image\n", slot->id);
                     img_sl.request_encode_image = true;
                     slot->images.push_back(img_sl);
                 }
@@ -874,10 +781,7 @@ struct llama_server_context
 
         all_slots_are_idle = false;
 
-        LOG_INFO("slot is processing task", {
-            {"slot_id", slot->id},
-            {"task_id", slot->task_id},
-        });
+        LOG_TEE("slot %i is processing [task id: %i]\n", slot->id, slot->task_id);
 
         return true;
     }
@@ -902,24 +806,10 @@ struct llama_server_context
                 llama_batch_add(batch, system_tokens[i], i, { 0 }, false);
             }
 
-            for (int32_t i = 0; i < (int32_t) batch.n_tokens; i += params.n_batch)
+            if (llama_decode(ctx, batch) != 0)
             {
-                const int32_t n_tokens = std::min(params.n_batch, (int32_t) (batch.n_tokens - i));
-                llama_batch batch_view = {
-                    n_tokens,
-                    batch.token    + i,
-                    nullptr,
-                    batch.pos      + i,
-                    batch.n_seq_id + i,
-                    batch.seq_id   + i,
-                    batch.logits   + i,
-                    0, 0, 0, // unused
-                };
-                if (llama_decode(ctx, batch_view) != 0)
-                {
-                    LOG_TEE("%s: llama_decode() failed\n", __func__);
-                    return;
-                }
+                LOG_TEE("%s: llama_decode() failed\n", __func__);
+                return;
             }
 
             // assign the system KV cache to all parallel sequences
@@ -1107,12 +997,43 @@ struct llama_server_context
             {
                 continue;
             }
-
-            if (!llava_image_embed_make_with_clip_img(clp_ctx, params.n_threads, img.img_data, &img.image_embedding, &img.image_tokens)) {
+            clip_image_f32_batch img_res_v;
+            img_res_v.size = 0;
+            img_res_v.data = nullptr;
+            if (!clip_image_preprocess(clp_ctx, img.img_data, img_res_v))
+            {
+                LOG_TEE("Error processing the given image");
+                clip_free(clp_ctx);
+                clip_image_f32_batch_free(img_res_v);
+                return false;
+            }
+            if (img_res_v.size == 0)
+            {
                 LOG_TEE("Error processing the given image");
                 return false;
             }
 
+            // note: assumes only one image was returned by clip_image_preprocess
+            clip_image_f32 * img_res = img_res_v.data;
+
+            img.image_tokens = clip_n_patches(clp_ctx);
+            img.image_embedding = (float *)malloc(clip_embd_nbytes(clp_ctx));
+            if (!img.image_embedding)
+            {
+                LOG_TEE("Unable to allocate memory for image embeddings\n");
+                clip_image_f32_batch_free(img_res_v);
+                clip_free(clp_ctx);
+                return false;
+            }
+            LOG_TEE("slot %i - encoding image [id: %i]\n", slot.id, img.id);
+            if (!clip_image_encode(clp_ctx, params.n_threads, img_res, img.image_embedding))
+            {
+                LOG_TEE("Unable to encode image\n");
+                clip_image_f32_batch_free(img_res_v);
+                return false;
+            }
+
+            clip_image_f32_batch_free(img_res_v);
 
             img.request_encode_image = false;
         }
@@ -1454,7 +1375,7 @@ struct llama_server_context
                 if (slot == nullptr)
                 {
                     // if no slot is available, we defer this task for processing later
-                    LOG_VERBOSE("no slot is available", {{"task_id", task.id}});
+                    LOG_VERBOSE("no slot is available", {});
                     queue_tasks.defer(task);
                     break;
                 }
@@ -1503,70 +1424,6 @@ struct llama_server_context
             case TASK_TYPE_NEXT_RESPONSE: {
                 // do nothing
             } break;
-            case TASK_TYPE_METRICS: {
-                json slots_data        = json::array();
-                int n_idle_slots       = 0;
-                int n_processing_slots = 0;
-
-                for (llama_client_slot &slot: slots) {
-                    json slot_data = get_formated_generation(slot);
-                    slot_data["id"] = slot.id;
-                    slot_data["task_id"] = slot.task_id;
-                    slot_data["state"] = slot.state;
-                    slot_data["prompt"] = slot.prompt;
-                    slot_data["next_token"] = {
-                            {"has_next_token", slot.has_next_token},
-                            {"n_remain", slot.n_remaining},
-                            {"num_tokens_predicted", slot.n_decoded},
-                            {"stopped_eos", slot.stopped_eos},
-                            {"stopped_word", slot.stopped_word},
-                            {"stopped_limit", slot.stopped_limit},
-                            {"stopping_word", slot.stopping_word},
-                    };
-                    if (slot_data["state"] == IDLE) {
-                        n_idle_slots++;
-                    } else {
-                        n_processing_slots++;
-                    }
-                    slots_data.push_back(slot_data);
-                }
-                LOG_INFO("slot data", {
-                    {"task_id",            task.id},
-                    {"n_idle_slots",       n_idle_slots},
-                    {"n_processing_slots", n_processing_slots}
-                });
-                LOG_VERBOSE("slot data", {
-                    {"task_id",            task.id},
-                    {"n_idle_slots",       n_idle_slots},
-                    {"n_processing_slots", n_processing_slots},
-                    {"slots",              slots_data}
-                });
-                task_result res;
-                res.id = task.id;
-                res.multitask_id = task.multitask_id;
-                res.stop = true;
-                res.error = false;
-                res.result_json = {
-                        { "idle",                            n_idle_slots       },
-                        { "processing",                      n_processing_slots },
-                        { "deferred",                        queue_tasks.queue_tasks_deferred.size() },
-
-                        { "n_prompt_tokens_processed_total", metrics.n_prompt_tokens_processed_total},
-                        { "n_tokens_predicted_total",        metrics.n_tokens_predicted_total},
-
-                        { "n_prompt_tokens_processed",       metrics.n_prompt_tokens_processed},
-                        { "t_prompt_processing",             metrics.t_prompt_processing},
-                        { "n_tokens_predicted",              metrics.n_tokens_predicted},
-                        { "t_tokens_generation",             metrics.t_tokens_generation},
-
-                        { "kv_cache_tokens_count",          llama_get_kv_cache_token_count(ctx)},
-                        { "kv_cache_used_cells",            llama_get_kv_cache_used_cells(ctx)},
-
-                        { "slots",                          slots_data },
-                };
-                metrics.reset_bucket();
-                queue_results.send(res);
-            } break;
         }
     }
 
@@ -1592,7 +1449,7 @@ struct llama_server_context
     bool update_slots() {
         if (system_need_update)
         {
-            LOG_INFO("updating system prompt", {});
+            LOG_TEE("updating system prompt\n");
             update_system_prompt();
         }
 
@@ -1602,13 +1459,12 @@ struct llama_server_context
         {
             if (system_prompt.empty() && clean_kv_cache)
             {
-                LOG_INFO("all slots are idle and system prompt is empty, clear the KV cache", {});
+                LOG_TEE("all slots are idle and system prompt is empty, clear the KV cache\n");
                 kv_cache_clear();
             }
             return true;
         }
 
-        LOG_VERBOSE("posting NEXT_RESPONSE", {});
         task_server task;
         task.type = TASK_TYPE_NEXT_RESPONSE;
         task.target_id = -1;
@@ -1621,25 +1477,14 @@ struct llama_server_context
                 if (slot.is_processing() && system_tokens.size() + slot.cache_tokens.size() >= (size_t) slot.n_ctx)
                 {
                     // Shift context
-                    const int n_keep    = slot.params.n_keep + add_bos_token;
-                    const int n_left    = (int) system_tokens.size() + slot.n_past - n_keep;
+                    const int n_left    = system_tokens.size() + slot.n_past - slot.params.n_keep - 1;
                     const int n_discard = n_left / 2;
 
-                    LOG_INFO("slot context shift", {
-                        {"slot_id",         slot.id},
-                        {"task_id",         slot.task_id},
-                        {"n_keep",          n_keep},
-                        {"n_left",          n_left},
-                        {"n_discard",       n_discard},
-                        {"n_ctx",           n_ctx},
-                        {"n_past",          slot.n_past},
-                        {"n_system_tokens", system_tokens.size()},
-                        {"n_cache_tokens",  slot.cache_tokens.size()}
-                    });
-                    llama_kv_cache_seq_rm (ctx, slot.id, n_keep            , n_keep + n_discard);
-                    llama_kv_cache_seq_add(ctx, slot.id, n_keep + n_discard, system_tokens.size() + slot.n_past, -n_discard);
+                    LOG_TEE("slot %d: context shift - n_keep = %d, n_left = %d, n_discard = %d\n", slot.id, slot.params.n_keep, n_left, n_discard);
+                    llama_kv_cache_seq_rm   (ctx, slot.id, slot.params.n_keep + 1            , slot.params.n_keep + n_discard + 1);
+                    llama_kv_cache_seq_shift(ctx, slot.id, slot.params.n_keep + 1 + n_discard, system_tokens.size() + slot.n_past, -n_discard);
 
-                    for (size_t i = n_keep + n_discard; i < slot.cache_tokens.size(); i++)
+                    for (size_t i = slot.params.n_keep + 1 + n_discard; i < slot.cache_tokens.size(); i++)
                     {
                         slot.cache_tokens[i - n_discard] = slot.cache_tokens[i];
                     }
@@ -1649,12 +1494,17 @@ struct llama_server_context
                     slot.n_past -= n_discard;
 
                     slot.truncated = true;
+
+                    LOG_VERBOSE("context shift", {
+                        { "n_ctx", n_ctx },
+                        { "n_keep", params.n_keep },
+                        { "n_left", n_left },
+                    });
                 }
             }
         }
 
         // decode any currently ongoing sequences
-        LOG_VERBOSE("decoding ongoing sequences", {});
         for (auto & slot : slots)
         {
             // release the slot
@@ -1664,15 +1514,7 @@ struct llama_server_context
                 slot.command = NONE;
                 slot.t_last_used = ggml_time_us();
 
-                LOG_INFO("slot released", {
-                    {"slot_id",         slot.id},
-                    {"task_id",         slot.task_id},
-                    {"n_ctx",           n_ctx},
-                    {"n_past",          slot.n_past},
-                    {"n_system_tokens", system_tokens.size()},
-                    {"n_cache_tokens",  slot.cache_tokens.size()},
-                    {"truncated",       slot.truncated}
-                });
+                LOG_TEE("slot %d released (%d tokens in cache)\n", slot.id, (int) slot.cache_tokens.size());
                 queue_tasks.notify_slot_changed();
 
                 continue;
@@ -1799,14 +1641,6 @@ struct llama_server_context
                         }
 
                         slot.n_past = common_part(slot.cache_tokens, prompt_tokens);
-
-                        // the last token of the cache is not in the KV cache until the next call to llama_decode
-                        // (it was sampled, pushed into the "cache_tokens", but not yet put in the context)
-                        if (slot.n_past > 0 && slot.n_past == (int32_t) slot.cache_tokens.size())
-                        {
-                            slot.n_past -= 1;
-                        }
-
                         slot.num_prompt_tokens_processed = slot.num_prompt_tokens - slot.n_past;
 
                         if (slot.ga_n != 1)
@@ -1828,12 +1662,7 @@ struct llama_server_context
                             slot.ga_i = ga_i;
                         }
 
-                        LOG_INFO("slot progression", {
-                            { "slot_id", slot.id },
-                            { "task_id", slot.task_id },
-                            { "n_past",  slot.n_past },
-                            { "num_prompt_tokens_processed", slot.num_prompt_tokens_processed }
-                        });
+                        LOG_TEE("slot %d : in cache: %i tokens | to process: %i tokens\n", slot.id, slot.n_past, slot.num_prompt_tokens_processed);
                     }
 
                     slot.cache_tokens = prompt_tokens;
@@ -1841,10 +1670,7 @@ struct llama_server_context
                     if (slot.n_past == slot.num_prompt_tokens && slot.n_past > 0)
                     {
                         // we have to evaluate at least 1 token to generate logits.
-                        LOG_INFO("we have to evaluate at least 1 token to generate logits", {
-                            { "slot_id", slot.id },
-                            { "task_id", slot.task_id }
-                        });
+                        LOG_TEE("slot %d : we have to evaluate at least 1 token to generate logits\n", slot.id);
                         slot.n_past--;
                         if (slot.ga_i > 0)
                         {
@@ -1852,13 +1678,9 @@ struct llama_server_context
                         }
                     }
 
-                    int p0 = (int) system_tokens.size() + slot.n_past;
-                    LOG_INFO("kv cache rm [p0, end)", {
-                        { "slot_id", slot.id },
-                        { "task_id", slot.task_id },
-                        { "p0",      p0 }
-                    });
-                    llama_kv_cache_seq_rm(ctx, slot.id, p0, -1);
+                    LOG_TEE("slot %d : kv cache rm - [%d, end)\n", slot.id, (int) system_tokens.size() + slot.n_past);
+
+                    llama_kv_cache_seq_rm(ctx, slot.id, system_tokens.size() + slot.n_past, -1);
 
                     LOG_VERBOSE("prompt ingested", {
                                                     {"n_past",  slot.n_past},
@@ -1893,13 +1715,7 @@ struct llama_server_context
 
                     if (has_images && !ingest_images(slot, n_batch))
                     {
-                        LOG_ERROR("failed processing images", {
-                            "slot_id", slot.id,
-                            "task_id", slot.task_id,
-                        });
-                        // FIXME @phymbert: to be properly tested
-                        //  early returning without changing the slot state will block the slot for ever
-                        // no one at the moment is checking the return value
+                        LOG_TEE("failed processing images\n");
                         return false;
                     }
 
@@ -1941,9 +1757,9 @@ struct llama_server_context
                         LOG_TEE("div:   [%6d, %6d] / %6d -> [%6d, %6d]\n", slot.ga_i + ib * bd, slot.ga_i + ib * bd + slot.ga_w, slot.ga_n, (slot.ga_i + ib * bd) / slot.ga_n, (slot.ga_i + ib * bd + slot.ga_w) / slot.ga_n);
                         LOG_TEE("shift: [%6d, %6d] + %6d -> [%6d, %6d]\n", slot.ga_i + ib * bd + slot.ga_w, slot.n_past_se + ib * bd, dd, slot.ga_i + ib * bd + slot.ga_w + dd, slot.n_past_se + ib * bd + dd);
 
-                        llama_kv_cache_seq_add(ctx, slot.id, slot.ga_i, slot.n_past_se, ib * bd);
+                        llama_kv_cache_seq_shift(ctx, slot.id, slot.ga_i, slot.n_past_se, ib * bd);
                         llama_kv_cache_seq_div(ctx, slot.id, slot.ga_i + ib * bd, slot.ga_i + ib * bd + slot.ga_w,slot.ga_n);
-                        llama_kv_cache_seq_add(ctx, slot.id, slot.ga_i + ib * bd + slot.ga_w,slot.n_past_se + ib * bd, dd);
+                        llama_kv_cache_seq_shift(ctx, slot.id, slot.ga_i + ib * bd + slot.ga_w,slot.n_past_se + ib * bd, dd);
 
                         slot.n_past_se -= bd;
 
@@ -1999,7 +1815,7 @@ struct llama_server_context
                     send_embedding(slot);
                     slot.release();
                     slot.i_batch = -1;
-                    continue;
+                    return true;
                 }
 
                 completion_token_output result;
@@ -2012,7 +1828,6 @@ struct llama_server_context
                 {
                     slot.t_start_genereration = ggml_time_us();
                     slot.t_prompt_processing = (slot.t_start_genereration - slot.t_start_process_prompt) / 1e3;
-                    metrics.on_prompt_eval(slot);
                 }
 
                 llama_token_data_array cur_p = { slot.ctx_sampling->cur.data(), slot.ctx_sampling->cur.size(), false };
@@ -2035,14 +1850,11 @@ struct llama_server_context
                     slot.release();
                     slot.print_timings();
                     send_final_response(slot);
-                    metrics.on_prediction(slot);
                 }
 
                 slot.i_batch = -1;
             }
         }
-
-        LOG_VERBOSE("slots updated", {});
         return true;
     }
 
@@ -2115,15 +1927,9 @@ static void server_print_usage(const char *argv0, const gpt_params &params,
     printf("  -cb, --cont-batching      enable continuous batching (a.k.a dynamic batching) (default: disabled)\n");
     printf("  -spf FNAME, --system-prompt-file FNAME\n");
     printf("                            set a file to load a system prompt (initial prompt of all slots), this is useful for chat applications.\n");
-    printf("  -ctk TYPE, --cache-type-k TYPE\n");
-    printf("                            KV cache data type for K (default: f16)\n");
-    printf("  -ctv TYPE, --cache-type-v TYPE\n");
-    printf("                            KV cache data type for V (default: f16)\n");
     printf("  --mmproj MMPROJ_FILE      path to a multimodal projector file for LLaVA.\n");
-    printf("  --log-format              log output format: json or text (default: json)\n");
     printf("  --log-disable             disables logging to a file.\n");
     printf("  --slots-endpoint-disable  disables slots monitoring endpoint.\n");
-    printf("  --metrics                 enable prometheus compatible metrics endpoint (default: %s).\n", sparams.metrics_endpoint ? "enabled" : "disabled");
     printf("\n");
     printf("  -n, --n-predict           maximum tokens to predict (default: %d)\n", params.n_predict);
     printf("  --override-kv KEY=TYPE:VALUE\n");
@@ -2131,9 +1937,8 @@ static void server_print_usage(const char *argv0, const gpt_params &params,
     printf("                            types: int, float, bool. example: --override-kv tokenizer.ggml.add_bos_token=bool:false\n");
     printf("  -gan N, --grp-attn-n N    set the group attention factor to extend context size through self-extend(default: 1=disabled), used together with group attention width `--grp-attn-w`");
     printf("  -gaw N, --grp-attn-w N    set the group attention width to extend context size through self-extend(default: 512), used together with group attention factor `--grp-attn-n`");
-    printf("  --chat-template JINJA_TEMPLATE\n");
-    printf("                            set custom jinja chat template (default: template taken from model's metadata)\n");
-    printf("                            Note: only commonly used templates are accepted, since we don't have jinja parser\n");
+    printf("  --chat-template FORMAT_NAME");
+    printf("                            set chat template, possible value is: llama2, chatml (default %s)", sparams.chat_template.c_str());
     printf("\n");
 }
 
@@ -2255,9 +2060,9 @@ static void server_params_parse(int argc, char **argv, server_params &sparams,
                 break;
             }
             std::string value(argv[i]);
-            /**/ if (value == "none")   { params.rope_scaling_type = LLAMA_ROPE_SCALING_TYPE_NONE; }
-            else if (value == "linear") { params.rope_scaling_type = LLAMA_ROPE_SCALING_TYPE_LINEAR; }
-            else if (value == "yarn")   { params.rope_scaling_type = LLAMA_ROPE_SCALING_TYPE_YARN; }
+            /**/ if (value == "none")   { params.rope_scaling_type = LLAMA_ROPE_SCALING_NONE; }
+            else if (value == "linear") { params.rope_scaling_type = LLAMA_ROPE_SCALING_LINEAR; }
+            else if (value == "yarn")   { params.rope_scaling_type = LLAMA_ROPE_SCALING_YARN; }
             else { invalid_param = true; break; }
         }
         else if (arg == "--rope-freq-base")
@@ -2381,15 +2186,15 @@ static void server_params_parse(int argc, char **argv, server_params &sparams,
             std::string arg_next = argv[i];
             if (arg_next == "none")
             {
-                params.split_mode = LLAMA_SPLIT_MODE_NONE;
+                params.split_mode = LLAMA_SPLIT_NONE;
             }
             else if (arg_next == "layer")
             {
-                params.split_mode = LLAMA_SPLIT_MODE_LAYER;
+                params.split_mode = LLAMA_SPLIT_LAYER;
             }
             else if (arg_next == "row")
             {
-                params.split_mode = LLAMA_SPLIT_MODE_ROW;
+                params.split_mode = LLAMA_SPLIT_ROW;
             }
             else {
                 invalid_param = true;
@@ -2559,12 +2364,6 @@ static void server_params_parse(int argc, char **argv, server_params &sparams,
             );
             llama.process_system_prompt_data(json::parse(systm_content));
         }
-        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 == "--mmproj")
         {
             if (++i >= argc)
@@ -2574,27 +2373,6 @@ static void server_params_parse(int argc, char **argv, server_params &sparams,
             }
             params.mmproj = argv[i];
         }
-        else if (arg == "--log-format")
-        {
-            if (++i >= argc)
-            {
-                invalid_param = true;
-                break;
-            }
-            if (std::strcmp(argv[i], "json") == 0)
-            {
-                server_log_json = true;
-            }
-            else if (std::strcmp(argv[i], "text") == 0)
-            {
-                server_log_json = false;
-            }
-            else
-            {
-                invalid_param = true;
-                break;
-            }
-        }
         else if (arg == "--log-disable")
         {
             log_set_target(stdout);
@@ -2604,10 +2382,6 @@ static void server_params_parse(int argc, char **argv, server_params &sparams,
         {
             sparams.slots_endpoint = false;
         }
-        else if (arg == "--metrics")
-        {
-            sparams.metrics_endpoint = true;
-        }
         else if (arg == "--chat-template")
         {
             if (++i >= argc)
@@ -2615,13 +2389,13 @@ static void server_params_parse(int argc, char **argv, server_params &sparams,
                 invalid_param = true;
                 break;
             }
-            if (!verify_custom_template(argv[i])) {
-                fprintf(stderr, "error: the supplied chat template is not supported: %s\n", argv[i]);
-                fprintf(stderr, "note: llama.cpp does not use jinja parser, we only support commonly used templates\n");
+            std::string value(argv[i]);
+            if (value != "chatml" && value != "llama2") {
+                fprintf(stderr, "error: chat template can be \"llama2\" or \"chatml\", but got: %s\n", value.c_str());
                 invalid_param = true;
                 break;
             }
-            sparams.chat_template = argv[i];
+            sparams.chat_template = value;
         }
         else if (arg == "--override-kv")
         {
@@ -2641,15 +2415,15 @@ static void server_params_parse(int argc, char **argv, server_params &sparams,
             sep++;
             if (strncmp(sep, "int:", 4) == 0) {
                 sep += 4;
-                kvo.tag = LLAMA_KV_OVERRIDE_TYPE_INT;
+                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_TYPE_FLOAT;
+                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_TYPE_BOOL;
+                kvo.tag = LLAMA_KV_OVERRIDE_BOOL;
                 if (std::strcmp(sep, "true") == 0) {
                     kvo.bool_value = true;
                 } else if (std::strcmp(sep, "false") == 0) {
@@ -2708,40 +2482,32 @@ static json format_partial_response(
 
 static json format_tokenizer_response(const std::vector<llama_token> &tokens)
 {
-    return json {
-        {"tokens", tokens}
-    };
+    return json{
+        {"tokens", tokens}};
 }
 
 static json format_detokenized_response(std::string content)
 {
-    return json {
-        {"content", content}
-    };
+    return json{
+        {"content", content}};
 }
 
 
 static void log_server_request(const httplib::Request &req, const httplib::Response &res)
 {
-    // skip GH copilot requests when using default port
-    if (req.path == "/v1/health" || req.path == "/v1/completions")
-    {
-        return;
-    }
-
     LOG_INFO("request", {
-        {"remote_addr", req.remote_addr},
-        {"remote_port", req.remote_port},
-        {"status",      res.status},
-        {"method",      req.method},
-        {"path",        req.path},
-        {"params",      req.params},
-    });
+                            {"remote_addr", req.remote_addr},
+                            {"remote_port", req.remote_port},
+                            {"status", res.status},
+                            {"method", req.method},
+                            {"path", req.path},
+                            {"params", req.params},
+                        });
 
     LOG_VERBOSE("request", {
-        {"request",  req.body},
-        {"response", res.body},
-    });
+                               {"request", req.body},
+                               {"response", res.body},
+                           });
 }
 
 struct token_translator
@@ -2816,44 +2582,40 @@ int main(int argc, char **argv)
         res.set_header("Access-Control-Allow-Headers", "*");
     });
 
-    svr.Get("/health", [&](const httplib::Request& req, httplib::Response& res) {
+    svr.Get("/health", [&](const httplib::Request&, httplib::Response& res) {
         server_state current_state = state.load();
         switch(current_state) {
-            case SERVER_STATE_READY: {
-                // request slots data using task queue
-                task_server task;
-                task.id   = llama.queue_tasks.get_new_id();
-                task.type = TASK_TYPE_METRICS;
-                task.target_id = -1;
-
-                llama.queue_results.add_waiting_task_id(task.id);
-                llama.queue_tasks.post(task);
-
-                // get the result
-                task_result result = llama.queue_results.recv(task.id);
-                llama.queue_results.remove_waiting_task_id(task.id);
-
-                int n_idle_slots       = result.result_json["idle"];
-                int n_processing_slots = result.result_json["processing"];
-
-                json health = {
-                        {"status",           "ok"},
-                        {"slots_idle",       n_idle_slots},
-                        {"slots_processing", n_processing_slots}};
-                res.status = 200; // HTTP OK
-                if (sparams.slots_endpoint && req.has_param("include_slots")) {
-                    health["slots"] = result.result_json["slots"];
-                }
-
-                if (n_idle_slots == 0) {
-                    health["status"] = "no slot available";
-                    if (req.has_param("fail_on_no_slot")) {
+            case SERVER_STATE_READY:
+                if (llama.all_slots_are_idle) {
+                    res.set_content(R"({"status": "ok"})", "application/json");
+                    res.status = 200; // HTTP OK
+                } else {
+                    int available_slots = 0;
+                    int processing_slots = 0;
+                    for (llama_client_slot & slot : llama.slots) {
+                        if (slot.available()) {
+                            available_slots++;
+                        } else {
+                            processing_slots++;
+                        }
+                    }
+                    if (available_slots > 0) {
+                        json health = {
+                                {"status",           "ok"},
+                                {"slots_idle",       available_slots},
+                                {"slots_processing", processing_slots}};
+                        res.set_content(health.dump(), "application/json");
+                        res.status = 200; // HTTP OK
+                    } else {
+                        json health = {
+                                {"status",           "no slot available"},
+                                {"slots_idle",       available_slots},
+                                {"slots_processing", processing_slots}};
+                        res.set_content(health.dump(), "application/json");
                         res.status = 503; // HTTP Service Unavailable
                     }
                 }
-                res.set_content(health.dump(), "application/json");
                 break;
-            }
             case SERVER_STATE_LOADING_MODEL:
                 res.set_content(R"({"status": "loading model"})", "application/json");
                 res.status = 503; // HTTP Service Unavailable
@@ -2867,101 +2629,26 @@ int main(int argc, char **argv)
 
     if (sparams.slots_endpoint) {
         svr.Get("/slots", [&](const httplib::Request&, httplib::Response& res) {
-            // request slots data using task queue
-            task_server task;
-            task.id = llama.queue_tasks.get_new_id();
-            task.type = TASK_TYPE_METRICS;
-            task.target_id = -1;
+            json slots;
+            for (llama_client_slot & slot : llama.slots) {
+                json slot_data = llama.get_formated_generation(slot);
+                slot_data["id"] = slot.id;
+                slot_data["task_id"] = slot.task_id;
+                slot_data["state"] = slot.state;
+                slot_data["prompt"] = slot.prompt;
+                slot_data["next_token"] = {
+                        {"has_next_token", slot.has_next_token},
+                        {"n_remain", slot.n_remaining},
+                        {"num_tokens_predicted", slot.n_decoded},
+                        {"stopped_eos", slot.stopped_eos},
+                        {"stopped_word", slot.stopped_word},
+                        {"stopped_limit", slot.stopped_limit},
+                        {"stopping_word", slot.stopping_word},
+                };
 
-            llama.queue_results.add_waiting_task_id(task.id);
-            llama.queue_tasks.post(task);
-
-            // get the result
-            task_result result = llama.queue_results.recv(task.id);
-            llama.queue_results.remove_waiting_task_id(task.id);
-
-            res.set_content(result.result_json["slots"].dump(), "application/json");
-            res.status = 200; // HTTP OK
-        });
-    }
-
-    if (sparams.metrics_endpoint) {
-        svr.Get("/metrics", [&](const httplib::Request&, httplib::Response& res) {
-            // request slots data using task queue
-            task_server task;
-            task.id = llama.queue_tasks.get_new_id();
-            task.type = TASK_TYPE_METRICS;
-            task.target_id = -1;
-
-            llama.queue_results.add_waiting_task_id(task.id);
-            llama.queue_tasks.post(task);
-
-            // get the result
-            task_result result = llama.queue_results.recv(task.id);
-            llama.queue_results.remove_waiting_task_id(task.id);
-
-            json data = result.result_json;
-
-            uint64_t n_prompt_tokens_processed = data["n_prompt_tokens_processed"];
-            uint64_t t_prompt_processing       = data["t_prompt_processing"];
-
-            uint64_t n_tokens_predicted       = data["n_tokens_predicted"];
-            uint64_t t_tokens_generation      = data["t_tokens_generation"];
-
-            int32_t kv_cache_used_cells = data["kv_cache_used_cells"];
-
-            // metrics definition: https://prometheus.io/docs/practices/naming/#metric-names
-            json all_metrics_def = json {
-                    {"counter", {{
-                            {"name",  "prompt_tokens_total"},
-                            {"help",  "Number of prompt tokens processed."},
-                            {"value",  data["n_prompt_tokens_processed_total"]}
-                    }, {
-                            {"name",  "tokens_predicted_total"},
-                            {"help",  "Number of generation tokens processed."},
-                            {"value",  data["n_tokens_predicted_total"]}
-                    }}},
-                    {"gauge", {{
-                            {"name",  "prompt_tokens_seconds"},
-                            {"help",  "Average prompt throughput in tokens/s."},
-                            {"value",  n_prompt_tokens_processed ? 1e3 / t_prompt_processing * n_prompt_tokens_processed : 0}
-                    },{
-                            {"name",  "predicted_tokens_seconds"},
-                            {"help",  "Average generation throughput in tokens/s."},
-                            {"value",  n_tokens_predicted ? 1e3 / t_tokens_generation * n_tokens_predicted : 0}
-                     },{
-                            {"name",  "kv_cache_usage_ratio"},
-                            {"help",  "KV-cache usage. 1 means 100 percent usage."},
-                            {"value",  1. * kv_cache_used_cells / params.n_ctx}
-                     },{
-                            {"name",  "kv_cache_tokens"},
-                            {"help",  "KV-cache tokens."},
-                            {"value",  data["kv_cache_tokens_count"]}
-                    },{
-                            {"name",  "requests_processing"},
-                            {"help",  "Number of request processing."},
-                            {"value",  data["processing"]}
-                  },{
-                            {"name",  "requests_deferred"},
-                            {"help",  "Number of request deferred."},
-                            {"value",  data["deferred"]}
-                  }}}
-            };
-
-            std::stringstream prometheus;
-            for (const auto& el : all_metrics_def.items()) {
-                const auto& type = el.key();
-                const auto& metrics_def = el.value();
-                for (const auto& metric_def : metrics_def) {
-                    std::string name = metric_def["name"];
-                    std::string help = metric_def["help"];
-                    prometheus << "# HELP llamacpp:" << name << " " << help                << "\n"
-                               << "# TYPE llamacpp:" << name << " " << type                << "\n"
-                               << "llamacpp:"        << name << " " << metric_def["value"] << "\n";
-                }
+                slots.push_back(slot_data);
             }
-
-            res.set_content(prometheus.str(), "text/plain; version=0.0.4");
+            res.set_content(slots.dump(), "application/json");
             res.status = 200; // HTTP OK
         });
     }
@@ -3018,6 +2705,9 @@ int main(int argc, char **argv)
     // Set the base directory for serving static files
     svr.set_base_dir(sparams.public_path);
 
+    // to make it ctrl+clickable:
+    LOG_TEE("\nllama server listening at http://%s:%d\n\n", sparams.hostname.c_str(), sparams.port);
+
     std::unordered_map<std::string, std::string> log_data;
     log_data["hostname"] = sparams.hostname;
     log_data["port"] = std::to_string(sparams.port);
@@ -3028,6 +2718,19 @@ int main(int argc, char **argv)
         log_data["api_key"] = "api_key: " + std::to_string(sparams.api_keys.size()) + " keys loaded";
     }
 
+    LOG_INFO("HTTP server listening", log_data);
+    // run the HTTP server in a thread - see comment below
+    std::thread t([&]()
+            {
+                if (!svr.listen_after_bind())
+                {
+                    state.store(SERVER_STATE_ERROR);
+                    return 1;
+                }
+
+                return 0;
+            });
+
     // load the model
     if (!llama.load_model(params))
     {
@@ -3039,11 +2742,6 @@ int main(int argc, char **argv)
         LOG_INFO("model loaded", {});
     }
 
-    if (sparams.chat_template.empty()) { // custom chat template is not supplied
-        // check if the template comes with the model is supported by us
-        llama.validate_model_chat_template(sparams);
-    }
-
     // 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
@@ -3215,7 +2913,7 @@ int main(int argc, char **argv)
                 if (!validate_api_key(req, res)) {
                     return;
                 }
-                json data = oaicompat_completion_params_parse(llama.model, json::parse(req.body), sparams.chat_template);
+                json data = oaicompat_completion_params_parse(json::parse(req.body), sparams.chat_template);
 
                 const int task_id = llama.queue_tasks.get_new_id();
                 llama.queue_results.add_waiting_task_id(task_id);
@@ -3495,19 +3193,6 @@ int main(int argc, char **argv)
     }*/
     //);
 
