minor : spacing

* Add LLMFarm (ui for iOS) to list

.devops/nix/docker.nix

@@ -0,0 +1,37 @@
+{
+  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 || useVulkan) lib.platforms.darwin;
+      badPlatforms = optionals (useCuda || useOpenCL) 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) || (useVulkan && effectiveStdenv.isDarwin);
+      broken = (useMetalKit && !effectiveStdenv.isDarwin);
 
       description = "Inference of LLaMA model in pure C/C++${descriptionSuffix}";
       homepage = "https://github.com/ggerganov/llama.cpp/";

.devops/nix/scope.nix

@@ -12,5 +12,8 @@ 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

@@ -0,0 +1,27 @@
+{
+  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

@@ -7,3 +7,5 @@ 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

@@ -37,6 +37,8 @@ jobs:
 
       - name: Build
         id: make_build
+        env:
+            LLAMA_FATAL_WARNINGS: 1
         run: |
           CC=gcc-8 make -j $(nproc)
 
@@ -65,7 +67,7 @@ jobs:
         run: |
           mkdir build
           cd build
-          cmake ..
+          cmake .. -DLLAMA_FATAL_WARNINGS=ON
           cmake --build . --config Release -j $(nproc)
 
       - name: Test
@@ -74,6 +76,17 @@ jobs:
           cd build
           ctest -L main --verbose --timeout 900
 
+      - name: Test llama2c conversion
+        id: llama2c_test
+        run: |
+          cd build
+          echo "Fetch tokenizer"
+          wget https://huggingface.co/karpathy/tinyllamas/resolve/main/stories260K/tok512.bin
+          echo "Fetch llama2c model"
+          wget https://huggingface.co/karpathy/tinyllamas/resolve/main/stories260K/stories260K.bin
+          ./bin/convert-llama2c-to-ggml --copy-vocab-from-model ./tok512.bin --llama2c-model stories260K.bin --llama2c-output-model stories260K.gguf
+          ./bin/main -m stories260K.gguf -p "One day, Lily met a Shoggoth" -n 500 -c 256
+
   ubuntu-latest-cmake-sanitizer:
     runs-on: ubuntu-latest
 
@@ -100,7 +113,7 @@ jobs:
         run: |
           mkdir build
           cd build
-          cmake .. -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON -DCMAKE_BUILD_TYPE=${{ matrix.build_type }}
+          cmake .. -DLLAMA_FATAL_WARNINGS=ON -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON -DCMAKE_BUILD_TYPE=${{ matrix.build_type }}
           cmake --build . --config ${{ matrix.build_type }} -j $(nproc)
 
       - name: Test
@@ -244,6 +257,8 @@ jobs:
 
       - name: Build
         id: make_build
+        env:
+            LLAMA_FATAL_WARNINGS: 1
         run: |
           LLAMA_NO_METAL=1 make -j $(sysctl -n hw.logicalcpu)
 
@@ -277,7 +292,7 @@ jobs:
           sysctl -a
           mkdir build
           cd build
-          cmake -DLLAMA_METAL=OFF ..
+          cmake -DLLAMA_FATAL_WARNINGS=ON -DLLAMA_METAL=OFF ..
           cmake --build . --config Release -j $(sysctl -n hw.logicalcpu)
 
       - name: Test
@@ -665,8 +680,7 @@ jobs:
         run: |
           cd examples/llama.android
 
-          # Skip armeabi-v7a for now (https://github.com/llvm/llvm-project/issues/65820).
-          ./gradlew build --no-daemon -Pskip-armeabi-v7a
+          ./gradlew build --no-daemon
 
 #  freeBSD-latest:
 #    runs-on: macos-12

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

@@ -19,7 +19,6 @@ on:
 
 jobs:
   nix-build-aarch64:
-    if: ${{ vars.CACHIX_NAME != '' }}
     runs-on: ubuntu-latest
     steps:
     - name: Checkout repository
@@ -37,8 +36,8 @@ jobs:
         extra-conf: |
           extra-platforms = aarch64-linux
           extra-system-features = nixos-test kvm
-          extra-substituters = https://${{ vars.CACHIX_NAME }}.cachix.org https://cuda-maintainers.cachix.org
-          extra-trusted-public-keys = ${{ vars.CACHIX_PUBLIC_KEY }} cuda-maintainers.cachix.org-1:0dq3bujKpuEPMCX6U4WylrUDZ9JyUG0VpVZa7CNfq5E=
+          extra-substituters = https://llama-cpp.cachix.org https://cuda-maintainers.cachix.org
+          extra-trusted-public-keys = llama-cpp.cachix.org-1:H75X+w83wUKTIPSO1KWy9ADUrzThyGs8P5tmAbkWhQc= cuda-maintainers.cachix.org-1:0dq3bujKpuEPMCX6U4WylrUDZ9JyUG0VpVZa7CNfq5E=
     - uses: DeterminateSystems/magic-nix-cache-action@v2
       with:
         upstream-cache: https://${{ matrix.cachixName }}.cachix.org
@@ -46,7 +45,7 @@ jobs:
       uses: cachix/cachix-action@v13
       with:
         authToken: '${{ secrets.CACHIX_AUTH_TOKEN }}'
-        name: ${{ vars.CACHIX_NAME }}
+        name: llama-cpp
     - 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: ${{ secrets.GITHUB_TOKEN }}
         extra-conf: |
-          extra-substituters = https://${{ vars.CACHIX_NAME }}.cachix.org https://cuda-maintainers.cachix.org
-          extra-trusted-public-keys = ${{ vars.CACHIX_PUBLIC_KEY }} cuda-maintainers.cachix.org-1:0dq3bujKpuEPMCX6U4WylrUDZ9JyUG0VpVZa7CNfq5E=
+          extra-substituters = https://llama-cpp.cachix.org https://cuda-maintainers.cachix.org
+          extra-trusted-public-keys = llama-cpp.cachix.org-1:H75X+w83wUKTIPSO1KWy9ADUrzThyGs8P5tmAbkWhQc= cuda-maintainers.cachix.org-1:0dq3bujKpuEPMCX6U4WylrUDZ9JyUG0VpVZa7CNfq5E=
     - uses: DeterminateSystems/magic-nix-cache-action@v2
       with:
         upstream-cache: https://${{ matrix.cachixName }}.cachix.org
@@ -37,7 +37,6 @@ jobs:
           --flake
           ".#packages.$(nix eval --raw --impure --expr builtins.currentSystem)"
   nix-build:
-    if: ${{ vars.CACHIX_NAME != '' }}
     strategy:
       fail-fast: false
       matrix:
@@ -51,8 +50,8 @@ jobs:
       with:
         github-token: ${{ secrets.GITHUB_TOKEN }}
         extra-conf: |
-          extra-substituters = https://${{ vars.CACHIX_NAME }}.cachix.org https://cuda-maintainers.cachix.org
-          extra-trusted-public-keys = ${{ vars.CACHIX_PUBLIC_KEY }} cuda-maintainers.cachix.org-1:0dq3bujKpuEPMCX6U4WylrUDZ9JyUG0VpVZa7CNfq5E=
+          extra-substituters = https://llama-cpp.cachix.org https://cuda-maintainers.cachix.org
+          extra-trusted-public-keys = llama-cpp.cachix.org-1:H75X+w83wUKTIPSO1KWy9ADUrzThyGs8P5tmAbkWhQc= cuda-maintainers.cachix.org-1:0dq3bujKpuEPMCX6U4WylrUDZ9JyUG0VpVZa7CNfq5E=
     - uses: DeterminateSystems/magic-nix-cache-action@v2
       with:
         upstream-cache: https://${{ matrix.cachixName }}.cachix.org
@@ -60,7 +59,7 @@ jobs:
       uses: cachix/cachix-action@v13
       with:
         authToken: '${{ secrets.CACHIX_AUTH_TOKEN }}'
-        name: ${{ vars.CACHIX_NAME }}
+        name: llama-cpp
     - name: Build
       run: >
           nix run github:Mic92/nix-fast-build

.github/workflows/server.yml

@@ -0,0 +1,83 @@
+# Server build and tests
+name: Server
+
+on:
+  workflow_dispatch: # allows manual triggering
+  push:
+    branches:
+      - master
+    paths: ['.github/workflows/server.yml', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.swift', '**/*.m', 'examples/server/tests/**.*']
+  pull_request:
+    types: [opened, synchronize, reopened]
+    paths: ['.github/workflows/server.yml', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.swift', '**/*.m', 'examples/server/tests/**.*']
+
+jobs:
+  server:
+    runs-on: ubuntu-latest
+
+    strategy:
+      matrix:
+        sanitizer: [ADDRESS, THREAD, UNDEFINED]
+        build_type: [Debug, Release]
+        include:
+          - build_type: Release
+            sanitizer: ""
+        exclude:
+          - build_type: Release
+            sanitizer: ADDRESS
+          - build_type: Release
+            sanitizer: THREAD
+          - build_type: Release
+            sanitizer: UNDEFINED
+
+    container:
+      image: ubuntu:latest
+      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 \
+            git \
+            cmake \
+            python3-pip \
+            wget \
+            psmisc
+
+      - name: Build
+        id: cmake_build
+        run: |
+          mkdir build
+          cd build
+          cmake .. \
+              -DLLAMA_NATIVE=OFF \
+              -DLLAMA_BUILD_SERVER=ON \
+              -DCMAKE_BUILD_TYPE=${{ matrix.build_type }} \
+              -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON ;
+          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
+        run: |
+          cd examples/server/tests
+          PORT=8888 ./tests.sh

CMakeLists.txt

@@ -55,6 +55,9 @@ option(LLAMA_ALL_WARNINGS               "llama: enable all compiler warnings"
 option(LLAMA_ALL_WARNINGS_3RD_PARTY     "llama: enable all compiler warnings in 3rd party libs" OFF)
 option(LLAMA_GPROF                      "llama: enable gprof"                                   OFF)
 
+# build
+option(LLAMA_FATAL_WARNINGS             "llama: enable -Werror flag"                            OFF)
+
 # sanitizers
 option(LLAMA_SANITIZE_THREAD            "llama: enable thread sanitizer"                        OFF)
 option(LLAMA_SANITIZE_ADDRESS           "llama: enable address sanitizer"                       OFF)
@@ -107,6 +110,7 @@ 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)
@@ -198,6 +202,29 @@ 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
@@ -515,7 +542,7 @@ if (LLAMA_SYCL)
 
     message(STATUS "SYCL found")
 
-    add_compile_definitions(GML_USE_SYCL)
+    add_compile_definitions(GGML_USE_SYCL)
 
     if (LLAMA_SYCL_F16)
         add_compile_definitions(GGML_SYCL_F16)
@@ -730,28 +757,30 @@ function(get_flags CCID CCVER)
         if (CCVER VERSION_GREATER_EQUAL 8.1.0)
             list(APPEND CXX_FLAGS -Wextra-semi)
         endif()
-    elseif (CCID MATCHES "Intel")
-        if (NOT LLAMA_SYCL)
-            # enable max optimization level when using Intel compiler
-            set(C_FLAGS   -ipo -O3 -static -fp-model=fast -flto -fno-stack-protector)
-            set(CXX_FLAGS -ipo -O3 -static -fp-model=fast -flto -fno-stack-protector)
-            add_link_options(-fuse-ld=lld -static-intel)
-        endif()
     endif()
 
     set(GF_C_FLAGS   ${C_FLAGS}   PARENT_SCOPE)
     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)
-        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 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(C_FLAGS   ${WARNING_FLAGS} ${C_FLAGS})
-        set(CXX_FLAGS ${WARNING_FLAGS} ${CXX_FLAGS})
+        list(APPEND C_FLAGS   ${WARNING_FLAGS})
+        list(APPEND CXX_FLAGS ${WARNING_FLAGS})
 
         get_flags(${CMAKE_CXX_COMPILER_ID} ${CMAKE_CXX_COMPILER_VERSION})
 
@@ -767,9 +796,10 @@ endif()
 set(CUDA_CXX_FLAGS "")
 
 if (LLAMA_CUBLAS)
-    set(CUDA_FLAGS ${CXX_FLAGS} -use_fast_math)
-    if (NOT MSVC)
-        list(APPEND CUDA_FLAGS -Wno-pedantic)
+    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)
@@ -803,7 +833,11 @@ if (LLAMA_CUBLAS)
         message("-- CUDA host compiler is ${CUDA_CCID} ${CUDA_CCVER}")
 
         get_flags(${CUDA_CCID} ${CUDA_CCVER})
-        list(APPEND CUDA_CXX_FLAGS ${GF_CXX_FLAGS})  # This is passed to -Xcompiler later
+        list(APPEND CUDA_CXX_FLAGS ${CXX_FLAGS} ${GF_CXX_FLAGS})  # This is passed to -Xcompiler later
+    endif()
+
+    if (NOT MSVC)
+        list(APPEND CUDA_CXX_FLAGS -Wno-pedantic)
     endif()
 endif()
 
@@ -902,10 +936,16 @@ 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 (${CMAKE_SYSTEM_PROCESSOR} MATCHES "armv7")
-            # Raspberry Pi 2
-            list(APPEND ARCH_FLAGS -mfpu=neon-fp-armv8 -mno-unaligned-access -funsafe-math-optimizations)
+            if ("${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()
         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

@@ -97,9 +97,10 @@ endif
 #
 
 # keep standard at C11 and C++11
-MK_CPPFLAGS = -I. -Icommon
-MK_CFLAGS   = -std=c11   -fPIC
-MK_CXXFLAGS = -std=c++11 -fPIC
+MK_CPPFLAGS  = -I. -Icommon
+MK_CFLAGS    = -std=c11   -fPIC
+MK_CXXFLAGS  = -std=c++11 -fPIC
+MK_NVCCFLAGS = -std=c++11
 
 # -Ofast tends to produce faster code, but may not be available for some compilers.
 ifdef LLAMA_FAST
@@ -172,7 +173,7 @@ ifdef LLAMA_DEBUG
 	MK_LDFLAGS  += -g
 
 	ifeq ($(UNAME_S),Linux)
-		MK_CXXFLAGS += -Wp,-D_GLIBCXX_ASSERTIONS
+		MK_CPPFLAGS += -D_GLIBCXX_ASSERTIONS
 	endif
 else
 	MK_CPPFLAGS += -DNDEBUG
@@ -215,6 +216,11 @@ MK_CFLAGS    += $(WARN_FLAGS) -Wshadow -Wstrict-prototypes -Wpointer-arith -Wmis
 				-Werror=implicit-function-declaration
 MK_CXXFLAGS  += $(WARN_FLAGS) -Wmissing-declarations -Wmissing-noreturn
 
+ifeq ($(LLAMA_FATAL_WARNINGS),1)
+	MK_CFLAGS   += -Werror
+	MK_CXXFLAGS += -Werror
+endif
+
 # this version of Apple ld64 is buggy
 ifneq '' '$(findstring dyld-1015.7,$(shell $(CC) $(LDFLAGS) -Wl,-v 2>&1))'
 	MK_CPPFLAGS += -DHAVE_BUGGY_APPLE_LINKER
@@ -379,6 +385,9 @@ ifdef LLAMA_CUBLAS
 	MK_LDFLAGS   += -lcuda -lcublas -lculibos -lcudart -lcublasLt -lpthread -ldl -lrt -L/usr/local/cuda/lib64 -L/opt/cuda/lib64 -L$(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
@@ -437,9 +446,9 @@ 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 $(NVCCFLAGS) -Xcompiler "$(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) $(CPPFLAGS) -Xcompiler "$(CUDA_CXXFLAGS)" -c $< -o $@
 else
-	$(NVCC) $(BASE_CXXFLAGS) $(NVCCFLAGS) -Wno-pedantic -Xcompiler "$(CUDA_CXXFLAGS)" -c $< -o $@
+	$(NVCC) $(NVCCFLAGS) $(CPPFLAGS) -Xcompiler "$(CUDA_CXXFLAGS)" -c $< -o $@
 endif # JETSON_EOL_MODULE_DETECT
 endif # LLAMA_CUBLAS
 
@@ -524,11 +533,29 @@ 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
@@ -540,9 +567,10 @@ GF_CC := $(CC)
 include scripts/get-flags.mk
 
 # combine build flags with cmdline overrides
-override CFLAGS    := $(MK_CPPFLAGS) $(CPPFLAGS) $(MK_CFLAGS) $(GF_CFLAGS) $(CFLAGS)
-BASE_CXXFLAGS      := $(MK_CPPFLAGS) $(CPPFLAGS) $(MK_CXXFLAGS) $(CXXFLAGS)
-override CXXFLAGS  := $(BASE_CXXFLAGS) $(HOST_CXXFLAGS) $(GF_CXXFLAGS)
+override 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 NVCCFLAGS := $(MK_NVCCFLAGS) $(NVCCFLAGS)
 override LDFLAGS   := $(MK_LDFLAGS) $(LDFLAGS)
 
@@ -550,7 +578,7 @@ override LDFLAGS   := $(MK_LDFLAGS) $(LDFLAGS)
 ifdef LLAMA_CUBLAS
 GF_CC := $(NVCC) $(NVCCFLAGS) 2>/dev/null .c -Xcompiler
 include scripts/get-flags.mk
-CUDA_CXXFLAGS := $(GF_CXXFLAGS)
+CUDA_CXXFLAGS := $(BASE_CXXFLAGS) $(GF_CXXFLAGS) -Wno-pedantic
 endif
 
 #
@@ -569,7 +597,7 @@ $(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
@@ -691,7 +719,7 @@ 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 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 examples/llava/llava.h examples/llava/llava.cpp 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)
@@ -862,3 +890,7 @@ tests/test-model-load-cancel: tests/test-model-load-cancel.cpp ggml.o llama.o te
 tests/test-autorelease: tests/test-autorelease.cpp ggml.o llama.o tests/get-model.cpp $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
+
+tests/test-chat-template: tests/test-chat-template.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)

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: **/opt/intel/oneapi**.
+Recommend to install to default folder: **C:\Program Files (x86)\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

@@ -10,13 +10,9 @@ Inference of Meta's [LLaMA](https://arxiv.org/abs/2302.13971) model (and others)
 
 ### Hot topics
 
-- 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
+- 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
 - Looking for contributions to improve and maintain the `server` example: https://github.com/ggerganov/llama.cpp/issues/4216
 
 ----
@@ -61,7 +57,7 @@ variety of hardware - locally and in the cloud.
 - Plain C/C++ implementation without any dependencies
 - Apple silicon is a first-class citizen - optimized via ARM NEON, Accelerate and Metal frameworks
 - AVX, AVX2 and AVX512 support for x86 architectures
-- 2-bit, 3-bit, 4-bit, 5-bit, 6-bit, and 8-bit integer quantization for faster inference and reduced memory use
+- 1.5-bit, 2-bit, 3-bit, 4-bit, 5-bit, 6-bit, and 8-bit integer quantization for faster inference and reduced memory use
 - Custom CUDA kernels for running LLMs on NVIDIA GPUs (support for AMD GPUs via HIP)
 - Vulkan, SYCL, and (partial) OpenCL backend support
 - CPU+GPU hybrid inference to partially accelerate models larger than the total VRAM capacity
@@ -107,6 +103,7 @@ 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:**
 
