py : add Gemma conversion from HF models (#5647)

* py : add gemma conversion from HF models

* Update convert-hf-to-gguf.py

Co-authored-by: Aarni Koskela <akx@iki.fi>

* Update convert-hf-to-gguf.py

Co-authored-by: Aarni Koskela <akx@iki.fi>

* Update convert-hf-to-gguf.py

Co-authored-by: Jared Van Bortel <jared@nomic.ai>

---------

Co-authored-by: Aarni Koskela <akx@iki.fi>
Co-authored-by: Jared Van Bortel <jared@nomic.ai>

.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/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
@@ -100,7 +102,7 @@ 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 +246,8 @@ jobs:
 
       - name: Build
         id: make_build
+        env:
+            LLAMA_FATAL_WARNINGS: 1
         run: |
           LLAMA_NO_METAL=1 make -j $(sysctl -n hw.logicalcpu)
 
@@ -277,7 +281,7 @@ 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

.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

.gitignore

@@ -23,11 +23,13 @@
 .clang-tidy
 .vs/
 .vscode/
+.idea/
 
 lcov-report/
 gcovr-report/
 
 build*
+cmake-build-*
 out/
 tmp/
 

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)
@@ -107,22 +110,20 @@ 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)
 option(LLAMA_SYCL                            "llama: use SYCL"                                  OFF)
 option(LLAMA_SYCL_F16                        "llama: use 16 bit floats for sycl calculations"   OFF)
+option(LLAMA_CPU_HBM                         "llama: use memkind for CPU HBM"                   OFF)
 
 option(LLAMA_BUILD_TESTS                     "llama: build tests"    ${LLAMA_STANDALONE})
 option(LLAMA_BUILD_EXAMPLES                  "llama: build examples" ${LLAMA_STANDALONE})
 option(LLAMA_BUILD_SERVER                    "llama: build server example"                      ON)
 
-
 # add perf arguments
 option(LLAMA_PERF                            "llama: enable perf"                               OFF)
-if (LLAMA_PERF)
-    add_definitions(-DGGML_PERF)
-endif()
 
 # Required for relocatable CMake package
 include(${CMAKE_CURRENT_SOURCE_DIR}/scripts/build-info.cmake)
@@ -130,6 +131,7 @@ include(${CMAKE_CURRENT_SOURCE_DIR}/scripts/build-info.cmake)
 #
 # Compile flags
 #
+
 if (LLAMA_SYCL)
     set(CMAKE_CXX_STANDARD 17)
 else()
@@ -140,6 +142,7 @@ set(CMAKE_CXX_STANDARD_REQUIRED true)
 set(CMAKE_C_STANDARD 11)
 set(CMAKE_C_STANDARD_REQUIRED true)
 set(THREADS_PREFER_PTHREAD_FLAG ON)
+
 find_package(Threads REQUIRED)
 include(CheckCXXCompilerFlag)
 
@@ -151,17 +154,17 @@ endif()
 if (NOT MSVC)
     if (LLAMA_SANITIZE_THREAD)
         add_compile_options(-fsanitize=thread)
-        link_libraries(-fsanitize=thread)
+        link_libraries     (-fsanitize=thread)
     endif()
 
     if (LLAMA_SANITIZE_ADDRESS)
         add_compile_options(-fsanitize=address -fno-omit-frame-pointer)
-        link_libraries(-fsanitize=address)
+        link_libraries     (-fsanitize=address)
     endif()
 
     if (LLAMA_SANITIZE_UNDEFINED)
         add_compile_options(-fsanitize=undefined)
-        link_libraries(-fsanitize=undefined)
+        link_libraries     (-fsanitize=undefined)
     endif()
 endif()
 
@@ -199,6 +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
@@ -298,14 +324,17 @@ if (LLAMA_BLAS)
         endif()
 
         message(STATUS "BLAS found, Includes: ${BLAS_INCLUDE_DIRS}")
+
         add_compile_options(${BLAS_LINKER_FLAGS})
+
         add_compile_definitions(GGML_USE_OPENBLAS)
+
         if (${BLAS_INCLUDE_DIRS} MATCHES "mkl" AND (${LLAMA_BLAS_VENDOR} MATCHES "Generic" OR ${LLAMA_BLAS_VENDOR} MATCHES "Intel"))
             add_compile_definitions(GGML_BLAS_USE_MKL)
         endif()
-        set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} ${BLAS_LIBRARIES})
-        set(LLAMA_EXTRA_INCLUDES ${LLAMA_EXTRA_INCLUDES} ${BLAS_INCLUDE_DIRS})
 
+        set(LLAMA_EXTRA_LIBS     ${LLAMA_EXTRA_LIBS}     ${BLAS_LIBRARIES})
+        set(LLAMA_EXTRA_INCLUDES ${LLAMA_EXTRA_INCLUDES} ${BLAS_INCLUDE_DIRS})
     else()
         message(WARNING "BLAS not found, please refer to "
         "https://cmake.org/cmake/help/latest/module/FindBLAS.html#blas-lapack-vendors"
@@ -330,9 +359,6 @@ if (LLAMA_CUBLAS)
         set(GGML_SOURCES_CUDA ggml-cuda.cu)
 
         add_compile_definitions(GGML_USE_CUBLAS)
-#        if (LLAMA_CUDA_CUBLAS)
-#            add_compile_definitions(GGML_CUDA_CUBLAS)
-#        endif()
         if (LLAMA_CUDA_FORCE_DMMV)
             add_compile_definitions(GGML_CUDA_FORCE_DMMV)
         endif()
@@ -387,15 +413,20 @@ if (LLAMA_MPI)
     find_package(MPI)
     if (MPI_C_FOUND)
         message(STATUS "MPI found")
+
         set(GGML_HEADERS_MPI ggml-mpi.h)
-        set(GGML_SOURCES_MPI ggml-mpi.c ggml-mpi.h)
+        set(GGML_SOURCES_MPI ggml-mpi.c)
+
         add_compile_definitions(GGML_USE_MPI)
         add_compile_definitions(${MPI_C_COMPILE_DEFINITIONS})
+
         if (NOT MSVC)
             add_compile_options(-Wno-cast-qual)
         endif()
+
         set(LLAMA_EXTRA_LIBS     ${LLAMA_EXTRA_LIBS}     ${MPI_C_LIBRARIES})
         set(LLAMA_EXTRA_INCLUDES ${LLAMA_EXTRA_INCLUDES} ${MPI_C_INCLUDE_DIRS})
+
         # Even if you're only using the C header, C++ programs may bring in MPI
         # C++ functions, so more linkage is needed
         if (MPI_CXX_FOUND)
@@ -427,31 +458,28 @@ if (LLAMA_VULKAN)
     if (Vulkan_FOUND)
         message(STATUS "Vulkan found")
 
-        add_library(ggml-vulkan OBJECT ggml-vulkan.cpp ggml-vulkan.h)
-        if (BUILD_SHARED_LIBS)
-            set_target_properties(ggml-vulkan PROPERTIES POSITION_INDEPENDENT_CODE ON)
-        endif()
-        target_link_libraries(ggml-vulkan PRIVATE Vulkan::Vulkan)
+        set(GGML_HEADERS_VULKAN ggml-vulkan.h)
+        set(GGML_SOURCES_VULKAN ggml-vulkan.cpp)
 
         add_compile_definitions(GGML_USE_VULKAN)
 
         if (LLAMA_VULKAN_CHECK_RESULTS)
-            target_compile_definitions(ggml-vulkan PRIVATE GGML_VULKAN_CHECK_RESULTS)
+            add_compile_definitions(GGML_VULKAN_CHECK_RESULTS)
         endif()
 
         if (LLAMA_VULKAN_DEBUG)
-            target_compile_definitions(ggml-vulkan PRIVATE GGML_VULKAN_DEBUG)
+            add_compile_definitions(GGML_VULKAN_DEBUG)
         endif()
 
         if (LLAMA_VULKAN_VALIDATE)
-            target_compile_definitions(ggml-vulkan PRIVATE GGML_VULKAN_VALIDATE)
+            add_compile_definitions(GGML_VULKAN_VALIDATE)
         endif()
 
         if (LLAMA_VULKAN_RUN_TESTS)
-            target_compile_definitions(ggml-vulkan PRIVATE GGML_VULKAN_RUN_TESTS)
+            add_compile_definitions(GGML_VULKAN_RUN_TESTS)
         endif()
 
-        set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} ggml-vulkan)
+        set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} Vulkan::Vulkan)
     else()
         message(WARNING "Vulkan not found")
     endif()
@@ -463,43 +491,45 @@ if (LLAMA_HIPBLAS)
     if (NOT ${CMAKE_C_COMPILER_ID} MATCHES "Clang")
         message(WARNING "Only LLVM is supported for HIP, hint: CC=/opt/rocm/llvm/bin/clang")
     endif()
+
     if (NOT ${CMAKE_CXX_COMPILER_ID} MATCHES "Clang")
         message(WARNING "Only LLVM is supported for HIP, hint: CXX=/opt/rocm/llvm/bin/clang++")
     endif()
 
-    find_package(hip)
-    find_package(hipblas)
-    find_package(rocblas)
+    find_package(hip     REQUIRED)
+    find_package(hipblas REQUIRED)
+    find_package(rocblas REQUIRED)
 
-    if (${hipblas_FOUND} AND ${hip_FOUND})
-        message(STATUS "HIP and hipBLAS found")
-        add_compile_definitions(GGML_USE_HIPBLAS GGML_USE_CUBLAS)
-        if (LLAMA_HIP_UMA)
-            add_compile_definitions(GGML_HIP_UMA)
-        endif()
-        add_library(ggml-rocm OBJECT ggml-cuda.cu ggml-cuda.h)
-        if (BUILD_SHARED_LIBS)
-            set_target_properties(ggml-rocm PROPERTIES POSITION_INDEPENDENT_CODE ON)
-        endif()
-        if (LLAMA_CUDA_FORCE_DMMV)
-            target_compile_definitions(ggml-rocm PRIVATE GGML_CUDA_FORCE_DMMV)
-        endif()
-        if (LLAMA_CUDA_FORCE_MMQ)
-            target_compile_definitions(ggml-rocm PRIVATE GGML_CUDA_FORCE_MMQ)
-        endif()
-        target_compile_definitions(ggml-rocm PRIVATE GGML_CUDA_DMMV_X=${LLAMA_CUDA_DMMV_X})
-        target_compile_definitions(ggml-rocm PRIVATE GGML_CUDA_MMV_Y=${LLAMA_CUDA_MMV_Y})
-        target_compile_definitions(ggml-rocm PRIVATE K_QUANTS_PER_ITERATION=${LLAMA_CUDA_KQUANTS_ITER})
-        set_source_files_properties(ggml-cuda.cu PROPERTIES LANGUAGE CXX)
-        target_link_libraries(ggml-rocm PRIVATE hip::device PUBLIC hip::host roc::rocblas roc::hipblas)
+    message(STATUS "HIP and hipBLAS found")
 
-        if (LLAMA_STATIC)
-            message(FATAL_ERROR "Static linking not supported for HIP/ROCm")
-        endif()
-        set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} ggml-rocm)
-    else()
-        message(WARNING "hipBLAS or HIP not found. Try setting CMAKE_PREFIX_PATH=/opt/rocm")
+    set(GGML_HEADERS_ROCM ggml-cuda.h)
+    set(GGML_SOURCES_ROCM ggml-cuda.cu)
+
+    add_compile_definitions(GGML_USE_HIPBLAS GGML_USE_CUBLAS)
+
+    if (LLAMA_HIP_UMA)
+        add_compile_definitions(GGML_HIP_UMA)
     endif()
+
+    if (LLAMA_CUDA_FORCE_DMMV)
+        add_compile_definitions(GGML_CUDA_FORCE_DMMV)
+    endif()
+
+    if (LLAMA_CUDA_FORCE_MMQ)
+        add_compile_definitions(GGML_CUDA_FORCE_MMQ)
+    endif()
+
+    add_compile_definitions(GGML_CUDA_DMMV_X=${LLAMA_CUDA_DMMV_X})
+    add_compile_definitions(GGML_CUDA_MMV_Y=${LLAMA_CUDA_MMV_Y})
+    add_compile_definitions(K_QUANTS_PER_ITERATION=${LLAMA_CUDA_KQUANTS_ITER})
+
+    set_source_files_properties(ggml-cuda.cu PROPERTIES LANGUAGE CXX)
+
+    if (LLAMA_STATIC)
+        message(FATAL_ERROR "Static linking not supported for HIP/ROCm")
+    endif()
+
+    set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} hip::device PUBLIC hip::host roc::rocblas roc::hipblas)
 endif()
 
 if (LLAMA_SYCL)
@@ -509,10 +539,14 @@ if (LLAMA_SYCL)
     #todo: AOT
 
     find_package(IntelSYCL REQUIRED)
+
+    message(STATUS "SYCL found")
+
+    add_compile_definitions(GGML_USE_SYCL)
+
     if (LLAMA_SYCL_F16)
         add_compile_definitions(GGML_SYCL_F16)
     endif()
-    add_compile_definitions(GGML_USE_SYCL)
 
     add_compile_options(-I./) #include DPCT
     add_compile_options(-I/${SYCL_INCLUDE_DIR})
@@ -521,7 +555,7 @@ if (LLAMA_SYCL)
     set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -O3")
     set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fsycl -L${MKLROOT}/lib")
 
-    set(GGML_HEADERS_SYCL ggml.h ggml-sycl.h)
+    set(GGML_HEADERS_SYCL ggml-sycl.h)
     set(GGML_SOURCES_SYCL ggml-sycl.cpp)
 
     if (WIN32)
@@ -540,61 +574,61 @@ if (LLAMA_KOMPUTE)
     endif()
 
     function(compile_shader)
-      set(options)
-      set(oneValueArgs)
-      set(multiValueArgs SOURCES)
-      cmake_parse_arguments(compile_shader "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
-      foreach(source ${compile_shader_SOURCES})
-        get_filename_component(filename ${source} NAME)
-        set(spv_file ${filename}.spv)
-        add_custom_command(
-            OUTPUT ${spv_file}
-            DEPENDS ${CMAKE_CURRENT_SOURCE_DIR}/${source}
-              ${CMAKE_CURRENT_SOURCE_DIR}/kompute-shaders/common.comp
-              ${CMAKE_CURRENT_SOURCE_DIR}/kompute-shaders/op_getrows.comp
-              ${CMAKE_CURRENT_SOURCE_DIR}/kompute-shaders/op_mul_mv_q_n_pre.comp
-              ${CMAKE_CURRENT_SOURCE_DIR}/kompute-shaders/op_mul_mv_q_n.comp
-              COMMAND ${glslc_executable} --target-env=vulkan1.2 -o ${spv_file} ${CMAKE_CURRENT_SOURCE_DIR}/${source}
-            COMMENT "Compiling ${source} to ${spv_file}"
-        )
+        set(options)
+        set(oneValueArgs)
+        set(multiValueArgs SOURCES)
+        cmake_parse_arguments(compile_shader "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
+        foreach(source ${compile_shader_SOURCES})
+            get_filename_component(filename ${source} NAME)
+            set(spv_file ${filename}.spv)
+            add_custom_command(
+                OUTPUT ${spv_file}
+                DEPENDS ${CMAKE_CURRENT_SOURCE_DIR}/${source}
+                ${CMAKE_CURRENT_SOURCE_DIR}/kompute-shaders/common.comp
+                ${CMAKE_CURRENT_SOURCE_DIR}/kompute-shaders/op_getrows.comp
+                ${CMAKE_CURRENT_SOURCE_DIR}/kompute-shaders/op_mul_mv_q_n_pre.comp
+                ${CMAKE_CURRENT_SOURCE_DIR}/kompute-shaders/op_mul_mv_q_n.comp
+                COMMAND ${glslc_executable} --target-env=vulkan1.2 -o ${spv_file} ${CMAKE_CURRENT_SOURCE_DIR}/${source}
+                COMMENT "Compiling ${source} to ${spv_file}"
+                )
 
-        get_filename_component(RAW_FILE_NAME ${spv_file} NAME)
-        set(FILE_NAME "shader${RAW_FILE_NAME}")
-        string(REPLACE ".comp.spv" ".h" HEADER_FILE ${FILE_NAME})
-        string(TOUPPER ${HEADER_FILE} HEADER_FILE_DEFINE)
-        string(REPLACE "." "_" HEADER_FILE_DEFINE "${HEADER_FILE_DEFINE}")
-        set(OUTPUT_HEADER_FILE "${HEADER_FILE}")
-        message(STATUS "${HEADER_FILE} generating ${HEADER_FILE_DEFINE}")
-        if(CMAKE_GENERATOR MATCHES "Visual Studio")
-            add_custom_command(
-              OUTPUT ${OUTPUT_HEADER_FILE}
-              COMMAND ${CMAKE_COMMAND} -E echo "/*THIS FILE HAS BEEN AUTOMATICALLY GENERATED - DO NOT EDIT*/" > ${OUTPUT_HEADER_FILE}
-              COMMAND ${CMAKE_COMMAND} -E echo \"\#ifndef ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
-              COMMAND ${CMAKE_COMMAND} -E echo \"\#define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
-              COMMAND ${CMAKE_COMMAND} -E echo "namespace kp {" >> ${OUTPUT_HEADER_FILE}
-              COMMAND ${CMAKE_COMMAND} -E echo "namespace shader_data {" >> ${OUTPUT_HEADER_FILE}
-              COMMAND ${CMAKE_BINARY_DIR}/bin/$<CONFIG>/xxd -i ${RAW_FILE_NAME} >> ${OUTPUT_HEADER_FILE}
-              COMMAND ${CMAKE_COMMAND} -E echo "}}" >> ${OUTPUT_HEADER_FILE}
-              COMMAND ${CMAKE_COMMAND} -E echo \"\#endif // define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
-              DEPENDS ${spv_file} xxd
-              COMMENT "Converting to hpp: ${FILE_NAME} ${CMAKE_BINARY_DIR}/bin/$<CONFIG>/xxd"
-            )
-        else()
-            add_custom_command(
-              OUTPUT ${OUTPUT_HEADER_FILE}
-              COMMAND ${CMAKE_COMMAND} -E echo "/*THIS FILE HAS BEEN AUTOMATICALLY GENERATED - DO NOT EDIT*/" > ${OUTPUT_HEADER_FILE}
-              COMMAND ${CMAKE_COMMAND} -E echo \"\#ifndef ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
-              COMMAND ${CMAKE_COMMAND} -E echo \"\#define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
-              COMMAND ${CMAKE_COMMAND} -E echo "namespace kp {" >> ${OUTPUT_HEADER_FILE}
-              COMMAND ${CMAKE_COMMAND} -E echo "namespace shader_data {" >> ${OUTPUT_HEADER_FILE}
-              COMMAND ${CMAKE_BINARY_DIR}/bin/xxd -i ${RAW_FILE_NAME} >> ${OUTPUT_HEADER_FILE}
-              COMMAND ${CMAKE_COMMAND} -E echo "}}" >> ${OUTPUT_HEADER_FILE}
-              COMMAND ${CMAKE_COMMAND} -E echo \"\#endif // define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
-              DEPENDS ${spv_file} xxd
-              COMMENT "Converting to hpp: ${FILE_NAME} ${CMAKE_BINARY_DIR}/bin/xxd"
-            )
-        endif()
-      endforeach()
+            get_filename_component(RAW_FILE_NAME ${spv_file} NAME)
+            set(FILE_NAME "shader${RAW_FILE_NAME}")
+            string(REPLACE ".comp.spv" ".h" HEADER_FILE ${FILE_NAME})
+            string(TOUPPER ${HEADER_FILE} HEADER_FILE_DEFINE)
+            string(REPLACE "." "_" HEADER_FILE_DEFINE "${HEADER_FILE_DEFINE}")
+            set(OUTPUT_HEADER_FILE "${HEADER_FILE}")
+            message(STATUS "${HEADER_FILE} generating ${HEADER_FILE_DEFINE}")
+            if(CMAKE_GENERATOR MATCHES "Visual Studio")
+                add_custom_command(
+                    OUTPUT ${OUTPUT_HEADER_FILE}
+                    COMMAND ${CMAKE_COMMAND} -E echo "/*THIS FILE HAS BEEN AUTOMATICALLY GENERATED - DO NOT EDIT*/" > ${OUTPUT_HEADER_FILE}
+                    COMMAND ${CMAKE_COMMAND} -E echo \"\#ifndef ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
+                    COMMAND ${CMAKE_COMMAND} -E echo \"\#define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
+                    COMMAND ${CMAKE_COMMAND} -E echo "namespace kp {" >> ${OUTPUT_HEADER_FILE}
+                    COMMAND ${CMAKE_COMMAND} -E echo "namespace shader_data {" >> ${OUTPUT_HEADER_FILE}
+                    COMMAND ${CMAKE_BINARY_DIR}/bin/$<CONFIG>/xxd -i ${RAW_FILE_NAME} >> ${OUTPUT_HEADER_FILE}
+                    COMMAND ${CMAKE_COMMAND} -E echo "}}" >> ${OUTPUT_HEADER_FILE}
+                    COMMAND ${CMAKE_COMMAND} -E echo \"\#endif // define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
+                    DEPENDS ${spv_file} xxd
+                    COMMENT "Converting to hpp: ${FILE_NAME} ${CMAKE_BINARY_DIR}/bin/$<CONFIG>/xxd"
+                    )
+            else()
+                add_custom_command(
+                    OUTPUT ${OUTPUT_HEADER_FILE}
+                    COMMAND ${CMAKE_COMMAND} -E echo "/*THIS FILE HAS BEEN AUTOMATICALLY GENERATED - DO NOT EDIT*/" > ${OUTPUT_HEADER_FILE}
+                    COMMAND ${CMAKE_COMMAND} -E echo \"\#ifndef ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
+                    COMMAND ${CMAKE_COMMAND} -E echo \"\#define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
+                    COMMAND ${CMAKE_COMMAND} -E echo "namespace kp {" >> ${OUTPUT_HEADER_FILE}
+                    COMMAND ${CMAKE_COMMAND} -E echo "namespace shader_data {" >> ${OUTPUT_HEADER_FILE}
+                    COMMAND ${CMAKE_BINARY_DIR}/bin/xxd -i ${RAW_FILE_NAME} >> ${OUTPUT_HEADER_FILE}
+                    COMMAND ${CMAKE_COMMAND} -E echo "}}" >> ${OUTPUT_HEADER_FILE}
+                    COMMAND ${CMAKE_COMMAND} -E echo \"\#endif // define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
+                    DEPENDS ${spv_file} xxd
+                    COMMENT "Converting to hpp: ${FILE_NAME} ${CMAKE_BINARY_DIR}/bin/xxd"
+                    )
+            endif()
+        endforeach()
     endfunction()
 
     if (EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/kompute/CMakeLists.txt")
@@ -604,66 +638,66 @@ if (LLAMA_KOMPUTE)
 
         # Compile our shaders
         compile_shader(SOURCES
-          kompute-shaders/op_scale.comp
-          kompute-shaders/op_scale_8.comp
-          kompute-shaders/op_add.comp
-          kompute-shaders/op_addrow.comp
-          kompute-shaders/op_mul.comp
-          kompute-shaders/op_silu.comp
-          kompute-shaders/op_relu.comp
-          kompute-shaders/op_gelu.comp
-          kompute-shaders/op_softmax.comp
-          kompute-shaders/op_norm.comp
-          kompute-shaders/op_rmsnorm.comp
-          kompute-shaders/op_diagmask.comp
-          kompute-shaders/op_mul_mat_mat_f32.comp
-          kompute-shaders/op_mul_mat_f16.comp
-          kompute-shaders/op_mul_mat_q8_0.comp
-          kompute-shaders/op_mul_mat_q4_0.comp
-          kompute-shaders/op_mul_mat_q4_1.comp
-          kompute-shaders/op_mul_mat_q6_k.comp
-          kompute-shaders/op_getrows_f16.comp
-          kompute-shaders/op_getrows_q4_0.comp
-          kompute-shaders/op_getrows_q4_1.comp
-          kompute-shaders/op_getrows_q6_k.comp
-          kompute-shaders/op_rope_f16.comp
-          kompute-shaders/op_rope_f32.comp
-          kompute-shaders/op_cpy_f16_f16.comp
-          kompute-shaders/op_cpy_f16_f32.comp
-          kompute-shaders/op_cpy_f32_f16.comp
-          kompute-shaders/op_cpy_f32_f32.comp
+            kompute-shaders/op_scale.comp
+            kompute-shaders/op_scale_8.comp
+            kompute-shaders/op_add.comp
+            kompute-shaders/op_addrow.comp
+            kompute-shaders/op_mul.comp
+            kompute-shaders/op_silu.comp
+            kompute-shaders/op_relu.comp
+            kompute-shaders/op_gelu.comp
+            kompute-shaders/op_softmax.comp
+            kompute-shaders/op_norm.comp
+            kompute-shaders/op_rmsnorm.comp
+            kompute-shaders/op_diagmask.comp
+            kompute-shaders/op_mul_mat_mat_f32.comp
+            kompute-shaders/op_mul_mat_f16.comp
+            kompute-shaders/op_mul_mat_q8_0.comp
+            kompute-shaders/op_mul_mat_q4_0.comp
+            kompute-shaders/op_mul_mat_q4_1.comp
+            kompute-shaders/op_mul_mat_q6_k.comp
+            kompute-shaders/op_getrows_f16.comp
+            kompute-shaders/op_getrows_q4_0.comp
+            kompute-shaders/op_getrows_q4_1.comp
+            kompute-shaders/op_getrows_q6_k.comp
+            kompute-shaders/op_rope_f16.comp
+            kompute-shaders/op_rope_f32.comp
+            kompute-shaders/op_cpy_f16_f16.comp
+            kompute-shaders/op_cpy_f16_f32.comp
+            kompute-shaders/op_cpy_f32_f16.comp
+            kompute-shaders/op_cpy_f32_f32.comp
         )
 
         # Create a custom target for our generated shaders
         add_custom_target(generated_shaders DEPENDS
-          shaderop_scale.h
-          shaderop_scale_8.h
-          shaderop_add.h
-          shaderop_addrow.h
-          shaderop_mul.h
-          shaderop_silu.h
-          shaderop_relu.h
-          shaderop_gelu.h
-          shaderop_softmax.h
-          shaderop_norm.h
-          shaderop_rmsnorm.h
-          shaderop_diagmask.h
-          shaderop_mul_mat_mat_f32.h
-          shaderop_mul_mat_f16.h
-          shaderop_mul_mat_q8_0.h
-          shaderop_mul_mat_q4_0.h
-          shaderop_mul_mat_q4_1.h
-          shaderop_mul_mat_q6_k.h
-          shaderop_getrows_f16.h
-          shaderop_getrows_q4_0.h
-          shaderop_getrows_q4_1.h
-          shaderop_getrows_q6_k.h
-          shaderop_rope_f16.h
-          shaderop_rope_f32.h
-          shaderop_cpy_f16_f16.h
-          shaderop_cpy_f16_f32.h
-          shaderop_cpy_f32_f16.h
-          shaderop_cpy_f32_f32.h
+            shaderop_scale.h
+            shaderop_scale_8.h
+            shaderop_add.h
+            shaderop_addrow.h
+            shaderop_mul.h
+            shaderop_silu.h
+            shaderop_relu.h
+            shaderop_gelu.h
+            shaderop_softmax.h
+            shaderop_norm.h
+            shaderop_rmsnorm.h
+            shaderop_diagmask.h
+            shaderop_mul_mat_mat_f32.h
+            shaderop_mul_mat_f16.h
+            shaderop_mul_mat_q8_0.h
+            shaderop_mul_mat_q4_0.h
+            shaderop_mul_mat_q4_1.h
+            shaderop_mul_mat_q6_k.h
+            shaderop_getrows_f16.h
+            shaderop_getrows_q4_0.h
+            shaderop_getrows_q4_1.h
+            shaderop_getrows_q6_k.h
+            shaderop_rope_f16.h
+            shaderop_rope_f32.h
+            shaderop_cpy_f16_f16.h
+            shaderop_cpy_f16_f32.h
+            shaderop_cpy_f32_f16.h
+            shaderop_cpy_f32_f32.h
         )
 
         # Create a custom command that depends on the generated_shaders
@@ -676,8 +710,10 @@ if (LLAMA_KOMPUTE)
 
         # Add the stamp to the main sources to ensure dependency tracking
         set(GGML_SOURCES_KOMPUTE ggml-kompute.cpp ${CMAKE_CURRENT_BINARY_DIR}/ggml-kompute.stamp)
-        set(GGML_HEADERS_KOMPUTE ggml-kompute.h ${CMAKE_CURRENT_BINARY_DIR}/ggml-kompute.stamp)
+        set(GGML_HEADERS_KOMPUTE ggml-kompute.h   ${CMAKE_CURRENT_BINARY_DIR}/ggml-kompute.stamp)
+
         add_compile_definitions(GGML_USE_KOMPUTE)
+
         set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} kompute)
         set(LLAMA_EXTRA_INCLUDES ${LLAMA_EXTRA_INCLUDES} ${CMAKE_BINARY_DIR})
     else()
@@ -685,6 +721,18 @@ if (LLAMA_KOMPUTE)
     endif()
 endif()
 
+if (LLAMA_CPU_HBM)
+    find_library(memkind memkind REQUIRED)
+
+    add_compile_definitions(GGML_USE_CPU_HBM)
+
+    target_link_libraries(ggml PUBLIC memkind)
+endif()
+
+if (LLAMA_PERF)
+    add_compile_definitions(GGML_PERF)
+endif()
+
 function(get_flags CCID CCVER)
     set(C_FLAGS "")
     set(CXX_FLAGS "")
@@ -709,28 +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})
 
@@ -746,9 +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)
@@ -782,7 +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()
 
@@ -821,6 +876,7 @@ execute_process(
     ERROR_VARIABLE output
     OUTPUT_QUIET
 )
+
 if (output MATCHES "dyld-1015\.7")
     add_compile_definitions(HAVE_BUGGY_APPLE_LINKER)
 endif()
@@ -830,10 +886,10 @@ endif()
 #       feel free to update the Makefile for your architecture and send a pull request or issue
 message(STATUS "CMAKE_SYSTEM_PROCESSOR: ${CMAKE_SYSTEM_PROCESSOR}")
 if (MSVC)
-  string(TOLOWER "${CMAKE_GENERATOR_PLATFORM}" CMAKE_GENERATOR_PLATFORM_LWR)
-  message(STATUS "CMAKE_GENERATOR_PLATFORM: ${CMAKE_GENERATOR_PLATFORM}")
+    string(TOLOWER "${CMAKE_GENERATOR_PLATFORM}" CMAKE_GENERATOR_PLATFORM_LWR)
+    message(STATUS "CMAKE_GENERATOR_PLATFORM: ${CMAKE_GENERATOR_PLATFORM}")
 else ()
-  set(CMAKE_GENERATOR_PLATFORM_LWR "")
+    set(CMAKE_GENERATOR_PLATFORM_LWR "")
 endif ()
 
 if (NOT MSVC)
@@ -855,11 +911,21 @@ if (CMAKE_OSX_ARCHITECTURES STREQUAL "arm64" OR CMAKE_GENERATOR_PLATFORM_LWR STR
      CMAKE_SYSTEM_PROCESSOR MATCHES "^(aarch64|arm.*|ARM64)$"))
     message(STATUS "ARM detected")
     if (MSVC)
+        add_compile_definitions(__aarch64__) # MSVC defines _M_ARM64 instead
         add_compile_definitions(__ARM_NEON)
         add_compile_definitions(__ARM_FEATURE_FMA)
-        add_compile_definitions(__ARM_FEATURE_DOTPROD)
-        # add_compile_definitions(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) # MSVC doesn't support vdupq_n_f16, vld1q_f16, vst1q_f16
-        add_compile_definitions(__aarch64__) # MSVC defines _M_ARM64 instead
+
+        set(CMAKE_REQUIRED_FLAGS_PREV ${CMAKE_REQUIRED_FLAGS})
+        string(JOIN " " CMAKE_REQUIRED_FLAGS ${CMAKE_REQUIRED_FLAGS} "/arch:armv8.2")
+        check_cxx_source_compiles("#include <arm_neon.h>\nint main() { int8x16_t _a, _b; int32x4_t _s = vdotq_s32(_s, _a, _b); return 0; }" GGML_COMPILER_SUPPORT_DOTPROD)
+        if (GGML_COMPILER_SUPPORT_DOTPROD)
+            add_compile_definitions(__ARM_FEATURE_DOTPROD)
+        endif ()
+        check_cxx_source_compiles("#include <arm_neon.h>\nint main() { float16_t _a; float16x8_t _s = vdupq_n_f16(_a); return 0; }" GGML_COMPILER_SUPPORT_FP16_VECTOR_ARITHMETIC)
+        if (GGML_COMPILER_SUPPORT_FP16_VECTOR_ARITHMETIC)
+            add_compile_definitions(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC)
+        endif ()
+        set(CMAKE_REQUIRED_FLAGS ${CMAKE_REQUIRED_FLAGS_PREV})
     else()
         check_cxx_compiler_flag(-mfp16-format=ieee COMPILER_SUPPORTS_FP16_FORMAT_I3E)
         if (NOT "${COMPILER_SUPPORTS_FP16_FORMAT_I3E}" STREQUAL "")
@@ -1017,11 +1083,6 @@ endif()
 
 # ggml
 
-if (GGML_USE_CPU_HBM)
-    add_definitions(-DGGML_USE_CPU_HBM)
-    find_library(memkind memkind REQUIRED)
-endif()
-
 add_library(ggml OBJECT
             ggml.c
             ggml.h
@@ -1038,16 +1099,17 @@ add_library(ggml OBJECT
             ${GGML_SOURCES_EXTRA}   ${GGML_HEADERS_EXTRA}
             ${GGML_SOURCES_SYCL}    ${GGML_HEADERS_SYCL}
             ${GGML_SOURCES_KOMPUTE} ${GGML_HEADERS_KOMPUTE}
+            ${GGML_SOURCES_VULKAN}  ${GGML_HEADERS_VULKAN}
+            ${GGML_SOURCES_ROCM}    ${GGML_HEADERS_ROCM}
             )
 
 target_include_directories(ggml PUBLIC . ${LLAMA_EXTRA_INCLUDES})
-target_compile_features(ggml PUBLIC c_std_11) # don't bump
+target_compile_features   (ggml PUBLIC c_std_11) # don't bump
+
 target_link_libraries(ggml PUBLIC Threads::Threads ${LLAMA_EXTRA_LIBS})
-if (GGML_USE_CPU_HBM)
-    target_link_libraries(ggml PUBLIC memkind)
-endif()
 
 add_library(ggml_static STATIC $<TARGET_OBJECTS:ggml>)
+
 if (BUILD_SHARED_LIBS)
     set_target_properties(ggml PROPERTIES POSITION_INDEPENDENT_CODE ON)
     add_library(ggml_shared SHARED $<TARGET_OBJECTS:ggml>)
@@ -1063,7 +1125,8 @@ add_library(llama
             )
 
 target_include_directories(llama PUBLIC .)
-target_compile_features(llama PUBLIC cxx_std_11) # don't bump
+target_compile_features   (llama PUBLIC cxx_std_11) # don't bump
+
 target_link_libraries(llama PRIVATE
     ggml
     ${LLAMA_EXTRA_LIBS}
@@ -1114,7 +1177,7 @@ install(FILES ${CMAKE_CURRENT_BINARY_DIR}/LlamaConfig.cmake
         DESTINATION ${CMAKE_INSTALL_LIBDIR}/cmake/Llama)
 
 set(GGML_PUBLIC_HEADERS "ggml.h" "ggml-alloc.h" "ggml-backend.h"
-        "${GGML_HEADERS_CUDA}" "${GGML_HEADERS_OPENCL}"
+        "${GGML_HEADERS_CUDA}"  "${GGML_HEADERS_OPENCL}"
         "${GGML_HEADERS_METAL}" "${GGML_HEADERS_MPI}" "${GGML_HEADERS_EXTRA}")
 
 set_target_properties(ggml PROPERTIES PUBLIC_HEADER "${GGML_PUBLIC_HEADERS}")

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
 
 #
@@ -569,6 +597,14 @@ $(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 := $(shell nvcc --version | grep -oP 'release (\K[0-9]+\.[0-9])')
+ifeq ($(shell awk -v "v=$(CUDA_VERSION)" 'BEGIN { print (v < 11.7) }'),1)
+ifndef CUDA_DOCKER_ARCH
+ifndef CUDA_POWER_ARCH
+$(error I ERROR: For CUDA versions < 11.7 a target CUDA architecture must be explicitly provided via CUDA_DOCKER_ARCH)
+endif # CUDA_POWER_ARCH
+endif # CUDA_DOCKER_ARCH
+endif # eq ($(shell echo "$(CUDA_VERSION) < 11.7" | bc),1)
 endif # LLAMA_CUBLAS
 $(info )
 
@@ -683,7 +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)
@@ -854,3 +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:**
 
@@ -145,6 +142,7 @@ 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 +154,7 @@ 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)
 