-    LOG_INFO("HTTP server listening", log_data);
-    // run the HTTP server in a thread - see comment below
-    std::thread t([&]()
-            {
-                if (!svr.listen_after_bind())
-                {
-                    state.store(SERVER_STATE_ERROR);
-                    return 1;
-                }
-
-                return 0;
-            });
-
     llama.queue_tasks.on_new_task(std::bind(
         &llama_server_context::process_single_task, &llama, std::placeholders::_1));
     llama.queue_tasks.on_finish_multitask(std::bind(

examples/server/tests/README.md

@@ -1,47 +0,0 @@
-# Server tests
-
-Python based server tests scenario using [BDD](https://en.wikipedia.org/wiki/Behavior-driven_development) and [behave](https://behave.readthedocs.io/en/latest/):
- * [issues.feature](./features/issues.feature) Pending issues scenario
- * [parallel.feature](./features/parallel.feature) Scenario involving multi slots and concurrent requests
- * [security.feature](./features/security.feature) Security, CORS and API Key
- * [server.feature](./features/server.feature) Server base scenario: completion, embedding, tokenization, etc...
-
-Tests target GitHub workflows job runners with 4 vCPU.
-
-Requests are using [aiohttp](https://docs.aiohttp.org/en/stable/client_reference.html), [asyncio](https://docs.python.org/fr/3/library/asyncio.html) based http client.
-
-Note: If the host architecture inference speed is faster than GitHub runners one, parallel scenario may randomly fail. To mitigate it, you can increase values in `n_predict`, `kv_size`.
-
-### Install dependencies
-`pip install -r requirements.txt`
-
-### Run tests
-1. Build the server
-```shell
-cd ../../..
-mkdir build
-cd build
-cmake ../
-cmake --build . --target server
-```
-2. download required models:
-   1. `../../../scripts/hf.sh --repo ggml-org/models --file tinyllamas/stories260K.gguf`
-3. Start the test: `./tests.sh`
-
-It's possible to override some scenario steps values with environment variables:
- - `PORT` -> `context.server_port` to set the listening port of the server during scenario, default: `8080`
- - `LLAMA_SERVER_BIN_PATH` -> to change the server binary path, default: `../../../build/bin/server`
- - `DEBUG` -> "ON" to enable steps and server verbose mode `--verbose`
- - `SERVER_LOG_FORMAT_JSON` -> if set switch server logs to json format
-
-### Run @bug, @wip or @wrong_usage annotated scenario
-
-Feature or Scenario must be annotated with `@llama.cpp` to be included in the default scope.
-- `@bug` annotation aims to link a scenario with a GitHub issue.
-- `@wrong_usage` are meant to show user issue that are actually an expected behavior
-- `@wip` to focus on a scenario working in progress
-
-To run a scenario annotated with `@bug`, start:
-`DEBUG=ON ./tests.sh --no-skipped --tags bug`
-
-After changing logic in `steps.py`, ensure that `@bug` and `@wrong_usage` scenario are updated.

examples/server/tests/features/environment.py

@@ -1,69 +0,0 @@
-import os
-import socket
-import subprocess
-import time
-from contextlib import closing
-from signal import SIGKILL
-
-
-def before_scenario(context, scenario):
-    print(f"\x1b[33;42mStarting new scenario: {scenario.name}!\x1b[0m")
-    port = 8080
-    if 'PORT' in os.environ:
-        port = int(os.environ['PORT'])
-    if is_server_listening("localhost", port):
-        assert False, "Server already started"
-
-
-def after_scenario(context, scenario):
-    if context.server_process is None:
-        return
-    if scenario.status == "failed":
-        if 'GITHUB_ACTIONS' in os.environ:
-            print(f"\x1b[33;101mSCENARIO FAILED: {scenario.name} server logs:\x1b[0m\n\n")
-            if os.path.isfile('llama.log'):
-                with closing(open('llama.log', 'r')) as f:
-                    for line in f:
-                        print(line)
-        if not is_server_listening(context.server_fqdn, context.server_port):
-            print("\x1b[33;101mERROR: Server stopped listening\x1b[0m")
-
-    if not pid_exists(context.server_process.pid):
-        assert False, f"Server not running pid={context.server_process.pid} ..."
-
-    print(f"stopping server pid={context.server_process.pid} ...")
-    context.server_process.kill()
-    # Wait few for socket to free up
-    time.sleep(0.05)
-
-    attempts = 0
-    while is_server_listening(context.server_fqdn, context.server_port):
-        print(f"stopping server pid={context.server_process.pid} ...")
-        os.kill(context.server_process.pid, SIGKILL)
-        time.sleep(0.1)
-        attempts += 1
-        if attempts > 5:
-            print(f"Server dangling exits, killing all {context.server_path} ...")
-            process = subprocess.run(['killall', '-9', context.server_path],
-                                     stderr=subprocess.PIPE,
-                                     universal_newlines=True)
-            print(process)
-
-
-def is_server_listening(server_fqdn, server_port):
-    with closing(socket.socket(socket.AF_INET, socket.SOCK_STREAM)) as sock:
-        result = sock.connect_ex((server_fqdn, server_port))
-        return result == 0
-
-
-def pid_exists(pid):
-    """Check whether pid exists in the current process table."""
-    import errno
-    if pid < 0:
-        return False
-    try:
-        os.kill(pid, 0)
-    except OSError as e:
-        return e.errno == errno.EPERM
-    else:
-        return True

examples/server/tests/features/issues.feature

@@ -1,4 +0,0 @@
-# List of ongoing issues
-@bug
-Feature: Issues
-  # No confirmed issue at the moment

examples/server/tests/features/parallel.feature

@@ -1,123 +0,0 @@
-@llama.cpp
-Feature: Parallel
-
-  Background: Server startup
-    Given a server listening on localhost:8080
-    And   a model file stories260K.gguf
-    And   a model alias tinyllama-2
-    And   42 as server seed
-    And   64 KV cache size
-    And   2 slots
-    And   embeddings extraction
-    And   continuous batching
-    Then  the server is starting
-    Then  the server is healthy
-
-  Scenario Outline: Multi users completion
-    Given a prompt:
-      """
-      Write a very long story about AI.
-      """
-    And a prompt:
-      """
-      Write another very long music lyrics.
-      """
-    And <n_predict> max tokens to predict
-    Given concurrent completion requests
-    Then the server is busy
-    Then the server is idle
-    And  all slots are idle
-    Then all prompts are predicted with <n_predict> tokens
-    Examples:
-      | n_predict |
-      | 128       |
-
-  Scenario Outline: Multi users OAI completions compatibility
-    Given a system prompt You are a writer.
-    And   a model tinyllama-2
-    Given a prompt:
-      """
-      Write a very long book.
-      """
-    And a prompt:
-      """
-      Write another a poem.
-      """
-    And <n_predict> max tokens to predict
-    And streaming is <streaming>
-    Given concurrent OAI completions requests
-    Then the server is busy
-    Then the server is idle
-    Then all prompts are predicted with <n_predict> tokens
-    Examples:
-      | streaming | n_predict |
-      | disabled  | 128       |
-      | enabled   | 64        |
-
-  Scenario:  Multi users with total number of tokens to predict exceeds the KV Cache size #3969
-    Given a prompt:
-      """
-      Write a very long story about AI.
-      """
-    And a prompt:
-      """
-      Write another very long music lyrics.
-      """
-    And a prompt:
-      """
-      Write a very long poem.
-      """
-    And a prompt:
-      """
-      Write a very long joke.
-      """
-    And 128 max tokens to predict
-    Given concurrent completion requests
-    Then the server is busy
-    Then the server is idle
-    Then all prompts are predicted
-
-  Scenario: Multi users embeddings
-    Given a prompt:
-      """
-      Write a very long story about AI.
-      """
-    And a prompt:
-      """
-      Write another very long music lyrics.
-      """
-    And a prompt:
-      """
-      Write a very long poem.
-      """
-    And a prompt:
-      """
-      Write a very long joke.
-      """
-    Given concurrent embedding requests
-    Then the server is busy
-    Then the server is idle
-    Then all embeddings are generated
-
-  Scenario: Multi users OAI compatibility embeddings
-    Given a prompt:
-      """
-      In which country Paris is located ?
-      """
-    And a prompt:
-      """
-      Is Madrid the capital of Spain ?
-      """
-    And a prompt:
-      """
-      What is the biggest US city ?
-      """
-    And a prompt:
-      """
-      What is the capital of Bulgaria ?
-      """
-    And   a model tinyllama-2
-    Given concurrent OAI embedding requests
-    Then the server is busy
-    Then the server is idle
-    Then all embeddings are generated

examples/server/tests/features/security.feature

@@ -1,50 +0,0 @@
-@llama.cpp
-Feature: Security
-
-  Background: Server startup with an api key defined
-    Given a server listening on localhost:8080
-    And   a model file stories260K.gguf
-    And   a server api key llama.cpp
-    Then  the server is starting
-    Then  the server is healthy
-
-  Scenario Outline: Completion with some user api key
-    Given a prompt test
-    And   a user api key <api_key>
-    And   4 max tokens to predict
-    And   a completion request with <api_error> api error
-
-    Examples: Prompts
-      | api_key   | api_error |
-      | llama.cpp | no        |
-      | llama.cpp | no        |
-      | hackeme   | raised    |
-      |           | raised    |
-
-  Scenario Outline: OAI Compatibility
-    Given a system prompt test
-    And   a user prompt test
-    And   a model test
-    And   2 max tokens to predict
-    And   streaming is disabled
-    And   a user api key <api_key>
-    Given an OAI compatible chat completions request with <api_error> api error
-
-    Examples: Prompts
-      | api_key   | api_error |
-      | llama.cpp | no        |
-      | llama.cpp | no        |
-      | hackme    | raised    |
-
-
-  Scenario Outline: CORS Options
-    When an OPTIONS request is sent from <origin>
-    Then CORS header <cors_header> is set to <cors_header_value>
-
-    Examples: Headers
-      | origin          | cors_header                      | cors_header_value |
-      | localhost       | Access-Control-Allow-Origin      | localhost         |
-      | web.mydomain.fr | Access-Control-Allow-Origin      | web.mydomain.fr   |
-      | origin          | Access-Control-Allow-Credentials | true              |
-      | web.mydomain.fr | Access-Control-Allow-Methods     | POST              |
-      | web.mydomain.fr | Access-Control-Allow-Headers     | *                 |

examples/server/tests/features/server.feature

@@ -1,84 +0,0 @@
-@llama.cpp
-Feature: llama.cpp server
-
-  Background: Server startup
-    Given a server listening on localhost:8080
-    And   a model file stories260K.gguf
-    And   a model alias tinyllama-2
-    And   42 as server seed
-      # KV Cache corresponds to the total amount of tokens
-      # that can be stored across all independent sequences: #4130
-      # see --ctx-size and #5568
-    And   32 KV cache size
-    And   1 slots
-    And   embeddings extraction
-    And   32 server max tokens to predict
-    And   prometheus compatible metrics exposed
-    Then  the server is starting
-    Then  the server is healthy
-
-  Scenario: Health
-    Then the server is ready
-    And  all slots are idle
-
-  Scenario Outline: Completion
-    Given a prompt <prompt>
-    And   <n_predict> max tokens to predict
-    And   a completion request with no api error
-    Then  <n_predicted> tokens are predicted matching <re_content>
-    And   prometheus metrics are exposed
-
-    Examples: Prompts
-      | prompt                           | n_predict | re_content                             | n_predicted |
-      | I believe the meaning of life is | 8         | (read<or>going)+                       | 8           |
-      | Write a joke about AI            | 64        | (park<or>friends<or>scared<or>always)+ | 32          |
-
-  Scenario Outline: OAI Compatibility
-    Given a model <model>
-    And   a system prompt <system_prompt>
-    And   a user prompt <user_prompt>
-    And   <max_tokens> max tokens to predict
-    And   streaming is <enable_streaming>
-    Given an OAI compatible chat completions request with no api error
-    Then  <n_predicted> tokens are predicted matching <re_content>
-
-    Examples: Prompts
-      | model        | system_prompt               | user_prompt                          | max_tokens | re_content                 | n_predicted | enable_streaming |
-      | llama-2      | Book                        | What is the best book                | 8          | (Mom<or>what)+             | 8           | disabled         |
-      | codellama70b | You are a coding assistant. | Write the fibonacci function in c++. | 64         | (thanks<or>happy<or>bird)+ | 32          | enabled          |
-
-  Scenario: Embedding
-    When embeddings are computed for:
-    """
-    What is the capital of Bulgaria ?
-    """
-    Then embeddings are generated
-
-  Scenario: OAI Embeddings compatibility
-    Given a model tinyllama-2
-    When an OAI compatible embeddings computation request for:
-    """
-    What is the capital of Spain ?
-    """
-    Then embeddings are generated
-
-  Scenario: OAI Embeddings compatibility with multiple inputs
-    Given a model tinyllama-2
-    Given a prompt:
-      """
-      In which country Paris is located ?
-      """
-    And a prompt:
-      """
-      Is Madrid the capital of Spain ?
-      """
-    When an OAI compatible embeddings computation request for multiple inputs
-    Then embeddings are generated
-
-
-  Scenario: Tokenize / Detokenize
-    When tokenizing:
-    """
-    What is the capital of France ?
-    """
-    Then tokens can be detokenize

examples/server/tests/features/steps/steps.py

@@ -1,803 +0,0 @@
-import asyncio
-import collections
-import json
-import os
-import re
-import socket
-import subprocess
-import time
-from contextlib import closing
-from re import RegexFlag
-
-import aiohttp
-import openai
-from behave import step
-from behave.api.async_step import async_run_until_complete
-from prometheus_client import parser
-
-
-@step(u"a server listening on {server_fqdn}:{server_port}")
-def step_server_config(context, server_fqdn, server_port):
-    context.server_fqdn = server_fqdn
-    context.server_port = int(server_port)
-    if 'PORT' in os.environ:
-        context.server_port = int(os.environ['PORT'])
-        print(f"$PORT set, overriding server port with to {context.server_port}")
-
-    context.base_url = f'http://{context.server_fqdn}:{context.server_port}'
-
-    context.debug = 'DEBUG' in os.environ and os.environ['DEBUG'] == 'ON'
-    context.model_alias = None
-    context.n_ctx = None
-    context.n_predict = None
-    context.n_server_predict = None
-    context.n_slots = None
-    context.server_api_key = None
-    context.server_continuous_batching = False
-    context.server_embeddings = False
-    context.server_metrics = False
-    context.server_process = None
-    context.server_seed = None
-    context.user_api_key = None
-
-    context.tasks_result = []
-    context.concurrent_tasks = []
-    context.prompts = []
-
-
-@step(u'a model file {model_file}')
-def step_model_file(context, model_file):
-    context.model_file = model_file
-
-
-@step(u'a model alias {model_alias}')
-def step_model_alias(context, model_alias):
-    context.model_alias = model_alias
-
-
-@step(u'{seed} as server seed')
-def step_seed(context, seed):
-    context.server_seed = int(seed)
-
-
-@step(u'{n_ctx} KV cache size')
-def step_n_ctx(context, n_ctx):
-    context.n_ctx = int(n_ctx)
-
-
-@step(u'{n_slots} slots')
-def step_n_slots(context, n_slots):
-    context.n_slots = int(n_slots)
-
-
-@step(u'{n_predict} server max tokens to predict')
-def step_server_n_predict(context, n_predict):
-    context.n_server_predict = int(n_predict)
-
-
-@step(u'continuous batching')
-def step_server_continuous_batching(context):
-    context.server_continuous_batching = True
-
-
-@step(u'embeddings extraction')
-def step_server_embeddings(context):
-    context.server_embeddings = True
-
-
-@step(u'prometheus compatible metrics exposed')
-def step_server_metrics(context):
-    context.server_metrics = True
-
-
-@step(u"the server is starting")
-def step_start_server(context):
-    start_server_background(context)
-    attempts = 0
-    while True:
-        with closing(socket.socket(socket.AF_INET, socket.SOCK_STREAM)) as sock:
-            result = sock.connect_ex((context.server_fqdn, context.server_port))
-            if result == 0:
-                print("\x1b[33;46mserver started!\x1b[0m")
-                return
-            attempts += 1
-            if attempts > 20:
-                assert False, "server not started"
-            print(f"waiting for server to start, connect error code = {result}...")
-            time.sleep(0.1)
-
-
-@step(u"the server is {expecting_status}")
-@async_run_until_complete
-async def step_wait_for_the_server_to_be_started(context, expecting_status):
-    match expecting_status:
-        case 'healthy':
-            await wait_for_health_status(context, context.base_url, 200, 'ok')
-
-        case 'ready' | 'idle':
-            await wait_for_health_status(context, context.base_url, 200, 'ok',
-                                         params={'fail_on_no_slot': 0, 'include_slots': 0},
-                                         slots_idle=context.n_slots,
-                                         slots_processing=0,
-                                         expected_slots=[{'id': slot_id, 'state': 0}
-                                                         for slot_id in range(context.n_slots)])
-        case 'busy':
-            await wait_for_health_status(context, context.base_url, 503,
-                                         'no slot available',
-                                         params={'fail_on_no_slot': 0, 'include_slots': 0},
-                                         slots_idle=0,
-                                         slots_processing=context.n_slots,
-                                         expected_slots=[{'id': slot_id, 'state': 1}
-                                                         for slot_id in range(context.n_slots)])
-        case _:
-            assert False, "unknown status"
-
-
-@step(u'all slots are {expected_slot_status_string}')
-@async_run_until_complete
-async def step_all_slots_status(context, expected_slot_status_string):
-    match expected_slot_status_string:
-        case 'idle':
-            expected_slot_status = 0
-        case 'busy':
-            expected_slot_status = 1
-        case _:
-            assert False, "unknown status"
-
-    expected_slots = [{'id': slot_id, 'state': expected_slot_status}
-                      for slot_id in range(context.n_slots)]
-    await request_slots_status(context, expected_slots)
-
-
-@step(u'a completion request with {api_error} api error')
-@async_run_until_complete
-async def step_request_completion(context, api_error):
-    expect_api_error = api_error == 'raised'
-    completion = await request_completion(context.prompts.pop(),
-                                          context.base_url,
-                                          debug=context.debug,
-                                          n_predict=context.n_predict,
-                                          server_seed=context.server_seed,
-                                          expect_api_error=expect_api_error,
-                                          user_api_key=context.user_api_key)
-    context.tasks_result.append(completion)
-    if context.debug:
-        print(f"Completion response: {completion}")
-    if expect_api_error:
-        assert completion == 401, f"completion must be an 401 status code: {completion}"
-
-
-@step(u'{predicted_n} tokens are predicted matching {re_content}')
-def step_n_tokens_predicted_with_content(context, predicted_n, re_content):
-    assert_n_tokens_predicted(context.tasks_result.pop(), int(predicted_n), re_content)
-
-
-@step(u'{predicted_n} tokens are predicted')
-def step_n_tokens_predicted(context, predicted_n):
-    assert_n_tokens_predicted(context.tasks_result.pop(), int(predicted_n))
-
-
-@step(u'a user prompt {user_prompt}')
-def step_user_prompt(context, user_prompt):
-    context.prompts.append(user_prompt)
-
-
-@step(u'a system prompt {system_prompt}')
-def step_system_prompt(context, system_prompt):
-    context.system_prompt = system_prompt
-
-
-@step(u'a model {model}')
-def step_model(context, model):
-    context.model = model
-
-
-@step(u'{max_tokens} max tokens to predict')
-def step_max_tokens(context, max_tokens):
-    context.n_predict = int(max_tokens)
-
-
-@step(u'streaming is {enable_streaming}')
-def step_streaming(context, enable_streaming):
-    context.enable_streaming = enable_streaming == 'enabled'
-
-
-@step(u'a user api key {user_api_key}')
-def step_user_api_key(context, user_api_key):
-    context.user_api_key = user_api_key
-
-
-@step(u'no user api key')
-def step_no_user_api_key(context):
-    context.user_api_key = None
-
-
-@step(u'a user api key ')
-def step_no_user_api_key_space(context):
-    context.user_api_key = None
-
-
-@step(u'a server api key {server_api_key}')
-def step_server_api_key(context, server_api_key):
-    context.server_api_key = server_api_key
-
-
-@step(u'an OAI compatible chat completions request with {api_error} api error')
-@async_run_until_complete
-async def step_oai_chat_completions(context, api_error):
-    if context.debug:
-        print(f"Submitting OAI compatible completions request...")
-    expect_api_error = api_error == 'raised'
-    completion = await oai_chat_completions(context.prompts.pop(),
-                                            context.system_prompt,
-                                            context.base_url,
-                                            False,
-                                            model=context.model if hasattr(context, 'model') else None,
-
-                                            n_predict=context.n_predict
-                                            if hasattr(context, 'n_predict') else None,
-
-                                            enable_streaming=context.enable_streaming
-                                            if hasattr(context, 'enable_streaming') else None,
-
-                                            server_seed=context.server_seed
-                                            if hasattr(context, 'server_seed') else None,
-
-                                            user_api_key=context.user_api_key
-                                            if hasattr(context, 'user_api_key') else None,
-
-                                            expect_api_error=expect_api_error)
-    context.tasks_result.append(completion)
-    if context.debug:
-        print(f"Completion response: {completion}")
-    if expect_api_error:
-        assert completion == 401, f"completion must be an 401 status code: {completion}"
-
-    if context.debug:
-        print(f"Completion response: {completion}")
-
-
-@step(u'a prompt')
-def step_a_prompt(context):
-    context.prompts.append(context.text)
-
-
-@step(u'a prompt {prompt}')
-def step_a_prompt_prompt(context, prompt):
-    context.prompts.append(prompt)
-
-
-@step(u'concurrent completion requests')
-@async_run_until_complete()
-async def step_concurrent_completion_requests(context):
-    await concurrent_requests(context,
-                              request_completion,
-                              # prompt is inserted automatically
-                              context.base_url,
-                              debug=context.debug,
-                              n_predict=context.n_predict if hasattr(context, 'n_predict') else None,
-                              server_seed=context.server_seed if hasattr(context, 'server_seed') else None,
-                              user_api_key=context.user_api_key if hasattr(context,
-                                                                           'user_api_key') else None)
-
-
-@step(u'concurrent OAI completions requests')
-@async_run_until_complete
-async def step_oai_chat_completions(context):
-    await concurrent_requests(context, oai_chat_completions,
-                              # user_prompt is inserted automatically
-                              context.system_prompt,
-                              context.base_url,
-                              True,  # async_client
-                              model=context.model
-                              if hasattr(context, 'model') else None,
-                              n_predict=context.n_predict
-                              if hasattr(context, 'n_predict') else None,
-                              enable_streaming=context.enable_streaming
-                              if hasattr(context, 'enable_streaming') else None,
-                              server_seed=context.server_seed
-                              if hasattr(context, 'server_seed') else None,
-                              user_api_key=context.user_api_key
-                              if hasattr(context, 'user_api_key') else None)
-
-
-@step(u'all prompts are predicted')
-@async_run_until_complete
-async def step_all_prompts_are_predicted(context):
-    await all_prompts_are_predicted(context)
-
-
-@step(u'all prompts are predicted with {n_predict} tokens')
-@async_run_until_complete
-async def step_all_prompts_are_predicted_with_n_tokens(context, n_predict):
-    expected_predicted_n = int(n_predict)
-    await all_prompts_are_predicted(context, expected_predicted_n)
-
-
-async def all_prompts_are_predicted(context, expected_predicted_n=None):
-    n_completions = await gather_tasks_results(context)
-    assert n_completions > 0
-    for i in range(n_completions):
-        assert_n_tokens_predicted(context.tasks_result.pop(), expected_predicted_n=expected_predicted_n)
-    assert len(context.concurrent_tasks) == 0, f"{len(context.concurrent_tasks)} pending requests"
-
-
-@step(u'embeddings are computed for')
-@async_run_until_complete
-async def step_compute_embedding(context):
-    context.embeddings = await request_embedding(context.text, base_url=context.base_url)
-
-
-@step(u'embeddings are generated')
-def step_assert_embeddings(context):
-    if len(context.prompts) == 0:
-        assert_embeddings(context.embeddings)
-    else:
-        assert len(context.embeddings) == len(context.prompts), (f"unexpected response:\n"
-                                                                 f"context.prompts={context.prompts}\n"
-                                                                 f"context.embeddings={context.embeddings}")
-        for embedding in context.embeddings:
-            context.prompts.pop()
-            assert_embeddings(embedding)
-
-
-@step(u'an OAI compatible embeddings computation request for')
-@async_run_until_complete
-async def step_oai_compute_embeddings(context):
-    context.embeddings = await request_oai_embeddings(context.text,
-                                                      base_url=context.base_url,
-                                                      user_api_key=context.user_api_key,
-                                                      model=context.model)
-
-
-@step(u'an OAI compatible embeddings computation request for multiple inputs')
-@async_run_until_complete
-async def step_oai_compute_embeddings_multiple_inputs(context):
-    context.embeddings = await request_oai_embeddings(context.prompts,
-                                                      base_url=context.base_url,
-                                                      user_api_key=context.user_api_key,
-                                                      model=context.model)
-
-
-@step(u'concurrent embedding requests')
-@async_run_until_complete()
-async def step_concurrent_embedding_requests(context):
-    await concurrent_requests(context,
-                              request_embedding,
-                              # prompt is inserted automatically
-                              base_url=context.base_url)
-
-
-@step(u'concurrent OAI embedding requests')
-@async_run_until_complete()
-async def step_concurrent_oai_embedding_requests(context):
-    await concurrent_requests(context,
-                              request_oai_embeddings,
-                              # prompt is inserted automatically
-                              base_url=context.base_url,
-                              async_client=True,
-                              model=context.model)
-
-
-@step(u'all embeddings are generated')
-@async_run_until_complete()
-async def all_embeddings_are_generated(context):
-    n_embedding_requests = await gather_tasks_results(context)
-    assert n_embedding_requests > 0
-    for i in range(n_embedding_requests):
-        assert_embeddings(context.tasks_result.pop())
-
-
-@step(u'tokenizing')
-@async_run_until_complete
-async def step_tokenize(context):
-    context.tokenized_text = context.text
-    async with aiohttp.ClientSession() as session:
-        async with session.post(f'{context.base_url}/tokenize',
-                                json={
-                                    "content": context.tokenized_text,
-                                }) as response:
-            assert response.status == 200
-            tokenize_json = await response.json()
-            context.tokens = tokenize_json['tokens']
-
-
-@step(u'tokens can be detokenize')
-@async_run_until_complete
-async def step_detokenize(context):
-    assert len(context.tokens) > 0
-    async with aiohttp.ClientSession() as session:
-        async with session.post(f'{context.base_url}/detokenize',
-                                json={
-                                    "tokens": context.tokens,
-                                }) as response:
-            assert response.status == 200
-            detokenize_json = await response.json()
-            # SPM tokenizer adds a whitespace prefix: https://github.com/google/sentencepiece/issues/15
-            assert context.tokenized_text == detokenize_json['content'].strip()
-
-
-@step(u'an OPTIONS request is sent from {origin}')
-@async_run_until_complete
-async def step_options_request(context, origin):
-    async with aiohttp.ClientSession() as session:
-        async with session.options(f'{context.base_url}/v1/chat/completions',
-                                   headers={"Origin": origin}) as response:
-            assert response.status == 200
-            context.options_response = response
-
-
-@step(u'CORS header {cors_header} is set to {cors_header_value}')
-def step_check_options_header_value(context, cors_header, cors_header_value):
-    assert context.options_response.headers[cors_header] == cors_header_value
-
-
-@step(u'prometheus metrics are exposed')
-@async_run_until_complete
-async def step_prometheus_metrics_exported(context):
-    async with aiohttp.ClientSession() as session:
-        async with await session.get(f'{context.base_url}/metrics') as metrics_response:
-            assert metrics_response.status == 200
-            assert metrics_response.headers['Content-Type'] == "text/plain; version=0.0.4"
-            metrics_raw = await metrics_response.text()
-            metric_exported = False
-            for metric in parser.text_string_to_metric_families(metrics_raw):
-                match metric.name:
-                    case "llamacpp:kv_cache_usage_ratio":
-                        assert len(metric.samples) > 0
-                        metric_exported = True
-            assert metric_exported, "No metrics exported"
-
-
-async def concurrent_requests(context, f_completion, *args, **kwargs):
-    n_prompts = len(context.prompts)
-    if context.debug:
-        print(f"starting {n_prompts} concurrent completion requests...")
-    assert n_prompts > 0
-    for prompt_no in range(n_prompts):
-        shifted_args = [context.prompts.pop(), *args]
-        context.concurrent_tasks.append(asyncio.create_task(f_completion(*shifted_args, **kwargs)))
-    await asyncio.sleep(0.1)
-
-
-async def request_completion(prompt,
-                             base_url,
-                             debug=False,
-                             n_predict=None,
-                             server_seed=None,
-                             expect_api_error=None,
-                             user_api_key=None):
-    if debug:
-        print(f"Sending completion request: {prompt}")
-    origin = "my.super.domain"
-    headers = {
-        'Origin': origin
-    }
-    if user_api_key is not None:
-        if debug:
-            print(f"Set user_api_key: {user_api_key}")
-        headers['Authorization'] = f'Bearer {user_api_key}'
-
-    async with aiohttp.ClientSession() as session:
-        async with session.post(f'{base_url}/completion',
-                                json={
-                                    "prompt": prompt,
-                                    "n_predict": int(n_predict) if n_predict is not None else -1,
-                                    "seed": server_seed if server_seed is not None else 42
-                                },
-                                headers=headers) as response:
-            if expect_api_error is None or not expect_api_error:
-                assert response.status == 200
-                assert response.headers['Access-Control-Allow-Origin'] == origin
-                return await response.json()
-            else:
-                return response.status
-
-
-async def oai_chat_completions(user_prompt,
-                               system_prompt,
-                               base_url,
-                               async_client,
-                               debug=False,
-                               model=None,
-                               n_predict=None,
-                               enable_streaming=None,
-                               server_seed=None,
-                               user_api_key=None,
-                               expect_api_error=None):
-    if debug:
-        print(f"Sending OAI Chat completions request: {user_prompt}")
-    # openai client always expects an api key
-    user_api_key = user_api_key if user_api_key is not None else 'nope'
-    seed = server_seed if server_seed is not None else 42
-    enable_streaming = enable_streaming if enable_streaming is not None else False
-    payload = {
-        "messages": [
-            {
-                "role": "system",
-                "content": system_prompt,
-            },
-            {
-                "role": "user",
-                "content": user_prompt,
-            }
-        ],
-        "model": model,
-        "max_tokens": n_predict,
-        "stream": enable_streaming,
-        "seed": seed
-    }
-    completion_response = {
-        'content': '',
-        'timings': {
-            'predicted_n': 0
-        }
-    }
-    if async_client:
-        origin = 'llama.cpp'
-        headers = {'Authorization': f'Bearer {user_api_key}', 'Origin': origin}
-        async with aiohttp.ClientSession() as session:
-            async with session.post(f'{base_url}/v1/chat/completions',
-                                    json=payload,
-                                    headers=headers) as response:
-                if enable_streaming:
-                    assert response.status == 200
-                    assert response.headers['Access-Control-Allow-Origin'] == origin
-                    assert response.headers['Content-Type'] == "text/event-stream"
-                    event_received = True
-                    while event_received:
-                        event_received = False
-                        async for line_in_bytes in response.content:
-                            line = line_in_bytes.decode('utf8')
-                            line = line.rstrip('\n').rstrip('\r')
-                            if line == '':
-                                continue
-                            event_data = line.split(': ', 1)
-                            assert event_data[0] == 'data', f'Bad event code received: ```{event_data}```'
-                            chunk_raw = event_data[1]
-
-                            chunk = json.loads(chunk_raw)
-                            assert len(chunk['choices']) == 1, f"no choices provided, line ```{line}```"
-                            delta = chunk['choices'][0]['delta']
-                            if 'content' in delta:
-                                completion_response['content'] += delta['content']
-                                completion_response['timings']['predicted_n'] += 1
-                else:
-                    if expect_api_error is None or not expect_api_error:
-                        assert response.status == 200
-                        assert response.headers['Access-Control-Allow-Origin'] == origin
-                        assert response.headers['Content-Type'] == "application/json; charset=utf-8"
-                        chat_completion_raw = await response.json()
-                        completion_response = {
-                            'content': chat_completion_raw['choices'][0]['message'],
-                            'timings': {
-                                'predicted_n': chat_completion_raw['usage']['completion_tokens']
-                            }
-                        }
-                    else:
-                        return response.status
-    else:
-        try:
-            openai.api_key = user_api_key
-            openai.api_base = f'{base_url}/v1/chat'
-            chat_completion = openai.Completion.create(
-                messages=payload['messages'],
-                model=model,
-                max_tokens=n_predict,
-                stream=enable_streaming,
-                seed=seed
-            )
-        except openai.error.APIError as e:
-            if expect_api_error is not None and expect_api_error:
-                return 401
-            else:
-                assert False, f'error raised: {e}'
-
-        if enable_streaming:
-            for chunk in chat_completion:
-                assert len(chunk.choices) == 1
-                delta = chunk.choices[0].delta
-                if 'content' in delta:
-                    completion_response['content'] += delta['content']
-                    completion_response['timings']['predicted_n'] += 1
-        else:
-            assert len(chat_completion.choices) == 1
-            completion_response = {
-                'content': chat_completion.choices[0].message.content,
-                'timings': {
-                    'predicted_n': chat_completion.usage.completion_tokens
-                }
-            }
-    if debug:
-        print("OAI response formatted to llama.cpp:", completion_response)
-    return completion_response
-
-
-async def request_embedding(content, base_url=None):
-    async with aiohttp.ClientSession() as session:
-        async with session.post(f'{base_url}/embedding',
-                                json={
-                                    "content": content,
-                                }) as response:
-            assert response.status == 200
-            response_json = await response.json()
-            return response_json['embedding']
-
-
-async def request_oai_embeddings(input,
-                                 base_url=None, user_api_key=None,
-                                 model=None, async_client=False):
-    # openai client always expects an api_key
-    user_api_key = user_api_key if user_api_key is not None else 'nope'
-    if async_client:
-        origin = 'llama.cpp'
-        if user_api_key is not None:
-            headers = {'Authorization': f'Bearer {user_api_key}', 'Origin': origin}
-        async with aiohttp.ClientSession() as session:
-            async with session.post(f'{base_url}/v1/embeddings',
-                                    json={
-                                        "input": input,
-                                        "model": model,
-                                    },
-                                    headers=headers) as response:
-                assert response.status == 200, f"received status code not expected: {response.status}"
-                assert response.headers['Access-Control-Allow-Origin'] == origin
-                assert response.headers['Content-Type'] == "application/json; charset=utf-8"
-                response_json = await response.json()
-                assert response_json['model'] == model, f"invalid model received: {response_json['model']}"
-                assert response_json['object'] == 'list'
-                return response_json['data']
-    else:
-        openai.api_key = user_api_key
-        openai.api_base = f'{base_url}/v1'
-        oai_embeddings = openai.Embedding.create(
-            model=model,
-            input=input,
-        )
-
-        if isinstance(input, collections.abc.Sequence):
-            embeddings = []
-            for an_oai_embeddings in oai_embeddings.data:
-                embeddings.append(an_oai_embeddings.embedding)
-        else:
-            embeddings = oai_embeddings.data.embedding
-        return embeddings
-
-
-def assert_n_tokens_predicted(completion_response, expected_predicted_n=None, re_content=None):
-    content = completion_response['content']
-    n_predicted = completion_response['timings']['predicted_n']
-    assert len(content) > 0, "no token predicted"
-    if expected_predicted_n is not None:
-        assert n_predicted == expected_predicted_n, (f'invalid number of tokens predicted:'
-                                                     f' {n_predicted} <> {expected_predicted_n}')
-    if re_content is not None:
-        re_content = '^.*' + re_content.replace('<or>', '|') + '.*$'
-        assert re.match(re_content, content, flags=RegexFlag.IGNORECASE | RegexFlag.MULTILINE | RegexFlag.DOTALL), (
-            f'invalid tokens predicted:'
-            f' ```\n{content}\n``` do not match /{re_content}/')
-
-
-async def gather_tasks_results(context):
-    n_tasks = len(context.concurrent_tasks)
-    if context.debug:
-        print(f"Waiting for all {n_tasks} tasks results...")
-    for task_no in range(n_tasks):
-        context.tasks_result.append(await context.concurrent_tasks.pop())
-    n_completions = len(context.tasks_result)
-    return n_completions
-
-
-async def wait_for_health_status(context,
-                                 base_url,
-                                 expected_http_status_code,
-                                 expected_health_status,
-                                 params=None,
-                                 slots_idle=None,
-                                 slots_processing=None,
-                                 expected_slots=None):
-    if context.debug:
-        print(f"Starting checking for health for expected_health_status={expected_health_status}")
-    timeout = 3  # seconds
-    if expected_health_status == 'ok':
-        timeout = 10 # CI slow inference
-    interval = 0.5
-    counter = 0
-    async with aiohttp.ClientSession() as session:
-        while True:
-            async with await session.get(f'{base_url}/health', params=params) as health_response:
-                status_code = health_response.status
-                health = await health_response.json()
-                if context.debug:
-                    print(f"HEALTH - response for expected health status='{expected_health_status}' on "
-                          f"'{base_url}/health'?{params} is {health}")
-                if (status_code == expected_http_status_code
-                        and health['status'] == expected_health_status
-                        and (slots_idle is None or health['slots_idle'] == slots_idle)
-                        and (slots_processing is None or health['slots_processing'] == slots_processing)):
-                    if expected_slots is not None:
-                        assert_slots_status(health['slots'], expected_slots)
-                    return
-                if (status_code == expected_http_status_code
-                        and health['status'] == expected_health_status
-                        and (slots_idle is None or health['slots_idle'] == slots_idle)
-                        and (slots_processing is None or health['slots_processing'] == slots_processing)):
-                    if expected_slots is not None:
-                        assert_slots_status(health['slots'], expected_slots)
-                    return
-            await asyncio.sleep(interval)
-
-            counter += interval
-            if counter >= timeout:
-                # Sometimes health requests are triggered after completions are predicted
-                if expected_http_status_code == 503:
-                    if len(context.tasks_result) == 0:
-                        print("\x1b[5;37;43mWARNING: forcing concurrent tasks,"
-                              " busy health check missed, probably too fast inference\x1b[0m")
-                        n_completions = await gather_tasks_results(context)
-                        if n_completions > 0:
-                            return
-
-                assert False, f'{expected_health_status} timeout exceeded {counter}s>={timeout}'
-
-
-def assert_embeddings(embeddings):
-    assert len(embeddings) > 0
-    embeddings_computed = False
-    for emb in embeddings:
-        if emb != 0:
-            embeddings_computed = True
-    assert embeddings_computed, f"Embeddings: {embeddings}"
-
-
-async def request_slots_status(context, expected_slots):
-    async with aiohttp.ClientSession() as session:
-        async with await session.get(f'{context.base_url}/slots') as slots_response:
-            assert slots_response.status == 200
-            slots = await slots_response.json()
-            assert_slots_status(slots, expected_slots)
-
-
-def assert_slots_status(slots, expected_slots):
-    assert len(slots) == len(expected_slots)
-    for slot_id, (expected, slot) in enumerate(zip(expected_slots, slots)):
-        for key in expected:
-            assert expected[key] == slot[key], (f"invalid slot {slot_id}"
-                                                f" expected[{key}] != slot[{key}]"
-                                                f" = {expected[key]} != {slot[key]}")
-
-
-def start_server_background(context):
-    context.server_path = '../../../build/bin/server'
-    if 'LLAMA_SERVER_BIN_PATH' in os.environ:
-        context.server_path = os.environ['LLAMA_SERVER_BIN_PATH']
-    server_args = [
-        '--host', context.server_fqdn,
-        '--port', context.server_port,
-        '--model', context.model_file
-    ]
-    if context.server_continuous_batching:
-        server_args.append('--cont-batching')
-    if context.server_embeddings:
-        server_args.append('--embedding')
-    if context.server_metrics:
-        server_args.append('--metrics')
-    if context.model_alias is not None:
-        server_args.extend(['--alias', context.model_alias])
-    if context.n_ctx is not None:
-        server_args.extend(['--ctx-size', context.n_ctx])
-    if context.n_slots is not None:
-        server_args.extend(['--parallel', context.n_slots])
-    if context.n_server_predict is not None:
-        server_args.extend(['--n-predict', context.n_server_predict])
-    if context.server_api_key is not None:
-        server_args.extend(['--api-key', context.server_api_key])
-    if context.debug:
-        server_args.append('--verbose')
-    if 'SERVER_LOG_FORMAT_JSON' not in os.environ:
-        server_args.extend(['--log-format', "text"])
-    print(f"starting server with: {context.server_path}", *server_args)
-    context.server_process = subprocess.Popen(
-        [str(arg) for arg in [context.server_path, *server_args]],
-        close_fds=True)
-    print(f"server pid={context.server_process.pid}")