@@ -117,6 +114,9 @@ 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:**
 
@@ -145,6 +145,7 @@ 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)
@@ -156,6 +157,9 @@ 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)
+- [LLMFarm](https://github.com/guinmoon/LLMFarm?tab=readme-ov-file) (MIT)
 
 ---
 
@@ -768,7 +772,7 @@ The time per token is measured on a MacBook M1 Pro 32GB RAM using 4 and 8 thread
 
 #### How to run
 
-1. Download/extract: https://s3.amazonaws.com/research.metamind.io/wikitext/wikitext-2-raw-v1.zip?ref=salesforce-research
+1. Download/extract: https://huggingface.co/datasets/ggml-org/ci/resolve/main/wikitext-2-raw-v1.zip
 2. Run `./perplexity -m models/7B/ggml-model-q4_0.gguf -f wiki.test.raw`
 3. Output:
 ```

build.zig

@@ -123,6 +123,7 @@ 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 });
@@ -131,7 +132,7 @@ 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 });
+    const server = make.exe("server", "examples/server/server.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common, buildinfo, sampling, grammar_parser, clip, llava });
     if (server.target.isWindows()) {
         server.linkSystemLibrary("ws2_32");
     }

ci/run.sh

@@ -33,7 +33,7 @@ sd=`dirname $0`
 cd $sd/../
 SRC=`pwd`
 
-CMAKE_EXTRA=""
+CMAKE_EXTRA="-DLLAMA_FATAL_WARNINGS=ON"
 
 if [ ! -z ${GG_BUILD_METAL} ]; then
     CMAKE_EXTRA="${CMAKE_EXTRA} -DLLAMA_METAL_SHADER_DEBUG=ON"
@@ -219,7 +219,7 @@ function gg_run_open_llama_3b_v2 {
     gg_wget models-mnt/open-llama/3B-v2/ https://huggingface.co/openlm-research/open_llama_3b_v2/resolve/main/pytorch_model.bin
     gg_wget models-mnt/open-llama/3B-v2/ https://huggingface.co/openlm-research/open_llama_3b_v2/raw/main/generation_config.json
 
-    gg_wget models-mnt/wikitext/ https://s3.amazonaws.com/research.metamind.io/wikitext/wikitext-2-raw-v1.zip
+    gg_wget models-mnt/wikitext/ https://huggingface.co/datasets/ggml-org/ci/resolve/main/wikitext-2-raw-v1.zip
     unzip -o models-mnt/wikitext/wikitext-2-raw-v1.zip -d models-mnt/wikitext/
     head -n 60 models-mnt/wikitext/wikitext-2-raw/wiki.test.raw > models-mnt/wikitext/wikitext-2-raw/wiki.test-60.raw
 
@@ -401,7 +401,7 @@ function gg_run_open_llama_7b_v2 {
     gg_wget models-mnt/open-llama/7B-v2/ https://huggingface.co/openlm-research/open_llama_7b_v2/resolve/main/pytorch_model-00002-of-00002.bin
     gg_wget models-mnt/open-llama/7B-v2/ https://huggingface.co/openlm-research/open_llama_7b_v2/raw/main/generation_config.json
 
-    gg_wget models-mnt/wikitext/ https://s3.amazonaws.com/research.metamind.io/wikitext/wikitext-2-raw-v1.zip
+    gg_wget models-mnt/wikitext/ https://huggingface.co/datasets/ggml-org/ci/resolve/main/wikitext-2-raw-v1.zip
     unzip -o models-mnt/wikitext/wikitext-2-raw-v1.zip -d models-mnt/wikitext/
 
     path_models="../models-mnt/open-llama/7B-v2"

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_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; }
+            /**/ 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; }
             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_NONE;
+                params.split_mode = LLAMA_SPLIT_MODE_NONE;
             } else if (arg_next == "layer") {
-                params.split_mode = LLAMA_SPLIT_LAYER;
+                params.split_mode = LLAMA_SPLIT_MODE_LAYER;
             } else if (arg_next == "row") {
-                params.split_mode = LLAMA_SPLIT_ROW;
+                params.split_mode = LLAMA_SPLIT_MODE_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_INT;
+                kvo.tag = LLAMA_KV_OVERRIDE_TYPE_INT;
                 kvo.int_value = std::atol(sep);
             } else if (strncmp(sep, "float:", 6) == 0) {
                 sep += 6;
-                kvo.tag = LLAMA_KV_OVERRIDE_FLOAT;
+                kvo.tag = LLAMA_KV_OVERRIDE_TYPE_FLOAT;
                 kvo.float_value = std::atof(sep);
             } else if (strncmp(sep, "bool:", 5) == 0) {
                 sep += 5;
-                kvo.tag = LLAMA_KV_OVERRIDE_BOOL;
+                kvo.tag = LLAMA_KV_OVERRIDE_TYPE_BOOL;
                 if (std::strcmp(sep, "true") == 0) {
                     kvo.bool_value = true;
                 } else if (std::strcmp(sep, "false") == 0) {
@@ -1704,6 +1704,7 @@ void dump_non_result_info_yaml(FILE * stream, const gpt_params & params, const l
     }
     fprintf(stream, "lora_base: %s\n", params.lora_base.c_str());
     fprintf(stream, "main_gpu: %d # default: 0\n", params.main_gpu);
+    fprintf(stream, "min_keep: %d # default: 0 (disabled)\n", sparams.min_keep);
     fprintf(stream, "mirostat: %d # default: 0 (disabled)\n", sparams.mirostat);
     fprintf(stream, "mirostat_ent: %f # default: 5.0\n", sparams.mirostat_tau);
     fprintf(stream, "mirostat_lr: %f # default: 0.1\n", sparams.mirostat_eta);
@@ -1741,7 +1742,7 @@ void dump_non_result_info_yaml(FILE * stream, const gpt_params & params, const l
 
     fprintf(stream, "rope_freq_base: %f # default: 10000.0\n", params.rope_freq_base);
     fprintf(stream, "rope_freq_scale: %f # default: 1.0\n", params.rope_freq_scale);
-    fprintf(stream, "seed: %d # default: -1 (random seed)\n", params.seed);
+    fprintf(stream, "seed: %u # default: -1 (random seed)\n", params.seed);
     fprintf(stream, "simple_io: %s # default: false\n", params.simple_io ? "true" : "false");
     fprintf(stream, "cont_batching: %s # default: false\n", params.cont_batching ? "true" : "false");
     fprintf(stream, "temp: %f # default: 0.8\n", sparams.temp);
@@ -1750,7 +1751,7 @@ void dump_non_result_info_yaml(FILE * stream, const gpt_params & params, const l
     dump_vector_float_yaml(stream, "tensor_split", tensor_split_vector);
 
     fprintf(stream, "tfs: %f # default: 1.0\n", sparams.tfs_z);
-    fprintf(stream, "threads: %d # default: %d\n", params.n_threads, std::thread::hardware_concurrency());
+    fprintf(stream, "threads: %d # default: %u\n", params.n_threads, std::thread::hardware_concurrency());
     fprintf(stream, "top_k: %d # default: 40\n", sparams.top_k);
     fprintf(stream, "top_p: %f # default: 0.95\n", sparams.top_p);
     fprintf(stream, "min_p: %f # default: 0.0\n", sparams.min_p);
@@ -1801,7 +1802,8 @@ void dump_kv_cache_view_seqs(const llama_kv_cache_view & view, int row_size) {
             if (cs_curr[j] < 0) { continue; }
             if (seqs.find(cs_curr[j]) == seqs.end()) {
                 if (seqs.size() + 1 >= sizeof(slot_chars)) { break; }
-                seqs[cs_curr[j]] = seqs.size();
+                const size_t sz = seqs.size();
+                seqs[cs_curr[j]] = sz;
             }
         }
         if (seqs.size() + 1 >= sizeof(slot_chars)) { break; }

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_LAYER; // how to split the model across GPUs
+    llama_split_mode split_mode   = LLAMA_SPLIT_MODE_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_UNSPECIFIED;
+    int32_t rope_scaling_type     = LLAMA_ROPE_SCALING_TYPE_UNSPECIFIED;
     ggml_numa_strategy numa       = GGML_NUMA_STRATEGY_DISABLED;
 
     // // sampling parameters

common/sampling.cpp

@@ -121,7 +121,7 @@ static void sampler_queue(
                    struct llama_context * ctx_main,
             const llama_sampling_params & params,
                  llama_token_data_array & cur_p,
-                                 size_t & min_keep) {
+                                 size_t   min_keep) {
     const float         temp              = params.temp;
     const float         dynatemp_range    = params.dynatemp_range;
     const float         dynatemp_exponent = params.dynatemp_exponent;
@@ -249,7 +249,7 @@ static llama_token llama_sampling_sample_impl(
             id = llama_sample_token_mirostat_v2(ctx_main, &cur_p, mirostat_tau, mirostat_eta, &ctx_sampling->mirostat_mu);
         } else {
             // temperature sampling
-            size_t min_keep = std::max(1, params.n_probs);
+            size_t min_keep = std::max(1, params.min_keep);
 
             sampler_queue(ctx_main, params, cur_p, min_keep);
 
@@ -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/sampling.h

@@ -22,6 +22,7 @@ enum class llama_sampler_type : char {
 typedef struct llama_sampling_params {
     int32_t     n_prev                = 64;       // number of previous tokens to remember
     int32_t     n_probs               = 0;        // if greater than 0, output the probabilities of top n_probs tokens.
+    int32_t     min_keep              = 0;        // 0 = disabled, otherwise samplers should return at least min_keep tokens
     int32_t     top_k                 = 40;       // <= 0 to use vocab size
     float       top_p                 = 0.95f;    // 1.0 = disabled
     float       min_p                 = 0.05f;    // 0.0 = disabled

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_ADAM);
+    state->opt->params = ggml_opt_default_params(GGML_OPT_TYPE_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_ADAM;
+        opt->params.type = GGML_OPT_TYPE_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_LBFGS;
+        opt->params.type = GGML_OPT_TYPE_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_ADAM:
+        case GGML_OPT_TYPE_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_LBFGS:
+        case GGML_OPT_TYPE_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 ("StableLMEpochForCausalLM", "LlavaStableLMEpochForCausalLM"):
+        if model_architecture in ("StableLmForCausalLM", "StableLMEpochForCausalLM", "LlavaStableLMEpochForCausalLM"):
             return StableLMModel
         if model_architecture == "QWenLMHeadModel":
             return QwenModel
@@ -218,6 +218,8 @@ class Model:
             return BertModel
         if model_architecture == "NomicBertModel":
             return NomicBertModel
+        if model_architecture == "GemmaForCausalLM":
+            return GemmaModel
         return Model
 
     def _is_model_safetensors(self) -> bool:
@@ -251,7 +253,7 @@ class Model:
             return gguf.MODEL_ARCH.REFACT
         if arch == "PersimmonForCausalLM":
             return gguf.MODEL_ARCH.PERSIMMON
-        if arch in ("StableLMEpochForCausalLM", "LlavaStableLMEpochForCausalLM"):
+        if arch in ("StableLmForCausalLM", "StableLMEpochForCausalLM", "LlavaStableLMEpochForCausalLM"):
             return gguf.MODEL_ARCH.STABLELM
         if arch == "QWenLMHeadModel":
             return gguf.MODEL_ARCH.QWEN
@@ -277,6 +279,8 @@ 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!')
 
@@ -618,11 +622,6 @@ 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):
@@ -655,6 +654,8 @@ 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):
@@ -1031,7 +1032,6 @@ 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()
@@ -1074,10 +1074,11 @@ 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"])
-        self.gguf_writer.add_rope_dimension_count(int(hparams["rope_pct"] * (hparams["hidden_size"] // hparams["num_attention_heads"])))
+        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_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(1e-5)
+        self.gguf_writer.add_layer_norm_eps(self.find_hparam(["layer_norm_eps", "norm_eps"]))
 
 
 class MixtralModel(Model):
@@ -1785,6 +1786,63 @@ 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

@@ -1533,27 +1533,28 @@ int main(int argc, char ** argv) {
 
         int n_past = 0;
 
-        ggml_cgraph gf = {};
+        struct ggml_cgraph * gf = NULL;
+        gf = ggml_new_graph_custom(ctx0, LLAMA_TRAIN_MAX_NODES, true);
 
         get_example_targets_batch(ctx0, 64*ex+0,  tokens_input, targets);
 
-        struct ggml_tensor * logits = forward_batch(&model, &kv_self, ctx0, &gf, tokens_input, n_tokens, n_past, n_batch);
+        struct ggml_tensor * logits = forward_batch(&model, &kv_self, ctx0, gf, tokens_input, n_tokens, n_past, n_batch);
         // struct ggml_tensor * e = cross_entropy_loss(ctx0, targets, logits);
         struct ggml_tensor * e = square_error_loss(ctx0, targets, logits);
 
-        ggml_build_forward_expand(&gf, e);
-        ggml_graph_compute_helper(work_buffer, &gf, /*n_threads*/ 1);
+        ggml_build_forward_expand(gf, e);
+        ggml_graph_compute_helper(work_buffer, gf, /*n_threads*/ 1);
 
         float error_before_opt = ggml_get_f32_1d(e, 0);
 
-        struct ggml_opt_params opt_params_lbfgs = ggml_opt_default_params(GGML_OPT_LBFGS);
+        struct ggml_opt_params opt_params_lbfgs = ggml_opt_default_params(GGML_OPT_TYPE_LBFGS);
         opt_params_lbfgs.print_forward_graph = false;
         opt_params_lbfgs.print_backward_graph = false;
         opt_params_lbfgs.lbfgs.n_iter = 16;
         ggml_opt(ctx0, opt_params_lbfgs, e);
         //
-        ggml_build_forward_expand(&gf, e);
-        ggml_graph_compute_helper(work_buffer, &gf, /*n_threads*/ 1);
+        ggml_build_forward_expand(gf, e);
+        ggml_graph_compute_helper(work_buffer, gf, /*n_threads*/ 1);
 
         float error_after_opt = ggml_get_f32_1d(e, 0);
 
@@ -1600,13 +1601,14 @@ int main(int argc, char ** argv) {
             };
             struct ggml_context * ctx0 = ggml_init(params);
 
-            ggml_cgraph gf = {};
+            struct ggml_cgraph * gf = NULL;
+            gf = ggml_new_graph_custom(ctx0, LLAMA_TRAIN_MAX_NODES, true);
 
             int n_past = 0;
-            struct ggml_tensor * logits = forward(&model, &kv_self, ctx0, &gf, tokens_input, sample_ctx, n_past);
+            struct ggml_tensor * logits = forward(&model, &kv_self, ctx0, gf, tokens_input, sample_ctx, n_past);
 
-            ggml_build_forward_expand(&gf, logits);
-            ggml_graph_compute_helper(work_buffer, &gf, /*n_threads*/ 1);
+            ggml_build_forward_expand(gf, logits);
+            ggml_graph_compute_helper(work_buffer, gf, /*n_threads*/ 1);
 
             struct ggml_tensor * best_samples = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, sample_ctx);
             struct ggml_tensor * probs        = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_vocab, sample_ctx);

examples/batched-bench/batched-bench.cpp

@@ -159,7 +159,7 @@ int main(int argc, char ** argv) {
     }
 
     LOG_TEE("\n");
-    LOG_TEE("%s: n_kv_max = %d, is_pp_shared = %d, n_gpu_layers = %d, mmq = %d, n_threads = %d, n_threads_batch = %d\n", __func__, n_kv_max, is_pp_shared, n_gpu_layers, mmq, ctx_params.n_threads, ctx_params.n_threads_batch);
+    LOG_TEE("%s: n_kv_max = %d, is_pp_shared = %d, n_gpu_layers = %d, mmq = %d, n_threads = %u, n_threads_batch = %u\n", __func__, n_kv_max, is_pp_shared, n_gpu_layers, mmq, ctx_params.n_threads, ctx_params.n_threads_batch);
     LOG_TEE("\n");
 
     LOG_TEE("|%6s | %6s | %4s | %6s | %8s | %8s | %8s | %8s | %8s | %8s |\n", "PP",     "TG",     "B",    "N_KV",     "T_PP s",   "S_PP t/s", "T_TG s",   "S_TG t/s", "T s",      "S t/s");

examples/batched/batched.cpp

@@ -92,7 +92,7 @@ int main(int argc, char ** argv) {
 
     const int n_ctx    = llama_n_ctx(ctx);
 
-    LOG_TEE("\n%s: n_len = %d, n_ctx = %d, n_batch = %d, n_parallel = %d, n_kv_req = %d\n", __func__, n_len, n_ctx, ctx_params.n_batch, n_parallel, n_kv_req);
+    LOG_TEE("\n%s: n_len = %d, n_ctx = %d, n_batch = %u, n_parallel = %d, n_kv_req = %d\n", __func__, n_len, n_ctx, ctx_params.n_batch, n_parallel, n_kv_req);
 
     // make sure the KV cache is big enough to hold all the prompt and generated tokens
     if (n_kv_req > n_ctx) {

examples/convert-llama2c-to-ggml/README.md

@@ -21,6 +21,8 @@ An example command using a model from [karpathy/tinyllamas](https://huggingface.
 
 `$ ./convert-llama2c-to-ggml --copy-vocab-from-model llama-2-7b-chat.gguf.q2_K.bin --llama2c-model stories42M.bin --llama2c-output-model stories42M.gguf.bin`
 
+Note: The vocabulary for `stories260K.bin` should be its own tokenizer `tok512.bin` found in [karpathy/tinyllamas/stories260K](https://huggingface.co/karpathy/tinyllamas/tree/main/stories260K).
+
 Now you can use the model with a command like:
 
 `$ ./main -m stories42M.gguf.bin -p "One day, Lily met a Shoggoth" -n 500 -c 256`

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

@@ -1,6 +1,7 @@
 #include "ggml.h"
 #include "llama.h"
 #include "common.h"
+#include "log.h"
 
 #include <unordered_map>
 #include <vector>
@@ -78,111 +79,101 @@ typedef struct {
 
 struct TransformerWeights {
     // token embedding table
-    float* token_embedding_table;    // (vocab_size, dim)
+    std::vector<float> token_embedding_table;    // (vocab_size, dim)
     // weights for rmsnorms
-    float* rms_att_weight; // (layer, dim) rmsnorm weights
-    float* rms_ffn_weight; // (layer, dim)
+    std::vector<float> rms_att_weight; // (layer, dim) rmsnorm weights