 ---
 
@@ -768,7 +767,7 @@ 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:
 ```
@@ -958,7 +957,7 @@ We have three Docker images available for this project:
 
 1. `ghcr.io/ggerganov/llama.cpp:full`: This image includes both the main executable file and the tools to convert LLaMA models into ggml and convert into 4-bit quantization. (platforms: `linux/amd64`, `linux/arm64`)
 2. `ghcr.io/ggerganov/llama.cpp:light`: This image only includes the main executable file. (platforms: `linux/amd64`, `linux/arm64`)
-3. `ghcr.io/ggerganov/llama.cpp:server`: This image only includes the server executabhle file. (platforms: `linux/amd64`, `linux/arm64`)
+3. `ghcr.io/ggerganov/llama.cpp:server`: This image only includes the server executable file. (platforms: `linux/amd64`, `linux/arm64`)
 
 Additionally, there the following images, similar to the above:
 

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"
@@ -568,6 +568,54 @@ function gg_sum_open_llama_7b_v2 {
     #gg_printf '- shakespeare (q8_0 / f16 base lora):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-lora-q8_0-f16.log)"
 }
 
+# bge-small
+
+function gg_run_embd_bge_small {
+    cd ${SRC}
+
+    gg_wget models-mnt/bge-small/ https://huggingface.co/BAAI/bge-small-en-v1.5/raw/main/config.json
+    gg_wget models-mnt/bge-small/ https://huggingface.co/BAAI/bge-small-en-v1.5/resolve/main/tokenizer.model
+    gg_wget models-mnt/bge-small/ https://huggingface.co/BAAI/bge-small-en-v1.5/raw/main/tokenizer_config.json
+    gg_wget models-mnt/bge-small/ https://huggingface.co/BAAI/bge-small-en-v1.5/raw/main/special_tokens_map.json
+    gg_wget models-mnt/bge-small/ https://huggingface.co/BAAI/bge-small-en-v1.5/resolve/main/pytorch_model.bin
+    gg_wget models-mnt/bge-small/ https://huggingface.co/BAAI/bge-small-en-v1.5/raw/main/sentence_bert_config.json
+    gg_wget models-mnt/bge-small/ https://huggingface.co/BAAI/bge-small-en-v1.5/raw/main/vocab.txt
+    gg_wget models-mnt/bge-small/ https://huggingface.co/BAAI/bge-small-en-v1.5/raw/main/modules.json
+    gg_wget models-mnt/bge-small/ https://huggingface.co/BAAI/bge-small-en-v1.5/raw/main/config.json
+
+    gg_wget models-mnt/bge-small/1_Pooling https://huggingface.co/BAAI/bge-small-en-v1.5/raw/main/1_Pooling/config.json
+
+    path_models="../models-mnt/bge-small"
+
+    rm -rf build-ci-release && mkdir build-ci-release && cd build-ci-release
+
+    set -e
+
+    (time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
+    (time make -j                                            ) 2>&1 | tee -a $OUT/${ci}-make.log
+
+    python3 ../convert-hf-to-gguf.py ${path_models}
+
+    model_f16="${path_models}/ggml-model-f16.gguf"
+    model_q8_0="${path_models}/ggml-model-q8_0.gguf"
+
+    ./bin/quantize ${model_f16} ${model_q8_0} q8_0
+
+    (time ./bin/embedding --model ${model_f16}  -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
+    (time ./bin/embedding --model ${model_q8_0} -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
+
+    set +e
+}
+
+function gg_sum_embd_bge_small {
+    gg_printf '### %s\n\n' "${ci}"
+
+    gg_printf 'BGE Small (BERT):\n'
+    gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
+    gg_printf '- f16: \n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-f16.log)"
+    gg_printf '- q8_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q8_0.log)"
+}
+
 ## main
 
 if [ -z ${GG_BUILD_LOW_PERF} ]; then
@@ -591,6 +639,8 @@ test $ret -eq 0 && gg_run ctest_debug
 test $ret -eq 0 && gg_run ctest_release
 
 if [ -z ${GG_BUILD_LOW_PERF} ]; then
+    test $ret -eq 0 && gg_run embd_bge_small
+
     if [ -z ${GG_BUILD_VRAM_GB} ] || [ ${GG_BUILD_VRAM_GB} -ge 8 ]; then
         if [ -z ${GG_BUILD_CUDA} ]; then
             test $ret -eq 0 && gg_run open_llama_3b_v2

common/common.cpp

@@ -341,7 +341,7 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
                 break;
             }
             const auto sampler_names = string_split(argv[i], ';');
-            sparams.samplers_sequence = sampler_types_from_names(sampler_names);
+            sparams.samplers_sequence = sampler_types_from_names(sampler_names, true);
         } else if (arg == "--sampling-seq") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -671,7 +671,15 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
         } else if (arg == "--no-mmap") {
             params.use_mmap = false;
         } else if (arg == "--numa") {
-            params.numa = true;
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            std::string value(argv[i]);
+            /**/ if (value == "distribute" || value == "") { params.numa = GGML_NUMA_STRATEGY_DISTRIBUTE; }
+            else if (value == "isolate") { params.numa = GGML_NUMA_STRATEGY_ISOLATE; }
+            else if (value == "numactl") { params.numa = GGML_NUMA_STRATEGY_NUMACTL; }
+            else { invalid_param = true; break; }
         } else if (arg == "--verbose-prompt") {
             params.verbose_prompt = true;
         } else if (arg == "--no-display-prompt") {
@@ -935,7 +943,7 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     printf("  -tb N, --threads-batch N\n");
     printf("                        number of threads to use during batch and prompt processing (default: same as --threads)\n");
     printf("  -td N, --threads-draft N");
-    printf("                        number of threads to use during generation (default: same as --threads)");
+    printf("                        number of threads to use during generation (default: same as --threads)\n");
     printf("  -tbd N, --threads-batch-draft N\n");
     printf("                        number of threads to use during batch and prompt processing (default: same as --threads-draft)\n");
     printf("  -p PROMPT, --prompt PROMPT\n");
@@ -956,7 +964,8 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     printf("  -n N, --n-predict N   number of tokens to predict (default: %d, -1 = infinity, -2 = until context filled)\n", params.n_predict);
     printf("  -c N, --ctx-size N    size of the prompt context (default: %d, 0 = loaded from model)\n", params.n_ctx);
     printf("  -b N, --batch-size N  batch size for prompt processing (default: %d)\n", params.n_batch);
-    printf("  --samplers            samplers that will be used for generation in the order, separated by \';\' (default: %s)\n", sampler_type_names.c_str());
+    printf("  --samplers            samplers that will be used for generation in the order, separated by \';\'\n");
+    printf("                        (default: %s)\n", sampler_type_names.c_str());
     printf("  --sampling-seq        simplified sequence for samplers that will be used (default: %s)\n", sampler_type_chars.c_str());
     printf("  --top-k N             top-k sampling (default: %d, 0 = disabled)\n", sparams.top_k);
     printf("  --top-p N             top-p sampling (default: %.1f, 1.0 = disabled)\n", (double)sparams.top_p);
@@ -1005,7 +1014,7 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     printf("  --winogrande-tasks N  number of tasks to use when computing the Winogrande score (default: %zu)\n", params.winogrande_tasks);
     printf("  --multiple-choice     compute multiple choice score over random tasks from datafile supplied with -f\n");
     printf("  --multiple-choice-tasks N number of tasks to use when computing the multiple choice score (default: %zu)\n", params.winogrande_tasks);
-    printf("  --kl-divergence       computes KL-divergence to logits provided via --kl-divergence-base");
+    printf("  --kl-divergence       computes KL-divergence to logits provided via --kl-divergence-base\n");
     printf("  --keep N              number of tokens to keep from the initial prompt (default: %d, -1 = all)\n", params.n_keep);
     printf("  --draft N             number of tokens to draft for speculative decoding (default: %d)\n", params.n_draft);
     printf("  --chunks N            max number of chunks to process (default: %d, -1 = all)\n", params.n_chunks);
@@ -1022,7 +1031,10 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     if (llama_supports_mmap()) {
         printf("  --no-mmap             do not memory-map model (slower load but may reduce pageouts if not using mlock)\n");
     }
-    printf("  --numa                attempt optimizations that help on some NUMA systems\n");
+    printf("  --numa TYPE           attempt optimizations that help on some NUMA systems\n");
+    printf("                          - distribute: spread execution evenly over all nodes\n");
+    printf("                          - isolate: only spawn threads on CPUs on the node that execution started on\n");
+    printf("                          - numactl: use the CPU map provided by numactl\n");
     printf("                        if run without this previously, it is recommended to drop the system page cache before using this\n");
     printf("                        see https://github.com/ggerganov/llama.cpp/issues/1437\n");
     if (llama_supports_gpu_offload()) {
@@ -1122,34 +1134,50 @@ std::vector<std::string> string_split(std::string input, char separator) {
     return parts;
 }
 
-std::vector<llama_sampler_type> sampler_types_from_names(const std::vector<std::string> & names) {
+std::vector<llama_sampler_type> sampler_types_from_names(const std::vector<std::string> & names, bool allow_alt_names) {
+    std::unordered_map<std::string, llama_sampler_type> sampler_canonical_name_map {
+        {"top_k",       llama_sampler_type::TOP_K},
+        {"top_p",       llama_sampler_type::TOP_P},
+        {"typical_p",   llama_sampler_type::TYPICAL_P},
+        {"min_p",       llama_sampler_type::MIN_P},
+        {"tfs_z",       llama_sampler_type::TFS_Z},
+        {"temperature", llama_sampler_type::TEMPERATURE}
+    };
+
     // since samplers names are written multiple ways
     // make it ready for both system names and input names
-    std::unordered_map<std::string, llama_sampler_type> sampler_name_map {
-        {"top_k",       llama_sampler_type::TOP_K},
+    std::unordered_map<std::string, llama_sampler_type> sampler_alt_name_map {
         {"top-k",       llama_sampler_type::TOP_K},
-        {"top_p",       llama_sampler_type::TOP_P},
         {"top-p",       llama_sampler_type::TOP_P},
         {"nucleus",     llama_sampler_type::TOP_P},
-        {"typical_p",   llama_sampler_type::TYPICAL_P},
         {"typical-p",   llama_sampler_type::TYPICAL_P},
         {"typical",     llama_sampler_type::TYPICAL_P},
-        {"min_p",       llama_sampler_type::MIN_P},
         {"min-p",       llama_sampler_type::MIN_P},
-        {"tfs_z",       llama_sampler_type::TFS_Z},
         {"tfs-z",       llama_sampler_type::TFS_Z},
         {"tfs",         llama_sampler_type::TFS_Z},
-        {"temp",        llama_sampler_type::TEMP},
-        {"temperature", llama_sampler_type::TEMP}
+        {"temp",        llama_sampler_type::TEMPERATURE}
     };
 
     std::vector<llama_sampler_type> sampler_types;
     sampler_types.reserve(names.size());
-    for (const auto& name : names) {
-        const auto sampler_item = sampler_name_map.find(name);
-        if (sampler_item != sampler_name_map.end()) {
+    for (const auto & name : names)
+    {
+        auto sampler_item = sampler_canonical_name_map.find(name);
+        if (sampler_item != sampler_canonical_name_map.end())
+        {
             sampler_types.push_back(sampler_item->second);
         }
+        else
+        {
+            if (allow_alt_names)
+            {
+                sampler_item = sampler_alt_name_map.find(name);
+                if (sampler_item != sampler_alt_name_map.end())
+                {
+                    sampler_types.push_back(sampler_item->second);
+                }
+            }
+        }
     }
     return sampler_types;
 }
@@ -1161,7 +1189,7 @@ std::vector<llama_sampler_type> sampler_types_from_chars(const std::string & nam
         {'y', llama_sampler_type::TYPICAL_P},
         {'m', llama_sampler_type::MIN_P},
         {'f', llama_sampler_type::TFS_Z},
-        {'t', llama_sampler_type::TEMP}
+        {'t', llama_sampler_type::TEMPERATURE}
     };
 
     std::vector<llama_sampler_type> sampler_types;
@@ -1177,12 +1205,12 @@ std::vector<llama_sampler_type> sampler_types_from_chars(const std::string & nam
 
 std::string sampler_type_to_name_string(llama_sampler_type sampler_type) {
     switch (sampler_type) {
-        case llama_sampler_type::TOP_K:     return "top_k";
-        case llama_sampler_type::TFS_Z:     return "tfs_z";
-        case llama_sampler_type::TYPICAL_P: return "typical_p";
-        case llama_sampler_type::TOP_P:     return "top_p";
-        case llama_sampler_type::MIN_P:     return "min_p";
-        case llama_sampler_type::TEMP:      return "temp";
+        case llama_sampler_type::TOP_K:       return "top_k";
+        case llama_sampler_type::TFS_Z:       return "tfs_z";
+        case llama_sampler_type::TYPICAL_P:   return "typical_p";
+        case llama_sampler_type::TOP_P:       return "top_p";
+        case llama_sampler_type::MIN_P:       return "min_p";
+        case llama_sampler_type::TEMPERATURE: return "temperature";
         default : return "";
     }
 }
@@ -1676,6 +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);
@@ -1689,7 +1718,6 @@ void dump_non_result_info_yaml(FILE * stream, const gpt_params & params, const l
     fprintf(stream, "no_mmap: %s # default: false\n", !params.use_mmap ? "true" : "false");
     fprintf(stream, "no_mul_mat_q: %s # default: false\n", !params.mul_mat_q ? "true" : "false");
     fprintf(stream, "no_penalize_nl: %s # default: false\n", !sparams.penalize_nl ? "true" : "false");
-    fprintf(stream, "numa: %s # default: false\n", params.numa ? "true" : "false");
     fprintf(stream, "ppl_output_type: %d # default: 0\n", params.ppl_output_type);
     fprintf(stream, "ppl_stride: %d # default: 0\n", params.ppl_stride);
     fprintf(stream, "presence_penalty: %f # default: 0.0\n", sparams.penalty_present);
@@ -1714,7 +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);
@@ -1723,7 +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);
@@ -1774,7 +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

@@ -76,6 +76,7 @@ struct gpt_params {
     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;
+    ggml_numa_strategy numa       = GGML_NUMA_STRATEGY_DISABLED;
 
     // // sampling parameters
     struct llama_sampling_params sparams;
@@ -134,7 +135,6 @@ struct gpt_params {
     bool logits_all        = false; // return logits for all tokens in the batch
     bool use_mmap          = true;  // use mmap for faster loads
     bool use_mlock         = false; // use mlock to keep model in memory
-    bool numa              = false; // attempt optimizations that help on some NUMA systems
     bool verbose_prompt    = false; // print prompt tokens before generation
     bool display_prompt    = true;  // print prompt before generation
     bool infill            = false; // use infill mode
@@ -165,7 +165,7 @@ void process_escapes(std::string& input);
 // String utils
 //
 
-std::vector<llama_sampler_type> sampler_types_from_names(const std::vector<std::string> & names);
+std::vector<llama_sampler_type> sampler_types_from_names(const std::vector<std::string> & names, bool allow_alt_names);
 std::vector<llama_sampler_type> sampler_types_from_chars(const std::string & names_string);
 std::vector<std::string> string_split(std::string input, char separator);
 std::string sampler_type_to_name_string(llama_sampler_type sampler_type);

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;
@@ -139,7 +139,7 @@ static void sampler_queue(
             case llama_sampler_type::TYPICAL_P: llama_sample_typical  (ctx_main, &cur_p, typical_p, min_keep); break;
             case llama_sampler_type::TOP_P    : llama_sample_top_p    (ctx_main, &cur_p, top_p,     min_keep); break;
             case llama_sampler_type::MIN_P    : llama_sample_min_p    (ctx_main, &cur_p, min_p,     min_keep); break;
-            case llama_sampler_type::TEMP:
+            case llama_sampler_type::TEMPERATURE:
                 if (dynatemp_range > 0) {
                     float dynatemp_min = std::max(0.0f, temp - dynatemp_range);
                     float dynatemp_max = std::max(0.0f, temp + dynatemp_range);
@@ -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);
 

common/sampling.h

@@ -10,18 +10,19 @@
 
 // sampler types
 enum class llama_sampler_type : char {
-    TOP_K     = 'k',
-    TOP_P     = 'p',
-    MIN_P     = 'm',
-    TFS_Z     = 'f',
-    TYPICAL_P = 'y',
-    TEMP      = 't'
+    TOP_K       = 'k',
+    TOP_P       = 'p',
+    MIN_P       = 'm',
+    TFS_Z       = 'f',
+    TYPICAL_P   = 'y',
+    TEMPERATURE = 't'
 };
 
 // sampling parameters
 typedef struct llama_sampling_params {
     int32_t     n_prev                = 64;       // number of previous tokens to remember
     int32_t     n_probs               = 0;        // if greater than 0, output the probabilities of top n_probs tokens.
+    int32_t     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
@@ -45,7 +46,7 @@ typedef struct llama_sampling_params {
         llama_sampler_type::TYPICAL_P,
         llama_sampler_type::TOP_P,
         llama_sampler_type::MIN_P,
-        llama_sampler_type::TEMP
+        llama_sampler_type::TEMPERATURE
     };
 
     std::string grammar;  // optional BNF-like grammar to constrain sampling

convert-hf-to-gguf.py

@@ -10,7 +10,7 @@ import re
 import sys
 from enum import IntEnum
 from pathlib import Path
-from typing import TYPE_CHECKING, Any, ContextManager, Iterator, cast
+from typing import TYPE_CHECKING, Any, ContextManager, Iterator, Sequence, cast
 
 import numpy as np
 import torch
@@ -25,15 +25,6 @@ import gguf
 from convert import HfVocab
 
 
-# check for any of the given keys in the dictionary and return the value of the first key found
-def get_key_opts(d, keys):
-    for k in keys:
-        if k in d:
-            return d[k]
-    print(f"Could not find any of {keys}")
-    sys.exit()
-
-
 ###### MODEL DEFINITIONS ######
 
 class SentencePieceTokenTypes(IntEnum):
@@ -58,6 +49,15 @@ class Model:
         self.hparams = Model.load_hparams(self.dir_model)
         self.model_arch = self._get_model_architecture()
         self.gguf_writer = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[self.model_arch], endianess=self.endianess, use_temp_file=False)
+        self.block_count = self.find_hparam(["n_layers", "num_hidden_layers", "n_layer"])
+
+    def find_hparam(self, keys: Sequence[str], optional: bool = False) -> Any:
+        key = next((k for k in keys if k in self.hparams), None)
+        if key is not None:
+            return self.hparams[key]
+        if optional:
+            return None
+        raise KeyError(f"could not find any of: {keys}")
 
     def set_vocab(self):
         self._set_vocab_gpt2()
@@ -79,28 +79,33 @@ class Model:
 
     def set_gguf_parameters(self):
         self.gguf_writer.add_name(self.dir_model.name)
-        self.gguf_writer.add_block_count(self.hparams.get(
-            "n_layers", self.hparams.get("num_hidden_layers", self.hparams.get("n_layer")),
-        ))
-        if (n_ctx := self.hparams.get("max_position_embeddings")) is not None:
+        self.gguf_writer.add_block_count(self.block_count)
+
+        if (n_ctx := self.find_hparam(["max_position_embeddings", "n_ctx"], optional=True)) is not None:
             self.gguf_writer.add_context_length(n_ctx)
-        if (n_embd := self.hparams.get("hidden_size")) is not None:
-            self.gguf_writer.add_embedding_length(n_embd)
-        if (n_ff := self.hparams.get("intermediate_size")) is not None:
+
+        n_embd = self.find_hparam(["hidden_size", "n_embd"])
+        self.gguf_writer.add_embedding_length(n_embd)
+
+        if (n_ff := self.find_hparam(["intermediate_size", "n_inner"], optional=True)) is not None:
             self.gguf_writer.add_feed_forward_length(n_ff)
-        if (n_head := self.hparams.get("num_attention_heads")) is not None:
-            self.gguf_writer.add_head_count(n_head)
+
+        n_head = self.find_hparam(["num_attention_heads", "n_head"])
+        self.gguf_writer.add_head_count(n_head)
+
         if (n_head_kv := self.hparams.get("num_key_value_heads")) is not None:
             self.gguf_writer.add_head_count_kv(n_head_kv)
 
-        if (n_rms_eps := self.hparams.get("rms_norm_eps")) is not None:
-            self.gguf_writer.add_layer_norm_rms_eps(n_rms_eps)
+        if (f_rms_eps := self.hparams.get("rms_norm_eps")) is not None:
+            self.gguf_writer.add_layer_norm_rms_eps(f_rms_eps)
+        if (f_norm_eps := self.find_hparam(["layer_norm_eps", "layer_norm_epsilon"], optional=True)) is not None:
+            self.gguf_writer.add_layer_norm_eps(f_norm_eps)
         if (n_experts := self.hparams.get("num_local_experts")) is not None:
             self.gguf_writer.add_expert_count(n_experts)
         if (n_experts_used := self.hparams.get("num_experts_per_tok")) is not None:
             self.gguf_writer.add_expert_used_count(n_experts_used)
 
-        self.gguf_writer.add_parallel_residual(self.hparams.get("use_parallel_residual", True))
+        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")))
@@ -211,6 +216,10 @@ class Model:
             return MiniCPMModel
         if model_architecture == "BertModel":
             return BertModel
+        if model_architecture == "NomicBertModel":
+            return NomicBertModel
+        if model_architecture == "GemmaForCausalLM":
+            return GemmaModel
         return Model
 
     def _is_model_safetensors(self) -> bool:
@@ -268,6 +277,10 @@ class Model:
             return gguf.MODEL_ARCH.MINICPM
         if arch == "BertModel":
             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!')
 
@@ -646,6 +659,8 @@ 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):
@@ -1022,7 +1037,6 @@ 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()
@@ -1297,21 +1311,21 @@ class GPT2Model(Model):
 
 class Phi2Model(Model):
     def set_gguf_parameters(self):
-        block_count = get_key_opts(self.hparams, ["num_hidden_layers", "n_layer"])
+        block_count = self.find_hparam(["num_hidden_layers", "n_layer"])
 
-        rot_pct = get_key_opts(self.hparams, ["partial_rotary_factor"])
-        n_embd = get_key_opts(self.hparams, ["hidden_size", "n_embd"])
-        n_head = get_key_opts(self.hparams, ["num_attention_heads", "n_head"])
+        rot_pct = self.find_hparam(["partial_rotary_factor"])
+        n_embd = self.find_hparam(["hidden_size", "n_embd"])
+        n_head = self.find_hparam(["num_attention_heads", "n_head"])
 
         self.gguf_writer.add_name("Phi2")
-        self.gguf_writer.add_context_length(get_key_opts(self.hparams, ["n_positions", "max_position_embeddings"]))
+        self.gguf_writer.add_context_length(self.find_hparam(["n_positions", "max_position_embeddings"]))
 
         self.gguf_writer.add_embedding_length(n_embd)
         self.gguf_writer.add_feed_forward_length(4 * n_embd)
         self.gguf_writer.add_block_count(block_count)
         self.gguf_writer.add_head_count(n_head)
         self.gguf_writer.add_head_count_kv(n_head)
-        self.gguf_writer.add_layer_norm_eps(get_key_opts(self.hparams, ["layer_norm_epsilon", "layer_norm_eps"]))
+        self.gguf_writer.add_layer_norm_eps(self.find_hparam(["layer_norm_epsilon", "layer_norm_eps"]))
         self.gguf_writer.add_rope_dimension_count(int(rot_pct * n_embd) // n_head)
         self.gguf_writer.add_file_type(self.ftype)
         self.gguf_writer.add_add_bos_token(False)
@@ -1636,19 +1650,34 @@ in chat mode so that the conversation can end normally.")
 class BertModel(Model):
     def __init__(self, *args, **kwargs):
         super().__init__(*args, **kwargs)
-        self.block_count = self.hparams["num_hidden_layers"]
+        self.vocab_size = None
 
     def set_gguf_parameters(self):
-        # TODO(cebtenzzre): merge with parent class
-        self.gguf_writer.add_name(self.dir_model.name)
-        self.gguf_writer.add_context_length(self.hparams["max_position_embeddings"])
-        self.gguf_writer.add_embedding_length(self.hparams["hidden_size"])
-        self.gguf_writer.add_feed_forward_length(self.hparams["intermediate_size"])
-        self.gguf_writer.add_block_count(self.block_count)
-        self.gguf_writer.add_head_count(self.hparams["num_attention_heads"])
-        self.gguf_writer.add_layer_norm_eps(self.hparams["layer_norm_eps"])
+        super().set_gguf_parameters()
         self.gguf_writer.add_causal_attention(False)
-        self.gguf_writer.add_file_type(self.ftype)
+
+        # get pooling path
+        with open(self.dir_model / "modules.json", encoding="utf-8") as f:
+            modules = json.load(f)
+        pooling_path = None
+        for mod in modules:
+            if mod["type"] == "sentence_transformers.models.Pooling":
+                pooling_path = mod["path"]
+                break
+
+        # get pooling type
+        pooling_type = gguf.PoolingType.NONE
+        if pooling_path is not None:
+            with open(self.dir_model / pooling_path / "config.json", encoding="utf-8") as f:
+                pooling = json.load(f)
+            if pooling["pooling_mode_mean_tokens"]:
+                pooling_type = gguf.PoolingType.MEAN
+            elif pooling["pooling_mode_cls_token"]:
+                pooling_type = gguf.PoolingType.CLS
+            else:
+                raise NotImplementedError("Only MEAN and CLS pooling types supported")
+
+        self.gguf_writer.add_pooling_type(pooling_type.value)
 
     def set_vocab(self):
         path = self.dir_model
@@ -1658,6 +1687,7 @@ class BertModel(Model):
         vocab = HfVocab(path, added_tokens_path)
         tokens, scores, toktypes = zip(*vocab.all_tokens())
         assert len(tokens) == vocab.vocab_size
+        self.vocab_size = vocab.vocab_size
 
         # we need this to validate the size of the token_type embeddings
         # though currently we are passing all zeros to the token_type embeddings
@@ -1671,7 +1701,7 @@ class BertModel(Model):
             if tok.startswith(b"##"):
                 return tok[2:]
             return b"\xe2\x96\x81" + tok
-        tokens = [phantom(t, y) for t, y in zip(tokens, toktypes)]
+        tokens = tuple(phantom(t, y) for t, y in zip(tokens, toktypes))
 
         # set up bos and eos tokens (cls and sep)
         self.gguf_writer.add_bos_token_id(vocab.tokenizer.cls_token_id)
@@ -1723,6 +1753,99 @@ class BertModel(Model):
             self.gguf_writer.add_tensor(new_name, data)
 
 
+class NomicBertModel(BertModel):
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+
+        # the HF config claims n_ctx=8192, but it uses RoPE scaling
+        self.hparams["n_ctx"] = 2048
+
+        # SwigLU activation
+        assert self.hparams["activation_function"] == "swiglu"
+        # this doesn't do anything in the HF version
+        assert self.hparams["causal"] is False
+        # no bias tensors
+        assert self.hparams["qkv_proj_bias"] is False
+        assert self.hparams["mlp_fc1_bias"] is False
+        assert self.hparams["mlp_fc2_bias"] is False
+        # norm at end of layer
+        assert self.hparams["prenorm"] is False
+        # standard RoPE
+        assert self.hparams["rotary_emb_fraction"] == 1.0
+        assert self.hparams["rotary_emb_interleaved"] is False
+        assert self.hparams["rotary_emb_scale_base"] is None
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        self.gguf_writer.add_rope_freq_base(self.hparams["rotary_emb_base"])
+
+    def get_tensors(self):
+        assert self.vocab_size is not None
+        for name, data in super().get_tensors():
+            # Nomic Embed's token embeddings tensor is padded, but llama.cpp wants tensor sizes to match exactly.
+            if name == 'embeddings.word_embeddings.weight' and data.shape[1] != self.vocab_size:
+                rounded_vocab_size = (self.vocab_size + 63) // 64 * 64
+                assert data.shape == (rounded_vocab_size, self.hparams["n_embd"])
+                data = data[:self.vocab_size, :]
+            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"])
+
+    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 ######
 
 

convert.py

@@ -1173,7 +1173,7 @@ def convert_to_output_type(model: LazyModel, output_type: GGMLFileType) -> LazyM
             for (name, tensor) in model.items()}
 
 
-def convert_model_names(model: LazyModel, params: Params) -> LazyModel:
+def convert_model_names(model: LazyModel, params: Params, skip_unknown: bool) -> LazyModel:
     tmap = gguf.TensorNameMap(ARCH, params.n_layer)
     should_skip: set[gguf.MODEL_TENSOR] = set(gguf.MODEL_TENSOR_SKIP.get(ARCH, []))
 
@@ -1199,7 +1199,11 @@ def convert_model_names(model: LazyModel, params: Params) -> LazyModel:
     for name, lazy_tensor in model.items():
         tensor_type, name_new = tmap.get_type_and_name(name, try_suffixes = (".weight", ".bias")) or (None, None)
         if name_new is None:
-            raise Exception(f"Unexpected tensor name: {name}")
+            if skip_unknown:
+                print(f"Unexpected tensor name: {name} - skipping")
+                continue
+            else:
+                raise Exception(f"Unexpected tensor name: {name}. Use --skip-unknown to ignore it (e.g. LLaVA)")
 
         if tensor_type in should_skip:
             print(f"skipping tensor {name_new}")
@@ -1377,19 +1381,20 @@ def main(args_in: list[str] | None = None) -> None:
         output_choices.append("q8_0")
     vocab_types = ["spm", "bpe", "hfft"]
     parser = argparse.ArgumentParser(description="Convert a LLaMa model to a GGML compatible file")
-    parser.add_argument("--awq-path",    type=Path,              help="Path to scale awq cache file", default=None)
-    parser.add_argument("--dump",        action="store_true",    help="don't convert, just show what's in the model")
-    parser.add_argument("--dump-single", action="store_true",    help="don't convert, just show what's in a single model file")
-    parser.add_argument("--vocab-only",  action="store_true",    help="extract only the vocab")
-    parser.add_argument("--outtype",     choices=output_choices, help="output format - note: q8_0 may be very slow (default: f16 or f32 based on input)")
-    parser.add_argument("--vocab-dir",   type=Path,              help="directory containing tokenizer.model, if separate from model file")
-    parser.add_argument("--vocab-type",  choices=vocab_types,    help="The vocabulary format used to define the tokenizer model (default: spm)", default="spm")
-    parser.add_argument("--outfile",     type=Path,              help="path to write to; default: based on input")
-    parser.add_argument("model",         type=Path,              help="directory containing model file, or model file itself (*.pth, *.pt, *.bin)")
-    parser.add_argument("--ctx",         type=int,               help="model training context (default: based on input)")
-    parser.add_argument("--concurrency", type=int,               help=f"concurrency used for conversion (default: {DEFAULT_CONCURRENCY})", default=DEFAULT_CONCURRENCY)
-    parser.add_argument("--big-endian",  action="store_true",    help="model is executed on big endian machine")
-    parser.add_argument("--pad-vocab",   action="store_true",    help="add pad tokens when model vocab expects more than tokenizer metadata provides")
+    parser.add_argument("--awq-path",     type=Path,              help="Path to scale awq cache file", default=None)
+    parser.add_argument("--dump",         action="store_true",    help="don't convert, just show what's in the model")
+    parser.add_argument("--dump-single",  action="store_true",    help="don't convert, just show what's in a single model file")
+    parser.add_argument("--vocab-only",   action="store_true",    help="extract only the vocab")
+    parser.add_argument("--outtype",      choices=output_choices, help="output format - note: q8_0 may be very slow (default: f16 or f32 based on input)")
+    parser.add_argument("--vocab-dir",    type=Path,              help="directory containing tokenizer.model, if separate from model file")
+    parser.add_argument("--vocab-type",   choices=vocab_types,    help="The vocabulary format used to define the tokenizer model (default: spm)", default="spm")
+    parser.add_argument("--outfile",      type=Path,              help="path to write to; default: based on input")
+    parser.add_argument("model",          type=Path,              help="directory containing model file, or model file itself (*.pth, *.pt, *.bin)")
+    parser.add_argument("--ctx",          type=int,               help="model training context (default: based on input)")
+    parser.add_argument("--concurrency",  type=int,               help=f"concurrency used for conversion (default: {DEFAULT_CONCURRENCY})", default=DEFAULT_CONCURRENCY)
+    parser.add_argument("--big-endian",   action="store_true",    help="model is executed on big endian machine")
+    parser.add_argument("--pad-vocab",    action="store_true",    help="add pad tokens when model vocab expects more than tokenizer metadata provides")
+    parser.add_argument("--skip-unknown", action="store_true",    help="skip unknown tensor names instead of failing")
 
     args = parser.parse_args(args_in)
     if args.awq_path:
@@ -1461,7 +1466,7 @@ def main(args_in: list[str] | None = None) -> None:
     print(f"Special vocab info: {special_vocab}")
 
     model   = model_plus.model
-    model   = convert_model_names(model, params)
+    model   = convert_model_names(model, params, args.skip_unknown)
     ftype   = pick_output_type(model, args.outtype)
     model   = convert_to_output_type(model, ftype)
     outfile = args.outfile or default_outfile(model_plus.paths, ftype)

examples/CMakeLists.txt

@@ -38,6 +38,7 @@ else()
     add_subdirectory(speculative)
     add_subdirectory(lookahead)
     add_subdirectory(lookup)
+    add_subdirectory(gguf)
     add_subdirectory(train-text-from-scratch)
     add_subdirectory(imatrix)
     if (LLAMA_BUILD_SERVER)

examples/baby-llama/baby-llama.cpp

@@ -1533,16 +1533,17 @@ 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);
 
@@ -1552,8 +1553,8 @@ int main(int argc, char ** argv) {
         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

@@ -82,7 +82,8 @@ int main(int argc, char ** argv) {
 
     // init LLM
 
-    llama_backend_init(params.numa);
+    llama_backend_init();
+    llama_numa_init(params.numa);
 
     // initialize the model
 
@@ -158,7 +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.swift/Sources/main.swift

@@ -17,7 +17,7 @@ let n_parallel: Int = arguments.count > 3 && Int(arguments[3]) != nil ? Int(argu
 let n_len: Int = 32
 
 // init LLM
-llama_backend_init(false)
+llama_backend_init()
 defer {
     llama_backend_free()
 }

examples/batched/batched.cpp

@@ -50,7 +50,8 @@ int main(int argc, char ** argv) {
 
     // init LLM
 
-    llama_backend_init(params.numa);
+    llama_backend_init();
+    llama_numa_init(params.numa);
 
     // initialize the model
 
@@ -91,7 +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/beam-search/beam-search.cpp

@@ -119,7 +119,8 @@ int main(int argc, char ** argv)
     // Init LLM :
     //---------------------------------
 
-    llama_backend_init(params.numa);
+    llama_backend_init();
+    llama_numa_init(params.numa);
 
     llama_model * model;
     llama_context * ctx;

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

@@ -325,14 +325,14 @@ 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);
+    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_mult:  %u\n", __func__, params->n_mult);
+    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 init_model(struct my_llama_model * model) {
@@ -350,25 +350,25 @@ 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);
+    printf("[%s:GG] Allocating [%u] x [%u] = [%u] 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);
+    printf("[%s:GG] Allocating [%u] 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);
+    printf("[%s:GG] Allocating [%u] x[%u] = [%u] 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 [%u] x[%u] = [%u] float space for layer.wq for [%u] layers\n",__func__, n_embd, n_embd, n_embd * n_embd, n_layer);
+    printf("[%s:GG] Allocating [%u] x[%u] = [%u] float space for layer.wk for [%u] layers\n",__func__, n_embd, n_embd, n_embd * n_embd, n_layer);
+    printf("[%s:GG] Allocating [%u] x[%u] = [%u] float space for layer.wv for [%u] layers\n",__func__, n_embd, n_embd, n_embd * n_embd, n_layer);
+    printf("[%s:GG] Allocating [%u] x[%u] = [%u] float space for layer.wo for [%u] 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 [%u] float space for layer.ffn_norm for [%u] 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);
+    printf("[%s:GG] Allocating [%u] x[%u] = [%u] float space for layer.w1 for [%u] layers\n",__func__, n_ff, n_embd, n_embd * n_ff, n_layer);
+    printf("[%s:GG] Allocating [%u] x[%u] = [%u] float space for layer.w2 for [%u] layers\n",__func__, n_embd, n_ff, n_ff * n_embd, n_layer);
+    printf("[%s:GG] Allocating [%u] x[%u] = [%u] float space for layer.w3 for [%u] 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");

examples/embedding/embedding.cpp

@@ -7,6 +7,51 @@
 #pragma warning(disable: 4244 4267) // possible loss of data
 #endif
 
+static std::vector<std::string> split_lines(const std::string & s) {
+    std::string line;
+    std::vector<std::string> lines;
+    std::stringstream ss(s);
+    while (std::getline(ss, line)) {
+        lines.push_back(line);
+    }
+    return lines;
+}
+
+static void batch_add_seq(llama_batch & batch, const std::vector<int32_t> & tokens, int seq_id) {
+    for (size_t i = 0; i < tokens.size(); i++) {
+        llama_batch_add(batch, tokens[i], i, { seq_id }, false);
+    }
+}
+
+static void normalize(float * vec, float * out, int n) {
+    float norm = 0;
+    for (int i = 0; i < n; i++) {
+        norm += vec[i] * vec[i];
+    }
+    norm = sqrt(norm);
+    for (int i = 0; i < n; i++) {
+        out[i] = vec[i] / norm;
+    }
+}
+
+static void batch_decode(llama_context * ctx, llama_batch & batch, float * output, int n_seq, int n_embd) {
+    // clear previous kv_cache values (irrelevant for embeddings)
+    llama_kv_cache_clear(ctx);
+
+    // run model
+    fprintf(stderr, "%s: n_tokens = %d, n_seq = %d\n", __func__, batch.n_tokens, n_seq);
+    if (llama_decode(ctx, batch) < 0) {
+        fprintf(stderr, "%s : failed to decode\n", __func__);
+    }
+
+    // normalize on copy
+    for (int k = 0; k < n_seq; k++) {
+        float * emb = llama_get_embeddings_ith(ctx, k);
+        float * out = output + k * n_embd;
+        normalize(emb, out, n_embd);
+    }
+}
+
 int main(int argc, char ** argv) {
     gpt_params params;
 
@@ -29,7 +74,8 @@ int main(int argc, char ** argv) {
         params.prompt = gpt_random_prompt(rng);
     }
 
-    llama_backend_init(params.numa);
+    llama_backend_init();
+    llama_numa_init(params.numa);
 
     llama_model * model;
     llama_context * ctx;
@@ -55,59 +101,84 @@ int main(int argc, char ** argv) {
         fprintf(stderr, "%s\n", get_system_info(params).c_str());
     }
 
-    int n_past = 0;
+    // split the prompt into lines
+    std::vector<std::string> prompts = split_lines(params.prompt);
 
-    // tokenize the prompt
-    auto embd_inp = ::llama_tokenize(ctx, params.prompt, true);
+    // max batch size
+    const uint64_t n_batch = params.n_batch;
+    GGML_ASSERT(params.n_batch == params.n_ctx);
 
+    // tokenize the prompts and trim
+    std::vector<std::vector<int32_t>> inputs;
+    for (const auto & prompt : prompts) {
+        auto inp = ::llama_tokenize(ctx, prompt, true);
+        if (inp.size() > n_batch) {
+            inp.resize(n_batch);
+        }
+        inputs.push_back(inp);
+    }
+
+    // tokenization stats
     if (params.verbose_prompt) {
-        fprintf(stderr, "\n");
-        fprintf(stderr, "%s: prompt: '%s'\n", __func__, params.prompt.c_str());
-        fprintf(stderr, "%s: number of tokens in prompt = %zu\n", __func__, embd_inp.size());
-        for (int i = 0; i < (int) embd_inp.size(); i++) {
-            fprintf(stderr, "%6d -> '%s'\n", embd_inp[i], llama_token_to_piece(ctx, embd_inp[i]).c_str());
+        for (int i = 0; i < (int) inputs.size(); i++) {
+            fprintf(stderr, "%s: prompt %d: '%s'\n", __func__, i, prompts[i].c_str());
+            fprintf(stderr, "%s: number of tokens in prompt = %zu\n", __func__, inputs[i].size());
+            for (int j = 0; j < (int) inputs[i].size(); j++) {
+                fprintf(stderr, "%6d -> '%s'\n", inputs[i][j], llama_token_to_piece(ctx, inputs[i][j]).c_str());
+            }
+            fprintf(stderr, "\n\n");
         }
-        fprintf(stderr, "\n");
     }
 
-    if (embd_inp.size() > (size_t)n_ctx) {
-        fprintf(stderr, "%s: error: prompt is longer than the context window (%zu tokens, n_ctx = %d)\n",
-                __func__, embd_inp.size(), n_ctx);
-        return 1;
-    }
-
-    while (!embd_inp.empty()) {
-        int n_tokens = std::min(params.n_batch, (int) embd_inp.size());
-        if (llama_decode(ctx, llama_batch_get_one(embd_inp.data(), n_tokens, n_past, 0))) {
-            fprintf(stderr, "%s : failed to eval\n", __func__);
-            return 1;
-        }
-        n_past += n_tokens;
-        embd_inp.erase(embd_inp.begin(), embd_inp.begin() + n_tokens);
-    }
+    // initialize batch
+    const int n_prompts = prompts.size();
+    struct llama_batch batch = llama_batch_init(n_batch, 0, n_prompts);
 
+    // allocate output
     const int n_embd = llama_n_embd(model);
-    auto * embeddings = llama_get_embeddings(ctx);
+    std::vector<float> embeddings(n_prompts * n_embd, 0);
+    float * emb = embeddings.data();
 
-    // l2-normalize embeddings
-    float norm = 0;
-    for (int i = 0; i < n_embd; i++) {
-        norm += embeddings[i] * embeddings[i];
-    }
-    norm = sqrt(norm);
-    for (int i = 0; i < n_embd; i++) {
-        embeddings[i] /= norm;
+    // break into batches
+    int p = 0; // number of prompts processed already
+    int s = 0; // number of prompts in current batch
+    for (int k = 0; k < n_prompts; k++) {
+        // clamp to n_batch tokens
+        auto & inp = inputs[k];
+        const uint64_t n_toks = inp.size();
+
+        // encode if at capacity
+        if (batch.n_tokens + n_toks > n_batch) {
+            float * out = emb + p * n_embd;
+            batch_decode(ctx, batch, out, s, n_embd);
+            llama_batch_clear(batch);
+            p += s;
+            s = 0;
+        }
+
+        // add to batch
+        batch_add_seq(batch, inp, s);
+        s += 1;
     }
 
-    for (int i = 0; i < n_embd; i++) {
-        printf("%f ", embeddings[i]);
-    }
-    printf("\n");
+    // final batch
+    float * out = emb + p * n_embd;
+    batch_decode(ctx, batch, out, s, n_embd);
 
+    // print first 3 embeddings
+    for (int j = 0; j < std::min(3, n_prompts); j++) {
+        fprintf(stderr, "embedding %d: ", j);
+        for (int i = 0; i < n_embd; i++) {
+            fprintf(stderr, "%f ", emb[j * n_embd + i]);
+        }
+        fprintf(stderr, "\n\n");
+    }
+    fprintf(stderr, "\n");
+
+    // clean up
     llama_print_timings(ctx);
     llama_free(ctx);
     llama_free_model(model);
-
     llama_backend_free();
 
     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/README.md

@@ -80,9 +80,9 @@ The LORA rank can be configured for each model tensor type separately with these
   --rank-wk N                LORA rank for wk tensor (default 4)
   --rank-wv N                LORA rank for wv tensor (default 4)
   --rank-wo N                LORA rank for wo tensor (default 4)
-  --rank-w1 N                LORA rank for w1 tensor (default 4)
-  --rank-w2 N                LORA rank for w2 tensor (default 4)
-  --rank-w3 N                LORA rank for w3 tensor (default 4)
+  --rank-ffn_gate N          LORA rank for ffn_gate tensor (default 4)
+  --rank-ffn_down N          LORA rank for ffn_down tensor (default 4)
+  --rank-ffn_up N            LORA rank for ffn_up tensor (default 4)
 ```
 
 The LORA rank of 'norm' tensors should always be 1.

examples/finetune/finetune.cpp

@@ -60,9 +60,9 @@ struct my_llama_layer {
     struct ggml_tensor * ffn_norm;
 
     // ff
-    struct ggml_tensor * w1;
-    struct ggml_tensor * w2;
-    struct ggml_tensor * w3;
+    struct ggml_tensor * ffn_gate; // w1
+    struct ggml_tensor * ffn_down; // w2
+    struct ggml_tensor * ffn_up;   // w3
 };
 
 struct my_llama_model {
@@ -85,9 +85,9 @@ struct my_llama_lora_hparams {
     uint32_t n_rank_wv = 4;
     uint32_t n_rank_wo = 4;
     uint32_t n_rank_ffn_norm = 1;
-    uint32_t n_rank_w1 = 4;
-    uint32_t n_rank_w2 = 4;
-    uint32_t n_rank_w3 = 4;
+    uint32_t n_rank_ffn_gate = 4;
+    uint32_t n_rank_ffn_down = 4;
+    uint32_t n_rank_ffn_up = 4;
     uint32_t n_rank_tok_embeddings = 4;
     uint32_t n_rank_norm = 1;
     uint32_t n_rank_output = 4;
@@ -117,12 +117,12 @@ struct my_llama_lora_layer {
     struct ggml_tensor * ffn_norm_b;
 
     // ff
-    struct ggml_tensor * w1_a;
-    struct ggml_tensor * w1_b;
-    struct ggml_tensor * w2_a;
-    struct ggml_tensor * w2_b;
-    struct ggml_tensor * w3_a;
-    struct ggml_tensor * w3_b;
+    struct ggml_tensor * ffn_gate_a;
+    struct ggml_tensor * ffn_gate_b;
+    struct ggml_tensor * ffn_down_a;
+    struct ggml_tensor * ffn_down_b;
+    struct ggml_tensor * ffn_up_a;
+    struct ggml_tensor * ffn_up_b;
 };
 
 struct my_llama_lora {
@@ -208,9 +208,9 @@ static void print_lora_params(struct my_llama_lora_hparams * params) {
     printf("%s: n_rank_wv             : %u\n", __func__, params->n_rank_wv);
     printf("%s: n_rank_wo             : %u\n", __func__, params->n_rank_wo);
     printf("%s: n_rank_ffn_norm       : %u\n", __func__, params->n_rank_ffn_norm);
-    printf("%s: n_rank_w1             : %u\n", __func__, params->n_rank_w1);
-    printf("%s: n_rank_w2             : %u\n", __func__, params->n_rank_w2);
-    printf("%s: n_rank_w3             : %u\n", __func__, params->n_rank_w3);
+    printf("%s: n_rank_ffn_gate       : %u\n", __func__, params->n_rank_ffn_gate);
+    printf("%s: n_rank_ffn_down       : %u\n", __func__, params->n_rank_ffn_down);
+    printf("%s: n_rank_ffn_up         : %u\n", __func__, params->n_rank_ffn_up);
     printf("%s: n_rank_tok_embeddings : %u\n", __func__, params->n_rank_tok_embeddings);
     printf("%s: n_rank_norm           : %u\n", __func__, params->n_rank_norm);
     printf("%s: n_rank_output         : %u\n", __func__, params->n_rank_output);
@@ -319,9 +319,9 @@ static void init_model(struct llama_model * input, struct my_llama_model * model
         layer.wv             = llama_get_model_tensor(input, tni(LLM_TENSOR_ATTN_V, i));
         layer.wo             = llama_get_model_tensor(input, tni(LLM_TENSOR_ATTN_OUT, i));
         layer.ffn_norm       = llama_get_model_tensor(input, tni(LLM_TENSOR_FFN_NORM, i));
-        layer.w1             = llama_get_model_tensor(input, tni(LLM_TENSOR_FFN_GATE, i));
-        layer.w2             = llama_get_model_tensor(input, tni(LLM_TENSOR_FFN_DOWN, i));
-        layer.w3             = llama_get_model_tensor(input, tni(LLM_TENSOR_FFN_UP, i));
+        layer.ffn_gate       = llama_get_model_tensor(input, tni(LLM_TENSOR_FFN_GATE, i));
+        layer.ffn_down       = llama_get_model_tensor(input, tni(LLM_TENSOR_FFN_DOWN, i));
+        layer.ffn_up         = llama_get_model_tensor(input, tni(LLM_TENSOR_FFN_UP, i));
 
         assert_shape_1d(layer.attention_norm, hparams.n_embd);
         assert_shape_2d(layer.wq,             hparams.n_embd, hparams.n_embd);
@@ -329,9 +329,9 @@ static void init_model(struct llama_model * input, struct my_llama_model * model
         assert_shape_2d(layer.wv,             hparams.n_embd, hparams.n_embd_gqa());
         assert_shape_2d(layer.wo,             hparams.n_embd, hparams.n_embd);
         assert_shape_1d(layer.ffn_norm,       hparams.n_embd);
-        assert_shape_2d(layer.w1,             hparams.n_embd, hparams.n_ff);
-        assert_shape_2d(layer.w2,             hparams.n_ff,   hparams.n_embd);
-        assert_shape_2d(layer.w3,             hparams.n_embd, hparams.n_ff);
+        assert_shape_2d(layer.ffn_gate,       hparams.n_embd, hparams.n_ff);
+        assert_shape_2d(layer.ffn_down,       hparams.n_ff,   hparams.n_embd);
+        assert_shape_2d(layer.ffn_up,         hparams.n_embd, hparams.n_ff);
     }
 }
 
@@ -362,12 +362,12 @@ static void set_param_lora(struct my_llama_lora * lora) {
         ggml_set_param(ctx, layer.wo_b);
         ggml_set_param(ctx, layer.ffn_norm_a);
         ggml_set_param(ctx, layer.ffn_norm_b);
-        ggml_set_param(ctx, layer.w1_a);
-        ggml_set_param(ctx, layer.w1_b);
-        ggml_set_param(ctx, layer.w2_a);
-        ggml_set_param(ctx, layer.w2_b);
-        ggml_set_param(ctx, layer.w3_a);
-        ggml_set_param(ctx, layer.w3_b);
+        ggml_set_param(ctx, layer.ffn_gate_a);
+        ggml_set_param(ctx, layer.ffn_gate_b);
+        ggml_set_param(ctx, layer.ffn_down_a);
+        ggml_set_param(ctx, layer.ffn_down_b);
+        ggml_set_param(ctx, layer.ffn_up_a);
+        ggml_set_param(ctx, layer.ffn_up_b);
     }
 }
 
@@ -435,12 +435,12 @@ static void init_lora(const struct my_llama_model * model, struct my_llama_lora
         layer.ffn_norm_a = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_ffn_norm, n_embd);
         layer.ffn_norm_b = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_ffn_norm, 1);
 