examples/server/tests/features/wrong_usages.feature

@@ -1,21 +0,0 @@
-# run with ./test.sh --tags wrong_usage
-@wrong_usage
-Feature: Wrong usage of llama.cpp server
-
-  #3969 The user must always set --n-predict option
-  # to cap the number of tokens any completion request can generate
-  # or pass n_predict/max_tokens in the request.
-  Scenario: Infinite loop
-    Given a server listening on localhost:8080
-    And   a model file stories260K.gguf
-    # Uncomment below to fix the issue
-    #And   64 server max tokens to predict
-    Then  the server is starting
-    Given a prompt:
-      """
-      Go to: infinite loop
-      """
-    # Uncomment below to fix the issue
-    #And   128 max tokens to predict
-    Given concurrent completion requests
-    Then all prompts are predicted

examples/server/tests/requirements.txt

@@ -1,4 +0,0 @@
-aiohttp~=3.9.3
-behave~=1.2.6
-openai~=0.25.0
-prometheus-client~=0.20.0

examples/server/tests/tests.sh

@@ -1,12 +0,0 @@
-#!/bin/bash
-
-set -eu
-
-if [ $# -lt 1 ]
-then
-  # Start @llama.cpp scenario
-  behave --summary --stop --no-capture --exclude 'issues|wrong_usages' --tags llama.cpp
-else
-  behave "$@"
-fi
-

examples/server/utils.hpp

@@ -14,7 +14,6 @@
 using json = nlohmann::json;
 
 extern bool server_verbose;
-extern bool server_log_json;
 
 #ifndef SERVER_VERBOSE
 #define SERVER_VERBOSE 1
@@ -28,14 +27,14 @@ extern bool server_log_json;
     {                                                                    \
         if (server_verbose)                                              \
         {                                                                \
-            server_log("VERB", __func__, __LINE__, MSG, __VA_ARGS__); \
+            server_log("VERBOSE", __func__, __LINE__, MSG, __VA_ARGS__); \
         }                                                                \
     } while (0)
 #endif
 
-#define LOG_ERROR(  MSG, ...) server_log("ERR",  __func__, __LINE__, MSG, __VA_ARGS__)
-#define LOG_WARNING(MSG, ...) server_log("WARN", __func__, __LINE__, MSG, __VA_ARGS__)
-#define LOG_INFO(   MSG, ...) server_log("INFO", __func__, __LINE__, MSG, __VA_ARGS__)
+#define LOG_ERROR(  MSG, ...) server_log("ERROR",   __func__, __LINE__, MSG, __VA_ARGS__)
+#define LOG_WARNING(MSG, ...) server_log("WARNING", __func__, __LINE__, MSG, __VA_ARGS__)
+#define LOG_INFO(   MSG, ...) server_log("INFO",    __func__, __LINE__, MSG, __VA_ARGS__)
 
 //
 // parallel
@@ -50,8 +49,7 @@ enum server_state {
 enum task_type {
     TASK_TYPE_COMPLETION,
     TASK_TYPE_CANCEL,
-    TASK_TYPE_NEXT_RESPONSE,
-    TASK_TYPE_METRICS
+    TASK_TYPE_NEXT_RESPONSE
 };
 
 struct task_server {
@@ -134,48 +132,26 @@ struct completion_token_output
     std::string text_to_send;
 };
 
-static inline void server_log(const char *level, const char *function, int line, const char *message, const nlohmann::ordered_json &extra)
+static inline void server_log(const char *level, const char *function, int line,
+                       const char *message, const nlohmann::ordered_json &extra)
 {
-    std::stringstream ss_tid;
-    ss_tid << std::this_thread::get_id();
-    json log = nlohmann::ordered_json{
-        {"tid", ss_tid.str()},
+    nlohmann::ordered_json log
+    {
         {"timestamp", time(nullptr)},
+        {"level",     level},
+        {"function",  function},
+        {"line",      line},
+        {"message",   message},
     };
 
-    if (server_log_json) {
-        log.merge_patch(
-                {
-                        {"level",     level},
-                        {"function",  function},
-                        {"line",      line},
-                        {"msg",       message},
-                });
-        if (!extra.empty()) {
-            log.merge_patch(extra);
-        }
-
-        std::cout << log.dump(-1, ' ', false, json::error_handler_t::replace) << "\n" << std::flush;
-    } else {
-        char buf[1024];
-        snprintf(buf, 1024, "%4s [%24s] %s", level, function, message);
-
-        if (!extra.empty()) {
-            log.merge_patch(extra);
-        }
-        std::stringstream ss;
-        ss << buf << " |";
-        for (const auto& el : log.items())
-        {
-            const std::string value = el.value().dump(-1, ' ', false, json::error_handler_t::replace);
-            snprintf(buf, 1024, " %s=%s", el.key().c_str(), value.c_str());
-            ss << buf;
-        }
-
-        const std::string str = ss.str();
-        printf("%.*s\n", (int)str.size(), str.data());
-        fflush(stdout);
+    if (!extra.empty())
+    {
+        log.merge_patch(extra);
     }
+
+    const std::string str = log.dump(-1, ' ', false, json::error_handler_t::replace);
+    printf("%.*s\n", (int)str.size(), str.data());
+    fflush(stdout);
 }
 
 //
@@ -191,47 +167,50 @@ static T json_value(const json &body, const std::string &key, const T &default_v
         : default_value;
 }
 
-// Check if the template supplied via "--chat-template" is supported or not. Returns true if it's valid
-inline bool verify_custom_template(const std::string & tmpl) {
-    llama_chat_message chat[] = {{"user", "test"}};
-    std::vector<char> buf(1);
-    int res = llama_chat_apply_template(nullptr, tmpl.c_str(), chat, 1, true, buf.data(), buf.size());
-    return res >= 0;
+inline std::string format_llama2(std::vector<json> messages)
+{
+    std::ostringstream output;
+    bool is_inside_turn = false;
+
+    for (auto it = messages.begin(); it != messages.end(); ++it) {
+        if (!is_inside_turn) {
+            output << "[INST] ";
+        }
+        std::string role    = json_value(*it, "role", std::string("user"));
+        std::string content = json_value(*it, "content", std::string(""));
+        if (role == "system") {
+            output << "<<SYS>>\n" << content << "\n<<SYS>>\n\n";
+            is_inside_turn = true;
+        } else if (role == "user") {
+            output << content << " [/INST]";
+            is_inside_turn = true;
+        } else {
+            output << " " << content << " </s>";
+            is_inside_turn = false;
+        }
+    }
+
+    LOG_VERBOSE("format_llama2", {{"text", output.str()}});
+
+    return output.str();
 }
 
-// Format given chat. If tmpl is empty, we take the template from model metadata
-inline std::string format_chat(const struct llama_model * model, const std::string & tmpl, const std::vector<json> & messages)
+inline std::string format_chatml(std::vector<json> messages)
 {
-    size_t alloc_size = 0;
-    // vector holding all allocated string to be passed to llama_chat_apply_template
-    std::vector<std::string> str(messages.size() * 2);
-    std::vector<llama_chat_message> chat(messages.size());
+    std::ostringstream chatml_msgs;
 
-    for (size_t i = 0; i < messages.size(); ++i) {
-        auto &curr_msg = messages[i];
-        str[i*2 + 0]    = json_value(curr_msg, "role",    std::string(""));
-        str[i*2 + 1]    = json_value(curr_msg, "content", std::string(""));
-        alloc_size     += str[i*2 + 1].length();
-        chat[i].role    = str[i*2 + 0].c_str();
-        chat[i].content = str[i*2 + 1].c_str();
+    for (auto it = messages.begin(); it != messages.end(); ++it) {
+        chatml_msgs << "<|im_start|>"
+                    << json_value(*it, "role",    std::string("user")) << '\n';
+        chatml_msgs << json_value(*it, "content", std::string(""))
+                    << "<|im_end|>\n";
     }
 
-    const char * ptr_tmpl = tmpl.empty() ? nullptr : tmpl.c_str();
-    std::vector<char> buf(alloc_size * 2);
+    chatml_msgs << "<|im_start|>assistant" << '\n';
 
-    // run the first time to get the total output length
-    int32_t res = llama_chat_apply_template(model, ptr_tmpl, chat.data(), chat.size(), true, buf.data(), buf.size());
+    LOG_VERBOSE("format_chatml", {{"text", chatml_msgs.str()}});
 
-    // if it turns out that our buffer is too small, we resize it
-    if ((size_t) res > buf.size()) {
-        buf.resize(res);
-        res = llama_chat_apply_template(model, ptr_tmpl, chat.data(), chat.size(), true, buf.data(), buf.size());
-    }
-
-    std::string formatted_chat(buf.data(), res);
-    LOG_VERBOSE("formatted_chat", {{"text", formatted_chat.c_str()}});
-
-    return formatted_chat;
+    return chatml_msgs.str();
 }
 
 //
@@ -257,7 +236,6 @@ struct llama_server_queue {
         std::unique_lock<std::mutex> lock(mutex_tasks);
         if (task.id == -1) {
             task.id = id++;
-            LOG_VERBOSE("new task id", {{"new_id", task.id}});
         }
         queue_tasks.push_back(std::move(task));
         condition_tasks.notify_one();
@@ -273,9 +251,7 @@ struct llama_server_queue {
     // Get the next id for creating anew task
     int get_new_id() {
         std::unique_lock<std::mutex> lock(mutex_tasks);
-        int new_id = id++;
-        LOG_VERBOSE("new task id", {{"new_id", new_id}});
-        return new_id;
+        return id++;
     }
 
     // Register function to process a new task
@@ -316,7 +292,8 @@ struct llama_server_queue {
     void start_loop() {
         running = true;
         while (true) {
-            LOG_VERBOSE("new task may arrive", {});
+            // new task arrived
+            LOG_VERBOSE("have new task", {});
             {
                 while (true)
                 {
@@ -328,7 +305,7 @@ struct llama_server_queue {
                     task_server task = queue_tasks.front();
                     queue_tasks.erase(queue_tasks.begin());
                     lock.unlock();
-                    LOG_VERBOSE("callback_new_task", {{"task_id", task.id}});
+                    LOG_VERBOSE("callback_new_task", {});
                     callback_new_task(task);
                 }
                 LOG_VERBOSE("callback_all_task_finished", {});
@@ -409,13 +386,11 @@ struct llama_server_response {
     std::condition_variable condition_results;
 
     void add_waiting_task_id(int task_id) {
-        LOG_VERBOSE("waiting for task id", {{"task_id", task_id}});
         std::unique_lock<std::mutex> lock(mutex_results);
         waiting_task_ids.insert(task_id);
     }
 
     void remove_waiting_task_id(int task_id) {
-        LOG_VERBOSE("remove waiting for task id", {{"task_id", task_id}});
         std::unique_lock<std::mutex> lock(mutex_results);
         waiting_task_ids.erase(task_id);
     }
@@ -428,6 +403,7 @@ struct llama_server_response {
             condition_results.wait(lock, [&]{
                 return !queue_results.empty();
             });
+            LOG_VERBOSE("condition_results unblock", {});
 
             for (int i = 0; i < (int) queue_results.size(); i++)
             {
@@ -452,22 +428,22 @@ struct llama_server_response {
     // Send a new result to a waiting task_id
     void send(task_result result) {
         std::unique_lock<std::mutex> lock(mutex_results);
-        LOG_VERBOSE("send new result", {{"task_id", result.id}});
+        LOG_VERBOSE("send new result", {});
         for (auto& task_id : waiting_task_ids) {
             // LOG_TEE("waiting task id %i \n", task_id);
             // for now, tasks that have associated parent multitasks just get erased once multitask picks up the result
             if (result.multitask_id == task_id)
             {
-                LOG_VERBOSE("callback_update_multitask", {{"task_id", task_id}});
+                LOG_VERBOSE("callback_update_multitask", {});
                 callback_update_multitask(task_id, result.id, result);
                 continue;
             }
 
             if (result.id == task_id)
             {
-                LOG_VERBOSE("queue_results.push_back", {{"task_id", task_id}});
+                LOG_VERBOSE("queue_results.push_back", {});
                 queue_results.push_back(result);
-                condition_results.notify_all();
+                condition_results.notify_one();
                 return;
             }
         }

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

@@ -960,7 +960,7 @@ int main(int argc, char ** argv) {
     struct ggml_opt_context * opt   = train->opt;
 
     // set opt params from command line
-    opt->params = ggml_opt_default_params(GGML_OPT_TYPE_ADAM);
+    opt->params = ggml_opt_default_params(GGML_OPT_ADAM);
     opt->params.print_forward_graph     = false;
     opt->params.print_backward_graph    = false;
     opt->params.graph_size              = LLAMA_TRAIN_MAX_NODES;

flake.lock

@@ -20,11 +20,11 @@
     },
     "nixpkgs": {
       "locked": {
-        "lastModified": 1708655239,
-        "narHash": "sha256-ZrP/yACUvDB+zbqYJsln4iwotbH6CTZiTkANJ0AgDv4=",
+        "lastModified": 1708118438,
+        "narHash": "sha256-kk9/0nuVgA220FcqH/D2xaN6uGyHp/zoxPNUmPCMmEE=",
         "owner": "NixOS",
         "repo": "nixpkgs",
-        "rev": "cbc4211f0afffe6dfd2478a62615dd5175a13f9a",
+        "rev": "5863c27340ba4de8f83e7e3c023b9599c3cb3c80",
         "type": "github"
       },
       "original": {

flake.nix

@@ -150,7 +150,6 @@
             packages =
               {
                 default = config.legacyPackages.llamaPackages.llama-cpp;
-                vulkan = config.packages.default.override { useVulkan = true; };
               }
               // lib.optionalAttrs pkgs.stdenv.isLinux {
                 opencl = config.packages.default.override { useOpenCL = true; };
@@ -158,6 +157,7 @@
 
                 mpi-cpu = config.packages.default.override { useMpi = true; };
                 mpi-cuda = config.packages.default.override { useMpi = true; };
+                vulkan = config.packages.default.override { useVulkan = true; };
               }
               // lib.optionalAttrs (system == "x86_64-linux") {
                 rocm = config.legacyPackages.llamaPackagesRocm.llama-cpp;

ggml-alloc.c

@@ -377,9 +377,6 @@ struct ggml_gallocr {
 
     struct node_alloc * node_allocs; // [n_nodes]
     int n_nodes;
-
-    struct tensor_alloc * leaf_allocs; // [n_leafs]
-    int n_leafs;
 };
 
 ggml_gallocr_t ggml_gallocr_new_n(ggml_backend_buffer_type_t * bufts, int n_bufs) {
@@ -430,7 +427,6 @@ void ggml_gallocr_free(ggml_gallocr_t galloc) {
     free(galloc->buffers);
     free(galloc->buf_tallocs);
     free(galloc->node_allocs);
-    free(galloc->leaf_allocs);
     free(galloc);
 }
 
@@ -468,7 +464,7 @@ static void ggml_gallocr_allocate_node(ggml_gallocr_t galloc, struct ggml_tensor
             for (int i = 0; i < GGML_MAX_SRC; i++) {
                 struct ggml_tensor * parent = node->src[i];
                 if (parent == NULL) {
-                    continue;
+                    break;
                 }
 
                 // if the node's data is external, then we cannot re-use it
@@ -548,8 +544,22 @@ static void ggml_gallocr_alloc_graph_impl(ggml_gallocr_t galloc, struct ggml_cgr
     memset(galloc->hash_set.keys, 0, galloc->hash_set.size * sizeof(struct ggml_tensor *));
     memset(galloc->hash_values,   0, galloc->hash_set.size * sizeof(struct hash_node));
 
+    // allocate all graph inputs first to avoid overwriting them
+    for (int i = 0; i < graph->n_nodes; i++) {
+        if (graph->nodes[i]->flags & GGML_TENSOR_FLAG_INPUT) {
+            ggml_gallocr_allocate_node(galloc, graph->nodes[i], get_node_buffer_id(node_buffer_ids, i));
+        }
+        for (int j = 0; j < GGML_MAX_SRC; j++) {
+            if (graph->nodes[i]->src[j] == NULL) {
+                continue;
+            }
+            if (graph->nodes[i]->src[j]->flags & GGML_TENSOR_FLAG_INPUT) {
+                ggml_gallocr_allocate_node(galloc, graph->nodes[i]->src[j], get_node_buffer_id(node_buffer_ids, i));
+            }
+        }
+    }
+
     // count number of children and views
-    // allocate all graph inputs and leafs first to avoid overwriting them
     for (int i = 0; i < graph->n_nodes; i++) {
         struct ggml_tensor * node = graph->nodes[i];
 
@@ -558,37 +568,14 @@ static void ggml_gallocr_alloc_graph_impl(ggml_gallocr_t galloc, struct ggml_cgr
             ggml_gallocr_hash_get(galloc, view_src)->n_views += 1;
         }
 
-        if (node->flags & GGML_TENSOR_FLAG_INPUT) {
-            ggml_gallocr_allocate_node(galloc, graph->nodes[i], get_node_buffer_id(node_buffer_ids, i));
-        }
-
         for (int j = 0; j < GGML_MAX_SRC; j++) {
-            struct ggml_tensor * src = node->src[j];
-            if (src == NULL) {
-                continue;
-            }
-
-            ggml_gallocr_hash_get(galloc, src)->n_children += 1;
-
-            // allocate explicit inputs and leafs
-            if (src->flags & GGML_TENSOR_FLAG_INPUT || src->op == GGML_OP_NONE) {
-                ggml_gallocr_allocate_node(galloc, src, get_node_buffer_id(node_buffer_ids, i));
+            struct ggml_tensor * parent = node->src[j];
+            if (parent == NULL) {
+                break;
             }
+            ggml_gallocr_hash_get(galloc, parent)->n_children += 1;
         }
-    }
-
-    // allocate the remaining leafs that are unused on the graph
-    // these are effectively static tensors that the application is not using in the graph, but may still want to allocate for other purposes
-    for (int i = 0; i < graph->n_leafs; i++) {
-        struct ggml_tensor * leaf = graph->leafs[i];
-        struct hash_node * hn = ggml_gallocr_hash_get(galloc, leaf);
-
-        if (hn->n_children == 0) {
-            assert(!hn->allocated);
-            // since buffer ids are only given for nodes, these leafs are always allocated in the first buffer
-            ggml_gallocr_allocate_node(galloc, leaf, 0);
-        }
-    }
+   }
 
     // allocate tensors
     for (int i = 0; i < graph->n_nodes; i++) {
@@ -599,7 +586,7 @@ static void ggml_gallocr_alloc_graph_impl(ggml_gallocr_t galloc, struct ggml_cgr
         for (int j = 0; j < GGML_MAX_SRC; j++) {
             struct ggml_tensor * parent = node->src[j];
             if (parent == NULL) {
-                continue;
+                break;
             }
             ggml_gallocr_allocate_node(galloc, parent, buffer_id);
         }
@@ -611,7 +598,7 @@ static void ggml_gallocr_alloc_graph_impl(ggml_gallocr_t galloc, struct ggml_cgr
         for (int j = 0; j < GGML_MAX_SRC; j++) {
             struct ggml_tensor * parent = node->src[j];
             if (parent == NULL) {
-                continue;
+                break;
             }
             AT_PRINTF("%s", parent->name);
             if (j < GGML_MAX_SRC - 1 && node->src[j + 1] != NULL) {
@@ -624,7 +611,7 @@ static void ggml_gallocr_alloc_graph_impl(ggml_gallocr_t galloc, struct ggml_cgr
         for (int j = 0; j < GGML_MAX_SRC; j++) {
             struct ggml_tensor * parent = node->src[j];
             if (parent == NULL) {
-                continue;
+                break;
             }
             struct hash_node * p_hn = ggml_gallocr_hash_get(galloc, parent);
             p_hn->n_children -= 1;
@@ -709,18 +696,6 @@ bool ggml_gallocr_reserve_n(ggml_gallocr_t galloc, struct ggml_cgraph * graph, c
             }
         }
     }
-    if (galloc->n_leafs < graph->n_leafs) {
-        free(galloc->leaf_allocs);
-        galloc->leaf_allocs = calloc(sizeof(struct tensor_alloc), graph->n_leafs);
-        GGML_ASSERT(galloc->leaf_allocs != NULL);
-    }
-    galloc->n_leafs = graph->n_leafs;
-    for (int i = 0; i < graph->n_leafs; i++) {
-        struct ggml_tensor * leaf = graph->leafs[i];
-        struct hash_node * hn = ggml_gallocr_hash_get(galloc, leaf);
-        galloc->leaf_allocs[i].offset = hn->offset;
-        galloc->leaf_allocs[i].size_max = ggml_backend_buft_get_alloc_size(galloc->bufts[hn->buffer_id], leaf);
-    }
 
     // reallocate buffers if needed
     for (int i = 0; i < galloc->n_buffers; i++) {
@@ -747,8 +722,8 @@ bool ggml_gallocr_reserve(ggml_gallocr_t galloc, struct ggml_cgraph *graph) {
     return ggml_gallocr_reserve_n(galloc, graph, NULL);
 }
 
-static void ggml_gallocr_init_tensor(ggml_gallocr_t galloc, struct ggml_tensor * node, int buffer_id, struct tensor_alloc * tensor_alloc) {
-    assert(node->data || node->view_src || ggml_backend_buffer_get_alloc_size(galloc->buffers[buffer_id], node) <= tensor_alloc->size_max);
+static void ggml_gallocr_init_tensor(ggml_gallocr_t galloc, struct ggml_tensor * node, struct node_alloc * node_alloc, struct tensor_alloc * tensor_alloc) {
+    assert(node->data || node->view_src || ggml_backend_buffer_get_alloc_size(galloc->buffers[node_alloc->buffer_id], node) <= tensor_alloc->size_max);
 
     if (node->view_src != NULL) {
         if (node->buffer == NULL) {
@@ -757,20 +732,29 @@ static void ggml_gallocr_init_tensor(ggml_gallocr_t galloc, struct ggml_tensor *
                 // this tensor was allocated without ggml-backend
                 return;
             }
-            ggml_backend_view_init(galloc->buffers[buffer_id], node);
+            ggml_backend_view_init(galloc->buffers[node_alloc->buffer_id], node);
         }
     } else {
         if (node->data == NULL) {
             assert(tensor_alloc->offset != SIZE_MAX);
-            assert(ggml_backend_buffer_get_alloc_size(galloc->buffers[buffer_id], node) <= tensor_alloc->size_max);
-            void * base = ggml_backend_buffer_get_base(galloc->buffers[buffer_id]);
+            assert(ggml_backend_buffer_get_alloc_size(galloc->buffers[node_alloc->buffer_id], node) <= tensor_alloc->size_max);
+            void * base = ggml_backend_buffer_get_base(galloc->buffers[node_alloc->buffer_id]);
             void * addr = (char *)base + tensor_alloc->offset;
-            ggml_backend_tensor_alloc(galloc->buffers[buffer_id], node, addr);
+            ggml_backend_tensor_alloc(galloc->buffers[node_alloc->buffer_id], node, addr);
         } else {
             if (node->buffer == NULL) {
                 // this tensor was allocated without ggml-backend
                 return;
             }
+
+#ifndef NDEBUG
+            size_t offset =
+                (char *)node->data -
+                (char *)ggml_backend_buffer_get_base(node->buffer);
+            size_t size = ggml_backend_buffer_get_alloc_size(node->buffer, node);
+            assert(tensor_alloc->offset == SIZE_MAX || offset == tensor_alloc->offset);
+            assert(tensor_alloc->offset == SIZE_MAX || size <= tensor_alloc->size_max);
+#endif
         }
     }
 }
@@ -789,13 +773,6 @@ static bool ggml_gallocr_needs_realloc(ggml_gallocr_t galloc, struct ggml_cgraph
         return true;
     }
 
-    if (galloc->n_leafs != graph->n_leafs) {
-#ifndef NDEBUG
-        fprintf(stderr, "%s: graph has different number of leafs\n", __func__);
-#endif
-        return true;
-    }
-
     for (int i = 0; i < graph->n_nodes; i++) {
         struct ggml_tensor * node = graph->nodes[i];
         struct node_alloc * node_alloc = &galloc->node_allocs[i];
@@ -850,7 +827,6 @@ bool ggml_gallocr_alloc_graph(ggml_gallocr_t galloc, struct ggml_cgraph * graph)
     }
 
     // allocate the graph tensors from the previous assignments
-    // nodes
     for (int i = 0; i < graph->n_nodes; i++) {
         struct ggml_tensor * node = graph->nodes[i];
         struct node_alloc * node_alloc = &galloc->node_allocs[i];
@@ -859,15 +835,9 @@ bool ggml_gallocr_alloc_graph(ggml_gallocr_t galloc, struct ggml_cgraph * graph)
             if (src == NULL) {
                 continue;
             }
-            ggml_gallocr_init_tensor(galloc, src, node_alloc->buffer_id, &node_alloc->src[j]);
+            ggml_gallocr_init_tensor(galloc, src, node_alloc, &node_alloc->src[j]);
         }
-        ggml_gallocr_init_tensor(galloc, node, node_alloc->buffer_id, &node_alloc->dst);
-    }
-    // leafs
-    for (int i = 0; i < graph->n_leafs; i++) {
-        struct ggml_tensor * leaf = graph->leafs[i];
-        struct tensor_alloc * leaf_alloc = &galloc->leaf_allocs[i];
-        ggml_gallocr_init_tensor(galloc, leaf, 0, leaf_alloc);
+        ggml_gallocr_init_tensor(galloc, node, node_alloc, &node_alloc->dst);
     }
 
     return true;

ggml-impl.h

@@ -53,23 +53,11 @@ extern "C" {
 //
 #include <arm_neon.h>
 
-#define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
-#define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x)
+#define GGML_COMPUTE_FP16_TO_FP32(x) ((float) (x))
+#define GGML_COMPUTE_FP32_TO_FP16(x) (x)
 
-#define GGML_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
-
-static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
-    __fp16 tmp;
-    memcpy(&tmp, &h, sizeof(ggml_fp16_t));
-    return (float)tmp;
-}
-
-static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
-    ggml_fp16_t res;
-    __fp16 tmp = f;
-    memcpy(&res, &tmp, sizeof(ggml_fp16_t));
-    return res;
-}
+#define GGML_FP16_TO_FP32(x) ((float) (x))
+#define GGML_FP32_TO_FP16(x) (x)
 