+    std::vector<float> rms_ffn_weight; // (layer, dim)
     // weights for matmuls
-    float* wq; // (layer, dim, dim)
-    float* wk; // (layer, dim, dim)
-    float* wv; // (layer, dim, dim)
-    float* wo; // (layer, dim, dim)
+    std::vector<float> wq; // (layer, dim, dim)
+    std::vector<float> wk; // (layer, dim, dim)
+    std::vector<float> wv; // (layer, dim, dim)
+    std::vector<float> wo; // (layer, dim, dim)
     // weights for ffn
-    float* w1; // (layer, hidden_dim, dim)
-    float* w2; // (layer, dim, hidden_dim)
-    float* w3; // (layer, hidden_dim, dim)
+    std::vector<float> w1; // (layer, hidden_dim, dim)
+    std::vector<float> w2; // (layer, dim, hidden_dim)
+    std::vector<float> w3; // (layer, hidden_dim, dim)
     // final rmsnorm
-    float* rms_final_weight; // (dim,)
+    std::vector<float> rms_final_weight; // (dim,)
     // freq_cis for RoPE relatively positional embeddings
-    // float* freq_cis_real; // (seq_len, dim/2)
-    // float* freq_cis_imag; // (seq_len, dim/2)
+    // std::vector<float> freq_cis_real; // (seq_len, dim/2)
+    // std::vector<float> freq_cis_imag; // (seq_len, dim/2)
     // (optional) classifier weights for the logits, on the last layer
-    float* wcls;
-
-    ~TransformerWeights() {
-        delete[] token_embedding_table;
-        delete[] rms_att_weight;
-        delete[] rms_ffn_weight;
-        delete[] wq;
-        delete[] wk;
-        delete[] wv;
-        delete[] wo;
-        delete[] w1;
-        delete[] w2;
-        delete[] w3;
-        delete[] rms_final_weight;
-        delete[] wcls;
-    }
+    std::vector<float> wcls;
 };
 
-static void malloc_weights(TransformerWeights* w, Config* p, bool shared_weights) {
-    // we calloc instead of malloc to keep valgrind happy
-    w->token_embedding_table = new float[p->vocab_size * p->dim]();
-    printf("[%s:AK] Allocating [%d] x [%d] = [%d] float space for w->token_embedding_table\n",__func__,p->vocab_size , p->dim, p->vocab_size * p->dim);
+static void alloc_weights(TransformerWeights * w, const Config * p, bool shared_weights) {
+    const int n_multiqueries = p->n_kv_heads <= 0 || p->n_kv_heads >= p->n_heads ? 1 : p->n_heads / p->n_kv_heads;
+    try {
+        w->token_embedding_table.resize(p->vocab_size * p->dim);
+        LOG("%s: Allocating [%d] x [%d] = [%d] float space for w->token_embedding_table\n",__func__,p->vocab_size , p->dim, p->vocab_size * p->dim);
 
-    w->rms_att_weight = new float[p->n_layers * p->dim]();
-    printf("[%s:AK] Allocating [%d] x [%d] = [%d] float space for w->rms_att_weight\n",__func__,p->n_layers, p->dim, p->n_layers * p->dim);
+        w->rms_att_weight.resize(p->n_layers * p->dim);
+        LOG("%s: Allocating [%d] x [%d] = [%d] float space for w->rms_att_weight\n",__func__,p->n_layers, p->dim, p->n_layers * p->dim);
 
-    w->rms_ffn_weight = new float[p->n_layers * p->dim]();
-    printf("[%s:AK] Allocating [%d] x [%d] = [%d] float space for w->rms_ffn_weight\n",__func__,p->n_layers , p->dim, p->n_layers * p->dim);
+        w->rms_ffn_weight.resize(p->n_layers * p->dim);
+        LOG("%s: Allocating [%d] x [%d] = [%d] float space for w->rms_ffn_weight\n",__func__,p->n_layers , p->dim, p->n_layers * p->dim);
 
-    w->wq = new float[p->n_layers * p->dim * p->dim]();
-    printf("[%s:AK] Allocating [%d] x [%d] x [%d] = [%d] float space for w->wq\n",__func__,p->n_layers, p->dim, p->dim, p->n_layers * p->dim * p->dim);
+        w->wq.resize(p->n_layers * p->dim * p->dim);
+        LOG("%s: Allocating [%d] x [%d] x [%d] = [%d] float space for w->wq\n",__func__,p->n_layers, p->dim, p->dim, p->n_layers * p->dim * p->dim);
 
-    w->wk = new float[p->n_layers * p->dim * p->dim]();
-    printf("[%s:AK] Allocating [%d] x [%d] x [%d] = [%d] float space for w->wk\n",__func__,p->n_layers, p->dim, p->dim, p->n_layers * p->dim * p->dim);
+        w->wk.resize(p->n_layers * p->dim * p->dim / n_multiqueries);
+        LOG("%s: Allocating [%d] x [%d] x [%d] = [%d] float space for w->wk\n",__func__,p->n_layers, p->dim, p->dim / n_multiqueries, p->n_layers * p->dim * p->dim / n_multiqueries);
 
-    w->wv = new float[p->n_layers * p->dim * p->dim]();
-    printf("[%s:AK] Allocating [%d] x [%d] x [%d] = [%d] float space for w->wv\n",__func__, p->n_layers, p->dim, p->dim, p->n_layers * p->dim * p->dim);
+        w->wv.resize(p->n_layers * p->dim * p->dim / n_multiqueries);
+        LOG("%s: Allocating [%d] x [%d] x [%d] = [%d] float space for w->wv\n",__func__, p->n_layers, p->dim, p->dim / n_multiqueries, p->n_layers * p->dim * p->dim / n_multiqueries);
 
-    w->wo = new float[p->n_layers * p->dim * p->dim]();
-    printf("[%s:AK] Allocating [%d] x [%d] x [%d] = [%d] float space for w->wo\n",__func__,p->n_layers, p->dim, p->dim, p->n_layers * p->dim * p->dim);
+        w->wo.resize(p->n_layers * p->dim * p->dim);
+        LOG("%s: Allocating [%d] x [%d] x [%d] = [%d] float space for w->wo\n",__func__,p->n_layers, p->dim, p->dim, p->n_layers * p->dim * p->dim);
 
-    w->w1 = new float[p->n_layers * p->hidden_dim * p->dim]();
-    printf("[%s:AK] Allocating [%d] x [%d] x [%d] = [%d] float space for w->w1\n",__func__,p->n_layers, p->hidden_dim, p->dim, p->n_layers * p->hidden_dim * p->dim);
+        w->w1.resize(p->n_layers * p->hidden_dim * p->dim);
+        LOG("%s: Allocating [%d] x [%d] x [%d] = [%d] float space for w->w1\n",__func__,p->n_layers, p->hidden_dim, p->dim, p->n_layers * p->hidden_dim * p->dim);
 
-    w->w2 = new float[p->n_layers * p->hidden_dim * p->dim]();
-    printf("[%s:AK] Allocating [%d] x [%d] x [%d] = [%d] float space for w->w2\n",__func__,p->n_layers, p->dim, p->hidden_dim, p->n_layers * p->hidden_dim * p->dim);
+        w->w2.resize(p->n_layers * p->hidden_dim * p->dim);
+        LOG("%s: Allocating [%d] x [%d] x [%d] = [%d] float space for w->w2\n",__func__,p->n_layers, p->dim, p->hidden_dim, p->n_layers * p->hidden_dim * p->dim);
 
-    w->w3 = new float[p->n_layers * p->hidden_dim * p->dim]();
-    printf("[%s:AK] Allocating [%d] x [%d] x [%d] = [%d] float space for w->w3\n",__func__,p->n_layers, p->hidden_dim, p->dim, p->n_layers * p->hidden_dim * p->dim);
+        w->w3.resize(p->n_layers * p->hidden_dim * p->dim);
+        LOG("%s: Allocating [%d] x [%d] x [%d] = [%d] float space for w->w3\n",__func__,p->n_layers, p->hidden_dim, p->dim, p->n_layers * p->hidden_dim * p->dim);
 
-    w->rms_final_weight = new float[p->dim]();
-    printf("[%s:AK] Allocating [%d] float space for w->rms_final_weight\n",__func__,p->dim);
+        w->rms_final_weight.resize(p->dim);
+        LOG("%s: Allocating [%d] float space for w->rms_final_weight\n",__func__,p->dim);
 
-    if (shared_weights) {
-        w->wcls = NULL;
-    } else {
-        w->wcls = new float[p->vocab_size * p->dim]();
-        printf("[%s:AK] Allocating [%d] x [%d] = [%d] float space for w->wcls\n",__func__,p->vocab_size , p->dim, p->vocab_size * p->dim);
+        if (shared_weights) {
+            w->wcls = {};
+        } else {
+            w->wcls.resize(p->vocab_size * p->dim);
+            LOG("%s: Allocating [%d] x [%d] = [%d] float space for w->wcls\n",__func__,p->vocab_size , p->dim, p->vocab_size * p->dim);
+        }
+    }
+    catch (std::length_error &) {
+        die("Invalid configuration. Failed to allocate memory for weights");
     }
 }
 
-static int checkpoint_init_weights(TransformerWeights *w, Config* p, FILE* f, bool shared_weights) {
-    if (fread(w->token_embedding_table, sizeof(float), p->vocab_size * p->dim, f) != static_cast<size_t>(p->vocab_size * p->dim)) return 1;
-    if (fread(w->rms_att_weight, sizeof(float), p->n_layers * p->dim, f) != static_cast<size_t>(p->n_layers * p->dim)) return 1;
-    if (fread(w->wq, sizeof(float), p->n_layers * p->dim * p->dim, f) != static_cast<size_t>(p->n_layers * p->dim * p->dim)) return 1;
-    if (fread(w->wk, sizeof(float), p->n_layers * p->dim * p->dim, f) != static_cast<size_t>(p->n_layers * p->dim * p->dim)) return 1;
-    if (fread(w->wv, sizeof(float), p->n_layers * p->dim * p->dim, f) != static_cast<size_t>(p->n_layers * p->dim * p->dim)) return 1;
-    if (fread(w->wo, sizeof(float), p->n_layers * p->dim * p->dim, f) != static_cast<size_t>(p->n_layers * p->dim * p->dim)) return 1;
-    if (fread(w->rms_ffn_weight, sizeof(float), p->n_layers * p->dim, f) != static_cast<size_t>(p->n_layers * p->dim)) return 1;
-    if (fread(w->w1, sizeof(float), p->n_layers * p->dim * p->hidden_dim, f) != static_cast<size_t>(p->n_layers * p->dim * p->hidden_dim)) return 1;
-    if (fread(w->w2, sizeof(float), p->n_layers * p->hidden_dim * p->dim, f) != static_cast<size_t>(p->n_layers * p->hidden_dim * p->dim)) return 1;
-    if (fread(w->w3, sizeof(float), p->n_layers * p->dim * p->hidden_dim, f) != static_cast<size_t>(p->n_layers * p->dim * p->hidden_dim)) return 1;
-    if (fread(w->rms_final_weight, sizeof(float), p->dim, f) != static_cast<size_t>(p->dim)) return 1;
+static int checkpoint_init_weights(TransformerWeights * w, const Config * p, FILE * f, bool shared_weights) {
+    if (fread(w->token_embedding_table.data(), sizeof(float), w->token_embedding_table.size(), f) != w->token_embedding_table.size()) return 1;
+    if (fread(w->rms_att_weight.data(), sizeof(float), w->rms_att_weight.size(), f) != w->rms_att_weight.size()) return 1;
+    if (fread(w->wq.data(), sizeof(float), w->wq.size(), f) != w->wq.size()) return 1;
+    if (fread(w->wk.data(), sizeof(float), w->wk.size(), f) != w->wk.size()) return 1;
+    if (fread(w->wv.data(), sizeof(float), w->wv.size(), f) != w->wv.size()) return 1;
+    if (fread(w->wo.data(), sizeof(float), w->wo.size(), f) != w->wo.size()) return 1;
+    if (fread(w->rms_ffn_weight.data(), sizeof(float), w->rms_ffn_weight.size(), f) != w->rms_ffn_weight.size()) return 1;
+    if (fread(w->w1.data(), sizeof(float), w->w1.size(), f) != w->w1.size()) return 1;
+    if (fread(w->w2.data(), sizeof(float), w->w2.size(), f) != w->w2.size()) return 1;
+    if (fread(w->w3.data(), sizeof(float), w->w3.size(), f) != w->w3.size()) return 1;
+    if (fread(w->rms_final_weight.data(), sizeof(float), w->rms_final_weight.size(), f) != w->rms_final_weight.size()) return 1;
 
     // Skip freq_cis_real & freq_cis_imag
     int head_size = p->dim / p->n_heads;
     fseek(f, p->seq_len * head_size * sizeof(float), SEEK_CUR);
 
-    if (!shared_weights && fread(w->wcls, sizeof(float), p->vocab_size * p->dim, f) != static_cast<size_t>(p->vocab_size * p->dim)) return 1;
+    if (!shared_weights && fread(w->wcls.data(), sizeof(float), w->wcls.size(), f) != w->wcls.size()) return 1;
 
     // Check we didn't forget to read anything
     auto curr = ftell(f);
     fseek(f, 0, SEEK_END);
     auto end = ftell(f);
     if (curr != end) {
-        printf("Error: failed to read the checkpoint file to the end (curr = %ld, end =  %ld)\n", curr, end);
+        LOG("%s: Error: failed to read the checkpoint file to the end (curr = %ld, end =  %ld)\n", __func__, curr, end);
         return 1;
     }
 
@@ -190,20 +181,20 @@ static int checkpoint_init_weights(TransformerWeights *w, Config* p, FILE* f, bo
 }
 
 static void print_sample_weights(TransformerWeights *w){
-    printf("----- Quick print of first of the weight vales of all the variables\n");
-    printf("%f\n", w->token_embedding_table[0]);
-    printf("%f\n", w->rms_att_weight[0]);
-    printf("%f\n", w->rms_ffn_weight[0]);
+    LOG("----- Quick print of first of the weight vales of all the variables\n");
+    LOG("%f\n", w->token_embedding_table[0]);
+    LOG("%f\n", w->rms_att_weight[0]);
+    LOG("%f\n", w->rms_ffn_weight[0]);
 
-    printf("%f\n", w->wq[0]);
-    printf("%f\n", w->wk[0]);
-    printf("%f\n", w->wv[0]);
-    printf("%f\n", w->wo[0]);
-    printf("%f\n", w->w1[0]);
-    printf("%f\n", w->w2[0]);
-    printf("%f\n", w->w3[0]);
-    printf("%f\n", w->rms_att_weight[0]);
-    if (w->wcls) printf("%f\n", w->wcls[0]);
+    LOG("%f\n", w->wq[0]);
+    LOG("%f\n", w->wk[0]);
+    LOG("%f\n", w->wv[0]);
+    LOG("%f\n", w->wo[0]);
+    LOG("%f\n", w->w1[0]);
+    LOG("%f\n", w->w2[0]);
+    LOG("%f\n", w->w3[0]);
+    LOG("%f\n", w->rms_att_weight[0]);
+    if (!w->wcls.empty()) LOG("%f\n", w->wcls[0]);
 }
 ////////////////////////////////////////////////////////////////////////////////////////////////////////////
 
@@ -225,14 +216,16 @@ struct llama_vocab {
 };
 
 struct my_llama_hparams {
-    uint32_t n_vocab = 32000;
-    uint32_t n_ctx   = 512;   // this is provided as user input?
-    uint32_t n_embd  = 4096;
-    uint32_t n_ff    = 11008;
-    uint32_t n_mult  = 4;
-    uint32_t n_head  = 32;
-    uint32_t n_layer = 32;
-    uint32_t n_rot   = 64;
+    uint32_t n_vocab   = 32000;
+    uint32_t n_ctx     = 512;   // this is provided as user input?
+    uint32_t n_embd    = 4096;
+    uint32_t n_ff      = 11008;
+    uint32_t n_mult    = 4;
+    uint32_t n_head    = 32;
+    uint32_t n_head_kv = 32;
+    uint32_t n_layer   = 32;
+    uint32_t n_rot     = 64;
+
     bool operator!=(const my_llama_hparams& other) const {
         return memcmp(this, &other, sizeof(my_llama_hparams));
     }
@@ -325,14 +318,30 @@ struct train_params {
 };
 
 static void print_params(struct my_llama_hparams * params) {
-    printf("%s: n_vocab: %d\n", __func__, params->n_vocab);
-    printf("%s: n_ctx:   %d\n", __func__, params->n_ctx);
-    printf("%s: n_embd:  %d\n", __func__, params->n_embd);
-    printf("%s: n_mult:  %d\n", __func__, params->n_mult);
-    printf("%s: n_head:  %d\n", __func__, params->n_head);
-    printf("%s: n_ff:    %d\n", __func__, params->n_ff);
-    printf("%s: n_layer: %d\n", __func__, params->n_layer);
-    printf("%s: n_rot:   %d\n", __func__, params->n_rot);
+    LOG("%s: n_vocab:   %u\n", __func__, params->n_vocab);
+    LOG("%s: n_ctx:     %u\n", __func__, params->n_ctx);
+    LOG("%s: n_embd:    %u\n", __func__, params->n_embd);
+    LOG("%s: n_mult:    %u\n", __func__, params->n_mult);
+    LOG("%s: n_head:    %u\n", __func__, params->n_head);
+    LOG("%s: n_head_kv: %u\n", __func__, params->n_head_kv);
+    LOG("%s: n_ff:      %u\n", __func__, params->n_ff);
+    LOG("%s: n_layer:   %u\n", __func__, params->n_layer);
+    LOG("%s: n_rot:     %u\n", __func__, params->n_rot);
+}
+
+static void print_tensor_info(const struct ggml_context * ctx) {
+    for (auto t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) {
+        LOG("%s: Allocating ", __func__);
+        int64_t total = 1;
+        int i = 0;
+        for (; i < ggml_n_dims(t); ++i) {
+            if (i > 0) LOG("x ");
+            LOG("[%" PRId64 "] ", t->ne[i]);
+            total *= t->ne[i];
+        }
+        if (i > 1) LOG("= [%" PRId64 "] ", total);
+        LOG("float space for %s\n", ggml_get_name(t));
+    }
 }
 
 static void init_model(struct my_llama_model * model) {
@@ -342,6 +351,8 @@ static void init_model(struct my_llama_model * model) {
     const uint32_t n_layer = hparams.n_layer;
     const uint32_t n_vocab = hparams.n_vocab;
 
+    const uint32_t n_multiqueries = hparams.n_head_kv <= 0 || hparams.n_head_kv >= hparams.n_head ? 1 : hparams.n_head / hparams.n_head_kv;
+
     const uint32_t n_ff = hparams.n_ff;
     struct ggml_context * ctx = model->ctx;
 
@@ -350,25 +361,8 @@ static void init_model(struct my_llama_model * model) {
     model->train_tokens = 0;
 
     model->tok_embeddings = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_vocab);
-    printf("[%s:GG] Allocating [%d] x [%d] = [%d] float space for model->tok_embeddings\n",__func__,n_embd , n_vocab, n_embd * n_vocab);
-
     model->norm           = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
-    printf("[%s:GG] Allocating [%d] float space for model->norm\n",__func__,n_embd);
-
     model->output         = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_vocab);
-    printf("[%s:GG] Allocating [%d] x[%d] = [%d] float space for model->output\n",__func__,n_embd, n_vocab, n_embd * n_vocab);
-
-    // printing the per-layer allocations here so we dont print in the for loop.
-    printf("[%s:GG] Allocating [%d] x[%d] = [%d] float space for layer.wq for [%d] layers\n",__func__, n_embd, n_embd, n_embd * n_embd, n_layer);
-    printf("[%s:GG] Allocating [%d] x[%d] = [%d] float space for layer.wk for [%d] layers\n",__func__, n_embd, n_embd, n_embd * n_embd, n_layer);
-    printf("[%s:GG] Allocating [%d] x[%d] = [%d] float space for layer.wv for [%d] layers\n",__func__, n_embd, n_embd, n_embd * n_embd, n_layer);
-    printf("[%s:GG] Allocating [%d] x[%d] = [%d] float space for layer.wo for [%d] layers\n",__func__, n_embd, n_embd, n_embd * n_embd, n_layer);