-        layer.w1_a = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_w1, n_embd);
-        layer.w1_b = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_w1, n_ff);
-        layer.w2_a = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_w2, n_ff);
-        layer.w2_b = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_w2, n_embd);
-        layer.w3_a = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_w3, n_embd);
-        layer.w3_b = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_w3, n_ff);
+        layer.ffn_gate_a = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_ffn_gate, n_embd);
+        layer.ffn_gate_b = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_ffn_gate, n_ff);
+        layer.ffn_down_a = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_ffn_down, n_ff);
+        layer.ffn_down_b = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_ffn_down, n_embd);
+        layer.ffn_up_a   = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_ffn_up,   n_embd);
+        layer.ffn_up_b   = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_ffn_up,   n_ff);
 
         ggml_set_name(layer.attention_norm_a, tni(LLM_TENSOR_ATTN_NORM, ".weight.lora_a", i));
         ggml_set_name(layer.attention_norm_b, tni(LLM_TENSOR_ATTN_NORM, ".weight.lora_b", i));
@@ -454,12 +454,12 @@ static void init_lora(const struct my_llama_model * model, struct my_llama_lora
         ggml_set_name(layer.wo_b,             tni(LLM_TENSOR_ATTN_OUT,  ".weight.lora_b", i));
         ggml_set_name(layer.ffn_norm_a,       tni(LLM_TENSOR_FFN_NORM,  ".weight.lora_a", i));
         ggml_set_name(layer.ffn_norm_b,       tni(LLM_TENSOR_FFN_NORM,  ".weight.lora_b", i));
-        ggml_set_name(layer.w1_a,             tni(LLM_TENSOR_FFN_GATE,  ".weight.lora_a", i));
-        ggml_set_name(layer.w1_b,             tni(LLM_TENSOR_FFN_GATE,  ".weight.lora_b", i));
-        ggml_set_name(layer.w2_a,             tni(LLM_TENSOR_FFN_DOWN,  ".weight.lora_a", i));
-        ggml_set_name(layer.w2_b,             tni(LLM_TENSOR_FFN_DOWN,  ".weight.lora_b", i));
-        ggml_set_name(layer.w3_a,             tni(LLM_TENSOR_FFN_UP,    ".weight.lora_a", i));
-        ggml_set_name(layer.w3_b,             tni(LLM_TENSOR_FFN_UP,    ".weight.lora_b", i));
+        ggml_set_name(layer.ffn_gate_a,       tni(LLM_TENSOR_FFN_GATE,  ".weight.lora_a", i));
+        ggml_set_name(layer.ffn_gate_b,       tni(LLM_TENSOR_FFN_GATE,  ".weight.lora_b", i));
+        ggml_set_name(layer.ffn_down_a,       tni(LLM_TENSOR_FFN_DOWN,  ".weight.lora_a", i));
+        ggml_set_name(layer.ffn_down_b,       tni(LLM_TENSOR_FFN_DOWN,  ".weight.lora_b", i));
+        ggml_set_name(layer.ffn_up_a,         tni(LLM_TENSOR_FFN_UP,    ".weight.lora_a", i));
+        ggml_set_name(layer.ffn_up_b,         tni(LLM_TENSOR_FFN_UP,    ".weight.lora_b", i));
     }
 
     set_param_lora(lora);
@@ -497,12 +497,12 @@ static void randomize_lora(struct my_llama_lora * lora, int seed, float mean, fl
         randomize_tensor_normal(layer.ffn_norm_a, rnd);
         ggml_set_zero(layer.ffn_norm_b);
 
-        randomize_tensor_normal(layer.w1_a, rnd);
-        ggml_set_zero(layer.w1_b);
-        randomize_tensor_normal(layer.w2_a, rnd);
-        ggml_set_zero(layer.w2_b);
-        randomize_tensor_normal(layer.w3_a, rnd);
-        ggml_set_zero(layer.w3_b);
+        randomize_tensor_normal(layer.ffn_gate_a, rnd);
+        ggml_set_zero(layer.ffn_gate_b);
+        randomize_tensor_normal(layer.ffn_down_a, rnd);
+        ggml_set_zero(layer.ffn_down_b);
+        randomize_tensor_normal(layer.ffn_up_a, rnd);
+        ggml_set_zero(layer.ffn_up_b);
     }
 
     free_random_normal_distribution(rnd);
@@ -610,13 +610,13 @@ static struct ggml_tensor * llama_build_lora_finetune_graphs(
 
         struct ggml_tensor * attention_norm = add_to_f32(ctx, layer.attention_norm, ggml_mul_mat(ctx, llayer.attention_norm_a, llayer.attention_norm_b));
         struct ggml_tensor * ffn_norm = add_to_f32(ctx, layer.ffn_norm, ggml_mul_mat(ctx, llayer.ffn_norm_a, llayer.ffn_norm_b));
-        struct ggml_tensor * wq = add_to_f32(ctx, layer.wq, ggml_mul_mat(ctx, llayer.wq_a, llayer.wq_b));
-        struct ggml_tensor * wk = add_to_f32(ctx, layer.wk, ggml_mul_mat(ctx, llayer.wk_a, llayer.wk_b));
-        struct ggml_tensor * wv = add_to_f32(ctx, layer.wv, ggml_mul_mat(ctx, llayer.wv_a, llayer.wv_b));
-        struct ggml_tensor * wo = add_to_f32(ctx, layer.wo, ggml_mul_mat(ctx, llayer.wo_a, llayer.wo_b));
-        struct ggml_tensor * w1 = add_to_f32(ctx, layer.w1, ggml_mul_mat(ctx, llayer.w1_a, llayer.w1_b));
-        struct ggml_tensor * w2 = add_to_f32(ctx, layer.w2, ggml_mul_mat(ctx, llayer.w2_a, llayer.w2_b));
-        struct ggml_tensor * w3 = add_to_f32(ctx, layer.w3, ggml_mul_mat(ctx, llayer.w3_a, llayer.w3_b));
+        struct ggml_tensor * wq       = add_to_f32(ctx, layer.wq, ggml_mul_mat(ctx, llayer.wq_a, llayer.wq_b));
+        struct ggml_tensor * wk       = add_to_f32(ctx, layer.wk, ggml_mul_mat(ctx, llayer.wk_a, llayer.wk_b));
+        struct ggml_tensor * wv       = add_to_f32(ctx, layer.wv, ggml_mul_mat(ctx, llayer.wv_a, llayer.wv_b));
+        struct ggml_tensor * wo       = add_to_f32(ctx, layer.wo, ggml_mul_mat(ctx, llayer.wo_a, llayer.wo_b));
+        struct ggml_tensor * ffn_gate = add_to_f32(ctx, layer.ffn_gate, ggml_mul_mat(ctx, llayer.ffn_gate_a, llayer.ffn_gate_b));
+        struct ggml_tensor * ffn_down = add_to_f32(ctx, layer.ffn_down, ggml_mul_mat(ctx, llayer.ffn_down_a, llayer.ffn_down_b));
+        struct ggml_tensor * ffn_up   = add_to_f32(ctx, layer.ffn_up, ggml_mul_mat(ctx, llayer.ffn_up_a, llayer.ffn_up_b));
 
         struct ggml_tensor * t02 = ggml_rms_norm     (ctx, cur, rms_norm_eps);                       set_name(t02, "t02");     assert_shape_2d(t02, n_embd, N*n_batch);
         struct ggml_tensor * t03 = ggml_repeat       (ctx, attention_norm, t02);                     set_name(t03, "t03");     assert_shape_2d(t03, n_embd, N*n_batch);
@@ -659,11 +659,11 @@ static struct ggml_tensor * llama_build_lora_finetune_graphs(
         struct ggml_tensor * t22 = ggml_rms_norm     (ctx, t21, rms_norm_eps);                       set_name(t22, "t22");     assert_shape_2d(t22, n_embd, N*n_batch);
         struct ggml_tensor * t23 = ggml_repeat       (ctx, ffn_norm, t22);                           set_name(t23, "t23");     assert_shape_2d(t23, n_embd, N*n_batch);
         struct ggml_tensor * t24 = ggml_mul          (ctx, t23, t22);                                set_name(t24, "t24");     assert_shape_2d(t24, n_embd, N*n_batch);
-        struct ggml_tensor * t25 = ggml_mul_mat      (ctx, w3, t24);                                 set_name(t25, "t25");     assert_shape_2d(t25, n_ff, N*n_batch);
-        struct ggml_tensor * t26 = ggml_mul_mat      (ctx, w1, t24);                                 set_name(t26, "t26");     assert_shape_2d(t26, n_ff, N*n_batch);
+        struct ggml_tensor * t25 = ggml_mul_mat      (ctx, ffn_up, t24);                             set_name(t25, "t25");     assert_shape_2d(t25, n_ff, N*n_batch);
+        struct ggml_tensor * t26 = ggml_mul_mat      (ctx, ffn_gate, t24);                           set_name(t26, "t26");     assert_shape_2d(t26, n_ff, N*n_batch);
         struct ggml_tensor * t27 = ggml_silu         (ctx, t26);                                     set_name(t27, "t27");     assert_shape_2d(t27, n_ff, N*n_batch);
         struct ggml_tensor * t28 = ggml_mul          (ctx, t27, t25);                                set_name(t28, "t28");     assert_shape_2d(t28, n_ff, N*n_batch);
-        struct ggml_tensor * t29 = ggml_mul_mat      (ctx, w2, t28);                                 set_name(t29, "t29");     assert_shape_2d(t29, n_embd, N*n_batch);
+        struct ggml_tensor * t29 = ggml_mul_mat      (ctx, ffn_down, t28);                           set_name(t29, "t29");     assert_shape_2d(t29, n_embd, N*n_batch);
         struct ggml_tensor * t30 = ggml_add          (ctx, t29, t21);                                set_name(t30, "t30");     assert_shape_2d(t30, n_embd, N*n_batch);
         cur = t30;
         if (enable_checkpointing) {
@@ -723,9 +723,9 @@ static struct ggml_tensor * llama_build_lora_finetune_graphs(
         ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.wk, 1.0f));
         ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.wv, 1.0f));
         ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.wo, 1.0f));
-        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.w1, 1.0f));
-        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.w2, 1.0f));
-        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.w3, 1.0f));
+        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.ffn_gate, 1.0f));
+        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.ffn_down, 1.0f));
+        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.ffn_up, 1.0f));
     }
 
     // allocating checkpoints in one block to reduce memory fragmentation
@@ -798,9 +798,9 @@ static void load_llama_lora_gguf(struct gguf_context * fctx, struct ggml_context
     GGUF_GET_KEY(fctx, lora->hparams.n_rank_wv,             gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_LORA_RANK_ATTN_V);
     GGUF_GET_KEY(fctx, lora->hparams.n_rank_wo,             gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_LORA_RANK_ATTN_OUT);
     GGUF_GET_KEY(fctx, lora->hparams.n_rank_ffn_norm,       gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_LORA_RANK_FFN_NORM);
-    GGUF_GET_KEY(fctx, lora->hparams.n_rank_w1,             gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_LORA_RANK_FFN_GATE);
-    GGUF_GET_KEY(fctx, lora->hparams.n_rank_w2,             gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_LORA_RANK_FFN_DOWN);
-    GGUF_GET_KEY(fctx, lora->hparams.n_rank_w3,             gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_LORA_RANK_FFN_UP);
+    GGUF_GET_KEY(fctx, lora->hparams.n_rank_ffn_gate,       gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_LORA_RANK_FFN_GATE);
+    GGUF_GET_KEY(fctx, lora->hparams.n_rank_ffn_down,       gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_LORA_RANK_FFN_DOWN);
+    GGUF_GET_KEY(fctx, lora->hparams.n_rank_ffn_up,         gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_LORA_RANK_FFN_UP);
 
     init_lora(model, lora);
 
@@ -825,12 +825,12 @@ static void load_llama_lora_gguf(struct gguf_context * fctx, struct ggml_context
         copy_tensor_by_name(layer.wo_b,             f_ggml_ctx, ggml_get_name(layer.wo_b));
         copy_tensor_by_name(layer.ffn_norm_a,       f_ggml_ctx, ggml_get_name(layer.ffn_norm_a));
         copy_tensor_by_name(layer.ffn_norm_b,       f_ggml_ctx, ggml_get_name(layer.ffn_norm_b));
-        copy_tensor_by_name(layer.w1_a,             f_ggml_ctx, ggml_get_name(layer.w1_a));
-        copy_tensor_by_name(layer.w1_b,             f_ggml_ctx, ggml_get_name(layer.w1_b));
-        copy_tensor_by_name(layer.w2_a,             f_ggml_ctx, ggml_get_name(layer.w2_a));
-        copy_tensor_by_name(layer.w2_b,             f_ggml_ctx, ggml_get_name(layer.w2_b));
-        copy_tensor_by_name(layer.w3_a,             f_ggml_ctx, ggml_get_name(layer.w3_a));
-        copy_tensor_by_name(layer.w3_b,             f_ggml_ctx, ggml_get_name(layer.w3_b));
+        copy_tensor_by_name(layer.ffn_gate_a,       f_ggml_ctx, ggml_get_name(layer.ffn_gate_a));
+        copy_tensor_by_name(layer.ffn_gate_b,       f_ggml_ctx, ggml_get_name(layer.ffn_gate_b));
+        copy_tensor_by_name(layer.ffn_down_a,       f_ggml_ctx, ggml_get_name(layer.ffn_down_a));
+        copy_tensor_by_name(layer.ffn_down_b,       f_ggml_ctx, ggml_get_name(layer.ffn_down_b));
+        copy_tensor_by_name(layer.ffn_up_a,         f_ggml_ctx, ggml_get_name(layer.ffn_up_a));
+        copy_tensor_by_name(layer.ffn_up_b,         f_ggml_ctx, ggml_get_name(layer.ffn_up_b));
     }
 }
 
@@ -868,9 +868,9 @@ static void save_llama_lora_gguf(struct gguf_context * fctx, struct my_llama_mod
     gguf_set_val_u32(fctx, LLM_KV_TRAINING_LORA_RANK_ATTN_V,       lora->hparams.n_rank_wv);
     gguf_set_val_u32(fctx, LLM_KV_TRAINING_LORA_RANK_ATTN_OUT,     lora->hparams.n_rank_wo);
     gguf_set_val_u32(fctx, LLM_KV_TRAINING_LORA_RANK_FFN_NORM,     lora->hparams.n_rank_ffn_norm);
-    gguf_set_val_u32(fctx, LLM_KV_TRAINING_LORA_RANK_FFN_GATE,     lora->hparams.n_rank_w1);
-    gguf_set_val_u32(fctx, LLM_KV_TRAINING_LORA_RANK_FFN_DOWN,     lora->hparams.n_rank_w2);
-    gguf_set_val_u32(fctx, LLM_KV_TRAINING_LORA_RANK_FFN_UP,       lora->hparams.n_rank_w3);
+    gguf_set_val_u32(fctx, LLM_KV_TRAINING_LORA_RANK_FFN_GATE,     lora->hparams.n_rank_ffn_gate);
+    gguf_set_val_u32(fctx, LLM_KV_TRAINING_LORA_RANK_FFN_DOWN,     lora->hparams.n_rank_ffn_down);
+    gguf_set_val_u32(fctx, LLM_KV_TRAINING_LORA_RANK_FFN_UP,       lora->hparams.n_rank_ffn_up);
 
     gguf_add_tensor(fctx, lora->tok_embeddings_a);
     gguf_add_tensor(fctx, lora->tok_embeddings_b);
@@ -894,12 +894,12 @@ static void save_llama_lora_gguf(struct gguf_context * fctx, struct my_llama_mod
         gguf_add_tensor(fctx, layer.wo_b);
         gguf_add_tensor(fctx, layer.ffn_norm_a);
         gguf_add_tensor(fctx, layer.ffn_norm_b);
-        gguf_add_tensor(fctx, layer.w1_a);
-        gguf_add_tensor(fctx, layer.w1_b);
-        gguf_add_tensor(fctx, layer.w2_a);
-        gguf_add_tensor(fctx, layer.w2_b);
-        gguf_add_tensor(fctx, layer.w3_a);
-        gguf_add_tensor(fctx, layer.w3_b);
+        gguf_add_tensor(fctx, layer.ffn_gate_a);
+        gguf_add_tensor(fctx, layer.ffn_gate_b);
+        gguf_add_tensor(fctx, layer.ffn_down_a);
+        gguf_add_tensor(fctx, layer.ffn_down_b);
+        gguf_add_tensor(fctx, layer.ffn_up_a);
+        gguf_add_tensor(fctx, layer.ffn_up_b);
     }
 }
 
@@ -1104,12 +1104,12 @@ static void save_as_llama_lora(const char * filename, struct my_llama_lora * lor
         write_tensor(&file, layer.wo_b,             tni(LLM_TENSOR_ATTN_OUT,  i, ".weight.loraB"));
         write_tensor(&file, layer.ffn_norm_a,       tni(LLM_TENSOR_FFN_NORM,  i, ".weight.loraA"));
         write_tensor(&file, layer.ffn_norm_b,       tni(LLM_TENSOR_FFN_NORM,  i, ".weight.loraB"));
-        write_tensor(&file, layer.w1_a,             tni(LLM_TENSOR_FFN_GATE,  i, ".weight.loraA"));
-        write_tensor(&file, layer.w1_b,             tni(LLM_TENSOR_FFN_GATE,  i, ".weight.loraB"));
-        write_tensor(&file, layer.w2_a,             tni(LLM_TENSOR_FFN_DOWN,  i, ".weight.loraA"));
-        write_tensor(&file, layer.w2_b,             tni(LLM_TENSOR_FFN_DOWN,  i, ".weight.loraB"));
-        write_tensor(&file, layer.w3_a,             tni(LLM_TENSOR_FFN_UP,    i, ".weight.loraA"));
-        write_tensor(&file, layer.w3_b,             tni(LLM_TENSOR_FFN_UP,    i, ".weight.loraB"));
+        write_tensor(&file, layer.ffn_gate_a,       tni(LLM_TENSOR_FFN_GATE,  i, ".weight.loraA"));
+        write_tensor(&file, layer.ffn_gate_b,       tni(LLM_TENSOR_FFN_GATE,  i, ".weight.loraB"));
+        write_tensor(&file, layer.ffn_down_a,       tni(LLM_TENSOR_FFN_DOWN,  i, ".weight.loraA"));
+        write_tensor(&file, layer.ffn_down_b,       tni(LLM_TENSOR_FFN_DOWN,  i, ".weight.loraB"));
+        write_tensor(&file, layer.ffn_up_a,         tni(LLM_TENSOR_FFN_UP,    i, ".weight.loraA"));
+        write_tensor(&file, layer.ffn_up_b,         tni(LLM_TENSOR_FFN_UP,    i, ".weight.loraB"));
     }
 }
 
@@ -1139,9 +1139,9 @@ struct train_params {
     uint32_t n_rank_wv;
     uint32_t n_rank_wo;
     uint32_t n_rank_ffn_norm;
-    uint32_t n_rank_w1;
-    uint32_t n_rank_w2;
-    uint32_t n_rank_w3;
+    uint32_t n_rank_ffn_gate;
+    uint32_t n_rank_ffn_down;
+    uint32_t n_rank_ffn_up;
     uint32_t n_rank_tok_embeddings;
     uint32_t n_rank_norm;
     uint32_t n_rank_output;
@@ -1152,9 +1152,9 @@ struct train_params {
     bool custom_n_rank_wv;
     bool custom_n_rank_wo;
     bool custom_n_rank_ffn_norm;
-    bool custom_n_rank_w1;
-    bool custom_n_rank_w2;
-    bool custom_n_rank_w3;
+    bool custom_n_rank_ffn_gate;
+    bool custom_n_rank_ffn_down;
+    bool custom_n_rank_ffn_up;
     bool custom_n_rank_tok_embeddings;
     bool custom_n_rank_norm;
     bool custom_n_rank_output;
@@ -1186,9 +1186,9 @@ static struct train_params get_default_train_params() {
     params.n_rank_wv             = 4;
     params.n_rank_wo             = 4;
     params.n_rank_ffn_norm       = 1;
-    params.n_rank_w1             = 4;
-    params.n_rank_w2             = 4;
-    params.n_rank_w3             = 4;
+    params.n_rank_ffn_gate       = 4;
+    params.n_rank_ffn_down       = 4;
+    params.n_rank_ffn_up         = 4;
     params.n_rank_tok_embeddings = 4;
     params.n_rank_norm           = 1;
     params.n_rank_output         = 4;
@@ -1199,9 +1199,9 @@ static struct train_params get_default_train_params() {
     params.custom_n_rank_wv             = false;
     params.custom_n_rank_wo             = false;
     params.custom_n_rank_ffn_norm       = false;
-    params.custom_n_rank_w1             = false;
-    params.custom_n_rank_w2             = false;
-    params.custom_n_rank_w3             = false;
+    params.custom_n_rank_ffn_gate       = false;
+    params.custom_n_rank_ffn_down       = false;
+    params.custom_n_rank_ffn_up         = false;
     params.custom_n_rank_tok_embeddings = false;
     params.custom_n_rank_norm           = false;
     params.custom_n_rank_output         = false;
@@ -1232,9 +1232,9 @@ static void train_print_usage(int argc, char ** argv, const struct train_params
     fprintf(stderr, "  --rank-wk N                LORA rank for wk tensor, overrides default rank.\n");
     fprintf(stderr, "  --rank-wv N                LORA rank for wv tensor, overrides default rank.\n");
     fprintf(stderr, "  --rank-wo N                LORA rank for wo tensor, overrides default rank.\n");
-    fprintf(stderr, "  --rank-w1 N                LORA rank for w1 tensor, overrides default rank.\n");
-    fprintf(stderr, "  --rank-w2 N                LORA rank for w2 tensor, overrides default rank.\n");
-    fprintf(stderr, "  --rank-w3 N                LORA rank for w3 tensor, overrides default rank.\n");
+    fprintf(stderr, "  --rank-ffn_gate N          LORA rank for ffn_gate tensor, overrides default rank.\n");
+    fprintf(stderr, "  --rank-ffn_down N          LORA rank for ffn_down tensor, overrides default rank.\n");
+    fprintf(stderr, "  --rank-ffn_up N            LORA rank for ffn_up tensor, overrides default rank.\n");
 
     print_common_train_usage(argc, argv, &params->common);
 }
@@ -1369,27 +1369,27 @@ static bool train_params_parse(int argc, char ** argv, struct train_params * par
             }
             params->n_rank_wo = std::stoi(argv[i]);
             params->custom_n_rank_wo = true;
-        } else if (arg == "--rank-w1") {
+        } else if (arg == "--rank-ffn_gate") {
             if (++i >= argc) {
                 invalid_param = true;
                 break;
             }
-            params->n_rank_w1 = std::stoi(argv[i]);
-            params->custom_n_rank_w1 = true;
-        } else if (arg == "--rank-w2") {
+            params->n_rank_ffn_gate = std::stoi(argv[i]);
+            params->custom_n_rank_ffn_gate = true;
+        } else if (arg == "--rank-ffn_down") {
             if (++i >= argc) {
                 invalid_param = true;
                 break;
             }
-            params->n_rank_w2 = std::stoi(argv[i]);
-            params->custom_n_rank_w2 = true;
-        } else if (arg == "--rank-w3") {
+            params->n_rank_ffn_down = std::stoi(argv[i]);
+            params->custom_n_rank_ffn_down = true;
+        } else if (arg == "--rank-ffn_up") {
             if (++i >= argc) {
                 invalid_param = true;
                 break;
             }
-            params->n_rank_w3 = std::stoi(argv[i]);
-            params->custom_n_rank_w3 = true;
+            params->n_rank_ffn_up = std::stoi(argv[i]);
+            params->custom_n_rank_ffn_up = true;
         } else {
             fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
             train_print_usage(argc, argv, &default_params);
@@ -1452,12 +1452,12 @@ static int64_t get_parameter_count(struct my_llama_lora* lora) {
         nx += ggml_nelements(layer.wo_b);
         nx += ggml_nelements(layer.ffn_norm_a);
         nx += ggml_nelements(layer.ffn_norm_b);
-        nx += ggml_nelements(layer.w1_a);
-        nx += ggml_nelements(layer.w1_b);
-        nx += ggml_nelements(layer.w2_a);
-        nx += ggml_nelements(layer.w2_b);
-        nx += ggml_nelements(layer.w3_a);
-        nx += ggml_nelements(layer.w3_b);
+        nx += ggml_nelements(layer.ffn_gate_a);
+        nx += ggml_nelements(layer.ffn_gate_b);
+        nx += ggml_nelements(layer.ffn_down_a);
+        nx += ggml_nelements(layer.ffn_down_b);
+        nx += ggml_nelements(layer.ffn_up_a);
+        nx += ggml_nelements(layer.ffn_up_b);
     }
     return nx;
 }
@@ -1511,9 +1511,9 @@ int main(int argc, char ** argv) {
     uint32_t n_rank_wv                 = params.custom_n_rank_wv             ? params.n_rank_wv             : params.lora_r;
     uint32_t n_rank_wo                 = params.custom_n_rank_wo             ? params.n_rank_wo             : params.lora_r;
     uint32_t n_rank_ffn_norm           = params.custom_n_rank_ffn_norm       ? params.n_rank_ffn_norm       : 1;
-    uint32_t n_rank_w1                 = params.custom_n_rank_w1             ? params.n_rank_w1             : params.lora_r;
-    uint32_t n_rank_w2                 = params.custom_n_rank_w2             ? params.n_rank_w2             : params.lora_r;
-    uint32_t n_rank_w3                 = params.custom_n_rank_w3             ? params.n_rank_w3             : params.lora_r;
+    uint32_t n_rank_ffn_gate           = params.custom_n_rank_ffn_gate       ? params.n_rank_ffn_gate       : params.lora_r;
+    uint32_t n_rank_ffn_down           = params.custom_n_rank_ffn_down       ? params.n_rank_ffn_down       : params.lora_r;
+    uint32_t n_rank_ffn_up             = params.custom_n_rank_ffn_up         ? params.n_rank_ffn_up         : params.lora_r;
     uint32_t n_rank_tok_embeddings     = params.custom_n_rank_tok_embeddings ? params.n_rank_tok_embeddings : params.lora_r;
     uint32_t n_rank_norm               = params.custom_n_rank_norm           ? params.n_rank_norm           : 1;
     uint32_t n_rank_output             = params.custom_n_rank_output         ? params.n_rank_output         : params.lora_r;
@@ -1523,9 +1523,9 @@ int main(int argc, char ** argv) {
     lora.hparams.n_rank_wv             = n_rank_wv;
     lora.hparams.n_rank_wo             = n_rank_wo;
     lora.hparams.n_rank_ffn_norm       = n_rank_ffn_norm;
-    lora.hparams.n_rank_w1             = n_rank_w1;
-    lora.hparams.n_rank_w2             = n_rank_w2;
-    lora.hparams.n_rank_w3             = n_rank_w3;
+    lora.hparams.n_rank_ffn_gate       = n_rank_ffn_gate;
+    lora.hparams.n_rank_ffn_down       = n_rank_ffn_down;
+    lora.hparams.n_rank_ffn_up         = n_rank_ffn_up;
     lora.hparams.n_rank_tok_embeddings = n_rank_tok_embeddings;
     lora.hparams.n_rank_norm           = n_rank_norm;
     lora.hparams.n_rank_output         = n_rank_output;
@@ -1566,9 +1566,9 @@ int main(int argc, char ** argv) {
         || (lora.hparams.n_rank_wv             != n_rank_wv)
         || (lora.hparams.n_rank_wo             != n_rank_wo)
         || (lora.hparams.n_rank_ffn_norm       != n_rank_ffn_norm)
-        || (lora.hparams.n_rank_w1             != n_rank_w1)
-        || (lora.hparams.n_rank_w2             != n_rank_w2)
-        || (lora.hparams.n_rank_w3             != n_rank_w3)
+        || (lora.hparams.n_rank_ffn_gate       != n_rank_ffn_gate)
+        || (lora.hparams.n_rank_ffn_down       != n_rank_ffn_down)
+        || (lora.hparams.n_rank_ffn_up         != n_rank_ffn_up)
         || (lora.hparams.n_rank_tok_embeddings != n_rank_tok_embeddings)
         || (lora.hparams.n_rank_norm           != n_rank_norm)
         || (lora.hparams.n_rank_output         != n_rank_output)

examples/imatrix/imatrix.cpp

@@ -568,7 +568,8 @@ int main(int argc, char ** argv) {
         params.prompt = gpt_random_prompt(rng);
     }
 
-    llama_backend_init(params.numa);
+    llama_backend_init();
+    llama_numa_init(params.numa);
 
     llama_model_params mparams = llama_model_params_from_gpt_params(params);
 

examples/infill/infill.cpp

@@ -202,7 +202,8 @@ int main(int argc, char ** argv) {
     std::mt19937 rng(params.seed);
 
     LOG("%s: llama backend init\n", __func__);
-    llama_backend_init(params.numa);
+    llama_backend_init();
+    llama_numa_init(params.numa);
 
     llama_model * model;
     llama_context * ctx;

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

@@ -1151,8 +1151,7 @@ int main(int argc, char ** argv) {
     if (!params.verbose) {
         llama_log_set(llama_null_log_callback, NULL);
     }
-    bool numa = false;
-    llama_backend_init(numa);
+    llama_backend_init();
 
     // initialize printer
     std::unique_ptr<printer> p;

examples/llama.android/app/src/main/cpp/llama-android.cpp

@@ -274,8 +274,8 @@ Java_com_example_llama_Llm_new_1batch(JNIEnv *, jobject, jint n_tokens, jint emb
 
 extern "C"
 JNIEXPORT void JNICALL
-Java_com_example_llama_Llm_backend_1init(JNIEnv *, jobject, jboolean numa) {
-    llama_backend_init(numa);
+Java_com_example_llama_Llm_backend_1init(JNIEnv *, jobject) {
+    llama_backend_init();
 }
 
 extern "C"

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

@@ -51,7 +51,7 @@ actor LlamaContext {
     }
 
     static func create_context(path: String) throws -> LlamaContext {
-        llama_backend_init(false)
+        llama_backend_init()
         var model_params = llama_model_default_params()
 
 #if targetEnvironment(simulator)

examples/llava/README.md

@@ -1,10 +1,12 @@
 # LLaVA
 
-Currently this implementation supports [llava-v1.5](https://huggingface.co/liuhaotian/llava-v1.5-7b) variants.
+Currently this implementation supports [llava-v1.5](https://huggingface.co/liuhaotian/llava-v1.5-7b) variants,
+as well as llava-1.6 [llava-v1.6](https://huggingface.co/collections/liuhaotian/llava-16-65b9e40155f60fd046a5ccf2) variants.
 
 The pre-converted [7b](https://huggingface.co/mys/ggml_llava-v1.5-7b)
 and [13b](https://huggingface.co/mys/ggml_llava-v1.5-13b)
 models are available.
+For llava-1.6 a variety of prepared gguf models are available as well [7b-34b](https://huggingface.co/cmp-nct/llava-1.6-gguf)
 
 After API is confirmed, more models will be supported / uploaded.
 
@@ -18,10 +20,11 @@ After building, run: `./llava-cli` to see the usage. For example:
 ```
 
 **note**: A lower temperature like 0.1 is recommended for better quality. add `--temp 0.1` to the command to do so.
+**note**: For GPU offloading ensure to use the `-ngl` flag just like usual
 
-## Model conversion
+## LLaVA 1.5
 
-- Clone `llava-v15-7b` and `clip-vit-large-patch14-336` locally:
+- Clone a LLaVA and a CLIP model ([available options](https://github.com/haotian-liu/LLaVA/blob/main/docs/MODEL_ZOO.md)). For example:
 
 ```sh
 git clone https://huggingface.co/liuhaotian/llava-v1.5-7b
@@ -50,13 +53,79 @@ 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) 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:
+```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) 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)
+
+## llava-cli templating and llava-1.6 prompting
+
+llava-1.5 models all use the same vicuna prompt, here you can just add your image question like `-p "Provide a full description."`
+For llava-1.5 models which are not vicuna (mistral and Yi) you need to adapt system prompt as well as user prompt, for this purpose llava-cli has a basic templating system:
+
+**For Mistral and using llava-cli binary:**
+Add this: `-p "<image>\nUSER:\nProvide a full description.\nASSISTANT:\n"`
+The mistral template for llava-1.6 seems to be no system print and a USER/ASSISTANT role
+
+**For the 34B this should work:**
+Add this: `-e -p <|im_start|>system\nAnswer the questions.<|im_end|><|im_start|>user\n<image>\nProvide a full description.<|im_end|><|im_start|>assistant\n`
+
+
+## How to know if you are running in llava-1.5 or llava-1.6 mode
+
+When running llava-cli you will see a visual information right before the prompt is being processed:
+
+**Llava-1.5:**
+`encode_image_with_clip: image embedding created: 576 tokens`
+
+**Llava-1.6 (anything above 576):**
+`encode_image_with_clip: image embedding created: 2880 tokens`
+
+
+Alternatively just pay notice to how many "tokens" have been used for your prompt, it will also show 1000+ tokens for llava-1.6
+
+
+
+
 ## TODO
 
-- [ ] Support non-CPU backend for the image encoding part.
+- [x] Support non-CPU backend for the image encoding part.
 - [ ] Support different sampling methods.
 - [ ] Support more model variants.

examples/llava/clip.h

@@ -24,25 +24,7 @@ struct clip_ctx;
 extern "C" {
 #endif
 
-struct clip_vision_hparams {
-    int32_t image_size;
-    int32_t patch_size;
-    int32_t hidden_size;
-    int32_t n_intermediate;
-    int32_t projection_dim;
-    int32_t n_head;
-    int32_t n_layer;
-    float eps;
-};
-
-CLIP_API struct clip_ctx * clip_model_load(const char * fname, int verbosity);
-
-CLIP_API void clip_free(struct clip_ctx * ctx);
-
-CLIP_API size_t clip_embd_nbytes(const struct clip_ctx * ctx);
-
-CLIP_API int clip_n_patches    (const struct clip_ctx * ctx);
-CLIP_API int clip_n_mmproj_embd(const struct clip_ctx * ctx);
+struct clip_ctx;
 
 struct clip_image_u8_batch {
     struct clip_image_u8 * data;
@@ -54,18 +36,43 @@ struct clip_image_f32_batch {
     size_t size;
 };
 
+CLIP_API struct clip_ctx * clip_model_load    (const char * fname, int verbosity);
+CLIP_API struct clip_ctx * clip_model_load_cpu(const char * fname, int verbosity);
+
+CLIP_API void clip_free(struct clip_ctx * ctx);
+
+CLIP_API size_t clip_embd_nbytes(const struct clip_ctx * ctx);
+
+CLIP_API int32_t clip_image_size (const struct clip_ctx * ctx);
+CLIP_API int32_t clip_patch_size (const struct clip_ctx * ctx);
+CLIP_API int32_t clip_hidden_size(const struct clip_ctx * ctx);
+
+// TODO: should be enum, not string
+CLIP_API const char * clip_patch_merge_type(const struct clip_ctx * ctx);
+
+CLIP_API const int32_t * clip_image_grid(const struct clip_ctx * ctx);
+
+CLIP_API int clip_n_patches    (const struct clip_ctx * ctx);
+CLIP_API int clip_n_mmproj_embd(const struct clip_ctx * ctx);
+
 CLIP_API struct clip_image_u8  * clip_image_u8_init ();
 CLIP_API struct clip_image_f32 * clip_image_f32_init();
 
-CLIP_API void clip_image_u8_free (struct clip_image_u8 * img);
+CLIP_API void clip_image_u8_free (struct clip_image_u8  * img);
 CLIP_API void clip_image_f32_free(struct clip_image_f32 * img);
+CLIP_API void clip_image_u8_batch_free (struct clip_image_u8_batch  & batch);
+CLIP_API void clip_image_f32_batch_free(struct clip_image_f32_batch & batch);
 
 CLIP_API bool clip_image_load_from_file(const char * fname, struct clip_image_u8 * img);
 
 /** interpret bytes as an image file with length bytes_length, and use the result to populate img */
 CLIP_API bool clip_image_load_from_bytes(const unsigned char * bytes, size_t bytes_length, struct clip_image_u8 * img);
 
-CLIP_API bool clip_image_preprocess  (struct clip_ctx * ctx, const struct clip_image_u8 * img, struct clip_image_f32 * res, bool pad2square);
+/** preprocess img and store the result in res_imgs, pad_to_square may be overriden to false depending on model configuration */
+CLIP_API bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, clip_image_f32_batch & res_imgs );
+
+CLIP_API struct ggml_tensor * clip_get_newline_tensor(const struct clip_ctx * ctx);
+
 CLIP_API bool clip_image_encode      (struct clip_ctx * ctx, int n_threads, struct clip_image_f32 * img, float * vec);
 CLIP_API bool clip_image_batch_encode(struct clip_ctx * ctx, int n_threads, const struct clip_image_f32_batch * imgs, float * vec);
 

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

@@ -78,18 +78,19 @@ ap.add_argument("--text-only", action="store_true", required=False,
                 help="Save a text-only model. It can't be used to encode images")
 ap.add_argument("--vision-only", action="store_true", required=False,
                 help="Save a vision-only model. It can't be used to encode texts")
-ap.add_argument("--clip_model_is_vision", action="store_true", required=False,
+ap.add_argument("--clip-model-is-vision", action="store_true", required=False,
                 help="The clip model is a pure vision model (ShareGPT4V vision extract for example)")
+ap.add_argument("--clip-model-is-openclip", action="store_true", required=False,
+                help="The clip model is from openclip (for ViT-SO400M type))")
 ap.add_argument("--llava-projector", help="Path to llava.projector file. If specified, save an image encoder for LLaVA models.")
 ap.add_argument("--projector-type", help="Type of projector. Possible values: mlp, ldp", choices=["mlp", "ldp"], default="mlp")
-ap.add_argument("--image-mean", nargs=3, type=float, required=False, help="Override image mean values")
-ap.add_argument("--image-std", nargs=3, type=float, required=False, help="Override image std values")
 ap.add_argument("-o", "--output-dir", help="Directory to save GGUF files. Default is the original model directory", default=None)
 # Example --image_mean 0.48145466 0.4578275 0.40821073 --image_std 0.26862954 0.26130258 0.27577711
+# Example --image_mean 0.5 0.5 0.5 --image_std 0.5 0.5 0.5
 default_image_mean = [0.48145466, 0.4578275, 0.40821073]
 default_image_std = [0.26862954, 0.26130258, 0.27577711]
-ap.add_argument('--image_mean', type=float, nargs='+', help='Mean of the images for normalization (overrides processor) ', default=None)
-ap.add_argument('--image_std', type=float, nargs='+', help='Standard deviation of the images for normalization (overrides processor)', default=None)
+ap.add_argument('--image-mean', type=float, nargs='+', help='Mean of the images for normalization (overrides processor) ', default=None)
+ap.add_argument('--image-std', type=float, nargs='+', help='Standard deviation of the images for normalization (overrides processor)', default=None)
 
 # with proper
 args = ap.parse_args()
@@ -105,7 +106,7 @@ if args.use_f32:
 # output in the same directory as the model if output_dir is None
 dir_model = args.model_dir
 
-if args.clip_model_is_vision:
+if args.clip_model_is_vision or not os.path.exists(dir_model + "/vocab.json") or args.clip_model_is_openclip:
     vocab = None
     tokens = None
 else:
@@ -133,7 +134,7 @@ ftype = 1
 if args.use_f32:
     ftype = 0
 
-if args.clip_model_is_vision:
+if args.clip_model_is_vision or args.clip_model_is_openclip:
     model = CLIPVisionModel.from_pretrained(dir_model)
     processor = None
 else:
@@ -202,6 +203,57 @@ if has_vision_encoder:
     fout.add_float32(k(KEY_ATTENTION_LAYERNORM_EPS, VISION), v_hparams["layer_norm_eps"])
     block_count = v_hparams["num_hidden_layers"] - 1 if has_llava_projector else v_hparams["num_hidden_layers"]
     fout.add_uint32(k(KEY_BLOCK_COUNT, VISION), block_count)
+                            #     /**
+                            #      "image_grid_pinpoints": [
+                            #         [
+                            #         336,
+                            #         672
+                            #         ],
+                            #         [
+                            #         672,
+                            #         336
+                            #         ],
+                            #         [
+                            #         672,
+                            #         672
+                            #         ],
+                            #         [
+                            #         1008,
+                            #         336
+                            #         ],
+                            #         [
+                            #         336,
+                            #         1008
+                            #         ]
+                            #     ],
+                            #     Flattened:
+                            #     [
+                            #         336, 672,
+                            #         672, 336,
+                            #         672, 672,
+                            #         1008, 336,
+                            #         336, 1008
+                            #     ]
+                            #  *
+                            #  */
+    if "image_grid_pinpoints" in v_hparams:
+        # flatten it
+        image_grid_pinpoints = []
+        for pinpoint in v_hparams["image_grid_pinpoints"]:
+            for p in pinpoint:
+                image_grid_pinpoints.append(p)
+        fout.add_array("clip.vision.image_grid_pinpoints", image_grid_pinpoints)
+    if "image_crop_resolution" in v_hparams:
+        fout.add_uint32("clip.vision.image_crop_resolution", v_hparams["image_crop_resolution"])
+    if "image_aspect_ratio" in v_hparams:
+        fout.add_string("clip.vision.image_aspect_ratio", v_hparams["image_aspect_ratio"])
+    if "image_split_resolution" in v_hparams:
+        fout.add_uint32("clip.vision.image_split_resolution", v_hparams["image_split_resolution"])
+    if "mm_patch_merge_type" in v_hparams:
+        fout.add_string("clip.vision.mm_patch_merge_type", v_hparams["mm_patch_merge_type"])
+    if "mm_projector_type" in v_hparams:
+        fout.add_string("clip.vision.mm_projector_type", v_hparams["mm_projector_type"])
+
 
     if processor is not None:
         image_mean = processor.image_processor.image_mean if args.image_mean is None or args.image_mean == default_image_mean else args.image_mean

examples/llava/llava-cli.cpp

@@ -155,11 +155,29 @@ static void process_prompt(struct llava_context * ctx_llava, struct llava_image_
         system_prompt = prompt.substr(0, image_pos);
         user_prompt = prompt.substr(image_pos + std::string("<image>").length());
         printf("system_prompt: %s\n", system_prompt.c_str());
+        if (params->verbose_prompt) {
+            auto tmp = ::llama_tokenize(ctx_llava->ctx_llama, system_prompt, true, true);
+            for (int i = 0; i < (int) tmp.size(); i++) {
+                printf("%6d -> '%s'\n", tmp[i], llama_token_to_piece(ctx_llava->ctx_llama, tmp[i]).c_str());
+            }
+        }
         printf("user_prompt: %s\n", user_prompt.c_str());
+        if (params->verbose_prompt) {
+            auto tmp = ::llama_tokenize(ctx_llava->ctx_llama, user_prompt, true, true);
+            for (int i = 0; i < (int) tmp.size(); i++) {
+                printf("%6d -> '%s'\n", tmp[i], llama_token_to_piece(ctx_llava->ctx_llama, tmp[i]).c_str());
+            }
+        }
     } else {
         // llava-1.5 native mode
         system_prompt = "A chat between a curious human and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the human's questions.\nUSER:";
         user_prompt = prompt + "\nASSISTANT:";
+        if (params->verbose_prompt) {
+            auto tmp = ::llama_tokenize(ctx_llava->ctx_llama, user_prompt, true, true);
+            for (int i = 0; i < (int) tmp.size(); i++) {
+                printf("%6d -> '%s'\n", tmp[i], llama_token_to_piece(ctx_llava->ctx_llama, tmp[i]).c_str());
+            }
+        }
     }
 
     eval_string(ctx_llava->ctx_llama, system_prompt.c_str(), params->n_batch, &n_past, add_bos);
@@ -171,13 +189,17 @@ static void process_prompt(struct llava_context * ctx_llava, struct llava_image_
     fprintf(stderr, "\n");
 
     struct llama_sampling_context * ctx_sampling = llama_sampling_init(params->sparams);
-
+    std::string response = "";
     for (int i = 0; i < max_tgt_len; i++) {
         const char * tmp = sample(ctx_sampling, ctx_llava->ctx_llama, &n_past);
+        response += tmp;
         if (strcmp(tmp, "</s>") == 0) break;
         if (strstr(tmp, "###")) break; // Yi-VL behavior
-
         printf("%s", tmp);
+        if (strstr(response.c_str(), "<|im_end|>")) break; // Yi-34B llava-1.6 - for some reason those decode not as the correct token (tokenizer works)
+        if (strstr(response.c_str(), "<|im_start|>")) break; // Yi-34B llava-1.6
+        if (strstr(response.c_str(), "USER:")) break; // mistral llava-1.6
+
         fflush(stdout);
     }
 
@@ -196,7 +218,8 @@ static struct llava_context * llava_init(gpt_params * params) {
 
     auto ctx_clip = clip_model_load(clip_path, /*verbosity=*/ 1);
 