 #else
 
@@ -226,7 +214,8 @@ extern float ggml_table_f32_f16[1 << 16];
 // On ARM NEON, it's quicker to directly convert x -> x instead of calling into ggml_lookup_fp16_to_fp32,
 // so we define GGML_FP16_TO_FP32 and GGML_FP32_TO_FP16 elsewhere for NEON.
 // This is also true for POWER9.
-#if !defined(GGML_FP16_TO_FP32)
+#if !defined(GGML_FP16_TO_FP32) || !defined(GGML_FP32_TO_FP16)
+
 inline static float ggml_lookup_fp16_to_fp32(ggml_fp16_t f) {
     uint16_t s;
     memcpy(&s, &f, sizeof(uint16_t));
@@ -234,10 +223,8 @@ inline static float ggml_lookup_fp16_to_fp32(ggml_fp16_t f) {
 }
 
 #define GGML_FP16_TO_FP32(x) ggml_lookup_fp16_to_fp32(x)
-#endif
-
-#if !defined(GGML_FP32_TO_FP16)
 #define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x)
+
 #endif
 
 #define GGML_HASHTABLE_FULL ((size_t)-1)

ggml-metal.m

@@ -61,9 +61,7 @@ enum ggml_metal_kernel_type {
     GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_XXS,
     GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_XS,
     GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ3_XXS,
-    GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ3_S,
     GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ1_S,
-    GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ4_NL,
     GGML_METAL_KERNEL_TYPE_GET_ROWS_I32,
     GGML_METAL_KERNEL_TYPE_RMS_NORM,
     GGML_METAL_KERNEL_TYPE_GROUP_NORM,
@@ -86,9 +84,7 @@ enum ggml_metal_kernel_type {
     GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_XXS_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_XS_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MV_IQ3_XXS_F32,
-    GGML_METAL_KERNEL_TYPE_MUL_MV_IQ3_S_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MV_IQ1_S_F32,
-    GGML_METAL_KERNEL_TYPE_MUL_MV_IQ4_NL_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F32_F32,
   //GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F16,
     GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F32,
@@ -107,9 +103,7 @@ enum ggml_metal_kernel_type {
     GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_XXS_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_XS_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ3_XXS_F32,
-    GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ3_S_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ1_S_F32,
-    GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ4_NL_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MM_F32_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MM_F16_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_0_F32,
@@ -125,9 +119,7 @@ enum ggml_metal_kernel_type {
     GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_XXS_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_XS_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MM_IQ3_XXS_F32,
-    GGML_METAL_KERNEL_TYPE_MUL_MM_IQ3_S_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MM_IQ1_S_F32,
-    GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_NL_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F32_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F16_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_0_F32,
@@ -143,9 +135,7 @@ enum ggml_metal_kernel_type {
     GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XXS_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XS_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ3_XXS_F32,
-    GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ3_S_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ1_S_F32,
-    GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_NL_F32,
     GGML_METAL_KERNEL_TYPE_ROPE_F32,
     GGML_METAL_KERNEL_TYPE_ROPE_F16,
     GGML_METAL_KERNEL_TYPE_ALIBI_F32,
@@ -156,6 +146,12 @@ enum ggml_metal_kernel_type {
     GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_ASC,
     GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_DESC,
     GGML_METAL_KERNEL_TYPE_LEAKY_RELU_F32,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H64,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H80,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H96,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H112,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H128,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H256,
     GGML_METAL_KERNEL_TYPE_CPY_F32_F16,
     GGML_METAL_KERNEL_TYPE_CPY_F32_F32,
     GGML_METAL_KERNEL_TYPE_CPY_F32_Q8_0,
@@ -287,14 +283,6 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
                 return NULL;
             }
         } else {
-#if GGML_METAL_EMBED_LIBRARY
-            GGML_METAL_LOG_INFO("%s: using embedded metal library\n", __func__);
-
-            extern const char ggml_metallib_start[];
-            extern const char ggml_metallib_end[];
-
-            NSString * src  = [[NSString alloc] initWithBytes:ggml_metallib_start length:(ggml_metallib_end-ggml_metallib_start) encoding:NSUTF8StringEncoding];
-#else
             GGML_METAL_LOG_INFO("%s: default.metallib not found, loading from source\n", __func__);
 
             NSString * sourcePath;
@@ -317,7 +305,6 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
                 GGML_METAL_LOG_ERROR("%s: error: %s\n", __func__, [[error description] UTF8String]);
                 return NULL;
             }
-#endif
 
             @autoreleasepool {
                 // dictionary of preprocessor macros
@@ -414,6 +401,9 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
             id<MTLFunction> metal_function = [metal_library newFunctionWithName:@"kernel_"#name]; \
             kernel->pipeline = [ctx->device newComputePipelineStateWithFunction:metal_function error:&error]; \
             [metal_function release]; \
+            GGML_METAL_LOG_INFO("%s: loaded %-32s %16p | th_max = %4d | th_width = %4d\n", __func__, "kernel_"#name, (void *) kernel->pipeline, \
+                    (int) kernel->pipeline.maxTotalThreadsPerThreadgroup, \
+                    (int) kernel->pipeline.threadExecutionWidth); \
             if (error) { \
                 GGML_METAL_LOG_ERROR("%s: error: load pipeline error: %s\n", __func__, [[error description] UTF8String]); \
                 [metal_library release]; \
@@ -425,145 +415,141 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
 
         // simd_sum and simd_max requires MTLGPUFamilyApple7
 
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ADD,                       add,                    true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ADD_ROW,                   add_row,                true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL,                       mul,                    true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_ROW,                   mul_row,                true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_DIV,                       div,                    true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_DIV_ROW,                   div_row,                true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SCALE,                     scale,                  true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SCALE_4,                   scale_4,                true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_TANH,                      tanh,                   true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_RELU,                      relu,                   true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GELU,                      gelu,                   true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GELU_QUICK,                gelu_quick,             true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SILU,                      silu,                   true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SOFT_MAX,                  soft_max,               ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SOFT_MAX_4,                soft_max_4,             ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_DIAG_MASK_INF,             diag_mask_inf,          true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_DIAG_MASK_INF_8,           diag_mask_inf_8,        true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_F32,              get_rows_f32,           true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_F16,              get_rows_f16,           true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q4_0,             get_rows_q4_0,          true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q4_1,             get_rows_q4_1,          true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q5_0,             get_rows_q5_0,          true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q5_1,             get_rows_q5_1,          true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q8_0,             get_rows_q8_0,          true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q2_K,             get_rows_q2_K,          true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q3_K,             get_rows_q3_K,          true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q4_K,             get_rows_q4_K,          true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q5_K,             get_rows_q5_K,          true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q6_K,             get_rows_q6_K,          true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_XXS,          get_rows_iq2_xxs,       true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_XS,           get_rows_iq2_xs,        true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ3_XXS,          get_rows_iq3_xxs,       true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ3_S,            get_rows_iq3_s,         true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ1_S,            get_rows_iq1_s,         true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ4_NL,           get_rows_iq4_nl,        true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_I32,              get_rows_i32,           true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_RMS_NORM,                  rms_norm,               ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GROUP_NORM,                group_norm,             ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_NORM,                      norm,                   true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F32_F32,            mul_mv_f32_f32,         ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F16,            mul_mv_f16_f16,         ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32,            mul_mv_f16_f32,         ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_1ROW,       mul_mv_f16_f32_1row,    ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_L4,         mul_mv_f16_f32_l4,      ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_0_F32,           mul_mv_q4_0_f32,        ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_1_F32,           mul_mv_q4_1_f32,        ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_0_F32,           mul_mv_q5_0_f32,        ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_1_F32,           mul_mv_q5_1_f32,        ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q8_0_F32,           mul_mv_q8_0_f32,        ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q2_K_F32,           mul_mv_q2_K_f32,        ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q3_K_F32,           mul_mv_q3_K_f32,        ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_K_F32,           mul_mv_q4_K_f32,        ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_K_F32,           mul_mv_q5_K_f32,        ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q6_K_F32,           mul_mv_q6_K_f32,        ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_XXS_F32,        mul_mv_iq2_xxs_f32,     ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_XS_F32,         mul_mv_iq2_xs_f32,      ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ3_XXS_F32,        mul_mv_iq3_xxs_f32,     ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ3_S_F32,          mul_mv_iq3_s_f32,       ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ1_S_F32,          mul_mv_iq1_s_f32,       ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ4_NL_F32,         mul_mv_iq4_nl_f32,      ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F32_F32,         mul_mv_id_f32_f32,      ctx->support_simdgroup_reduction);
-      //GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F16,         mul_mv_id_f16_f16,      ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F32,         mul_mv_id_f16_f32,      ctx->support_simdgroup_reduction);
-      //GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F32_1ROW,    mul_mv_id_f16_f32_1row, ctx->support_simdgroup_reduction);
-      //GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F32_L4,      mul_mv_id_f16_f32_l4,   ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q4_0_F32,        mul_mv_id_q4_0_f32,     ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q4_1_F32,        mul_mv_id_q4_1_f32,     ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q5_0_F32,        mul_mv_id_q5_0_f32,     ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q5_1_F32,        mul_mv_id_q5_1_f32,     ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q8_0_F32,        mul_mv_id_q8_0_f32,     ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q2_K_F32,        mul_mv_id_q2_K_f32,     ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q3_K_F32,        mul_mv_id_q3_K_f32,     ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q4_K_F32,        mul_mv_id_q4_K_f32,     ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q5_K_F32,        mul_mv_id_q5_K_f32,     ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q6_K_F32,        mul_mv_id_q6_K_f32,     ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_XXS_F32,     mul_mv_id_iq2_xxs_f32,  ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_XS_F32,      mul_mv_id_iq2_xs_f32,   ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ3_XXS_F32,     mul_mv_id_iq3_xxs_f32,  ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ3_S_F32,       mul_mv_id_iq3_s_f32,    ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ1_S_F32,       mul_mv_id_iq1_s_f32,    ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ4_NL_F32,      mul_mv_id_iq4_nl_f32,   ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_F32_F32,            mul_mm_f32_f32,         ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_F16_F32,            mul_mm_f16_f32,         ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_0_F32,           mul_mm_q4_0_f32,        ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_1_F32,           mul_mm_q4_1_f32,        ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_0_F32,           mul_mm_q5_0_f32,        ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_1_F32,           mul_mm_q5_1_f32,        ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q8_0_F32,           mul_mm_q8_0_f32,        ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q2_K_F32,           mul_mm_q2_K_f32,        ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q3_K_F32,           mul_mm_q3_K_f32,        ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_K_F32,           mul_mm_q4_K_f32,        ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_K_F32,           mul_mm_q5_K_f32,        ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q6_K_F32,           mul_mm_q6_K_f32,        ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_XXS_F32,        mul_mm_iq2_xxs_f32,     ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_XS_F32,         mul_mm_iq2_xs_f32,      ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ3_XXS_F32,        mul_mm_iq3_xxs_f32,     ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ3_S_F32,          mul_mm_iq3_s_f32,       ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ1_S_F32,          mul_mm_iq1_s_f32,       ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_NL_F32,         mul_mm_iq4_nl_f32,      ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F32_F32,         mul_mm_id_f32_f32,      ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F16_F32,         mul_mm_id_f16_f32,      ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_0_F32,        mul_mm_id_q4_0_f32,     ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_1_F32,        mul_mm_id_q4_1_f32,     ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_0_F32,        mul_mm_id_q5_0_f32,     ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_1_F32,        mul_mm_id_q5_1_f32,     ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q8_0_F32,        mul_mm_id_q8_0_f32,     ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q2_K_F32,        mul_mm_id_q2_K_f32,     ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q3_K_F32,        mul_mm_id_q3_K_f32,     ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_K_F32,        mul_mm_id_q4_K_f32,     ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_K_F32,        mul_mm_id_q5_K_f32,     ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q6_K_F32,        mul_mm_id_q6_K_f32,     ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XXS_F32,     mul_mm_id_iq2_xxs_f32,  ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XS_F32,      mul_mm_id_iq2_xs_f32,   ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ3_XXS_F32,     mul_mm_id_iq3_xxs_f32,  ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ3_S_F32,       mul_mm_id_iq3_s_f32,    ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ1_S_F32,       mul_mm_id_iq1_s_f32,    ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_NL_F32,      mul_mm_id_iq4_nl_f32,   ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ROPE_F32,                  rope_f32,               true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ROPE_F16,                  rope_f16,               true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ALIBI_F32,                 alibi_f32,              true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_IM2COL_F16,                im2col_f16,             true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_IM2COL_F32,                im2col_f32,             true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_UPSCALE_F32,               upscale_f32,            true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_PAD_F32,                   pad_f32,                true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_ASC,       argsort_f32_i32_asc,    true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_DESC,      argsort_f32_i32_desc,   true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_LEAKY_RELU_F32,            leaky_relu_f32,         true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_F16,               cpy_f32_f16,            true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_F32,               cpy_f32_f32,            true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_Q8_0,              cpy_f32_q8_0,           true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_Q4_0,              cpy_f32_q4_0,           true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_Q4_1,              cpy_f32_q4_1,           true);
-      //GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_Q5_0,              cpy_f32_q5_0,           true);
-      //GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_Q5_1,              cpy_f32_q5_1,           true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F16_F16,               cpy_f16_f16,            true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F16_F32,               cpy_f16_f32,            true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CONCAT,                    concat,                 true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SQR,                       sqr,                    true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SUM_ROWS,                  sum_rows,               true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ADD,                       add,                     true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ADD_ROW,                   add_row,                 true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL,                       mul,                     true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_ROW,                   mul_row,                 true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_DIV,                       div,                     true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_DIV_ROW,                   div_row,                 true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SCALE,                     scale,                   true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SCALE_4,                   scale_4,                 true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_TANH,                      tanh,                    true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_RELU,                      relu,                    true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GELU,                      gelu,                    true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GELU_QUICK,                gelu_quick,              true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SILU,                      silu,                    true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SOFT_MAX,                  soft_max,                ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SOFT_MAX_4,                soft_max_4,              ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_DIAG_MASK_INF,             diag_mask_inf,           true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_DIAG_MASK_INF_8,           diag_mask_inf_8,         true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_F32,              get_rows_f32,            true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_F16,              get_rows_f16,            true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q4_0,             get_rows_q4_0,           true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q4_1,             get_rows_q4_1,           true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q5_0,             get_rows_q5_0,           true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q5_1,             get_rows_q5_1,           true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q8_0,             get_rows_q8_0,           true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q2_K,             get_rows_q2_K,           true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q3_K,             get_rows_q3_K,           true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q4_K,             get_rows_q4_K,           true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q5_K,             get_rows_q5_K,           true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q6_K,             get_rows_q6_K,           true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_XXS,          get_rows_iq2_xxs,        true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_XS,           get_rows_iq2_xs,         true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ3_XXS,          get_rows_iq3_xxs,        true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ1_S,            get_rows_iq1_s,          true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_I32,              get_rows_i32,            true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_RMS_NORM,                  rms_norm,                ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GROUP_NORM,                group_norm,              ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_NORM,                      norm,                    true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F32_F32,            mul_mv_f32_f32,          ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F16,            mul_mv_f16_f16,          ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32,            mul_mv_f16_f32,          ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_1ROW,       mul_mv_f16_f32_1row,     ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_L4,         mul_mv_f16_f32_l4,       ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_0_F32,           mul_mv_q4_0_f32,         ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_1_F32,           mul_mv_q4_1_f32,         ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_0_F32,           mul_mv_q5_0_f32,         ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_1_F32,           mul_mv_q5_1_f32,         ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q8_0_F32,           mul_mv_q8_0_f32,         ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q2_K_F32,           mul_mv_q2_K_f32,         ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q3_K_F32,           mul_mv_q3_K_f32,         ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_K_F32,           mul_mv_q4_K_f32,         ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_K_F32,           mul_mv_q5_K_f32,         ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q6_K_F32,           mul_mv_q6_K_f32,         ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_XXS_F32,        mul_mv_iq2_xxs_f32,      ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_XS_F32,         mul_mv_iq2_xs_f32,       ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ3_XXS_F32,        mul_mv_iq3_xxs_f32,      ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ1_S_F32,          mul_mv_iq1_s_f32,        ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F32_F32,         mul_mv_id_f32_f32,       ctx->support_simdgroup_reduction);
+      //GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F16,         mul_mv_id_f16_f16,       ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F32,         mul_mv_id_f16_f32,       ctx->support_simdgroup_reduction);
+      //GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F32_1ROW,    mul_mv_id_f16_f32_1row,  ctx->support_simdgroup_reduction);
+      //GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F32_L4,      mul_mv_id_f16_f32_l4,    ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q4_0_F32,        mul_mv_id_q4_0_f32,      ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q4_1_F32,        mul_mv_id_q4_1_f32,      ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q5_0_F32,        mul_mv_id_q5_0_f32,      ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q5_1_F32,        mul_mv_id_q5_1_f32,      ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q8_0_F32,        mul_mv_id_q8_0_f32,      ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q2_K_F32,        mul_mv_id_q2_K_f32,      ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q3_K_F32,        mul_mv_id_q3_K_f32,      ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q4_K_F32,        mul_mv_id_q4_K_f32,      ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q5_K_F32,        mul_mv_id_q5_K_f32,      ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q6_K_F32,        mul_mv_id_q6_K_f32,      ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_XXS_F32,     mul_mv_id_iq2_xxs_f32,   ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_XS_F32,      mul_mv_id_iq2_xs_f32,    ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ3_XXS_F32,     mul_mv_id_iq3_xxs_f32,   ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ1_S_F32,       mul_mv_id_iq1_s_f32,     ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_F32_F32,            mul_mm_f32_f32,          ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_F16_F32,            mul_mm_f16_f32,          ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_0_F32,           mul_mm_q4_0_f32,         ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_1_F32,           mul_mm_q4_1_f32,         ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_0_F32,           mul_mm_q5_0_f32,         ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_1_F32,           mul_mm_q5_1_f32,         ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q8_0_F32,           mul_mm_q8_0_f32,         ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q2_K_F32,           mul_mm_q2_K_f32,         ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q3_K_F32,           mul_mm_q3_K_f32,         ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_K_F32,           mul_mm_q4_K_f32,         ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_K_F32,           mul_mm_q5_K_f32,         ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q6_K_F32,           mul_mm_q6_K_f32,         ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_XXS_F32,        mul_mm_iq2_xxs_f32,      ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_XS_F32,         mul_mm_iq2_xs_f32,       ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ3_XXS_F32,        mul_mm_iq3_xxs_f32,      ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ1_S_F32,          mul_mm_iq1_s_f32,        ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F32_F32,         mul_mm_id_f32_f32,       ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F16_F32,         mul_mm_id_f16_f32,       ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_0_F32,        mul_mm_id_q4_0_f32,      ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_1_F32,        mul_mm_id_q4_1_f32,      ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_0_F32,        mul_mm_id_q5_0_f32,      ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_1_F32,        mul_mm_id_q5_1_f32,      ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q8_0_F32,        mul_mm_id_q8_0_f32,      ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q2_K_F32,        mul_mm_id_q2_K_f32,      ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q3_K_F32,        mul_mm_id_q3_K_f32,      ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_K_F32,        mul_mm_id_q4_K_f32,      ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_K_F32,        mul_mm_id_q5_K_f32,      ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q6_K_F32,        mul_mm_id_q6_K_f32,      ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XXS_F32,     mul_mm_id_iq2_xxs_f32,   ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XS_F32,      mul_mm_id_iq2_xs_f32,    ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ3_XXS_F32,     mul_mm_id_iq3_xxs_f32,   ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ1_S_F32,       mul_mm_id_iq1_s_f32,     ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ROPE_F32,                  rope_f32,                true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ROPE_F16,                  rope_f16,                true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ALIBI_F32,                 alibi_f32,               true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_IM2COL_F16,                im2col_f16,              true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_IM2COL_F32,                im2col_f32,              true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_UPSCALE_F32,               upscale_f32,             true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_PAD_F32,                   pad_f32,                 true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_ASC,       argsort_f32_i32_asc,     true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_DESC,      argsort_f32_i32_desc,    true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_LEAKY_RELU_F32,            leaky_relu_f32,          true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H64,    flash_attn_ext_f16_h64,  true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H80,    flash_attn_ext_f16_h80,  true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H96,    flash_attn_ext_f16_h96,  true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H112,   flash_attn_ext_f16_h112, true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H128,   flash_attn_ext_f16_h128, true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H256,   flash_attn_ext_f16_h256, true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_F16,               cpy_f32_f16,             true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_F32,               cpy_f32_f32,             true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_Q8_0,              cpy_f32_q8_0,            true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_Q4_0,              cpy_f32_q4_0,            true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_Q4_1,              cpy_f32_q4_1,            true);
+      //GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_Q5_0,              cpy_f32_q5_0,            true);
+      //GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_Q5_1,              cpy_f32_q5_1,            true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F16_F16,               cpy_f16_f16,             true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F16_F32,               cpy_f16_f32,             true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CONCAT,                    concat,                  true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SQR,                       sqr,                     true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SUM_ROWS,                  sum_rows,                true);
     }
 
     [metal_library release];
@@ -679,6 +665,7 @@ static bool ggml_metal_supports_op(const struct ggml_metal_context * ctx, const
         case GGML_OP_PAD:
         case GGML_OP_ARGSORT:
         case GGML_OP_LEAKY_RELU:
+        case GGML_OP_FLASH_ATTN_EXT:
             return true;
         case GGML_OP_MUL_MAT:
         case GGML_OP_MUL_MAT_ID:
@@ -1214,6 +1201,8 @@ static bool ggml_metal_graph_compute(
                     } break;
                 case GGML_OP_SOFT_MAX:
                     {
+                        GGML_ASSERT(!src1 || src1->type == GGML_TYPE_F16);
+
                         int nth = 32; // SIMD width
 
                         id<MTLComputePipelineState> pipeline = nil;
@@ -1357,9 +1346,7 @@ static bool ggml_metal_graph_compute(
                                 case GGML_TYPE_IQ2_XXS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_XXS_F32].pipeline; break;
                                 case GGML_TYPE_IQ2_XS:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_XS_F32 ].pipeline; break;
                                 case GGML_TYPE_IQ3_XXS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ3_XXS_F32].pipeline; break;
-                                case GGML_TYPE_IQ3_S:   pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ3_S_F32  ].pipeline; break;
                                 case GGML_TYPE_IQ1_S:   pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ1_S_F32  ].pipeline; break;
-                                case GGML_TYPE_IQ4_NL:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_NL_F32 ].pipeline; break;
                                 default: GGML_ASSERT(false && "MUL MAT-MAT not implemented");
                             }
 
@@ -1494,24 +1481,12 @@ static bool ggml_metal_graph_compute(
                                         nth1 = 16;
                                         pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ3_XXS_F32].pipeline;
                                     } break;
-                                case GGML_TYPE_IQ3_S:
-                                    {
-                                        nth0 = 4;
-                                        nth1 = 16;
-                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ3_S_F32].pipeline;
-                                    } break;
                                 case GGML_TYPE_IQ1_S:
                                     {
                                         nth0 = 4;
                                         nth1 = 16;
                                         pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ1_S_F32].pipeline;
                                     } break;
-                                case GGML_TYPE_IQ4_NL:
-                                    {
-                                        nth0 = 4;
-                                        nth1 = 16;
-                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ4_NL_F32].pipeline;
-                                    } break;
                                 default:
                                     {
                                         GGML_METAL_LOG_ERROR("Asserting on type %d\n", (int)src0t);
@@ -1554,16 +1529,11 @@ static bool ggml_metal_graph_compute(
                                 [encoder setThreadgroupMemoryLength:mem_size atIndex:0];
                                 [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
                             }
-                            else if (src0t == GGML_TYPE_IQ3_XXS || src0t == GGML_TYPE_IQ3_S) {
-                                const int mem_size = src0t == GGML_TYPE_IQ3_XXS ? 256*4+128 : 512*4;
+                            else if (src0t == GGML_TYPE_IQ3_XXS) {
+                                const int mem_size = 256*4+128;
                                 [encoder setThreadgroupMemoryLength:mem_size atIndex:0];
                                 [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
                             }
-                            else if (src0t == GGML_TYPE_IQ4_NL) {
-                                const int mem_size = 32*sizeof(float);
-                                [encoder setThreadgroupMemoryLength:mem_size atIndex:0];
-                                [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
-                            }
                             else if (src0t == GGML_TYPE_Q4_K) {
                                 [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
                             }
@@ -1657,9 +1627,7 @@ static bool ggml_metal_graph_compute(
                                 case GGML_TYPE_IQ2_XXS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XXS_F32].pipeline; break;
                                 case GGML_TYPE_IQ2_XS:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XS_F32 ].pipeline; break;
                                 case GGML_TYPE_IQ3_XXS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ3_XXS_F32].pipeline; break;
-                                case GGML_TYPE_IQ3_S:   pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ3_S_F32  ].pipeline; break;
                                 case GGML_TYPE_IQ1_S:   pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ1_S_F32  ].pipeline; break;
-                                case GGML_TYPE_IQ4_NL:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_NL_F32 ].pipeline; break;
                                 default: GGML_ASSERT(false && "MUL_MAT_ID not implemented");
                             }
 
@@ -1797,24 +1765,12 @@ static bool ggml_metal_graph_compute(
                                         nth1 = 16;
                                         pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ3_XXS_F32].pipeline;
                                     } break;
-                                case GGML_TYPE_IQ3_S:
-                                    {
-                                        nth0 = 4;
-                                        nth1 = 16;
-                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ3_S_F32].pipeline;
-                                    } break;
                                 case GGML_TYPE_IQ1_S:
                                     {
                                         nth0 = 4;
                                         nth1 = 16;
                                         pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ1_S_F32].pipeline;
                                     } break;
-                                case GGML_TYPE_IQ4_NL:
-                                    {
-                                        nth0 = 4;
-                                        nth1 = 16;
-                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ4_NL_F32].pipeline;
-                                    } break;
                                 default:
                                     {
                                         GGML_METAL_LOG_ERROR("Asserting on type %d\n", (int)src2t);
@@ -1873,16 +1829,11 @@ static bool ggml_metal_graph_compute(
                                 [encoder setThreadgroupMemoryLength:mem_size atIndex:0];
                                 [encoder dispatchThreadgroups:MTLSizeMake((ne21 + 7)/8, _ne1, ne01*ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
                             }
-                            else if (src2t == GGML_TYPE_IQ3_XXS || src2t == GGML_TYPE_IQ3_S) {
-                                const int mem_size = src2t == GGML_TYPE_IQ3_XXS ? 256*4+128 : 512*4;
+                            else if (src2t == GGML_TYPE_IQ3_XXS) {
+                                const int mem_size = 256*4+128;
                                 [encoder setThreadgroupMemoryLength:mem_size atIndex:0];
                                 [encoder dispatchThreadgroups:MTLSizeMake((ne21 + 7)/8, _ne1, ne01*ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
                             }
-                            else if (src2t == GGML_TYPE_IQ4_NL) {
-                                const int mem_size = 32*sizeof(float);
-                                [encoder setThreadgroupMemoryLength:mem_size atIndex:0];
-                                [encoder dispatchThreadgroups:MTLSizeMake((ne21 + 3)/4, _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)];
                             }
@@ -1924,9 +1875,7 @@ static bool ggml_metal_graph_compute(
                             case GGML_TYPE_IQ2_XXS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_XXS].pipeline; break;
                             case GGML_TYPE_IQ2_XS:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_XS ].pipeline; break;
                             case GGML_TYPE_IQ3_XXS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ3_XXS].pipeline; break;
-                            case GGML_TYPE_IQ3_S:   pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ3_S  ].pipeline; break;
                             case GGML_TYPE_IQ1_S:   pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ1_S  ].pipeline; break;
-                            case GGML_TYPE_IQ4_NL:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ4_NL ].pipeline; break;
                             case GGML_TYPE_I32:     pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_I32    ].pipeline; break;
                             default: GGML_ASSERT(false && "not implemented");
                         }
@@ -2262,8 +2211,8 @@ static bool ggml_metal_graph_compute(
                         id<MTLComputePipelineState> pipeline = nil;
 
                         switch (order) {
-                            case GGML_SORT_ORDER_ASC:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_ASC].pipeline;  break;
-                            case GGML_SORT_ORDER_DESC: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_DESC].pipeline; break;
+                            case GGML_SORT_ASC:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_ASC].pipeline;  break;
+                            case GGML_SORT_DESC: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_DESC].pipeline; break;
                             default: GGML_ASSERT(false);
                         };
 
@@ -2292,6 +2241,106 @@ static bool ggml_metal_graph_compute(
 
                         [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
                     } break;
+                case GGML_OP_FLASH_ATTN_EXT:
+                    {
+                        GGML_ASSERT(ne00 % 4 == 0);
+                        GGML_ASSERT(src0->type == GGML_TYPE_F32);
+
+                        struct ggml_tensor * src3 = gf->nodes[i]->src[3];
+
+                        GGML_ASSERT(ggml_are_same_shape(src1, src2));
+                        GGML_ASSERT(src3);
+
+                        size_t offs_src3 = 0;
+
+                        id<MTLBuffer> id_src3 = src3 ? ggml_metal_get_buffer(src3, &offs_src3) : nil;
+
+                        GGML_ASSERT(!src3 || src3->type == GGML_TYPE_F16);
+                        GGML_ASSERT(!src3 || src3->ne[1] >= GGML_PAD(src0->ne[1], 8) &&
+                                "the Flash-Attention Metal kernel requires the mask to be padded to 8 and at least n_queries big");
+
+                        const int64_t  ne30 = src3 ? src3->ne[0] : 0; GGML_UNUSED(ne30);
+                        const int64_t  ne31 = src3 ? src3->ne[1] : 0;
+                        const int64_t  ne32 = src3 ? src3->ne[2] : 0; GGML_UNUSED(ne32);
+                        const int64_t  ne33 = src3 ? src3->ne[3] : 0; GGML_UNUSED(ne33);
+
+                        const uint64_t nb30 = src3 ? src3->nb[0] : 0; GGML_UNUSED(nb30);
+                        const uint64_t nb31 = src3 ? src3->nb[1] : 0;
+                        const uint64_t nb32 = src3 ? src3->nb[2] : 0; GGML_UNUSED(nb32);
+                        const uint64_t nb33 = src3 ? src3->nb[3] : 0; GGML_UNUSED(nb33);
+
+                        const enum ggml_type src2t = src2 ? src2->type : GGML_TYPE_COUNT; GGML_UNUSED(src2t);
+
+                        float scale;
+                        memcpy(&scale, dst->op_params, sizeof(float));
+
+                        id<MTLComputePipelineState> pipeline = nil;
+
+                        switch (ne00) {
+                            case 64:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H64 ].pipeline; break;
+                            case 80:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H80 ].pipeline; break;
+                            case 96:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H96 ].pipeline; break;
+                            case 112: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H112].pipeline; break;
+                            case 128: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H128].pipeline; break;
+                            case 256: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H256].pipeline; break;
+                            default:
+                                {
+                                    GGML_METAL_LOG_ERROR("unsupported size: %lld\n", ne00);
+                                    GGML_METAL_LOG_ERROR("add template specialization for this size\n");
+                                    GGML_ASSERT(false && "add template specialization for this size");
+                                }
+                        }
+
+                        // TODO: extend if necessary
+                        [encoder setComputePipelineState:pipeline];
+                        [encoder setBuffer:id_src0 offset:offs_src0        atIndex:0];
+                        [encoder setBuffer:id_src1 offset:offs_src1        atIndex:1];
+                        [encoder setBuffer:id_src2 offset:offs_src2        atIndex:2];
+                        [encoder setBuffer:id_src3 offset:offs_src3        atIndex:3];
+                        [encoder setBuffer:id_dst  offset:offs_dst         atIndex:4];
+                        [encoder setBytes:&ne00    length:sizeof( int64_t) atIndex:5];
+                        [encoder setBytes:&ne01    length:sizeof( int64_t) atIndex:6];
+                        [encoder setBytes:&ne02    length:sizeof( int64_t) atIndex:7];
+                        [encoder setBytes:&ne03    length:sizeof( int64_t) atIndex:8];
+                        [encoder setBytes:&nb00    length:sizeof(uint64_t) atIndex:9];
+                        [encoder setBytes:&nb01    length:sizeof(uint64_t) atIndex:10];
+                        [encoder setBytes:&nb02    length:sizeof(uint64_t) atIndex:11];
+                        [encoder setBytes:&nb03    length:sizeof(uint64_t) atIndex:12];
+                        [encoder setBytes:&ne10    length:sizeof( int64_t) atIndex:13];
+                        [encoder setBytes:&ne11    length:sizeof( int64_t) atIndex:14];
+                        [encoder setBytes:&ne12    length:sizeof( int64_t) atIndex:15];
+                        [encoder setBytes:&ne13    length:sizeof( int64_t) atIndex:16];
+                        [encoder setBytes:&nb10    length:sizeof(uint64_t) atIndex:17];
+                        [encoder setBytes:&nb11    length:sizeof(uint64_t) atIndex:18];
+                        [encoder setBytes:&nb12    length:sizeof(uint64_t) atIndex:19];
+                        [encoder setBytes:&nb13    length:sizeof(uint64_t) atIndex:20];
+                        [encoder setBytes:&ne31    length:sizeof( int64_t) atIndex:21];
+                        [encoder setBytes:&nb31    length:sizeof(uint64_t) atIndex:22];
+                        [encoder setBytes:&ne0     length:sizeof( int64_t) atIndex:23];
+                        [encoder setBytes:&ne1     length:sizeof( int64_t) atIndex:24];
+                        [encoder setBytes:&ne2     length:sizeof( int64_t) atIndex:25];
+                        [encoder setBytes:&ne3     length:sizeof( int64_t) atIndex:26];
+                        [encoder setBytes:&scale   length:sizeof(   float) atIndex:27];
+
+                        const int64_t nqptg = 8;  // queries per threadgroup    !! sync with kernel template arguments !!
+                        const int64_t ncpsg = 32; // cache values per simdgroup !! sync with kernel template arguments !!
+
+                        GGML_ASSERT(nqptg <= 32);
+                        GGML_ASSERT(nqptg  % 8  == 0);
+                        GGML_ASSERT(ncpsg  % 32 == 0);
+
+                        // simdgroups per threadgroup (a.k.a. warps)
+                        // for small batches use more simdgroups (needs more tests, to confirm if it's worth it)
+                        const int64_t nsg = ne01 <= nqptg ? MAX(4, MIN(ne11/ncpsg, (int64_t) pipeline.maxTotalThreadsPerThreadgroup/32)) : 4;
+
+                        const size_t smem = nqptg*(ne00 + nsg*(ncpsg + nqptg))*(sizeof(float)/2);
+
+                        //printf("smem: %zu, max: %zu\n", smem, ctx->device.maxThreadgroupMemoryLength);
+                        GGML_ASSERT(smem <= ctx->device.maxThreadgroupMemoryLength);
+                        [encoder setThreadgroupMemoryLength:smem atIndex:0];
+
+                        [encoder dispatchThreadgroups:MTLSizeMake((ne01 + nqptg - 1)/nqptg, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(32, nsg, 1)];
+                    } break;
                 case GGML_OP_DUP:
                 case GGML_OP_CPY:
                 case GGML_OP_CONT:
@@ -2494,10 +2543,13 @@ GGML_CALL static const char * ggml_backend_metal_buffer_type_get_name(ggml_backe
     UNUSED(buft);
 }
 