-
-    printf("[%s:GG] Allocating [%d] float space for layer.ffn_norm for [%d] layers\n",__func__,n_embd, n_layer);
-
-    printf("[%s:GG] Allocating [%d] x[%d] = [%d] float space for layer.w1 for [%d] layers\n",__func__, n_ff, n_embd, n_embd * n_ff, n_layer);
-    printf("[%s:GG] Allocating [%d] x[%d] = [%d] float space for layer.w2 for [%d] layers\n",__func__, n_embd, n_ff, n_ff * n_embd, n_layer);
-    printf("[%s:GG] Allocating [%d] x[%d] = [%d] float space for layer.w3 for [%d] layers\n",__func__, n_ff, n_embd, n_embd * n_ff, n_layer);
 
     ggml_set_name(model->tok_embeddings, "tok_embeddings.weight");
     ggml_set_name(model->norm,           "norm.weight");
@@ -383,8 +377,8 @@ static void init_model(struct my_llama_model * model) {
         layer.attention_norm = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
 
         layer.wq = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_embd);
-        layer.wk = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_embd);
-        layer.wv = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_embd);
+        layer.wk = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_embd / n_multiqueries);
+        layer.wv = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_embd / n_multiqueries);
         layer.wo = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_embd);
 
         layer.ffn_norm = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
@@ -406,6 +400,8 @@ static void init_model(struct my_llama_model * model) {
         ggml_format_name(layer.w2, "%s.feed_forward.w2.weight", layers_i.c_str());
         ggml_format_name(layer.w3, "%s.feed_forward.w3.weight", layers_i.c_str());
     }
+
+    print_tensor_info(ctx);
 }
 
 static float get_f32_2d(struct ggml_tensor * tensor, int64_t i0, int64_t i1) {
@@ -421,9 +417,9 @@ static int32_t get_i32_2d(struct ggml_tensor * tensor, int64_t i0, int64_t i1) {
 static void print_row(struct ggml_tensor * probs, int i) {
     for (int k = 0; k < probs->ne[0]; ++k) {
         float p = get_f32_2d(probs, k, i);
-        printf(" %f", p);
+        LOG(" %f", p);
     }
-    printf("\n");
+    LOG("\n");
 }
 
 static void print_matrix(struct ggml_tensor * probs) {
@@ -431,33 +427,12 @@ static void print_matrix(struct ggml_tensor * probs) {
     for (int i = 0; i < probs->ne[1]; ++i) {
         for (int k = 0; k < probs->ne[0]; ++k) {
             float p = get_f32_2d(probs, k, i);
-            printf(" %.2f", p);
+            LOG(" %.2f", p);
         }
-        printf("\n");
+        LOG("\n");
     }
 }
 
-#ifdef __GNUC__
-#ifdef __MINGW32__
-__attribute__((format(gnu_printf, 1, 2)))
-#else
-__attribute__((format(printf, 1, 2)))
-#endif
-#endif
-static std::string format(const char * fmt, ...) {
-    va_list ap, ap2;
-    va_start(ap, fmt);
-    va_copy(ap2, ap);
-    int size = vsnprintf(NULL, 0, fmt, ap);
-    GGML_ASSERT(size >= 0 && size < INT_MAX);
-    std::vector<char> buf(size + 1);
-    int size2 = vsnprintf(buf.data(), size + 1, fmt, ap2);
-    GGML_ASSERT(size2 == size);
-    va_end(ap2);
-    va_end(ap);
-    return std::string(buf.data(), size);
-}
-
 struct llama_file {
     // use FILE * so we don't have to re-open the file to mmap
     FILE * fp;
@@ -549,8 +524,9 @@ static std::string llama_escape_whitespaces(const std::string & text) {
     return out.str();
 }
 
-static void load_vocab(const char *filename, Config *config, struct llama_vocab *vocab) {
+static void load_vocab(const char * filename, const Config * config, struct llama_vocab * vocab) {
     if (is_ggml_file(filename)) {
+        LOG("%s: Loading vocabulary from gguf file %s\n", __func__, filename);
         struct ggml_context * ctx_data = NULL;
 
         struct gguf_init_params params = {
@@ -578,6 +554,9 @@ static void load_vocab(const char *filename, Config *config, struct llama_vocab
         const int * toktypes = (const int * ) gguf_get_arr_data(ctx, toktype_idx);
 
         const uint32_t n_vocab = gguf_get_arr_n(ctx, token_idx);
+        if (n_vocab != static_cast<uint32_t>(config->vocab_size)) {
+            die_fmt("vocab size mismatch: (gguf) %u != (llama2c) %d", n_vocab, config->vocab_size);
+        }
 
         vocab->id_to_token.resize(n_vocab);
 
@@ -595,7 +574,7 @@ static void load_vocab(const char *filename, Config *config, struct llama_vocab
         gguf_free(ctx);
     } else {
         // assume llama2.c vocabulary
-        printf("Assuming llama2.c vocabulary since %s is not a gguf file\n", filename);
+        LOG("%s: Assuming llama2.c vocabulary since %s is not a gguf file\n", __func__, filename);
         llama_file file(filename, "rb");
         if (!file.fp) {
             die_fmt("%s: %s", strerror(errno), filename);
@@ -638,38 +617,15 @@ static void load_vocab(const char *filename, Config *config, struct llama_vocab
 }
 
 static void convert_weights_ak_to_gg(struct ggml_tensor * gg_weights, const float * karpathy_weights) {
-    int ct;
-    switch (ggml_n_dims(gg_weights)) {
-        case 1:
-            ct = 0;
-            for (int i0 = 0; i0 < gg_weights->ne[0]; i0++){
-                float * ptr = (float *) ((char *) gg_weights->data + i0*gg_weights->nb[0]);
-                *ptr = karpathy_weights[ct];
-                ct++;
-            }
-            break;
-        case 2:
-            ct = 0;
-            for (int i1 = 0; i1 < gg_weights->ne[1]; i1++) {
-                for (int i0 = 0; i0 < gg_weights->ne[0]; i0++) {
-                    float * ptr = (float *) ((char *) gg_weights->data + i0*gg_weights->nb[0] + i1*gg_weights->nb[1]);
-                    *ptr = karpathy_weights[ct];
-                    ct++;
-                }
-            }
-            break;
-        case 3:
-            ct = 0;
-            for (int i2 = 0; i2 < gg_weights->ne[2]; i2++) {
-                for (int i1 = 0; i1 < gg_weights->ne[1]; i1++) {
-                    for (int i0 = 0; i0 < gg_weights->ne[0]; i0++) {
-                        float * ptr = (float *) ((char *) gg_weights->data + i0*gg_weights->nb[0] + i1*gg_weights->nb[1] + i2*gg_weights->nb[2]);
-                        *ptr = karpathy_weights[ct];
-                        ct++;
-                    }
-                }
-            }
-            break;
+    int size = 1;
+    for (int dim = 0; dim < ggml_n_dims(gg_weights); ++dim) {
+        size *= gg_weights->ne[dim];
+    }
+    for (int ct = 0; ct < size; ++ct) {
+        int64_t i0 = 0; int64_t i1 = 0;
+        int64_t i2 = 0; int64_t i3 = 0;
+        ggml_unravel_index(gg_weights, ct, &i0, &i1, &i2, &i3);
+        ggml_set_f32_nd(gg_weights, i0, i1, i2, i3, karpathy_weights[ct]);
     }
 }
 
@@ -679,16 +635,18 @@ static void save_as_llama_model(
     // convert AK weights into GG weights one by one.
     // w->token_embedding_table -> model->tok_embeddings
     // float*                   -> struct ggml_tensor
-    convert_weights_ak_to_gg(model->tok_embeddings, w->token_embedding_table);
-    convert_weights_ak_to_gg(model->output, w->wcls ? w->wcls : w->token_embedding_table);
+    convert_weights_ak_to_gg(model->tok_embeddings, w->token_embedding_table.data());
+    convert_weights_ak_to_gg(model->output, !w->wcls.empty() ? w->wcls.data() : w->token_embedding_table.data());
 
-    convert_weights_ak_to_gg(model->norm, w->rms_final_weight);
+    convert_weights_ak_to_gg(model->norm, w->rms_final_weight.data());
     //print_row(model->norm, 0);
 
     // for rms-att-weight
     int row_length = model->hparams.n_embd;
     int n_ff = model->hparams.n_ff;
 
+    const uint32_t n_multiqueries = model->hparams.n_head_kv <= 0 || model->hparams.n_head_kv >= model->hparams.n_head ? 1 : model->hparams.n_head / model->hparams.n_head_kv;
+
     for (uint32_t i = 0; i < model->hparams.n_layer; ++i){
         auto & layer = model->layers[i];
         // 1d
@@ -697,9 +655,10 @@ static void save_as_llama_model(
 
         // from 3d matrix layer x dim x dim to 2d matrix dim x dim
         convert_weights_ak_to_gg(layer.wq            , &w->wq[i*row_length*row_length]);
-        convert_weights_ak_to_gg(layer.wk            , &w->wk[i*row_length*row_length]);
-        convert_weights_ak_to_gg(layer.wv            , &w->wv[i*row_length*row_length]);
         convert_weights_ak_to_gg(layer.wo            , &w->wo[i*row_length*row_length]);
+        // from 3d matrix layer x dim x dim to 2d matrix dim x dim / n_multiqueries
+        convert_weights_ak_to_gg(layer.wk            , &w->wk[i*row_length*row_length/n_multiqueries]);
+        convert_weights_ak_to_gg(layer.wv            , &w->wv[i*row_length*row_length/n_multiqueries]);
 
         convert_weights_ak_to_gg(layer.w1            , &w->w1[i*row_length*n_ff]);
         convert_weights_ak_to_gg(layer.w2            , &w->w2[i*n_ff*row_length]);
@@ -736,8 +695,8 @@ static void save_as_llama_model(
     gguf_set_val_u32(ctx, KV_EMBEDDING_LENGTH, model->hparams.n_embd);
     gguf_set_val_u32(ctx, KV_FEED_FORWARD_LENGTH, model->hparams.n_ff);
     gguf_set_val_u32(ctx, KV_ATTENTION_HEAD_COUNT, model->hparams.n_head);
-    // n_head_kv is optional, default to n_head
-    // gguf_set_val_u32(ctx, KV_ATTENTION_HEAD_COUNT_KV, ...);
+    gguf_set_val_u32(ctx, KV_ATTENTION_HEAD_COUNT, model->hparams.n_head);
+    gguf_set_val_u32(ctx, KV_ATTENTION_HEAD_COUNT_KV, model->hparams.n_head_kv);
     gguf_set_val_u32(ctx, KV_BLOCK_COUNT, model->hparams.n_layer);
     gguf_set_val_u32(ctx, KV_ROPE_DIMENSION_COUNT, model->hparams.n_rot);
     gguf_set_val_f32(ctx, KV_ATTENTION_LAYERNORM_RMS_EPS, 1e-5f);
@@ -789,12 +748,12 @@ static void save_as_llama_model(
 
 static struct train_params get_default_train_params() {
     struct train_params params;
-    params.fn_vocab_model    = "models/7B/ggml-model-f16.gguf";
+    params.fn_vocab_model          = "models/7B/ggml-model-f16.gguf";
     params.fn_llama2c_output_model = "ak_llama_model.bin";
-    params.fn_train_data     = "shakespeare.txt";
-    params.fn_checkpoint_in  = "checkpoint.bin";
-    params.fn_checkpoint_out = "checkpoint.bin";
-    params.fn_model_out      = "ggml-checkpoint-f32.bin";
+    params.fn_train_data           = "shakespeare.txt";
+    params.fn_checkpoint_in        = "checkpoint.bin";
+    params.fn_checkpoint_out       = "checkpoint.bin";
+    params.fn_model_out            = "ggml-checkpoint-f32.bin";
 
     params.seed       =   -1;
 
@@ -829,8 +788,8 @@ static struct train_params get_default_train_params() {
     params.adam_alpha        = 1e-3f;
     params.adam_decay        = 1e-3f;
 
-    params.mem_model_gb   = 2;
-    params.mem_compute_gb = 24;
+    params.mem_model_gb    = 2;
+    params.mem_compute_gb  = 24;
     params.mem_compute0_gb = 8;
     params.mem_compute1_gb = 2;
 
@@ -916,19 +875,30 @@ int main(int argc, char ** argv) {
     if (!params_parse(argc, argv, &params)) {
         return 1;
     }
+    log_set_target(stdout);
     Config config;
     TransformerWeights weights = {};
     {
-        FILE *file = fopen(params.fn_llama2c_model, "rb");
-        if (!file) { printf("Unable to open the checkpoint file %s!\n", params.fn_llama2c_model); return 1; }
+        LOG("%s: Loading llama2c model from %s\n", __func__, params.fn_llama2c_model);
+        FILE *file = fopen(params.fn_llama2c_model, "r");
+        if (!file) {
+            LOG("%s: Unable to open the checkpoint file %s!\n", __func__, params.fn_llama2c_model);
+            return 1;
+        }
         // read in the config header
-        if(fread(&config, sizeof(Config), 1, file) != 1) { return 1; }
+        if (fread(&config, sizeof(Config), 1, file) != 1) {
+            LOG("%s: Unable to read llama2c config from %s!\n",__func__,params.fn_llama2c_model);
+            return 1;
+        }
         auto shared_weights = config.vocab_size > 0;
         config.vocab_size = abs(config.vocab_size);
 
         // read in the Transformer weights
-        malloc_weights(&weights, &config, shared_weights);
-        if(checkpoint_init_weights(&weights, &config, file, shared_weights)) { return 1; }
+        alloc_weights(&weights, &config, shared_weights);
+        if (checkpoint_init_weights(&weights, &config, file, shared_weights)) {
+            LOG("%s: Unable to initialize transformer weights from %s!",__func__,params.fn_llama2c_model);
+            return 1;
+        }
         fclose(file);
     }
 
@@ -936,15 +906,18 @@ int main(int argc, char ** argv) {
     load_vocab(params.fn_vocab_model, &config, &vocab);
 
     struct my_llama_model model;
-    model.hparams.n_vocab = config.vocab_size; //llama_n_vocab(lctx);
-    model.hparams.n_ctx   = params.n_ctx;
-    model.hparams.n_embd  = config.dim; //params.n_embd;
-    model.hparams.n_ff    = config.hidden_dim;
-    model.hparams.n_mult  = 32;//params.n_mult;
-    model.hparams.n_head  = config.n_heads; //params.n_head;
-    model.hparams.n_layer = config.n_layers; //params.n_layer;
-    model.hparams.n_rot   = std::min((uint32_t)params.n_rotmax, model.hparams.n_embd / model.hparams.n_head);
+    model.hparams.n_vocab   = config.vocab_size; //llama_n_vocab(lctx);
+    model.hparams.n_ctx     = params.n_ctx;
+    model.hparams.n_embd    = config.dim; //params.n_embd;
+    model.hparams.n_ff      = config.hidden_dim;
+    model.hparams.n_mult    = 32;//params.n_mult;
+    model.hparams.n_head    = config.n_heads; //params.n_head;
+    model.hparams.n_head_kv = config.n_kv_heads;
+    model.hparams.n_layer   = config.n_layers; //params.n_layer;
+    model.hparams.n_rot     = std::min((uint32_t)params.n_rotmax, model.hparams.n_embd / model.hparams.n_head);
+
     print_params(&model.hparams);
+
     struct ggml_init_params lcparams;
     lcparams.mem_size   = 1024ll*1024ll*1024ll*((size_t) params.mem_model_gb);
     lcparams.mem_buffer = NULL;
@@ -956,7 +929,7 @@ int main(int argc, char ** argv) {
     model.name = basename(params.fn_llama2c_model);
     save_as_llama_model(&vocab, &model, &weights, params.fn_llama2c_output_model);
 
-    printf("Saving llama.c model file %s in ggml format at %s\n", params.fn_llama2c_model, params.fn_llama2c_output_model);
+    LOG("%s: Saving llama.c model file %s in ggml format at %s\n", __func__, params.fn_llama2c_model, params.fn_llama2c_output_model);
 
     ggml_free(model.ctx);
     return 0;

examples/export-lora/export-lora.cpp

@@ -7,8 +7,6 @@
 #include <string>
 #include <thread>
 
-static const size_t tensor_alignment = 32;
-
 struct lora_info {
     std::string filename;
     float scale;

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_ADAM);
+    opt->params = ggml_opt_default_params(GGML_OPT_TYPE_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_shift(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_add(ctx, 0, params.n_keep + 1 + n_discard, n_past, -n_discard);
 
                 n_past -= n_discard;
 

examples/json-schema-to-grammar.py

@@ -87,7 +87,21 @@ 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')
-            rule = f'"[" space ({item_rule_name} ("," space {item_rule_name})*)? "]" space'
+            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'
             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_NONE:  return "none";
-        case LLAMA_SPLIT_LAYER: return "layer";
-        case LLAMA_SPLIT_ROW:   return "row";
+        case LLAMA_SPLIT_MODE_NONE:  return "none";
+        case LLAMA_SPLIT_MODE_LAYER: return "layer";
+        case LLAMA_SPLIT_MODE_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_LAYER},
+    /* split_mode    */ {LLAMA_SPLIT_MODE_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_NONE;
+                    mode = LLAMA_SPLIT_MODE_NONE;
                 } else if (m == "layer") {
-                    mode = LLAMA_SPLIT_LAYER;
+                    mode = LLAMA_SPLIT_MODE_LAYER;
                 } else if (m == "row") {
-                    mode = LLAMA_SPLIT_ROW;
+                    mode = LLAMA_SPLIT_MODE_ROW;
                 } else {
                     invalid_param = true;
                     break;

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

@@ -21,12 +21,8 @@ android {
             useSupportLibrary = true
         }
         ndk {
-            // 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")
-            }
+            // Add NDK properties if wanted, e.g.
+            // abiFilters += listOf("arm64-v8a")
         }
         externalNativeBuild {
             cmake {

examples/llava/README.md

@@ -53,20 +53,45 @@ python ./examples/llava/convert-image-encoder-to-gguf.py -m ../clip-vit-large-pa
 5. Use `convert.py` to convert the LLaMA part of LLaVA to GGUF:
 