-    llama_backend_init(params->numa);
+    llama_backend_init();
+    llama_numa_init(params->numa);
 
     llama_model_params model_params = llama_model_params_from_gpt_params(*params);
 

examples/llava/llava-surgery-v2.py

@@ -0,0 +1,155 @@
+import argparse
+import glob
+import os
+import torch
+from safetensors.torch import load as safe_load, save as safe_save, safe_open, save_file
+
+# Function to determine if file is a SafeTensor file
+def is_safetensor_file(file_path):
+    return file_path.endswith('.safetensors')
+
+
+# Unified loading function
+def load_model(file_path):
+    if is_safetensor_file(file_path):
+        tensors = {}
+        with safe_open(file_path, framework="pt", device="cpu") as f:
+            for key in f.keys():
+                tensors[key] = f.get_tensor(key).clone()
+                # output shape
+                print(f"{key} : {tensors[key].shape}")
+        return tensors, 'safetensor'
+    else:
+        return torch.load(file_path, map_location=torch.device('cpu')), 'pytorch'
+
+
+# Unified saving function
+def save_model(model, file_path, file_type):
+    if file_type == 'safetensor':
+        # safe_save(model, file_path)
+        save_file(model, file_path)
+    else:
+        torch.save(model, file_path)
+
+
+# Adapted function to clean vision tower from checkpoint
+def clean_vision_tower_from_checkpoint(checkpoint_path):
+    checkpoint, file_type = load_model(checkpoint_path)
+    # file_type = 'pytorch'
+    model_path = os.path.dirname(checkpoint_path)
+    print(f"Searching for vision tower tensors in {checkpoint_path}")
+    clip_tensors = [k for k, v in checkpoint.items() if (k.startswith("model.vision_tower") or k.startswith("vit."))]
+
+    if len(clip_tensors) > 0:
+        print(f"Found {len(clip_tensors)} tensors to extract from {checkpoint_path}")
+        # Adapted for file type
+        clip_path = os.path.join(model_path, "llava.clip")
+
+        if os.path.exists(clip_path):
+            print(f"Loading existing llava.clip from {clip_path}")
+            existing_clip, _ = load_model(clip_path)
+        else:
+            print(f"Creating new llava.clip at {clip_path}")
+            existing_clip = {}
+        # Update existing_clip with new tensors, avoid duplicates
+        for name in clip_tensors:
+            simple_name = name[name.index('vision_model.'):] if 'vision_model.' in name else name
+            print(f"Adding {simple_name} to llava.clip")
+            if simple_name not in existing_clip:
+                existing_clip[simple_name] = checkpoint[name]
+
+        # Save the updated clip tensors back to llava.clip
+        save_model(existing_clip, clip_path, 'pytorch')
+
+        # Remove the tensors from the original checkpoint
+        for name in clip_tensors:
+            del checkpoint[name]
+
+        checkpoint_path = checkpoint_path
+        return True
+    return False
+
+def find_relevant_checkpoints(checkpoint_paths, newline_criteria, projector):
+    newline_checkpoint_path = None
+    projector_checkpoint_path = None
+
+    for path in checkpoint_paths:
+        checkpoint, _ = load_model(path)
+        if newline_criteria(checkpoint) and newline_checkpoint_path is None:
+            newline_checkpoint_path = path
+        if projector(checkpoint):
+            projector_checkpoint_path = path
+
+    return newline_checkpoint_path, projector_checkpoint_path
+
+def newline_criteria(checkpoint):
+    return any(k.startswith("model.image_newline") for k in checkpoint.keys())
+
+def proj_criteria(checkpoint):
+    return any(k.startswith("model.mm_projector") or k.startswith("vision_proj.") for k in checkpoint.keys())
+
+
+# Command-line interface setup
+ap = argparse.ArgumentParser()
+ap.add_argument("-m", "--model", required=True, help="Path to LLaVA v1.5+ model")
+ap.add_argument("-C", "--clean-vision-tower", action="store_true", help="Remove any vision tower from the model files")
+args = ap.parse_args()
+
+if args.clean_vision_tower:
+    # Generalized to handle both PyTorch and SafeTensors models
+    model_files = sorted(glob.glob(f"{args.model}/*"), key=os.path.getmtime, reverse=True)
+    # checkpoint_paths = [path for path in model_files if (path.endswith('.bin') and path.startswith('pytorch')) or (path.endswith('.safetensors') and path.startswith('model'))]
+    checkpoint_paths = [path for path in model_files if (path.endswith('.bin') and 'pytorch' in path.split('/')[-1].split('\\')[-1]) or (path.endswith('.safetensors') and 'model' in path.split('/')[-1].split('\\')[-1])]
+    for projector_checkpoint_path in checkpoint_paths:
+        print(f"Cleaning {projector_checkpoint_path}")
+        if not clean_vision_tower_from_checkpoint(projector_checkpoint_path):
+            print(f"No vision tower found in {projector_checkpoint_path}")
+            # we break once none is found, so far all models append them at the end
+            # break
+    print("Done! All vision tower tensors are removed from the model files and stored in llava.clip file.")
+
+# Now we look for the projector in the last checkpoint
+model_files = sorted(glob.glob(f"{args.model}/*"), key=os.path.getmtime, reverse=True)
+checkpoint_paths = [path for path in model_files if (path.endswith('.bin') and 'pytorch' in path.split('/')[-1].split('\\')[-1]) or (path.endswith('.safetensors') and 'model' in path.split('/')[-1].split('\\')[-1])]
+# last_checkpoint_path = checkpoint_paths[0]
+# first_checkpoint_path = checkpoint_paths[-1]
+newline_checkpoint_path, projector_checkpoint_path = find_relevant_checkpoints(checkpoint_paths, newline_criteria, proj_criteria)
+
+print(f"Taking projector from {projector_checkpoint_path}")
+first_mm_tensors = []
+first_checkpoint = None
+if newline_checkpoint_path is not None:
+    print(f"Taking newline from {newline_checkpoint_path}")
+    first_checkpoint, file_type = load_model(newline_checkpoint_path)
+    first_mm_tensors = [k for k, v in first_checkpoint.items() if k.startswith("model.image_newline")]
+
+# Load the checkpoint
+mm_tensors = []
+last_checkpoint = None
+if projector_checkpoint_path is not None:
+    last_checkpoint, file_type = load_model(projector_checkpoint_path)
+    mm_tensors = [k for k, v in last_checkpoint.items() if k.startswith("model.mm_projector") or k.startswith("vision_proj.")]
+
+if len(mm_tensors) == 0:
+    if last_checkpoint is not None:
+        for k, v in last_checkpoint.items():
+            print(k)
+    print(f"Found {len(mm_tensors)} tensors to extract out of {len(last_checkpoint)} tensors.")
+    print("No tensors found. Is this a LLaVA model?")
+    exit()
+
+print(f"Found {len(mm_tensors)} tensors to extract.")
+print(f"Found additional {len(first_mm_tensors)} tensors to extract.")
+# projector = {name: checkpoint.[name].float() for name in mm_tensors}
+projector = {}
+for name in mm_tensors:
+    projector[name] = last_checkpoint[name].float()
+for name in first_mm_tensors:
+    projector[name] = first_checkpoint[name].float()
+
+if len(projector) > 0:
+    save_model(projector, f"{args.model}/llava.projector", 'pytorch')
+
+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

@@ -2,32 +2,296 @@
 #include "common.h"
 #include "llama.h"
 #include "llava.h"
+#include "base64.hpp"
 
 #include <cstdio>
 #include <cstdlib>
 #include <vector>
+#include <numeric>
+
+// RGB uint8 image
+struct clip_image_u8 {
+    int nx;
+    int ny;
+
+    std::vector<uint8_t> buf;
+};
+
+// RGB float32 image (NHWC)
+// Memory layout: RGBRGBRGB...
+struct clip_image_f32 {
+    int nx;
+    int ny;
+
+    std::vector<float> buf;
+};
+
+struct clip_image_grid_shape {
+    int first;
+    int second;
+};
+
+/**
+ * Selects the best resolution from a list of possible resolutions based on the original size.
+ *
+ * @param original_size The original size of the image in the format (width, height).
+ * @param possible_resolutions A list of possible resolutions in the format [(width1, height1), (width2, height2), ...].
+ * @return The best fit resolution in the format (width, height).
+ */
+static std::pair<int, int> select_best_resolution(const std::pair<int, int>& original_size, const std::vector<std::pair<int, int>>& possible_resolutions) {
+    int original_width  = original_size.first;
+    int original_height = original_size.second;
+
+    std::pair<int, int> best_fit;
+    int max_effective_resolution = 0;
+    int min_wasted_resolution = std::numeric_limits<int>::max();
+
+    for (const auto& resolution : possible_resolutions) {
+        int width = resolution.first;
+        int height = resolution.second;
+        float scale = std::min(static_cast<float>(width) / original_width, static_cast<float>(height) / original_height);
+        int downscaled_width  = static_cast<int>(original_width * scale);
+        int downscaled_height = static_cast<int>(original_height * scale);
+        int effective_resolution = std::min(downscaled_width * downscaled_height, original_width * original_height);
+        int wasted_resolution = (width * height) - effective_resolution;
+        // fprintf(stderr, "resolution: %d %d, scale: %f, downscaled: %d %d, effective: %d, wasted: %d\n", width, height, scale, downscaled_width, downscaled_height, effective_resolution, wasted_resolution);
+        if (effective_resolution > max_effective_resolution || (effective_resolution == max_effective_resolution && wasted_resolution < min_wasted_resolution)) {
+            max_effective_resolution = effective_resolution;
+            min_wasted_resolution = wasted_resolution;
+            best_fit = resolution;
+        }
+    }
+
+    return best_fit;
+}
+
+/**
+ * @brief Get the anyres image grid shape object
+ *
+ * @param image_size
+ * @param grid_pinpoints
+ * @param image_patch_size
+ * @return <int, int>
+ */
+static struct clip_image_grid_shape get_anyres_image_grid_shape(const std::pair<int, int> & image_size, const std::vector<std::pair<int, int>> & grid_pinpoints, int image_patch_size) {
+    /**
+        Conversion from gguf flat array to vector:
+        std::vector<std::pair<int, int>> possible_resolutions;
+        for (int i = 0; i < 32 && params.image_grid_pinpoints[i] != 0; i+=2) {
+            possible_resolutions.push_back({params.image_grid_pinpoints[i], params.image_grid_pinpoints[i+1]});
+        }
+     */
+    auto best_resolution = select_best_resolution(image_size, grid_pinpoints);
+    return {best_resolution.first / image_patch_size, best_resolution.second / image_patch_size};
+}
+
+// Take the image segments in a grid configuration and return the embeddings and the number of embeddings into preallocated memory (image_embd_out)
+static bool clip_llava_handle_patches(clip_ctx * ctx_clip, std::vector<float *> & image_embd_v, struct clip_image_grid_shape grid_shape, float * image_embd_out, int * n_img_pos_out) {
+    struct {
+        struct ggml_tensor * newline;
+        struct ggml_context * ctx;
+    } model;
+
+    const int32_t image_size = clip_image_size(ctx_clip);
+    const int32_t patch_size = clip_patch_size(ctx_clip);
+
+    int32_t num_patches_per_side = image_size / patch_size; // 336 / 14 = 24 - used for embedding-patching boxes (24*24 = 576 patches)
+
+    int num_patches_width  = grid_shape.first;  // grid 1-4
+    int num_patches_height = grid_shape.second; // grid 1-4
+
+    const size_t num_images = num_patches_width * num_patches_height + 1;
+
+    // TODO: size calculation is not calculated - it's only tens of MB
+    size_t ctx_size = 0;
+
+    {
+        ctx_size += clip_embd_nbytes(ctx_clip) * num_images * 8; // image_features
+        ctx_size += 1024*1024 * ggml_type_size(GGML_TYPE_F32);
+    }
+
+    struct ggml_init_params params {
+        /*.mem_size   =*/ ctx_size,
+        /*.mem_buffer =*/ NULL,
+        /*.no_alloc   =*/ false, // NOTE: this should be false when using the legacy API
+    };
+
+    // Python reference code for full unpad:
+    /*
+        base_image_feature = image_feature[0]
+        image_feature = image_feature[1:]
+        image_feature = image_feature.permute(4, 0, 2, 1, 3).contiguous()
+        image_feature = image_feature.flatten(1, 2).flatten(2, 3)
+        image_feature = unpad_image(image_feature, image_sizes[image_idx])
+        image_feature = torch.cat((
+            image_feature,
+            self.model.image_newline[:, None, None].expand(*image_feature.shape[:-1], 1)
+        ), dim=-1)
+        image_feature = image_feature.flatten(1, 2).transpose(0, 1)
+        image_feature = torch.cat((base_image_feature, image_feature), dim=0)
+    */
+    // We now have two options: unpad or no unpad. Unpad removes tokens for faster llm eval.
+    // In terms of result quality it appears to make no difference, so we'll start with the easier approach given 5D tensors are not supported in ggml yet.
+    // Without unpad we have to split the sub-image embeddings into patches of 24 features each and permute them.
+    // Once all images are processed to prepended the base_image_features without any changes.
+
+    // Pytorch reference simplified, modified for ggml compatibility - confirmed identical output in python (for a 2x2 grid image (676x676 scaling))
+    /*
+        image_feature = image_feature.view(2, 2, 24, 24, 4096)
+        image_feature = image_feature.permute(0, 2, 1, 3, 4).contiguous()
+        image_feature = image_feature.view(2, 24, 2, 24, 4096)
+        image_feature = image_feature.flatten(0, 3)
+
+        // Reshape to 4D tensor by merging the last two dimensions
+        image_feature = image_feature.view(2, 2, 24, 24*4096)
+        image_feature = image_feature.permute(0, 2, 1, 3).contiguous()
+        image_feature = image_feature.view(-1, 4096)
+    */
+
+    model.ctx = ggml_init(params);
+
+    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->buffer == NULL) {
+            printf("newline_tmp tensor buffer is NULL\n");
+        }
+        ggml_backend_tensor_get(newline_tmp, model.newline->data, 0, ggml_nbytes(newline_tmp));
+    } else {
+        model.newline->data = newline_tmp->data;
+        if (model.newline->data == NULL) {
+            printf("newline_tmp tensor data is NULL\n");
+        }
+    }
+
+    struct ggml_tensor * image_features = ggml_new_tensor_3d(model.ctx, GGML_TYPE_F32, clip_n_mmproj_embd(ctx_clip), clip_n_patches(ctx_clip), num_images - 1); // example: 4096 x 576 x 4
+    // ggml_tensor_printf(image_features,"image_features",__LINE__,false,false);
+    // fill it with the image embeddings, ignoring the base
+    for (size_t i = 1; i < num_images; i++) {
+        size_t offset = (i-1) * clip_embd_nbytes(ctx_clip);
+        memcpy((uint8_t *)(image_features->data) + offset, image_embd_v[i], clip_embd_nbytes(ctx_clip));
+    }
+
+    struct ggml_cgraph  * gf = ggml_new_graph(model.ctx);
+    size_t size_ele = ggml_type_size(GGML_TYPE_F32);
+
+    struct ggml_tensor *image_features_patchview = ggml_view_4d(model.ctx, image_features,
+                                                                num_patches_per_side * clip_n_mmproj_embd(ctx_clip),
+                                                                num_patches_per_side,
+                                                                num_patches_width,
+                                                                num_patches_height,
+                                                                size_ele * num_patches_per_side * clip_n_mmproj_embd(ctx_clip),
+                                                                size_ele * num_patches_per_side * clip_n_mmproj_embd(ctx_clip) * num_patches_per_side,
+                                                                size_ele * num_patches_per_side * clip_n_mmproj_embd(ctx_clip) * num_patches_per_side * num_patches_width, 0);
+    // ggml_tensor_printf(image_features_patchview,"image_features_patchview",__LINE__,false,false);
+    struct ggml_tensor *permuted_cont = ggml_cont(model.ctx, ggml_permute(model.ctx, image_features_patchview, 0, 2, 1, 3));
+    /**
+     At the end of each row we have to add the row_end embeddings, which are the same as the newline embeddings
+         image_feature = torch.cat((
+        image_feature,
+        self.model.image_newline[:, None, None].expand(*image_feature.shape[:-1], 1).to(image_feature.device)
+    ), dim=-1)
+     *
+     */
+
+    // ggml_tensor_printf(permuted_cont,"permuted_cont",__LINE__,false,false);
+    struct ggml_tensor *flatten = ggml_view_2d(model.ctx, permuted_cont, clip_n_mmproj_embd(ctx_clip), num_patches_height * num_patches_width * num_patches_per_side * num_patches_per_side,  size_ele * clip_n_mmproj_embd(ctx_clip), 0);
+    // ggml_tensor_printf(flatten,"flatten",__LINE__,false,false);
+    ggml_build_forward_expand(gf, flatten);
+    ggml_graph_compute_with_ctx(model.ctx, gf, 1);
+    struct ggml_tensor* result = gf->nodes[gf->n_nodes - 1];
+
+    memcpy(image_embd_out, image_embd_v[0], clip_embd_nbytes(ctx_clip)); // main image as global context
+    // append without newline tokens (default behavior in llava_arch when not using unpad ):
+    memcpy(image_embd_out + clip_n_patches(ctx_clip) * clip_n_mmproj_embd(ctx_clip), (float*)result->data, clip_embd_nbytes(ctx_clip) * (num_images-1)); // grid patches
+    *n_img_pos_out = static_cast<int>(result->ne[1]+clip_n_patches(ctx_clip));
+
+    // Debug: Test single segments
+    // Current findings: sending base image, sending a segment embedding all works similar to python
+    // However, permuted embeddings do not work yet (stride issue?)
+    // memcpy(image_embd_out, image_embd_v[0], clip_embd_nbytes(ctx_clip)); // main image as context
+    // memcpy(image_embd_out, (float*)prepared_cont->data, clip_embd_nbytes(ctx_clip)); // main image as context
+    // *n_img_pos_out=576;
+
+    ggml_free(model.ctx);
+    return true;
+}
 
-#include "base64.hpp"
 
 static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const clip_image_u8 * img, float * image_embd, int * n_img_pos) {
-    clip_image_f32 * img_res = clip_image_f32_init();
-    if (!clip_image_preprocess(ctx_clip, img, img_res, /*pad2square =*/ true)) {
+    // std::vector<clip_image_f32*> img_res_v; // format VectN x H x W x RGB (N x 336 x 336 x 3), so interleaved RGB - different to the python implementation which is N x 3 x 336 x 336
+    clip_image_f32_batch img_res_v;
+    img_res_v.size = 0;
+    img_res_v.data = nullptr;
+    if (!clip_image_preprocess(ctx_clip, img, img_res_v)) {
         fprintf(stderr, "%s: unable to preprocess image\n", __func__);
-        clip_image_f32_free(img_res);
+        delete[] img_res_v.data;
         return false;
     }
 
-    *n_img_pos = clip_n_patches(ctx_clip);
-
     const int64_t t_img_enc_start_us = ggml_time_us();
-    bool encoded = clip_image_encode(ctx_clip, n_threads, img_res, image_embd);
-    clip_image_f32_free(img_res);
-    if (!encoded) {
-        fprintf(stderr, "Unable to encode image\n");
 
-        return false;
+    const char * mm_patch_merge_type = clip_patch_merge_type(ctx_clip);
+
+    if (strcmp(mm_patch_merge_type, "spatial_unpad") != 0) {
+        // flat / default llava-1.5 type embedding
+        *n_img_pos = clip_n_patches(ctx_clip);
+        bool encoded = clip_image_encode(ctx_clip, n_threads, &img_res_v.data[0], image_embd); // image_embd shape is 576 x 4096
+        delete[] img_res_v.data;
+        if (!encoded) {
+            fprintf(stderr, "Unable to encode image\n");
+
+            return false;
+        }
+    } else {
+        // spatial_unpad llava-1.6 type embedding
+        // TODO: CLIP needs batching support - in HF the llm projection is separate after encoding, which might be a solution to quickly get batching working
+        std::vector<float *> image_embd_v;
+        image_embd_v.resize(img_res_v.size);
+        for (size_t i = 0; i < img_res_v.size; i++) {
+            image_embd_v[i] = (float *)malloc(clip_embd_nbytes(ctx_clip)); // 576 patches * 4096 embeddings * 4 bytes = 9437184
+            const bool encoded = clip_image_encode(ctx_clip, n_threads, &img_res_v.data[i], image_embd_v[i]); // image data is in 3x336x336 format and will be converted to 336x336x3 inside
+            if (!encoded) {
+                fprintf(stderr, "Unable to encode image - spatial_unpad - subimage %d of %d\n", (int) i+1, (int) img_res_v.size);
+                return false;
+            }
+        }
+        const int64_t t_img_enc_batch_us = ggml_time_us();
+        printf("%s: %d segments encoded in %8.2f ms\n", __func__, (int)img_res_v.size, (t_img_enc_batch_us - t_img_enc_start_us) / 1000.0);
+
+        const int32_t * image_grid = clip_image_grid(ctx_clip);
+
+        std::vector<std::pair<int, int>> grid_pinpoints;
+        for (int i = 0; i < 32 && image_grid[i] != 0; i += 2) {
+            grid_pinpoints.push_back({image_grid[i], image_grid[i+1]});
+        }
+
+        // free all img_res_v - not needed anymore
+        delete[] img_res_v.data;
+        img_res_v.size = 0;
+        img_res_v.data = nullptr;
+
+        const int32_t image_size = clip_image_size(ctx_clip);
+
+        struct clip_image_grid_shape grid_shape = get_anyres_image_grid_shape({img->nx,img->ny}, grid_pinpoints, image_size);
+
+        int n_img_pos_out;
+        clip_llava_handle_patches(ctx_clip, image_embd_v, grid_shape, image_embd, &n_img_pos_out);
+        *n_img_pos = n_img_pos_out;
+
+        for (size_t i = 0; i < image_embd_v.size(); i++) {
+            free(image_embd_v[i]);
+        }
+        image_embd_v.clear();
+
+        // debug image/segment/normalization content:
+        // clip_image_u8 * tmp = clip_image_u8_init();
+        // clip_image_convert_f32_to_u8(*image_feature, *tmp);
+        // clip_image_save_to_bmp(*tmp, "image_feature.bmp");
     }
 
+    printf("%s: image embedding created: %d tokens\n", __func__, *n_img_pos);
+
     const int64_t t_img_enc_end_us = ggml_time_us();
     float t_img_enc_ms = (t_img_enc_end_us - t_img_enc_start_us) / 1000.0;
 
@@ -47,11 +311,10 @@ 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) {
-    float * image_embd = (float *)malloc(clip_embd_nbytes(ctx_clip));
+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");
-        free(image_embd);
         return false;
     }
 
@@ -85,7 +348,7 @@ bool llava_eval_image_embed(llama_context * ctx_llama, const struct llava_image_
     return true;
 }
 
-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) {
+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) {
     clip_image_u8 * img = clip_image_u8_init();
     if (!clip_image_load_from_bytes(image_bytes, image_bytes_length, img)) {
         clip_image_u8_free(img);
@@ -142,7 +405,7 @@ static bool load_file_to_bytes(const char* path, unsigned char** bytesOut, long
     return true;
 }
 
-LLAVA_API struct llava_image_embed * llava_image_embed_make_with_filename(struct clip_ctx * ctx_clip, int n_threads, const char * image_path) {
+struct llava_image_embed * llava_image_embed_make_with_filename(struct clip_ctx * ctx_clip, int n_threads, const char * image_path) {
     unsigned char* image_bytes;
     long image_bytes_length;
     auto loaded = load_file_to_bytes(image_path, &image_bytes, &image_bytes_length);
@@ -151,13 +414,13 @@ LLAVA_API struct llava_image_embed * llava_image_embed_make_with_filename(struct
         return NULL;
     }
 
-    auto embed = llava_image_embed_make_with_bytes(ctx_clip, n_threads, image_bytes, image_bytes_length);
+    llava_image_embed *embed = llava_image_embed_make_with_bytes(ctx_clip, n_threads, image_bytes, image_bytes_length);
     free(image_bytes);
 
     return embed;
 }
 
-LLAVA_API void llava_image_embed_free(struct llava_image_embed * embed) {
+void llava_image_embed_free(struct llava_image_embed * embed) {
     free(embed->embed);
     free(embed);
 }

examples/llava/llava.h

@@ -3,7 +3,6 @@
 
 #include "ggml.h"
 
-
 #ifdef LLAMA_SHARED
 #    if defined(_WIN32) && !defined(__MINGW32__)
 #        ifdef LLAMA_BUILD
@@ -32,6 +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 */
@@ -42,7 +43,6 @@ LLAVA_API void llava_image_embed_free(struct llava_image_embed * embed);
 /** write the image represented by embed into the llama context with batch size n_batch, starting at context pos n_past. on completion, n_past points to the next position in the context after the image embed. */
 LLAVA_API bool llava_eval_image_embed(struct llama_context * ctx_llama, const struct llava_image_embed * embed, int n_batch, int * n_past);
 
-
 #ifdef __cplusplus
 }
 #endif

examples/lookahead/lookahead.cpp

@@ -54,7 +54,8 @@ int main(int argc, char ** argv) {
 #endif // LOG_DISABLE_LOGS
 
     // init llama.cpp
-    llama_backend_init(params.numa);
+    llama_backend_init();
+    llama_numa_init(params.numa);
 
     llama_model * model = NULL;
     llama_context * ctx = NULL;

examples/lookup/lookup.cpp

@@ -31,7 +31,8 @@ int main(int argc, char ** argv){
 #endif // LOG_DISABLE_LOGS
 
     // init llama.cpp
-    llama_backend_init(params.numa);
+    llama_backend_init();
+    llama_numa_init(params.numa);
 
     llama_model * model = NULL;
     llama_context * ctx = NULL;

examples/main/README.md

@@ -283,7 +283,11 @@ These options help improve the performance and memory usage of the LLaMA models.
 
 ### NUMA support
 
--   `--numa`: Attempt optimizations that help on some systems with non-uniform memory access. This currently consists of pinning an equal proportion of the threads to the cores on each NUMA node, and disabling prefetch and readahead for mmap. The latter causes mapped pages to be faulted in on first access instead of all at once, and in combination with pinning threads to NUMA nodes, more of the pages end up on the NUMA node where they are used. Note that if the model is already in the system page cache, for example because of a previous run without this option, this will have little effect unless you drop the page cache first. This can be done by rebooting the system or on Linux by writing '3' to '/proc/sys/vm/drop_caches' as root.
+-   `--numa distribute`: Pin an equal proportion of the threads to the cores on each NUMA node. This will spread the load amongst all cores on the system, utilitizing all memory channels at the expense of potentially requiring memory to travel over the slow links between nodes.
+-   `--numa isolate`: Pin all threads to the NUMA node that the program starts on. This limits the number of cores and amount of memory that can be used, but guarantees all memory access remains local to the NUMA node.
+-   `--numa numactl`: Pin threads to the CPUMAP that is passed to the program by starting it with the numactl utility. This is the most flexible mode, and allow arbitraty core usage patterns, for example a map that uses all the cores on one NUMA nodes, and just enough cores on a second node to saturate the inter-node memory bus.
+
+ These flags attempt optimizations that help on some systems with non-uniform memory access. This currently consists of one of the above strategies, and disabling prefetch and readahead for mmap. The latter causes mapped pages to be faulted in on first access instead of all at once, and in combination with pinning threads to NUMA nodes, more of the pages end up on the NUMA node where they are used. Note that if the model is already in the system page cache, for example because of a previous run without this option, this will have little effect unless you drop the page cache first. This can be done by rebooting the system or on Linux by writing '3' to '/proc/sys/vm/drop_caches' as root.
 
 ### Memory Float 32
 

examples/main/main.cpp

@@ -185,7 +185,8 @@ int main(int argc, char ** argv) {
     }
 
     LOG("%s: llama backend init\n", __func__);
-    llama_backend_init(params.numa);
+    llama_backend_init();
+    llama_numa_init(params.numa);
 
     llama_model * model;
     llama_context * ctx;
@@ -333,6 +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
@@ -382,8 +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());
             }
@@ -539,14 +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_shift(ctx, 0, params.n_keep + n_discard, n_past, -n_discard);
 
                     n_past -= n_discard;
 

examples/parallel/parallel.cpp

@@ -122,7 +122,8 @@ int main(int argc, char ** argv) {
 #endif // LOG_DISABLE_LOGS
 
     // init llama.cpp
-    llama_backend_init(params.numa);
+    llama_backend_init();
+    llama_numa_init(params.numa);
 
     llama_model * model = NULL;
     llama_context * ctx = NULL;

examples/passkey/passkey.cpp

@@ -71,7 +71,8 @@ int main(int argc, char ** argv) {
 
     // init LLM
 
-    llama_backend_init(params.numa);
+    llama_backend_init();
+    llama_numa_init(params.numa);
 
     // initialize the model
 

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);
     }
 
@@ -1809,7 +1809,8 @@ int main(int argc, char ** argv) {
         params.prompt = gpt_random_prompt(rng);
     }
 
-    llama_backend_init(params.numa);
+    llama_backend_init();
+    llama_numa_init(params.numa);
 
     llama_model * model;
     llama_context * ctx;

examples/quantize/quantize.cpp

@@ -23,6 +23,7 @@ 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",            },
@@ -31,6 +32,7 @@ static const std::vector<struct quant_option> QUANT_OPTIONS = {
     { "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", },
@@ -237,7 +239,7 @@ int main(int argc, char ** argv) {
         params.imatrix = &imatrix_data;
     }
 
-    llama_backend_init(false);
+    llama_backend_init();
 
     // parse command line arguments
     const std::string fname_inp = argv[arg_idx];
@@ -287,9 +289,10 @@ 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

@@ -16,6 +16,13 @@ Command line options:
 - `--memory-f32`: Use 32-bit floats instead of 16-bit floats for memory key+value. Not recommended.
 - `--mlock`: Lock the model in memory, preventing it from being swapped out when memory-mapped.
 - `--no-mmap`: Do not memory-map the model. By default, models are mapped into memory, which allows the system to load only the necessary parts of the model as needed.
+- `--numa STRATEGY`: Attempt one of the below optimization strategies  that help on some NUMA systems
+- `--numa distribute`: Spread execution evenly over all nodes
+- `--numa isolate`: Only spawn threads on CPUs on the node that execution started on
+- `--numa numactl`: Use the CPU map provided by numactl
+if run without this previously, it is recommended to drop the system page cache before using this
+see https://github.com/ggerganov/llama.cpp/issues/1437
+
 - `--numa`: Attempt optimizations that help on some NUMA systems.
 - `--lora FNAME`: Apply a LoRA (Low-Rank Adaptation) adapter to the model (implies --no-mmap). This allows you to adapt the pretrained model to specific tasks or domains.
 - `--lora-base FNAME`: Optional model to use as a base for the layers modified by the LoRA adapter. This flag is used in conjunction with the `--lora` flag, and specifies the base model for the adaptation.
@@ -32,6 +39,9 @@ Command line options:
 - `--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-predict`: Set the maximum tokens to predict (default: -1)
+- `--slots-endpoint-disable`: To disable slots state monitoring endpoint. Slots state may contain user data, prompts included.
+- `--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)
 
 ## Build
 
@@ -125,9 +135,13 @@ 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.
 
@@ -137,7 +151,7 @@ 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).
 
@@ -189,14 +203,18 @@ 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)
 
+    `samplers`: The order the samplers should be applied in. An array of strings representing sampler type names. If a sampler is not set, it will not be used. If a sampler is specified more than once, it will be applied multiple times. (default: `["top_k", "tfs_z", "typical_p", "top_p", "min_p", "temperature"]` - these are all the available values)
+
 ### Result JSON
 
 - Note: When using streaming mode (`stream`) only `content` and `stop` will be returned until end of completion.
@@ -224,7 +242,7 @@ 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
@@ -370,6 +388,69 @@ 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
+    }
+]
+```
+
 ## 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/server.cpp

@@ -5,6 +5,7 @@
 #include "oai.hpp"
 
 #include "../llava/clip.h"
+#include "../llava/llava.h"
 
 #include "stb_image.h"
 
@@ -28,6 +29,7 @@
 #include <chrono>
 #include <condition_variable>
 #include <atomic>
+#include <signal.h>
 
 using json = nlohmann::json;
 
@@ -36,10 +38,11 @@ struct server_params
     std::string hostname = "127.0.0.1";
     std::vector<std::string> api_keys;
     std::string public_path = "examples/server/public";
-    std::string chat_template = "chatml";
+    std::string chat_template = "";
     int32_t port = 8080;
     int32_t read_timeout = 600;
     int32_t write_timeout = 600;
+    bool slots_endpoint = true;
 };
 
 bool server_verbose = false;
@@ -158,6 +161,7 @@ struct llama_client_slot
     int32_t n_decoded   = 0;
     int32_t n_remaining = -1;
     int32_t i_batch     = -1;
+    int32_t n_predict   = -1;
 
     int32_t num_prompt_tokens           = 0;
     int32_t num_prompt_tokens_processed = 0;
@@ -396,6 +400,16 @@ struct llama_server_context
         return true;
     }
 
+    void validate_model_chat_template(server_params & sparams) {
+        llama_chat_message chat[] = {{"user", "test"}};
+        std::vector<char> buf(1);
+        int res = llama_chat_apply_template(model, nullptr, chat, 1, true, buf.data(), buf.size());
+        if (res < 0) {
+            LOG_ERROR("The chat template comes with this model is not yet supported, falling back to chatml. This may cause the model to output suboptimal responses", {});
+            sparams.chat_template = "<|im_start|>"; // llama_chat_apply_template only checks if <|im_start|> exist in the template
+        }
+    }
+
     void initialize() {
         // create slots
         all_slots_are_idle = true;
@@ -409,6 +423,7 @@ struct llama_server_context
 
             slot.id = i;
             slot.n_ctx = n_ctx_slot;
+            slot.n_predict = params.n_predict;
 
             LOG_TEE(" -> Slot %i - max context: %i\n", slot.id, n_ctx_slot);
 
@@ -436,10 +451,6 @@ struct llama_server_context
         default_generation_settings_for_props["seed"] = -1;
 
         batch = llama_batch_init(n_ctx, 0, params.n_parallel);
-
-        // empty system prompt
-        system_prompt = "";
-        system_tokens.clear();
     }
 
     std::vector<llama_token> tokenize(const json & json_prompt, bool add_bos) const
@@ -548,6 +559,16 @@ struct llama_server_context
         slot->params.seed               = json_value(data, "seed",              default_params.seed);
         slot->sparams.grammar           = json_value(data, "grammar",           default_sparams.grammar);
         slot->sparams.n_probs           = json_value(data, "n_probs",           default_sparams.n_probs);
+        slot->sparams.min_keep          = json_value(data, "min_keep",          default_sparams.min_keep);
+
+        if (slot->n_predict > 0 && slot->params.n_predict > slot->n_predict) {
+            // Might be better to reject the request with a 400 ?
+            LOG_WARNING("Max tokens to predict exceeds server configuration", {
+                {"params.n_predict", slot->params.n_predict},
+                {"slot.n_predict", slot->n_predict},
+            });
+            slot->params.n_predict = slot->n_predict;
+        }
 
         // infill
         if (data.count("input_prefix") != 0)
@@ -676,6 +697,24 @@ struct llama_server_context
             }
         }
 
+        const auto &samplers_sequence = data.find("samplers");
+        if (samplers_sequence != data.end() && samplers_sequence->is_array())
+        {
+            std::vector<std::string> sampler_names;
+            for (const auto &sampler_name : *samplers_sequence)
+            {
+                if (sampler_name.is_string())
+                {
+                    sampler_names.emplace_back(sampler_name);
+                }
+            }
+            slot->sparams.samplers_sequence = sampler_types_from_names(sampler_names, false);
+        }
+        else
+        {
+            slot->sparams.samplers_sequence = default_sparams.samplers_sequence;
+        }
+
         if (multimodal)
         {
             const auto &images_data = data.find("image_data");
@@ -765,27 +804,30 @@ struct llama_server_context
     }
 
     void update_system_prompt() {
-        system_tokens = ::llama_tokenize(ctx, system_prompt, add_bos_token);
-
-        llama_batch_clear(batch);
-
         kv_cache_clear();
+        system_tokens.clear();
 
-        for (int i = 0; i < (int) system_tokens.size(); ++i)
-        {
-            llama_batch_add(batch, system_tokens[i], i, { 0 }, false);
-        }
+        if (!system_prompt.empty()) {
+            system_tokens = ::llama_tokenize(ctx, system_prompt, add_bos_token);
 
-        if (llama_decode(ctx, batch) != 0)
-        {
-            LOG_TEE("%s: llama_decode() failed\n", __func__);
-            return;
-        }
+            llama_batch_clear(batch);
 
-        // assign the system KV cache to all parallel sequences
-        for (int32_t i = 1; i < params.n_parallel; ++i)
-        {
-            llama_kv_cache_seq_cp(ctx, 0, i, 0, system_tokens.size());
+            for (int i = 0; i < (int)system_tokens.size(); ++i)
+            {
+                llama_batch_add(batch, system_tokens[i], i, { 0 }, false);
+            }
+
+            if (llama_decode(ctx, batch) != 0)
+            {
+                LOG_TEE("%s: llama_decode() failed\n", __func__);
+                return;
+            }
+
+            // assign the system KV cache to all parallel sequences
+            for (int32_t i = 1; i < params.n_parallel; ++i)
+            {
+                llama_kv_cache_seq_cp(ctx, 0, i, 0, system_tokens.size());
+            }
         }
 
         LOG_TEE("system prompt updated\n");
@@ -807,10 +849,8 @@ struct llama_server_context
         name_user      = sys_props.value("anti_prompt", "");
         name_assistant = sys_props.value("assistant_name", "");
 
-        if (slots.size() > 0)
-        {
-            notify_system_prompt_changed();
-        }
+
+        notify_system_prompt_changed();
     }
 
     static size_t find_stopping_strings(const std::string &text, const size_t last_token_size,
@@ -968,28 +1008,13 @@ struct llama_server_context
             {
                 continue;
             }
-            clip_image_f32 * img_res = clip_image_f32_init();
-            if (!clip_image_preprocess(clp_ctx, img.img_data, img_res, /*pad2square =*/ true))
-            {
+
+            if (!llava_image_embed_make_with_clip_img(clp_ctx, params.n_threads, img.img_data, &img.image_embedding, &img.image_tokens)) {
                 LOG_TEE("Error processing the given image");
-                clip_free(clp_ctx);
                 return false;
             }
-            img.image_tokens = clip_n_patches(clp_ctx);
-            img.image_embedding = (float *)malloc(clip_embd_nbytes(clp_ctx));
-            if (!img.image_embedding)
-            {
-                LOG_TEE("Unable to allocate memory for image embeddings\n");
-                clip_free(clp_ctx);
-                return false;
-            }
-            LOG_TEE("slot %i - encoding image [id: %i]\n", slot.id, img.id);
-            if (!clip_image_encode(clp_ctx, params.n_threads, img_res, img.image_embedding))
-            {
-                LOG_TEE("Unable to encode image\n");
-                return false;
-            }
-            clip_image_f32_free(img_res);
+
+
             img.request_encode_image = false;
         }
 
@@ -1013,8 +1038,15 @@ struct llama_server_context
         const auto eos_bias = slot.sparams.logit_bias.find(llama_token_eos(model));
         const bool ignore_eos = eos_bias != slot.sparams.logit_bias.end() &&
                                 eos_bias->second < 0.0f && std::isinf(eos_bias->second);
+        std::vector<std::string> samplers_sequence;
+        for (const auto &sampler_type : slot.sparams.samplers_sequence)
+        {
+            samplers_sequence.emplace_back(sampler_type_to_name_string(sampler_type));
+        }
+
         return json {
             {"n_ctx",             slot.n_ctx},
+            {"n_predict",         slot.n_predict},
             {"model",             params.model_alias},
             {"seed",              slot.params.seed},
             {"temperature",       slot.sparams.temp},
@@ -1042,7 +1074,9 @@ struct llama_server_context
             {"stream",            slot.params.stream},
             {"logit_bias",        slot.sparams.logit_bias},
             {"n_probs",           slot.sparams.n_probs},
+            {"min_keep",          slot.sparams.min_keep},
             {"grammar",           slot.sparams.grammar},
+            {"samplers",          samplers_sequence}
         };
     }
 
@@ -1370,6 +1404,46 @@ struct llama_server_context
             case TASK_TYPE_NEXT_RESPONSE: {
                 // do nothing
             } break;
+            case TASK_TYPE_SLOTS_DATA: {
+                json slots_data        = json::array();
+                int n_idle_slots       = 0;
+                int n_processing_slots = 0;
+
+                for (llama_client_slot &slot: slots) {
+                    if (slot.available()) {
+                        n_idle_slots++;
+                    } else {
+                        n_processing_slots++;
+                    }
+                    json slot_data = get_formated_generation(slot);
+                    slot_data["id"] = slot.id;
+                    slot_data["task_id"] = slot.task_id;
+                    slot_data["state"] = slot.state;
+                    slot_data["prompt"] = slot.prompt;
+                    slot_data["next_token"] = {
+                            {"has_next_token", slot.has_next_token},
+                            {"n_remain", slot.n_remaining},
+                            {"num_tokens_predicted", slot.n_decoded},
+                            {"stopped_eos", slot.stopped_eos},
+                            {"stopped_word", slot.stopped_word},
+                            {"stopped_limit", slot.stopped_limit},
+                            {"stopping_word", slot.stopping_word},
+                    };
+                    slots_data.push_back(slot_data);
+                }
+                LOG_TEE("task %i - slots data: idle=%i processing=%i\n", task.id, n_idle_slots, n_processing_slots);
+                task_result res;
+                res.id = task.id;
+                res.multitask_id = task.multitask_id;
+                res.stop = true;
+                res.error = false;
+                res.result_json = {
+                        { "idle",       n_idle_slots       },
+                        { "processing", n_processing_slots },
+                        { "slots",      slots_data         }
+                };
+                queue_results.send(res);
+            } break;
         }
     }
 
@@ -1423,14 +1497,15 @@ struct llama_server_context
                 if (slot.is_processing() && system_tokens.size() + slot.cache_tokens.size() >= (size_t) slot.n_ctx)
                 {
                     // Shift context
-                    const int n_left    = system_tokens.size() + slot.n_past - slot.params.n_keep - 1;
+                    const int n_keep    = slot.params.n_keep + add_bos_token;
+                    const int n_left    = system_tokens.size() + slot.n_past - n_keep;
                     const int n_discard = n_left / 2;
 
-                    LOG_TEE("slot %d: context shift - n_keep = %d, n_left = %d, n_discard = %d\n", slot.id, slot.params.n_keep, n_left, n_discard);
-                    llama_kv_cache_seq_rm   (ctx, slot.id, slot.params.n_keep + 1            , slot.params.n_keep + n_discard + 1);
-                    llama_kv_cache_seq_shift(ctx, slot.id, slot.params.n_keep + 1 + n_discard, system_tokens.size() + slot.n_past, -n_discard);
+                    LOG_TEE("slot %d: context shift - n_keep = %d, n_left = %d, n_discard = %d\n", slot.id, n_keep, n_left, n_discard);
+                    llama_kv_cache_seq_rm   (ctx, slot.id, n_keep            , n_keep + n_discard);
+                    llama_kv_cache_seq_shift(ctx, slot.id, n_keep + n_discard, system_tokens.size() + slot.n_past, -n_discard);
 
-                    for (size_t i = slot.params.n_keep + 1 + n_discard; i < slot.cache_tokens.size(); i++)
+                    for (size_t i = n_keep + n_discard; i < slot.cache_tokens.size(); i++)
                     {
                         slot.cache_tokens[i - n_discard] = slot.cache_tokens[i];
                     }
@@ -1443,7 +1518,7 @@ struct llama_server_context
 
                     LOG_VERBOSE("context shift", {
                         { "n_ctx", n_ctx },
-                        { "n_keep", params.n_keep },
+                        { "n_keep", n_keep },
                         { "n_left", n_left },
                     });
                 }
@@ -1839,7 +1914,10 @@ static void server_print_usage(const char *argv0, const gpt_params &params,
     {
         printf("  --no-mmap                 do not memory-map model (slower load but may reduce pageouts if not using mlock)\n");
     }
-    printf("  --numa                    attempt optimizations that help on some NUMA systems\n");
+    printf("  --numa TYPE               attempt optimizations that help on some NUMA systems\n");
+    printf("                              - distribute: spread execution evenly over all nodes\n");
+    printf("                              - isolate: only spawn threads on CPUs on the node that execution started on\n");
+    printf("                              - numactl: use the CPU map provided my numactl\n");
     if (llama_supports_gpu_offload()) {
         printf("  -ngl N, --n-gpu-layers N\n");
         printf("                            number of layers to store in VRAM\n");
@@ -1872,14 +1950,17 @@ static void server_print_usage(const char *argv0, const gpt_params &params,
     printf("                            set a file to load a system prompt (initial prompt of all slots), this is useful for chat applications.\n");
     printf("  --mmproj MMPROJ_FILE      path to a multimodal projector file for LLaVA.\n");
     printf("  --log-disable             disables logging to a file.\n");
+    printf("  --slots-endpoint-disable  disables slots monitoring endpoint.\n");
     printf("\n");
+    printf("  -n, --n-predict           maximum tokens to predict (default: %d)\n", params.n_predict);
     printf("  --override-kv KEY=TYPE:VALUE\n");
     printf("                            advanced option to override model metadata by key. may be specified multiple times.\n");
     printf("                            types: int, float, bool. example: --override-kv tokenizer.ggml.add_bos_token=bool:false\n");
     printf("  -gan N, --grp-attn-n N    set the group attention factor to extend context size through self-extend(default: 1=disabled), used together with group attention width `--grp-attn-w`");
     printf("  -gaw N, --grp-attn-w N    set the group attention width to extend context size through self-extend(default: 512), used together with group attention factor `--grp-attn-n`");
-    printf("  --chat-template FORMAT_NAME");
-    printf("                            set chat template, possible valus is: llama2, chatml (default %s)", sparams.chat_template.c_str());
+    printf("  --chat-template JINJA_TEMPLATE\n");
+    printf("                            set custom jinja chat template (default: template taken from model's metadata)\n");
+    printf("                            Note: only commonly used templates are accepted, since we don't have jinja parser\n");
     printf("\n");
 }
 