-static void ggml_backend_metal_log_allocated_size(id<MTLDevice> device) {
+static void ggml_backend_metal_log_allocated_size(id<MTLDevice> device, size_t size_aligned) {
+#ifndef GGML_METAL_NDEBUG
 #if TARGET_OS_OSX || (TARGET_OS_IOS && __clang_major__ >= 15)
     if (@available(macOS 10.12, iOS 16.0, *)) {
-        GGML_METAL_LOG_INFO(", (%8.2f / %8.2f)",
+        GGML_METAL_LOG_INFO("%s: allocated buffer, size = %8.2f MiB, (%8.2f / %8.2f)",
+                __func__,
+                size_aligned / 1024.0 / 1024.0,
                 device.currentAllocatedSize / 1024.0 / 1024.0,
                 device.recommendedMaxWorkingSetSize / 1024.0 / 1024.0);
 
@@ -2507,10 +2559,15 @@ static void ggml_backend_metal_log_allocated_size(id<MTLDevice> device) {
             GGML_METAL_LOG_INFO("\n");
         }
     } else {
-        GGML_METAL_LOG_INFO(", (%8.2f)\n", device.currentAllocatedSize / 1024.0 / 1024.0);
+        GGML_METAL_LOG_INFO("%s: allocated buffer, size = %8.2f MiB, (%8.2f)\n",
+                __func__,
+                size_aligned / 1024.0 / 1024.0,
+                device.currentAllocatedSize / 1024.0 / 1024.0);
     }
+#endif
 #endif
     UNUSED(device);
+    UNUSED(size_aligned);
 }
 
 GGML_CALL static ggml_backend_buffer_t ggml_backend_metal_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
@@ -2544,8 +2601,7 @@ GGML_CALL static ggml_backend_buffer_t ggml_backend_metal_buffer_type_alloc_buff
         return NULL;
     }
 
-    GGML_METAL_LOG_INFO("%s: allocated buffer, size = %8.2f MiB", __func__, size_aligned / 1024.0 / 1024.0);
-    ggml_backend_metal_log_allocated_size(device);
+    ggml_backend_metal_log_allocated_size(device, size_aligned);
 
     return ggml_backend_buffer_init(buft, ggml_backend_metal_buffer_i, ctx, size);
 }
@@ -2632,7 +2688,7 @@ GGML_CALL ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void * data,
             return false;
         }
 
-        GGML_METAL_LOG_INFO("%s: allocated buffer, size = %8.2f MiB", __func__, size_aligned / 1024.0 / 1024.0);
+        ggml_backend_metal_log_allocated_size(device, size_aligned);
 
         ++ctx->n_buffers;
     } else {
@@ -2655,7 +2711,8 @@ GGML_CALL ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void * data,
                 return false;
             }
 
-            GGML_METAL_LOG_INFO("%s: allocated buffer, size = %8.2f MiB, offs = %12ld", __func__, size_step_aligned / 1024.0 / 1024.0, i);
+            ggml_backend_metal_log_allocated_size(device, size_step_aligned);
+
             if (i + size_step < size) {
                 GGML_METAL_LOG_INFO("\n");
             }
@@ -2664,8 +2721,6 @@ GGML_CALL ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void * data,
         }
     }
 
-    ggml_backend_metal_log_allocated_size(device);
-
     return ggml_backend_buffer_init(ggml_backend_metal_buffer_type(), ggml_backend_metal_buffer_i, ctx, size);
 }
 

ggml-opencl.cpp

@@ -1354,7 +1354,7 @@ static void ggml_cl_pool_free(cl_mem mem, size_t size) {
 }
 
 void ggml_cl_free_data(const struct ggml_tensor* tensor) {
-    if (tensor->backend != GGML_BACKEND_TYPE_GPU) {
+    if (tensor->backend != GGML_BACKEND_GPU) {
         return;
     }
 
@@ -1412,7 +1412,7 @@ static cl_int ggml_cl_h2d_tensor_2d(cl_command_queue queue, cl_mem dst, size_t o
 }
 
 static void ggml_cl_mul_f32(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
-    GGML_ASSERT(src1->backend == GGML_BACKEND_TYPE_GPU);
+    GGML_ASSERT(src1->backend == GGML_BACKEND_GPU);
     const int64_t ne00 = src0->ne[0];
     const int64_t ne01 = src0->ne[1];
     const int64_t ne02 = src0->ne[2];
@@ -1476,7 +1476,7 @@ void ggml_cl_mul(const struct ggml_tensor * src0, const struct ggml_tensor * src
 }
 
 static void ggml_cl_add_f32(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
-    GGML_ASSERT(src1->backend == GGML_BACKEND_TYPE_GPU);
+    GGML_ASSERT(src1->backend == GGML_BACKEND_GPU);
     const int64_t ne00 = src0->ne[0];
     const int64_t ne01 = src0->ne[1];
     const int64_t ne02 = src0->ne[2];
@@ -1566,13 +1566,13 @@ static void ggml_cl_mul_mat_f32(const ggml_tensor * src0, const ggml_tensor * sr
     size_t y_size;
     size_t d_size;
     cl_mem d_X;
-    if (src0->backend == GGML_BACKEND_TYPE_GPU) { // NOLINT
+    if (src0->backend == GGML_BACKEND_GPU) { // NOLINT
         d_X = (cl_mem) src0->extra;
     } else {
         d_X = ggml_cl_pool_malloc(sizeof(float) * x_ne, &x_size);
     }
-    cl_mem d_Y = src1->backend == GGML_BACKEND_TYPE_GPU ? (cl_mem) src1->extra : ggml_cl_pool_malloc(sizeof(float) * y_ne, &y_size);
-    cl_mem d_D =  dst->backend == GGML_BACKEND_TYPE_GPU ? (cl_mem)  dst->extra : ggml_cl_pool_malloc(sizeof(float) * d_ne, &d_size);
+    cl_mem d_Y = src1->backend == GGML_BACKEND_GPU ? (cl_mem) src1->extra : ggml_cl_pool_malloc(sizeof(float) * y_ne, &y_size);
+    cl_mem d_D =  dst->backend == GGML_BACKEND_GPU ? (cl_mem)  dst->extra : ggml_cl_pool_malloc(sizeof(float) * d_ne, &d_size);
 
     size_t x_offset = 0;
 
@@ -1580,7 +1580,7 @@ static void ggml_cl_mul_mat_f32(const ggml_tensor * src0, const ggml_tensor * sr
         // TODO: copy src0 here when r3>1
         for (int64_t i13 = i03 * r3, e13 = i13 + r3; i13 < e13; i13++) {
             for (int64_t i02 = 0; i02 < ne02; i02++) {
-                if (src0->backend == GGML_BACKEND_TYPE_GPU) {
+                if (src0->backend == GGML_BACKEND_GPU) {
                     x_offset = (i03 * ne02 + i02) * x_ne;
                 } else {
                     // copy src0 to device
@@ -1589,7 +1589,7 @@ static void ggml_cl_mul_mat_f32(const ggml_tensor * src0, const ggml_tensor * sr
 
                 for (int64_t i12 = i02 * r2, e12 = i12 + r2; i12 < e12; i12++) {
                     // copy src1 to device
-                    if (src1->backend == GGML_BACKEND_TYPE_CPU) {
+                    if (src1->backend == GGML_BACKEND_CPU) {
                         CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_Y, 0, src1, i13, i12, NULL));
                     }
 
@@ -1612,7 +1612,7 @@ static void ggml_cl_mul_mat_f32(const ggml_tensor * src0, const ggml_tensor * sr
                     }
 
                     // copy dst to host
-                    if (dst->backend == GGML_BACKEND_TYPE_CPU) {
+                    if (dst->backend == GGML_BACKEND_CPU) {
                         float * d = (float *) ((char *) dst->data + i12*nb2 + i13*nb3);
                         CL_CHECK(clEnqueueReadBuffer(queue, d_D, true, 0, sizeof(float) * d_ne, d, 1, &ev_sgemm, NULL));
                     }
@@ -1621,13 +1621,13 @@ static void ggml_cl_mul_mat_f32(const ggml_tensor * src0, const ggml_tensor * sr
         }
     }
 
-    if (src0->backend != GGML_BACKEND_TYPE_GPU) {
+    if (src0->backend != GGML_BACKEND_GPU) {
         ggml_cl_pool_free(d_X, x_size);
     }
-    if (src1->backend != GGML_BACKEND_TYPE_GPU) {
+    if (src1->backend != GGML_BACKEND_GPU) {
         ggml_cl_pool_free(d_Y, y_size);
     }
-    if (dst->backend != GGML_BACKEND_TYPE_GPU) {
+    if (dst->backend != GGML_BACKEND_GPU) {
         ggml_cl_pool_free(d_D, d_size);
     }
 }
@@ -1670,7 +1670,7 @@ static void ggml_cl_mul_mat_f16(const ggml_tensor * src0, const ggml_tensor * sr
     size_t y_size;
     size_t d_size;
     cl_mem d_X;
-    if (src0->backend == GGML_BACKEND_TYPE_GPU) { // NOLINT
+    if (src0->backend == GGML_BACKEND_GPU) { // NOLINT
         d_X = (cl_mem) src0->extra;
     } else {
         d_X = ggml_cl_pool_malloc(sizeof(ggml_fp16_t) * x_ne, &x_size);
@@ -1687,7 +1687,7 @@ static void ggml_cl_mul_mat_f16(const ggml_tensor * src0, const ggml_tensor * sr
         // TODO: copy src0 here when r3>1
         for (int64_t i13 = i03 * r3, e13 = i13 + r3; i13 < e13; i13++) {
             for (int64_t i02 = 0; i02 < ne02; i02++) {
-                if (src0->backend == GGML_BACKEND_TYPE_GPU) {
+                if (src0->backend == GGML_BACKEND_GPU) {
                     x_offset = (i03 * ne02 + i02) * x_ne;
                 } else {
                     // copy src0 to device
@@ -1741,7 +1741,7 @@ static void ggml_cl_mul_mat_f16(const ggml_tensor * src0, const ggml_tensor * sr
                     }
 
                     // copy dst to host, then convert to float
-                    if (dst->backend == GGML_BACKEND_TYPE_CPU) {
+                    if (dst->backend == GGML_BACKEND_CPU) {
                         CL_CHECK(clEnqueueReadBuffer(queue, d_D, true, 0, sizeof(ggml_fp16_t) * d_ne, tmp, 1, &ev_sgemm, NULL));
                         float * d = (float *) ((char *) dst->data + i12*nb2 + i13*nb3);
                         ggml_fp16_to_fp32_row(tmp, d, d_ne);
@@ -1753,7 +1753,7 @@ static void ggml_cl_mul_mat_f16(const ggml_tensor * src0, const ggml_tensor * sr
         }
     }
 
-    if (src0->backend != GGML_BACKEND_TYPE_GPU) {
+    if (src0->backend != GGML_BACKEND_GPU) {
         ggml_cl_pool_free(d_X, x_size);
     }
     ggml_cl_pool_free(d_Y, y_size);
@@ -1798,7 +1798,7 @@ static void ggml_cl_mul_mat_q_f32(const ggml_tensor * src0, const ggml_tensor *
     cl_mem d_Y = ggml_cl_pool_malloc(sizeof(float) * y_ne, &y_size);
     cl_mem d_D = ggml_cl_pool_malloc(sizeof(float) * d_ne, &d_size);
     cl_mem d_Q;
-    if (src0->backend == GGML_BACKEND_TYPE_CPU) {
+    if (src0->backend == GGML_BACKEND_CPU) {
         d_Q = ggml_cl_pool_malloc(q_sz, &q_size);
     }
 
@@ -1817,10 +1817,10 @@ static void ggml_cl_mul_mat_q_f32(const ggml_tensor * src0, const ggml_tensor *
         for (int64_t i13 = i03 * r3, e13 = i13 + r3; i13 < e13; i13++) {
             for (int64_t i02 = 0; i02 < ne02; i02++) {
                 // copy src0 to device if necessary
-                if (src0->backend == GGML_BACKEND_TYPE_CPU) {
+                if (src0->backend == GGML_BACKEND_CPU) {
                     events.emplace_back();
                     CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_Q, 0, src0, i03, i02, events.data() + ev_idx++));
-                } else if (src0->backend == GGML_BACKEND_TYPE_GPU) {
+                } else if (src0->backend == GGML_BACKEND_GPU) {
                     d_Q = (cl_mem) src0->extra;
                 } else {
                     GGML_ASSERT(false);
@@ -1829,7 +1829,7 @@ static void ggml_cl_mul_mat_q_f32(const ggml_tensor * src0, const ggml_tensor *
                 if (!mul_mat_vec) {
                     // convert src0 to fp32 on device
                     const size_t global = x_ne / global_denom;
-                    const size_t offset = src0->backend == GGML_BACKEND_TYPE_GPU ? (i03 * ne02 + i02) * x_bps : 0;
+                    const size_t offset = src0->backend == GGML_BACKEND_GPU ? (i03 * ne02 + i02) * x_bps : 0;
                     CL_CHECK(clSetKernelArg(*to_fp32_cl, 0, sizeof(cl_mem), &d_Q));
                     CL_CHECK(clSetKernelArg(*to_fp32_cl, 1, sizeof(cl_mem), &d_X));
                     CL_CHECK(clEnqueueNDRangeKernel(queue, *to_fp32_cl, 1, &offset, &global, local > 0 ? &local : NULL, events.size(), !events.empty() ? events.data() : NULL, NULL));
@@ -1843,7 +1843,7 @@ static void ggml_cl_mul_mat_q_f32(const ggml_tensor * src0, const ggml_tensor *
 
                         // compute
                         const size_t global = ne01 * local;
-                        const size_t offset = src0->backend == GGML_BACKEND_TYPE_GPU ? (i03 * ne02 + i02) * x_bps : 0;
+                        const size_t offset = src0->backend == GGML_BACKEND_GPU ? (i03 * ne02 + i02) * x_bps : 0;
                         const cl_int ncols = ne00;
                         events.emplace_back();
                         CL_CHECK(clSetKernelArg(*dmmv, 0, sizeof(cl_mem), &d_Q));
@@ -1895,7 +1895,7 @@ static void ggml_cl_mul_mat_q_f32(const ggml_tensor * src0, const ggml_tensor *
     }
     ggml_cl_pool_free(d_Y, y_size);
     ggml_cl_pool_free(d_D, d_size);
-    if (src0->backend == GGML_BACKEND_TYPE_CPU) {
+    if (src0->backend == GGML_BACKEND_CPU) {
         ggml_cl_pool_free(d_Q, q_size);
     }
 }
@@ -1911,7 +1911,7 @@ bool ggml_cl_can_mul_mat(const struct ggml_tensor * src0, const struct ggml_tens
     if ((src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 || ggml_is_quantized(src0->type)) &&
         src1->type == GGML_TYPE_F32 &&
         dst->type == GGML_TYPE_F32 &&
-        ((ne0 >= 32 && ne1 >= 32 && ne10 >= 32) || src0->backend == GGML_BACKEND_TYPE_GPU)) {
+        ((ne0 >= 32 && ne1 >= 32 && ne10 >= 32) || src0->backend == GGML_BACKEND_GPU)) {
         return true;
     }
 
@@ -1993,7 +1993,7 @@ void ggml_cl_transform_tensor(void * data, ggml_tensor * tensor) {
     CL_CHECK(clFinish(queue));
 
     tensor->extra = dst;
-    GGML_ASSERT(tensor->backend == GGML_BACKEND_TYPE_GPU);
+    GGML_ASSERT(tensor->backend == GGML_BACKEND_GPU);
 }
 
 // ggml-backend
@@ -2045,7 +2045,7 @@ static void ggml_backend_opencl_buffer_init_tensor(ggml_backend_buffer_t buffer,
         ctx->sub_buffers.push_back(sub_buffer);
         tensor->extra = sub_buffer;
     }
-    tensor->backend = GGML_BACKEND_TYPE_GPU;
+    tensor->backend = GGML_BACKEND_GPU;
 }
 
 static void ggml_backend_opencl_buffer_set_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor, const void * data, size_t offset, size_t size) {

ggml-quants.h

@@ -191,21 +191,6 @@ typedef struct {
 } block_iq3_xxs;
 static_assert(sizeof(block_iq3_xxs) == sizeof(ggml_fp16_t) + 3*(QK_K/8), "wrong iq3_xxs block size/padding");
 
-// 3.4375 bpw
-#if QK_K == 64
-#define IQ3S_N_SCALE 2
-#else
-#define IQ3S_N_SCALE QK_K/64
-#endif
-typedef struct {
-    ggml_fp16_t d;
-    uint8_t qs[QK_K/4];
-    uint8_t qh[QK_K/32];
-    uint8_t signs[QK_K/8];
-    uint8_t scales[IQ3S_N_SCALE];
-} block_iq3_s;
-static_assert(sizeof(block_iq3_s) == sizeof(ggml_fp16_t) + 13*(QK_K/32) + IQ3S_N_SCALE, "wrong iq3_s block size/padding");
-
 typedef struct {
     ggml_fp16_t d;
     uint8_t qs[QK_K/8];
@@ -213,14 +198,6 @@ typedef struct {
 } block_iq1_s;
 static_assert(sizeof(block_iq1_s) == sizeof(ggml_fp16_t) + QK_K/8 + QK_K/16, "wrong iq1_s block size/padding");
 
-// Non-linear quants
-#define QK4_NL 32
-typedef struct {
-    ggml_fp16_t d;
-    uint8_t qs[QK4_NL/2];
-} block_iq4_nl;
-static_assert(sizeof(block_iq4_nl) == sizeof(ggml_fp16_t) + QK4_NL/2, "wrong iq4_nl block size/padding");
-
 #ifdef __cplusplus
 extern "C" {
 #endif
@@ -240,8 +217,6 @@ void quantize_row_q5_K_reference(const float * GGML_RESTRICT x, block_q5_K * GGM
 void quantize_row_q6_K_reference(const float * GGML_RESTRICT x, block_q6_K * GGML_RESTRICT y, int k);
 void quantize_row_q8_K_reference(const float * GGML_RESTRICT x, block_q8_K * GGML_RESTRICT y, int k);
 void quantize_row_iq3_xxs_reference(const float * GGML_RESTRICT x, block_iq3_xxs * GGML_RESTRICT y, int k);
-void quantize_row_iq4_nl_reference (const float * GGML_RESTRICT x, block_iq4_nl  * GGML_RESTRICT y, int k);
-void quantize_row_iq3_s_reference  (const float * GGML_RESTRICT x, block_iq3_s   * GGML_RESTRICT y, int k);
 
 void quantize_row_q4_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int k);
 void quantize_row_q4_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int k);
@@ -257,8 +232,6 @@ void quantize_row_q5_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, in
 void quantize_row_q6_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int k);
 void quantize_row_q8_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int k);
 void quantize_row_iq3_xxs(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int k);
-void quantize_row_iq4_nl (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int k);
-void quantize_row_iq3_s  (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int k);
 
 // Dequantization
 void dequantize_row_q4_0(const block_q4_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int k);
@@ -278,8 +251,6 @@ void dequantize_row_iq2_xxs(const block_iq2_xxs * GGML_RESTRICT x, float * GGML_
 void dequantize_row_iq2_xs (const block_iq2_xs  * GGML_RESTRICT x, float * GGML_RESTRICT y, int k);
 void dequantize_row_iq3_xxs(const block_iq3_xxs * GGML_RESTRICT x, float * GGML_RESTRICT y, int k);
 void dequantize_row_iq1_s  (const block_iq1_s   * GGML_RESTRICT x, float * GGML_RESTRICT y, int k);
-void dequantize_row_iq4_nl (const block_iq4_nl  * GGML_RESTRICT x, float * GGML_RESTRICT y, int k);
-void dequantize_row_iq3_s  (const block_iq3_s   * GGML_RESTRICT x, float * GGML_RESTRICT y, int k);
 
 // Dot product
 void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
@@ -297,8 +268,6 @@ void ggml_vec_dot_iq2_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const
 void ggml_vec_dot_iq2_xs_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
 void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
 void ggml_vec_dot_iq1_s_q8_K  (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_iq4_nl_q8_0 (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_iq3_s_q8_K  (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
 
 //
 // Quantization utilizing an importance matrix (a.k.a. "Activation aWare Quantization")
@@ -307,8 +276,6 @@ size_t quantize_iq2_xxs(const float * src, void * dst, int nrows, int n_per_row,
 size_t quantize_iq2_xs (const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix);
 size_t quantize_iq3_xxs(const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix);
 size_t quantize_iq1_s  (const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix);
-size_t quantize_iq4_nl (const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix);
-size_t quantize_iq3_s  (const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix);
 size_t quantize_q2_K   (const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix);
 size_t quantize_q3_K   (const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix);
 size_t quantize_q4_K   (const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix);

ggml-sycl.cpp

@@ -3338,7 +3338,7 @@ void print_ggml_tensor(const char*name, struct ggml_tensor *src){
 
     size_t total_elements = ggml_nelements(src);
 
-    const bool src_on_device = src->backend == GGML_BACKEND_TYPE_GPU || src->backend == GGML_BACKEND_TYPE_GPU_SPLIT;
+    const bool src_on_device = src->backend == GGML_BACKEND_GPU || src->backend == GGML_BACKEND_GPU_SPLIT;
     float *src_data =NULL;
     if(src_on_device) {
         ggml_tensor_extra_gpu * src_extra = (ggml_tensor_extra_gpu *)  src->extra;
@@ -8086,11 +8086,11 @@ static void k_argsort_f32_i32(const float * x, int * dst, const int ncols,
             int ixj = col ^ j;
             if (ixj > col) {
                 if ((col & k) == 0) {
-                    if (order == GGML_SORT_ORDER_ASC ? x_row[dst_row[col]] > x_row[dst_row[ixj]] : x_row[dst_row[col]] < x_row[dst_row[ixj]]) {
+                    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_ORDER_ASC ? x_row[dst_row[col]] < x_row[dst_row[ixj]] : x_row[dst_row[col]] > x_row[dst_row[ixj]]) {
+                    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]);
                     }
                 }
@@ -9188,22 +9188,174 @@ static void convert_mul_mat_vec_f16_sycl(const void *vx, const dfloat *y,
     }
 }
 
-template <int qk, int qi, typename block_q_t, int vdr,
-          vec_dot_q_sycl_t vec_dot_q_sycl>
-static void mul_mat_vec_q_sycl_submitter(const void *vx, const void *vy,
-                                         float *dst, const int ncols,
-                                         const int nrows,
-                                         dpct::queue_ptr stream) {
-  GGML_ASSERT(ncols % QK4_0 == 0);
-  const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y;
-  const sycl::range<3> block_nums(1, 1, block_num_y);
-  const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
-  stream->parallel_for(
-      sycl::nd_range<3>(block_nums * block_dims, block_dims), [=
-  ](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] {
-        mul_mat_vec_q<qk, qi, block_q_t, vdr, vec_dot_q_sycl>(
-            vx, vy, dst, ncols, nrows, item_ct1);
-      });
+static void mul_mat_vec_q4_0_q8_1_sycl(const void *vx, const void *vy,
+                                       float *dst, const int ncols,
+                                       const int nrows,
+                                       dpct::queue_ptr stream) {
+    GGML_ASSERT(ncols % QK4_0 == 0);
+    const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y;
+    const sycl::range<3> block_nums(1, 1, block_num_y);
+    const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
+    stream->parallel_for(
+        sycl::nd_range<3>(block_nums * block_dims, block_dims),
+        [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] {
+            mul_mat_vec_q<QK4_0, QI4_0, block_q4_0, VDR_Q4_0_Q8_1_MMVQ,
+                          vec_dot_q4_0_q8_1>(vx, vy, dst, ncols, nrows,
+                                             item_ct1);
+        });
+}
+
+static void mul_mat_vec_q4_1_q8_1_sycl(const void *vx, const void *vy,
+                                       float *dst, const int ncols,
+                                       const int nrows,
+                                       dpct::queue_ptr stream) {
+    GGML_ASSERT(ncols % QK4_1 == 0);
+    const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y;
+    const sycl::range<3> block_nums(1, 1, block_num_y);
+    const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
+    stream->parallel_for(
+        sycl::nd_range<3>(block_nums * block_dims, block_dims),
+        [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] {
+            mul_mat_vec_q<QK4_0, QI4_1, block_q4_1, VDR_Q4_1_Q8_1_MMVQ,
+                          vec_dot_q4_1_q8_1>(vx, vy, dst, ncols, nrows,
+                                             item_ct1);
+        });
+}
+
+static void mul_mat_vec_q5_0_q8_1_sycl(const void *vx, const void *vy,
+                                       float *dst, const int ncols,
+                                       const int nrows,
+                                       dpct::queue_ptr stream) {
+    GGML_ASSERT(ncols % QK5_0 == 0);
+    const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y;
+    const sycl::range<3> block_nums(1, 1, block_num_y);
+    const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
+    stream->parallel_for(
+        sycl::nd_range<3>(block_nums * block_dims, block_dims),
+        [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] {
+            mul_mat_vec_q<QK5_0, QI5_0, block_q5_0, VDR_Q5_0_Q8_1_MMVQ,
+                          vec_dot_q5_0_q8_1>(vx, vy, dst, ncols, nrows,
+                                             item_ct1);
+        });
+}
+
+static void mul_mat_vec_q5_1_q8_1_sycl(const void *vx, const void *vy,
+                                       float *dst, const int ncols,
+                                       const int nrows,
+                                       dpct::queue_ptr stream) {
+    GGML_ASSERT(ncols % QK5_1 == 0);
+    const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y;
+    const sycl::range<3> block_nums(1, 1, block_num_y);
+    const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
+    stream->parallel_for(
+        sycl::nd_range<3>(block_nums * block_dims, block_dims),
+        [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] {
+            mul_mat_vec_q<QK5_1, QI5_1, block_q5_1, VDR_Q5_1_Q8_1_MMVQ,
+                          vec_dot_q5_1_q8_1>(vx, vy, dst, ncols, nrows,
+                                             item_ct1);
+        });
+}
+
+static void mul_mat_vec_q8_0_q8_1_sycl(const void *vx, const void *vy,
+                                       float *dst, const int ncols,
+                                       const int nrows,
+                                       dpct::queue_ptr stream) {
+    GGML_ASSERT(ncols % QK8_0 == 0);
+    const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y;
+    const sycl::range<3> block_nums(1, 1, block_num_y);
+    const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
+    stream->parallel_for(
+        sycl::nd_range<3>(block_nums * block_dims, block_dims),
+        [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] {
+            mul_mat_vec_q<QK8_0, QI8_0, block_q8_0, VDR_Q8_0_Q8_1_MMVQ,
+                          vec_dot_q8_0_q8_1>(vx, vy, dst, ncols, nrows,
+                                             item_ct1);
+        });
+}
+
+static void mul_mat_vec_q2_K_q8_1_sycl(const void *vx, const void *vy,
+                                       float *dst, const int ncols,
+                                       const int nrows,
+                                       dpct::queue_ptr stream) {
+    GGML_ASSERT(ncols % QK_K == 0);
+    const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y;
+    const sycl::range<3> block_nums(1, 1, block_num_y);
+    const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
+    stream->parallel_for(
+        sycl::nd_range<3>(block_nums * block_dims, block_dims),
+        [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] {
+            mul_mat_vec_q<QK_K, QI2_K, block_q2_K, VDR_Q2_K_Q8_1_MMVQ,
+                          vec_dot_q2_K_q8_1>(vx, vy, dst, ncols, nrows,
+                                             item_ct1);
+        });
+}
+
+static void mul_mat_vec_q3_K_q8_1_sycl(const void *vx, const void *vy,
+                                       float *dst, const int ncols,
+                                       const int nrows,
+                                       dpct::queue_ptr stream) {
+    GGML_ASSERT(ncols % QK_K == 0);
+    const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y;
+    const sycl::range<3> block_nums(1, 1, block_num_y);
+    const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
+    stream->parallel_for(
+        sycl::nd_range<3>(block_nums * block_dims, block_dims),
+        [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] {
+            mul_mat_vec_q<QK_K, QI3_K, block_q3_K, VDR_Q3_K_Q8_1_MMVQ,
+                          vec_dot_q3_K_q8_1>(vx, vy, dst, ncols, nrows,
+                                             item_ct1);
+        });
+}
+
+static void mul_mat_vec_q4_K_q8_1_sycl(const void *vx, const void *vy,
+                                       float *dst, const int ncols,
+                                       const int nrows,
+                                       dpct::queue_ptr stream) {
+    GGML_ASSERT(ncols % QK_K == 0);
+    const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y;
+    const sycl::range<3> block_nums(1, 1, block_num_y);
+    const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
+    stream->parallel_for(
+        sycl::nd_range<3>(block_nums * block_dims, block_dims),
+        [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] {
+            mul_mat_vec_q<QK_K, QI4_K, block_q4_K, VDR_Q4_K_Q8_1_MMVQ,
+                          vec_dot_q4_K_q8_1>(vx, vy, dst, ncols, nrows,
+                                             item_ct1);
+        });
+}
+
+static void mul_mat_vec_q5_K_q8_1_sycl(const void *vx, const void *vy,
+                                       float *dst, const int ncols,
+                                       const int nrows,
+                                       dpct::queue_ptr stream) {
+    GGML_ASSERT(ncols % QK_K == 0);
+    const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y;
+    const sycl::range<3> block_nums(1, 1, block_num_y);
+    const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
+    stream->parallel_for(
+        sycl::nd_range<3>(block_nums * block_dims, block_dims),
+        [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] {
+            mul_mat_vec_q<QK_K, QI5_K, block_q5_K, VDR_Q5_K_Q8_1_MMVQ,
+                          vec_dot_q5_K_q8_1>(vx, vy, dst, ncols, nrows,
+                                             item_ct1);
+        });
+}
+
+static void mul_mat_vec_q6_K_q8_1_sycl(const void *vx, const void *vy,
+                                       float *dst, const int ncols,
+                                       const int nrows,
+                                       dpct::queue_ptr stream) {
+    GGML_ASSERT(ncols % QK_K == 0);
+    const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y;
+    const sycl::range<3> block_nums(1, 1, block_num_y);
+    const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
+    stream->parallel_for(
+        sycl::nd_range<3>(block_nums * block_dims, block_dims),
+        [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] {
+            mul_mat_vec_q<QK_K, QI6_K, block_q6_K, VDR_Q6_K_Q8_1_MMVQ,
+                          vec_dot_q6_K_q8_1>(vx, vy, dst, ncols, nrows,
+                                             item_ct1);
+        });
 }
 
 int get_device_index_by_id(int id){
@@ -10825,7 +10977,7 @@ static void argsort_f32_i32_sycl(const float *x, int *dst, const int ncols,
 
     const sycl::range<3> block_dims(1, 1, ncols);
     const sycl::range<3> block_nums(1, nrows, 1);
-    if (order == GGML_SORT_ORDER_ASC) {
+    if (order == GGML_SORT_ASC) {
         /*
         DPCT1049:44: The work-group size passed to the SYCL kernel may exceed
         the limit. To get the device limit, query
@@ -10834,9 +10986,9 @@ static void argsort_f32_i32_sycl(const float *x, int *dst, const int ncols,
         stream->parallel_for(
             sycl::nd_range<3>(block_nums * block_dims, block_dims),
             [=](sycl::nd_item<3> item_ct1) {
-                k_argsort_f32_i32<GGML_SORT_ORDER_ASC>(x, dst, ncols, item_ct1);
+                k_argsort_f32_i32<GGML_SORT_ASC>(x, dst, ncols, item_ct1);
             });
-    } else if (order == GGML_SORT_ORDER_DESC) {
+    } else if (order == GGML_SORT_DESC) {
         /*
         DPCT1049:45: The work-group size passed to the SYCL kernel may exceed
         the limit. To get the device limit, query
@@ -10845,7 +10997,7 @@ static void argsort_f32_i32_sycl(const float *x, int *dst, const int ncols,
         stream->parallel_for(
             sycl::nd_range<3>(block_nums * block_dims, block_dims),
             [=](sycl::nd_item<3> item_ct1) {
-                k_argsort_f32_i32<GGML_SORT_ORDER_DESC>(x, dst, ncols, item_ct1);
+                k_argsort_f32_i32<GGML_SORT_DESC>(x, dst, ncols, item_ct1);
             });
     } else {
         GGML_ASSERT(false);
@@ -11407,12 +11559,12 @@ static dpct::err0 ggml_sycl_cpy_tensor_2d(void *dst,
 
     dpct::memcpy_direction kind;
     char * src_ptr;
-    if (src->backend == GGML_BACKEND_TYPE_CPU) {
+    if (src->backend == GGML_BACKEND_CPU) {
         kind = dpct::host_to_device;
         src_ptr = (char *) src->data;
-        // GGML_SYCL_DEBUG("ggml_sycl_cpy_tensor_2d  GGML_BACKEND_TYPE_CPU src_ptr %p\n", src_ptr);
-    } else if (src->backend == GGML_BACKEND_TYPE_GPU || src->backend == GGML_BACKEND_TYPE_GPU_SPLIT) {
-        GGML_ASSERT(src->backend != GGML_BACKEND_TYPE_GPU_SPLIT || (i1_low == 0 && i1_high == src->ne[1]));
+        // GGML_SYCL_DEBUG("ggml_sycl_cpy_tensor_2d  GGML_BACKEND_CPU src_ptr %p\n", src_ptr);
+    } else if (src->backend == GGML_BACKEND_GPU || src->backend == GGML_BACKEND_GPU_SPLIT) {
+        GGML_ASSERT(src->backend != GGML_BACKEND_GPU_SPLIT || (i1_low == 0 && i1_high == src->ne[1]));
         kind = dpct::device_to_device;
         ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) src->extra;
         int id;
@@ -11846,7 +11998,7 @@ inline void ggml_sycl_op_mul_mat_q(
 