 ```sh
-python ./convert.py ../llava-v1.5-7b
+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) 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:
+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/
+```
 - 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 (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`
+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
+```
 - 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) Everything else as usual: convert.py the hf model, quantize as needed
+
+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
+```
+
 **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_cont(ctx0, ggml_permute(ctx0, KQV, 0, 2, 1, 3));
+            KQV = ggml_permute(ctx0, KQV, 0, 2, 1, 3);
 
-            cur = ggml_cpy(ctx0, KQV, ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, hidden_size, num_positions, batch_size));
+            cur = ggml_cont_3d(ctx0, KQV, hidden_size, num_positions, batch_size);
         }
 
         // attention output

examples/llava/llava-surgery-v2.py

@@ -65,9 +65,7 @@ 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
 
@@ -152,16 +150,6 @@ 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-surgery.py

@@ -19,19 +19,12 @@ mm_tensors = [k for k, v in checkpoint.items() if k.startswith("model.mm_project
 projector = {name: checkpoint[name].float() for name in mm_tensors}
 torch.save(projector, f"{args.model}/llava.projector")
 
-# remove these tensors from the checkpoint and save it again
-for name in mm_tensors:
-    del checkpoint[name]
-
 # BakLLaVA models contain CLIP tensors in it
 clip_tensors = [k for k, v in checkpoint.items() if k.startswith("model.vision_tower")]
 if len(clip_tensors) > 0:
     clip = {name.replace("vision_tower.vision_tower.", ""): checkpoint[name].float() for name in clip_tensors}
     torch.save(clip, f"{args.model}/llava.clip")
 
-    # remove these tensors
-    for name in clip_tensors:
-        del checkpoint[name]
 
     # added tokens should be removed to be able to convert Mistral models
     if os.path.exists(f"{args.model}/added_tokens.json"):
@@ -39,7 +32,6 @@ if len(clip_tensors) > 0:
             f.write("{}\n")
 
 
-torch.save(checkpoint, path)
 
 print("Done!")
 print(f"Now you can convert {args.model} to a regular LLaMA 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_CPU) {
+    if (newline_tmp->backend != GGML_BACKEND_TYPE_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;
 }
 
-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) {
+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

@@ -31,6 +31,8 @@ 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

@@ -334,6 +334,8 @@ 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
@@ -383,8 +385,8 @@ int main(int argc, char ** argv) {
             }
         }
 
-        if (params.n_keep > 0) {
-        LOG_TEE("%s: static prompt based on n_keep: '", __func__);
+        if (params.n_keep > add_bos) {
+            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());
             }
@@ -540,14 +542,14 @@ int main(int argc, char ** argv) {
                         break;
                     }
 
-                    const int n_left    = n_past - params.n_keep - 1;
+                    const int n_left    = n_past - params.n_keep;
                     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 + 1            , params.n_keep + n_discard + 1);
-                    llama_kv_cache_seq_shift(ctx, 0, params.n_keep + 1 + n_discard, n_past, -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);
 
                     n_past -= n_discard;
 
@@ -574,9 +576,9 @@ 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_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);
+                    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);
 
                     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", __func__, n_len, n_ctx, n_kv_req, n_grp, n_batch);
+    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);
 
     // print the prompt token-by-token
 
@@ -146,10 +146,11 @@ 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_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);
+            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);
 
-            n_past -= bd;
+            n_past = llama_kv_cache_seq_pos_max(ctx, 0) + 1;
         }
 
         llama_batch_clear(batch);
@@ -179,10 +180,12 @@ 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_shift(ctx, 0, n_keep + n_discard, n_ctx,  -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);
 
-        n_past -= n_discard;
+        n_past = llama_kv_cache_seq_pos_max(ctx, 0) + 1;
 
         llama_batch_clear(batch);
 
@@ -208,10 +211,12 @@ 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_shift(ctx, 0, n_keep + n_discard, n_ctx,  -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);
 
-            n_past -= n_discard;
+            n_past = llama_kv_cache_seq_pos_max(ctx, 0) + 1;
         }
     }
 

examples/perplexity/perplexity.cpp

@@ -309,7 +309,7 @@ static void process_logits(int n_vocab, const float * logits, const int * tokens
 }
 
 static results_perplexity perplexity_v2(llama_context * ctx, const gpt_params & params) {
-    // Download: https://s3.amazonaws.com/research.metamind.io/wikitext/wikitext-2-raw-v1.zip?ref=salesforce-research
+    // Download: https://huggingface.co/datasets/ggml-org/ci/resolve/main/wikitext-2-raw-v1.zip
     // Run `./perplexity -m models/7B/ggml-model-q4_0.bin -f wiki.test.raw`
     // Output: `perplexity: 13.5106 [114/114]`
     // BOS tokens will be added for each chunk before eval
@@ -447,7 +447,7 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par
         return perplexity_v2(ctx, params);
     }
 
-    // Download: https://s3.amazonaws.com/research.metamind.io/wikitext/wikitext-2-raw-v1.zip?ref=salesforce-research
+    // Download: https://huggingface.co/datasets/ggml-org/ci/resolve/main/wikitext-2-raw-v1.zip
     // Run `./perplexity -m models/7B/ggml-model-q4_0.bin -f wiki.test.raw`
     // Output: `perplexity: 13.5106 [114/114]`
     // BOS tokens will be added for each chunk before eval
@@ -1623,7 +1623,7 @@ static void kl_divergence(llama_context * ctx, const gpt_params & params) {
     uint32_t n_ctx;
     in.read((char *)&n_ctx, sizeof(n_ctx));
     if (n_ctx > llama_n_ctx(ctx)) {
-        fprintf(stderr, "%s: %s has been computed with %d, while the current context is %d. Increase it with -c and retry\n",
+        fprintf(stderr, "%s: %s has been computed with %u, while the current context is %d. Increase it with -c and retry\n",
                 __func__, params.logits_file.c_str(), n_ctx, params.n_ctx);
     }
 

examples/quantize/quantize.cpp

@@ -23,14 +23,18 @@ static const std::vector<struct quant_option> QUANT_OPTIONS = {
     { "Q5_1",   LLAMA_FTYPE_MOSTLY_Q5_1,   " 4.70G, +0.0349 ppl @ LLaMA-v1-7B", },
     { "IQ2_XXS",LLAMA_FTYPE_MOSTLY_IQ2_XXS," 2.06 bpw quantization",            },
     { "IQ2_XS", LLAMA_FTYPE_MOSTLY_IQ2_XS, " 2.31 bpw quantization",            },
+    { "IQ1_S",  LLAMA_FTYPE_MOSTLY_IQ1_S,  " 1.56 bpw quantization",            },
     { "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", },
@@ -287,9 +291,10 @@ int main(int argc, char ** argv) {
         }
     }
 
-    if ((params.ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS || params.ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS || params.ftype == LLAMA_FTYPE_MOSTLY_Q2_K_S) && imatrix_data.empty()) {
+    if ((params.ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS || params.ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS ||
+         params.ftype == LLAMA_FTYPE_MOSTLY_Q2_K_S || params.ftype == LLAMA_FTYPE_MOSTLY_IQ1_S) && imatrix_data.empty()) {
         fprintf(stderr, "\n===============================================================================================\n");
-        fprintf(stderr, "Please do not use IQ2_XXS, IQ2_XS or Q2_K_S quantization without an importance matrix\n");
+        fprintf(stderr, "Please do not use IQ1_S, IQ2_XXS, IQ2_XS or Q2_K_S quantization without an importance matrix\n");
         fprintf(stderr, "===============================================================================================\n\n\n");
         return 1;
     }

examples/server/README.md

@@ -1,8 +1,20 @@
-# llama.cpp/example/server
+# LLaMA.cpp HTTP Server
 
-This example demonstrates a simple HTTP API server and a simple web front end to interact with llama.cpp.
+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**.
 
-Command line options:
+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:**
 
 - `--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.
@@ -39,6 +51,12 @@ 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)
+- `--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
 
@@ -95,6 +113,12 @@ 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.
@@ -132,9 +156,13 @@ node index.js
 ## API Endpoints
 
 - **GET** `/health`: Returns the current state of the server:
-  - `{"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.
+  - 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.
 
 - **POST** `/completion`: Given a `prompt`, it returns the predicted completion.
 
@@ -144,7 +172,7 @@ node index.js
 
     `temperature`: Adjust the randomness of the generated text (default: 0.8).
 
-    `dynatemp_range`: Dynamic temperature range (default: 0.0, 0.0 = disabled).
+    `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_exponent`: Dynamic temperature exponent (default: 1.0).
 
@@ -196,11 +224,13 @@ node index.js
 
     `n_probs`: If greater than 0, the response also contains the probabilities of top N tokens for each generated token (default: 0)
 
+    `min_keep`: If greater than 0, force samplers to return N possible tokens at minimum (default: 0)
+
     `image_data`: An array of objects to hold base64-encoded image `data` and its `id`s to be reference in `prompt`. You can determine the place of the image in the prompt as in the following: `USER:[img-12]Describe the image in detail.\nASSISTANT:`. In this case, `[img-12]` will be replaced by the embeddings of the image with id `12` in the following `image_data` array: `{..., "image_data": [{"data": "<BASE64_STRING>", "id": 12}]}`. Use `image_data` only with multimodal models, e.g., LLaVA.
 
     `slot_id`: Assign the completion task to an specific slot. If is -1 the task will be assigned to a Idle slot (default: -1)
 
-    `cache_prompt`: Save the prompt and generation for avoid reprocess entire prompt if a part of this isn't change (default: false)
+    `cache_prompt`: Re-use previously cached prompt from the last request if possible. This may prevent re-caching the prompt from scratch. (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)
 
@@ -233,7 +263,7 @@ 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`
+- `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.).
 - `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
@@ -379,6 +409,81 @@ Notice that each `probs` is an array of length `n_probs`.
     }'
     ```
 
+- **GET** `/slots`: Returns the current slots processing state. Can be disabled with `--slots-endpoint-disable`.
+
+### Result JSON
+
+```json
+[
+    {
+        "dynatemp_exponent": 1.0,
+        "dynatemp_range": 0.0,
+        "frequency_penalty": 0.0,
+        "grammar": "",
+        "id": 0,
+        "ignore_eos": false,
+        "logit_bias": [],
+        "min_p": 0.05000000074505806,
+        "mirostat": 0,
+        "mirostat_eta": 0.10000000149011612,
+        "mirostat_tau": 5.0,
+        "model": "llama-2-7b-32k-instruct.Q2_K.gguf",
+        "n_ctx": 2048,
+        "n_keep": 0,
+        "n_predict": 100000,
+        "n_probs": 0,
+        "next_token": {
+            "has_next_token": true,
+            "n_remain": -1,
+            "num_tokens_predicted": 0,
+            "stopped_eos": false,
+            "stopped_limit": false,
+            "stopped_word": false,
+            "stopping_word": ""
+        },
+        "penalize_nl": true,
+        "penalty_prompt_tokens": [],
+        "presence_penalty": 0.0,
+        "prompt": "Say hello to llama.cpp",
+        "repeat_last_n": 64,
+        "repeat_penalty": 1.100000023841858,
+        "samplers": [
+            "top_k",
+            "tfs_z",
+            "typical_p",
+            "top_p",
+            "min_p",
+            "temperature"
+        ],
+        "seed": 42,
+        "state": 1,
+        "stop": [
+            "\n"
+        ],
+        "stream": false,
+        "task_id": 0,
+        "temperature": 0.0,
+        "tfs_z": 1.0,
+        "top_k": 40,
+        "top_p": 0.949999988079071,
+        "typical_p": 1.0,
+        "use_penalty_prompt_tokens": false
+    }
+]
+```
+
+- **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

@@ -15,13 +15,11 @@
 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;
 
@@ -34,7 +32,7 @@ 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"]            = formatted_prompt;
+    llama_params["prompt"]            = format_chat(model, chat_template, body["messages"]);
     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/public/index.html

@@ -234,6 +234,7 @@
       mirostat_eta: 0.1, // learning rate
       grammar: '',
       n_probs: 0, // no completion_probabilities,
+      min_keep: 0, // min probs from each sampler,
       image_data: [],
       cache_prompt: true,
       api_key: ''
@@ -791,6 +792,9 @@
             <fieldset>
               ${IntField({ label: "Show Probabilities", max: 10, min: 0, name: "n_probs", value: params.value.n_probs })}
             </fieldset>
+            <fieldset>
+              ${IntField({ label: "Min Probabilities from each Sampler", max: 10, min: 0, name: "min_keep", value: params.value.min_keep })}
+            </fieldset>
             <fieldset>
               <label for="api_key">API Key</label>
               <input type="text" name="api_key" value="${params.value.api_key}" placeholder="Enter API key" oninput=${updateParams} />

examples/server/tests/README.md

@@ -0,0 +1,47 @@
+# 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

@@ -0,0 +1,69 @@
+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

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

examples/server/tests/features/parallel.feature

@@ -0,0 +1,123 @@
+@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

@@ -0,0 +1,50 @@
+@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

@@ -0,0 +1,84 @@
+@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

@@ -0,0 +1,803 @@
+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

@@ -0,0 +1,21 @@
+# 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

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

examples/server/tests/tests.sh

@@ -0,0 +1,12 @@
+#!/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

@@ -14,6 +14,7 @@
 using json = nlohmann::json;
 
 extern bool server_verbose;
+extern bool server_log_json;
 
 #ifndef SERVER_VERBOSE
 #define SERVER_VERBOSE 1
@@ -27,14 +28,14 @@ extern bool server_verbose;
     {                                                                    \
         if (server_verbose)                                              \
         {                                                                \
-            server_log("VERBOSE", __func__, __LINE__, MSG, __VA_ARGS__); \
+            server_log("VERB", __func__, __LINE__, MSG, __VA_ARGS__); \
         }                                                                \
     } while (0)
 #endif
 
-#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__)
+#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__)
 
 //
 // parallel
@@ -49,7 +50,8 @@ enum server_state {
 enum task_type {
     TASK_TYPE_COMPLETION,
     TASK_TYPE_CANCEL,
-    TASK_TYPE_NEXT_RESPONSE
+    TASK_TYPE_NEXT_RESPONSE,
+    TASK_TYPE_METRICS
 };
 
 struct task_server {
@@ -132,26 +134,48 @@ 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)
 {
-    nlohmann::ordered_json log
-    {
+    std::stringstream ss_tid;
+    ss_tid << std::this_thread::get_id();
+    json log = nlohmann::ordered_json{
+        {"tid", ss_tid.str()},
         {"timestamp", time(nullptr)},
-        {"level",     level},
-        {"function",  function},
-        {"line",      line},
-        {"message",   message},
     };
 
-    if (!extra.empty())
-    {
-        log.merge_patch(extra);
-    }
+    if (server_log_json) {
+        log.merge_patch(
+                {
+                        {"level",     level},
+                        {"function",  function},
+                        {"line",      line},
+                        {"msg",       message},
+                });
+        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);
+        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);
+    }
 }
 
 //
@@ -167,50 +191,47 @@ static T json_value(const json &body, const std::string &key, const T &default_v
         : default_value;
 }
 
-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();
+// 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_chatml(std::vector<json> messages)
+// 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)
 {
-    std::ostringstream chatml_msgs;
+    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());
 
-    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";
+    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();
     }
 
-    chatml_msgs << "<|im_start|>assistant" << '\n';
+    const char * ptr_tmpl = tmpl.empty() ? nullptr : tmpl.c_str();
+    std::vector<char> buf(alloc_size * 2);
 
-    LOG_VERBOSE("format_chatml", {{"text", chatml_msgs.str()}});
+    // 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());
 
-    return 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;
 }
 
 //
@@ -220,6 +241,7 @@ inline std::string format_chatml(std::vector<json> messages)
 struct llama_server_queue {
     int id = 0;
     std::mutex mutex_tasks;
+    bool running;
     // queues
     std::vector<task_server> queue_tasks;
     std::vector<task_server> queue_tasks_deferred;
@@ -235,6 +257,7 @@ 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();
@@ -250,7 +273,9 @@ struct llama_server_queue {
     // Get the next id for creating anew task
     int get_new_id() {
         std::unique_lock<std::mutex> lock(mutex_tasks);
-        return id++;
+        int new_id = id++;
+        LOG_VERBOSE("new task id", {{"new_id", new_id}});
+        return new_id;
     }
 
     // Register function to process a new task
@@ -278,12 +303,20 @@ struct llama_server_queue {
         queue_tasks_deferred.clear();
     }
 
-    // Start the main loop. This call is blocking
-    [[noreturn]]
+    // end the start_loop routine
+    void terminate() {
+        {
+            std::unique_lock<std::mutex> lock(mutex_tasks);
+            running = false;
+        }
+        condition_tasks.notify_all();
+    }
+
+    // Start the main loop.
     void start_loop() {
+        running = true;
         while (true) {
-            // new task arrived
-            LOG_VERBOSE("have new task", {});
+            LOG_VERBOSE("new task may arrive", {});
             {
                 while (true)
                 {
@@ -295,7 +328,7 @@ struct llama_server_queue {
                     task_server task = queue_tasks.front();
                     queue_tasks.erase(queue_tasks.begin());
                     lock.unlock();
-                    LOG_VERBOSE("callback_new_task", {});
+                    LOG_VERBOSE("callback_new_task", {{"task_id", task.id}});
                     callback_new_task(task);
                 }
                 LOG_VERBOSE("callback_all_task_finished", {});
@@ -324,8 +357,12 @@ struct llama_server_queue {
             {
                 std::unique_lock<std::mutex> lock(mutex_tasks);
                 if (queue_tasks.empty()) {
+                    if (!running) {
+                        LOG_VERBOSE("ending start_loop", {});
+                        return;
+                    }
                     condition_tasks.wait(lock, [&]{
-                        return !queue_tasks.empty();
+                        return (!queue_tasks.empty() || !running);
                     });
                 }
             }
@@ -372,11 +409,13 @@ 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);
     }
@@ -389,7 +428,6 @@ 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++)
             {
@@ -414,22 +452,22 @@ 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", {});
+        LOG_VERBOSE("send new result", {{"task_id", result.id}});
         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", {});
+                LOG_VERBOSE("callback_update_multitask", {{"task_id", task_id}});
                 callback_update_multitask(task_id, result.id, result);
                 continue;
             }
 
             if (result.id == task_id)
             {
-                LOG_VERBOSE("queue_results.push_back", {});
+                LOG_VERBOSE("queue_results.push_back", {{"task_id", task_id}});
                 queue_results.push_back(result);
-                condition_results.notify_one();
+                condition_results.notify_all();
                 return;
             }
         }

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

@@ -111,13 +111,13 @@ static const char * LLM_TENSOR_FFN_DOWN      = "blk.%d.ffn_down";
 static const char * LLM_TENSOR_FFN_UP        = "blk.%d.ffn_up";
 
 static void print_params(struct my_llama_hparams * params) {
-    printf("%s: n_vocab: %d\n", __func__, params->n_vocab);
-    printf("%s: n_ctx:   %d\n", __func__, params->n_ctx);
-    printf("%s: n_embd:  %d\n", __func__, params->n_embd);
-    printf("%s: n_head:  %d\n", __func__, params->n_head);
-    printf("%s: n_ff:    %d\n", __func__, params->n_ff);
-    printf("%s: n_layer: %d\n", __func__, params->n_layer);
-    printf("%s: n_rot:   %d\n", __func__, params->n_rot);
+    printf("%s: n_vocab: %u\n", __func__, params->n_vocab);
+    printf("%s: n_ctx:   %u\n", __func__, params->n_ctx);
+    printf("%s: n_embd:  %u\n", __func__, params->n_embd);
+    printf("%s: n_head:  %u\n", __func__, params->n_head);
+    printf("%s: n_ff:    %u\n", __func__, params->n_ff);
+    printf("%s: n_layer: %u\n", __func__, params->n_layer);
+    printf("%s: n_rot:   %u\n", __func__, params->n_rot);
 }
 
 static void set_param_model(struct my_llama_model * model) {
@@ -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_ADAM);
+    opt->params = ggml_opt_default_params(GGML_OPT_TYPE_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": 1707268954,
-        "narHash": "sha256-2en1kvde3cJVc3ZnTy8QeD2oKcseLFjYPLKhIGDanQ0=",
+        "lastModified": 1708655239,
+        "narHash": "sha256-ZrP/yACUvDB+zbqYJsln4iwotbH6CTZiTkANJ0AgDv4=",
         "owner": "NixOS",
         "repo": "nixpkgs",
-        "rev": "f8e2ebd66d097614d51a56a755450d4ae1632df1",
+        "rev": "cbc4211f0afffe6dfd2478a62615dd5175a13f9a",
         "type": "github"
       },
       "original": {

flake.nix

@@ -150,6 +150,7 @@
             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; };
@@ -157,7 +158,6 @@
 