@@ -2248,9 +2329,17 @@ static void server_params_parse(int argc, char **argv, server_params &sparams,
         {
             params.use_mmap = false;
         }
-        else if (arg == "--numa")
-        {
-            params.numa = true;
+        else if (arg == "--numa") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            } else {
+                std::string value(argv[i]);
+                /**/ if (value == "distribute" || value == "" ) { params.numa = GGML_NUMA_STRATEGY_DISTRIBUTE; }
+                else if (value == "isolate") { params.numa = GGML_NUMA_STRATEGY_ISOLATE; }
+                else if (value == "numactl") { params.numa = GGML_NUMA_STRATEGY_NUMACTL; }
+                else { invalid_param = true; break; }
+            }
         }
         else if (arg == "--embedding")
         {
@@ -2311,6 +2400,10 @@ static void server_params_parse(int argc, char **argv, server_params &sparams,
             log_set_target(stdout);
             LOG_INFO("logging to file is disabled.", {});
         }
+        else if (arg == "--slots-endpoint-disable")
+        {
+            sparams.slots_endpoint = false;
+        }
         else if (arg == "--chat-template")
         {
             if (++i >= argc)
@@ -2318,13 +2411,13 @@ static void server_params_parse(int argc, char **argv, server_params &sparams,
                 invalid_param = true;
                 break;
             }
-            std::string value(argv[i]);
-            if (value != "chatml" && value != "llama2") {
-                fprintf(stderr, "error: chat template can be \"llama2\" or \"chatml\", but got: %s\n", value.c_str());
+            if (!verify_custom_template(argv[i])) {
+                fprintf(stderr, "error: the supplied chat template is not supported: %s\n", argv[i]);
+                fprintf(stderr, "note: llama.cpp does not use jinja parser, we only support commonly used templates\n");
                 invalid_param = true;
                 break;
             }
-            sparams.chat_template = value;
+            sparams.chat_template = argv[i];
         }
         else if (arg == "--override-kv")
         {
@@ -2462,6 +2555,9 @@ static void append_to_generated_text_from_generated_token_probs(llama_server_con
     }
 }
 
+std::function<void(int)> shutdown_handler;
+inline void signal_handler(int signal) { shutdown_handler(signal); }
+
 int main(int argc, char **argv)
 {
 #if SERVER_VERBOSE != 1
@@ -2481,7 +2577,8 @@ int main(int argc, char **argv)
         params.model_alias = params.model;
     }
 
-    llama_backend_init(params.numa);
+    llama_backend_init();
+    llama_numa_init(params.numa);
 
     LOG_INFO("build info", {{"build", LLAMA_BUILD_NUMBER},
                             {"commit", LLAMA_COMMIT}});
@@ -2507,13 +2604,44 @@ int main(int argc, char **argv)
         res.set_header("Access-Control-Allow-Headers", "*");
     });
 
-    svr.Get("/health", [&](const httplib::Request&, httplib::Response& res) {
+    svr.Get("/health", [&](const httplib::Request& req, httplib::Response& res) {
         server_state current_state = state.load();
         switch(current_state) {
-            case SERVER_STATE_READY:
-                res.set_content(R"({"status": "ok"})", "application/json");
+            case SERVER_STATE_READY: {
+                // request slots data using task queue
+                task_server task;
+                task.id   = llama.queue_tasks.get_new_id();
+                task.type = TASK_TYPE_SLOTS_DATA;
+                task.target_id = -1;
+
+                llama.queue_results.add_waiting_task_id(task.id);
+                llama.queue_tasks.post(task);
+
+                // get the result
+                task_result result = llama.queue_results.recv(task.id);
+                llama.queue_results.remove_waiting_task_id(task.id);
+
+                int n_idle_slots       = result.result_json["idle"];
+                int n_processing_slots = result.result_json["processing"];
+
+                json health = {
+                        {"status",           "ok"},
+                        {"slots_idle",       n_idle_slots},
+                        {"slots_processing", n_processing_slots}};
                 res.status = 200; // HTTP OK
+                if (sparams.slots_endpoint && req.has_param("include_slots")) {
+                    health["slots"] = result.result_json["slots"];
+                }
+
+                if (n_idle_slots == 0) {
+                    health["status"] = "no slot available";
+                    if (req.has_param("fail_on_no_slot")) {
+                        res.status = 503; // HTTP Service Unavailable
+                    }
+                }
+                res.set_content(health.dump(), "application/json");
                 break;
+            }
             case SERVER_STATE_LOADING_MODEL:
                 res.set_content(R"({"status": "loading model"})", "application/json");
                 res.status = 503; // HTTP Service Unavailable
@@ -2525,6 +2653,26 @@ int main(int argc, char **argv)
         }
     });
 
+    if (sparams.slots_endpoint) {
+        svr.Get("/slots", [&](const httplib::Request&, httplib::Response& res) {
+            // request slots data using task queue
+            task_server task;
+            task.id = llama.queue_tasks.get_new_id();
+            task.type = TASK_TYPE_SLOTS_DATA;
+            task.target_id = -1;
+
+            llama.queue_results.add_waiting_task_id(task.id);
+            llama.queue_tasks.post(task);
+
+            // get the result
+            task_result result = llama.queue_results.recv(task.id);
+            llama.queue_results.remove_waiting_task_id(task.id);
+
+            res.set_content(result.result_json["slots"].dump(), "application/json");
+            res.status = 200; // HTTP OK
+        });
+    }
+
     svr.set_logger(log_server_request);
 
     svr.set_exception_handler([](const httplib::Request &, httplib::Response &res, std::exception_ptr ep)
@@ -2614,6 +2762,11 @@ int main(int argc, char **argv)
         LOG_INFO("model loaded", {});
     }
 
+    if (sparams.chat_template.empty()) { // custom chat template is not supplied
+        // check if the template comes with the model is supported by us
+        llama.validate_model_chat_template(sparams);
+    }
+
     // Middleware for API key validation
     auto validate_api_key = [&sparams](const httplib::Request &req, httplib::Response &res) -> bool {
         // If API key is not set, skip validation
@@ -2785,7 +2938,7 @@ int main(int argc, char **argv)
                 if (!validate_api_key(req, res)) {
                     return;
                 }
-                json data = oaicompat_completion_params_parse(json::parse(req.body), sparams.chat_template);
+                json data = oaicompat_completion_params_parse(llama.model, json::parse(req.body), sparams.chat_template);
 
                 const int task_id = llama.queue_tasks.get_new_id();
                 llama.queue_results.add_waiting_task_id(task_id);
@@ -3078,8 +3231,25 @@ int main(int argc, char **argv)
         std::placeholders::_2,
         std::placeholders::_3
     ));
-    llama.queue_tasks.start_loop();
 
+    shutdown_handler = [&](int) {
+        llama.queue_tasks.terminate();
+    };
+
+#if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__))
+    struct sigaction sigint_action;
+    sigint_action.sa_handler = signal_handler;
+    sigemptyset (&sigint_action.sa_mask);
+    sigint_action.sa_flags = 0;
+    sigaction(SIGINT, &sigint_action, NULL);
+#elif defined (_WIN32)
+    auto console_ctrl_handler = +[](DWORD ctrl_type) -> BOOL {
+        return (ctrl_type == CTRL_C_EVENT) ? (signal_handler(SIGINT), true) : false;
+    };
+    SetConsoleCtrlHandler(reinterpret_cast<PHANDLER_ROUTINE>(console_ctrl_handler), true);
+#endif
+    llama.queue_tasks.start_loop();
+    svr.stop();
     t.join();
 
     llama_backend_free();

examples/server/utils.hpp

@@ -49,7 +49,8 @@ enum server_state {
 enum task_type {
     TASK_TYPE_COMPLETION,
     TASK_TYPE_CANCEL,
-    TASK_TYPE_NEXT_RESPONSE
+    TASK_TYPE_NEXT_RESPONSE,
+    TASK_TYPE_SLOTS_DATA
 };
 
 struct task_server {
@@ -167,50 +168,47 @@ 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 +218,7 @@ 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;
@@ -278,9 +277,18 @@ 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", {});
@@ -324,8 +332,12 @@ 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);
                     });
                 }
             }

examples/simple/simple.cpp

@@ -31,7 +31,8 @@ int main(int argc, char ** argv) {
 
     // init LLM
 
-    llama_backend_init(params.numa);
+    llama_backend_init();
+    llama_numa_init(params.numa);
 
     // initialize the model
 

examples/speculative/speculative.cpp

@@ -50,7 +50,8 @@ int main(int argc, char ** argv) {
 #endif // LOG_DISABLE_LOGS
 
     // init llama.cpp
-    llama_backend_init(params.numa);
+    llama_backend_init();
+    llama_numa_init(params.numa);
 
     llama_model * model_tgt = NULL;
     llama_model * model_dft = NULL;

examples/tokenize/tokenize.cpp

@@ -17,7 +17,7 @@ int main(int argc, char ** argv) {
 
     const bool printing_ids = argc > 3 && std::string(argv[3]) == "--ids";
 
-    llama_backend_init(false);
+    llama_backend_init();
 
     llama_model_params model_params = llama_model_default_params();
     model_params.vocab_only = true;

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

@@ -50,9 +50,9 @@ struct my_llama_layer {
     struct ggml_tensor * ffn_norm;
 
     // ff
-    struct ggml_tensor * w1;
-    struct ggml_tensor * w2;
-    struct ggml_tensor * w3;
+    struct ggml_tensor * ffn_gate; // w1
+    struct ggml_tensor * ffn_down; // w2
+    struct ggml_tensor * ffn_up;   // w3
 };
 
 struct my_llama_model {
@@ -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) {
@@ -140,9 +140,9 @@ static void set_param_model(struct my_llama_model * model) {
         ggml_set_param(ctx, layer.wv);
         ggml_set_param(ctx, layer.wo);
         ggml_set_param(ctx, layer.ffn_norm);
-        ggml_set_param(ctx, layer.w1);
-        ggml_set_param(ctx, layer.w2);
-        ggml_set_param(ctx, layer.w3);
+        ggml_set_param(ctx, layer.ffn_gate);
+        ggml_set_param(ctx, layer.ffn_down);
+        ggml_set_param(ctx, layer.ffn_up);
     }
 }
 
@@ -198,9 +198,9 @@ static void init_model(struct my_llama_model * model) {
 
         layer.ffn_norm = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
 
-        layer.w1 = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd,   n_ff);
-        layer.w2 = ggml_new_tensor_2d(ctx, GGML_TYPE_F32,   n_ff, n_embd);
-        layer.w3 = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd,   n_ff);
+        layer.ffn_gate = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd,   n_ff);
+        layer.ffn_down = ggml_new_tensor_2d(ctx, GGML_TYPE_F32,   n_ff, n_embd);
+        layer.ffn_up   = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd,   n_ff);
 
         ggml_set_name(layer.attention_norm, tni(LLM_TENSOR_ATTN_NORM, i));
 
@@ -211,9 +211,9 @@ static void init_model(struct my_llama_model * model) {
 
         ggml_set_name(layer.ffn_norm,       tni(LLM_TENSOR_FFN_NORM, i));
 
-        ggml_set_name(layer.w1,             tni(LLM_TENSOR_FFN_GATE, i));
-        ggml_set_name(layer.w2,             tni(LLM_TENSOR_FFN_DOWN, i));
-        ggml_set_name(layer.w3,             tni(LLM_TENSOR_FFN_UP, i));
+        ggml_set_name(layer.ffn_gate,       tni(LLM_TENSOR_FFN_GATE, i));
+        ggml_set_name(layer.ffn_down,       tni(LLM_TENSOR_FFN_DOWN, i));
+        ggml_set_name(layer.ffn_up,         tni(LLM_TENSOR_FFN_UP, i));
     }
 
     set_param_model(model);
@@ -244,9 +244,9 @@ static void randomize_model(struct my_llama_model * model, int seed, float mean,
 
         randomize_tensor_normal(layer.ffn_norm, rnd);
 
-        randomize_tensor_normal(layer.w1, rnd);
-        randomize_tensor_normal(layer.w2, rnd);
-        randomize_tensor_normal(layer.w3, rnd);
+        randomize_tensor_normal(layer.ffn_gate, rnd);
+        randomize_tensor_normal(layer.ffn_down, rnd);
+        randomize_tensor_normal(layer.ffn_up,   rnd);
     }
 
     free_random_normal_distribution(rnd);
@@ -356,11 +356,11 @@ static struct ggml_tensor * llama_build_train_graphs(
         struct ggml_tensor * t22 = ggml_rms_norm     (ctx, t21, f_norm_rms_eps);                    set_name(t22, "t22");     assert_shape_2d(t22, n_embd, N*n_batch);
         struct ggml_tensor * t23 = ggml_repeat       (ctx, layer.ffn_norm, t22);                    set_name(t23, "t23");     assert_shape_2d(t23, n_embd, N*n_batch);
         struct ggml_tensor * t24 = ggml_mul          (ctx, t23, t22);                               set_name(t24, "t24");     assert_shape_2d(t24, n_embd, N*n_batch);
-        struct ggml_tensor * t25 = ggml_mul_mat      (ctx, layer.w3, t24);                          set_name(t25, "t25");     assert_shape_2d(t25, n_ff, N*n_batch);
-        struct ggml_tensor * t26 = ggml_mul_mat      (ctx, layer.w1, t24);                          set_name(t26, "t26");     assert_shape_2d(t26, n_ff, N*n_batch);
+        struct ggml_tensor * t25 = ggml_mul_mat      (ctx, layer.ffn_up, t24);                      set_name(t25, "t25");     assert_shape_2d(t25, n_ff, N*n_batch);
+        struct ggml_tensor * t26 = ggml_mul_mat      (ctx, layer.ffn_gate, t24);                    set_name(t26, "t26");     assert_shape_2d(t26, n_ff, N*n_batch);
         struct ggml_tensor * t27 = ggml_silu         (ctx, t26);                                    set_name(t27, "t27");     assert_shape_2d(t27, n_ff, N*n_batch);
         struct ggml_tensor * t28 = ggml_mul          (ctx, t27, t25);                               set_name(t28, "t28");     assert_shape_2d(t28, n_ff, N*n_batch);
-        struct ggml_tensor * t29 = ggml_mul_mat      (ctx, layer.w2, t28);                          set_name(t29, "t29");     assert_shape_2d(t29, n_embd, N*n_batch);
+        struct ggml_tensor * t29 = ggml_mul_mat      (ctx, layer.ffn_down, t28);                    set_name(t29, "t29");     assert_shape_2d(t29, n_embd, N*n_batch);
         struct ggml_tensor * t30 = ggml_add          (ctx, t29, t21);                               set_name(t30, "t30");     assert_shape_2d(t30, n_embd, N*n_batch);
         cur = t30;
         checkpoints.push_back(cur);
@@ -521,9 +521,9 @@ static void load_llama_model_gguf(struct gguf_context * fctx, struct ggml_contex
         copy_tensor_by_name(layer.wv,             f_ggml_ctx, tni(LLM_TENSOR_ATTN_V, i));
         copy_tensor_by_name(layer.wo,             f_ggml_ctx, tni(LLM_TENSOR_ATTN_OUT, i));
         copy_tensor_by_name(layer.ffn_norm,       f_ggml_ctx, tni(LLM_TENSOR_FFN_NORM, i));
-        copy_tensor_by_name(layer.w1,             f_ggml_ctx, tni(LLM_TENSOR_FFN_GATE, i));
-        copy_tensor_by_name(layer.w2,             f_ggml_ctx, tni(LLM_TENSOR_FFN_DOWN, i));
-        copy_tensor_by_name(layer.w3,             f_ggml_ctx, tni(LLM_TENSOR_FFN_UP, i));
+        copy_tensor_by_name(layer.ffn_gate,       f_ggml_ctx, tni(LLM_TENSOR_FFN_GATE, i));
+        copy_tensor_by_name(layer.ffn_down,       f_ggml_ctx, tni(LLM_TENSOR_FFN_DOWN, i));
+        copy_tensor_by_name(layer.ffn_up,         f_ggml_ctx, tni(LLM_TENSOR_FFN_UP, i));
     }
 }
 
@@ -664,9 +664,9 @@ static void save_llama_model_gguf(struct gguf_context * fctx, const char * fn_vo
         gguf_add_tensor(fctx, layer.wv);
         gguf_add_tensor(fctx, layer.wo);
         gguf_add_tensor(fctx, layer.ffn_norm);
-        gguf_add_tensor(fctx, layer.w1);
-        gguf_add_tensor(fctx, layer.w2);
-        gguf_add_tensor(fctx, layer.w3);
+        gguf_add_tensor(fctx, layer.ffn_gate);
+        gguf_add_tensor(fctx, layer.ffn_down);
+        gguf_add_tensor(fctx, layer.ffn_up);
     }
 }
 
@@ -915,9 +915,9 @@ static int64_t get_parameter_count(struct my_llama_model* model) {
         nx += ggml_nelements(layer.wv);
         nx += ggml_nelements(layer.wo);
         nx += ggml_nelements(layer.ffn_norm);
-        nx += ggml_nelements(layer.w1);
-        nx += ggml_nelements(layer.w2);
-        nx += ggml_nelements(layer.w3);
+        nx += ggml_nelements(layer.ffn_gate);
+        nx += ggml_nelements(layer.ffn_down);
+        nx += ggml_nelements(layer.ffn_up);
     }
     return nx;
 }

flake.lock

@@ -20,11 +20,11 @@
     },
     "nixpkgs": {
       "locked": {
-        "lastModified": 1707268954,
-        "narHash": "sha256-2en1kvde3cJVc3ZnTy8QeD2oKcseLFjYPLKhIGDanQ0=",
+        "lastModified": 1708118438,
+        "narHash": "sha256-kk9/0nuVgA220FcqH/D2xaN6uGyHp/zoxPNUmPCMmEE=",
         "owner": "NixOS",
         "repo": "nixpkgs",
-        "rev": "f8e2ebd66d097614d51a56a755450d4ae1632df1",
+        "rev": "5863c27340ba4de8f83e7e3c023b9599c3cb3c80",
         "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

@@ -219,6 +219,10 @@ GGML_CALL void ggml_backend_tensor_set(struct ggml_tensor * tensor, const void *
     GGML_ASSERT(buf != NULL && "tensor buffer not set");
     GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor write out of bounds");
 
+    if (!size) {
+        return;
+    }
+
     tensor->buffer->iface.set_tensor(buf, tensor, data, offset, size);
 }
 
@@ -229,6 +233,10 @@ GGML_CALL void ggml_backend_tensor_get(const struct ggml_tensor * tensor, void *
     GGML_ASSERT(tensor->buffer != NULL && "tensor buffer not set");
     GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor read out of bounds");
 
+    if (!size) {
+        return;
+    }
+
     tensor->buffer->iface.get_tensor(buf, tensor, data, offset, size);
 }
 
@@ -748,7 +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:
@@ -998,6 +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
@@ -1032,7 +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);
@@ -1079,7 +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,
@@ -1135,7 +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);
@@ -1247,7 +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) {
@@ -1306,7 +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
@@ -1353,7 +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 */) {
@@ -1659,7 +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);
     }
@@ -1688,7 +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-cuda.cu

@@ -1,3 +1,7 @@
+#include "ggml-cuda.h"
+#include "ggml.h"
+#include "ggml-backend-impl.h"
+
 #include <algorithm>
 #include <assert.h>
 #include <atomic>
@@ -54,6 +58,8 @@
 #define cudaDeviceProp hipDeviceProp_t
 #define cudaDeviceSynchronize hipDeviceSynchronize
 #define cudaError_t hipError_t
+#define cudaErrorPeerAccessAlreadyEnabled hipErrorPeerAccessAlreadyEnabled
+#define cudaErrorPeerAccessNotEnabled hipErrorPeerAccessNotEnabled
 #define cudaEventCreateWithFlags hipEventCreateWithFlags
 #define cudaEventDisableTiming hipEventDisableTiming
 #define cudaEventRecord hipEventRecord
@@ -119,11 +125,6 @@
 
 #endif // defined(GGML_USE_HIPBLAS)
 
-// ggml-cuda need half type so keep ggml headers include at last
-#include "ggml-cuda.h"
-#include "ggml.h"
-#include "ggml-backend-impl.h"
-
 #define CUDART_HMAX     11070 // CUDA 11.7, min. ver. for which __hmax and __hmax2 are known to work (may be higher than needed)
 
 #define CC_PASCAL     600
@@ -517,6 +518,24 @@ 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");
 
+#define QR1_S 8
+#define QI1_S (QK_K / (4*QR1_S))
+typedef struct {
+    half 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");
+
+#define QK4_NL 32
+#define QR4_NL 2
+#define QI4_NL (QK4_NL / (4*QR4_NL))
+typedef struct {
+    half 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");
+
 #define WARP_SIZE 32
 #define MATRIX_ROW_PADDING 512 // last row of quant. matrices is a multiple of this to avoid out-of-bounds memory accesses
 
@@ -642,18 +661,18 @@ static __device__ __forceinline__ float2 warp_reduce_sum(float2 a) {
     return a;
 }
 
-static __device__ __forceinline__ half2 warp_reduce_sum(half2 a) {
-#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL
-#pragma unroll
-    for (int mask = 16; mask > 0; mask >>= 1) {
-        a = __hadd2(a, __shfl_xor_sync(0xffffffff, a, mask, 32));
-    }
-    return a;
-#else
-    (void) a;
-    NO_DEVICE_CODE;
-#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL
-}
+//static __device__ __forceinline__ half2 warp_reduce_sum(half2 a) {
+//#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL
+//#pragma unroll
+//    for (int mask = 16; mask > 0; mask >>= 1) {
+//        a = __hadd2(a, __shfl_xor_sync(0xffffffff, a, mask, 32));
+//    }
+//    return a;
+//#else
+//    (void) a;
+//    NO_DEVICE_CODE;
+//#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL
+//}
 
 static __device__ __forceinline__ float warp_reduce_max(float x) {
 #pragma unroll
@@ -663,18 +682,18 @@ static __device__ __forceinline__ float warp_reduce_max(float x) {
     return x;
 }
 
-static __device__ __forceinline__ half2 warp_reduce_max(half2 x) {
-#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL && CUDART_VERSION >= CUDART_HMAX
-#pragma unroll
-    for (int mask = 16; mask > 0; mask >>= 1) {
-        x = __hmax2(x, __shfl_xor_sync(0xffffffff, x, mask, 32));
-    }
-    return x;
-#else
-    (void) x;
-    NO_DEVICE_CODE;
-#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL && CUDART_VERSION >= CUDART_HMAX
-}
+//static __device__ __forceinline__ half2 warp_reduce_max(half2 x) {
+//#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL && CUDART_VERSION >= CUDART_HMAX
+//#pragma unroll
+//    for (int mask = 16; mask > 0; mask >>= 1) {
+//        x = __hmax2(x, __shfl_xor_sync(0xffffffff, x, mask, 32));
+//    }
+//    return x;
+//#else
+//    (void) x;
+//    NO_DEVICE_CODE;
+//#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL && CUDART_VERSION >= CUDART_HMAX
+//}
 
 static __device__ __forceinline__ float op_repeat(const float a, const float b) {
     return b;
@@ -1681,6 +1700,137 @@ static const __device__ uint32_t iq3xxs_grid[256] = {
     0x3e1c1c1c, 0x3e1c3404, 0x3e24140c, 0x3e24240c, 0x3e2c0404, 0x3e2c0414, 0x3e2c1424, 0x3e341c04,
 };
 
+static const __device__ uint64_t iq1s_grid[512] = {
+    0xffffffffffff0101, 0xffffffffff01ff00, 0xffffffffff010100, 0xffffffff00000000,
+    0xffffffff01ff00ff, 0xffffffff01ff0001, 0xffffffff0101ffff, 0xffffffff0101ff01,
+    0xffffff00ff000000, 0xffffff000000ff00, 0xffffff00000000ff, 0xffffff0000000100,
+    0xffffff0000010000, 0xffffff0001000000, 0xffffff01ffff00ff, 0xffffff01ff01ff00,
+    0xffffff01ff010100, 0xffffff0100000001, 0xffffff0101ffff00, 0xffffff0101ff0101,
+    0xffffff0101010100, 0xffff00ffff00ff01, 0xffff00ffff0000ff, 0xffff00ff00ff0100,
+    0xffff00ff0100ff00, 0xffff00ff010001ff, 0xffff0000ff0101ff, 0xffff000000ffff00,
+    0xffff000000000000, 0xffff00000001ff01, 0xffff000001000101, 0xffff0000010100ff,
+    0xffff0001ffff0100, 0xffff00010000ff00, 0xffff000100010101, 0xffff000101000000,
+    0xffff01ffffff0000, 0xffff01ffff01ffff, 0xffff01ffff010100, 0xffff01ff00000000,
+    0xffff01ff01ffffff, 0xffff01ff01ff0001, 0xffff01ff0101ffff, 0xffff01ff01010001,
+    0xffff0100ffffff01, 0xffff01000000ffff, 0xffff010000000100, 0xffff010001ff01ff,
+    0xffff010001000000, 0xffff0101ff000000, 0xffff0101000101ff, 0xffff010101ffff01,
+    0xffff01010101ff00, 0xff00ffffff000000, 0xff00ffff00ffff00, 0xff00ffff00000001,
+    0xff00ffff000001ff, 0xff00ffff01010000, 0xff00ff00ffff0000, 0xff00ff00ff00ff00,
+    0xff00ff00ff0000ff, 0xff00ff00ff000100, 0xff00ff00ff010001, 0xff00ff0000ff0001,
+    0xff00ff000000ffff, 0xff00ff0000000000, 0xff00ff000001ff00, 0xff00ff0000010100,
+    0xff00ff0001ff0000, 0xff00ff000100ff00, 0xff00ff0001000100, 0xff00ff01ff000000,
+    0xff00ff0100ff0000, 0xff00ff01000001ff, 0xff00ff0101010001, 0xff0000ff00000000,
+    0xff0000ff0001ff00, 0xff0000ff00010100, 0xff000000ffff0101, 0xff000000ff000000,
+    0xff000000ff01ff00, 0xff00000000ff0000, 0xff0000000000ff00, 0xff000000000000ff,
+    0xff00000000000000, 0xff00000000000001, 0xff00000000000100, 0xff0000000001ffff,
+    0xff00000000010000, 0xff00000001000000, 0xff00000001010100, 0xff000001ff00ff01,
+    0xff000001ff0100ff, 0xff00000100000000, 0xff0000010001ff00, 0xff00000101ff0100,
+    0xff0000010100ff00, 0xff0001ff00ff00ff, 0xff0001ff00000101, 0xff0001ff000100ff,
+    0xff0001ff01000000, 0xff000100ff0001ff, 0xff0001000000ff01, 0xff00010000000000,
+    0xff00010000010001, 0xff00010000010100, 0xff00010001ffff00, 0xff00010001ff0101,
+    0xff00010001010000, 0xff000101ffffffff, 0xff000101ff000101, 0xff00010101ff00ff,
+    0xff00010101000001, 0xff000101010100ff, 0xff01ffffff000101, 0xff01ffffff01ffff,
+    0xff01ffffff01ff01, 0xff01ffffff0101ff, 0xff01ffff00000000, 0xff01ffff01ff0001,
+    0xff01ffff0101ff01, 0xff01ff00ff000000, 0xff01ff0000ff0100, 0xff01ff000000ff01,
+    0xff01ff0000010000, 0xff01ff00010000ff, 0xff01ff01ff01ff00, 0xff01ff0100000101,
+    0xff0100ffffff0000, 0xff0100ffff010000, 0xff0100ff01ff00ff, 0xff0100ff01000100,
+    0xff0100ff010100ff, 0xff010000ffffff01, 0xff01000000000000, 0xff0100000101ff00,
+    0xff010001ffff00ff, 0xff010001ff000100, 0xff01000100ffff00, 0xff01000100010001,
+    0xff01000101ff0001, 0xff010001010001ff, 0xff0101ffffffffff, 0xff0101ffff01ffff,
+    0xff0101ffff010101, 0xff0101ff0000ff00, 0xff0101ff01010001, 0xff010100ff000000,
+    0xff010100ff01ff01, 0xff01010000ff0001, 0xff01010000000100, 0xff01010001000000,
+    0xff0101010100ffff, 0x00ffffff0000ff01, 0x00ffffff000000ff, 0x00ffffff00000100,
+    0x00ffffff00010000, 0x00ffff00ffff0001, 0x00ffff00ff0000ff, 0x00ffff00ff000100,
+    0x00ffff0000000000, 0x00ffff0001000100, 0x00ffff0001010001, 0x00ffff01ff00ff01,
+    0x00ffff0100ff0100, 0x00ffff010000ff00, 0x00ffff01000100ff, 0x00ffff0101ff00ff,
+    0x00ffff010101ff00, 0x00ff00ffffffffff, 0x00ff00ffffff01ff, 0x00ff00ffff000101,
+    0x00ff00ff00000000, 0x00ff00ff000101ff, 0x00ff00ff01010101, 0x00ff0000ff000000,
+    0x00ff0000ff01ffff, 0x00ff000000ff0000, 0x00ff00000000ff00, 0x00ff0000000000ff,
+    0x00ff000000000000, 0x00ff000000000001, 0x00ff000000000100, 0x00ff000000010000,
+    0x00ff000001ffff01, 0x00ff000001000000, 0x00ff0001ff000101, 0x00ff000100ffffff,
+    0x00ff000100000000, 0x00ff0001010001ff, 0x00ff01ffff000000, 0x00ff01ff0001ff00,
+    0x00ff01ff01ff0100, 0x00ff0100ff01ff01, 0x00ff010000ff00ff, 0x00ff010000ff0101,
+    0x00ff010000000000, 0x00ff010000010101, 0x00ff01000100ff00, 0x00ff010001010000,
+    0x00ff0101ffffff00, 0x00ff01010000ff01, 0x00ff010100000100, 0x00ff010101ff0000,
+    0x0000ffffffff0100, 0x0000ffffff00ff00, 0x0000ffffff0000ff, 0x0000ffffff010000,
+    0x0000ffff00000000, 0x0000ffff00010101, 0x0000ffff01ffff01, 0x0000ffff01000100,
+    0x0000ff00ff000000, 0x0000ff00ff01ff00, 0x0000ff00ff0101ff, 0x0000ff0000ff0000,
+    0x0000ff000000ff00, 0x0000ff00000000ff, 0x0000ff0000000000, 0x0000ff0000000001,
+    0x0000ff0000000100, 0x0000ff0000010000, 0x0000ff0001ffffff, 0x0000ff0001ff01ff,
+    0x0000ff0001000000, 0x0000ff000101ffff, 0x0000ff01ffff0101, 0x0000ff01ff010000,
+    0x0000ff0100000000, 0x0000ff0101000101, 0x000000ffffff0001, 0x000000ffff000000,
+    0x000000ff00ff0000, 0x000000ff0000ff00, 0x000000ff000000ff, 0x000000ff00000000,
+    0x000000ff00000001, 0x000000ff00000100, 0x000000ff00010000, 0x000000ff01000000,
+    0x000000ff0101ff00, 0x00000000ffff0000, 0x00000000ff00ff00, 0x00000000ff0000ff,
+    0x00000000ff000000, 0x00000000ff000001, 0x00000000ff000100, 0x00000000ff010000,
+    0x0000000000ffff00, 0x0000000000ff00ff, 0x0000000000ff0000, 0x0000000000ff0001,
+    0x0000000000ff0100, 0x000000000000ffff, 0x000000000000ff00, 0x000000000000ff01,
+    0x00000000000000ff, 0x0000000000000001, 0x00000000000001ff, 0x0000000000000100,
+    0x0000000000000101, 0x000000000001ff00, 0x00000000000100ff, 0x0000000000010000,
+    0x0000000000010001, 0x0000000000010100, 0x0000000001ff0000, 0x000000000100ff00,
+    0x00000000010000ff, 0x0000000001000000, 0x0000000001000001, 0x0000000001000100,
+    0x0000000001010000, 0x00000001ffff01ff, 0x00000001ff000000, 0x0000000100ff0000,
+    0x000000010000ff00, 0x00000001000000ff, 0x0000000100000000, 0x0000000100000001,
+    0x0000000100000100, 0x0000000100010000, 0x0000000101000000, 0x000001ffff00ff00,
+    0x000001ffff010001, 0x000001ffff0101ff, 0x000001ff00ffff01, 0x000001ff0000ffff,
+    0x000001ff00000000, 0x000001ff010000ff, 0x000001ff01010100, 0x00000100ffff0100,
+    0x00000100ff000000, 0x0000010000ff0000, 0x000001000000ff00, 0x00000100000000ff,
+    0x0000010000000000, 0x0000010000000001, 0x0000010000000100, 0x0000010000010000,
+    0x0000010001000000, 0x000001000101ff01, 0x00000101ffff0001, 0x00000101ff01ffff,
+    0x0000010100000000, 0x0000010101010100, 0x0001ffffff000000, 0x0001ffff00ffffff,
+    0x0001ffff00000100, 0x0001ffff0001ff00, 0x0001ffff01000000, 0x0001ff00ffffff00,
+    0x0001ff00ffff01ff, 0x0001ff00ff010000, 0x0001ff0000000000, 0x0001ff0000010001,
+    0x0001ff0001ff0000, 0x0001ff0001010100, 0x0001ff01ff0000ff, 0x0001ff01ff000001,
+    0x0001ff0100ffffff, 0x0001ff010001ffff, 0x0001ff01000101ff, 0x0001ff010100ff01,
+    0x000100ffff00ffff, 0x000100ffff00ff01, 0x000100ffff000100, 0x000100ff00000000,
+    0x000100ff000101ff, 0x000100ff01ff0101, 0x000100ff0100ffff, 0x000100ff01010101,
+    0x00010000ff000000, 0x00010000ff010100, 0x0001000000ff0000, 0x000100000000ff00,
+    0x00010000000000ff, 0x0001000000000000, 0x0001000000000001, 0x0001000000000100,
+    0x0001000000010000, 0x0001000001ffff01, 0x0001000001000000, 0x0001000100ff0101,
+    0x0001000100000000, 0x00010001010100ff, 0x000101ffffff01ff, 0x000101ffffff0101,
+    0x000101ff00010000, 0x000101ff01ff0000, 0x000101ff0100ff01, 0x00010100ffff0000,
+    0x0001010000000000, 0x000101000001ffff, 0x0001010000010101, 0x00010100010001ff,
+    0x00010101ff00ff00, 0x00010101ff010001, 0x0001010100ffffff, 0x0001010100ff01ff,
+    0x00010101000101ff, 0x0001010101ff0000, 0x000101010100ff01, 0x0001010101000101,
+    0x01ffffffffff0101, 0x01ffffffff01ffff, 0x01ffffffff01ff01, 0x01ffffffff0101ff,
+    0x01ffffffff010101, 0x01ffffff00000000, 0x01ffffff01ff01ff, 0x01ffffff01000101,
+    0x01ffffff0101ff01, 0x01ffffff010100ff, 0x01ffff000000ff00, 0x01ffff0000000001,
+    0x01ffff00000001ff, 0x01ffff0000010000, 0x01ffff0001ff0000, 0x01ffff01ffffffff,
+    0x01ffff01ffff01ff, 0x01ffff01ff000000, 0x01ffff01ff01ffff, 0x01ffff01ff0101ff,
+    0x01ffff010100ffff, 0x01ff00ffffff0000, 0x01ff00ffff010000, 0x01ff00ff00ffff01,
+    0x01ff0000ff0000ff, 0x01ff000000000000, 0x01ff00000001ff01, 0x01ff000001ffffff,
+    0x01ff000001010100, 0x01ff0001ffffff01, 0x01ff0001ff010001, 0x01ff000101ff0100,
+    0x01ff000101000001, 0x01ff0001010100ff, 0x01ff01ffff00ffff, 0x01ff01ff00010001,
+    0x01ff01ff01000000, 0x01ff01ff010101ff, 0x01ff0100ff000001, 0x01ff010000ffff00,
+    0x01ff010000000100, 0x01ff010001ff01ff, 0x01ff01000101ffff, 0x01ff0101ffff00ff,
+    0x01ff0101ffff0101, 0x01ff0101ff0101ff, 0x01ff010100010000, 0x0100ffff00ff00ff,
+    0x0100ffff00ff0001, 0x0100ffff00000100, 0x0100ffff0100ff00, 0x0100ff00ffff0000,
+    0x0100ff00ff00ffff, 0x0100ff00ff00ff01, 0x0100ff00ff000100, 0x0100ff00ff010000,
+    0x0100ff0000000000, 0x0100ff00000100ff, 0x0100ff0001ff0101, 0x0100ff0001010101,
+    0x0100ff0100ff00ff, 0x0100ff0100ff0001, 0x0100ff0100000100, 0x0100ff0100010001,
+    0x0100ff0101000000, 0x010000ffff00ff00, 0x010000ff0000ffff, 0x010000ff00000000,
+    0x010000ff010001ff, 0x010000ff01010001, 0x01000000ffffff00, 0x01000000ffff0101,
+    0x01000000ff000000, 0x01000000ff0100ff, 0x01000000ff010101, 0x0100000000ff0000,
+    0x010000000000ff00, 0x01000000000000ff, 0x0100000000000000, 0x0100000000000001,
+    0x0100000000000100, 0x0100000000010000, 0x0100000001000000, 0x0100000100000000,
+    0x01000001000101ff, 0x0100000101ffff01, 0x010001ffff000101, 0x010001ff00ff0100,
+    0x010001ff0000ff00, 0x010001ff000100ff, 0x010001ff01ffffff, 0x01000100ffff0000,
+    0x01000100ff0001ff, 0x0100010000000000, 0x010001000001ff00, 0x0100010001ff0000,
+    0x01000100010000ff, 0x0100010001000101, 0x01000101ff00ff01, 0x0100010100ff0100,
+    0x010001010000ffff, 0x0100010101010001, 0x0101ffffffff0101, 0x0101ffffff0001ff,
+    0x0101ffffff01ffff, 0x0101ffffff010101, 0x0101ffff00000000, 0x0101ffff0101ffff,
+    0x0101ffff010101ff, 0x0101ff00ff000000, 0x0101ff0000ff0100, 0x0101ff000000ff00,
+    0x0101ff0000010000, 0x0101ff00010000ff, 0x0101ff0001000001, 0x0101ff01ff010101,
+    0x0101ff0100000000, 0x0101ff010101ff00, 0x010100ffffff0000, 0x010100ffff010000,
+    0x010100ff00ff01ff, 0x010100ff000000ff, 0x010100ff00000101, 0x010100ff01ffff00,
+    0x01010000ffffff01, 0x01010000ff000100, 0x01010000ff01ff01, 0x0101000000000000,
+    0x01010000000100ff, 0x010100000101ff01, 0x01010001ffff0000, 0x01010001ff00ffff,
+    0x01010001ff010000, 0x0101000101ffffff, 0x0101000101ff01ff, 0x0101000101010101,
+    0x010101ffff01ffff, 0x010101ff00000000, 0x010101ff0001ff01, 0x010101ff0101ffff,
+    0x010101ff010101ff, 0x01010100ffffffff, 0x01010100ff000001, 0x010101000000ff00,
+    0x0101010001010000, 0x0101010100ff0001, 0x010101010001ff01, 0x010101010101ffff,
+};
+
 static const __device__ uint8_t ksigns_iq2xs[128] = {
       0, 129, 130,   3, 132,   5,   6, 135, 136,   9,  10, 139,  12, 141, 142,  15,
     144,  17,  18, 147,  20, 149, 150,  23,  24, 153, 154,  27, 156,  29,  30, 159,
@@ -1823,6 +1973,49 @@ static __global__ void dequantize_block_iq3_xxs(const void * __restrict__ vx, ds
 
 }
 
+template<typename dst_t>
+static __global__ void dequantize_block_iq1_s(const void * __restrict__ vx, dst_t * __restrict__ yy) {
+
+    const int i   = blockIdx.x;
+    const block_iq1_s * x = (const block_iq1_s  *) vx;
+
+    const int tid = threadIdx.x;
+#if QK_K == 256
+    const int il = tid/8; // 0...3
+    const int ib = tid%8; // 0...7
+    dst_t * y = yy + i*QK_K + 32*ib + 8*il;
+    const int i8 = 4*ib+il;
+    uint8_t h = x[i].scales[i8/2] >> 4*(i8%2);
+    const int8_t * grid = (const int8_t *)(iq1s_grid + (x[i].qs[i8] | ((h & 8) << 5)));
+    const float d = (float)x[i].d * (2*(h & 7) + 1);
+    for (int j = 0; j < 8; ++j) y[j] = d * grid[j];
+#else
+    assert(false);
+#endif
+
+}
+
+static const __device__ int8_t kvalues_iq4nl[16] = {-127, -104, -83, -65, -49, -35, -22, -10, 1, 13, 25, 38, 53, 69, 89, 113};
+
+template<typename dst_t>
+static __global__ void dequantize_block_iq4_nl(const void * __restrict__ vx, dst_t * __restrict__ yy) {
+
+    const int i   = blockIdx.x;
+    const block_iq4_nl * x = (const block_iq4_nl *) vx + i*(QK_K/QK4_NL);
+
+    const int tid = threadIdx.x;
+    const int il = tid/8; // 0...3
+    const int ib = tid%8; // 0...7
+    dst_t * y = yy + i*QK_K + 32*ib + 4*il;
+    const uint8_t  * q4 = x[ib].qs + 4*il;
+    const float d = (float)x[ib].d;
+    for (int j = 0; j < 4; ++j) {
+        y[j+ 0] = d * kvalues_iq4nl[q4[j] & 0xf];
+        y[j+16] = d * kvalues_iq4nl[q4[j] >>  4];
+    }
+
+}
+
 static __global__ void dequantize_mul_mat_vec_q2_k(const void * __restrict__ vx, const float * __restrict__ yy, float * __restrict__ dst, const int ncols, int nrows) {
 
     static_assert(16%K_QUANTS_PER_ITERATION == 0, "16 must be divisible by K_QUANTS_PER_ITERATION");
@@ -4478,10 +4671,12 @@ static __device__ __forceinline__ float vec_dot_iq2_xs_q8_1(
     const float d = (float)bq2->d * __low2float(bq8_1[ib32].ds) * 0.25f;
     return d * ((0.5f + ls1) * sumi1 + (0.5f + ls2) * sumi2);
 #else
+    (void) ksigns64;
     assert(false);
     return 0.f;
 #endif
 #else
+    (void) ksigns64;
     assert(false);
     return 0.f;
 #endif
@@ -4522,6 +4717,99 @@ static __device__ __forceinline__ float vec_dot_iq3_xxs_q8_1(
 #endif
 }
 