     // the main device has a larger memory buffer to hold the results from all GPUs
     // nrows_dst == nrows of the matrix that the dequantize_mul_mat kernel writes into
-    const int64_t nrows_dst = dst->backend == GGML_BACKEND_TYPE_GPU && device_id == g_main_device ? ne0 : row_diff;
+    const int64_t nrows_dst = dst->backend == GGML_BACKEND_GPU && device_id == g_main_device ? ne0 : row_diff;
 
     switch (src0->type) {
         case GGML_TYPE_Q4_0:
@@ -11943,63 +12095,37 @@ inline void ggml_sycl_op_mul_mat_vec_q(
     const int64_t ne00 = src0->ne[0];
     const int64_t row_diff = row_high - row_low;
 
-    // TODO: support these quantization types
-    GGML_ASSERT(!(src0->type == GGML_TYPE_IQ2_XXS ||
-                  src0->type == GGML_TYPE_IQ2_XS ||
-                  src0->type == GGML_TYPE_IQ3_XXS ||
-                  src0->type == GGML_TYPE_IQ1_S));
-
     switch (src0->type) {
         case GGML_TYPE_Q4_0:
-          mul_mat_vec_q_sycl_submitter<QK4_0, QI4_0, block_q4_0,
-                                       VDR_Q4_0_Q8_1_MMVQ, vec_dot_q4_0_q8_1>(
-              src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream);
-          break;
+            mul_mat_vec_q4_0_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream);
+            break;
         case GGML_TYPE_Q4_1:
-          mul_mat_vec_q_sycl_submitter<QK4_1, QI4_1, block_q4_1,
-                                       VDR_Q4_1_Q8_1_MMVQ, vec_dot_q4_1_q8_1>(
-              src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream);
-          break;
+            mul_mat_vec_q4_1_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream);
+            break;
         case GGML_TYPE_Q5_0:
-          mul_mat_vec_q_sycl_submitter<QK5_0, QI5_0, block_q5_0,
-                                       VDR_Q5_0_Q8_1_MMVQ, vec_dot_q5_0_q8_1>(
-              src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream);
-          break;
+            mul_mat_vec_q5_0_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream);
+            break;
         case GGML_TYPE_Q5_1:
-          mul_mat_vec_q_sycl_submitter<QK5_1, QI5_1, block_q5_1,
-                                       VDR_Q5_1_Q8_1_MMVQ, vec_dot_q5_1_q8_1>(
-              src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream);
-          break;
+            mul_mat_vec_q5_1_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream);
+            break;
         case GGML_TYPE_Q8_0:
-          mul_mat_vec_q_sycl_submitter<QK8_0, QI8_0, block_q8_0,
-                                       VDR_Q8_0_Q8_1_MMVQ, vec_dot_q8_0_q8_1>(
-              src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream);
-          break;
+            mul_mat_vec_q8_0_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream);
+            break;
         case GGML_TYPE_Q2_K:
-          mul_mat_vec_q_sycl_submitter<QK_K, QI2_K, block_q2_K,
-                                       VDR_Q2_K_Q8_1_MMVQ, vec_dot_q2_K_q8_1>(
-              src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream);
-          break;
+            mul_mat_vec_q2_K_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream);
+            break;
         case GGML_TYPE_Q3_K:
-          mul_mat_vec_q_sycl_submitter<QK_K, QI3_K, block_q3_K,
-                                       VDR_Q3_K_Q8_1_MMVQ, vec_dot_q3_K_q8_1>(
-              src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream);
-          break;
+            mul_mat_vec_q3_K_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream);
+            break;
         case GGML_TYPE_Q4_K:
-          mul_mat_vec_q_sycl_submitter<QK_K, QI4_K, block_q4_K,
-                                       VDR_Q4_K_Q8_1_MMVQ, vec_dot_q4_K_q8_1>(
-              src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream);
-          break;
+            mul_mat_vec_q4_K_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream);
+            break;
         case GGML_TYPE_Q5_K:
-          mul_mat_vec_q_sycl_submitter<QK_K, QI5_K, block_q5_K,
-                                       VDR_Q5_K_Q8_1_MMVQ, vec_dot_q5_K_q8_1>(
-              src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream);
-          break;
+            mul_mat_vec_q5_K_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream);
+            break;
         case GGML_TYPE_Q6_K:
-          mul_mat_vec_q_sycl_submitter<QK_K, QI6_K, block_q6_K,
-                                       VDR_Q6_K_Q8_1_MMVQ, vec_dot_q6_K_q8_1>(
-              src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream);
-          break;
+            mul_mat_vec_q6_K_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream);
+            break;
         default:
             GGML_ASSERT(false);
             break;
@@ -12019,7 +12145,7 @@ inline void ggml_sycl_op_dequantize_mul_mat_vec(
     const int64_t src1_ncols, const int64_t src1_padded_row_size,
     const dpct::queue_ptr &stream) {
 
-    GGML_TENSOR_BINARY_OP_LOCALS;
+    GGML_TENSOR_BINARY_OP_LOCALS
 
     const int64_t row_diff = row_high - row_low;
 
@@ -12119,7 +12245,7 @@ inline void ggml_sycl_op_mul_mat_sycl(
 
     // the main device has a larger memory buffer to hold the results from all GPUs
     // ldc == nrows of the matrix that cuBLAS writes into
-    int ldc = dst->backend == GGML_BACKEND_TYPE_GPU && device_id == g_main_device ? ne0 : row_diff;
+    int ldc = dst->backend == GGML_BACKEND_GPU && device_id == g_main_device ? ne0 : row_diff;
 
 #ifdef GGML_SYCL_F16
     bool use_fp16 = true;  // TODO(Yu) SYCL capability check
@@ -12501,16 +12627,16 @@ static void ggml_sycl_op_flatten(const ggml_tensor *src0,
     const bool use_src1 = src1 != nullptr;
     const int64_t nrows1 = use_src1 ? ggml_nrows(src1) : 1;
 
-    GGML_ASSERT(!use_src1 || src1->backend != GGML_BACKEND_TYPE_GPU_SPLIT);
-    GGML_ASSERT(              dst->backend != GGML_BACKEND_TYPE_GPU_SPLIT);
+    GGML_ASSERT(!use_src1 || src1->backend != GGML_BACKEND_GPU_SPLIT);
+    GGML_ASSERT(              dst->backend != GGML_BACKEND_GPU_SPLIT);
 
     ggml_tensor_extra_gpu * src0_extra =            (ggml_tensor_extra_gpu *) src0->extra;
     ggml_tensor_extra_gpu * src1_extra = use_src1 ? (ggml_tensor_extra_gpu *) src1->extra : nullptr;
     ggml_tensor_extra_gpu * dst_extra  =            (ggml_tensor_extra_gpu *)  dst->extra;
 
-    const bool src0_on_device =             src0->backend == GGML_BACKEND_TYPE_GPU || src0->backend == GGML_BACKEND_TYPE_GPU_SPLIT;
-    const bool src1_on_device = use_src1 && src1->backend == GGML_BACKEND_TYPE_GPU;
-    const bool  dst_on_device =              dst->backend == GGML_BACKEND_TYPE_GPU;
+    const bool src0_on_device =             src0->backend == GGML_BACKEND_GPU || src0->backend == GGML_BACKEND_GPU_SPLIT;
+    const bool src1_on_device = use_src1 && src1->backend == GGML_BACKEND_GPU;
+    const bool  dst_on_device =              dst->backend == GGML_BACKEND_GPU;
 
     // dd = data device
     float * src0_ddf = nullptr;
@@ -12565,7 +12691,7 @@ static void ggml_sycl_op_flatten(const ggml_tensor *src0,
             main_stream->memcpy(dst->data, dst_ddf, ggml_nbytes(dst))));
     }
 
-    if (dst->backend == GGML_BACKEND_TYPE_CPU) {
+    if (dst->backend == GGML_BACKEND_CPU) {
         SYCL_CHECK(CHECK_TRY_ERROR(
             dpct::get_current_device().queues_wait_and_throw()));
     }
@@ -12640,8 +12766,8 @@ static void ggml_sycl_op_mul_mat(const ggml_tensor *src0,
     const int nb2 = dst->nb[2];
     const int nb3 = dst->nb[3];
 
-    GGML_ASSERT(dst->backend != GGML_BACKEND_TYPE_GPU_SPLIT);
-    GGML_ASSERT(src1->backend != GGML_BACKEND_TYPE_GPU_SPLIT);
+    GGML_ASSERT(dst->backend != GGML_BACKEND_GPU_SPLIT);
+    GGML_ASSERT(src1->backend != GGML_BACKEND_GPU_SPLIT);
 
     GGML_ASSERT(ne12 >= ne02 && ne12 % ne02 == 0);
 
@@ -12656,13 +12782,13 @@ static void ggml_sycl_op_mul_mat(const ggml_tensor *src0,
     ggml_tensor_extra_gpu * src1_extra = (ggml_tensor_extra_gpu *) src1->extra;
     ggml_tensor_extra_gpu *  dst_extra = (ggml_tensor_extra_gpu *)  dst->extra;
 
-    const bool src0_on_device = src0->backend == GGML_BACKEND_TYPE_GPU || src0->backend == GGML_BACKEND_TYPE_GPU_SPLIT;
+    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);
 
     int64_t src1_padded_col_size = GGML_PAD(ne10, MATRIX_ROW_PADDING);
 
-    const bool split = src0->backend == GGML_BACKEND_TYPE_GPU_SPLIT;
+    const bool split = src0->backend == GGML_BACKEND_GPU_SPLIT;
     GGML_ASSERT(!(split && ne02 > 1));
     GGML_ASSERT(!(split && ne03 > 1));
     GGML_ASSERT(!(split && ne02 < ne12));
@@ -12717,8 +12843,8 @@ static void ggml_sycl_op_mul_mat(const ggml_tensor *src0,
 
         used_devices++;
 
-        const bool src1_on_device = src1->backend == GGML_BACKEND_TYPE_GPU && id == g_main_device_index;
-        const bool  dst_on_device =  dst->backend == GGML_BACKEND_TYPE_GPU && id == g_main_device_index;
+        const bool src1_on_device = src1->backend == GGML_BACKEND_GPU && id == g_main_device_index;
+        const bool  dst_on_device =  dst->backend == GGML_BACKEND_GPU && id == g_main_device_index;
 
         ggml_sycl_set_device(get_device_id_by_index(id));
         const dpct::queue_ptr stream = g_syclStreams[id][0];
@@ -12782,8 +12908,8 @@ static void ggml_sycl_op_mul_mat(const ggml_tensor *src0,
                 continue;
             }
 
-            const bool src1_on_device = src1->backend == GGML_BACKEND_TYPE_GPU && id == g_main_device_index;
-            const bool  dst_on_device =  dst->backend == GGML_BACKEND_TYPE_GPU && id == g_main_device_index;
+            const bool src1_on_device = src1->backend == GGML_BACKEND_GPU && id == g_main_device_index;
+            const bool  dst_on_device =  dst->backend == GGML_BACKEND_GPU && id == g_main_device_index;
             const int64_t row_diff = row_high[id] - row_low[id];
 
             ggml_sycl_set_device(get_device_id_by_index(id));
@@ -12809,12 +12935,12 @@ static void ggml_sycl_op_mul_mat(const ggml_tensor *src0,
 
                 // the main device memory buffer can be on VRAM scratch, with space for all partial results
                 // in that case an offset on dst_ddf_i is needed
-                if (dst->backend == GGML_BACKEND_TYPE_GPU && id == g_main_device_index) {
+                if (dst->backend == GGML_BACKEND_GPU && id == g_main_device_index) {
                     dst_dd_i += row_low[id]; // offset is 0 if no tensor split
                 }
 
                 // copy src0, src1 to device if necessary
-                if (src1->backend == GGML_BACKEND_TYPE_GPU && src1_is_contiguous) {
+                if (src1->backend == GGML_BACKEND_GPU && src1_is_contiguous) {
                     if (id != g_main_device_index) {
                         if (convert_src1_to_q8_1) {
                             char * src1_ddq_i_source = src1_ddq[g_main_device_index] + src1_ddq_i_offset;
@@ -12830,14 +12956,14 @@ static void ggml_sycl_op_mul_mat(const ggml_tensor *src0,
                                 src1_ncols * ne10 * sizeof(float))));
                         }
                     }
-                } else if (src1->backend == GGML_BACKEND_TYPE_CPU || (src1_on_device && !src1_is_contiguous)) {
+                } else if (src1->backend == GGML_BACKEND_CPU || (src1_on_device && !src1_is_contiguous)) {
                     SYCL_CHECK(ggml_sycl_cpy_tensor_2d(
                                    src1_ddf_i, src1, i03, i02, src1_col_0, src1_col_0+src1_ncols, stream));
                 } else {
                     GGML_ASSERT(false);
                 }
 
-                if (convert_src1_to_q8_1 && (src1->backend == GGML_BACKEND_TYPE_CPU || !src1_is_contiguous)) {
+                if (convert_src1_to_q8_1 && (src1->backend == GGML_BACKEND_CPU || !src1_is_contiguous)) {
                     quantize_row_q8_1_sycl(src1_ddf_i, src1_ddq_i, ne10, src1_ncols, src1_padded_col_size, stream);
                     /*
                     DPCT1010:92: SYCL uses exceptions to report errors and does
@@ -12867,10 +12993,10 @@ static void ggml_sycl_op_mul_mat(const ggml_tensor *src0,
                 if (!dst_on_device) {
                     void * dst_off_device;
                     dpct::memcpy_direction kind;
-                    if (dst->backend == GGML_BACKEND_TYPE_CPU) {
+                    if (dst->backend == GGML_BACKEND_CPU) {
                         dst_off_device = dst->data;
                         kind = dpct::device_to_host;
-                    } else if (dst->backend == GGML_BACKEND_TYPE_GPU) {
+                    } else if (dst->backend == GGML_BACKEND_GPU) {
                         dst_off_device = dst_extra->data_device[g_main_device_index];
                         kind = dpct::device_to_device;
                     } else {
@@ -12954,7 +13080,7 @@ static void ggml_sycl_op_mul_mat(const ggml_tensor *src0,
         }
     }
 
-    if (dst->backend == GGML_BACKEND_TYPE_CPU) {
+    if (dst->backend == GGML_BACKEND_CPU) {
         SYCL_CHECK(ggml_sycl_set_device(g_main_device));
         SYCL_CHECK(CHECK_TRY_ERROR(
             dpct::get_current_device().queues_wait_and_throw()));
@@ -13091,7 +13217,7 @@ static void ggml_sycl_mul_mat_vec_p021(const ggml_tensor *src0,
                                        const ggml_tensor *src1,
                                        ggml_tensor *dst) try {
     GGML_ASSERT(ggml_is_permuted(src0) && ggml_is_permuted(src1));
-    GGML_ASSERT(src0->backend != GGML_BACKEND_TYPE_GPU_SPLIT);
+    GGML_ASSERT(src0->backend != GGML_BACKEND_GPU_SPLIT);
     GGML_ASSERT(src0->nb[0] <= src0->nb[1] && src0->nb[2] <= src0->nb[3]); // 0213 permutation
     GGML_ASSERT(src1->nb[0] <= src1->nb[1] && src1->nb[2] <= src1->nb[3]); // 0213 permutation
     GGML_ASSERT(src0->type == GGML_TYPE_F16);
@@ -13129,7 +13255,7 @@ static void ggml_sycl_mul_mat_vec_nc(const ggml_tensor *src0,
     GGML_ASSERT(!ggml_is_transposed(src0));
     GGML_ASSERT(!ggml_is_transposed(src1));
     GGML_ASSERT(!ggml_is_permuted(src0));
-    GGML_ASSERT(src0->backend != GGML_BACKEND_TYPE_GPU_SPLIT);
+    GGML_ASSERT(src0->backend != GGML_BACKEND_GPU_SPLIT);
     GGML_ASSERT(src0->type == GGML_TYPE_F16);
     GGML_ASSERT(src1->type == GGML_TYPE_F32);
 
@@ -13196,7 +13322,7 @@ static void ggml_sycl_mul_mat_mat_batched_sycl(const ggml_tensor *src0,
     GGML_ASSERT(!ggml_is_transposed(src0));
     GGML_ASSERT(!ggml_is_transposed(src1));
 
-    GGML_ASSERT(src0->backend != GGML_BACKEND_TYPE_GPU_SPLIT);
+    GGML_ASSERT(src0->backend != GGML_BACKEND_GPU_SPLIT);
     GGML_ASSERT(src0->type == GGML_TYPE_F16);
     GGML_ASSERT(src1->type == GGML_TYPE_F32);
 
@@ -13372,11 +13498,11 @@ catch (sycl::exception const &exc) {
 
 static void ggml_sycl_mul_mat(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     const bool all_on_device =
-        (src0->backend == GGML_BACKEND_TYPE_GPU || src0->backend == GGML_BACKEND_TYPE_GPU_SPLIT) &&
-        (src1->backend == GGML_BACKEND_TYPE_GPU) &&
-        ( dst->backend == GGML_BACKEND_TYPE_GPU);
+        (src0->backend == GGML_BACKEND_GPU || src0->backend == GGML_BACKEND_GPU_SPLIT) &&
+        (src1->backend == GGML_BACKEND_GPU) &&
+        ( dst->backend == GGML_BACKEND_GPU);
 
-    const bool split = src0->backend == GGML_BACKEND_TYPE_GPU_SPLIT;
+    const bool split = src0->backend == GGML_BACKEND_GPU_SPLIT;
 
     int64_t min_compute_capability = INT_MAX;
     for (int64_t id = 0; id < g_device_count; ++id) {
@@ -13505,7 +13631,7 @@ static void ggml_sycl_mul_mat_id_sycl(ggml_tensor * dst) {
     GGML_ASSERT(!ggml_is_transposed(src00));
     GGML_ASSERT(!ggml_is_transposed(src1));
 
-    GGML_ASSERT(src00->backend != GGML_BACKEND_TYPE_GPU_SPLIT);
+    GGML_ASSERT(src00->backend != GGML_BACKEND_GPU_SPLIT);
     GGML_ASSERT(src1->type == GGML_TYPE_F32);
 
     GGML_TENSOR_LOCALS(int64_t, ne0, src00, ne);
@@ -13643,7 +13769,7 @@ static void ggml_sycl_mul_mat_id(const ggml_tensor *src0,
 
     const dpct::queue_ptr stream = g_syclStreams[g_main_device_index][0];
 
-    if (ids->backend == GGML_BACKEND_TYPE_GPU) {
+    if (ids->backend == GGML_BACKEND_GPU) {
         const char * ids_dev = (const char *)((const ggml_tensor_extra_gpu *)ids->extra)->data_device[g_main_device_index];
         SYCL_CHECK(CHECK_TRY_ERROR(
             stream->memcpy(ids_host.data(), ids_dev, ggml_nbytes(ids))));
@@ -13661,20 +13787,20 @@ static void ggml_sycl_mul_mat_id(const ggml_tensor *src0,
     ggml_tensor src1_row = *src1;
     ggml_tensor dst_row = *dst;
 
-    src1_row.backend = GGML_BACKEND_TYPE_GPU;
-    dst_row.backend  = GGML_BACKEND_TYPE_GPU;
+    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_TYPE_CPU ?
+    char * src1_original = src1->backend == GGML_BACKEND_CPU ?
         (char *) src1->data : (char *) src1_extra->data_device[g_main_device_index];
-    char * dst_original  =  dst->backend == GGML_BACKEND_TYPE_CPU ?
+    char * dst_original  =  dst->backend == GGML_BACKEND_CPU ?
         (char *)  dst->data : (char *)  dst_extra->data_device[g_main_device_index];
 
     if (src1->ne[1] == 1) {
-        GGML_ASSERT(src1->backend == GGML_BACKEND_TYPE_GPU);
-        GGML_ASSERT(dst->backend  == GGML_BACKEND_TYPE_GPU);
+        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;
@@ -13756,7 +13882,7 @@ static void ggml_sycl_mul_mat_id(const ggml_tensor *src0,
         }
     }
 
-    if (dst->backend == GGML_BACKEND_TYPE_CPU) {
+    if (dst->backend == GGML_BACKEND_CPU) {
         SYCL_CHECK(CHECK_TRY_ERROR(stream->wait()));
     }
 }
@@ -13779,8 +13905,8 @@ static void ggml_sycl_cpy(const ggml_tensor *src0, const ggml_tensor *src1,
     const int64_t ne = ggml_nelements(src0);
     GGML_ASSERT(ne == ggml_nelements(src1));
 
-    GGML_ASSERT(src0->backend == GGML_BACKEND_TYPE_GPU);
-    GGML_ASSERT(src1->backend == GGML_BACKEND_TYPE_GPU);
+    GGML_ASSERT(src0->backend == GGML_BACKEND_GPU);
+    GGML_ASSERT(src1->backend == GGML_BACKEND_GPU);
 
     GGML_ASSERT(ggml_nbytes(src0) <= INT_MAX);
     GGML_ASSERT(ggml_nbytes(src1) <= INT_MAX);
@@ -13887,17 +14013,17 @@ void ggml_sycl_transform_tensor(void *data, struct ggml_tensor *tensor) try {
     memset(extra, 0, sizeof(*extra));
 
     for (int64_t id = 0; id < g_device_count; ++id) {
-        if (backend == GGML_BACKEND_TYPE_GPU && id != g_main_device_index) {
+        if (backend == GGML_BACKEND_GPU && id != g_main_device_index) {
             continue;
         }
         ggml_sycl_set_device(get_device_id_by_index(id));
         const dpct::queue_ptr stream = g_syclStreams[id][0];
 
         int64_t row_low, row_high;
-        if (backend == GGML_BACKEND_TYPE_GPU) {
+        if (backend == GGML_BACKEND_GPU) {
             row_low = 0;
             row_high = nrows;
-        } else if (backend == GGML_BACKEND_TYPE_GPU_SPLIT) {
+        } else if (backend == GGML_BACKEND_GPU_SPLIT) {
             const int64_t rounding = get_row_rounding(tensor->type);
 
             row_low = id == 0 ? 0 : nrows*g_tensor_split[id];
@@ -13946,7 +14072,7 @@ void ggml_sycl_transform_tensor(void *data, struct ggml_tensor *tensor) try {
 
         extra->data_device[id] = buf;
 
-        if (backend == GGML_BACKEND_TYPE_GPU_SPLIT) {
+        if (backend == GGML_BACKEND_GPU_SPLIT) {
             for (int64_t is = 0; is < MAX_STREAMS; ++is) {
                 SYCL_CHECK(CHECK_TRY_ERROR(extra->events[id][is] =
                                                 new sycl::event()));
@@ -13963,7 +14089,7 @@ catch (sycl::exception const &exc) {
 }
 
 void ggml_sycl_free_data(struct ggml_tensor *tensor) try {
-    if (!tensor || !tensor->extra || (tensor->backend != GGML_BACKEND_TYPE_GPU && tensor->backend != GGML_BACKEND_TYPE_GPU_SPLIT) ) {
+    if (!tensor || !tensor->extra || (tensor->backend != GGML_BACKEND_GPU && tensor->backend != GGML_BACKEND_GPU_SPLIT) ) {
         return;
     }
 
@@ -14016,15 +14142,15 @@ static void ggml_sycl_assign_buffers_impl(struct ggml_tensor *tensor,
         return;
     }
 
-    tensor->backend = GGML_BACKEND_TYPE_GPU;
+    tensor->backend = GGML_BACKEND_GPU;
 
-    if (tensor->src[0] != nullptr && tensor->src[0]->backend == GGML_BACKEND_TYPE_CPU) {
+    if (tensor->src[0] != nullptr && tensor->src[0]->backend == GGML_BACKEND_CPU) {
         const ggml_op src0_op = tensor->src[0]->op;
         if (src0_op == GGML_OP_RESHAPE || src0_op == GGML_OP_TRANSPOSE || src0_op == GGML_OP_VIEW || src0_op == GGML_OP_PERMUTE) {
             ggml_sycl_assign_buffers_impl(tensor->src[0], scratch, force_inplace, no_alloc);
         }
     }
-    if (tensor->op == GGML_OP_CPY && tensor->src[1]->backend == GGML_BACKEND_TYPE_CPU) {
+    if (tensor->op == GGML_OP_CPY && tensor->src[1]->backend == GGML_BACKEND_CPU) {
         ggml_sycl_assign_buffers_impl(tensor->src[1], scratch, force_inplace, no_alloc);
     }
 
@@ -14042,7 +14168,7 @@ static void ggml_sycl_assign_buffers_impl(struct ggml_tensor *tensor,
     SYCL_CHECK(ggml_sycl_set_device(g_main_device));
     const dpct::queue_ptr stream = g_syclStreams[g_main_device_index][0];
 
-    if (inplace && (tensor->src[0]->backend == GGML_BACKEND_TYPE_GPU || tensor->src[0]->backend == GGML_BACKEND_TYPE_GPU_SPLIT)) {
+    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;
         char * src0_ddc = (char *) src0_extra->data_device[g_main_device_index];
         size_t offset = 0;
@@ -14111,7 +14237,7 @@ void ggml_sycl_assign_scratch_offset(struct ggml_tensor *tensor,
 
     const bool inplace = tensor->view_src != nullptr;
 
-    if (inplace && (tensor->view_src->backend == GGML_BACKEND_TYPE_GPU || tensor->view_src->backend == GGML_BACKEND_TYPE_GPU_SPLIT)) {
+    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_index];
         size_t view_offset = 0;
@@ -14132,7 +14258,7 @@ catch (sycl::exception const &exc) {
 }
 
 void ggml_sycl_copy_to_device(struct ggml_tensor *tensor) try {
-    GGML_ASSERT(tensor->backend == GGML_BACKEND_TYPE_GPU);
+    GGML_ASSERT(tensor->backend == GGML_BACKEND_GPU);
     GGML_ASSERT(ggml_is_contiguous(tensor));
 
     ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) tensor->extra;
@@ -14219,9 +14345,9 @@ bool ggml_sycl_compute_forward(struct ggml_compute_params * params, struct ggml_
     if (!g_sycl_loaded) return false;
 
     ggml_sycl_func_t func;
-    const bool any_on_device = tensor->backend == GGML_BACKEND_TYPE_GPU
-        || (tensor->src[0] != nullptr && (tensor->src[0]->backend == GGML_BACKEND_TYPE_GPU || tensor->src[0]->backend == GGML_BACKEND_TYPE_GPU_SPLIT))
-        || (tensor->src[1] != nullptr && tensor->src[1]->backend == GGML_BACKEND_TYPE_GPU);
+    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 && tensor->op != GGML_OP_MUL_MAT_ID) {
         return false;
@@ -14359,14 +14485,14 @@ bool ggml_sycl_compute_forward(struct ggml_compute_params * params, struct ggml_
             return false;
     }
 
-    if (tensor->src[0] != nullptr && tensor->src[0]->backend == GGML_BACKEND_TYPE_GPU_SPLIT) {
+    if (tensor->src[0] != nullptr && tensor->src[0]->backend == GGML_BACKEND_GPU_SPLIT) {
         ggml_sycl_set_peer_access(tensor->src[1]->ne[1]);
     }
 
     if (params->ith != 0) {
         return true;
     }
-    if (params->type == GGML_TASK_TYPE_INIT || params->type == GGML_TASK_TYPE_FINALIZE) {
+    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
         return true;
     }
     func(tensor->src[0], tensor->src[1], tensor);
@@ -14517,7 +14643,7 @@ static void ggml_backend_sycl_buffer_init_tensor(ggml_backend_buffer_t buffer,
 
     extra->data_device[ctx->device] = tensor->data;
 
-    tensor->backend = GGML_BACKEND_TYPE_GPU;
+    tensor->backend = GGML_BACKEND_GPU;
     tensor->extra = extra;
 
     if (ggml_is_quantized(tensor->type)) {
@@ -14548,7 +14674,7 @@ static void ggml_backend_sycl_buffer_set_tensor(ggml_backend_buffer_t buffer,
                                                 ggml_tensor *tensor,
                                                 const void *data, size_t offset,
                                                 size_t size) try {
-    GGML_ASSERT(tensor->backend == GGML_BACKEND_TYPE_GPU);
+    GGML_ASSERT(tensor->backend == GGML_BACKEND_GPU);
 
      ggml_backend_sycl_buffer_context * ctx = ( ggml_backend_sycl_buffer_context *)buffer->context;
 
@@ -14573,7 +14699,7 @@ static void ggml_backend_sycl_buffer_get_tensor(ggml_backend_buffer_t buffer,
                                                 const ggml_tensor *tensor,
                                                 void *data, size_t offset,
                                                 size_t size) try {
-    GGML_ASSERT(tensor->backend == GGML_BACKEND_TYPE_GPU);
+    GGML_ASSERT(tensor->backend == GGML_BACKEND_GPU);
 
      ggml_backend_sycl_buffer_context * ctx = ( ggml_backend_sycl_buffer_context *)buffer->context;
 
@@ -14642,8 +14768,7 @@ GGML_CALL static const char * ggml_backend_sycl_buffer_type_name(ggml_backend_bu
 static ggml_backend_buffer_t
 ggml_backend_sycl_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft,
                                            size_t size) try {
-    ggml_backend_sycl_buffer_type_context * buft_ctx = (ggml_backend_sycl_buffer_type_context *)buft->context;
-    int device = (int) buft_ctx->device;
+    int device = (int) (intptr_t) buft->context;
 
     ggml_sycl_set_device(device);
     int device_index = get_device_index_by_id(device);
@@ -14721,7 +14846,7 @@ ggml_backend_buffer_type_t ggml_backend_sycl_buffer_type(int device) {
         for (int i = 0; i < GGML_SYCL_MAX_DEVICES; i++) {
             ggml_backend_sycl_buffer_types[i] = {
                 /* .iface    = */ ggml_backend_sycl_buffer_type_interface,
-                /* .context  = */ new ggml_backend_sycl_buffer_type_context{i, GGML_SYCL_NAME + std::to_string(i)},
+                /* .context  = */ (ggml_backend_buffer_type_context_t) (intptr_t) i,
             };
         }
         ggml_backend_sycl_buffer_type_initialized = true;
@@ -14783,6 +14908,10 @@ ggml_backend_buffer_type_t ggml_backend_sycl_host_buffer_type() {
 
 // backend
 
+struct ggml_backend_context_sycl {
+    int device;
+};
+
 static const char * ggml_backend_sycl_name(ggml_backend_t backend) {
     return GGML_SYCL_NAME;
 
@@ -14790,14 +14919,14 @@ static const char * ggml_backend_sycl_name(ggml_backend_t backend) {
 }
 
 static void ggml_backend_sycl_free(ggml_backend_t backend) {
-    ggml_backend_sycl_context * sycl_ctx = (ggml_backend_sycl_context *)backend->context;
+    ggml_backend_context_sycl * sycl_ctx = (ggml_backend_context_sycl *)backend->context;
 
     delete sycl_ctx;
     delete backend;
 }
 
 static ggml_backend_buffer_type_t ggml_backend_sycl_get_default_buffer_type(ggml_backend_t backend) {
-    ggml_backend_sycl_context * sycl_ctx = (ggml_backend_sycl_context *)backend->context;
+    ggml_backend_context_sycl * sycl_ctx = (ggml_backend_context_sycl *)backend->context;
 
     return ggml_backend_sycl_buffer_type(sycl_ctx->device);
 }
@@ -14806,10 +14935,10 @@ static void ggml_backend_sycl_set_tensor_async(ggml_backend_t backend,
                                                ggml_tensor *tensor,
                                                const void *data, size_t offset,
                                                size_t size) try {
-    ggml_backend_sycl_context * sycl_ctx = (ggml_backend_sycl_context *)backend->context;
+    ggml_backend_context_sycl * sycl_ctx = (ggml_backend_context_sycl *)backend->context;
 
     GGML_ASSERT(tensor->buffer->buft == ggml_backend_sycl_buffer_type(sycl_ctx->device) && "unsupported buffer type");
-    GGML_ASSERT(tensor->backend == GGML_BACKEND_TYPE_GPU);
+    GGML_ASSERT(tensor->backend == GGML_BACKEND_GPU);
 