                 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

@@ -377,6 +377,9 @@ 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) {
@@ -427,6 +430,7 @@ 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);
 }
 
@@ -464,7 +468,7 @@ 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) {
-                    break;
+                    continue;
                 }
 
                 // if the node's data is external, then we cannot re-use it
@@ -544,22 +548,8 @@ 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) {
-                break;
-            }
-            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];
 
@@ -568,14 +558,37 @@ static void ggml_gallocr_alloc_graph_impl(ggml_gallocr_t galloc, struct ggml_cgr
             ggml_gallocr_hash_get(galloc, view_src)->n_views += 1;
         }
 
-        for (int j = 0; j < GGML_MAX_SRC; j++) {
-            struct ggml_tensor * parent = node->src[j];
-            if (parent == NULL) {
-                break;
-            }
-            ggml_gallocr_hash_get(galloc, parent)->n_children += 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));
+            }
+        }
+    }
+
+    // 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++) {
@@ -586,7 +599,7 @@ 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) {
-                break;
+                continue;
             }
             ggml_gallocr_allocate_node(galloc, parent, buffer_id);
         }
@@ -598,7 +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) {
-                break;
+                continue;
             }
             AT_PRINTF("%s", parent->name);
             if (j < GGML_MAX_SRC - 1 && node->src[j + 1] != NULL) {
@@ -611,7 +624,7 @@ 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) {
-                break;
+                continue;
             }
             struct hash_node * p_hn = ggml_gallocr_hash_get(galloc, parent);
             p_hn->n_children -= 1;
@@ -696,6 +709,18 @@ 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++) {
@@ -722,8 +747,8 @@ 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, 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);
+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);
 
     if (node->view_src != NULL) {
         if (node->buffer == NULL) {
@@ -732,29 +757,20 @@ 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[node_alloc->buffer_id], node);
+            ggml_backend_view_init(galloc->buffers[buffer_id], node);
         }
     } else {
         if (node->data == NULL) {
             assert(tensor_alloc->offset != SIZE_MAX);
-            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]);
+            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]);
             void * addr = (char *)base + tensor_alloc->offset;
-            ggml_backend_tensor_alloc(galloc->buffers[node_alloc->buffer_id], node, addr);
+            ggml_backend_tensor_alloc(galloc->buffers[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
         }
     }
 }
@@ -773,6 +789,13 @@ 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];
@@ -787,7 +810,7 @@ static bool ggml_gallocr_needs_realloc(ggml_gallocr_t galloc, struct ggml_cgraph
         for (int j = 0; j < GGML_MAX_SRC; j++) {
             struct ggml_tensor * src = node->src[j];
             if (src == NULL) {
-                break;
+                continue;
             }
             if (!ggml_gallocr_node_needs_realloc(galloc, src, node_alloc, &node_alloc->src[j])) {
 #ifndef NDEBUG
@@ -827,17 +850,24 @@ 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];
         for (int j = 0; j < GGML_MAX_SRC; j++) {
             struct ggml_tensor * src = node->src[j];
             if (src == NULL) {
-                break;
+                continue;
             }
-            ggml_gallocr_init_tensor(galloc, src, node_alloc, &node_alloc->src[j]);
+            ggml_gallocr_init_tensor(galloc, src, node_alloc->buffer_id, &node_alloc->src[j]);
         }
-        ggml_gallocr_init_tensor(galloc, node, node_alloc, &node_alloc->dst);
+        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);
     }
 
     return true;

ggml-backend.c

@@ -756,7 +756,7 @@ GGML_CALL static bool ggml_backend_cpu_graph_compute(ggml_backend_t backend, str
 GGML_CALL static bool ggml_backend_cpu_supports_op(ggml_backend_t backend, const struct ggml_tensor * op) {
     switch (op->op) {
         case GGML_OP_CPY:
-            return op->type != GGML_TYPE_IQ2_XXS && op->type != GGML_TYPE_IQ2_XS; // missing type_traits.from_float
+            return op->type != GGML_TYPE_IQ2_XXS && op->type != GGML_TYPE_IQ2_XS && op->type != GGML_TYPE_IQ1_S; // missing type_traits.from_float
         case GGML_OP_MUL_MAT:
             return op->src[1]->type == GGML_TYPE_F32 || op->src[1]->type == ggml_internal_get_type_traits(op->src[0]->type).vec_dot_type;
         default:
@@ -1006,6 +1006,7 @@ static int ggml_backend_sched_backend_from_buffer(ggml_backend_sched_t sched, gg
         }
     }
     GGML_ASSERT(false && "tensor buffer type not supported by any backend");
+    return -1; // silence warning
 }
 
 #if 0
@@ -1040,7 +1041,7 @@ static int ggml_backend_sched_backend_id_from_cur(ggml_backend_sched_t sched, st
     for (int i = 0; i < GGML_MAX_SRC; i++) {
         const struct ggml_tensor * src = tensor->src[i];
         if (src == NULL) {
-            break;
+            continue;
         }
         if (src->buffer != NULL && src->buffer->usage == GGML_BACKEND_BUFFER_USAGE_WEIGHTS) {
             int src_backend = ggml_backend_sched_backend_from_buffer(sched, src->buffer);
@@ -1087,7 +1088,7 @@ static void ggml_backend_sched_print_assignments(ggml_backend_sched_t sched, str
         for (int j = 0; j < GGML_MAX_SRC; j++) {
             struct ggml_tensor * src = node->src[j];
             if (src == NULL) {
-                break;
+                continue;
             }
             ggml_backend_t src_backend = tensor_backend(src);
             fprintf(stderr, " %20.20s (%5.5s) [%5.5s %8.8s]", src->name,
@@ -1143,7 +1144,7 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
         for (int j = 0; j < GGML_MAX_SRC; j++) {
             struct ggml_tensor * src = node->src[j];
             if (src == NULL) {
-                break;
+                continue;
             }
             if (tensor_backend_id(src) == -1) {
                 tensor_backend_id(src) = ggml_backend_sched_backend_id_from_cur(sched, src);
@@ -1255,7 +1256,7 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
         for (int j = 0; j < GGML_MAX_SRC; j++) {
             struct ggml_tensor * src = node->src[j];
             if (src == NULL) {
-                break;
+                continue;
             }
             int src_backend_id = tensor_backend_id(src);
             if (src_backend_id == -1) {
@@ -1314,7 +1315,7 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
             for (int j = 0; j < GGML_MAX_SRC; j++) {
                 struct ggml_tensor * src = node->src[j];
                 if (src == NULL) {
-                    break;
+                    continue;
                 }
                 int src_backend_id = tensor_backend_id(src);
                 assert(src_backend_id != -1); // all inputs should be assigned by now
@@ -1361,7 +1362,7 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
         for (int j = 0; j < GGML_MAX_SRC; j++) {
             struct ggml_tensor * src = node->src[j];
             if (src == NULL) {
-                break;
+                continue;
             }
             ggml_backend_t src_backend = tensor_backend(src);
             if (src_backend != tensor_backend /* && src_backend != NULL */) {
@@ -1667,7 +1668,7 @@ static struct ggml_tensor * graph_copy_dup_tensor(struct ggml_hash_set hash_set,
     for (int i = 0; i < GGML_MAX_SRC; i++) {
         struct ggml_tensor * s = src->src[i];
         if (s == NULL) {
-            break;
+            continue;
         }
         dst->src[i] = graph_copy_dup_tensor(hash_set, node_copies, ctx_allocated, ctx_unallocated, s);
     }
@@ -1696,7 +1697,7 @@ static void graph_copy_init_tensor(struct ggml_hash_set hash_set, struct ggml_te
     for (int i = 0; i < GGML_MAX_SRC; i++) {
         struct ggml_tensor * s = src->src[i];
         if (s == NULL) {
-            break;
+            continue;
         }
         graph_copy_init_tensor(hash_set, node_copies, node_init, s);
     }

ggml-impl.h

@@ -53,11 +53,23 @@ extern "C" {
 //
 #include <arm_neon.h>
 
-#define GGML_COMPUTE_FP16_TO_FP32(x) ((float) (x))
-#define GGML_COMPUTE_FP32_TO_FP16(x) (x)
+#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_FP16_TO_FP32(x) ((float) (x))
-#define GGML_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;
+}
 
 #else
 
@@ -214,8 +226,7 @@ extern float ggml_table_f32_f16[1 << 16];
 // On ARM NEON, it's quicker to directly convert x -> x instead of calling into ggml_lookup_fp16_to_fp32,
 // so we define GGML_FP16_TO_FP32 and GGML_FP32_TO_FP16 elsewhere for NEON.
 // This is also true for POWER9.
-#if !defined(GGML_FP16_TO_FP32) || !defined(GGML_FP32_TO_FP16)
-
+#if !defined(GGML_FP16_TO_FP32)
 inline static float ggml_lookup_fp16_to_fp32(ggml_fp16_t f) {
     uint16_t s;
     memcpy(&s, &f, sizeof(uint16_t));
@@ -223,8 +234,10 @@ inline static float ggml_lookup_fp16_to_fp32(ggml_fp16_t f) {
 }
 
 #define GGML_FP16_TO_FP32(x) ggml_lookup_fp16_to_fp32(x)
-#define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x)
+#endif
 
+#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

@@ -61,6 +61,9 @@ 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,
@@ -83,6 +86,9 @@ 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,
@@ -101,6 +107,9 @@ 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,
@@ -116,6 +125,9 @@ 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,
@@ -131,6 +143,9 @@ 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,
@@ -176,7 +191,7 @@ struct ggml_metal_context {
 // MSL code
 // TODO: move the contents here when ready
 //       for now it is easier to work in a separate file
-//static NSString * const msl_library_source = @"see metal.metal";
+// static NSString * const msl_library_source = @"see metal.metal";
 
 // Here to assist with NSBundle Path Hack
 @interface GGMLMetalClass : NSObject
@@ -272,6 +287,14 @@ 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;
@@ -294,6 +317,7 @@ 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
@@ -433,6 +457,9 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
         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);
@@ -455,6 +482,9 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
         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);
@@ -473,6 +503,9 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
         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);
@@ -488,6 +521,9 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
         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);
@@ -503,6 +539,9 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
         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);
@@ -728,6 +767,7 @@ static bool ggml_metal_graph_compute(
 
         size_t offs_src0 = 0;
         size_t offs_src1 = 0;
+        size_t offs_src2 = 0;
         size_t offs_dst  = 0;
 
         id<MTLCommandBuffer> command_buffer  = command_buffers[cb_idx];
@@ -746,6 +786,7 @@ static bool ggml_metal_graph_compute(
 
             struct ggml_tensor * src0 = gf->nodes[i]->src[0];
             struct ggml_tensor * src1 = gf->nodes[i]->src[1];
+            struct ggml_tensor * src2 = gf->nodes[i]->src[2];
             struct ggml_tensor * dst  = gf->nodes[i];
 
             switch (dst->op) {
@@ -807,6 +848,7 @@ static bool ggml_metal_graph_compute(
 
             id<MTLBuffer> id_src0 = src0 ? ggml_metal_get_buffer(src0, &offs_src0) : nil;
             id<MTLBuffer> id_src1 = src1 ? ggml_metal_get_buffer(src1, &offs_src1) : nil;
+            id<MTLBuffer> id_src2 = src2 ? ggml_metal_get_buffer(src2, &offs_src2) : nil;
             id<MTLBuffer> id_dst  = dst  ? ggml_metal_get_buffer(dst,  &offs_dst)  : nil;
 
             //GGML_METAL_LOG_INFO("%s: op - %s\n", __func__, ggml_op_name(dst->op));
@@ -1188,7 +1230,16 @@ static bool ggml_metal_graph_compute(
                             pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SOFT_MAX].pipeline;
                         }
 
-                        const float scale = ((float *) dst->op_params)[0];
+                        const float scale    = ((float *) dst->op_params)[0];
+                        const float max_bias = ((float *) dst->op_params)[1];
+
+                        const int64_t nrows_x = ggml_nrows(src0);
+                        const int64_t nrows_y = src0->ne[1];
+                        const uint32_t n_head_kv   = nrows_x/nrows_y;
+                        const uint32_t n_head_log2 = 1u << (uint32_t) floorf(log2f((float) n_head_kv));
+
+                        const float m0 = powf(2.0f, -(max_bias       ) / n_head_log2);
+                        const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2);
 
                         [encoder setComputePipelineState:pipeline];
                         [encoder setBuffer:id_src0 offset:offs_src0   atIndex:0];
@@ -1197,11 +1248,20 @@ static bool ggml_metal_graph_compute(
                         } else {
                             [encoder setBuffer:id_src0 offset:offs_src0   atIndex:1];
                         }
-                        [encoder setBuffer:id_dst  offset:offs_dst    atIndex:2];
-                        [encoder setBytes:&ne00  length:sizeof(ne00)  atIndex:3];
-                        [encoder setBytes:&ne01  length:sizeof(ne01)  atIndex:4];
-                        [encoder setBytes:&ne02  length:sizeof(ne02)  atIndex:5];
-                        [encoder setBytes:&scale length:sizeof(scale) atIndex:6];
+                        if (id_src2) {
+                            [encoder setBuffer:id_src2 offset:offs_src2   atIndex:2];
+                        } else {
+                            [encoder setBuffer:id_src0 offset:offs_src0   atIndex:2];
+                        }
+                        [encoder setBuffer:id_dst   offset:offs_dst          atIndex:3];
+                        [encoder setBytes:&ne00     length:sizeof(ne00)      atIndex:4];
+                        [encoder setBytes:&ne01     length:sizeof(ne01)      atIndex:5];
+                        [encoder setBytes:&ne02     length:sizeof(ne02)      atIndex:6];
+                        [encoder setBytes:&scale    length:sizeof(scale)     atIndex:7];
+                        [encoder setBytes:&max_bias length:sizeof(max_bias)  atIndex:8];
+                        [encoder setBytes:&m0       length:sizeof(m0)        atIndex:9];
+                        [encoder setBytes:&m1       length:sizeof(m1)        atIndex:10];
+                        [encoder setBytes:&n_head_log2 length:sizeof(n_head_log2) atIndex:11];
                         [encoder setThreadgroupMemoryLength:32*sizeof(float) atIndex:0];
 
                         [encoder dispatchThreadgroups:MTLSizeMake(ne01*ne02*ne03, 1, 1) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
@@ -1297,6 +1357,9 @@ 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");
                             }
 
@@ -1431,6 +1494,24 @@ 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);
@@ -1465,7 +1546,7 @@ static bool ggml_metal_graph_compute(
 
                             if (src0t == GGML_TYPE_Q4_0 || src0t == GGML_TYPE_Q4_1 ||
                                 src0t == GGML_TYPE_Q5_0 || src0t == GGML_TYPE_Q5_1 || src0t == GGML_TYPE_Q8_0 ||
-                                src0t == GGML_TYPE_Q2_K) { // || src0t == GGML_TYPE_Q4_K) {
+                                src0t == GGML_TYPE_Q2_K || src0t == GGML_TYPE_IQ1_S) { // || src0t == GGML_TYPE_Q4_K) {
                                 [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
                             }
                             else if (src0t == GGML_TYPE_IQ2_XXS || src0t == GGML_TYPE_IQ2_XS) {
@@ -1473,11 +1554,16 @@ 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) {
-                                const int mem_size = 256*4+128;
+                            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;
                                 [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)];
                             }
@@ -1514,8 +1600,6 @@ static bool ggml_metal_graph_compute(
                         // max size of the src1ids array in the kernel stack
                         GGML_ASSERT(ne11 <= 512);
 
-                        struct ggml_tensor * src2 = gf->nodes[i]->src[2];
-
                         const int64_t  ne20 = src2 ? src2->ne[0] : 0;
                         const int64_t  ne21 = src2 ? src2->ne[1] : 0;
                         const int64_t  ne22 = src2 ? src2->ne[2] : 0;
@@ -1573,6 +1657,9 @@ 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");
                             }
 
@@ -1710,6 +1797,24 @@ 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);
@@ -1760,7 +1865,7 @@ static bool ggml_metal_graph_compute(
 
                             if (src2t == GGML_TYPE_Q4_0 || src2t == GGML_TYPE_Q4_1 ||
                                 src2t == GGML_TYPE_Q5_0 || src2t == GGML_TYPE_Q5_1 || src2t == GGML_TYPE_Q8_0 ||
-                                src2t == GGML_TYPE_Q2_K) { // || src2t == GGML_TYPE_Q4_K) {
+                                src2t == GGML_TYPE_Q2_K || src2t == GGML_TYPE_IQ1_S) { // || src2t == GGML_TYPE_Q4_K) {
                                 [encoder dispatchThreadgroups:MTLSizeMake((ne21 + 7)/8, _ne1, ne01*ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
                             }
                             else if (src2t == GGML_TYPE_IQ2_XXS || src2t == GGML_TYPE_IQ2_XS) {
@@ -1768,11 +1873,16 @@ 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) {
-                                const int mem_size = 256*4+128;
+                            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;
                                 [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)];
                             }
@@ -1814,6 +1924,9 @@ 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");
                         }
@@ -2149,8 +2262,8 @@ static bool ggml_metal_graph_compute(
                         id<MTLComputePipelineState> pipeline = nil;
 
                         switch (order) {
-                            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;
+                            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;
                             default: GGML_ASSERT(false);
                         };
 

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_GPU) {
+    if (tensor->backend != GGML_BACKEND_TYPE_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_GPU);
+    GGML_ASSERT(src1->backend == GGML_BACKEND_TYPE_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_GPU);
+    GGML_ASSERT(src1->backend == GGML_BACKEND_TYPE_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_GPU) { // NOLINT
+    if (src0->backend == GGML_BACKEND_TYPE_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_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);
+    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);
 
     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_GPU) {
+                if (src0->backend == GGML_BACKEND_TYPE_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_CPU) {
+                    if (src1->backend == GGML_BACKEND_TYPE_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_CPU) {
+                    if (dst->backend == GGML_BACKEND_TYPE_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_GPU) {
+    if (src0->backend != GGML_BACKEND_TYPE_GPU) {
         ggml_cl_pool_free(d_X, x_size);
     }
-    if (src1->backend != GGML_BACKEND_GPU) {
+    if (src1->backend != GGML_BACKEND_TYPE_GPU) {
         ggml_cl_pool_free(d_Y, y_size);
     }
-    if (dst->backend != GGML_BACKEND_GPU) {
+    if (dst->backend != GGML_BACKEND_TYPE_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_GPU) { // NOLINT
+    if (src0->backend == GGML_BACKEND_TYPE_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_GPU) {
+                if (src0->backend == GGML_BACKEND_TYPE_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_CPU) {
+                    if (dst->backend == GGML_BACKEND_TYPE_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_GPU) {
+    if (src0->backend != GGML_BACKEND_TYPE_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_CPU) {
+    if (src0->backend == GGML_BACKEND_TYPE_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_CPU) {
+                if (src0->backend == GGML_BACKEND_TYPE_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_GPU) {
+                } else if (src0->backend == GGML_BACKEND_TYPE_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_GPU ? (i03 * ne02 + i02) * x_bps : 0;
+                    const size_t offset = src0->backend == GGML_BACKEND_TYPE_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_GPU ? (i03 * ne02 + i02) * x_bps : 0;
+                        const size_t offset = src0->backend == GGML_BACKEND_TYPE_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_CPU) {
+    if (src0->backend == GGML_BACKEND_TYPE_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_GPU)) {
+        ((ne0 >= 32 && ne1 >= 32 && ne10 >= 32) || src0->backend == GGML_BACKEND_TYPE_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_GPU);
+    GGML_ASSERT(tensor->backend == GGML_BACKEND_TYPE_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_GPU;
+    tensor->backend = GGML_BACKEND_TYPE_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