+static __device__ __forceinline__ float vec_dot_iq1_s_q8_1(
+    const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
+#if QK_K == 256
+    const block_iq1_s * bq1 = (const block_iq1_s *) vbq;
+
+    const int ib32 = iqs;
+    int sumi1 = 0, sumi2 = 0, sumi3 = 0, sumi4 = 0;
+    const uint8_t h1 = bq1->scales[2*ib32+0];
+    const uint8_t h2 = bq1->scales[2*ib32+1];
+#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
+    const int * q8 = (const int *)bq8_1[ib32].qs;
+    const int * grid1 = (const int *)(iq1s_grid + (bq1->qs[4*ib32+0] | ((h1 & 0x08) << 5)));
+    const int * grid2 = (const int *)(iq1s_grid + (bq1->qs[4*ib32+1] | ((h1 & 0x80) << 1)));
+    const int * grid3 = (const int *)(iq1s_grid + (bq1->qs[4*ib32+2] | ((h2 & 0x08) << 5)));
+    const int * grid4 = (const int *)(iq1s_grid + (bq1->qs[4*ib32+3] | ((h2 & 0x80) << 1)));
+    for (int j = 0; j < 2; ++j) {
+        sumi1 = __dp4a(q8[j+0], grid1[j], sumi1);
+        sumi2 = __dp4a(q8[j+2], grid2[j], sumi2);
+        sumi3 = __dp4a(q8[j+4], grid3[j], sumi3);
+        sumi4 = __dp4a(q8[j+6], grid4[j], sumi4);
+    }
+#else
+    const int8_t   * q8 = bq8_1[ib32].qs;
+    const int8_t * grid1 = (const int8_t *)(iq1s_grid + (bq1->qs[4*ib32+0] | ((h1 & 0x08) << 5)));
+    const int8_t * grid2 = (const int8_t *)(iq1s_grid + (bq1->qs[4*ib32+1] | ((h1 & 0x80) << 1)));
+    const int8_t * grid3 = (const int8_t *)(iq1s_grid + (bq1->qs[4*ib32+2] | ((h2 & 0x08) << 5)));
+    const int8_t * grid4 = (const int8_t *)(iq1s_grid + (bq1->qs[4*ib32+3] | ((h2 & 0x80) << 1)));
+    for (int j = 0; j < 8; ++j) {
+        sumi1 += q8[j+ 0] * grid1[j];
+        sumi2 += q8[j+ 8] * grid2[j];
+        sumi3 += q8[j+16] * grid3[j];
+        sumi4 += q8[j+24] * grid4[j];
+    }
+#endif
+    const float d = (float)bq1->d * __low2float(bq8_1[ib32].ds);
+    return d * (sumi1 * (2*(h1 & 7) + 1) + sumi2 * (2*((h1 >> 4) & 7) + 1) +
+                sumi3 * (2*(h2 & 7) + 1) + sumi4 * (2*((h2 >> 4) & 7) + 1));
+#else
+    assert(false);
+    return 0.f;
+#endif
+}
+
+#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
+static __device__ __forceinline__ void get_int_from_table_16(const uint32_t & q4, const uint8_t * values,
+        int & val1, int & val2) {
+
+    uint32_t aux32; const uint8_t * q8 = (const uint8_t *)&aux32;
+    aux32 = q4 & 0x0f0f0f0f;
+    uint16_t v1 = values[q8[0]] | (values[q8[1]] << 8);
+    uint16_t v2 = values[q8[2]] | (values[q8[3]] << 8);
+    val1 = v1 | (v2 << 16);
+    aux32 = (q4 >> 4) & 0x0f0f0f0f;
+    v1 = values[q8[0]] | (values[q8[1]] << 8);
+    v2 = values[q8[2]] | (values[q8[3]] << 8);
+    val2 = v1 | (v2 << 16);
+}
+#endif
+
+static __device__ __forceinline__ float vec_dot_iq4_nl_q8_1(
+    const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
+
+    const block_iq4_nl * bq = (const block_iq4_nl *) vbq;
+
+#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
+    const uint16_t * q4 = (const uint16_t *)bq->qs + 2*iqs;
+    const int32_t  * q8 = (const int32_t  *)bq8_1->qs + iqs;
+
+    const uint8_t * values = (const uint8_t *)kvalues_iq4nl;
+
+    int v1, v2;
+    int sumi1 = 0, sumi2 = 0;
+    for (int l = 0; l < VDR_Q4_0_Q8_1_MMVQ; ++l) {
+        const uint32_t aux = q4[2*l] | (q4[2*l+1] << 16);
+        get_int_from_table_16(aux, values, v1, v2);
+        sumi1 = __dp4a(v1, q8[l+0], sumi1);
+        sumi2 = __dp4a(v2, q8[l+4], sumi2);
+    }
+
+#else
+    const uint8_t * q4 = bq->qs + 4*iqs;
+    const int8_t  * q8 = bq8_1->qs + 4*iqs;
+
+    int sumi1 = 0, sumi2 = 0;
+    for (int l = 0; l < 4*VDR_Q4_0_Q8_1_MMVQ; ++l) {
+        sumi1 += q8[l+ 0] * kvalues_iq4nl[q4[l] & 0xf];
+        sumi2 += q8[l+16] * kvalues_iq4nl[q4[l] >>  4];
+    }
+#endif
+    const float d = (float)bq->d * __low2float(bq8_1->ds);
+    return d * (sumi1 + sumi2);
+}
+
 template <int qk, int qr, int qi, bool need_sum, typename block_q_t, int mmq_x, int mmq_y, int nwarps,
               allocate_tiles_cuda_t allocate_tiles, load_tiles_cuda_t load_tiles, int vdr, vec_dot_q_mul_mat_cuda_t vec_dot>
 static __device__ __forceinline__ void mul_mat_q(
@@ -5956,149 +6244,31 @@ static __global__ void diag_mask_inf_f32(const float * x, float * dst, const int
     dst[i] = x[i] - (col > n_past + row % rows_per_channel) * FLT_MAX;
 }
 
-template <bool vals_smem, int ncols_template, int block_size_template, bool need_check>
-static __global__ void soft_max_f16(const float * x, const float * y, float * dst, const int ncols_par, const int nrows_y, const float scale) {
-#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL && CUDART_VERSION >= CUDART_HMAX
-    const int ncols_data = ncols_template == 0 ? ncols_par : ncols_template;
-    const int ncols_smem = GGML_PAD(ncols_data, 2*WARP_SIZE)/2;
-
-    const int tid  = threadIdx.x;
-    const int rowx = blockIdx.x;
-    const int rowy = rowx % nrows_y; // broadcast the mask (y) in the row dimension
-
-    const int block_size = block_size_template == 0 ? blockDim.x : block_size_template;
-
-    const int warp_id = threadIdx.x / WARP_SIZE;
-    const int lane_id = threadIdx.x % WARP_SIZE;
-
-    extern __shared__ half data_soft_max_f16[];
-    half  * buf_iw = data_soft_max_f16 + 0; // shared memory buffer for inter-warp communication
-    // (shared memory) buffer to cache values between iterations:
-    half2 * vals = vals_smem ? (half2 *) (buf_iw + WARP_SIZE) : (half2 *) (dst + rowx*ncols_data);
-    // if the buffer is larger than max. shared memory per block, use dst as temp. buffer instead
-    // in that case col_smem == col_data must be enforced to avoid race conditions
-
-    half2 max_val = make_half2(-INFINITY, -INFINITY);
-
-#pragma unroll
-    for (int col0 = 0; col0 < ncols_smem; col0 += block_size) {
-        const int col_data = 2*col0 + 2*WARP_SIZE*warp_id + lane_id;
-        const int col_smem = vals_smem ? col0 + tid : col_data;
-
-        const int ix = rowx*ncols_data + col_data;
-        const int iy = rowy*ncols_data + col_data;
-
-        half2 val;
-        if (need_check && col_data + 0 >= ncols_data) {
-            val.x = -INFINITY;
-        } else {
-            val.x = x[ix + 0]*scale + (y ? y[iy + 0] : 0.0f);
-        }
-        if (need_check && col_data + WARP_SIZE >= ncols_data) {
-            val.y = -INFINITY;
-        } else {
-            val.y = x[ix + WARP_SIZE]*scale + (y ? y[iy + WARP_SIZE] : 0.0f);
-        }
-        if (!need_check || col_smem < (vals_smem ? ncols_smem : ncols_data)) {
-            vals[col_smem] = val;
-        }
-        max_val = __hmax2(max_val, val);
-    }
-
-    // find the max value in the block
-    max_val = warp_reduce_max(max_val);
-    if (block_size > WARP_SIZE) {
-        if (warp_id == 0) {
-            buf_iw[lane_id] = -INFINITY;
-        }
-        __syncthreads();
-
-        if (lane_id == 0) {
-            buf_iw[warp_id] = __hmax(max_val.x, max_val.y);
-        }
-        __syncthreads();
-
-        max_val = __half2half2(buf_iw[lane_id]);
-        max_val = warp_reduce_max(max_val);
-    } else {
-        max_val = __half2half2(__hmax(max_val.x, max_val.y));
-    }
-
-    half2 tmp = make_half2(0.0f, 0.0f); // partial sums
-
-#pragma unroll
-    for (int col0 = 0; col0 < ncols_smem; col0 += block_size) {
-        const int col_smem = vals_smem ? col0 + tid : 2*col0 + 2*warp_id*WARP_SIZE + lane_id;
-
-        if (ncols_template == 0 && col_smem >= (vals_smem ? ncols_smem : ncols_data)) {
-            break;
-        }
-
-        const half2 val = h2exp(vals[col_smem] - max_val);
-
-        tmp += val;
-        vals[col_smem] = val;
-    }
-
-    // find the sum of exps in the block
-    tmp = warp_reduce_sum(tmp);
-    if (block_size > WARP_SIZE) {
-        if (warp_id == 0) {
-            buf_iw[lane_id] = 0.0f;
-        }
-        __syncthreads();
-
-        if (lane_id == 0) {
-            buf_iw[warp_id] = tmp.x + tmp.y;
-        }
-        __syncthreads();
-
-        tmp = __half2half2(buf_iw[lane_id]);
-        tmp = warp_reduce_sum(tmp);
-    } else {
-        tmp = __half2half2(tmp.x + tmp.y);
-    }
-
-    const half2 inv_sum = make_half2(1.0f, 1.0f) / tmp;
-
-#pragma unroll
-    for (int col0 = 0; col0 < ncols_smem; col0 += block_size) {
-        const int col_data = 2*col0 + 2*WARP_SIZE*warp_id + lane_id;
-        const int col_smem = vals_smem ? col0 + tid : col_data;
-
-        const int idst = rowx*ncols_data + col_data;
-        const half2 result = vals[col_smem] * inv_sum;
-
-        if (need_check && col_data + 0 >= ncols_data) {
-            return;
-        }
-        dst[idst] = result.x;
-
-        if (need_check && col_data + WARP_SIZE >= ncols_data) {
-            return;
-        }
-
-        dst[idst + WARP_SIZE] = result.y;
-    }
-#else
-    (void) x; (void) y; (void) dst; (void) ncols_par; (void) nrows_y; (void) scale;
-    NO_DEVICE_CODE;
-#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL && CUDART_VERSION >= CUDART_HMAX
-}
-
 template <bool vals_smem, int ncols_template, int block_size_template>
-static __global__ void soft_max_f32(const float * x, const float * y, float * dst, const int ncols_par, const int nrows_y, const float scale) {
+static __global__ void soft_max_f32(const float * x, const float * mask, const float * pos, float * dst, const int ncols_par, const int nrows_y, const float scale, const float max_bias, const float m0, const float m1, uint32_t n_head_log2) {
     const int ncols = ncols_template == 0 ? ncols_par : ncols_template;
 
     const int tid  = threadIdx.x;
     const int rowx = blockIdx.x;
-    const int rowy = rowx % nrows_y; // broadcast the mask (y) in the row dimension
+    const int rowy = rowx % nrows_y; // broadcast the mask in the row dimension
 
     const int block_size = block_size_template == 0 ? blockDim.x : block_size_template;
 
     const int warp_id = threadIdx.x / WARP_SIZE;
     const int lane_id = threadIdx.x % WARP_SIZE;
 
+    float slope = 0.0f;
+
+    // ALiBi
+    if (max_bias > 0.0f) {
+        const int h = rowx/nrows_y; // head index
+
+        const float base = h < n_head_log2 ? m0 : m1;
+        const int   exp  = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1;
+
+        slope = powf(base, exp);
+    }
+
     extern __shared__ float data_soft_max_f32[];
     float * buf_iw = data_soft_max_f32; // shared memory buffer for inter-warp communication
     // shared memory buffer to cache values between iterations:
@@ -6117,7 +6287,8 @@ static __global__ void soft_max_f32(const float * x, const float * y, float * ds
         const int ix = rowx*ncols + col;
         const int iy = rowy*ncols + col;
 
-        const float val = x[ix]*scale + (y ? y[iy] : 0.0f);
+        const float val = x[ix]*scale + (mask ? mask[iy] : 0.0f) + (pos ? slope*pos[col] : 0.0f);
+
         vals[col] = val;
         max_val = max(max_val, val);
     }
@@ -6678,6 +6849,18 @@ static void dequantize_row_iq3_xxs_cuda(const void * vx, dst_t * y, const int k,
     dequantize_block_iq3_xxs<<<nb, 32, 0, stream>>>(vx, y);
 }
 
+template<typename dst_t>
+static void dequantize_row_iq1_s_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
+    const int nb = k / QK_K;
+    dequantize_block_iq1_s<<<nb, 32, 0, stream>>>(vx, y);
+}
+
+template<typename dst_t>
+static void dequantize_row_iq4_nl_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
+    const int nb = (k + QK_K - 1) / QK_K;
+    dequantize_block_iq4_nl<<<nb, 32, 0, stream>>>(vx, y);
+}
+
 template <typename src_t, typename dst_t>
 static void convert_unary_cuda(const void * __restrict__ vx, dst_t * __restrict__ y, const int k, cudaStream_t stream) {
     const int num_blocks = (k + CUDA_DEQUANTIZE_BLOCK_SIZE - 1) / CUDA_DEQUANTIZE_BLOCK_SIZE;
@@ -6717,6 +6900,10 @@ static to_fp16_cuda_t ggml_get_to_fp16_cuda(ggml_type type) {
             return dequantize_row_iq2_xs_cuda;
         case GGML_TYPE_IQ3_XXS:
             return dequantize_row_iq3_xxs_cuda;
+        case GGML_TYPE_IQ1_S:
+            return dequantize_row_iq1_s_cuda;
+        case GGML_TYPE_IQ4_NL:
+            return dequantize_row_iq4_nl_cuda;
         case GGML_TYPE_F32:
             return convert_unary_cuda<float>;
         default:
@@ -6752,6 +6939,10 @@ static to_fp32_cuda_t ggml_get_to_fp32_cuda(ggml_type type) {
             return dequantize_row_iq2_xs_cuda;
         case GGML_TYPE_IQ3_XXS:
             return dequantize_row_iq3_xxs_cuda;
+        case GGML_TYPE_IQ1_S:
+            return dequantize_row_iq1_s_cuda;
+        case GGML_TYPE_IQ4_NL:
+            return dequantize_row_iq4_nl_cuda;
         case GGML_TYPE_F16:
             return convert_unary_cuda<half>;
         default:
@@ -7589,89 +7780,53 @@ static void diag_mask_inf_f32_cuda(const float * x, float * dst, const int ncols
     diag_mask_inf_f32<<<block_nums, block_dims, 0, stream>>>(x, dst, ncols_x, rows_per_channel, n_past);
 }
 
-static void soft_max_f16_cuda(const float * x, const float * y, float * dst, const int ncols_x, const int nrows_x, const int nrows_y, const float scale, cudaStream_t stream) {
-    int nth = WARP_SIZE;
-    while (nth < ncols_x/2 && nth < CUDA_SOFT_MAX_BLOCK_SIZE) nth *= 2;
-    const dim3 block_dims(nth,     1, 1);
-    const dim3 block_nums(nrows_x, 1, 1);
-    const size_t shmem = (GGML_PAD(ncols_x, 2*WARP_SIZE) + WARP_SIZE)*sizeof(half);
-    static_assert(CUDA_SOFT_MAX_BLOCK_SIZE == 1024, "These values need to be adjusted.");
-    if (shmem <= g_device_caps[g_main_device].smpb) {
-        switch (ncols_x) {
-            case 32:
-                soft_max_f16<true, 32, 32, true><<<block_nums, block_dims, shmem, stream>>>(x, y, dst, ncols_x, nrows_y, scale);
-                break;
-            case 64:
-                soft_max_f16<true, 64, 32, false><<<block_nums, block_dims, shmem, stream>>>(x, y, dst, ncols_x, nrows_y, scale);
-                break;
-            case 128:
-                soft_max_f16<true, 128, 64, false><<<block_nums, block_dims, shmem, stream>>>(x, y, dst, ncols_x, nrows_y, scale);
-                break;
-            case 256:
-                soft_max_f16<true, 256, 128, false><<<block_nums, block_dims, shmem, stream>>>(x, y, dst, ncols_x, nrows_y, scale);
-                break;
-            case 512:
-                soft_max_f16<true, 512, 256, false><<<block_nums, block_dims, shmem, stream>>>(x, y, dst, ncols_x, nrows_y, scale);
-                break;
-            case 1024:
-                soft_max_f16<true, 1024, 512, false><<<block_nums, block_dims, shmem, stream>>>(x, y, dst, ncols_x, nrows_y, scale);
-                break;
-            case 2048:
-                soft_max_f16<true, 2048, 1024, false><<<block_nums, block_dims, shmem, stream>>>(x, y, dst, ncols_x, nrows_y, scale);
-                break;
-            case 4096:
-                soft_max_f16<true, 4096, 1024, false><<<block_nums, block_dims, shmem, stream>>>(x, y, dst, ncols_x, nrows_y, scale);
-                break;
-            default:
-                soft_max_f16<true, 0, 0, true><<<block_nums, block_dims, shmem, stream>>>(x, y, dst, ncols_x, nrows_y, scale);
-                break;
-        }
-    } else {
-        const size_t shmem_low = WARP_SIZE*sizeof(half);
-        soft_max_f16<false, 0, 0, true><<<block_nums, block_dims, shmem_low, stream>>>(x, y, dst, ncols_x, nrows_y, scale);
-    }
-}
-
-static void soft_max_f32_cuda(const float * x, const float * y, float * dst, const int ncols_x, const int nrows_x, const int nrows_y, const float scale, cudaStream_t stream) {
+static void soft_max_f32_cuda(const float * x, const float * mask, const float * pos, float * dst, const int ncols_x, const int nrows_x, const int nrows_y, const float scale, const float max_bias, cudaStream_t stream) {
     int nth = WARP_SIZE;
     while (nth < ncols_x && nth < CUDA_SOFT_MAX_BLOCK_SIZE) nth *= 2;
     const dim3 block_dims(nth,     1, 1);
     const dim3 block_nums(nrows_x, 1, 1);
     const size_t shmem = (GGML_PAD(ncols_x, WARP_SIZE) + WARP_SIZE)*sizeof(float);
     static_assert(CUDA_SOFT_MAX_BLOCK_SIZE == 1024, "These values need to be adjusted.");
+
+    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);
+
     if (shmem < g_device_caps[g_main_device].smpb) {
         switch (ncols_x) {
             case 32:
-                soft_max_f32<true, 32, 32><<<block_nums, block_dims, shmem, stream>>>(x, y, dst, ncols_x, nrows_y, scale);
+                soft_max_f32<true, 32, 32><<<block_nums, block_dims, shmem, stream>>>(x, mask, pos, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
                 break;
             case 64:
-                soft_max_f32<true, 64, 64><<<block_nums, block_dims, shmem, stream>>>(x, y, dst, ncols_x, nrows_y, scale);
+                soft_max_f32<true, 64, 64><<<block_nums, block_dims, shmem, stream>>>(x, mask, pos, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
                 break;
             case 128:
-                soft_max_f32<true, 128, 128><<<block_nums, block_dims, shmem, stream>>>(x, y, dst, ncols_x, nrows_y, scale);
+                soft_max_f32<true, 128, 128><<<block_nums, block_dims, shmem, stream>>>(x, mask, pos, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
                 break;
             case 256:
-                soft_max_f32<true, 256, 256><<<block_nums, block_dims, shmem, stream>>>(x, y, dst, ncols_x, nrows_y, scale);
+                soft_max_f32<true, 256, 256><<<block_nums, block_dims, shmem, stream>>>(x, mask, pos, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
                 break;
             case 512:
-                soft_max_f32<true, 512, 512><<<block_nums, block_dims, shmem, stream>>>(x, y, dst, ncols_x, nrows_y, scale);
+                soft_max_f32<true, 512, 512><<<block_nums, block_dims, shmem, stream>>>(x, mask, pos, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
                 break;
             case 1024:
-                soft_max_f32<true, 1024, 1024><<<block_nums, block_dims, shmem, stream>>>(x, y, dst, ncols_x, nrows_y, scale);
+                soft_max_f32<true, 1024, 1024><<<block_nums, block_dims, shmem, stream>>>(x, mask, pos, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
                 break;
             case 2048:
-                soft_max_f32<true, 2048, 1024><<<block_nums, block_dims, shmem, stream>>>(x, y, dst, ncols_x, nrows_y, scale);
+                soft_max_f32<true, 2048, 1024><<<block_nums, block_dims, shmem, stream>>>(x, mask, pos, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
                 break;
             case 4096:
-                soft_max_f32<true, 4096, 1024><<<block_nums, block_dims, shmem, stream>>>(x, y, dst, ncols_x, nrows_y, scale);
+                soft_max_f32<true, 4096, 1024><<<block_nums, block_dims, shmem, stream>>>(x, mask, pos, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
                 break;
             default:
-                soft_max_f32<true, 0, 0><<<block_nums, block_dims, shmem, stream>>>(x, y, dst, ncols_x, nrows_y, scale);
+                soft_max_f32<true, 0, 0><<<block_nums, block_dims, shmem, stream>>>(x, mask, pos, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
                 break;
         }
     } else {
         const size_t shmem_low = WARP_SIZE*sizeof(float);
-        soft_max_f32<false, 0, 0><<<block_nums, block_dims, shmem_low, stream>>>(x, y, dst, ncols_x, nrows_y, scale);
+        soft_max_f32<false, 0, 0><<<block_nums, block_dims, shmem_low, stream>>>(x, mask, pos, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
     }
 }
 
@@ -7943,6 +8098,7 @@ GGML_CALL void ggml_init_cublas() {
         if (cudaGetDeviceCount(&g_device_count) != cudaSuccess) {
             initialized = true;
             g_cublas_loaded = false;
+            fprintf(stderr, "%s: no " GGML_CUDA_NAME " devices found, " GGML_CUDA_NAME " will be disabled\n", __func__);
             return;
         }
 
@@ -8530,6 +8686,8 @@ static int64_t get_row_rounding(ggml_type type, const std::array<float, GGML_CUD
         case GGML_TYPE_IQ2_XXS:
         case GGML_TYPE_IQ2_XS:
         case GGML_TYPE_IQ3_XXS:
+        case GGML_TYPE_IQ1_S:
+        case GGML_TYPE_IQ4_NL:
             return max_compute_capability >= CC_RDNA2 ? 128 : 64;
         default:
             GGML_ASSERT(false);
@@ -8553,6 +8711,8 @@ static int64_t get_row_rounding(ggml_type type, const std::array<float, GGML_CUD
         case GGML_TYPE_IQ2_XXS:
         case GGML_TYPE_IQ2_XS:
         case GGML_TYPE_IQ3_XXS:
+        case GGML_TYPE_IQ1_S:
+        case GGML_TYPE_IQ4_NL:
             return max_compute_capability >= CC_VOLTA ? 128 : 64;
         case GGML_TYPE_Q6_K:
             return 64;
@@ -8650,6 +8810,14 @@ static void ggml_cuda_op_mul_mat_vec_q(
             mul_mat_vec_q_cuda<QK_K, QI3_XXS, block_iq3_xxs, 1, vec_dot_iq3_xxs_q8_1>
                 (src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
             break;
+        case GGML_TYPE_IQ1_S:
+            mul_mat_vec_q_cuda<QK_K, QI1_S, block_iq1_s, 1, vec_dot_iq1_s_q8_1>
+                (src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
+            break;
+        case GGML_TYPE_IQ4_NL:
+            mul_mat_vec_q_cuda<QK4_NL, QI4_NL, block_iq4_nl, VDR_Q4_0_Q8_1_MMVQ, vec_dot_iq4_nl_q8_1>
+                (src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
+            break;
         default:
             GGML_ASSERT(false);
             break;
@@ -9089,30 +9257,36 @@ static void ggml_cuda_op_soft_max(
 
     GGML_ASSERT(!src1 || src1->type == GGML_TYPE_F32); // src1 contains mask and it is optional
 
-    const int64_t ne00 = src0->ne[0];
+    const int64_t ne00    = src0->ne[0];
     const int64_t nrows_x = ggml_nrows(src0);
-    const int64_t nrows_y = src1 ? ggml_nrows(src1) : 1;
+    const int64_t nrows_y = src0->ne[1];
 
-    float scale = 1.0f;
-    memcpy(&scale, dst->op_params, sizeof(float));
+    float scale    = 1.0f;
+    float max_bias = 0.0f;
 
-#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && CUDART_VERSION >= CUDART_HMAX
-#ifdef GGML_CUDA_F16
-    const bool use_f16_soft_max = true;
-#else
-    const bool use_f16_soft_max = false;
-#endif // GGML_CUDA_F16
-#else
-    const bool use_f16_soft_max = false;
-#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) && CUDART_VERSION >= CUDART_HMAX
+    memcpy(&scale,    (float *) dst->op_params + 0, sizeof(float));
+    memcpy(&max_bias, (float *) dst->op_params + 1, sizeof(float));
 
-    if (use_f16_soft_max) {
-        soft_max_f16_cuda(src0_dd, src1 ? src1_dd : nullptr, dst_dd, ne00, nrows_x, nrows_y, scale, main_stream);
-    } else {
-        soft_max_f32_cuda(src0_dd, src1 ? src1_dd : nullptr, dst_dd, ne00, nrows_x, nrows_y, scale, main_stream);
+    // positions tensor
+    float * src2_dd = nullptr;
+    cuda_pool_alloc<float> src2_f;
+
+    ggml_tensor * src2 = dst->src[2];
+    const bool use_src2 = src2 != nullptr;
+
+    if (use_src2) {
+        const bool src2_on_device = src2->backend == GGML_BACKEND_GPU;
+
+        if (src2_on_device) {
+            ggml_tensor_extra_gpu * src2_extra = (ggml_tensor_extra_gpu *) src2->extra;
+            src2_dd = (float *) src2_extra->data_device[g_main_device];
+        } else {
+            src2_dd = src2_f.alloc(ggml_nelements(src2));
+            CUDA_CHECK(ggml_cuda_cpy_tensor_2d(src2_dd, src2, 0, 0, 0, 1, main_stream));
+        }
     }
 
-    (void) dst;
+    soft_max_f32_cuda(src0_dd, src1 ? src1_dd : nullptr, src2_dd, dst_dd, ne00, nrows_x, nrows_y, scale, max_bias, main_stream);
 }
 
 static void ggml_cuda_op_scale(
@@ -9247,9 +9421,15 @@ static void ggml_cuda_set_peer_access(const int n_tokens) {
             CUDA_CHECK(cudaDeviceCanAccessPeer(&can_access_peer, id, id_other));
             if (can_access_peer) {
                 if (enable_peer_access) {
-                    CUDA_CHECK(cudaDeviceEnablePeerAccess(id_other, 0));
+                    cudaError_t err = cudaDeviceEnablePeerAccess(id_other, 0);
+                    if (err != cudaErrorPeerAccessAlreadyEnabled) {
+                        CUDA_CHECK(err);
+                    }
                 } else {
-                    CUDA_CHECK(cudaDeviceDisablePeerAccess(id_other));
+                    cudaError_t err = cudaDeviceDisablePeerAccess(id_other);
+                    if (err != cudaErrorPeerAccessNotEnabled) {
+                        CUDA_CHECK(err);
+                    }
                 }
             }
         }
@@ -10921,10 +11101,10 @@ GGML_CALL static const char * ggml_backend_cuda_split_buffer_get_name(ggml_backe
     UNUSED(buffer);
 }
 
-// unused at the moment
-//static bool ggml_backend_buffer_is_cuda_split(ggml_backend_buffer_t buffer) {
-//    return buffer->iface.get_name == ggml_backend_cuda_split_buffer_get_name;
-//}
+static bool ggml_backend_buffer_is_cuda_split(ggml_backend_buffer_t buffer) {
+    return buffer->iface.get_name == ggml_backend_cuda_split_buffer_get_name;
+    UNUSED(ggml_backend_buffer_is_cuda_split); // only used in debug builds currently, avoid unused function warning in release builds
+}
 
 GGML_CALL static void ggml_backend_cuda_split_buffer_free_buffer(ggml_backend_buffer_t buffer) {
     ggml_backend_cuda_split_buffer_context * ctx = (ggml_backend_cuda_split_buffer_context *)buffer->context;
@@ -11312,7 +11492,7 @@ GGML_CALL static bool ggml_backend_cuda_graph_compute(ggml_backend_t backend, gg
         for (int j = 0; j < GGML_MAX_SRC; j++) {
             if (node->src[j] != nullptr) {
                 assert(node->src[j]->backend == GGML_BACKEND_GPU || node->src[j]->backend == GGML_BACKEND_GPU_SPLIT);
-                assert(node->src[j]->buffer->buft == ggml_backend_cuda_buffer_type(cuda_ctx->device));
+                assert(node->src[j]->buffer->buft == ggml_backend_cuda_buffer_type(cuda_ctx->device) || ggml_backend_buffer_is_cuda_split(node->src[j]->buffer));
                 assert(node->src[j]->extra != nullptr);
             }
         }
@@ -11360,7 +11540,8 @@ GGML_CALL static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, cons
                     return false;
                 }
                 ggml_type a_type = a->type;
-                if (a_type == GGML_TYPE_IQ2_XXS || a_type == GGML_TYPE_IQ2_XS || a_type == GGML_TYPE_IQ3_XXS) {
+                if (a_type == GGML_TYPE_IQ2_XXS || a_type == GGML_TYPE_IQ2_XS || a_type == GGML_TYPE_IQ3_XXS ||
+                    a_type == GGML_TYPE_IQ1_S   || a_type == GGML_TYPE_IQ4_NL) {
                     if (b->ne[1] == 1 && ggml_nrows(b) > 1) {
                         return false;
                     }

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 +61,8 @@ 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_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 +85,8 @@ 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_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 +105,8 @@ 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_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 +122,8 @@ 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_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 +139,8 @@ 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_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 +186,7 @@ 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 +282,14 @@ 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 +312,7 @@ 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 +452,8 @@ 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_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 +476,8 @@ 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_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 +496,8 @@ 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_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 +513,8 @@ 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_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 +530,8 @@ 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_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 +757,7 @@ 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 +776,7 @@ 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 +838,7 @@ 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 +1220,16 @@ 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 +1238,20 @@ 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 +1347,8 @@ 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_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 +1483,18 @@ 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_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 +1529,7 @@ 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) {
@@ -1478,6 +1542,11 @@ static bool ggml_metal_graph_compute(
                                 [encoder setThreadgroupMemoryLength:mem_size atIndex:0];
                                 [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
                             }
+                            else if (src0t == GGML_TYPE_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 +1583,6 @@ 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 +1640,8 @@ 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_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 +1779,18 @@ 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_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 +1841,7 @@ 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) {
@@ -1773,6 +1854,11 @@ static bool ggml_metal_graph_compute(
                                 [encoder setThreadgroupMemoryLength:mem_size atIndex:0];
                                 [encoder dispatchThreadgroups:MTLSizeMake((ne21 + 7)/8, _ne1, ne01*ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
                             }
+                            else if (src2t == GGML_TYPE_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 +1900,8 @@ 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_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");
                         }

ggml-metal.metal

@@ -351,12 +351,17 @@ kernel void kernel_sum_rows(
 kernel void kernel_soft_max(
         device const float * src0,
         device const float * src1,
+        device const float * src2,
         device       float * dst,
         constant   int64_t & ne00,
         constant   int64_t & ne01,
         constant   int64_t & ne02,
         constant     float & scale,
-        threadgroup float  * buf [[threadgroup(0)]],
+        constant     float & max_bias,
+        constant     float & m0,
+        constant     float & m1,
+        constant  uint32_t & n_head_log2,
+        threadgroup  float * buf [[threadgroup(0)]],
         uint  tgpig[[threadgroup_position_in_grid]],
         uint  tpitg[[thread_position_in_threadgroup]],
         uint  sgitg[[simdgroup_index_in_threadgroup]],
@@ -368,13 +373,26 @@ kernel void kernel_soft_max(
 
     device const float * psrc0 =         src0 + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
     device const float * pmask = src1 != src0 ? src1                               + i01*ne00 : nullptr;
+    device const float * ppos  = src2 != src0 ? src2                                          : nullptr;
     device       float * pdst  =         dst  + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
 
+    float slope = 0.0f;
+
+    // ALiBi
+    if (max_bias > 0.0f) {
+        const int64_t h = i02;
+
+        const float base = h < n_head_log2 ? m0 : m1;
+        const int   exp  = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1;
+
+        slope = pow(base, exp);
+    }
+
     // parallel max
     float lmax = -INFINITY;
 
     for (int i00 = tpitg; i00 < ne00; i00 += ntg) {
-        lmax = MAX(lmax, psrc0[i00]*scale + (pmask ? pmask[i00] : 0.0f));
+        lmax = MAX(lmax, psrc0[i00]*scale + (pmask ? pmask[i00] : 0.0f) + (ppos ? slope*ppos[i00] : 0.0f));
     }
 
     // find the max value in the block
@@ -399,7 +417,7 @@ kernel void kernel_soft_max(
     // parallel sum
     float lsum = 0.0f;
     for (int i00 = tpitg; i00 < ne00; i00 += ntg) {
-        const float exp_psrc0 = exp((psrc0[i00]*scale + (pmask ? pmask[i00] : 0.0f)) - max_val);
+        const float exp_psrc0 = exp((psrc0[i00]*scale + (pmask ? pmask[i00] : 0.0f) + (ppos ? slope*ppos[i00] : 0.0f)) - max_val);
         lsum += exp_psrc0;
         pdst[i00] = exp_psrc0;
     }
@@ -437,12 +455,17 @@ kernel void kernel_soft_max(
 kernel void kernel_soft_max_4(
         device const float * src0,
         device const float * src1,
+        device const float * src2,
         device       float * dst,
         constant   int64_t & ne00,
         constant   int64_t & ne01,
         constant   int64_t & ne02,
         constant     float & scale,
-        threadgroup float  * buf [[threadgroup(0)]],
+        constant     float & max_bias,
+        constant     float & m0,
+        constant     float & m1,
+        constant  uint32_t & n_head_log2,
+        threadgroup  float * buf [[threadgroup(0)]],
         uint  tgpig[[threadgroup_position_in_grid]],
         uint  tpitg[[thread_position_in_threadgroup]],
         uint  sgitg[[simdgroup_index_in_threadgroup]],
@@ -454,13 +477,25 @@ kernel void kernel_soft_max_4(
 
     device const float4 * psrc4 =                (device const float4 *)(src0 + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00);
     device const float4 * pmask = src1 != src0 ? (device const float4 *)(src1 +                                      i01*ne00) : nullptr;
+    device const float4 * ppos  = src2 != src0 ? (device const float4 *)(src2)                                                 : nullptr;
     device       float4 * pdst4 =                (device       float4 *)(dst  + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00);
 
+    float slope = 0.0f;
+
+    if (max_bias > 0.0f) {
+        const int64_t h = i02;
+
+        const float base = h < n_head_log2 ? m0 : m1;
+        const int   exp  = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1;
+
+        slope = pow(base, exp);
+    }
+
     // parallel max
     float4 lmax4 = -INFINITY;
 
     for (int i00 = tpitg; i00 < ne00/4; i00 += ntg) {
-        lmax4 = fmax(lmax4, psrc4[i00]*scale + (pmask ? pmask[i00] : 0.0f));
+        lmax4 = fmax(lmax4, psrc4[i00]*scale + (pmask ? pmask[i00] : 0.0f) + (ppos ? slope*ppos[i00] : 0.0f));
     }
 
     const float lmax = MAX(MAX(lmax4[0], lmax4[1]), MAX(lmax4[2], lmax4[3]));
@@ -486,7 +521,7 @@ kernel void kernel_soft_max_4(
     // parallel sum
     float4 lsum4 = 0.0f;
     for (int i00 = tpitg; i00 < ne00/4; i00 += ntg) {
-        const float4 exp_psrc4 = exp((psrc4[i00]*scale + (pmask ? pmask[i00] : 0.0f)) - max_val);
+        const float4 exp_psrc4 = exp((psrc4[i00]*scale + (pmask ? pmask[i00] : 0.0f) + (ppos ? slope*ppos[i00] : 0.0f)) - max_val);
         lsum4 += exp_psrc4;
         pdst4[i00] = exp_psrc4;
     }
@@ -2490,6 +2525,19 @@ typedef struct {
 } block_iq3_xxs;
 // 98 bytes / block for QK_K = 256, so 3.0625 bpw
 
+typedef struct {
+    half d;
+    uint8_t qs[QK_K/8];
+    uint8_t scales[QK_K/16];
+} block_iq1_s;
+
+// Non-linear quants
+#define QK4_NL 32
+typedef struct {
+    half    d;
+    uint8_t qs[QK4_NL/2];
+} block_iq4_nl;
+
 //====================================== dot products =========================
 
 void kernel_mul_mv_q2_K_f32_impl(
@@ -3747,6 +3795,137 @@ constexpr constant static uint32_t iq3xxs_grid[256] = {
     0x3e1c1c1c, 0x3e1c3404, 0x3e24140c, 0x3e24240c, 0x3e2c0404, 0x3e2c0414, 0x3e2c1424, 0x3e341c04,
 };
 
+#define NGRID_IQ1S 512
+constexpr constant static uint64_t iq1s_grid[NGRID_IQ1S] = {
+    0xffffffffffff0101, 0xffffffffff01ff00, 0xffffffffff010100, 0xffffffff00000000,
+    0xffffffff01ff00ff, 0xffffffff01ff0001, 0xffffffff0101ffff, 0xffffffff0101ff01,
+    0xffffff00ff000000, 0xffffff000000ff00, 0xffffff00000000ff, 0xffffff0000000100,
+    0xffffff0000010000, 0xffffff0001000000, 0xffffff01ffff00ff, 0xffffff01ff01ff00,
+    0xffffff01ff010100, 0xffffff0100000001, 0xffffff0101ffff00, 0xffffff0101ff0101,
+    0xffffff0101010100, 0xffff00ffff00ff01, 0xffff00ffff0000ff, 0xffff00ff00ff0100,
+    0xffff00ff0100ff00, 0xffff00ff010001ff, 0xffff0000ff0101ff, 0xffff000000ffff00,
+    0xffff000000000000, 0xffff00000001ff01, 0xffff000001000101, 0xffff0000010100ff,
+    0xffff0001ffff0100, 0xffff00010000ff00, 0xffff000100010101, 0xffff000101000000,
+    0xffff01ffffff0000, 0xffff01ffff01ffff, 0xffff01ffff010100, 0xffff01ff00000000,
+    0xffff01ff01ffffff, 0xffff01ff01ff0001, 0xffff01ff0101ffff, 0xffff01ff01010001,
+    0xffff0100ffffff01, 0xffff01000000ffff, 0xffff010000000100, 0xffff010001ff01ff,
+    0xffff010001000000, 0xffff0101ff000000, 0xffff0101000101ff, 0xffff010101ffff01,
+    0xffff01010101ff00, 0xff00ffffff000000, 0xff00ffff00ffff00, 0xff00ffff00000001,
+    0xff00ffff000001ff, 0xff00ffff01010000, 0xff00ff00ffff0000, 0xff00ff00ff00ff00,
+    0xff00ff00ff0000ff, 0xff00ff00ff000100, 0xff00ff00ff010001, 0xff00ff0000ff0001,
+    0xff00ff000000ffff, 0xff00ff0000000000, 0xff00ff000001ff00, 0xff00ff0000010100,
+    0xff00ff0001ff0000, 0xff00ff000100ff00, 0xff00ff0001000100, 0xff00ff01ff000000,
+    0xff00ff0100ff0000, 0xff00ff01000001ff, 0xff00ff0101010001, 0xff0000ff00000000,
+    0xff0000ff0001ff00, 0xff0000ff00010100, 0xff000000ffff0101, 0xff000000ff000000,
+    0xff000000ff01ff00, 0xff00000000ff0000, 0xff0000000000ff00, 0xff000000000000ff,
+    0xff00000000000000, 0xff00000000000001, 0xff00000000000100, 0xff0000000001ffff,
+    0xff00000000010000, 0xff00000001000000, 0xff00000001010100, 0xff000001ff00ff01,
+    0xff000001ff0100ff, 0xff00000100000000, 0xff0000010001ff00, 0xff00000101ff0100,
+    0xff0000010100ff00, 0xff0001ff00ff00ff, 0xff0001ff00000101, 0xff0001ff000100ff,
+    0xff0001ff01000000, 0xff000100ff0001ff, 0xff0001000000ff01, 0xff00010000000000,
+    0xff00010000010001, 0xff00010000010100, 0xff00010001ffff00, 0xff00010001ff0101,
+    0xff00010001010000, 0xff000101ffffffff, 0xff000101ff000101, 0xff00010101ff00ff,
+    0xff00010101000001, 0xff000101010100ff, 0xff01ffffff000101, 0xff01ffffff01ffff,
+    0xff01ffffff01ff01, 0xff01ffffff0101ff, 0xff01ffff00000000, 0xff01ffff01ff0001,
+    0xff01ffff0101ff01, 0xff01ff00ff000000, 0xff01ff0000ff0100, 0xff01ff000000ff01,
+    0xff01ff0000010000, 0xff01ff00010000ff, 0xff01ff01ff01ff00, 0xff01ff0100000101,
+    0xff0100ffffff0000, 0xff0100ffff010000, 0xff0100ff01ff00ff, 0xff0100ff01000100,
+    0xff0100ff010100ff, 0xff010000ffffff01, 0xff01000000000000, 0xff0100000101ff00,
+    0xff010001ffff00ff, 0xff010001ff000100, 0xff01000100ffff00, 0xff01000100010001,
+    0xff01000101ff0001, 0xff010001010001ff, 0xff0101ffffffffff, 0xff0101ffff01ffff,
+    0xff0101ffff010101, 0xff0101ff0000ff00, 0xff0101ff01010001, 0xff010100ff000000,
+    0xff010100ff01ff01, 0xff01010000ff0001, 0xff01010000000100, 0xff01010001000000,
+    0xff0101010100ffff, 0x00ffffff0000ff01, 0x00ffffff000000ff, 0x00ffffff00000100,
+    0x00ffffff00010000, 0x00ffff00ffff0001, 0x00ffff00ff0000ff, 0x00ffff00ff000100,
+    0x00ffff0000000000, 0x00ffff0001000100, 0x00ffff0001010001, 0x00ffff01ff00ff01,
+    0x00ffff0100ff0100, 0x00ffff010000ff00, 0x00ffff01000100ff, 0x00ffff0101ff00ff,
+    0x00ffff010101ff00, 0x00ff00ffffffffff, 0x00ff00ffffff01ff, 0x00ff00ffff000101,
+    0x00ff00ff00000000, 0x00ff00ff000101ff, 0x00ff00ff01010101, 0x00ff0000ff000000,
+    0x00ff0000ff01ffff, 0x00ff000000ff0000, 0x00ff00000000ff00, 0x00ff0000000000ff,
+    0x00ff000000000000, 0x00ff000000000001, 0x00ff000000000100, 0x00ff000000010000,
+    0x00ff000001ffff01, 0x00ff000001000000, 0x00ff0001ff000101, 0x00ff000100ffffff,
+    0x00ff000100000000, 0x00ff0001010001ff, 0x00ff01ffff000000, 0x00ff01ff0001ff00,
+    0x00ff01ff01ff0100, 0x00ff0100ff01ff01, 0x00ff010000ff00ff, 0x00ff010000ff0101,
+    0x00ff010000000000, 0x00ff010000010101, 0x00ff01000100ff00, 0x00ff010001010000,
+    0x00ff0101ffffff00, 0x00ff01010000ff01, 0x00ff010100000100, 0x00ff010101ff0000,
+    0x0000ffffffff0100, 0x0000ffffff00ff00, 0x0000ffffff0000ff, 0x0000ffffff010000,
+    0x0000ffff00000000, 0x0000ffff00010101, 0x0000ffff01ffff01, 0x0000ffff01000100,
+    0x0000ff00ff000000, 0x0000ff00ff01ff00, 0x0000ff00ff0101ff, 0x0000ff0000ff0000,
+    0x0000ff000000ff00, 0x0000ff00000000ff, 0x0000ff0000000000, 0x0000ff0000000001,
+    0x0000ff0000000100, 0x0000ff0000010000, 0x0000ff0001ffffff, 0x0000ff0001ff01ff,
+    0x0000ff0001000000, 0x0000ff000101ffff, 0x0000ff01ffff0101, 0x0000ff01ff010000,
+    0x0000ff0100000000, 0x0000ff0101000101, 0x000000ffffff0001, 0x000000ffff000000,
+    0x000000ff00ff0000, 0x000000ff0000ff00, 0x000000ff000000ff, 0x000000ff00000000,
+    0x000000ff00000001, 0x000000ff00000100, 0x000000ff00010000, 0x000000ff01000000,
+    0x000000ff0101ff00, 0x00000000ffff0000, 0x00000000ff00ff00, 0x00000000ff0000ff,
+    0x00000000ff000000, 0x00000000ff000001, 0x00000000ff000100, 0x00000000ff010000,
+    0x0000000000ffff00, 0x0000000000ff00ff, 0x0000000000ff0000, 0x0000000000ff0001,
+    0x0000000000ff0100, 0x000000000000ffff, 0x000000000000ff00, 0x000000000000ff01,
+    0x00000000000000ff, 0x0000000000000001, 0x00000000000001ff, 0x0000000000000100,
+    0x0000000000000101, 0x000000000001ff00, 0x00000000000100ff, 0x0000000000010000,
+    0x0000000000010001, 0x0000000000010100, 0x0000000001ff0000, 0x000000000100ff00,
+    0x00000000010000ff, 0x0000000001000000, 0x0000000001000001, 0x0000000001000100,
+    0x0000000001010000, 0x00000001ffff01ff, 0x00000001ff000000, 0x0000000100ff0000,
+    0x000000010000ff00, 0x00000001000000ff, 0x0000000100000000, 0x0000000100000001,
+    0x0000000100000100, 0x0000000100010000, 0x0000000101000000, 0x000001ffff00ff00,
+    0x000001ffff010001, 0x000001ffff0101ff, 0x000001ff00ffff01, 0x000001ff0000ffff,
+    0x000001ff00000000, 0x000001ff010000ff, 0x000001ff01010100, 0x00000100ffff0100,
+    0x00000100ff000000, 0x0000010000ff0000, 0x000001000000ff00, 0x00000100000000ff,
+    0x0000010000000000, 0x0000010000000001, 0x0000010000000100, 0x0000010000010000,
+    0x0000010001000000, 0x000001000101ff01, 0x00000101ffff0001, 0x00000101ff01ffff,
+    0x0000010100000000, 0x0000010101010100, 0x0001ffffff000000, 0x0001ffff00ffffff,
+    0x0001ffff00000100, 0x0001ffff0001ff00, 0x0001ffff01000000, 0x0001ff00ffffff00,
+    0x0001ff00ffff01ff, 0x0001ff00ff010000, 0x0001ff0000000000, 0x0001ff0000010001,
+    0x0001ff0001ff0000, 0x0001ff0001010100, 0x0001ff01ff0000ff, 0x0001ff01ff000001,
+    0x0001ff0100ffffff, 0x0001ff010001ffff, 0x0001ff01000101ff, 0x0001ff010100ff01,
+    0x000100ffff00ffff, 0x000100ffff00ff01, 0x000100ffff000100, 0x000100ff00000000,
+    0x000100ff000101ff, 0x000100ff01ff0101, 0x000100ff0100ffff, 0x000100ff01010101,
+    0x00010000ff000000, 0x00010000ff010100, 0x0001000000ff0000, 0x000100000000ff00,
+    0x00010000000000ff, 0x0001000000000000, 0x0001000000000001, 0x0001000000000100,
+    0x0001000000010000, 0x0001000001ffff01, 0x0001000001000000, 0x0001000100ff0101,
+    0x0001000100000000, 0x00010001010100ff, 0x000101ffffff01ff, 0x000101ffffff0101,
+    0x000101ff00010000, 0x000101ff01ff0000, 0x000101ff0100ff01, 0x00010100ffff0000,
+    0x0001010000000000, 0x000101000001ffff, 0x0001010000010101, 0x00010100010001ff,
+    0x00010101ff00ff00, 0x00010101ff010001, 0x0001010100ffffff, 0x0001010100ff01ff,
+    0x00010101000101ff, 0x0001010101ff0000, 0x000101010100ff01, 0x0001010101000101,
+    0x01ffffffffff0101, 0x01ffffffff01ffff, 0x01ffffffff01ff01, 0x01ffffffff0101ff,
+    0x01ffffffff010101, 0x01ffffff00000000, 0x01ffffff01ff01ff, 0x01ffffff01000101,
+    0x01ffffff0101ff01, 0x01ffffff010100ff, 0x01ffff000000ff00, 0x01ffff0000000001,
+    0x01ffff00000001ff, 0x01ffff0000010000, 0x01ffff0001ff0000, 0x01ffff01ffffffff,
+    0x01ffff01ffff01ff, 0x01ffff01ff000000, 0x01ffff01ff01ffff, 0x01ffff01ff0101ff,
+    0x01ffff010100ffff, 0x01ff00ffffff0000, 0x01ff00ffff010000, 0x01ff00ff00ffff01,
+    0x01ff0000ff0000ff, 0x01ff000000000000, 0x01ff00000001ff01, 0x01ff000001ffffff,
+    0x01ff000001010100, 0x01ff0001ffffff01, 0x01ff0001ff010001, 0x01ff000101ff0100,
+    0x01ff000101000001, 0x01ff0001010100ff, 0x01ff01ffff00ffff, 0x01ff01ff00010001,
+    0x01ff01ff01000000, 0x01ff01ff010101ff, 0x01ff0100ff000001, 0x01ff010000ffff00,
+    0x01ff010000000100, 0x01ff010001ff01ff, 0x01ff01000101ffff, 0x01ff0101ffff00ff,
+    0x01ff0101ffff0101, 0x01ff0101ff0101ff, 0x01ff010100010000, 0x0100ffff00ff00ff,
+    0x0100ffff00ff0001, 0x0100ffff00000100, 0x0100ffff0100ff00, 0x0100ff00ffff0000,
+    0x0100ff00ff00ffff, 0x0100ff00ff00ff01, 0x0100ff00ff000100, 0x0100ff00ff010000,
+    0x0100ff0000000000, 0x0100ff00000100ff, 0x0100ff0001ff0101, 0x0100ff0001010101,
+    0x0100ff0100ff00ff, 0x0100ff0100ff0001, 0x0100ff0100000100, 0x0100ff0100010001,
+    0x0100ff0101000000, 0x010000ffff00ff00, 0x010000ff0000ffff, 0x010000ff00000000,
+    0x010000ff010001ff, 0x010000ff01010001, 0x01000000ffffff00, 0x01000000ffff0101,
+    0x01000000ff000000, 0x01000000ff0100ff, 0x01000000ff010101, 0x0100000000ff0000,
+    0x010000000000ff00, 0x01000000000000ff, 0x0100000000000000, 0x0100000000000001,
+    0x0100000000000100, 0x0100000000010000, 0x0100000001000000, 0x0100000100000000,
+    0x01000001000101ff, 0x0100000101ffff01, 0x010001ffff000101, 0x010001ff00ff0100,
+    0x010001ff0000ff00, 0x010001ff000100ff, 0x010001ff01ffffff, 0x01000100ffff0000,
+    0x01000100ff0001ff, 0x0100010000000000, 0x010001000001ff00, 0x0100010001ff0000,
+    0x01000100010000ff, 0x0100010001000101, 0x01000101ff00ff01, 0x0100010100ff0100,
+    0x010001010000ffff, 0x0100010101010001, 0x0101ffffffff0101, 0x0101ffffff0001ff,
+    0x0101ffffff01ffff, 0x0101ffffff010101, 0x0101ffff00000000, 0x0101ffff0101ffff,
+    0x0101ffff010101ff, 0x0101ff00ff000000, 0x0101ff0000ff0100, 0x0101ff000000ff00,
+    0x0101ff0000010000, 0x0101ff00010000ff, 0x0101ff0001000001, 0x0101ff01ff010101,
+    0x0101ff0100000000, 0x0101ff010101ff00, 0x010100ffffff0000, 0x010100ffff010000,
+    0x010100ff00ff01ff, 0x010100ff000000ff, 0x010100ff00000101, 0x010100ff01ffff00,
+    0x01010000ffffff01, 0x01010000ff000100, 0x01010000ff01ff01, 0x0101000000000000,
+    0x01010000000100ff, 0x010100000101ff01, 0x01010001ffff0000, 0x01010001ff00ffff,
+    0x01010001ff010000, 0x0101000101ffffff, 0x0101000101ff01ff, 0x0101000101010101,
+    0x010101ffff01ffff, 0x010101ff00000000, 0x010101ff0001ff01, 0x010101ff0101ffff,
+    0x010101ff010101ff, 0x01010100ffffffff, 0x01010100ff000001, 0x010101000000ff00,
+    0x0101010001010000, 0x0101010100ff0001, 0x010101010001ff01, 0x010101010101ffff,
+};
 