     SYCL_CHECK(CHECK_TRY_ERROR(g_syclStreams[sycl_ctx->device][0]->memcpy(
         (char *)tensor->data + offset, data, size)));
@@ -14824,10 +14953,10 @@ static void ggml_backend_sycl_get_tensor_async(ggml_backend_t backend,
                                                const ggml_tensor *tensor,
                                                void *data, size_t offset,
                                                size_t size) try {
-    ggml_backend_sycl_context * sycl_ctx = (ggml_backend_sycl_context *)backend->context;
+    ggml_backend_context_sycl * sycl_ctx = (ggml_backend_context_sycl *)backend->context;
 
     GGML_ASSERT(tensor->buffer->buft == ggml_backend_sycl_buffer_type(sycl_ctx->device) && "unsupported buffer type");
-    GGML_ASSERT(tensor->backend == GGML_BACKEND_TYPE_GPU);
+    GGML_ASSERT(tensor->backend == GGML_BACKEND_GPU);
 
     SYCL_CHECK(CHECK_TRY_ERROR(g_syclStreams[sycl_ctx->device][0]->memcpy(
         data, (const char *)tensor->data + offset, size)));
@@ -14839,7 +14968,7 @@ catch (sycl::exception const &exc) {
 }
 
 static void ggml_backend_sycl_synchronize(ggml_backend_t backend) try {
-    ggml_backend_sycl_context * sycl_ctx = (ggml_backend_sycl_context *)backend->context;
+    ggml_backend_context_sycl * sycl_ctx = (ggml_backend_context_sycl *)backend->context;
 
     SYCL_CHECK(CHECK_TRY_ERROR(g_syclStreams[sycl_ctx->device][0]->wait()));
 
@@ -14875,12 +15004,12 @@ static void ggml_backend_sycl_graph_plan_compute(ggml_backend_t backend, ggml_ba
 }
 
 static bool ggml_backend_sycl_graph_compute(ggml_backend_t backend, ggml_cgraph * cgraph) {
-    ggml_backend_sycl_context * sycl_ctx = (ggml_backend_sycl_context *)backend->context;
+    ggml_backend_context_sycl * sycl_ctx = (ggml_backend_context_sycl *)backend->context;
 
     ggml_sycl_set_main_device(sycl_ctx->device);
 
     ggml_compute_params params = {};
-    params.type = GGML_TASK_TYPE_COMPUTE;
+    params.type = GGML_TASK_COMPUTE;
     params.ith = 0;
     for (int i = 0; i < cgraph->n_nodes; i++) {
         ggml_tensor * node = cgraph->nodes[i];
@@ -14888,13 +15017,13 @@ static bool ggml_backend_sycl_graph_compute(ggml_backend_t backend, ggml_cgraph
         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_TYPE_GPU);
+        assert(node->backend == GGML_BACKEND_GPU);
         assert(node->buffer->buft == ggml_backend_sycl_buffer_type(sycl_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_TYPE_GPU);
+                assert(node->src[j]->backend == GGML_BACKEND_GPU);
                 assert(node->src[j]->buffer->buft == ggml_backend_sycl_buffer_type(sycl_ctx->device));
                 assert(node->src[j]->extra != nullptr);
             }
@@ -14964,12 +15093,6 @@ static bool ggml_backend_sycl_supports_op(ggml_backend_t backend, const ggml_ten
                     return false;
                 }
 
-                if (a->type == GGML_TYPE_IQ1_S) {
-                    return false;
-                }
-                if (a->type == GGML_TYPE_IQ3_XXS) {
-                  return false;
-                }
                 if (a->type == GGML_TYPE_IQ2_XXS) {
                     return false;
                 }
@@ -15089,9 +15212,8 @@ ggml_backend_t ggml_backend_sycl_init(int device) {
     // not strictly necessary, but it may reduce the overhead of the first graph_compute
     ggml_sycl_set_main_device(device);
 
-    ggml_backend_sycl_context * ctx = new ggml_backend_sycl_context {
-        /* .device = */ device,
-        /* .name   = */ GGML_SYCL_NAME + std::to_string(device),
+    ggml_backend_context_sycl * ctx = new ggml_backend_context_sycl {
+        /* .device = */ device
     };
 
     ggml_backend_t sycl_backend = new ggml_backend {

ggml-vulkan.cpp

@@ -1091,10 +1091,7 @@ static void ggml_vk_print_gpu_info(size_t idx) {
     }
 }
 
-static bool ggml_vk_instance_validation_ext_available(const std::vector<vk::ExtensionProperties>& instance_extensions);
-static bool ggml_vk_instance_portability_enumeration_ext_available(const std::vector<vk::ExtensionProperties>& instance_extensions);
-
-void ggml_vk_instance_init() {
+static void ggml_vk_instance_init() {
     if (vk_instance_initialized) {
         return;
     }
@@ -1103,42 +1100,28 @@ void ggml_vk_instance_init() {
 #endif
 
     vk::ApplicationInfo app_info{ "ggml-vulkan", 1, nullptr, 0, VK_API_VERSION };
+    const std::vector<const char*> layers = {
+#ifdef GGML_VULKAN_VALIDATE
+        "VK_LAYER_KHRONOS_validation",
+#endif
+    };
+    const std::vector<const char*> extensions = {
+#ifdef GGML_VULKAN_VALIDATE
+        "VK_EXT_validation_features",
+#endif
+    };
+    vk::InstanceCreateInfo instance_create_info(vk::InstanceCreateFlags(), &app_info, layers, extensions);
+#ifdef GGML_VULKAN_VALIDATE
+    const std::vector<vk::ValidationFeatureEnableEXT> features_enable = { vk::ValidationFeatureEnableEXT::eBestPractices };
+    vk::ValidationFeaturesEXT validation_features = {
+        features_enable,
+        {},
+    };
+    validation_features.setPNext(nullptr);
+    instance_create_info.setPNext(&validation_features);
 
-    const std::vector<vk::ExtensionProperties> instance_extensions = vk::enumerateInstanceExtensionProperties();
-    const bool validation_ext = ggml_vk_instance_validation_ext_available(instance_extensions);
-    const bool portability_enumeration_ext = ggml_vk_instance_portability_enumeration_ext_available(instance_extensions);
-
-    std::vector<const char*> layers;
-
-    if (validation_ext) {
-        layers.push_back("VK_LAYER_KHRONOS_validation");
-    }
-    std::vector<const char*> extensions;
-    if (validation_ext) {
-        extensions.push_back("VK_EXT_validation_features");
-    }
-    if (portability_enumeration_ext) {
-        extensions.push_back("VK_KHR_portability_enumeration");
-    }
-    vk::InstanceCreateInfo instance_create_info(vk::InstanceCreateFlags{}, &app_info, layers, extensions);
-    if (portability_enumeration_ext) {
-        instance_create_info.flags |= vk::InstanceCreateFlagBits::eEnumeratePortabilityKHR;
-    }
-
-    std::vector<vk::ValidationFeatureEnableEXT> features_enable;
-    vk::ValidationFeaturesEXT validation_features;
-
-    if (validation_ext) {
-        features_enable = { vk::ValidationFeatureEnableEXT::eBestPractices };
-        validation_features = {
-            features_enable,
-            {},
-        };
-        validation_features.setPNext(nullptr);
-        instance_create_info.setPNext(&validation_features);
-
-        std::cerr << "ggml_vulkan: Validation layers enabled" << std::endl;
-    }
+    std::cerr << "ggml_vulkan: Validation layers enabled" << std::endl;
+#endif
     vk_instance.instance = vk::createInstance(instance_create_info);
 
     memset(vk_instance.initialized, 0, sizeof(bool) * GGML_VK_MAX_DEVICES);
@@ -1185,12 +1168,12 @@ static void ggml_vk_init(ggml_backend_vk_context * ctx, size_t idx) {
     vk_instance.devices[idx] = std::make_shared<vk_device>();
     ctx->device = vk_instance.devices[idx];
     ctx->device.lock()->physical_device = devices[dev_num];
-    const std::vector<vk::ExtensionProperties> ext_props = ctx->device.lock()->physical_device.enumerateDeviceExtensionProperties();
+    std::vector<vk::ExtensionProperties> ext_props = ctx->device.lock()->physical_device.enumerateDeviceExtensionProperties();
 
     bool maintenance4_support = false;
 
     // Check if maintenance4 is supported
-    for (const auto& properties : ext_props) {
+    for (auto properties : ext_props) {
         if (strcmp("VK_KHR_maintenance4", properties.extensionName) == 0) {
             maintenance4_support = true;
         }
@@ -1221,7 +1204,7 @@ static void ggml_vk_init(ggml_backend_vk_context * ctx, size_t idx) {
     bool fp16_storage = false;
     bool fp16_compute = false;
 
-    for (const auto& properties : ext_props) {
+    for (auto properties : ext_props) {
         if (strcmp("VK_KHR_16bit_storage", properties.extensionName) == 0) {
             fp16_storage = true;
         } else if (strcmp("VK_KHR_shader_float16_int8", properties.extensionName) == 0) {
@@ -2320,8 +2303,8 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context * su
         src1_uma = d_Qy != nullptr;
     }
 
-    const bool load_x = src0->backend != GGML_BACKEND_TYPE_GPU && !src0_uma;
-    const bool load_y = src1->backend != GGML_BACKEND_TYPE_GPU && !src1_uma;
+    const bool load_x = src0->backend != GGML_BACKEND_GPU && !src0_uma;
+    const bool load_y = src1->backend != GGML_BACKEND_GPU && !src1_uma;
 
     const bool x_non_contig = !load_x && !ggml_vk_dim01_contiguous(src0);
     const bool y_non_contig = !load_y && !ggml_vk_dim01_contiguous(src1);
@@ -2453,7 +2436,7 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context * su
     // compute
     ggml_vk_matmul(ctx, subctx, *pipeline, { d_X, x_buf_offset, x_sz * ne02 * ne03 }, { d_Y, y_buf_offset, y_sz * ne12 * ne13 }, { d_D, d_buf_offset, d_sz * ne12 * ne13 }, { ctx->prealloc_split_k, 0, d_sz * ne12 * ne13 * split_k }, ne01, ne11, ne10, ne10, ne10, ne01, split_k, ne12*ne13, ne02, ne12, r2, r3, stride_batch_x, stride_batch_y, ne20*ne21);  // NOLINT
 
-    if (dst->backend == GGML_BACKEND_TYPE_CPU) {
+    if (dst->backend == GGML_BACKEND_CPU) {
         // copy dst to host
         float * d = (float *) ((char *) dst->data);
         ggml_vk_buffer_read_async(ctx, subctx, d_D, 0, d, sizeof(float) * d_ne * ne12 * ne13);
@@ -2506,8 +2489,8 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context
         src1_uma = d_Qy != nullptr;
     }
 
-    const bool load_x = src0->backend != GGML_BACKEND_TYPE_GPU && !src0_uma;
-    const bool load_y = src1->backend != GGML_BACKEND_TYPE_GPU && !src1_uma;
+    const bool load_x = src0->backend != GGML_BACKEND_GPU && !src0_uma;
+    const bool load_y = src1->backend != GGML_BACKEND_GPU && !src1_uma;
 
     const bool x_non_contig = !load_x && !ggml_vk_dim01_contiguous(src0);
     const bool y_non_contig = !load_y && !ggml_vk_dim01_contiguous(src1);
@@ -2630,7 +2613,7 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context
             ggml_vk_sync_buffers(subctx);
             ggml_vk_dispatch_pipeline(ctx, subctx, *dmmv, { { d_X, x_offset, x_sz }, { d_Y, y_buffer_offset, y_sz + y_shader_offset }, { d_D, d_buffer_offset, d_sz + d_shader_offset } }, 3 * sizeof(int), &pc, { (uint32_t)ne01, 1, 1});
 
-            if (dst->backend == GGML_BACKEND_TYPE_CPU) {
+            if (dst->backend == GGML_BACKEND_CPU) {
                 // copy dst to host
                 float * d = (float *) ((char *) dst->data + i12*nb2 + i13*nb3);
                 ggml_vk_sync_buffers(subctx);
@@ -2647,7 +2630,7 @@ static void ggml_vk_mul_mat_vec_p021_f16_f32(ggml_backend_vk_context * ctx, vk_c
     std::cerr << "), (" << dst << ", name=" << dst->name << ", type=" << dst->type << ",  backend=" << dst->backend << ", ne0=" << dst->ne[0] << ", ne1=" << dst->ne[1] << ", ne2=" << dst->ne[2] << ", ne3=" << dst->ne[3] << ", nb0=" << dst->nb[0] << ", nb1=" << dst->nb[1] << ", nb2=" << dst->nb[2] << ", nb3=" << dst->nb[3] << "),)" << std::endl;
 #endif
     GGML_ASSERT(ggml_is_permuted(src0) && ggml_is_permuted(src1));
-    GGML_ASSERT(src0->backend == GGML_BACKEND_TYPE_GPU);
+    GGML_ASSERT(src0->backend == GGML_BACKEND_GPU);
     GGML_ASSERT(src0->nb[0] <= src0->nb[1] && src0->nb[2] <= src0->nb[3]);  // NOLINT
     GGML_ASSERT(src1->nb[0] <= src1->nb[1] && src1->nb[2] <= src1->nb[3]);  // NOLINT
     GGML_ASSERT(src0->type == GGML_TYPE_F16);
@@ -2679,7 +2662,7 @@ static void ggml_vk_mul_mat_vec_p021_f16_f32(ggml_backend_vk_context * ctx, vk_c
         src1_uma = d_Qy != nullptr;
     }
 
-    const bool load_y = src1->backend != GGML_BACKEND_TYPE_GPU && !src1_uma;
+    const bool load_y = src1->backend != GGML_BACKEND_GPU && !src1_uma;
 
     const uint64_t x_ne = ne00 * ne01 * ne02;
     const uint64_t y_ne = ne10 * ne11 * ne12;
@@ -2721,7 +2704,7 @@ static void ggml_vk_mul_mat_vec_p021_f16_f32(ggml_backend_vk_context * ctx, vk_c
     ggml_vk_sync_buffers(subctx);
     ggml_vk_dispatch_pipeline(ctx, subctx, ctx->pipeline_mul_mat_vec_p021_f16_f32, { { d_Qx, qx_buf_offset, qx_sz }, { d_Qy, qy_buffer_offset, qy_sz + qy_shader_offset }, { d_D, d_buffer_offset, d_sz + d_shader_offset } }, 6 * sizeof(uint32_t), &pc, { 1, (uint32_t)ne01, (uint32_t)ne12 });
 
-    if (dst->backend == GGML_BACKEND_TYPE_CPU) {
+    if (dst->backend == GGML_BACKEND_CPU) {
         // copy dst to host
         float * d = (float *) dst->data;
         ggml_vk_sync_buffers(subctx);
@@ -2738,7 +2721,7 @@ static void ggml_vk_mul_mat_vec_nc_f16_f32(ggml_backend_vk_context * ctx, vk_con
     GGML_ASSERT(!ggml_is_transposed(src0));
     GGML_ASSERT(!ggml_is_transposed(src1));
     GGML_ASSERT(!ggml_is_permuted(src0));
-    GGML_ASSERT(src0->backend == GGML_BACKEND_TYPE_GPU);
+    GGML_ASSERT(src0->backend == GGML_BACKEND_GPU);
     GGML_ASSERT(src0->type == GGML_TYPE_F16);
     GGML_ASSERT(src1->type == GGML_TYPE_F32);
 
@@ -2771,7 +2754,7 @@ static void ggml_vk_mul_mat_vec_nc_f16_f32(ggml_backend_vk_context * ctx, vk_con
         src1_uma = d_Qy != nullptr;
     }
 
-    const bool load_y = src1->backend != GGML_BACKEND_TYPE_GPU && !src1_uma;
+    const bool load_y = src1->backend != GGML_BACKEND_GPU && !src1_uma;
 
     const uint64_t d_ne = ne01 * ne11 * ne12;
 
@@ -2814,7 +2797,7 @@ static void ggml_vk_mul_mat_vec_nc_f16_f32(ggml_backend_vk_context * ctx, vk_con
     ggml_vk_sync_buffers(subctx);
     ggml_vk_dispatch_pipeline(ctx, subctx, ctx->pipeline_mul_mat_vec_nc_f16_f32, { { d_Qx, qx_buf_offset, qx_sz }, { d_Qy, qy_buffer_offset, qy_sz + qy_shader_offset }, { d_D, d_buffer_offset, d_sz + d_shader_offset } }, 7 * sizeof(uint32_t), &pc, { 1, (uint32_t)ne01, (uint32_t)ne12 });
 
-    if (dst->backend == GGML_BACKEND_TYPE_CPU) {
+    if (dst->backend == GGML_BACKEND_CPU) {
         // copy dst to host
         float * d = (float *) dst->data;
         ggml_vk_sync_buffers(subctx);
@@ -2832,7 +2815,7 @@ static bool ggml_vk_can_mul_mat(const ggml_tensor * src0, const ggml_tensor * sr
     return (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 || ggml_is_quantized(src0->type)) &&
            (src1->type == GGML_TYPE_F32 || src1->type == GGML_TYPE_F16 || ggml_is_quantized(src1->type)) &&
            dst->type == GGML_TYPE_F32 &&
-           ((ne0 >= 32 && ne1 >= 32 && ne10 >= 32) || src0->backend == GGML_BACKEND_TYPE_GPU);
+           ((ne0 >= 32 && ne1 >= 32 && ne10 >= 32) || src0->backend == GGML_BACKEND_GPU);
 }
 
 static void ggml_vk_mul_mat(ggml_backend_vk_context * ctx, vk_context * subctx, const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst) {
@@ -2880,8 +2863,8 @@ static void ggml_vk_op_repeat(ggml_backend_vk_context * ctx, vk_context * subctx
     // TODO: support for transposed / permuted tensors
     GGML_ASSERT(nb0  == sizeof(float));
     GGML_ASSERT(nb00 == sizeof(float));
-    GGML_ASSERT(src0->backend == GGML_BACKEND_TYPE_GPU);
-    GGML_ASSERT(dst->backend == GGML_BACKEND_TYPE_GPU);
+    GGML_ASSERT(src0->backend == GGML_BACKEND_GPU);
+    GGML_ASSERT(dst->backend == GGML_BACKEND_GPU);
 
     ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) dst->extra;
     ggml_tensor_extra_gpu * extra_src0 = (ggml_tensor_extra_gpu *) src0->extra;
@@ -3110,8 +3093,8 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context * subctx, c
         }
     }
 
-    const bool transfer_src0 = src0->backend != GGML_BACKEND_TYPE_GPU && !src0_uma;
-    const bool transfer_src1 = use_src1 && src1->backend != GGML_BACKEND_TYPE_GPU && !src1_uma;
+    const bool transfer_src0 = src0->backend != GGML_BACKEND_GPU && !src0_uma;
+    const bool transfer_src1 = use_src1 && src1->backend != GGML_BACKEND_GPU && !src1_uma;
 
     uint64_t x_sz = ggml_vk_align_size(ggml_type_size(src0->type) * ne0, ctx->device.lock()->properties.limits.minStorageBufferOffsetAlignment);
     uint64_t y_sz = use_src1 ? ggml_vk_align_size(ggml_type_size(src1->type) * ne1, ctx->device.lock()->properties.limits.minStorageBufferOffsetAlignment) : 0;
@@ -3120,7 +3103,7 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context * subctx, c
     vk_buffer d_D = extra->buffer_gpu.lock();
 
     // Workaround for tiny tensor inputs on ROPE
-    if (use_src1 && src1->backend == GGML_BACKEND_TYPE_GPU && y_sz > d_D->size) {
+    if (use_src1 && src1->backend == GGML_BACKEND_GPU && y_sz > d_D->size) {
         y_sz = VK_WHOLE_SIZE;
     }
 
@@ -3209,9 +3192,9 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context * subctx, c
             ggml_vk_sync_buffers(subctx);
             ggml_vk_dispatch_pipeline(ctx, subctx, *pipeline, { { d_X, x_buf_offset, x_sz }, { d_D, d_buf_offset, d_sz } }, sizeof(PC), &pc, elements);
         }
-        if (dst->backend == GGML_BACKEND_TYPE_CPU && op == GGML_OP_CPY) {
+        if (dst->backend == GGML_BACKEND_CPU && op == GGML_OP_CPY) {
             ggml_vk_d2h_tensor_2d(ctx, subctx, d_D, 0, dst);
-        } else if(dst->backend == GGML_BACKEND_TYPE_CPU) {
+        } else if(dst->backend == GGML_BACKEND_CPU) {
             // copy dst to host
             float * d = (float *) dst->data;
             ggml_vk_buffer_read_async(ctx, subctx, d_D, 0, d, d_sz);
@@ -3253,7 +3236,7 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context * subctx, c
                     ggml_vk_sync_buffers(subctx);
                     ggml_vk_dispatch_pipeline(ctx, subctx, *pipeline, { { d_X, x_buf_offset + x_offset, x_sz }, { d_D, d_buf_offset + d_offset, d_sz } }, sizeof(PC), &pc, elements);
                 }
-                if (dst->backend == GGML_BACKEND_TYPE_CPU) {
+                if (dst->backend == GGML_BACKEND_CPU) {
                     // copy dst to host
                     ggml_vk_buffer_read_async(ctx, subctx, d_D, d_buf_offset + d_offset, (char *) dst->data + i02*nb2 + i03*nb3, d_sz);
                 }
@@ -3359,7 +3342,7 @@ static void ggml_vk_rope(ggml_backend_vk_context * ctx, vk_context * subctx, con
 
 static void ggml_vk_nop(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, ggml_tensor * dst) {
     // If backend is CPU, data from src0 has to be copied off the device
-    if (dst->backend == GGML_BACKEND_TYPE_CPU) {
+    if (dst->backend == GGML_BACKEND_CPU) {
         ggml_tensor_extra_gpu * extra_src0 = (ggml_tensor_extra_gpu *) src0->extra;
         vk_buffer d_D = extra_src0->buffer_gpu.lock();
         ggml_vk_sync_buffers(subctx);
@@ -3994,9 +3977,9 @@ static void ggml_vk_preallocate_buffers_graph(ggml_backend_vk_context * ctx, ggm
 #ifdef GGML_VULKAN_DEBUG
     std::cerr << "ggml_vk_preallocate_buffers_graph(" << node << ")" << std::endl;
 #endif
-    const bool any_on_device = node->backend == GGML_BACKEND_TYPE_GPU
-        || (node->src[0] != nullptr && (node->src[0]->backend == GGML_BACKEND_TYPE_GPU || node->src[0]->backend == GGML_BACKEND_TYPE_GPU_SPLIT))
-        || (node->src[1] != nullptr && (node->src[1]->backend == GGML_BACKEND_TYPE_GPU));
+    const bool any_on_device = node->backend == GGML_BACKEND_GPU
+        || (node->src[0] != nullptr && (node->src[0]->backend == GGML_BACKEND_GPU || node->src[0]->backend == GGML_BACKEND_GPU_SPLIT))
+        || (node->src[1] != nullptr && (node->src[1]->backend == GGML_BACKEND_GPU));
 
     if (ctx->disable || (!any_on_device && node->op != GGML_OP_MUL_MAT)) {
         return;
@@ -4215,9 +4198,9 @@ static void ggml_vk_preallocate_buffers(ggml_backend_vk_context * ctx) {
 }
 
 static void ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * node, bool last_node){
-    const bool any_on_device = node->backend == GGML_BACKEND_TYPE_GPU
-        || (node->src[0] != nullptr && (node->src[0]->backend == GGML_BACKEND_TYPE_GPU || node->src[0]->backend == GGML_BACKEND_TYPE_GPU_SPLIT))
-        || (node->src[1] != nullptr && node->src[1]->backend == GGML_BACKEND_TYPE_GPU);
+    const bool any_on_device = node->backend == GGML_BACKEND_GPU
+        || (node->src[0] != nullptr && (node->src[0]->backend == GGML_BACKEND_GPU || node->src[0]->backend == GGML_BACKEND_GPU_SPLIT))
+        || (node->src[1] != nullptr && node->src[1]->backend == GGML_BACKEND_GPU);
 
     if (ctx->disable || (!any_on_device && node->op != GGML_OP_MUL_MAT) || (node->op == GGML_OP_MUL_MAT && !any_on_device && !ggml_vk_can_mul_mat(node->src[0], node->src[1], node))) {
         return;
@@ -4371,7 +4354,7 @@ static void ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod
     last_node = true;
 #endif
 
-    if (node->backend == GGML_BACKEND_TYPE_CPU || last_node) {
+    if (node->backend == GGML_BACKEND_CPU || last_node) {
         ggml_vk_ctx_end(ctx->compute_ctx);
         ctx->compute_ctx->exit_tensor = node;
         ctx->compute_ctx = nullptr;
@@ -4379,9 +4362,9 @@ static void ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod
 }
 
 static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_compute_params * params, ggml_tensor * tensor){
-    const bool any_on_device = tensor->backend == GGML_BACKEND_TYPE_GPU
-        || (tensor->src[0] != nullptr && (tensor->src[0]->backend == GGML_BACKEND_TYPE_GPU || tensor->src[0]->backend == GGML_BACKEND_TYPE_GPU_SPLIT))
-        || (tensor->src[1] != nullptr && tensor->src[1]->backend == GGML_BACKEND_TYPE_GPU);
+    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 (ctx->disable || (!any_on_device && tensor->op != GGML_OP_MUL_MAT)) {
         return false;
@@ -4442,7 +4425,7 @@ static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_compute_
     if (params->ith != 0) {
         return true;
     }
-    if (params->type == GGML_TASK_TYPE_INIT || params->type == GGML_TASK_TYPE_FINALIZE) {
+    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
         return true;
     }
 
@@ -4745,7 +4728,7 @@ GGML_CALL static void ggml_backend_vk_buffer_init_tensor(ggml_backend_buffer_t b
         extra->offset = (uint8_t *) tensor->data - (uint8_t *) vk_ptr_base;
     }
 
-    tensor->backend = GGML_BACKEND_TYPE_GPU;
+    tensor->backend = GGML_BACKEND_GPU;
     tensor->extra = extra;
 }
 
@@ -4753,7 +4736,7 @@ GGML_CALL static void ggml_backend_vk_buffer_set_tensor(ggml_backend_buffer_t bu
 #ifdef GGML_VULKAN_DEBUG
     std::cerr << "ggml_backend_vk_buffer_set_tensor(" << buffer << ", " << tensor << ", " << data << ", " << offset << ", " << size << ")" << std::endl;
 #endif
-    GGML_ASSERT(tensor->backend == GGML_BACKEND_TYPE_GPU);
+    GGML_ASSERT(tensor->backend == GGML_BACKEND_GPU);
 
     ggml_backend_vk_buffer_context * ctx = (ggml_backend_vk_buffer_context *)buffer->context;
 
@@ -4768,7 +4751,7 @@ GGML_CALL static void ggml_backend_vk_buffer_get_tensor(ggml_backend_buffer_t bu
 #ifdef GGML_VULKAN_DEBUG
     std::cerr << "ggml_backend_vk_buffer_get_tensor(" << buffer << ", " << tensor << ", " << data << ", " << offset << ", " << size << ")" << std::endl;
 #endif
-    GGML_ASSERT(tensor->backend == GGML_BACKEND_TYPE_GPU);
+    GGML_ASSERT(tensor->backend == GGML_BACKEND_GPU);
 
     ggml_backend_vk_buffer_context * ctx = (ggml_backend_vk_buffer_context *)buffer->context;
 
@@ -4999,7 +4982,7 @@ GGML_CALL static void ggml_backend_vk_set_tensor_async(ggml_backend_t backend, g
 #endif
     ggml_backend_vk_context * ctx = (ggml_backend_vk_context *)backend->context;
     GGML_ASSERT((tensor->buffer->buft == ggml_backend_vk_buffer_type(ctx->idx) || tensor->buffer->buft == ggml_backend_vk_host_buffer_type()) && "unsupported buffer type");
-    GGML_ASSERT(tensor->backend == GGML_BACKEND_TYPE_GPU);
+    GGML_ASSERT(tensor->backend == GGML_BACKEND_GPU);
 
     ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) tensor->extra;
 
@@ -5020,7 +5003,7 @@ GGML_CALL static void ggml_backend_vk_get_tensor_async(ggml_backend_t backend, c
 #endif
     ggml_backend_vk_context * ctx = (ggml_backend_vk_context *)backend->context;
     GGML_ASSERT((tensor->buffer->buft == ggml_backend_vk_buffer_type(ctx->idx) || tensor->buffer->buft == ggml_backend_vk_host_buffer_type()) && "unsupported buffer type");
-    GGML_ASSERT(tensor->backend == GGML_BACKEND_TYPE_GPU);
+    GGML_ASSERT(tensor->backend == GGML_BACKEND_GPU);
 
     ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) tensor->extra;
 
@@ -5097,7 +5080,7 @@ GGML_CALL static bool ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml
     int last_node = cgraph->n_nodes - 1;
 
     // If the last op in the cgraph isn't backend GPU, the command buffer doesn't get closed properly
-    while (last_node > 0 && cgraph->nodes[last_node]->backend != GGML_BACKEND_TYPE_GPU) {
+    while (last_node > 0 && cgraph->nodes[last_node]->backend != GGML_BACKEND_GPU) {
         last_node -= 1;
     }
 
@@ -5106,7 +5089,7 @@ GGML_CALL static bool ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml
     }
 
     ggml_compute_params params = {};
-    params.type = GGML_TASK_TYPE_COMPUTE;
+    params.type = GGML_TASK_COMPUTE;
     params.ith = 0;
     for (int i = 0; i < cgraph->n_nodes; i++) {
         ggml_tensor * node = cgraph->nodes[i];
@@ -5318,42 +5301,6 @@ GGML_CALL int ggml_backend_vk_reg_devices() {
     return vk_instance.device_indices.size();
 }
 
-// Extension availability
-static bool ggml_vk_instance_validation_ext_available(const std::vector<vk::ExtensionProperties>& instance_extensions) {
-#ifdef GGML_VULKAN_VALIDATE
-    bool portability_enumeration_ext = false;
-    // Check for portability enumeration extension for MoltenVK support
-    for (const auto& properties : instance_extensions) {
-        if (strcmp("VK_KHR_portability_enumeration", properties.extensionName) == 0) {
-            return true;
-        }
-    }
-    if (!portability_enumeration_ext) {
-        std::cerr << "ggml_vulkan: WARNING: Instance extension VK_KHR_portability_enumeration not found." << std::endl;
-    }
-#endif
-    return false;
-
-    UNUSED(instance_extensions);
-}
-static bool ggml_vk_instance_portability_enumeration_ext_available(const std::vector<vk::ExtensionProperties>& instance_extensions) {
-#ifdef __APPLE__
-    bool portability_enumeration_ext = false;
-    // Check for portability enumeration extension for MoltenVK support
-    for (const auto& properties : instance_extensions) {
-        if (strcmp("VK_KHR_portability_enumeration", properties.extensionName) == 0) {
-            return true;
-        }
-    }
-    if (!portability_enumeration_ext) {
-        std::cerr << "ggml_vulkan: WARNING: Instance extension VK_KHR_portability_enumeration not found." << std::endl;
-    }
-#endif
-    return false;
-
-    UNUSED(instance_extensions);
-}
-
 // checks
 
 #ifdef GGML_VULKAN_CHECK_RESULTS
@@ -5410,7 +5357,7 @@ static void ggml_vk_print_tensor_area(const ggml_tensor * tensor, const void * d
 static void ggml_vk_print_tensor(ggml_backend_vk_context * ctx, const ggml_tensor * tensor, const char * name) {
     void * tensor_data = tensor->data;
 
-    if (tensor->backend == GGML_BACKEND_TYPE_GPU) {
+    if (tensor->backend == GGML_BACKEND_GPU) {
         const size_t tensor_size = ggml_nbytes(tensor);
         tensor_data = malloc(tensor_size);
 
@@ -5436,14 +5383,14 @@ static void ggml_vk_print_tensor(ggml_backend_vk_context * ctx, const ggml_tenso
     std::vector<const ggml_tensor *> done;
     ggml_vk_print_graph_origin(tensor, done);
 
-    if (tensor->backend == GGML_BACKEND_TYPE_GPU) {
+    if (tensor->backend == GGML_BACKEND_GPU) {
         free(tensor_data);
     }
 }
 
 static void ggml_vk_check_tensor(const std::string& name, const ggml_tensor * tensor) {
     return;
-    GGML_ASSERT(tensor->backend == GGML_BACKEND_TYPE_CPU);
+    GGML_ASSERT(tensor->backend == GGML_BACKEND_CPU);
     if (tensor->type != GGML_TYPE_F32 && tensor->type != GGML_TYPE_F16) {
         return;
     }
@@ -5481,7 +5428,7 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_compute_
     if (params->ith != 0) {
         return;
     }
-    if (params->type == GGML_TASK_TYPE_INIT || params->type == GGML_TASK_TYPE_FINALIZE || tensor->op == GGML_OP_TRANSPOSE) {
+    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE || tensor->op == GGML_OP_TRANSPOSE) {
         return;
     }
 