@@ -191,6 +191,36 @@ 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];
+    uint8_t scales[QK_K/16];
+} 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
@@ -210,6 +240,8 @@ 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);
@@ -225,6 +257,8 @@ 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);
@@ -243,6 +277,9 @@ void dequantize_row_q8_K(const block_q8_K * GGML_RESTRICT x, float * GGML_RESTRI
 void dequantize_row_iq2_xxs(const block_iq2_xxs * GGML_RESTRICT x, float * GGML_RESTRICT y, int k);
 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);
@@ -259,6 +296,9 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
 void ggml_vec_dot_iq2_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_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")
@@ -266,6 +306,9 @@ void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const
 size_t quantize_iq2_xxs(const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix);
 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);
@@ -276,8 +319,8 @@ size_t quantize_q4_1   (const float * src, void * dst, int nrows, int n_per_row,
 size_t quantize_q5_0   (const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix);
 size_t quantize_q5_1   (const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix);
 
-void iq2xs_init_impl(int grid_size);
-void iq2xs_free_impl(int grid_size);
+void iq2xs_init_impl(enum ggml_type type);
+void iq2xs_free_impl(enum ggml_type type);
 void iq3xs_init_impl(int grid_size);
 void iq3xs_free_impl(int grid_size);
 

ggml-vulkan.cpp

@@ -1091,7 +1091,10 @@ static void ggml_vk_print_gpu_info(size_t idx) {
     }
 }
 
-static void ggml_vk_instance_init() {
+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() {
     if (vk_instance_initialized) {
         return;
     }
@@ -1100,28 +1103,42 @@ static 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);
 
-    std::cerr << "ggml_vulkan: Validation layers enabled" << std::endl;
-#endif
+    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;
+    }
     vk_instance.instance = vk::createInstance(instance_create_info);
 
     memset(vk_instance.initialized, 0, sizeof(bool) * GGML_VK_MAX_DEVICES);
@@ -1168,12 +1185,12 @@ 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];
-    std::vector<vk::ExtensionProperties> ext_props = ctx->device.lock()->physical_device.enumerateDeviceExtensionProperties();
+    const std::vector<vk::ExtensionProperties> ext_props = ctx->device.lock()->physical_device.enumerateDeviceExtensionProperties();
 
     bool maintenance4_support = false;
 
     // Check if maintenance4 is supported
-    for (auto properties : ext_props) {
+    for (const auto& properties : ext_props) {
         if (strcmp("VK_KHR_maintenance4", properties.extensionName) == 0) {
             maintenance4_support = true;
         }
@@ -1204,7 +1221,7 @@ static void ggml_vk_init(ggml_backend_vk_context * ctx, size_t idx) {
     bool fp16_storage = false;
     bool fp16_compute = false;
 
-    for (auto properties : ext_props) {
+    for (const 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) {
@@ -2303,8 +2320,8 @@ 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_GPU && !src0_uma;
-    const bool load_y = src1->backend != GGML_BACKEND_GPU && !src1_uma;
+    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 x_non_contig = !load_x && !ggml_vk_dim01_contiguous(src0);
     const bool y_non_contig = !load_y && !ggml_vk_dim01_contiguous(src1);
@@ -2436,7 +2453,7 @@ 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_CPU) {
+    if (dst->backend == GGML_BACKEND_TYPE_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);
@@ -2489,8 +2506,8 @@ 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_GPU && !src0_uma;
-    const bool load_y = src1->backend != GGML_BACKEND_GPU && !src1_uma;
+    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 x_non_contig = !load_x && !ggml_vk_dim01_contiguous(src0);
     const bool y_non_contig = !load_y && !ggml_vk_dim01_contiguous(src1);
@@ -2613,7 +2630,7 @@ 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_CPU) {
+            if (dst->backend == GGML_BACKEND_TYPE_CPU) {
                 // copy dst to host
                 float * d = (float *) ((char *) dst->data + i12*nb2 + i13*nb3);
                 ggml_vk_sync_buffers(subctx);
@@ -2630,7 +2647,7 @@ 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_GPU);
+    GGML_ASSERT(src0->backend == GGML_BACKEND_TYPE_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);
@@ -2662,7 +2679,7 @@ 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_GPU && !src1_uma;
+    const bool load_y = src1->backend != GGML_BACKEND_TYPE_GPU && !src1_uma;
 
     const uint64_t x_ne = ne00 * ne01 * ne02;
     const uint64_t y_ne = ne10 * ne11 * ne12;
@@ -2704,7 +2721,7 @@ 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_CPU) {
+    if (dst->backend == GGML_BACKEND_TYPE_CPU) {
         // copy dst to host
         float * d = (float *) dst->data;
         ggml_vk_sync_buffers(subctx);
@@ -2721,7 +2738,7 @@ 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_GPU);
+    GGML_ASSERT(src0->backend == GGML_BACKEND_TYPE_GPU);
     GGML_ASSERT(src0->type == GGML_TYPE_F16);
     GGML_ASSERT(src1->type == GGML_TYPE_F32);
 
@@ -2754,7 +2771,7 @@ 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_GPU && !src1_uma;
+    const bool load_y = src1->backend != GGML_BACKEND_TYPE_GPU && !src1_uma;
 
     const uint64_t d_ne = ne01 * ne11 * ne12;
 
@@ -2797,7 +2814,7 @@ 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_CPU) {
+    if (dst->backend == GGML_BACKEND_TYPE_CPU) {
         // copy dst to host
         float * d = (float *) dst->data;
         ggml_vk_sync_buffers(subctx);
@@ -2815,7 +2832,7 @@ 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_GPU);
+           ((ne0 >= 32 && ne1 >= 32 && ne10 >= 32) || src0->backend == GGML_BACKEND_TYPE_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) {
@@ -2863,8 +2880,8 @@ 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_GPU);
-    GGML_ASSERT(dst->backend == GGML_BACKEND_GPU);
+    GGML_ASSERT(src0->backend == GGML_BACKEND_TYPE_GPU);
+    GGML_ASSERT(dst->backend == GGML_BACKEND_TYPE_GPU);
 
     ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) dst->extra;
     ggml_tensor_extra_gpu * extra_src0 = (ggml_tensor_extra_gpu *) src0->extra;
@@ -3093,8 +3110,8 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context * subctx, c
         }
     }
 
-    const bool transfer_src0 = src0->backend != GGML_BACKEND_GPU && !src0_uma;
-    const bool transfer_src1 = use_src1 && src1->backend != GGML_BACKEND_GPU && !src1_uma;
+    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;
 
     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;
@@ -3103,7 +3120,7 @@ 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_GPU && y_sz > d_D->size) {
+    if (use_src1 && src1->backend == GGML_BACKEND_TYPE_GPU && y_sz > d_D->size) {
         y_sz = VK_WHOLE_SIZE;
     }
 
@@ -3192,9 +3209,9 @@ 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_CPU && op == GGML_OP_CPY) {
+        if (dst->backend == GGML_BACKEND_TYPE_CPU && op == GGML_OP_CPY) {
             ggml_vk_d2h_tensor_2d(ctx, subctx, d_D, 0, dst);
-        } else if(dst->backend == GGML_BACKEND_CPU) {
+        } else if(dst->backend == GGML_BACKEND_TYPE_CPU) {
             // copy dst to host
             float * d = (float *) dst->data;
             ggml_vk_buffer_read_async(ctx, subctx, d_D, 0, d, d_sz);
@@ -3236,7 +3253,7 @@ 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_CPU) {
+                if (dst->backend == GGML_BACKEND_TYPE_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);
                 }
@@ -3342,7 +3359,7 @@ 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_CPU) {
+    if (dst->backend == GGML_BACKEND_TYPE_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);
@@ -3977,9 +3994,9 @@ 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_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));
+    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));
 
     if (ctx->disable || (!any_on_device && node->op != GGML_OP_MUL_MAT)) {
         return;
@@ -4198,9 +4215,9 @@ 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_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);
+    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);
 
     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;
@@ -4354,7 +4371,7 @@ static void ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod
     last_node = true;
 #endif
 
-    if (node->backend == GGML_BACKEND_CPU || last_node) {
+    if (node->backend == GGML_BACKEND_TYPE_CPU || last_node) {
         ggml_vk_ctx_end(ctx->compute_ctx);
         ctx->compute_ctx->exit_tensor = node;
         ctx->compute_ctx = nullptr;
@@ -4362,9 +4379,9 @@ 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_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);
+    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);
 
     if (ctx->disable || (!any_on_device && tensor->op != GGML_OP_MUL_MAT)) {
         return false;
@@ -4425,7 +4442,7 @@ static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_compute_
     if (params->ith != 0) {
         return true;
     }
-    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
+    if (params->type == GGML_TASK_TYPE_INIT || params->type == GGML_TASK_TYPE_FINALIZE) {
         return true;
     }
 
@@ -4728,7 +4745,7 @@ 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_GPU;
+    tensor->backend = GGML_BACKEND_TYPE_GPU;
     tensor->extra = extra;
 }
 
@@ -4736,7 +4753,7 @@ 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_GPU);
+    GGML_ASSERT(tensor->backend == GGML_BACKEND_TYPE_GPU);
 
     ggml_backend_vk_buffer_context * ctx = (ggml_backend_vk_buffer_context *)buffer->context;
 
@@ -4751,7 +4768,7 @@ 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_GPU);
+    GGML_ASSERT(tensor->backend == GGML_BACKEND_TYPE_GPU);
 
     ggml_backend_vk_buffer_context * ctx = (ggml_backend_vk_buffer_context *)buffer->context;
 
@@ -4982,7 +4999,7 @@ 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_GPU);
+    GGML_ASSERT(tensor->backend == GGML_BACKEND_TYPE_GPU);
 
     ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) tensor->extra;
 
@@ -5003,7 +5020,7 @@ 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_GPU);
+    GGML_ASSERT(tensor->backend == GGML_BACKEND_TYPE_GPU);
 
     ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) tensor->extra;
 
@@ -5080,7 +5097,7 @@ 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_GPU) {
+    while (last_node > 0 && cgraph->nodes[last_node]->backend != GGML_BACKEND_TYPE_GPU) {
         last_node -= 1;
     }
 
@@ -5089,7 +5106,7 @@ GGML_CALL static bool ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml
     }
 
     ggml_compute_params params = {};
-    params.type = GGML_TASK_COMPUTE;
+    params.type = GGML_TASK_TYPE_COMPUTE;
     params.ith = 0;
     for (int i = 0; i < cgraph->n_nodes; i++) {
         ggml_tensor * node = cgraph->nodes[i];
@@ -5301,6 +5318,42 @@ 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
@@ -5357,7 +5410,7 @@ 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_GPU) {
+    if (tensor->backend == GGML_BACKEND_TYPE_GPU) {
         const size_t tensor_size = ggml_nbytes(tensor);
         tensor_data = malloc(tensor_size);
 
@@ -5383,14 +5436,14 @@ 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_GPU) {
+    if (tensor->backend == GGML_BACKEND_TYPE_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_CPU);
+    GGML_ASSERT(tensor->backend == GGML_BACKEND_TYPE_CPU);
     if (tensor->type != GGML_TYPE_F32 && tensor->type != GGML_TYPE_F16) {
         return;
     }
@@ -5428,7 +5481,7 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_compute_
     if (params->ith != 0) {
         return;
     }
-    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE || tensor->op == GGML_OP_TRANSPOSE) {
+    if (params->type == GGML_TASK_TYPE_INIT || params->type == GGML_TASK_TYPE_FINALIZE || tensor->op == GGML_OP_TRANSPOSE) {
         return;
     }
 
@@ -5465,10 +5518,10 @@ 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_CPU) {
+        if (src0->backend == GGML_BACKEND_TYPE_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_GPU) {
+        } else if (src0->backend == GGML_BACKEND_TYPE_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)) {
@@ -5508,10 +5561,10 @@ 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_CPU) {
+        if (src1->backend == GGML_BACKEND_TYPE_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_GPU) {
+        } else if (src1->backend == GGML_BACKEND_TYPE_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)) {
@@ -5670,7 +5723,7 @@ static void ggml_vk_check_results_1(ggml_backend_vk_context * ctx, ggml_compute_
     if (params->ith != 0) {
         return;
     }
-    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE || tensor->op == GGML_OP_TRANSPOSE) {
+    if (params->type == GGML_TASK_TYPE_INIT || params->type == GGML_TASK_TYPE_FINALIZE || tensor->op == GGML_OP_TRANSPOSE) {
         return;
     }
     if (!(vk_output_tensor > 0 && vk_output_tensor == check_counter) && check_counter <= vk_skip_checks) {
@@ -5682,7 +5735,7 @@ static void ggml_vk_check_results_1(ggml_backend_vk_context * ctx, ggml_compute_
 
     void * tensor_data = tensor->data;
 
-    if (tensor->backend == GGML_BACKEND_GPU) {
+    if (tensor->backend == GGML_BACKEND_TYPE_GPU) {
         size_t tensor_size = ggml_nbytes(tensor);
         tensor_data = malloc(tensor_size);
 
@@ -5815,7 +5868,7 @@ 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_GPU) {
+    if (tensor->backend == GGML_BACKEND_TYPE_GPU) {
         free(tensor_data);
     }
 }

ggml.h

@@ -315,13 +315,7 @@
 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);
@@ -354,6 +348,9 @@ extern "C" {
         GGML_TYPE_IQ2_XXS = 16,
         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,
@@ -367,9 +364,9 @@ extern "C" {
     };
 
     enum ggml_backend_type {
-        GGML_BACKEND_CPU = 0,
-        GGML_BACKEND_GPU = 10,
-        GGML_BACKEND_GPU_SPLIT = 20,
+        GGML_BACKEND_TYPE_CPU = 0,
+        GGML_BACKEND_TYPE_GPU = 10,
+        GGML_BACKEND_TYPE_GPU_SPLIT = 20,
     };
 
     // model file types
@@ -391,6 +388,9 @@ extern "C" {
         GGML_FTYPE_MOSTLY_IQ2_XXS = 15, // except 1d tensors
         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:
@@ -498,9 +498,9 @@ extern "C" {
     };
 
     enum ggml_object_type {
-        GGML_OBJECT_TENSOR,
-        GGML_OBJECT_GRAPH,
-        GGML_OBJECT_WORK_BUFFER
+        GGML_OBJECT_TYPE_TENSOR,
+        GGML_OBJECT_TYPE_GRAPH,
+        GGML_OBJECT_TYPE_WORK_BUFFER
     };
 
     enum ggml_log_level {
@@ -642,9 +642,9 @@ 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_INIT = 0,
-        GGML_TASK_COMPUTE,
-        GGML_TASK_FINALIZE,
+        GGML_TASK_TYPE_INIT = 0,
+        GGML_TASK_TYPE_COMPUTE,
+        GGML_TASK_TYPE_FINALIZE,
     };
 
     struct ggml_compute_params {
@@ -1383,13 +1383,17 @@ extern "C" {
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
 
-    // fused soft_max(a*scale + mask)
+    // fused soft_max(a*scale + mask + pos[i]*(ALiBi slope))
     // mask is optional
+    // pos is required when max_bias > 0.0f
+    // max_bias = 0.0f for no ALiBi
     GGML_API struct ggml_tensor * ggml_soft_max_ext(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
             struct ggml_tensor  * mask,
-            float                 scale);
+            struct ggml_tensor  * pos,
+            float                 scale,
+            float                 max_bias);
 
     GGML_API struct ggml_tensor * ggml_soft_max_back(
             struct ggml_context * ctx,
@@ -1491,12 +1495,13 @@ extern "C" {
 
     // alibi position embedding
     // in-place, returns view(a)
-    GGML_API struct ggml_tensor * ggml_alibi(
+    GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_alibi(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
             int                   n_past,
             int                   n_head,
-            float                 bias_max);
+            float                 bias_max),
+        "use ggml_soft_max_ext instead (will be removed in Mar 2024)");
 
     // clamp
     // in-place, returns view(a)
@@ -1644,8 +1649,8 @@ extern "C" {
 
     // sort rows
     enum ggml_sort_order {
-        GGML_SORT_ASC,
-        GGML_SORT_DESC,
+        GGML_SORT_ORDER_ASC,
+        GGML_SORT_ORDER_DESC,
     };
 
     GGML_API struct ggml_tensor * ggml_argsort(
@@ -1938,8 +1943,8 @@ extern "C" {
 
     // optimization methods
     enum ggml_opt_type {
-        GGML_OPT_ADAM,
-        GGML_OPT_LBFGS,
+        GGML_OPT_TYPE_ADAM,
+        GGML_OPT_TYPE_LBFGS,
     };
 
     // linesearch methods
@@ -1953,12 +1958,12 @@ extern "C" {
 
     // optimization return values
     enum ggml_opt_result {
-        GGML_OPT_OK = 0,
-        GGML_OPT_DID_NOT_CONVERGE,
-        GGML_OPT_NO_CONTEXT,
-        GGML_OPT_INVALID_WOLFE,
-        GGML_OPT_FAIL,
-        GGML_OPT_CANCEL,
+        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_LINESEARCH_FAIL = -128,
         GGML_LINESEARCH_MINIMUM_STEP,

gguf-py/gguf/constants.py

@@ -111,6 +111,7 @@ class MODEL_ARCH(IntEnum):
     ORION      = auto()
     INTERNLM2  = auto()
     MINICPM    = auto()
+    GEMMA      = auto()
 
 
 class MODEL_TENSOR(IntEnum):
@@ -167,6 +168,7 @@ 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] = {
@@ -511,6 +513,19 @@ 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

@@ -64,6 +64,15 @@ 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,
@@ -100,28 +109,32 @@ extern "C" {
         LLAMA_FTYPE_MOSTLY_Q2_K_S        = 21, // except 1d tensors
         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_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,
+        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,
     };
 
     enum llama_pooling_type {
-        LLAMA_POOLING_NONE = 0,
-        LLAMA_POOLING_MEAN = 1,
-        LLAMA_POOLING_CLS  = 2,
+        LLAMA_POOLING_TYPE_NONE = 0,
+        LLAMA_POOLING_TYPE_MEAN = 1,
+        LLAMA_POOLING_TYPE_CLS  = 2,
     };
 
     enum llama_split_mode {
-        LLAMA_SPLIT_NONE    = 0, // single GPU
-        LLAMA_SPLIT_LAYER   = 1, // split layers and KV across GPUs
-        LLAMA_SPLIT_ROW     = 2, // split rows across GPUs
+        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
     };
 
     typedef struct llama_token_data {
@@ -169,9 +182,9 @@ extern "C" {
     } llama_batch;
 
     enum llama_model_kv_override_type {
-        LLAMA_KV_OVERRIDE_INT,
-        LLAMA_KV_OVERRIDE_FLOAT,
-        LLAMA_KV_OVERRIDE_BOOL,
+        LLAMA_KV_OVERRIDE_TYPE_INT,
+        LLAMA_KV_OVERRIDE_TYPE_FLOAT,
+        LLAMA_KV_OVERRIDE_TYPE_BOOL,
     };
 
     struct llama_model_kv_override {
@@ -304,6 +317,12 @@ extern "C" {
         int32_t n_eval;
     };
 