 constexpr constant static uint8_t ksigns_iq2xs[128] = {
       0, 129, 130,   3, 132,   5,   6, 135, 136,   9,  10, 139,  12, 141, 142,  15,
@@ -3854,7 +4033,10 @@ void kernel_mul_mv_iq2_xxs_f32_impl(
         y4 += 32 * 32;
     }
 #else
-    // TODO
+    (void) x;
+    (void) y;
+    (void) yl;
+    (void) nb32;
 #endif
 
     for (int row = 0; row < N_DST; ++row) {
@@ -3997,7 +4179,10 @@ void kernel_mul_mv_iq2_xs_f32_impl(
         y4 += 32 * 32;
     }
 #else
-    // TODO
+    (void) x;
+    (void) y;
+    (void) yl;
+    (void) nb32;
 #endif
 
     for (int row = 0; row < N_DST; ++row) {
@@ -4133,7 +4318,10 @@ void kernel_mul_mv_iq3_xxs_f32_impl(
         y4 += 32 * 32;
     }
 #else
-    // TODO
+    (void) x;
+    (void) y;
+    (void) yl;
+    (void) nb32;
 #endif
 
     for (int row = 0; row < N_DST; ++row) {
@@ -4173,6 +4361,250 @@ kernel void kernel_mul_mv_iq3_xxs_f32(
     kernel_mul_mv_iq3_xxs_f32_impl(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3, shared_values, tgpig, tiisg, sgitg);
 }
 
+void kernel_mul_mv_iq1_s_f32_impl(
+        device const  void * src0,
+        device const float * src1,
+        device       float * dst,
+        constant   int64_t & ne00,
+        constant   int64_t & ne01,
+        constant   int64_t & ne02,
+        constant   int64_t & ne10,
+        constant   int64_t & ne12,
+        constant   int64_t & ne0,
+        constant   int64_t & ne1,
+        constant   uint    & r2,
+        constant   uint    & r3,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint  tiisg[[thread_index_in_simdgroup]],
+        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
+
+    const int nb = ne00/QK_K;
+    const int r0 = tgpig.x;
+    const int r1 = tgpig.y;
+    const int im = tgpig.z;
+
+    const int first_row = (r0 * N_SIMDGROUP + sgitg) * N_DST;
+    const int ib_row = first_row * nb;
+
+    const uint i12 = im%ne12;
+    const uint i13 = im/ne12;
+
+    const uint offset0 = (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
+    device const block_iq1_s * x = (device const block_iq1_s *) src0 + ib_row + offset0;
+    device const float       * y = (device const float       *) src1 + r1*ne10 + im*ne00*ne1;
+
+    float yl[16];
+    float sumf[N_DST]={0.f}, all_sum;
+
+    const int nb32 = nb * (QK_K / 32);
+
+#if QK_K == 256
+    const int ix = tiisg/2;
+    const int il = tiisg%2;
+
+    device const float * y4 = y + 32 * ix + 16 * il;
+
+    for (int ib32 = ix; ib32 < nb32; ib32 += 16) {
+
+        for (int i = 0; i < 16; ++i) {
+            yl[i] = y4[i];
+        }
+
+        const int ibl = ib32 / (QK_K / 32);
+        const int ib  = ib32 % (QK_K / 32);
+
+        device const block_iq1_s * xr = x + ibl;
+        device const uint8_t * qs = xr->qs + 4 * ib + 2 * il;
+        device const uint8_t * sc = xr->scales + 2 * ib + il;
+        device const half    * dh = &xr->d;
+
+        for (int row = 0; row < N_DST; row++) {
+
+            constant int8_t * grid1 = (constant int8_t *)(iq1s_grid + (qs[0] | ((sc[0] & 0x08) << 5)));
+            constant int8_t * grid2 = (constant int8_t *)(iq1s_grid + (qs[1] | ((sc[0] & 0x80) << 1)));
+
+            float2 sum = {0};
+            for (int j = 0; j < 8; ++j) {
+                sum[0] += yl[j+ 0] * grid1[j];
+                sum[1] += yl[j+ 8] * grid2[j];
+            }
+            sumf[row] += (float)dh[0] * (sum[0] * (2*(sc[0] & 7) + 1) + sum[1] * (2*((sc[0] >> 4) & 7) + 1));
+
+            dh += nb*sizeof(block_iq1_s)/2;
+            qs += nb*sizeof(block_iq1_s);
+            sc += nb*sizeof(block_iq1_s);
+        }
+
+        y4 += 16 * 32;
+    }
+#else
+    (void) x;
+    (void) y;
+    (void) yl;
+    (void) nb32;
+#endif
+
+    for (int row = 0; row < N_DST; ++row) {
+        all_sum = simd_sum(sumf[row]);
+        if (tiisg == 0) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + row] = all_sum;
+        }
+    }
+}
+
+constexpr constant static float kvalues_iq4nl_f[16] = {
+    -127.f, -104.f, -83.f, -65.f, -49.f, -35.f, -22.f, -10.f, 1.f, 13.f, 25.f, 38.f, 53.f, 69.f, 89.f, 113.f
+};
+
+void kernel_mul_mv_iq4_nl_f32_impl(
+        device const  void * src0,
+        device const float * src1,
+        device       float * dst,
+        constant   int64_t & ne00,
+        constant   int64_t & ne01,
+        constant   int64_t & ne02,
+        constant   int64_t & ne10,
+        constant   int64_t & ne12,
+        constant   int64_t & ne0,
+        constant   int64_t & ne1,
+        constant   uint    & r2,
+        constant   uint    & r3,
+        threadgroup float  * shared_values [[threadgroup(0)]],
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint  tiisg[[thread_index_in_simdgroup]],
+        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
+
+    const int nb = ne00/QK4_NL;
+    const int r0 = tgpig.x;
+    const int r1 = tgpig.y;
+    const int im = tgpig.z;
+    const int first_row = (r0 * 2 + sgitg) * 2;
+    const int ib_row = first_row * nb;
+
+    const uint i12 = im%ne12;
+    const uint i13 = im/ne12;
+
+    const uint offset0 = (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
+    device const block_iq4_nl * x = (device const block_iq4_nl *) src0 + ib_row + offset0;
+    device const float        * y = (device const float        *) src1 + r1*ne10 + im*ne00*ne1;
+
+    const int ix = tiisg/2;  // 0...15
+    const int it = tiisg%2;  // 0 or 1
+
+    shared_values[tiisg] = kvalues_iq4nl_f[tiisg%16];
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    float4 yl[4];
+    float sumf[2]={0.f}, all_sum;
+
+    device const float * yb = y + ix * QK4_NL + it * 8;
+
+    uint32_t aux32[2];
+    thread const uint8_t * q8 = (thread const uint8_t *)aux32;
+
+    float4 qf1, qf2;
+
+    for (int ib = ix; ib < nb; ib += 16) {
+
+        device const float4 * y4 = (device const float4 *)yb;
+        yl[0] = y4[0]; yl[1] = y4[4]; yl[2] = y4[1]; yl[3] = y4[5];
+
+        for (int row = 0; row < 2; ++row) {
+
+            device const block_iq4_nl & xb = x[row*nb + ib];
+            device const uint16_t * q4 = (device const uint16_t *)(xb.qs + 8*it);
+
+            float4 acc1 = {0.f}, acc2 = {0.f};
+
+            aux32[0] = q4[0] | (q4[1] << 16);
+            aux32[1] = (aux32[0] >> 4) & 0x0f0f0f0f;
+            aux32[0] &= 0x0f0f0f0f;
+            qf1 = {shared_values[q8[0]], shared_values[q8[1]], shared_values[q8[2]], shared_values[q8[3]]};
+            qf2 = {shared_values[q8[4]], shared_values[q8[5]], shared_values[q8[6]], shared_values[q8[7]]};
+            acc1 += yl[0] * qf1;
+            acc2 += yl[1] * qf2;
+
+            aux32[0] = q4[2] | (q4[3] << 16);
+            aux32[1] = (aux32[0] >> 4) & 0x0f0f0f0f;
+            aux32[0] &= 0x0f0f0f0f;
+            qf1 = {shared_values[q8[0]], shared_values[q8[1]], shared_values[q8[2]], shared_values[q8[3]]};
+            qf2 = {shared_values[q8[4]], shared_values[q8[5]], shared_values[q8[6]], shared_values[q8[7]]};
+            acc1 += yl[2] * qf1;
+            acc2 += yl[3] * qf2;
+
+            acc1 += acc2;
+
+            sumf[row] += (float)xb.d * (acc1[0] + acc1[1] + acc1[2] + acc1[3]);
+
+        }
+
+        yb += 16 * QK4_NL;
+    }
+
+    for (int row = 0; row < 2; ++row) {
+        all_sum = simd_sum(sumf[row]);
+        if (tiisg == 0) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + row] = all_sum;
+        }
+    }
+}
+
+[[host_name("kernel_mul_mv_iq1_s_f32")]]
+kernel void kernel_mul_mv_iq1_s_f32(
+        device const  void * src0,
+        device const float * src1,
+        device       float * dst,
+        constant   int64_t & ne00,
+        constant   int64_t & ne01,
+        constant   int64_t & ne02,
+        constant  uint64_t & nb00,
+        constant  uint64_t & nb01,
+        constant  uint64_t & nb02,
+        constant   int64_t & ne10,
+        constant   int64_t & ne11,
+        constant   int64_t & ne12,
+        constant  uint64_t & nb10,
+        constant  uint64_t & nb11,
+        constant  uint64_t & nb12,
+        constant   int64_t & ne0,
+        constant   int64_t & ne1,
+        constant   uint    & r2,
+        constant   uint    & r3,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint  tiisg[[thread_index_in_simdgroup]],
+        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
+
+    kernel_mul_mv_iq1_s_f32_impl(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3, tgpig, tiisg, sgitg);
+}
+
+[[host_name("kernel_mul_mv_iq4_nl_f32")]]
+kernel void kernel_mul_mv_iq4_nl_f32(
+        device const  void * src0,
+        device const float * src1,
+        device       float * dst,
+        constant   int64_t & ne00,
+        constant   int64_t & ne01,
+        constant   int64_t & ne02,
+        constant  uint64_t & nb00,
+        constant  uint64_t & nb01,
+        constant  uint64_t & nb02,
+        constant   int64_t & ne10,
+        constant   int64_t & ne11,
+        constant   int64_t & ne12,
+        constant  uint64_t & nb10,
+        constant  uint64_t & nb11,
+        constant  uint64_t & nb12,
+        constant   int64_t & ne0,
+        constant   int64_t & ne1,
+        constant   uint    & r2,
+        constant   uint    & r3,
+        threadgroup float * shared_values [[threadgroup(0)]],
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint tiisg[[thread_index_in_simdgroup]],
+        uint sgitg[[simdgroup_index_in_threadgroup]]) {
+
+    kernel_mul_mv_iq4_nl_f32_impl(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3, shared_values, tgpig, tiisg, sgitg);
+}
 
 //============================= templates and their specializations =============================
 
@@ -4369,6 +4801,8 @@ void dequantize_q4_K(device const block_q4_K *xb, short il, thread type4x4 & reg
     const float dl = d * sc[0];
     const float ml = min * sc[1];
 #else
+    (void) get_scale_min_k4_just2;
+
     q = q + 16 * (il&1);
     device const uint8_t * s = xb->scales;
     device const half2 * dh = (device const half2 *)xb->d;
@@ -4518,6 +4952,37 @@ void dequantize_iq3_xxs(device const block_iq3_xxs * xb, short il, thread type4x
     }
 }
 
+template <typename type4x4>
+void dequantize_iq1_s(device const block_iq1_s * xb, short il, thread type4x4 & reg) {
+    // il is 0...15 for QK_K = 256 => index of block of 32 is il/2
+    const float d = xb->d;
+    device const uint8_t * qs = xb->qs + 2*il;
+    device const uint8_t * sc = xb->scales + il;
+    const float dl1 = d * (2*(sc[0] & 7) + 1);
+    const float dl2 = d * (2*((sc[0] >> 4) & 7) + 1);
+    constant int8_t * grid1 = (constant int8_t *)(iq1s_grid + (qs[0] | ((sc[0] & 0x08) << 5)));
+    constant int8_t * grid2 = (constant int8_t *)(iq1s_grid + (qs[1] | ((sc[0] & 0x80) << 1)));
+    for (int i = 0; i < 8; ++i) {
+        reg[i/4+0][i%4] = dl1 * grid1[i];
+        reg[i/4+2][i%4] = dl2 * grid2[i];
+    }
+}
+
+template <typename type4x4>
+void dequantize_iq4_nl(device const block_iq4_nl * xb, short il, thread type4x4 & reg) {
+    device const uint16_t * q4 = (device const uint16_t *)xb->qs;
+    const float d = xb->d;
+    uint32_t aux32;
+    thread const uint8_t * q8 = (thread const uint8_t *)&aux32;
+    for (int i = 0; i < 4; ++i) {
+        aux32 = ((q4[2*i] | (q4[2*i+1] << 16)) >> 4*il) & 0x0f0f0f0f;
+        reg[i][0] = d * kvalues_iq4nl_f[q8[0]];
+        reg[i][1] = d * kvalues_iq4nl_f[q8[1]];
+        reg[i][2] = d * kvalues_iq4nl_f[q8[2]];
+        reg[i][3] = d * kvalues_iq4nl_f[q8[3]];
+    }
+}
+
 template<typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread float4x4 &)>
 kernel void kernel_get_rows(
         device const  void * src0,
@@ -5060,6 +5525,8 @@ template [[host_name("kernel_get_rows_q6_K")]] kernel get_rows_t kernel_get_rows
 template [[host_name("kernel_get_rows_iq2_xxs")]] kernel get_rows_t kernel_get_rows<block_iq2_xxs, QK_NL, dequantize_iq2_xxs>;
 template [[host_name("kernel_get_rows_iq2_xs")]]  kernel get_rows_t kernel_get_rows<block_iq2_xs,  QK_NL, dequantize_iq2_xs>;
 template [[host_name("kernel_get_rows_iq3_xxs")]] kernel get_rows_t kernel_get_rows<block_iq3_xxs, QK_NL, dequantize_iq3_xxs>;
+template [[host_name("kernel_get_rows_iq1_s")]]   kernel get_rows_t kernel_get_rows<block_iq1_s,   QK_NL, dequantize_iq1_s>;
+template [[host_name("kernel_get_rows_iq4_nl")]]  kernel get_rows_t kernel_get_rows<block_iq4_nl,  2, dequantize_iq4_nl>;
 
 //
 // matrix-matrix multiplication
@@ -5099,6 +5566,8 @@ template [[host_name("kernel_mul_mm_q6_K_f32")]] kernel mat_mm_t kernel_mul_mm<b
 template [[host_name("kernel_mul_mm_iq2_xxs_f32")]] kernel mat_mm_t kernel_mul_mm<block_iq2_xxs, QK_NL, dequantize_iq2_xxs>;
 template [[host_name("kernel_mul_mm_iq2_xs_f32")]]  kernel mat_mm_t kernel_mul_mm<block_iq2_xs,  QK_NL, dequantize_iq2_xs>;
 template [[host_name("kernel_mul_mm_iq3_xxs_f32")]] kernel mat_mm_t kernel_mul_mm<block_iq3_xxs, QK_NL, dequantize_iq3_xxs>;
+template [[host_name("kernel_mul_mm_iq1_s_f32")]]   kernel mat_mm_t kernel_mul_mm<block_iq1_s,   QK_NL, dequantize_iq1_s>;
+template [[host_name("kernel_mul_mm_iq4_nl_f32")]]  kernel mat_mm_t kernel_mul_mm<block_iq4_nl,  2, dequantize_iq4_nl>;
 
 //
 // indirect matrix-matrix multiplication
@@ -5150,6 +5619,8 @@ template [[host_name("kernel_mul_mm_id_q6_K_f32")]] kernel mat_mm_id_t kernel_mu
 template [[host_name("kernel_mul_mm_id_iq2_xxs_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_iq2_xxs, QK_NL, dequantize_iq2_xxs>;
 template [[host_name("kernel_mul_mm_id_iq2_xs_f32")]]  kernel mat_mm_id_t kernel_mul_mm_id<block_iq2_xs,  QK_NL, dequantize_iq2_xs>;
 template [[host_name("kernel_mul_mm_id_iq3_xxs_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_iq3_xxs, QK_NL, dequantize_iq3_xxs>;
+template [[host_name("kernel_mul_mm_id_iq1_s_f32")]]   kernel mat_mm_id_t kernel_mul_mm_id<block_iq1_s,   QK_NL, dequantize_iq1_s>;
+template [[host_name("kernel_mul_mm_id_iq4_nl_f32")]]  kernel mat_mm_id_t kernel_mul_mm_id<block_iq4_nl,  2, dequantize_iq4_nl>;
 
 //
 // matrix-vector multiplication
@@ -6117,3 +6588,131 @@ kernel void kernel_mul_mv_id_iq3_xxs_f32(
         tiisg,
         sgitg);
 }
+
+[[host_name("kernel_mul_mv_id_iq1_s_f32")]]
+kernel void kernel_mul_mv_id_iq1_s_f32(
+        device const    char * ids,
+        device const    char * src1,
+        device         float * dst,
+        constant    uint64_t & nbi1,
+        constant     int64_t & ne00,
+        constant     int64_t & ne01,
+        constant     int64_t & ne02,
+        constant    uint64_t & nb00,
+        constant    uint64_t & nb01,
+        constant    uint64_t & nb02,
+        constant     int64_t & ne10,
+        constant     int64_t & ne11,
+        constant     int64_t & ne12,
+        constant     int64_t & ne13,
+        constant    uint64_t & nb10,
+        constant    uint64_t & nb11,
+        constant    uint64_t & nb12,
+        constant     int64_t & ne0,
+        constant     int64_t & ne1,
+        constant    uint64_t & nb1,
+        constant        uint & r2,
+        constant        uint & r3,
+        constant         int & idx,
+        device const    char * src00,
+        device const    char * src01,
+        device const    char * src02,
+        device const    char * src03,
+        device const    char * src04,
+        device const    char * src05,
+        device const    char * src06,
+        device const    char * src07,
+        uint3                  tgpig[[threadgroup_position_in_grid]],
+        uint                   tiitg[[thread_index_in_threadgroup]],
+        uint                   tiisg[[thread_index_in_simdgroup]],
+        uint                   sgitg[[simdgroup_index_in_threadgroup]]) {
+    device const char * src0[8] = {src00, src01, src02, src03, src04, src05, src06, src07};
+
+    const int64_t bid = tgpig.z/(ne12*ne13);
+
+    tgpig.z = tgpig.z%(ne12*ne13);
+
+    const int32_t id = ((device int32_t *) (ids + bid*nbi1))[idx];
+
+    kernel_mul_mv_iq1_s_f32_impl(
+        src0[id],
+        (device const float *) (src1 + bid*nb11),
+        dst + bid*ne0,
+        ne00,
+        ne01,
+        ne02,
+        ne10,
+        ne12,
+        ne0,
+        ne1,
+        r2,
+        r3,
+        tgpig,
+        tiisg,
+        sgitg);
+}
+
+[[host_name("kernel_mul_mv_id_iq4_nl_f32")]]
+kernel void kernel_mul_mv_id_iq4_nl_f32(
+        device const    char * ids,
+        device const    char * src1,
+        device         float * dst,
+        constant    uint64_t & nbi1,
+        constant     int64_t & ne00,
+        constant     int64_t & ne01,
+        constant     int64_t & ne02,
+        constant    uint64_t & nb00,
+        constant    uint64_t & nb01,
+        constant    uint64_t & nb02,
+        constant     int64_t & ne10,
+        constant     int64_t & ne11,
+        constant     int64_t & ne12,
+        constant     int64_t & ne13,
+        constant    uint64_t & nb10,
+        constant    uint64_t & nb11,
+        constant    uint64_t & nb12,
+        constant     int64_t & ne0,
+        constant     int64_t & ne1,
+        constant    uint64_t & nb1,
+        constant        uint & r2,
+        constant        uint & r3,
+        constant         int & idx,
+        device const    char * src00,
+        device const    char * src01,
+        device const    char * src02,
+        device const    char * src03,
+        device const    char * src04,
+        device const    char * src05,
+        device const    char * src06,
+        device const    char * src07,
+        threadgroup float    * shared_values [[threadgroup(0)]],
+        uint3                  tgpig[[threadgroup_position_in_grid]],
+        uint                   tiitg[[thread_index_in_threadgroup]],
+        uint                   tiisg[[thread_index_in_simdgroup]],
+        uint                   sgitg[[simdgroup_index_in_threadgroup]]) {
+    device const char * src0[8] = {src00, src01, src02, src03, src04, src05, src06, src07};
+
+    const int64_t bid = tgpig.z/(ne12*ne13);
+
+    tgpig.z = tgpig.z%(ne12*ne13);
+
+    const int32_t id = ((device int32_t *) (ids + bid*nbi1))[idx];
+
+    kernel_mul_mv_iq4_nl_f32_impl(
+        src0[id],
+        (device const float *) (src1 + bid*nb11),
+        dst + bid*ne0,
+        ne00,
+        ne01,
+        ne02,
+        ne10,
+        ne12,
+        ne0,
+        ne1,
+        r2,
+        r3,
+        shared_values,
+        tgpig,
+        tiisg,
+        sgitg);
+}

ggml-quants.h

@@ -191,6 +191,21 @@ 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");
 
+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 +225,7 @@ 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_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 +241,7 @@ 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);
 
 // Dequantization
 void dequantize_row_q4_0(const block_q4_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int k);
@@ -243,6 +260,8 @@ 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);
 
 // 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 +278,8 @@ 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);
 
 //
 // Quantization utilizing an importance matrix (a.k.a. "Activation aWare Quantization")
@@ -266,6 +287,8 @@ 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_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 +299,8 @@ 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-sycl.cpp

@@ -9188,174 +9188,22 @@ static void convert_mul_mat_vec_f16_sycl(const void *vx, const dfloat *y,
     }
 }
 
-static void mul_mat_vec_q4_0_q8_1_sycl(const void *vx, const void *vy,
-                                       float *dst, const int ncols,
-                                       const int nrows,
-                                       dpct::queue_ptr stream) {
-    GGML_ASSERT(ncols % QK4_0 == 0);
-    const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y;
-    const sycl::range<3> block_nums(1, 1, block_num_y);
-    const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
-    stream->parallel_for(
-        sycl::nd_range<3>(block_nums * block_dims, block_dims),
-        [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] {
-            mul_mat_vec_q<QK4_0, QI4_0, block_q4_0, VDR_Q4_0_Q8_1_MMVQ,
-                          vec_dot_q4_0_q8_1>(vx, vy, dst, ncols, nrows,
-                                             item_ct1);
-        });
-}
-
-static void mul_mat_vec_q4_1_q8_1_sycl(const void *vx, const void *vy,
-                                       float *dst, const int ncols,
-                                       const int nrows,
-                                       dpct::queue_ptr stream) {
-    GGML_ASSERT(ncols % QK4_1 == 0);
-    const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y;
-    const sycl::range<3> block_nums(1, 1, block_num_y);
-    const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
-    stream->parallel_for(
-        sycl::nd_range<3>(block_nums * block_dims, block_dims),
-        [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] {
-            mul_mat_vec_q<QK4_0, QI4_1, block_q4_1, VDR_Q4_1_Q8_1_MMVQ,
-                          vec_dot_q4_1_q8_1>(vx, vy, dst, ncols, nrows,
-                                             item_ct1);
-        });
-}
-
-static void mul_mat_vec_q5_0_q8_1_sycl(const void *vx, const void *vy,
-                                       float *dst, const int ncols,
-                                       const int nrows,
-                                       dpct::queue_ptr stream) {
-    GGML_ASSERT(ncols % QK5_0 == 0);
-    const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y;
-    const sycl::range<3> block_nums(1, 1, block_num_y);
-    const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
-    stream->parallel_for(
-        sycl::nd_range<3>(block_nums * block_dims, block_dims),
-        [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] {
-            mul_mat_vec_q<QK5_0, QI5_0, block_q5_0, VDR_Q5_0_Q8_1_MMVQ,
-                          vec_dot_q5_0_q8_1>(vx, vy, dst, ncols, nrows,
-                                             item_ct1);
-        });
-}
-
-static void mul_mat_vec_q5_1_q8_1_sycl(const void *vx, const void *vy,
-                                       float *dst, const int ncols,
-                                       const int nrows,
-                                       dpct::queue_ptr stream) {
-    GGML_ASSERT(ncols % QK5_1 == 0);
-    const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y;
-    const sycl::range<3> block_nums(1, 1, block_num_y);
-    const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
-    stream->parallel_for(
-        sycl::nd_range<3>(block_nums * block_dims, block_dims),
-        [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] {
-            mul_mat_vec_q<QK5_1, QI5_1, block_q5_1, VDR_Q5_1_Q8_1_MMVQ,
-                          vec_dot_q5_1_q8_1>(vx, vy, dst, ncols, nrows,
-                                             item_ct1);
-        });
-}
-
-static void mul_mat_vec_q8_0_q8_1_sycl(const void *vx, const void *vy,
-                                       float *dst, const int ncols,
-                                       const int nrows,
-                                       dpct::queue_ptr stream) {
-    GGML_ASSERT(ncols % QK8_0 == 0);
-    const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y;
-    const sycl::range<3> block_nums(1, 1, block_num_y);
-    const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
-    stream->parallel_for(
-        sycl::nd_range<3>(block_nums * block_dims, block_dims),
-        [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] {
-            mul_mat_vec_q<QK8_0, QI8_0, block_q8_0, VDR_Q8_0_Q8_1_MMVQ,
-                          vec_dot_q8_0_q8_1>(vx, vy, dst, ncols, nrows,
-                                             item_ct1);
-        });
-}
-
-static void mul_mat_vec_q2_K_q8_1_sycl(const void *vx, const void *vy,
-                                       float *dst, const int ncols,
-                                       const int nrows,
-                                       dpct::queue_ptr stream) {
-    GGML_ASSERT(ncols % QK_K == 0);
-    const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y;
-    const sycl::range<3> block_nums(1, 1, block_num_y);
-    const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
-    stream->parallel_for(
-        sycl::nd_range<3>(block_nums * block_dims, block_dims),
-        [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] {
-            mul_mat_vec_q<QK_K, QI2_K, block_q2_K, VDR_Q2_K_Q8_1_MMVQ,
-                          vec_dot_q2_K_q8_1>(vx, vy, dst, ncols, nrows,
-                                             item_ct1);
-        });
-}
-
-static void mul_mat_vec_q3_K_q8_1_sycl(const void *vx, const void *vy,
-                                       float *dst, const int ncols,
-                                       const int nrows,
-                                       dpct::queue_ptr stream) {
-    GGML_ASSERT(ncols % QK_K == 0);
-    const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y;
-    const sycl::range<3> block_nums(1, 1, block_num_y);
-    const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
-    stream->parallel_for(
-        sycl::nd_range<3>(block_nums * block_dims, block_dims),
-        [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] {
-            mul_mat_vec_q<QK_K, QI3_K, block_q3_K, VDR_Q3_K_Q8_1_MMVQ,
-                          vec_dot_q3_K_q8_1>(vx, vy, dst, ncols, nrows,
-                                             item_ct1);
-        });
-}
-
-static void mul_mat_vec_q4_K_q8_1_sycl(const void *vx, const void *vy,
-                                       float *dst, const int ncols,
-                                       const int nrows,
-                                       dpct::queue_ptr stream) {
-    GGML_ASSERT(ncols % QK_K == 0);
-    const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y;
-    const sycl::range<3> block_nums(1, 1, block_num_y);
-    const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
-    stream->parallel_for(
-        sycl::nd_range<3>(block_nums * block_dims, block_dims),
-        [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] {
-            mul_mat_vec_q<QK_K, QI4_K, block_q4_K, VDR_Q4_K_Q8_1_MMVQ,
-                          vec_dot_q4_K_q8_1>(vx, vy, dst, ncols, nrows,
-                                             item_ct1);
-        });
-}
-
-static void mul_mat_vec_q5_K_q8_1_sycl(const void *vx, const void *vy,
-                                       float *dst, const int ncols,
-                                       const int nrows,
-                                       dpct::queue_ptr stream) {
-    GGML_ASSERT(ncols % QK_K == 0);
-    const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y;
-    const sycl::range<3> block_nums(1, 1, block_num_y);
-    const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
-    stream->parallel_for(
-        sycl::nd_range<3>(block_nums * block_dims, block_dims),
-        [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] {
-            mul_mat_vec_q<QK_K, QI5_K, block_q5_K, VDR_Q5_K_Q8_1_MMVQ,
-                          vec_dot_q5_K_q8_1>(vx, vy, dst, ncols, nrows,
-                                             item_ct1);
-        });
-}
-
-static void mul_mat_vec_q6_K_q8_1_sycl(const void *vx, const void *vy,
-                                       float *dst, const int ncols,
-                                       const int nrows,
-                                       dpct::queue_ptr stream) {
-    GGML_ASSERT(ncols % QK_K == 0);
-    const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y;
-    const sycl::range<3> block_nums(1, 1, block_num_y);
-    const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
-    stream->parallel_for(
-        sycl::nd_range<3>(block_nums * block_dims, block_dims),
-        [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] {
-            mul_mat_vec_q<QK_K, QI6_K, block_q6_K, VDR_Q6_K_Q8_1_MMVQ,
-                          vec_dot_q6_K_q8_1>(vx, vy, dst, ncols, nrows,
-                                             item_ct1);
-        });
+template <int qk, int qi, typename block_q_t, int vdr,
+          vec_dot_q_sycl_t vec_dot_q_sycl>
+static void mul_mat_vec_q_sycl_submitter(const void *vx, const void *vy,
+                                         float *dst, const int ncols,
+                                         const int nrows,
+                                         dpct::queue_ptr stream) {
+  GGML_ASSERT(ncols % QK4_0 == 0);
+  const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y;
+  const sycl::range<3> block_nums(1, 1, block_num_y);
+  const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
+  stream->parallel_for(
+      sycl::nd_range<3>(block_nums * block_dims, block_dims), [=
+  ](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] {
+        mul_mat_vec_q<qk, qi, block_q_t, vdr, vec_dot_q_sycl>(
+            vx, vy, dst, ncols, nrows, item_ct1);
+      });
 }
 
 int get_device_index_by_id(int id){
@@ -12095,37 +11943,63 @@ inline void ggml_sycl_op_mul_mat_vec_q(
     const int64_t ne00 = src0->ne[0];
     const int64_t row_diff = row_high - row_low;
 
+    // TODO: support these quantization types
+    GGML_ASSERT(!(src0->type == GGML_TYPE_IQ2_XXS ||
+                  src0->type == GGML_TYPE_IQ2_XS ||
+                  src0->type == GGML_TYPE_IQ3_XXS ||
+                  src0->type == GGML_TYPE_IQ1_S));
+
     switch (src0->type) {
         case GGML_TYPE_Q4_0:
-            mul_mat_vec_q4_0_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream);
-            break;
+          mul_mat_vec_q_sycl_submitter<QK4_0, QI4_0, block_q4_0,
+                                       VDR_Q4_0_Q8_1_MMVQ, vec_dot_q4_0_q8_1>(
+              src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream);
+          break;
         case GGML_TYPE_Q4_1:
-            mul_mat_vec_q4_1_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream);
-            break;
+          mul_mat_vec_q_sycl_submitter<QK4_1, QI4_1, block_q4_1,
+                                       VDR_Q4_1_Q8_1_MMVQ, vec_dot_q4_1_q8_1>(
+              src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream);
+          break;
         case GGML_TYPE_Q5_0:
-            mul_mat_vec_q5_0_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream);
-            break;
+          mul_mat_vec_q_sycl_submitter<QK5_0, QI5_0, block_q5_0,
+                                       VDR_Q5_0_Q8_1_MMVQ, vec_dot_q5_0_q8_1>(
+              src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream);
+          break;
         case GGML_TYPE_Q5_1:
-            mul_mat_vec_q5_1_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream);
-            break;
+          mul_mat_vec_q_sycl_submitter<QK5_1, QI5_1, block_q5_1,
+                                       VDR_Q5_1_Q8_1_MMVQ, vec_dot_q5_1_q8_1>(
+              src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream);
+          break;
         case GGML_TYPE_Q8_0:
-            mul_mat_vec_q8_0_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream);
-            break;
+          mul_mat_vec_q_sycl_submitter<QK8_0, QI8_0, block_q8_0,
+                                       VDR_Q8_0_Q8_1_MMVQ, vec_dot_q8_0_q8_1>(
+              src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream);
+          break;
         case GGML_TYPE_Q2_K:
-            mul_mat_vec_q2_K_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream);
-            break;
+          mul_mat_vec_q_sycl_submitter<QK_K, QI2_K, block_q2_K,
+                                       VDR_Q2_K_Q8_1_MMVQ, vec_dot_q2_K_q8_1>(
+              src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream);
+          break;
         case GGML_TYPE_Q3_K:
-            mul_mat_vec_q3_K_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream);
-            break;
+          mul_mat_vec_q_sycl_submitter<QK_K, QI3_K, block_q3_K,
+                                       VDR_Q3_K_Q8_1_MMVQ, vec_dot_q3_K_q8_1>(
+              src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream);
+          break;
         case GGML_TYPE_Q4_K:
-            mul_mat_vec_q4_K_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream);
-            break;
+          mul_mat_vec_q_sycl_submitter<QK_K, QI4_K, block_q4_K,
+                                       VDR_Q4_K_Q8_1_MMVQ, vec_dot_q4_K_q8_1>(
+              src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream);
+          break;
         case GGML_TYPE_Q5_K:
-            mul_mat_vec_q5_K_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream);
-            break;
+          mul_mat_vec_q_sycl_submitter<QK_K, QI5_K, block_q5_K,
+                                       VDR_Q5_K_Q8_1_MMVQ, vec_dot_q5_K_q8_1>(
+              src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream);
+          break;
         case GGML_TYPE_Q6_K:
-            mul_mat_vec_q6_K_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream);
-            break;
+          mul_mat_vec_q_sycl_submitter<QK_K, QI6_K, block_q6_K,
+                                       VDR_Q6_K_Q8_1_MMVQ, vec_dot_q6_K_q8_1>(
+              src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream);
+          break;
         default:
             GGML_ASSERT(false);
             break;
@@ -12145,7 +12019,7 @@ inline void ggml_sycl_op_dequantize_mul_mat_vec(
     const int64_t src1_ncols, const int64_t src1_padded_row_size,
     const dpct::queue_ptr &stream) {
 
-    GGML_TENSOR_BINARY_OP_LOCALS
+    GGML_TENSOR_BINARY_OP_LOCALS;
 
     const int64_t row_diff = row_high - row_low;
 
@@ -14768,7 +14642,8 @@ GGML_CALL static const char * ggml_backend_sycl_buffer_type_name(ggml_backend_bu
 static ggml_backend_buffer_t
 ggml_backend_sycl_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft,
                                            size_t size) try {
-    int device = (int) (intptr_t) buft->context;
+    ggml_backend_sycl_buffer_type_context * buft_ctx = (ggml_backend_sycl_buffer_type_context *)buft->context;
+    int device = (int) buft_ctx->device;
 
     ggml_sycl_set_device(device);
     int device_index = get_device_index_by_id(device);
@@ -14846,7 +14721,7 @@ ggml_backend_buffer_type_t ggml_backend_sycl_buffer_type(int device) {
         for (int i = 0; i < GGML_SYCL_MAX_DEVICES; i++) {
             ggml_backend_sycl_buffer_types[i] = {
                 /* .iface    = */ ggml_backend_sycl_buffer_type_interface,
-                /* .context  = */ (ggml_backend_buffer_type_context_t) (intptr_t) i,
+                /* .context  = */ new ggml_backend_sycl_buffer_type_context{i, GGML_SYCL_NAME + std::to_string(i)},
             };
         }
         ggml_backend_sycl_buffer_type_initialized = true;
@@ -14908,10 +14783,6 @@ ggml_backend_buffer_type_t ggml_backend_sycl_host_buffer_type() {
 
 // backend
 
-struct ggml_backend_context_sycl {
-    int device;
-};
-
 static const char * ggml_backend_sycl_name(ggml_backend_t backend) {
     return GGML_SYCL_NAME;
 
@@ -14919,14 +14790,14 @@ static const char * ggml_backend_sycl_name(ggml_backend_t backend) {
 }
 
 static void ggml_backend_sycl_free(ggml_backend_t backend) {
-    ggml_backend_context_sycl * sycl_ctx = (ggml_backend_context_sycl *)backend->context;
+    ggml_backend_sycl_context * sycl_ctx = (ggml_backend_sycl_context *)backend->context;
 
     delete sycl_ctx;
     delete backend;
 }
 
 static ggml_backend_buffer_type_t ggml_backend_sycl_get_default_buffer_type(ggml_backend_t backend) {
-    ggml_backend_context_sycl * sycl_ctx = (ggml_backend_context_sycl *)backend->context;
+    ggml_backend_sycl_context * sycl_ctx = (ggml_backend_sycl_context *)backend->context;
 
     return ggml_backend_sycl_buffer_type(sycl_ctx->device);
 }
@@ -14935,7 +14806,7 @@ static void ggml_backend_sycl_set_tensor_async(ggml_backend_t backend,
                                                ggml_tensor *tensor,
                                                const void *data, size_t offset,
                                                size_t size) try {
-    ggml_backend_context_sycl * sycl_ctx = (ggml_backend_context_sycl *)backend->context;
+    ggml_backend_sycl_context * sycl_ctx = (ggml_backend_sycl_context *)backend->context;
 