@@ -5518,10 +5465,10 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_compute_
 
         src0_buffer = malloc(src0_size);
         src0_clone->data = src0_buffer;
-        if (src0->backend == GGML_BACKEND_TYPE_CPU) {
+        if (src0->backend == GGML_BACKEND_CPU) {
             memcpy(src0_clone->data, src0->data, src0_size);
             memcpy(src0_clone->nb, src0->nb, sizeof(size_t) * GGML_MAX_DIMS);
-        } else if (src0->backend == GGML_BACKEND_TYPE_GPU) {
+        } else if (src0->backend == GGML_BACKEND_GPU) {
             ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) src0->extra;
             uint64_t offset = extra->offset;
             if (!ggml_is_contiguous(src0) && ggml_vk_dim01_contiguous(src0)) {
@@ -5561,10 +5508,10 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_compute_
 
         src1_buffer = malloc(src1_size);
         src1_clone->data = src1_buffer;
-        if (src1->backend == GGML_BACKEND_TYPE_CPU) {
+        if (src1->backend == GGML_BACKEND_CPU) {
             memcpy(src1_clone->data, src1->data, src1_size);
             memcpy(src1_clone->nb, src1->nb, sizeof(size_t) * GGML_MAX_DIMS);
-        } else if (src1->backend == GGML_BACKEND_TYPE_GPU) {
+        } else if (src1->backend == GGML_BACKEND_GPU) {
             ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) src1->extra;
             uint64_t offset = extra->offset;
             if (!ggml_is_contiguous(src1) && ggml_vk_dim01_contiguous(src1)) {
@@ -5723,7 +5670,7 @@ static void ggml_vk_check_results_1(ggml_backend_vk_context * ctx, ggml_compute_
     if (params->ith != 0) {
         return;
     }
-    if (params->type == GGML_TASK_TYPE_INIT || params->type == GGML_TASK_TYPE_FINALIZE || tensor->op == GGML_OP_TRANSPOSE) {
+    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE || tensor->op == GGML_OP_TRANSPOSE) {
         return;
     }
     if (!(vk_output_tensor > 0 && vk_output_tensor == check_counter) && check_counter <= vk_skip_checks) {
@@ -5735,7 +5682,7 @@ static void ggml_vk_check_results_1(ggml_backend_vk_context * ctx, ggml_compute_
 
     void * tensor_data = tensor->data;
 
-    if (tensor->backend == GGML_BACKEND_TYPE_GPU) {
+    if (tensor->backend == GGML_BACKEND_GPU) {
         size_t tensor_size = ggml_nbytes(tensor);
         tensor_data = malloc(tensor_size);
 
@@ -5868,7 +5815,7 @@ static void ggml_vk_check_results_1(ggml_backend_vk_context * ctx, ggml_compute_
     comp_result = nullptr;
     comp_size = 0;
 
-    if (tensor->backend == GGML_BACKEND_TYPE_GPU) {
+    if (tensor->backend == GGML_BACKEND_GPU) {
         free(tensor_data);
     }
 }

ggml.h

@@ -315,7 +315,13 @@
 extern "C" {
 #endif
 
+#if defined(__ARM_NEON) && defined(__CUDACC__)
+    typedef half ggml_fp16_t;
+#elif defined(__ARM_NEON) && !defined(_MSC_VER)
+    typedef __fp16 ggml_fp16_t;
+#else
     typedef uint16_t ggml_fp16_t;
+#endif
 
     // convert FP16 <-> FP32
     GGML_API float       ggml_fp16_to_fp32(ggml_fp16_t x);
@@ -349,8 +355,6 @@ extern "C" {
         GGML_TYPE_IQ2_XS  = 17,
         GGML_TYPE_IQ3_XXS = 18,
         GGML_TYPE_IQ1_S   = 19,
-        GGML_TYPE_IQ4_NL  = 20,
-        GGML_TYPE_IQ3_S   = 21,
         GGML_TYPE_I8,
         GGML_TYPE_I16,
         GGML_TYPE_I32,
@@ -364,9 +368,9 @@ extern "C" {
     };
 
     enum ggml_backend_type {
-        GGML_BACKEND_TYPE_CPU = 0,
-        GGML_BACKEND_TYPE_GPU = 10,
-        GGML_BACKEND_TYPE_GPU_SPLIT = 20,
+        GGML_BACKEND_CPU = 0,
+        GGML_BACKEND_GPU = 10,
+        GGML_BACKEND_GPU_SPLIT = 20,
     };
 
     // model file types
@@ -389,8 +393,6 @@ extern "C" {
         GGML_FTYPE_MOSTLY_IQ2_XS  = 16, // except 1d tensors
         GGML_FTYPE_MOSTLY_IQ3_XXS = 17, // except 1d tensors
         GGML_FTYPE_MOSTLY_IQ1_S   = 18, // except 1d tensors
-        GGML_FTYPE_MOSTLY_IQ4_NL  = 19, // except 1d tensors
-        GGML_FTYPE_MOSTLY_IQ3_S   = 20, // except 1d tensors
     };
 
     // available tensor operations:
@@ -454,6 +456,7 @@ extern "C" {
         GGML_OP_LEAKY_RELU,
 
         GGML_OP_FLASH_ATTN,
+        GGML_OP_FLASH_ATTN_EXT,
         GGML_OP_FLASH_FF,
         GGML_OP_FLASH_ATTN_BACK,
         GGML_OP_WIN_PART,
@@ -498,9 +501,9 @@ extern "C" {
     };
 
     enum ggml_object_type {
-        GGML_OBJECT_TYPE_TENSOR,
-        GGML_OBJECT_TYPE_GRAPH,
-        GGML_OBJECT_TYPE_WORK_BUFFER
+        GGML_OBJECT_TENSOR,
+        GGML_OBJECT_GRAPH,
+        GGML_OBJECT_WORK_BUFFER
     };
 
     enum ggml_log_level {
@@ -642,9 +645,9 @@ extern "C" {
     // NOTE: the INIT or FINALIZE pass is not scheduled unless explicitly enabled.
     // This behavior was changed since https://github.com/ggerganov/llama.cpp/pull/1995.
     enum ggml_task_type {
-        GGML_TASK_TYPE_INIT = 0,
-        GGML_TASK_TYPE_COMPUTE,
-        GGML_TASK_TYPE_FINALIZE,
+        GGML_TASK_INIT = 0,
+        GGML_TASK_COMPUTE,
+        GGML_TASK_FINALIZE,
     };
 
     struct ggml_compute_params {
@@ -1649,8 +1652,8 @@ extern "C" {
 
     // sort rows
     enum ggml_sort_order {
-        GGML_SORT_ORDER_ASC,
-        GGML_SORT_ORDER_DESC,
+        GGML_SORT_ASC,
+        GGML_SORT_DESC,
     };
 
     GGML_API struct ggml_tensor * ggml_argsort(
@@ -1671,6 +1674,25 @@ extern "C" {
             struct ggml_tensor  * v,
             bool                  masked);
 
+#define GGML_KQ_MASK_PAD 32
+
+    // q:    [n_embd, n_batch,     n_head,    1]
+    // k:    [n_embd, n_kv,        n_head_kv, 1]
+    // v:    [n_embd, n_kv,        n_head_kv, 1] !! not transposed !!
+    // mask: [n_kv,   n_batch_pad, 1,         1] !! n_batch_pad = GGML_PAD(n_batch, GGML_KQ_MASK_PAD) !!
+    // res:  [n_embd, n_head,      n_batch,   1] !! permuted !!
+    GGML_API struct ggml_tensor * ggml_flash_attn_ext(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * q,
+            struct ggml_tensor  * k,
+            struct ggml_tensor  * v,
+            struct ggml_tensor  * mask,
+            float                 scale);
+
+    GGML_API void ggml_flash_attn_ext_set_prec(
+            struct ggml_tensor * a,
+            enum ggml_prec       prec);
+
     GGML_API struct ggml_tensor * ggml_flash_attn_back(
            struct ggml_context * ctx,
            struct ggml_tensor  * q,
@@ -1943,8 +1965,8 @@ extern "C" {
 
     // optimization methods
     enum ggml_opt_type {
-        GGML_OPT_TYPE_ADAM,
-        GGML_OPT_TYPE_LBFGS,
+        GGML_OPT_ADAM,
+        GGML_OPT_LBFGS,
     };
 
     // linesearch methods
@@ -1958,12 +1980,12 @@ extern "C" {
 
     // optimization return values
     enum ggml_opt_result {
-        GGML_OPT_RESULT_OK = 0,
-        GGML_OPT_RESULT_DID_NOT_CONVERGE,
-        GGML_OPT_RESULT_NO_CONTEXT,
-        GGML_OPT_RESULT_INVALID_WOLFE,
-        GGML_OPT_RESULT_FAIL,
-        GGML_OPT_RESULT_CANCEL,
+        GGML_OPT_OK = 0,
+        GGML_OPT_DID_NOT_CONVERGE,
+        GGML_OPT_NO_CONTEXT,
+        GGML_OPT_INVALID_WOLFE,
+        GGML_OPT_FAIL,
+        GGML_OPT_CANCEL,
 
         GGML_LINESEARCH_FAIL = -128,
         GGML_LINESEARCH_MINIMUM_STEP,

gguf-py/gguf/constants.py

@@ -111,7 +111,6 @@ class MODEL_ARCH(IntEnum):
     ORION      = auto()
     INTERNLM2  = auto()
     MINICPM    = auto()
-    GEMMA      = auto()
 
 
 class MODEL_TENSOR(IntEnum):
@@ -168,7 +167,6 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
     MODEL_ARCH.ORION:          "orion",
     MODEL_ARCH.INTERNLM2:      "internlm2",
     MODEL_ARCH.MINICPM:        "minicpm",
-    MODEL_ARCH.GEMMA:          "gemma",
 }
 
 TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
@@ -513,19 +511,6 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.FFN_DOWN_EXP,
         MODEL_TENSOR.FFN_UP_EXP,
     ],
-    MODEL_ARCH.GEMMA: [
-        MODEL_TENSOR.TOKEN_EMBD,
-        MODEL_TENSOR.OUTPUT_NORM,
-        MODEL_TENSOR.ATTN_NORM,
-        MODEL_TENSOR.ATTN_Q,
-        MODEL_TENSOR.ATTN_K,
-        MODEL_TENSOR.ATTN_V,
-        MODEL_TENSOR.ATTN_OUT,
-        MODEL_TENSOR.FFN_GATE,
-        MODEL_TENSOR.FFN_DOWN,
-        MODEL_TENSOR.FFN_UP,
-        MODEL_TENSOR.FFN_NORM,
-    ],
     # TODO
 }
 

llama.h

@@ -64,15 +64,6 @@ extern "C" {
         LLAMA_VOCAB_TYPE_WPM = 2, // WordPiece
     };
 
-    // note: these values should be synchronized with ggml_rope
-    // TODO: maybe move this enum to ggml.h (ggml_rope_type)
-    enum llama_rope_type {
-        LLAMA_ROPE_TYPE_NONE = -1,
-        LLAMA_ROPE_TYPE_NORM =  0,
-        LLAMA_ROPE_TYPE_NEOX =  2,
-        LLAMA_ROPE_TYPE_GLM  =  4,
-    };
-
     enum llama_token_type {
         LLAMA_TOKEN_TYPE_UNDEFINED    = 0,
         LLAMA_TOKEN_TYPE_NORMAL       = 1,
@@ -110,31 +101,28 @@ extern "C" {
         LLAMA_FTYPE_MOSTLY_Q3_K_XS       = 22, // except 1d tensors
         LLAMA_FTYPE_MOSTLY_IQ3_XXS       = 23, // except 1d tensors
         LLAMA_FTYPE_MOSTLY_IQ1_S         = 24, // except 1d tensors
-        LLAMA_FTYPE_MOSTLY_IQ4_NL        = 25, // except 1d tensors
-        LLAMA_FTYPE_MOSTLY_IQ3_S         = 26, // except 1d tensors
-        LLAMA_FTYPE_MOSTLY_IQ3_M         = 27, // except 1d tensors
 
         LLAMA_FTYPE_GUESSED = 1024, // not specified in the model file
     };
 
     enum llama_rope_scaling_type {
-        LLAMA_ROPE_SCALING_TYPE_UNSPECIFIED = -1,
-        LLAMA_ROPE_SCALING_TYPE_NONE        = 0,
-        LLAMA_ROPE_SCALING_TYPE_LINEAR      = 1,
-        LLAMA_ROPE_SCALING_TYPE_YARN        = 2,
-        LLAMA_ROPE_SCALING_TYPE_MAX_VALUE   = LLAMA_ROPE_SCALING_TYPE_YARN,
+        LLAMA_ROPE_SCALING_UNSPECIFIED = -1,
+        LLAMA_ROPE_SCALING_NONE        = 0,
+        LLAMA_ROPE_SCALING_LINEAR      = 1,
+        LLAMA_ROPE_SCALING_YARN        = 2,
+        LLAMA_ROPE_SCALING_MAX_VALUE   = LLAMA_ROPE_SCALING_YARN,
     };
 
     enum llama_pooling_type {
-        LLAMA_POOLING_TYPE_NONE = 0,
-        LLAMA_POOLING_TYPE_MEAN = 1,
-        LLAMA_POOLING_TYPE_CLS  = 2,
+        LLAMA_POOLING_NONE = 0,
+        LLAMA_POOLING_MEAN = 1,
+        LLAMA_POOLING_CLS  = 2,
     };
 
     enum llama_split_mode {
-        LLAMA_SPLIT_MODE_NONE    = 0, // single GPU
-        LLAMA_SPLIT_MODE_LAYER   = 1, // split layers and KV across GPUs
-        LLAMA_SPLIT_MODE_ROW     = 2, // split rows across GPUs
+        LLAMA_SPLIT_NONE    = 0, // single GPU
+        LLAMA_SPLIT_LAYER   = 1, // split layers and KV across GPUs
+        LLAMA_SPLIT_ROW     = 2, // split rows across GPUs
     };
 
     typedef struct llama_token_data {
@@ -182,9 +170,9 @@ extern "C" {
     } llama_batch;
 
     enum llama_model_kv_override_type {
-        LLAMA_KV_OVERRIDE_TYPE_INT,
-        LLAMA_KV_OVERRIDE_TYPE_FLOAT,
-        LLAMA_KV_OVERRIDE_TYPE_BOOL,
+        LLAMA_KV_OVERRIDE_INT,
+        LLAMA_KV_OVERRIDE_FLOAT,
+        LLAMA_KV_OVERRIDE_BOOL,
     };
 
     struct llama_model_kv_override {
@@ -369,7 +357,6 @@ extern "C" {
     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);
-    LLAMA_API enum llama_rope_type  llama_rope_type (const struct llama_model * model);
 
     LLAMA_API int32_t llama_n_vocab    (const struct llama_model * model);
     LLAMA_API int32_t llama_n_ctx_train(const struct llama_model * model);
@@ -524,12 +511,10 @@ extern "C" {
                     llama_seq_id   seq_id);
 
     // Adds relative position "delta" to all tokens that belong to the specified sequence and have positions in [p0, p1)
-    // If the KV cache is RoPEd, the KV data is updated accordingly:
-    //   - lazily on next llama_decode()
-    //   - explicitly with llama_kv_cache_update()
+    // If the KV cache is RoPEd, the KV data is updated accordingly
     // p0 < 0 : [0,  p1]
     // p1 < 0 : [p0, inf)
-    LLAMA_API void llama_kv_cache_seq_add(
+    LLAMA_API void llama_kv_cache_seq_shift(
             struct llama_context * ctx,
                     llama_seq_id   seq_id,
                        llama_pos   p0,
@@ -537,9 +522,7 @@ extern "C" {
                        llama_pos   delta);
 
     // Integer division of the positions by factor of `d > 1`
-    // If the KV cache is RoPEd, the KV data is updated accordingly:
-    //   - lazily on next llama_decode()
-    //   - explicitly with llama_kv_cache_update()
+    // If the KV cache is RoPEd, the KV data is updated accordingly
     // p0 < 0 : [0,  p1]
     // p1 < 0 : [p0, inf)
     LLAMA_API void llama_kv_cache_seq_div(
@@ -549,20 +532,6 @@ extern "C" {
                        llama_pos   p1,
                              int   d);
 
-    // Returns the largest position present in the KV cache for the specified sequence
-    LLAMA_API llama_pos llama_kv_cache_seq_pos_max(
-            struct llama_context * ctx,
-                    llama_seq_id   seq_id);
-
-    // Defragment the KV cache
-    // This will be applied:
-    //   - lazily on next llama_decode()
-    //   - explicitly with llama_kv_cache_update()
-    LLAMA_API void llama_kv_cache_defrag(struct llama_context * ctx);
-
-    // Apply the KV cache updates (such as K-shifts, defragmentation, etc.)
-    LLAMA_API void llama_kv_cache_update(struct llama_context * ctx);
-
     //
     // State / sessions
     //
@@ -738,7 +707,7 @@ extern "C" {
 
     /// Apply chat template. Inspired by hf apply_chat_template() on python.
     /// Both "model" and "custom_template" are optional, but at least one is required. "custom_template" has higher precedence than "model"
-    /// NOTE: This function does not use a jinja parser. It only support a pre-defined list of template. See more: https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template
+    /// NOTE: This function only support some known jinja templates. It is not a jinja parser.
     /// @param tmpl A Jinja template to use for this chat. If this is nullptr, the model’s default chat template will be used instead.
     /// @param chat Pointer to a list of multiple llama_chat_message
     /// @param n_msg Number of llama_chat_message in this chat

scripts/sync-ggml.last

@@ -1 +1 @@
-8cdf783f288a98eddf521b0ab1b4d405be9e18ba
+5070f078a67c18c11736e78316ab715ca9afde16

tests/test-backend-ops.cpp

@@ -572,9 +572,19 @@ struct test_case {
         // 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;
+#if 0
         for (int i = 1; i < n_runs; i++) {
             gf->nodes[gf->n_nodes++] = out;
         }
+#else
+        int n_nodes = gf->n_nodes;
+        n_runs = 1000;
+        for (int i = 1; i < n_runs; i++) {
+            for (int j = 0; j < n_nodes; j++) {
+                gf->nodes[gf->n_nodes++] = gf->nodes[j];
+            }
+        }
+#endif
 
         // calculate memory
         size_t mem = n_runs * op_size(out);
@@ -1264,7 +1274,7 @@ struct test_argsort : public test_case {
 
     test_argsort(ggml_type type = GGML_TYPE_F32,
             std::array<int64_t, 4> ne = {16, 10, 10, 10},
-            ggml_sort_order order = GGML_SORT_ORDER_ASC)
+            ggml_sort_order order = GGML_SORT_ASC)
         : type(type), ne(ne), order(order) {}
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
@@ -1434,6 +1444,76 @@ struct test_leaky_relu : public test_case {
     }
 };
 
+// GGML_OP_FLASH_ATTN_EXT
+struct test_flash_attn_ext : public test_case {
+    const int64_t hs; // head size
+    const int64_t nh; // num heads
+    const int64_t kv; // kv size
+    const int64_t nb; // batch size
+
+    std::string vars() override {
+        return VARS_TO_STR4(hs, nh, kv, nb);
+    }
+
+    double max_nmse_err() override {
+        return 5e-4;
+    }
+
+    test_flash_attn_ext(int64_t hs = 128, int64_t nh = 32, int64_t kv = 96, int64_t nb = 8)
+        : hs(hs), nh(nh), kv(kv), nb(nb) {}
+
+    ggml_tensor * build_graph(ggml_context * ctx) override {
+        ggml_tensor * q = ggml_new_tensor_4d(ctx, GGML_TYPE_F32, hs, nb, nh, 1);
+        ggml_tensor * k = ggml_new_tensor_4d(ctx, GGML_TYPE_F16, hs, kv, nh, 1);
+        ggml_tensor * v = ggml_new_tensor_4d(ctx, GGML_TYPE_F16, hs, kv, nh, 1);
+        ggml_tensor * mask = ggml_new_tensor_4d(ctx, GGML_TYPE_F16, kv, GGML_PAD(nb, GGML_KQ_MASK_PAD), 1, 1);
+        ggml_tensor * out = ggml_flash_attn_ext(ctx, q, k, v, mask, 1.0f/sqrtf(hs));
+        return out;
+    }
+};
+
+// Attention
+struct test_attn : public test_case {
+    const int64_t hs; // head size
+    const int64_t nh; // num heads
+    const int64_t kv; // kv size
+    const int64_t nb; // batch size
+
+    std::string op_desc(ggml_tensor * t) override {
+        return "ATTN";
+
+        GGML_UNUSED(t);
+    }
+
+    std::string vars() override {
+        return VARS_TO_STR4(hs, nh, kv, nb);
+    }
+
+    double max_nmse_err() override {
+        return 5e-4;
+    }
+
+    test_attn(int64_t hs = 128, int64_t nh = 32, int64_t kv = 96, int64_t nb = 8)
+        : hs(hs), nh(nh), kv(kv), nb(nb) {}
+
+    ggml_tensor * build_graph(ggml_context * ctx) override {
+        ggml_tensor * q = ggml_new_tensor_4d(ctx, GGML_TYPE_F32, hs, nb, nh, 1);
+        ggml_tensor * k = ggml_new_tensor_4d(ctx, GGML_TYPE_F16, hs, kv, nh, 1);
+        ggml_tensor * v = ggml_new_tensor_4d(ctx, GGML_TYPE_F16, kv, hs, nh, 1); // transposed
+        ggml_tensor * mask = ggml_new_tensor_4d(ctx, GGML_TYPE_F16, kv, nb, 1, 1);
+
+        struct ggml_tensor * cur;
+
+        cur = ggml_mul_mat     (ctx, k, q);
+        cur = ggml_soft_max_ext(ctx, cur, mask, nullptr, 1.0f/sqrtf(hs), 0.0f);
+        cur = ggml_mul_mat     (ctx, v, cur);
+        cur = ggml_permute     (ctx, cur, 0, 2, 1, 3);
+        cur = ggml_cont_2d     (ctx, cur, hs*nh, nb);
+
+        return cur;
+    }
+};
+
 // Mixtral MOE
 struct test_moe : public test_case {
     const int n_experts;
@@ -1706,7 +1786,7 @@ struct test_llama : public test_llm {
         struct ggml_tensor * inp_pos = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, hp.n_tokens);
 
         // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
-        struct ggml_tensor * KQ_mask = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, hp.n_kv, hp.n_tokens, 1);
+        struct ggml_tensor * KQ_mask = ggml_new_tensor_3d(ctx, GGML_TYPE_F16, hp.n_kv, hp.n_tokens, 1);
 
         ggml_tensor * k_l = ggml_new_tensor_1d(ctx, GGML_TYPE_F16, 1638400);
         ggml_tensor * v_l = ggml_new_tensor_1d(ctx, GGML_TYPE_F16, 1638400);
@@ -1828,7 +1908,7 @@ struct test_falcon : public test_llm {
         struct ggml_tensor * inp_pos = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, hp.n_tokens);
 
         // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
-        struct ggml_tensor * KQ_mask = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, hp.n_kv, hp.n_tokens, 1);
+        struct ggml_tensor * KQ_mask = ggml_new_tensor_3d(ctx, GGML_TYPE_F16, hp.n_kv, hp.n_tokens, 1);
 
         ggml_tensor * k_l = ggml_new_tensor_1d(ctx, GGML_TYPE_F16, 1638400);
         ggml_tensor * v_l = ggml_new_tensor_1d(ctx, GGML_TYPE_F16, 1638400);
@@ -1918,7 +1998,6 @@ static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op
         GGML_TYPE_Q6_K,
         GGML_TYPE_IQ2_XXS, GGML_TYPE_IQ2_XS,
         GGML_TYPE_IQ3_XXS, GGML_TYPE_IQ1_S,
-        GGML_TYPE_IQ4_NL, GGML_TYPE_IQ3_S,
     };
 
     // unary ops
@@ -2116,7 +2195,7 @@ static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op
     test_cases.emplace_back(new test_concat(GGML_TYPE_F32));
     test_cases.emplace_back(new test_concat(GGML_TYPE_I32));
 
-    for (ggml_sort_order order : {GGML_SORT_ORDER_ASC, GGML_SORT_ORDER_DESC}) {
+    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));
     }
@@ -2128,6 +2207,30 @@ static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op
     test_cases.emplace_back(new test_pad());
     test_cases.emplace_back(new test_leaky_relu());
 
+#if 1
+    for (int hs : { 128, 64, 80, }) {
+        for (int nh : { 32, }) {
+            for (int kv : { 512, 1024, 2048, 4096, }) {
+                for (int nb : { 1, 2, 4, 8, 512, 1024, 2048, }) {
+                    test_cases.emplace_back(new test_attn          (hs, nh, kv, nb));
+                    test_cases.emplace_back(new test_flash_attn_ext(hs, nh, kv, nb));
+                }
+            }
+        }
+    }
+#else
+    for (int hs : { 128, }) {
+        for (int nh : { 32, }) {
+            for (int kv : { 512, 1024, }) {
+                for (int nb : { 1, 2, 4, 8, 512 }) {
+                    test_cases.emplace_back(new test_attn          (hs, nh, kv, nb));
+                    test_cases.emplace_back(new test_flash_attn_ext(hs, nh, kv, nb));
+                }
+            }
+        }
+    }
+#endif
+
     // these tests are disabled to save execution time, but they can be handy for debugging
 #if 0
 #if !defined(__SANITIZE_THREAD__)

tests/test-chat-template.cpp

@@ -27,24 +27,12 @@ int main(void) {
         "{%- for idx in range(0, messages|length) -%}\\n{%- if messages[idx]['role'] == 'user' -%}\\n{%- if idx > 1 -%}\\n{{- bos_token + '[INST] ' + messages[idx]['content'] + ' [/INST]' -}}\\n{%- else -%}\\n{{- messages[idx]['content'] + ' [/INST]' -}}\\n{%- endif -%}\\n{% elif messages[idx]['role'] == 'system' %}\\n{{- '[INST] <<SYS>>\\\\n' + messages[idx]['content'] + '\\\\n<</SYS>>\\\\n\\\\n' -}}\\n{%- elif messages[idx]['role'] == 'assistant' -%}\\n{{- ' '  + messages[idx]['content'] + ' ' + eos_token -}}\\n{% endif %}\\n{% endfor %}",
         // bofenghuang/vigogne-2-70b-chat
         "{{ bos_token }}{% if messages[0]['role'] == 'system' %}{% set loop_messages = messages[1:] %}{% set system_message = messages[0]['content'] %}{% elif true == true and not '<<SYS>>' in messages[0]['content'] %}{% set loop_messages = messages %}{% set system_message = 'Vous êtes Vigogne, un assistant IA créé par Zaion Lab. Vous suivez extrêmement bien les instructions. Aidez autant que vous le pouvez.' %}{% else %}{% set loop_messages = messages %}{% set system_message = false %}{% endif %}{% for message in loop_messages %}{% if (message['role'] == 'user') != (loop.index0 % 2 == 0) %}{{ raise_exception('Conversation roles must alternate user/assistant/user/assistant/...') }}{% endif %}{% if loop.index0 == 0 and system_message != false %}{% set content = '<<SYS>>\\\\n' + system_message + '\\\\n<</SYS>>\\\\n\\\\n' + message['content'] %}{% else %}{% set content = message['content'] %}{% endif %}{% if message['role'] == 'user' %}{{ '[INST] ' + content.strip() + ' [/INST]' }}{% elif message['role'] == 'system' %}{{ '<<SYS>>\\\\n' + content.strip() + '\\\\n<</SYS>>\\\\n\\\\n' }}{% elif message['role'] == 'assistant' %}{{ ' '  + content.strip() + ' ' + eos_token }}{% endif %}{% endfor %}",
-        // mlabonne/AlphaMonarch-7B
-        "{% for message in messages %}{{bos_token + message['role'] + '\\n' + message['content'] + eos_token + '\\n'}}{% endfor %}{% if add_generation_prompt %}{{ bos_token + 'assistant\\n' }}{% endif %}",
-        // google/gemma-7b-it
-        "{% if messages[0]['role'] == 'system' %}{{ raise_exception('System role not supported') }}{% endif %}{% for message in messages %}{% if (message['role'] == 'user') != (loop.index0 % 2 == 0) %}{{ raise_exception('Conversation roles must alternate user/assistant/user/assistant/...') }}{% endif %}{% if (message['role'] == 'assistant') %}{% set role = 'model' %}{% else %}{% set role = message['role'] %}{% endif %}{{ '<start_of_turn>' + role + '\\n' + message['content'] | trim + '<end_of_turn>\\n' }}{% endfor %}{% if add_generation_prompt %}{{'<start_of_turn>model\\n'}}{% endif %}",
     };
-    std::vector<std::string> expected_output = {
-        // teknium/OpenHermes-2.5-Mistral-7B
-        "<|im_start|>system\nYou are a helpful assistant<|im_end|>\n<|im_start|>user\nHello<|im_end|>\n<|im_start|>assistant\nHi there<|im_end|>\n<|im_start|>user\nWho are you<|im_end|>\n<|im_start|>assistant\n   I am an assistant   <|im_end|>\n<|im_start|>user\nAnother question<|im_end|>\n<|im_start|>assistant\n",
-        // mistralai/Mistral-7B-Instruct-v0.2
-        "[INST] You are a helpful assistant\nHello [/INST]Hi there</s>[INST] Who are you [/INST]   I am an assistant   </s>[INST] Another question [/INST]",
-        // TheBloke/FusionNet_34Bx2_MoE-AWQ
-        "[INST] <<SYS>>\nYou are a helpful assistant\n<</SYS>>\n\nHello [/INST] Hi there </s><s>[INST] Who are you [/INST]    I am an assistant    </s><s>[INST] Another question [/INST]",
-        // bofenghuang/vigogne-2-70b-chat
-        "[INST] <<SYS>>\nYou are a helpful assistant\n<</SYS>>\n\nHello [/INST] Hi there </s>[INST] Who are you [/INST] I am an assistant </s>[INST] Another question [/INST]",
-        // mlabonne/AlphaMonarch-7B
-        "system\nYou are a helpful assistant</s>\n<s>user\nHello</s>\n<s>assistant\nHi there</s>\n<s>user\nWho are you</s>\n<s>assistant\n   I am an assistant   </s>\n<s>user\nAnother question</s>\n<s>assistant\n",
-        // google/gemma-7b-it
-        "<start_of_turn>user\nYou are a helpful assistant\n\nHello<end_of_turn>\n<start_of_turn>model\nHi there<end_of_turn>\n<start_of_turn>user\nWho are you<end_of_turn>\n<start_of_turn>model\nI am an assistant<end_of_turn>\n<start_of_turn>user\nAnother question<end_of_turn>\n<start_of_turn>model\n",
+    std::vector<std::string> expected_substr = {
+        "<|im_start|>assistant\n   I am an assistant   <|im_end|>\n<|im_start|>user\nAnother question<|im_end|>\n<|im_start|>assistant",
+        "[/INST]Hi there</s>[INST] Who are you [/INST]   I am an assistant   </s>[INST] Another question [/INST]",
+        "</s><s>[INST] Who are you [/INST]    I am an assistant    </s><s>[INST] Another question [/INST]",
+        "[/INST] Hi there </s>[INST] Who are you [/INST] I am an assistant </s>[INST] Another question [/INST]",
     };
     std::vector<char> formatted_chat(1024);
     int32_t res;
@@ -55,7 +43,7 @@ int main(void) {
 
     for (size_t i = 0; i < templates.size(); i++) {
         std::string custom_template = templates[i];
-        std::string expected = expected_output[i];
+        std::string substr = expected_substr[i];
         formatted_chat.resize(1024);
         res = llama_chat_apply_template(
             nullptr,
@@ -69,7 +57,8 @@ int main(void) {
         formatted_chat.resize(res);
         std::string output(formatted_chat.data(), formatted_chat.size());
         std::cout << output << "\n-------------------------\n";
-        assert(output == expected);
+        // expect the "formatted_chat" to contain pre-defined strings
+        assert(output.find(substr) != std::string::npos);
     }
     return 0;
 }

tests/test-opt.cpp

@@ -118,7 +118,7 @@ int main(void) {
     const float fe = ggml_get_f32_1d(e, 0);
     printf("%s: e = %.4f\n", __func__, fe);
 
-    struct ggml_opt_params opt_params = ggml_opt_default_params(GGML_OPT_TYPE_ADAM);
+    struct ggml_opt_params opt_params = ggml_opt_default_params(GGML_OPT_ADAM);
 
     ggml_opt(ctx, opt_params, e);
 

tests/test-quantize-fns.cpp

@@ -151,7 +151,6 @@ int main(int argc, char * argv[]) {
             const float max_quantization_error =
                 type == GGML_TYPE_Q2_K    ? MAX_QUANTIZATION_TOTAL_ERROR_2BITS :
                 type == GGML_TYPE_Q3_K    ? MAX_QUANTIZATION_TOTAL_ERROR_3BITS :
-                type == GGML_TYPE_IQ3_S   ? MAX_QUANTIZATION_TOTAL_ERROR_3BITS :
                 type == GGML_TYPE_IQ3_XXS ? MAX_QUANTIZATION_TOTAL_ERROR_3BITS_XXS : MAX_QUANTIZATION_TOTAL_ERROR;
             failed = !(total_error < max_quantization_error);
             num_failed += failed;
@@ -168,8 +167,7 @@ int main(int argc, char * argv[]) {
 
             const float vec_dot_error = dot_product_error(qfns, test_size, test_data.data(), test_data2.data());
             const float max_allowed_error = type == GGML_TYPE_Q2_K || type == GGML_TYPE_IQ2_XS || type == GGML_TYPE_IQ2_XXS ||
-                                            type == GGML_TYPE_IQ3_XXS || type == GGML_TYPE_IQ3_S ? MAX_DOT_PRODUCT_ERROR_LOWBIT
-                                          : MAX_DOT_PRODUCT_ERROR;
+                                            type == GGML_TYPE_IQ3_XXS ? MAX_DOT_PRODUCT_ERROR_LOWBIT : MAX_DOT_PRODUCT_ERROR;
             failed = !(vec_dot_error < max_allowed_error);
             num_failed += failed;
             if (failed || verbose) {