+    // used in chat template
+    typedef struct llama_chat_message {
+        const char * role;
+        const char * content;
+    } llama_chat_message;
+
     // Helpers for getting default parameters
     LLAMA_API struct llama_model_params llama_model_default_params(void);
     LLAMA_API struct llama_context_params llama_context_default_params(void);
@@ -350,6 +369,7 @@ 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);
@@ -504,10 +524,12 @@ 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
+    // If the KV cache is RoPEd, the KV data is updated accordingly:
+    //   - lazily on next llama_decode()
+    //   - explicitly with llama_kv_cache_update()
     // p0 < 0 : [0,  p1]
     // p1 < 0 : [p0, inf)
-    LLAMA_API void llama_kv_cache_seq_shift(
+    LLAMA_API void llama_kv_cache_seq_add(
             struct llama_context * ctx,
                     llama_seq_id   seq_id,
                        llama_pos   p0,
@@ -515,7 +537,9 @@ 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
+    // If the KV cache is RoPEd, the KV data is updated accordingly:
+    //   - lazily on next llama_decode()
+    //   - explicitly with llama_kv_cache_update()
     // p0 < 0 : [0,  p1]
     // p1 < 0 : [p0, inf)
     LLAMA_API void llama_kv_cache_seq_div(
@@ -525,6 +549,20 @@ 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
     //
@@ -698,6 +736,25 @@ extern "C" {
                                   char * buf,
                                int32_t   length);
 
+    /// 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
+    /// @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
+    /// @param add_ass Whether to end the prompt with the token(s) that indicate the start of an assistant message.
+    /// @param buf A buffer to hold the output formatted prompt. The recommended alloc size is 2 * (total number of characters of all messages)
+    /// @param length The size of the allocated buffer
+    /// @return The total number of bytes of the formatted prompt. If is it larger than the size of buffer, you may need to re-alloc it and then re-apply the template.
+    LLAMA_API int32_t llama_chat_apply_template(
+              const struct llama_model * model,
+                            const char * tmpl,
+       const struct llama_chat_message * chat,
+                                size_t   n_msg,
+                                  bool   add_ass,
+                                  char * buf,
+                               int32_t   length);
+
     //
     // Grammar
     //

scripts/get-flags.mk

@@ -1,6 +1,6 @@
 ifeq '' '$(findstring clang,$(shell $(GF_CC) --version))'
 	GF_CC_IS_GCC = 1
-	GF_CC_VER := $(shell { $(GF_CC) -dumpfullversion 2>/dev/null || $(GF_CC) -dumpversion; } | awk -F. '{ printf("%02d%02d%02d", $$1, $$2, $$3) }')
+	GF_CC_VER := $(shell { $(GF_CC) -dumpfullversion 2>/dev/null; echo; $(GF_CC) -dumpversion; } | awk -F. '/./ { printf("%02d%02d%02d", $$1, $$2, $$3); exit }')
 else
 	GF_CC_IS_CLANG = 1
 	ifeq '' '$(findstring Apple,$(shell $(GF_CC) --version))'

scripts/get-wikitext-2.sh

@@ -1,6 +1,6 @@
 #!/bin/bash
 
-wget https://s3.amazonaws.com/research.metamind.io/wikitext/wikitext-2-raw-v1.zip
+wget https://huggingface.co/datasets/ggml-org/ci/resolve/main/wikitext-2-raw-v1.zip
 
 echo "Usage:"
 echo ""

scripts/sync-ggml.last

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

tests/CMakeLists.txt

@@ -28,6 +28,7 @@ endfunction()
 llama_build_and_test_executable(test-quantize-fns.cpp)
 llama_build_and_test_executable(test-quantize-perf.cpp)
 llama_build_and_test_executable(test-sampling.cpp)
+llama_build_and_test_executable(test-chat-template.cpp)
 
 llama_build_executable(test-tokenizer-0-llama.cpp)
 llama_test_executable (test-tokenizer-0-llama test-tokenizer-0-llama.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-llama.gguf)

tests/test-backend-ops.cpp

@@ -1085,24 +1085,32 @@ struct test_diag_mask_inf : public test_case {
 struct test_soft_max : public test_case {
     const ggml_type type;
     const std::array<int64_t, 4> ne;
-    const float scale;
     const bool mask;
+    const float scale;
+    const float max_bias;
 
     std::string vars() override {
-        return VARS_TO_STR4(type, ne, scale, mask);
+        return VARS_TO_STR5(type, ne, mask, scale, max_bias);
     }
 
     test_soft_max(ggml_type type = GGML_TYPE_F32,
             std::array<int64_t, 4> ne = {10, 10, 10, 10},
+            bool mask = false,
             float scale = 1.0f,
-            bool mask = false)
-        : type(type), ne(ne), scale(scale), mask(mask) {}
+            float max_bias = 0.0f)
+        : type(type), ne(ne), mask(mask), scale(scale), max_bias(max_bias) {}
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
         ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data());
-        ggml_tensor * b = nullptr;
-        if (mask) { b = ggml_new_tensor_2d(ctx, type, ne[0], ne[1]); }
-        ggml_tensor * out = ggml_soft_max_ext(ctx, a, b, scale);
+        ggml_tensor * mask = nullptr;
+        if (this->mask) {
+            mask = ggml_new_tensor_2d(ctx, type, ne[0], ne[1]);
+        }
+        ggml_tensor * pos = nullptr;
+        if (max_bias > 0.0f) {
+            pos = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, ne[0]);
+        }
+        ggml_tensor * out = ggml_soft_max_ext(ctx, a, mask, pos, scale, max_bias);
         return out;
     }
 };
@@ -1147,30 +1155,6 @@ struct test_rope : public test_case {
     }
 };
 
-// GGML_OP_ALIBI
-struct test_alibi : public test_case {
-    const ggml_type type;
-    const std::array<int64_t, 4> ne;
-    int n_past;
-    int n_head;
-    float bias_max;
-
-    std::string vars() override {
-        return VARS_TO_STR5(type, ne, n_past, n_head, bias_max);
-    }
-
-    test_alibi(ggml_type type = GGML_TYPE_F32,
-            std::array<int64_t, 4> ne = {10, 10, 10, 10},
-            int n_past = 512, int n_head = 10, float bias_max = 0.5f)
-        : type(type), ne(ne), n_past(n_past), n_head(n_head), bias_max(bias_max) {}
-
-    ggml_tensor * build_graph(ggml_context * ctx) override {
-        ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data());
-        ggml_tensor * out = ggml_alibi(ctx, a, n_past, n_head, bias_max);
-        return out;
-    }
-};
-
 // GGML_OP_POOL2D
 struct test_pool2d : public test_case {
     enum ggml_op_pool pool_type;
@@ -1280,7 +1264,7 @@ 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_ASC)
+            ggml_sort_order order = GGML_SORT_ORDER_ASC)
         : type(type), ne(ne), order(order) {}
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
@@ -1488,7 +1472,7 @@ struct test_moe : public test_case {
         ggml_tensor * cur = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_tokens);
 
         ggml_tensor * logits = ggml_mul_mat(ctx, ffn_gate_inp, cur);
-        ggml_tensor * probs = ggml_soft_max_ext(ctx, logits, nullptr, 1.0f/sqrtf(n_embd));
+        ggml_tensor * probs = ggml_soft_max_ext(ctx, logits, nullptr, nullptr, 1.0f/sqrtf(n_embd), 0.0f);
 
         // select experts
         ggml_tensor * selected_experts = ggml_top_k(ctx, probs, n_experts_per_tok);
@@ -1617,7 +1601,6 @@ public:
         ggml_cpy(ctx, v_cur_t, v_cache_view);
     }
 
-    // if max_alibi_bias > 0 then apply ALiBi
     struct ggml_tensor * llm_build_kqv(
             struct ggml_context * ctx,
              struct ggml_tensor * k_l,
@@ -1636,7 +1619,7 @@ public:
 
         struct ggml_tensor * kq = ggml_mul_mat(ctx, k, q);
 
-        kq = ggml_soft_max_ext(ctx, kq, kq_mask, kq_scale);
+        kq = ggml_soft_max_ext(ctx, kq, kq_mask, nullptr, kq_scale, 0.0f);
 
         // split cached v into n_head heads
         struct ggml_tensor * v =
@@ -1934,7 +1917,8 @@ static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op
         GGML_TYPE_Q4_K, GGML_TYPE_Q5_K,
         GGML_TYPE_Q6_K,
         GGML_TYPE_IQ2_XXS, GGML_TYPE_IQ2_XS,
-        GGML_TYPE_IQ3_XXS,
+        GGML_TYPE_IQ3_XXS, GGML_TYPE_IQ1_S,
+        GGML_TYPE_IQ4_NL, GGML_TYPE_IQ3_S,
     };
 
     // unary ops
@@ -2083,6 +2067,7 @@ static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op
     test_cases.emplace_back(new test_diag_mask_inf(GGML_TYPE_F32, {10, 10, 10,  1}, 5));
     test_cases.emplace_back(new test_diag_mask_inf(GGML_TYPE_F32, {10, 10, 10, 10}, 5));
 
+#if 0
     std::uniform_int_distribution<> dist_ne1(1, 50);
     int exponent = 1;
     while (exponent < (1 << 17)) {
@@ -2091,14 +2076,29 @@ static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op
         for (int n = 0; n < 10; ++n) {
             int64_t ne0 = dist_ne0(rng);
             int64_t ne1 = dist_ne1(rng);
-            test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {ne0, ne1, 1, 1}));
+            test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {ne0, ne1, 1, 1}, n/2 == 0, 0.1f, ne0 < 1000 ? 4.0f : 0.0f));
         }
 
         exponent <<= 1;
     }
+#endif
+    for (bool mask : {false, true}) {
+        for (float max_bias : {0.0f, 8.0f}) {
+            for (float scale : {1.0f, 0.1f}) {
+                for (int64_t ne0 : {16, 1024}) {
+                    for (int64_t ne1 : {16, 1024}) {
+                        test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {ne0, ne1, 1, 1}, mask, scale, max_bias));
+                        test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {ne0-1, ne1-1, 1, 1}, mask, scale, max_bias));
+                    }
+                }
+            }
+        }
+    }
 
-    test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {16, 2, 32, 1}, 0.1f));
-    test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {32, 2, 32, 1}, 0.1f, true));
+    test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {16, 2, 32, 1}, false, 0.1f, 0.0f));
+    test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {32, 2, 32, 1}, true,  0.1f, 0.0f));
+    test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {16, 2, 32, 1}, false, 0.1f, 8.0f));
+    test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {32, 2, 32, 1}, true,  0.1f, 8.0f));
 
     for (ggml_type type : {GGML_TYPE_F32, GGML_TYPE_F16}) {
         test_cases.emplace_back(new test_rope(type, {128,  32, 10, 1}, 128, 0, 512)); // llama 7B
@@ -2113,11 +2113,10 @@ static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op
         test_cases.emplace_back(new test_rope(type, { 80,  32, 10, 1},  32, 2, 512)); // neox (phi-2)
     }
 
-    test_cases.emplace_back(new test_alibi());
     test_cases.emplace_back(new test_concat(GGML_TYPE_F32));
     test_cases.emplace_back(new test_concat(GGML_TYPE_I32));
 
-    for (ggml_sort_order order : {GGML_SORT_ASC, GGML_SORT_DESC}) {
+    for (ggml_sort_order order : {GGML_SORT_ORDER_ASC, GGML_SORT_ORDER_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));
     }

tests/test-chat-template.cpp

@@ -0,0 +1,75 @@
+#include <iostream>
+#include <string>
+#include <vector>
+#include <sstream>
+
+#undef NDEBUG
+#include <cassert>
+
+#include "llama.h"
+
+int main(void) {
+    llama_chat_message conversation[] = {
+        {"system", "You are a helpful assistant"},
+        {"user", "Hello"},
+        {"assistant", "Hi there"},
+        {"user", "Who are you"},
+        {"assistant", "   I am an assistant   "},
+        {"user", "Another question"},
+    };
+    size_t message_count = 6;
+    std::vector<std::string> templates = {
+        // teknium/OpenHermes-2.5-Mistral-7B
+        "{% for message in messages %}{{'<|im_start|>' + message['role'] + '\\n' + message['content'] + '<|im_end|>' + '\\n'}}{% endfor %}{% if add_generation_prompt %}{{ '<|im_start|>assistant\\n' }}{% endif %}",
+        // mistralai/Mistral-7B-Instruct-v0.2
+        "{{ bos_token }}{% 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'] == 'user' %}{{ '[INST] ' + message['content'] + ' [/INST]' }}{% elif message['role'] == 'assistant' %}{{ message['content'] + eos_token}}{% else %}{{ raise_exception('Only user and assistant roles are supported!') }}{% endif %}{% endfor %}",
+        // TheBloke/FusionNet_34Bx2_MoE-AWQ
+        "{%- 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<char> formatted_chat(1024);
+    int32_t res;
+
+    // test invalid chat template
+    res = llama_chat_apply_template(nullptr, "INVALID TEMPLATE", conversation, message_count, true, formatted_chat.data(), formatted_chat.size());
+    assert(res < 0);
+
+    for (size_t i = 0; i < templates.size(); i++) {
+        std::string custom_template = templates[i];
+        std::string expected = expected_output[i];
+        formatted_chat.resize(1024);
+        res = llama_chat_apply_template(
+            nullptr,
+            custom_template.c_str(),
+            conversation,
+            message_count,
+            true,
+            formatted_chat.data(),
+            formatted_chat.size()
+        );
+        formatted_chat.resize(res);
+        std::string output(formatted_chat.data(), formatted_chat.size());
+        std::cout << output << "\n-------------------------\n";
+        assert(output == expected);
+    }
+    return 0;
+}

tests/test-grammar-parser.cpp

@@ -38,8 +38,8 @@ term  ::= [0-9]+)""";
         // pretty print error message before asserting
         if (expected_pair.first != key || expected_pair.second != value)
         {
-            fprintf(stderr, "expected_pair: %s, %d\n", expected_pair.first.c_str(), expected_pair.second);
-            fprintf(stderr, "actual_pair: %s, %d\n", key.c_str(), value);
+            fprintf(stderr, "expected_pair: %s, %u\n", expected_pair.first.c_str(), expected_pair.second);
+            fprintf(stderr, "actual_pair: %s, %u\n", key.c_str(), value);
             fprintf(stderr, "expected_pair != actual_pair\n");
         }
 
@@ -96,9 +96,9 @@ term  ::= [0-9]+)""";
             // pretty print error message before asserting
             if (expected_element.type != element.type || expected_element.value != element.value)
             {
-                fprintf(stderr, "index: %d\n", index);
-                fprintf(stderr, "expected_element: %d, %d\n", expected_element.type, expected_element.value);
-                fprintf(stderr, "actual_element: %d, %d\n", element.type, element.value);
+                fprintf(stderr, "index: %u\n", index);
+                fprintf(stderr, "expected_element: %d, %u\n", expected_element.type, expected_element.value);
+                fprintf(stderr, "actual_element: %d, %u\n", element.type, element.value);
                 fprintf(stderr, "expected_element != actual_element\n");
             }
 
@@ -144,8 +144,8 @@ term  ::= [0-9]+)""";
         // pretty print error message before asserting
         if (expected_pair.first != key || expected_pair.second != value)
         {
-            fprintf(stderr, "expected_pair: %s, %d\n", expected_pair.first.c_str(), expected_pair.second);
-            fprintf(stderr, "actual_pair: %s, %d\n", key.c_str(), value);
+            fprintf(stderr, "expected_pair: %s, %u\n", expected_pair.first.c_str(), expected_pair.second);
+            fprintf(stderr, "actual_pair: %s, %u\n", key.c_str(), value);
             fprintf(stderr, "expected_pair != actual_pair\n");
         }
 
@@ -235,9 +235,9 @@ term  ::= [0-9]+)""";
             // pretty print error message before asserting
             if (expected_element.type != element.type || expected_element.value != element.value)
             {
-                fprintf(stderr, "index: %d\n", index);
-                fprintf(stderr, "expected_element: %d, %d\n", expected_element.type, expected_element.value);
-                fprintf(stderr, "actual_element: %d, %d\n", element.type, element.value);
+                fprintf(stderr, "index: %u\n", index);
+                fprintf(stderr, "expected_element: %d, %u\n", expected_element.type, expected_element.value);
+                fprintf(stderr, "actual_element: %d, %u\n", element.type, element.value);
                 fprintf(stderr, "expected_element != actual_element\n");
             }
 

tests/test-llama-grammar.cpp

@@ -180,8 +180,8 @@ int main()
             if (expected_element.type != element->type || expected_element.value != element->value)
             {
                 fprintf(stderr, "index: %d\n", index);
-                fprintf(stderr, "expected_element: %d, %d\n", expected_element.type, expected_element.value);
-                fprintf(stderr, "actual_element: %d, %d\n", element->type, element->value);
+                fprintf(stderr, "expected_element: %d, %u\n", expected_element.type, expected_element.value);
+                fprintf(stderr, "actual_element: %d, %u\n", element->type, element->value);
                 fprintf(stderr, "expected_element != actual_element\n");
             }
 

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_ADAM);
+    struct ggml_opt_params opt_params = ggml_opt_default_params(GGML_OPT_TYPE_ADAM);
 
     ggml_opt(ctx, opt_params, e);
 

tests/test-quantize-fns.cpp

@@ -151,6 +151,7 @@ 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;
@@ -167,7 +168,8 @@ 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 ? MAX_DOT_PRODUCT_ERROR_LOWBIT : MAX_DOT_PRODUCT_ERROR;
+                                            type == GGML_TYPE_IQ3_XXS || type == GGML_TYPE_IQ3_S ? MAX_DOT_PRODUCT_ERROR_LOWBIT
+                                          : MAX_DOT_PRODUCT_ERROR;
             failed = !(vec_dot_error < max_allowed_error);
             num_failed += failed;
             if (failed || verbose) {

unicode.h

@@ -404,7 +404,8 @@ static std::unordered_map<uint32_t, int> codepoint_type_map() {
 
 static int codepoint_type(uint32_t cp) {
     static std::unordered_map<uint32_t, int> codepoint_types = codepoint_type_map();
-    return codepoint_types.find(cp) == codepoint_types.end() ? CODEPOINT_TYPE_UNIDENTIFIED : codepoint_types.at(cp);
+    const auto it = codepoint_types.find(cp);
+    return it == codepoint_types.end() ? CODEPOINT_TYPE_UNIDENTIFIED : it->second;
 }
 
 static int codepoint_type(const std::string & utf8) {