     GGML_ASSERT(tensor->buffer->buft == ggml_backend_sycl_buffer_type(sycl_ctx->device) && "unsupported buffer type");
     GGML_ASSERT(tensor->backend == GGML_BACKEND_GPU);
@@ -14953,7 +14824,7 @@ static void ggml_backend_sycl_get_tensor_async(ggml_backend_t backend,
                                                const ggml_tensor *tensor,
                                                void *data, size_t offset,
                                                size_t size) try {
-    ggml_backend_context_sycl * sycl_ctx = (ggml_backend_context_sycl *)backend->context;
+    ggml_backend_sycl_context * sycl_ctx = (ggml_backend_sycl_context *)backend->context;
 
     GGML_ASSERT(tensor->buffer->buft == ggml_backend_sycl_buffer_type(sycl_ctx->device) && "unsupported buffer type");
     GGML_ASSERT(tensor->backend == GGML_BACKEND_GPU);
@@ -14968,7 +14839,7 @@ catch (sycl::exception const &exc) {
 }
 
 static void ggml_backend_sycl_synchronize(ggml_backend_t backend) try {
-    ggml_backend_context_sycl * sycl_ctx = (ggml_backend_context_sycl *)backend->context;
+    ggml_backend_sycl_context * sycl_ctx = (ggml_backend_sycl_context *)backend->context;
 
     SYCL_CHECK(CHECK_TRY_ERROR(g_syclStreams[sycl_ctx->device][0]->wait()));
 
@@ -15004,7 +14875,7 @@ static void ggml_backend_sycl_graph_plan_compute(ggml_backend_t backend, ggml_ba
 }
 
 static bool ggml_backend_sycl_graph_compute(ggml_backend_t backend, ggml_cgraph * cgraph) {
-    ggml_backend_context_sycl * sycl_ctx = (ggml_backend_context_sycl *)backend->context;
+    ggml_backend_sycl_context * sycl_ctx = (ggml_backend_sycl_context *)backend->context;
 
     ggml_sycl_set_main_device(sycl_ctx->device);
 
@@ -15093,6 +14964,12 @@ static bool ggml_backend_sycl_supports_op(ggml_backend_t backend, const ggml_ten
                     return false;
                 }
 
+                if (a->type == GGML_TYPE_IQ1_S) {
+                    return false;
+                }
+                if (a->type == GGML_TYPE_IQ3_XXS) {
+                  return false;
+                }
                 if (a->type == GGML_TYPE_IQ2_XXS) {
                     return false;
                 }
@@ -15212,8 +15089,9 @@ ggml_backend_t ggml_backend_sycl_init(int device) {
     // not strictly necessary, but it may reduce the overhead of the first graph_compute
     ggml_sycl_set_main_device(device);
 
-    ggml_backend_context_sycl * ctx = new ggml_backend_context_sycl {
-        /* .device = */ device
+    ggml_backend_sycl_context * ctx = new ggml_backend_sycl_context {
+        /* .device = */ device,
+        /* .name   = */ GGML_SYCL_NAME + std::to_string(device),
     };
 
     ggml_backend_t sycl_backend = new ggml_backend {

ggml-vulkan.cpp

@@ -707,9 +707,21 @@ static void ggml_vk_queue_cleanup(ggml_backend_vk_context * ctx, vk_queue& q) {
     q.cmd_buffer_idx = 0;
 }
 
-static vk_buffer ggml_vk_create_buffer(ggml_backend_vk_context * ctx, size_t size, vk::MemoryPropertyFlags req_flags) {
+static uint32_t find_properties(const vk::PhysicalDeviceMemoryProperties* mem_props, vk::MemoryRequirements* mem_req, vk::MemoryPropertyFlags flags) {
+    for (uint32_t i = 0; i < mem_props->memoryTypeCount; ++i) {
+        vk::MemoryType memory_type = mem_props->memoryTypes[i];
+        if ((mem_req->memoryTypeBits & ((uint64_t)1 << i)) &&
+            (flags & memory_type.propertyFlags) == flags &&
+            mem_props->memoryHeaps[memory_type.heapIndex].size >= mem_req->size) {
+            return static_cast<int32_t>(i);
+        }
+    }
+    return UINT32_MAX;
+}
+
+static vk_buffer ggml_vk_create_buffer(ggml_backend_vk_context * ctx, size_t size, vk::MemoryPropertyFlags req_flags, vk::MemoryPropertyFlags fallback_flags = vk::MemoryPropertyFlags(0)) {
 #ifdef GGML_VULKAN_DEBUG
-    std::cerr << "ggml_vk_create_buffer(" << size << ", " << to_string(req_flags) << ")" << std::endl;
+    std::cerr << "ggml_vk_create_buffer(" << size << ", " << to_string(req_flags) << ", " << to_string(fallback_flags) << ")" << std::endl;
 #endif
     vk_buffer buf = std::make_shared<vk_buffer_struct>();
 
@@ -736,15 +748,15 @@ static vk_buffer ggml_vk_create_buffer(ggml_backend_vk_context * ctx, size_t siz
 
     uint32_t memory_type_index = UINT32_MAX;
 
-    for (uint32_t i = 0; i < mem_props.memoryTypeCount; ++i) {
-        vk::MemoryType memory_type = mem_props.memoryTypes[i];
-        if ((mem_req.memoryTypeBits & ((uint64_t)1 << i)) && (req_flags & memory_type.propertyFlags) == req_flags && mem_props.memoryHeaps[memory_type.heapIndex].size >= mem_req.size) {
-            memory_type_index = i;
-            break;
-        }
+    memory_type_index = find_properties(&mem_props, &mem_req, req_flags);
+    buf->memory_property_flags = req_flags;
+
+    if (memory_type_index == UINT32_MAX && fallback_flags) {
+        memory_type_index = find_properties(&mem_props, &mem_req, fallback_flags);
+        buf->memory_property_flags = fallback_flags;
     }
 
-    if (memory_type_index >= mem_props.memoryTypeCount) {
+    if (memory_type_index == UINT32_MAX) {
         ctx->device.lock()->device.destroyBuffer(buf->buffer);
         buf->size = 0;
         throw vk::OutOfDeviceMemoryError("No suitable memory type found");
@@ -758,10 +770,9 @@ static vk_buffer ggml_vk_create_buffer(ggml_backend_vk_context * ctx, size_t siz
         buf->size = 0;
         throw e;
     }
-    buf->memory_property_flags = req_flags;
     buf->ptr = nullptr;
 
-    if (req_flags & vk::MemoryPropertyFlagBits::eHostVisible) {
+    if (buf->memory_property_flags & vk::MemoryPropertyFlagBits::eHostVisible) {
         buf->ptr = ctx->device.lock()->device.mapMemory(buf->device_memory, 0, VK_WHOLE_SIZE);
     }
 
@@ -778,9 +789,9 @@ static vk_buffer ggml_vk_create_buffer(ggml_backend_vk_context * ctx, size_t siz
     return buf;
 }
 
-static vk_buffer ggml_vk_create_buffer_check(ggml_backend_vk_context * ctx, size_t size, vk::MemoryPropertyFlags req_flags) {
+static vk_buffer ggml_vk_create_buffer_check(ggml_backend_vk_context * ctx, size_t size, vk::MemoryPropertyFlags req_flags, vk::MemoryPropertyFlags fallback_flags = vk::MemoryPropertyFlags(0)) {
     try {
-        return ggml_vk_create_buffer(ctx, size, req_flags);
+        return ggml_vk_create_buffer(ctx, size, req_flags, fallback_flags);
     } catch (const vk::SystemError& e) {
         std::cerr << "ggml_vulkan: Memory allocation of size " << size << " failed." << std::endl;
         std::cerr << "ggml_vulkan: " << e.what() << std::endl;
@@ -791,16 +802,16 @@ static vk_buffer ggml_vk_create_buffer_check(ggml_backend_vk_context * ctx, size
 static vk_buffer ggml_vk_create_buffer_device(ggml_backend_vk_context * ctx, size_t size) {
     vk_buffer buf;
     try {
-        buf = ggml_vk_create_buffer(ctx, size, vk::MemoryPropertyFlagBits::eDeviceLocal);
-    } catch (const vk::SystemError& e) {
         if (ctx->device.lock()->uma) {
             // Fall back to host memory type
-            buf = ggml_vk_create_buffer_check(ctx, size, vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent);
+            buf = ggml_vk_create_buffer(ctx, size, vk::MemoryPropertyFlagBits::eDeviceLocal, vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent);
         } else {
-            std::cerr << "ggml_vulkan: Device memory allocation of size " << size << " failed." << std::endl;
-            std::cerr << "ggml_vulkan: " << e.what() << std::endl;
-            throw e;
+            buf = ggml_vk_create_buffer(ctx, size, vk::MemoryPropertyFlagBits::eDeviceLocal);
         }
+    } catch (const vk::SystemError& e) {
+        std::cerr << "ggml_vulkan: Device memory allocation of size " << size << " failed." << std::endl;
+        std::cerr << "ggml_vulkan: " << e.what() << std::endl;
+        throw e;
     }
 
     return buf;
@@ -1080,6 +1091,9 @@ static void ggml_vk_print_gpu_info(size_t idx) {
     }
 }
 
+static bool ggml_vk_instance_validation_ext_available(const std::vector<vk::ExtensionProperties>& instance_extensions);
+static bool ggml_vk_instance_portability_enumeration_ext_available(const std::vector<vk::ExtensionProperties>& instance_extensions);
+
 void ggml_vk_instance_init() {
     if (vk_instance_initialized) {
         return;
@@ -1089,28 +1103,42 @@ 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);
@@ -1139,7 +1167,7 @@ void ggml_vk_instance_init() {
     vk_instance_initialized = true;
 }
 
-void ggml_vk_init(ggml_backend_vk_context * ctx, size_t idx) {
+static void ggml_vk_init(ggml_backend_vk_context * ctx, size_t idx) {
     GGML_ASSERT(idx < vk_instance.device_indices.size());
     size_t dev_num = vk_instance.device_indices[idx];
 #ifdef GGML_VULKAN_DEBUG
@@ -1157,12 +1185,12 @@ 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;
         }
@@ -1193,7 +1221,7 @@ 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) {
@@ -1422,7 +1450,9 @@ static void * ggml_vk_host_malloc(ggml_backend_vk_context * ctx, size_t size) {
 #ifdef GGML_VULKAN_DEBUG
     std::cerr << "ggml_vk_host_malloc(" << size << ")" << std::endl;
 #endif
-    vk_buffer buf = ggml_vk_create_buffer(ctx, size, vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent | vk::MemoryPropertyFlagBits::eHostCached);
+    vk_buffer buf = ggml_vk_create_buffer(ctx, size,
+        vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent | vk::MemoryPropertyFlagBits::eHostCached,
+        vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent);
 
     if(!(buf->memory_property_flags & vk::MemoryPropertyFlagBits::eHostVisible)) {
         fprintf(stderr, "WARNING: failed to allocate %.2f MB of pinned memory\n",
@@ -1568,7 +1598,9 @@ static void deferred_memcpy(void * dst, const void * src, size_t size, std::vect
 static void ggml_vk_ensure_sync_staging_buffer(ggml_backend_vk_context * ctx, size_t size) {
     if (ctx->sync_staging == nullptr || ctx->sync_staging->size < size) {
         ggml_vk_destroy_buffer(ctx->sync_staging);
-        ctx->sync_staging = ggml_vk_create_buffer_check(ctx, size, vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent | vk::MemoryPropertyFlagBits::eHostCached);
+        ctx->sync_staging = ggml_vk_create_buffer_check(ctx, size,
+            vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent | vk::MemoryPropertyFlagBits::eHostCached,
+            vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent);
     }
 }
 
@@ -4082,7 +4114,9 @@ static void ggml_vk_preallocate_buffers(ggml_backend_vk_context * ctx) {
     std::cerr << "ggml_vk_preallocate_buffers(qx_size: " << ctx->prealloc_size_qx << " qy_size: " << ctx->prealloc_size_qy << " x_size: " << ctx->prealloc_size_x << " y_size: " << ctx->prealloc_size_y << " split_k_size: " << ctx->prealloc_size_split_k << ")" << std::endl;
 #endif
 #if defined(GGML_VULKAN_RUN_TESTS)
-    ctx->staging = ggml_vk_create_buffer_check(ctx, 100ul * 1024ul * 1024ul, vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent | vk::MemoryPropertyFlagBits::eHostCached);
+    ctx->staging = ggml_vk_create_buffer_check(ctx, 100ul * 1024ul * 1024ul,
+        vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent | vk::MemoryPropertyFlagBits::eHostCached
+        vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent);
     ggml_vk_test_transfer(ctx, 8192 * 1000, false);
     ggml_vk_test_transfer(ctx, 8192 * 1000, true);
 
@@ -4174,7 +4208,9 @@ static void ggml_vk_preallocate_buffers(ggml_backend_vk_context * ctx) {
         if (ctx->staging != nullptr) {
             ggml_vk_destroy_buffer(ctx->staging);
         }
-        ctx->staging = ggml_vk_create_buffer_check(ctx, ctx->staging_size, vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent | vk::MemoryPropertyFlagBits::eHostCached);
+        ctx->staging = ggml_vk_create_buffer_check(ctx, ctx->staging_size,
+            vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent | vk::MemoryPropertyFlagBits::eHostCached,
+            vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent);
     }
 }
 
@@ -4537,13 +4573,13 @@ static void ggml_vk_cleanup(ggml_backend_vk_context * ctx) {
     }
 }
 
-GGML_CALL int ggml_vk_get_device_count() {
+GGML_CALL static int ggml_vk_get_device_count() {
     ggml_vk_instance_init();
 
     return vk_instance.device_indices.size();
 }
 
-GGML_CALL void ggml_vk_get_device_description(int device, char * description, size_t description_size) {
+GGML_CALL static void ggml_vk_get_device_description(int device, char * description, size_t description_size) {
     ggml_vk_instance_init();
 
     std::vector<vk::PhysicalDevice> devices = vk_instance.instance.enumeratePhysicalDevices();
@@ -4561,7 +4597,7 @@ void ggml_vk_init_cpu_assist() {
 
     std::cerr << "ggml_vulkan: Found " << ggml_vk_get_device_count() << " Vulkan devices:" << std::endl;
 
-    for (size_t i = 0; i < ggml_vk_get_device_count(); i++) {
+    for (int i = 0; i < ggml_vk_get_device_count(); i++) {
         ggml_vk_print_gpu_info(i);
     }
     // Initialize the first backend to make sure CPU matrix multiplications can be offloaded.
@@ -5248,7 +5284,7 @@ GGML_CALL void ggml_backend_vk_get_device_description(int device, char * descrip
 }
 
 GGML_CALL void ggml_backend_vk_get_device_memory(int device, size_t * free, size_t * total) {
-    GGML_ASSERT(device < vk_instance.device_indices.size());
+    GGML_ASSERT(device < (int) vk_instance.device_indices.size());
 
     vk::PhysicalDevice vkdev = vk_instance.instance.enumeratePhysicalDevices()[vk_instance.device_indices[device]];
 
@@ -5282,6 +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

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 +348,8 @@ 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_I8,
         GGML_TYPE_I16,
         GGML_TYPE_I32,
@@ -391,6 +387,8 @@ 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
     };
 
     // available tensor operations:
@@ -658,6 +656,16 @@ extern "C" {
         void * wdata;
     };
 
+    // numa strategies
+    enum ggml_numa_strategy {
+        GGML_NUMA_STRATEGY_DISABLED   = 0,
+        GGML_NUMA_STRATEGY_DISTRIBUTE = 1,
+        GGML_NUMA_STRATEGY_ISOLATE    = 2,
+        GGML_NUMA_STRATEGY_NUMACTL    = 3,
+        GGML_NUMA_STRATEGY_MIRROR     = 4,
+        GGML_NUMA_STRATEGY_COUNT
+    };
+
     // misc
 
     GGML_API void    ggml_time_init(void); // call this once at the beginning of the program
@@ -668,7 +676,7 @@ extern "C" {
 
     GGML_API void    ggml_print_backtrace(void);
 
-    GGML_API void    ggml_numa_init(void); // call once for better performance on NUMA systems
+    GGML_API void    ggml_numa_init(enum ggml_numa_strategy numa); // call once for better performance on NUMA systems
     GGML_API bool    ggml_is_numa(void); // true if init detected that system has >1 NUMA node
 
     GGML_API void    ggml_print_object (const struct ggml_object * obj);
@@ -1373,13 +1381,17 @@ 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,
@@ -1481,12 +1493,13 @@ 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)

gguf-py/examples/reader.py

@@ -0,0 +1,45 @@
+#!/usr/bin/env python3
+import sys
+from pathlib import Path
+from gguf.gguf_reader import GGUFReader
+
+
+sys.path.insert(0, str(Path(__file__).parent.parent))
+
+
+def read_gguf_file(gguf_file_path):
+    """
+    Reads and prints key-value pairs and tensor information from a GGUF file in an improved format.
+
+    Parameters:
+    - gguf_file_path: Path to the GGUF file.
+    """
+
+    reader = GGUFReader(gguf_file_path)
+
+    # List all key-value pairs in a columnized format
+    print("Key-Value Pairs:")
+    max_key_length = max(len(key) for key in reader.fields.keys())
+    for key, field in reader.fields.items():
+        value = field.parts[field.data[0]]
+        print(f"{key:{max_key_length}} : {value}")
+    print("----")
+
+    # List all tensors
+    print("Tensors:")
+    tensor_info_format = "{:<30} | Shape: {:<15} | Size: {:<12} | Quantization: {}"
+    print(tensor_info_format.format("Tensor Name", "Shape", "Size", "Quantization"))
+    print("-" * 80)
+    for tensor in reader.tensors:
+        shape_str = "x".join(map(str, tensor.shape))
+        size_str = str(tensor.n_elements)
+        quantization_str = tensor.tensor_type.name
+        print(tensor_info_format.format(tensor.name, shape_str, size_str, quantization_str))
+
+
+if __name__ == '__main__':
+    if len(sys.argv) < 2:
+        print("Usage: reader.py <path_to_gguf_file>")
+        sys.exit(1)
+    gguf_file_path = sys.argv[1]
+    read_gguf_file(gguf_file_path)

gguf-py/gguf/constants.py

@@ -40,6 +40,7 @@ class Keys:
         TENSOR_DATA_LAYOUT    = "{arch}.tensor_data_layout"
         EXPERT_COUNT          = "{arch}.expert_count"
         EXPERT_USED_COUNT     = "{arch}.expert_used_count"
+        POOLING_TYPE          = "{arch}.pooling_type"
 
     class Attention:
         HEAD_COUNT        = "{arch}.attention.head_count"
@@ -72,6 +73,8 @@ class Keys:
         UNK_ID           = "tokenizer.ggml.unknown_token_id"
         SEP_ID           = "tokenizer.ggml.seperator_token_id"
         PAD_ID           = "tokenizer.ggml.padding_token_id"
+        CLS_ID           = "tokenizer.ggml.cls_token_id"
+        MASK_ID          = "tokenizer.ggml.mask_token_id"
         ADD_BOS          = "tokenizer.ggml.add_bos_token"
         ADD_EOS          = "tokenizer.ggml.add_eos_token"
         ADD_PREFIX       = "tokenizer.ggml.add_space_prefix"
@@ -86,27 +89,29 @@ class Keys:
 
 
 class MODEL_ARCH(IntEnum):
-    LLAMA     = auto()
-    FALCON    = auto()
-    BAICHUAN  = auto()
-    GPT2      = auto()
-    GPTJ      = auto()
-    GPTNEOX   = auto()
-    MPT       = auto()
-    STARCODER = auto()
-    PERSIMMON = auto()
-    REFACT    = auto()
-    BERT      = auto()
-    BLOOM     = auto()
-    STABLELM  = auto()
-    QWEN      = auto()
-    QWEN2     = auto()
-    PHI2      = auto()
-    PLAMO     = auto()
-    CODESHELL = auto()
-    ORION     = auto()
+    LLAMA      = auto()
+    FALCON     = auto()
+    BAICHUAN   = auto()
+    GPT2       = auto()
+    GPTJ       = auto()
+    GPTNEOX    = auto()
+    MPT        = auto()
+    STARCODER  = auto()
+    PERSIMMON  = auto()
+    REFACT     = auto()
+    BERT       = auto()
+    NOMIC_BERT = auto()
+    BLOOM      = auto()
+    STABLELM   = auto()
+    QWEN       = auto()
+    QWEN2      = auto()
+    PHI2       = auto()
+    PLAMO      = auto()
+    CODESHELL  = auto()
+    ORION      = auto()
     INTERNLM2  = auto()
-    MINICPM   = auto()
+    MINICPM    = auto()
+    GEMMA      = auto()
 
 
 class MODEL_TENSOR(IntEnum):
@@ -152,6 +157,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
     MODEL_ARCH.PERSIMMON:      "persimmon",
     MODEL_ARCH.REFACT:         "refact",
     MODEL_ARCH.BERT:           "bert",
+    MODEL_ARCH.NOMIC_BERT:     "nomic-bert",
     MODEL_ARCH.BLOOM:          "bloom",
     MODEL_ARCH.STABLELM:       "stablelm",
     MODEL_ARCH.QWEN:           "qwen",
@@ -162,6 +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] = {
@@ -281,6 +288,20 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.FFN_UP,
         MODEL_TENSOR.LAYER_OUT_NORM,
     ],
+    MODEL_ARCH.NOMIC_BERT: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.TOKEN_EMBD_NORM,
+        MODEL_TENSOR.TOKEN_TYPES,
+        MODEL_TENSOR.POS_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.ATTN_OUT_NORM,
+        MODEL_TENSOR.ATTN_QKV,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.FFN_GATE,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+        MODEL_TENSOR.LAYER_OUT_NORM,
+    ],
     MODEL_ARCH.MPT: [
         MODEL_TENSOR.TOKEN_EMBD,
         MODEL_TENSOR.OUTPUT_NORM,
@@ -492,6 +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
 }
 
@@ -542,6 +576,12 @@ class RopeScalingType(Enum):
     YARN   = 'yarn'
 
 
+class PoolingType(IntEnum):
+    NONE = 0
+    MEAN = 1
+    CLS  = 2
+
+
 class GGMLQuantizationType(IntEnum):
     F32  = 0
     F16  = 1
@@ -668,5 +708,7 @@ KEY_TOKENIZER_EOS_ID     = Keys.Tokenizer.EOS_ID
 KEY_TOKENIZER_UNK_ID     = Keys.Tokenizer.UNK_ID
 KEY_TOKENIZER_SEP_ID     = Keys.Tokenizer.SEP_ID
 KEY_TOKENIZER_PAD_ID     = Keys.Tokenizer.PAD_ID
+KEY_TOKENIZER_CLS_ID     = Keys.Tokenizer.CLS_ID
+KEY_TOKENIZER_MASK_ID    = Keys.Tokenizer.MASK_ID
 KEY_TOKENIZER_HF_JSON    = Keys.Tokenizer.HF_JSON
 KEY_TOKENIZER_RWKV       = Keys.Tokenizer.RWKV

gguf-py/gguf/gguf_writer.py

@@ -19,6 +19,7 @@ from .constants import (
     GGUFValueType,
     Keys,
     RopeScalingType,
+    PoolingType,
     TokenType,
 )
 
@@ -360,6 +361,9 @@ class GGUFWriter:
     def add_causal_attention(self, value: bool) -> None:
         self.add_bool(Keys.Attention.CAUSAL.format(arch=self.arch), value)
 
+    def add_pooling_type(self, value: PoolingType) -> None:
+        self.add_uint32(Keys.LLM.POOLING_TYPE.format(arch=self.arch), value)
+
     def add_rope_dimension_count(self, count: int) -> None:
         self.add_uint32(Keys.Rope.DIMENSION_COUNT.format(arch=self.arch), count)
 
@@ -411,6 +415,12 @@ class GGUFWriter:
     def add_pad_token_id(self, id: int) -> None:
         self.add_uint32(Keys.Tokenizer.PAD_ID, id)
 
+    def add_cls_token_id(self, id: int) -> None:
+        self.add_uint32(Keys.Tokenizer.CLS_ID, id)
+
+    def add_mask_token_id(self, id: int) -> None:
+        self.add_uint32(Keys.Tokenizer.MASK_ID, id)
+
     def add_add_bos_token(self, value: bool) -> None:
         self.add_bool(Keys.Tokenizer.ADD_BOS, value)
 

gguf-py/gguf/tensor_mapping.py

@@ -15,7 +15,7 @@ class TensorNameMap:
             "word_embeddings",                           # bloom
             "model.embed_tokens",                        # llama-hf
             "tok_embeddings",                            # llama-pth
-            "embeddings.word_embeddings",                # bert
+            "embeddings.word_embeddings",                # bert nomic-bert
             "language_model.embedding.word_embeddings",  # persimmon
             "wte",                                       # gpt2
             "transformer.embd.wte",                      # phi2
@@ -24,13 +24,14 @@ class TensorNameMap:
 
         # Token type embeddings
         MODEL_TENSOR.TOKEN_TYPES: (
-            "embeddings.token_type_embeddings",  # bert
+            "embeddings.token_type_embeddings",  # bert nomic-bert
         ),
 
         # Normalization of token embeddings
         MODEL_TENSOR.TOKEN_EMBD_NORM: (
             "word_embeddings_layernorm",  # bloom
             "embeddings.LayerNorm",       # bert
+            "emb_ln",                     # nomic-bert
         ),
 
         # Position embeddings
@@ -103,6 +104,7 @@ class TensorNameMap:
             "model.layers.{bid}.self_attn.query_key_value",                        # persimmon
             "h.{bid}.attn.c_attn",                                                 # gpt2
             "transformer.h.{bid}.mixer.Wqkv",                                      # phi2
+            "encoder.layers.{bid}.attn.Wqkv",                                      # nomic-bert
         ),
 
         # Attention query
@@ -152,11 +154,13 @@ class TensorNameMap:
             "transformer.h.{bid}.mixer.out_proj",                        # phi2
             "model.layers.layers.{bid}.self_attn.o_proj",                # plamo
             "model.layers.{bid}.attention.wo",                           # internlm2
+            "encoder.layers.{bid}.attn.out_proj",                        # nomic-bert
         ),
 
         # Attention output norm
         MODEL_TENSOR.ATTN_OUT_NORM: (
             "encoder.layer.{bid}.attention.output.LayerNorm",  # bert
+            "encoder.layers.{bid}.norm1",                      # nomic-bert
         ),
 
         # Rotary embeddings
@@ -205,6 +209,7 @@ class TensorNameMap:
             "model.layers.{bid}.mlp.fc1",                             # phi2
             "model.layers.layers.{bid}.mlp.up_proj",                  # plamo
             "model.layers.{bid}.feed_forward.w3",                     # internlm2
+            "encoder.layers.{bid}.mlp.fc11",                          # nomic-bert
         ),
 
         MODEL_TENSOR.FFN_UP_EXP: (
@@ -224,6 +229,7 @@ class TensorNameMap:
             "transformer.h.{bid}.mlp.w2",                 # qwen
             "model.layers.layers.{bid}.mlp.gate_proj",    # plamo
             "model.layers.{bid}.feed_forward.w1",         # internlm2
+            "encoder.layers.{bid}.mlp.fc12",              # nomic-bert
         ),
 
         MODEL_TENSOR.FFN_GATE_EXP: (
@@ -249,6 +255,7 @@ class TensorNameMap:
             "model.layers.{bid}.mlp.fc2",                             # phi2
             "model.layers.layers.{bid}.mlp.down_proj",                # plamo
             "model.layers.{bid}.feed_forward.w2",                     # internlm2
+            "encoder.layers.{bid}.mlp.fc2",                           # nomic-bert
         ),
 
         MODEL_TENSOR.FFN_DOWN_EXP: (
@@ -272,6 +279,7 @@ class TensorNameMap:
 
         MODEL_TENSOR.LAYER_OUT_NORM: (
             "encoder.layer.{bid}.output.LayerNorm",  # bert
+            "encoder.layers.{bid}.norm2",            # nomic-bert
         )
     }
 

gguf-py/gguf/vocab.py

@@ -29,7 +29,7 @@ class SpecialVocab:
         if special_token_types is not None:
             self.special_token_types = special_token_types
         else:
-            self.special_token_types = ('bos', 'eos', 'unk', 'sep', 'pad')
+            self.special_token_types = ('bos', 'eos', 'unk', 'sep', 'pad', 'cls', 'mask')
         self._load(Path(path))
 
     def __repr__(self) -> str:
@@ -152,10 +152,6 @@ class SpecialVocab:
             add_entry = tokenizer_config.get(f'add_{typ}_token')
             if isinstance(add_entry, bool):
                 self.add_special_token[typ] = add_entry
-            if not added_tokens:
-                # We will need this to get the content for the token, so if it's empty
-                # may as well just give up.
-                continue
             entry = tokenizer_config.get(f'{typ}_token')
             if isinstance(entry, str):
                 tc_content = entry

llama.h

@@ -100,6 +100,8 @@ 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_GUESSED = 1024, // not specified in the model file
     };
@@ -112,6 +114,12 @@ extern "C" {
         LLAMA_ROPE_SCALING_MAX_VALUE   = LLAMA_ROPE_SCALING_YARN,
     };
 
+    enum llama_pooling_type {
+        LLAMA_POOLING_NONE = 0,
+        LLAMA_POOLING_MEAN = 1,
+        LLAMA_POOLING_CLS  = 2,
+    };
+
     enum llama_split_mode {
         LLAMA_SPLIT_NONE    = 0, // single GPU
         LLAMA_SPLIT_LAYER   = 1, // split layers and KV across GPUs
@@ -236,6 +244,7 @@ extern "C" {
         bool logits_all;  // the llama_eval() call computes all logits, not just the last one (DEPRECATED - set llama_batch.logits instead)
         bool embedding;   // embedding mode only
         bool offload_kqv; // whether to offload the KQV ops (including the KV cache) to GPU
+        bool do_pooling;  // whether to pool (sum) embedding results by sequence id (ignored if no pooling layer)
     };
 
     // model quantization parameters
@@ -297,6 +306,12 @@ 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);
@@ -305,7 +320,10 @@ extern "C" {
     // Initialize the llama + ggml backend
     // If numa is true, use NUMA optimizations
     // Call once at the start of the program
-    LLAMA_API void llama_backend_init(bool numa);
+    LLAMA_API void llama_backend_init(void);
+
+    //optional:
+    LLAMA_API void llama_numa_init(enum ggml_numa_strategy numa);
 
     // Call once at the end of the program - currently only used for MPI
     LLAMA_API void llama_backend_free(void);
@@ -628,6 +646,10 @@ extern "C" {
     // shape: [n_embd] (1-dimensional)
     LLAMA_API float * llama_get_embeddings(struct llama_context * ctx);
 
+    // Get the embeddings for the ith sequence
+    // llama_get_embeddings(ctx) + i*n_embd
+    LLAMA_API float * llama_get_embeddings_ith(struct llama_context * ctx, int32_t i);
+
     //
     // Vocab
     //
@@ -684,6 +706,25 @@ 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/compare-commits.sh

@@ -0,0 +1,37 @@
+#!/bin/bash
+
+if [ $# -lt 2 ]; then
+    echo "usage: ./scripts/compare-commits.sh <commit1> <commit2> [additional llama-bench arguments]"
+    exit 1
+fi
+
+set -e
+set -x
+
+bench_args="${@:3}"
+
+rm -f llama-bench.sqlite
+
+backend="cpu"
+
+if [[ "$OSTYPE" == "darwin"* ]]; then
+    backend="metal"
+elif command -v nvcc &> /dev/null; then
+    backend="cuda"
+fi
+
+make_opts=""
+
+if [[ "$backend" == "cuda" ]]; then
+    make_opts="LLAMA_CUBLAS=1"
+fi
+
+git checkout $1
+make clean && make -j32 $make_opts llama-bench
+./llama-bench -o sql $bench_args | tee /dev/tty | sqlite3 llama-bench.sqlite
+
+git checkout $2
+make clean && make -j32 $make_opts llama-bench
+./llama-bench -o sql $bench_args | tee /dev/tty | sqlite3 llama-bench.sqlite
+
+./scripts/compare-llama-bench.py -b $1 -c $2

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/hf.sh

@@ -0,0 +1,107 @@
+#!/bin/bash
+#
+# Shortcut for downloading HF models
+#
+# Usage:
+#   ./main -m $(./examples/hf.sh https://huggingface.co/TheBloke/Mixtral-8x7B-v0.1-GGUF/resolve/main/mixtral-8x7b-v0.1.Q4_K_M.gguf)
+#   ./main -m $(./examples/hf.sh --url https://huggingface.co/TheBloke/Mixtral-8x7B-v0.1-GGUF/blob/main/mixtral-8x7b-v0.1.Q4_K_M.gguf)
+#   ./main -m $(./examples/hf.sh --repo TheBloke/Mixtral-8x7B-v0.1-GGUF --file mixtral-8x7b-v0.1.Q4_K_M.gguf)
+#
+
+# all logs go to stderr
+function log {
+    echo "$@" 1>&2
+}
+
+function usage {
+    log "Usage: $0 [[--url] <url>] [--repo <repo>] [--file <file>] [-h|--help]"
+    exit 1
+}
+
+# check for curl or wget
+function has_cmd {
+    if ! [ -x "$(command -v $1)" ]; then
+        return 1
+    fi
+}
+
+if has_cmd wget; then
+    cmd="wget -q --show-progress -c -O %s %s"
+elif has_cmd curl; then
+    cmd="curl -C - -f -o %s -L %s"
+else
+    log "[E] curl or wget not found"
+    exit 1
+fi
+
+url=""
+repo=""
+file=""
+
+# parse args
+while [[ $# -gt 0 ]]; do
+    case "$1" in
+        --url)
+            url="$2"
+            shift 2
+            ;;
+        --repo)
+            repo="$2"
+            shift 2
+            ;;
+        --file)
+            file="$2"
+            shift 2
+            ;;
+        -h|--help)
+            usage
+            ;;
+        *)
+            url="$1"
+            shift
+            ;;
+    esac
+done
+
+if [ -n "$repo" ] && [ -n "$file" ]; then
+    url="https://huggingface.co/$repo/resolve/main/$file"
+fi
+
+if [ -z "$url" ]; then
+    log "[E] missing --url"
+    usage
+fi
+
+# check if the URL is a HuggingFace model, and if so, try to download it
+is_url=false
+
+if [[ ${#url} -gt 22 ]]; then
+    if [[ ${url:0:22} == "https://huggingface.co" ]]; then
+        is_url=true
+    fi
+fi
+
+if [ "$is_url" = false ]; then
+    log "[E] invalid URL, must start with https://huggingface.co"
+    exit 0
+fi
+
+# replace "blob/main" with "resolve/main"
+url=${url/blob\/main/resolve\/main}
+
+basename=$(basename $url)
+
+log "[+] attempting to download $basename"
+
+if [ -n "$cmd" ]; then
+    cmd=$(printf "$cmd" "$basename" "$url")
+    log "[+] $cmd"
+    if $cmd; then
+        echo $basename
+        exit 0
+    fi
+fi
+
+log "[-] failed to download"
+
+exit 1

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-autorelease.cpp

@@ -12,7 +12,7 @@ int main(int argc, char ** argv) {
     auto * model_path = get_model_or_exit(argc, argv);
 
     std::thread([&model_path]() {
-        llama_backend_init(false);
+        llama_backend_init();
         auto * model = llama_load_model_from_file(model_path, llama_model_default_params());
         auto * ctx = llama_new_context_with_model(model, llama_context_default_params());
         llama_free(ctx);

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;
@@ -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,
     };
 
     // 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
@@ -2113,7 +2113,6 @@ 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));
 
@@ -2129,14 +2128,13 @@ static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op
     test_cases.emplace_back(new test_pad());
     test_cases.emplace_back(new test_leaky_relu());
 
+    // these tests are disabled to save execution time, but they can be handy for debugging
+#if 0
 #if !defined(__SANITIZE_THREAD__)
     // FIXME: these tests use too much memory with thread sanitizer
     test_cases.emplace_back(new test_moe(8, 2, 1, 4096, 8*1024));
     //test_cases.emplace_back(new test_moe(8, 2, 8, 4096, 14336));
 #endif
-
-    // these tests are disabled to save execution time, but they can be handy for debugging
-#if 0
     test_cases.emplace_back(new test_llama(1));
     test_cases.emplace_back(new test_llama(2));
     test_cases.emplace_back(new test_falcon(1));

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-model-load-cancel.cpp

@@ -14,7 +14,7 @@ int main(int argc, char *argv[] ) {
     fprintf(stderr, "using '%s'\n", model_path);
     fclose(file);
 
-    llama_backend_init(false);
+    llama_backend_init();
     auto params = llama_model_params{};
     params.use_mmap = false;
     params.progress_callback = [](float progress, void * ctx){

tests/test-tokenizer-0-falcon.cpp

@@ -61,7 +61,7 @@ int main(int argc, char **argv) {
     llama_model * model;
     llama_context * ctx;
 
-    llama_backend_init(false);
+    llama_backend_init();
 
     // load the vocab
     {

tests/test-tokenizer-0-llama.cpp

@@ -60,7 +60,7 @@ int main(int argc, char **argv) {
     llama_model * model;
     llama_context * ctx;
 
-    llama_backend_init(false);
+    llama_backend_init();
 
     // load the vocab
     {

tests/test-tokenizer-1-bpe.cpp

@@ -4,13 +4,13 @@
 #include "console.h"
 
 #include <cassert>
+#include <codecvt>
 #include <cstdio>
 #include <cstring>
-#include <string>
-#include <codecvt>
-#include <map>
-#include <vector>
 #include <locale>
+#include <string>
+#include <thread>
+#include <vector>
 
 int main(int argc, char **argv) {
     if (argc < 2) {
@@ -25,7 +25,7 @@ int main(int argc, char **argv) {
     llama_model * model;
     llama_context * ctx;
 
-    llama_backend_init(false);
+    llama_backend_init();
 
     // load the vocab
     {
@@ -74,45 +74,46 @@ int main(int argc, char **argv) {
             }
         }
         catch (const std::invalid_argument &) {
-            fprintf(stderr, "%s : info: utf8 conversion %d '%s'\n", __func__, i, str.c_str());
+            //fprintf(stderr, "%s : info: utf8 conversion %d '%s'\n", __func__, i, str.c_str());
         }
     }
 
-    for (uint32_t cp = 0x0000; cp < 0xffff; ++cp) {
-        // NOTE: these exceptions seem to be necessary, because the GPT2 tokenizer doesn't want to interfere with some ASCII control characters
-        if ((cp < 0x03 || cp > 0x05) && cp != 0x0b && cp != 0x11 && (cp < 0x13 || cp > 0x17) && cp != 0x19 && (cp < 0x1c || cp > 0x1e) && (cp < 0xd800 || cp > 0xdfff)) {
-            std::string str = " " + codepoint_to_utf8(cp);
-            std::vector<llama_token> tokens = llama_tokenize(ctx, str, false);
-            std::string check = llama_detokenize_bpe(ctx, tokens);
-            if (str != check) {
-                fprintf(stderr, "%s : error: codepoint %x detokenizes to '%s'(%zu) instead of '%s'(%zu)\n",
-                    __func__, cp, check.c_str(), check.length(), str.c_str(), str.length());
-                return 3;
-            }
-        }
-    }
-    // Restrict to assigned unicode planes
-    // for (uint32_t cp = 0x10000; cp < 0x0010ffff; ++cp) {
-    for (uint32_t cp = 0x10000; cp < 0x00040000; ++cp) {
-        std::string str = codepoint_to_utf8(cp);
-        std::vector<llama_token> tokens = llama_tokenize(ctx, str, false);
-        std::string check = llama_detokenize_bpe(ctx, tokens);
-        if (str != check) {
-            fprintf(stderr, "%s : error: codepoint %x detokenizes to '%s'(%zu) instead of '%s'(%zu)\n",
-                __func__, cp, check.c_str(), check.length(), str.c_str(), str.length());
-            return 4;
-        }
-    }
-    for (uint32_t cp = 0x000e0000; cp < 0x0010ffff; ++cp) {
-        std::string str = codepoint_to_utf8(cp);
-        std::vector<llama_token> tokens = llama_tokenize(ctx, str, false);
-        std::string check = llama_detokenize_bpe(ctx, tokens);
-        if (str != check) {
-            fprintf(stderr, "%s : error: codepoint %x detokenizes to '%s'(%zu) instead of '%s'(%zu)\n",
-                __func__, cp, check.c_str(), check.length(), str.c_str(), str.length());
-            return 4;
+    // unicode
+    {
+        const int nthread = std::thread::hardware_concurrency();
+
+        std::vector<std::thread> threads(nthread);
+
+        for (int i = 0; i < nthread; ++i) {
+            threads[i] = std::thread([i, nthread, ctx]() {
+                for (uint32_t cp = i; cp < 0x0010ffff; cp += nthread) {
+                    if (!( // NOLINT
+                                (cp < 0x03       || cp >  0x05)   && cp != 0x0b && cp != 0x11 &&
+                                (cp < 0x13       || cp >  0x17)   && cp != 0x19 &&
+                                (cp < 0x1c       || cp >  0x1e)   &&
+                                (cp < 0xd800     || cp >  0xdfff) &&
+                                (cp < 0x00040000 || cp >= 0x000e0000)
+                        )) {
+                        continue;
+                    }
+
+                    std::string str = codepoint_to_utf8(cp);
+                    std::vector<llama_token> tokens = llama_tokenize(ctx, str, false);
+                    std::string check = llama_detokenize_bpe(ctx, tokens);
+                    if (cp != 9601 && str != check) {
+                        fprintf(stderr, "error: codepoint %x detokenizes to '%s'(%zu) instead of '%s'(%zu)\n",
+                                cp, check.c_str(), check.length(), str.c_str(), str.length());
+                        std::exit(3);
+                    }
+                }
+            });
+        }
+
+        for (auto & t : threads) {
+            t.join();
         }
     }
+
     llama_free_model(model);
     llama_free(ctx);
 

tests/test-tokenizer-1-llama.cpp

@@ -4,13 +4,13 @@
 #include "console.h"
 
 #include <cassert>
+#include <codecvt>
 #include <cstdio>
 #include <cstring>
-#include <string>
-#include <codecvt>
-#include <map>
-#include <vector>
 #include <locale>
+#include <string>
+#include <thread>
+#include <vector>
 
 int main(int argc, char **argv) {
     if (argc < 2) {
@@ -25,7 +25,7 @@ int main(int argc, char **argv) {
     llama_model * model;
     llama_context * ctx;
 
-    llama_backend_init(false);
+    llama_backend_init();
 
     // load the vocab
     {
@@ -72,26 +72,33 @@ int main(int argc, char **argv) {
         }
     }
 
-    for (uint32_t cp = 0x0000; cp < 0xffff; ++cp) {
-        if (cp < 0xd800 || cp > 0xdfff) {
-            std::string str = codepoint_to_utf8(cp);
-            std::vector<llama_token> tokens = llama_tokenize(ctx, str, false);
-            std::string check = llama_detokenize_spm(ctx, tokens);
-            if (cp != 9601 && str != check) {
-                fprintf(stderr, "%s : error: codepoint %d detokenizes to '%s'(%zu) instead of '%s'(%zu)\n",
-                    __func__, cp, check.c_str(), check.length(), str.c_str(), str.length());
-                return 3;
-            }
+    // unicode
+    {
+        const int nthread = std::thread::hardware_concurrency();
+
+        std::vector<std::thread> threads(nthread);
+
+        for (int i = 0; i < nthread; ++i) {
+            threads[i] = std::thread([i, nthread, ctx]() {
+                for (uint32_t cp = i; cp < 0x0010ffff; cp += nthread) {
+                    if (cp >= 0xd800 && cp <= 0xdfff) {
+                        continue;
+                    }
+
+                    std::string str = codepoint_to_utf8(cp);
+                    std::vector<llama_token> tokens = llama_tokenize(ctx, str, false);
+                    std::string check = llama_detokenize_spm(ctx, tokens);
+                    if (cp != 9601 && str != check) {
+                        fprintf(stderr, "error: codepoint %x detokenizes to '%s'(%zu) instead of '%s'(%zu)\n",
+                                cp, check.c_str(), check.length(), str.c_str(), str.length());
+                        std::exit(3);
+                    }
+                }
+            });
         }
-    }
-    for (uint32_t cp = 0x10000; cp < 0x0010ffff; ++cp) {
-        std::string str = codepoint_to_utf8(cp);
-        std::vector<llama_token> tokens = llama_tokenize(ctx, str, false);
-        std::string check = llama_detokenize_spm(ctx, tokens);
-        if (str != check) {
-            fprintf(stderr, "%s : error: codepoint %d detokenizes to '%s'(%zu) instead of '%s'(%zu)\n",
-                __func__, cp, check.c_str(), check.length(), str.c_str(), str.length());
-            return 4;
+
+        for (auto & t : threads) {
+            t.join();
         }
     }