llama : remove MPI backend (#7395)

* Fix empty Vulkan host buffers

Add fp32 fp16 matmul shader

Fix matmul shader alignment

* Remove deprecated tensor->backend uses

* Fix Vulkan validation errors on embedding models with no offloaded layers

* Fix Vulkan llava segfault when not offloading layers

.devops/nix/package.nix

@@ -214,7 +214,6 @@ effectiveStdenv.mkDerivation (
         (cmakeBool "LLAMA_CUDA" useCuda)
         (cmakeBool "LLAMA_HIPBLAS" useRocm)
         (cmakeBool "LLAMA_METAL" useMetalKit)
-        (cmakeBool "LLAMA_MPI" useMpi)
         (cmakeBool "LLAMA_VULKAN" useVulkan)
         (cmakeBool "LLAMA_STATIC" enableStatic)
       ]
@@ -227,20 +226,20 @@ effectiveStdenv.mkDerivation (
         )
       ]
       ++ optionals useRocm [
-        (cmakeFeature "CMAKE_C_COMPILER" "hipcc")
-        (cmakeFeature "CMAKE_CXX_COMPILER" "hipcc")
-
-        # Build all targets supported by rocBLAS. When updating search for TARGET_LIST_ROCM
-        # in https://github.com/ROCmSoftwarePlatform/rocBLAS/blob/develop/CMakeLists.txt
-        # and select the line that matches the current nixpkgs version of rocBLAS.
-        # Should likely use `rocmPackages.clr.gpuTargets`.
-        "-DAMDGPU_TARGETS=gfx803;gfx900;gfx906:xnack-;gfx908:xnack-;gfx90a:xnack+;gfx90a:xnack-;gfx940;gfx941;gfx942;gfx1010;gfx1012;gfx1030;gfx1100;gfx1101;gfx1102"
+        (cmakeFeature "CMAKE_HIP_COMPILER" "${rocmPackages.llvm.clang}/bin/clang")
+        (cmakeFeature "CMAKE_HIP_ARCHITECTURES" (builtins.concatStringsSep ";" rocmPackages.clr.gpuTargets))
       ]
       ++ optionals useMetalKit [
         (lib.cmakeFeature "CMAKE_C_FLAGS" "-D__ARM_FEATURE_DOTPROD=1")
         (cmakeBool "LLAMA_METAL_EMBED_LIBRARY" (!precompileMetalShaders))
       ];
 
+    # Environment variables needed for ROCm
+    env = optionals useRocm {
+      ROCM_PATH = "${rocmPackages.clr}";
+      HIP_DEVICE_LIB_PATH = "${rocmPackages.rocm-device-libs}/amdgcn/bitcode";
+    };
+
     # TODO(SomeoneSerge): It's better to add proper install targets at the CMake level,
     # if they haven't been added yet.
     postInstall = ''

.github/labeler.yml

@@ -0,0 +1,73 @@
+# https://github.com/actions/labeler
+
+SYCL:
+    - changed-files:
+        - any-glob-to-any-file:
+            - ggml-sycl.h
+            - ggml-sycl.cpp
+            - README-sycl.md
+Nvidia GPU:
+    - changed-files:
+        - any-glob-to-any-file:
+            - ggml-cuda/**
+Vulkan:
+    - changed-files:
+        - any-glob-to-any-file:
+            - ggml_vk_generate_shaders.py
+            - ggml-vulkan*
+documentation:
+    - changed-files:
+        - any-glob-to-any-file:
+            - docs/**
+            - media/**
+testing:
+    - changed-files:
+        - any-glob-to-any-file:
+            - tests/**
+build:
+    - changed-files:
+        - any-glob-to-any-file:
+            - cmake/**
+            - CMakeLists.txt
+            - CMakePresets.json
+            - codecov.yml
+examples:
+    - changed-files:
+        - any-glob-to-any-file: examples/**
+devops:
+    - changed-files:
+        - any-glob-to-any-file:
+            - .devops/**
+            - .github/**
+            - ci/**
+python:
+    - changed-files:
+        - any-glob-to-any-file:
+            - "**/*.py"
+            - requirements/**
+            - gguf-py/**
+            - .flake8
+script:
+    - changed-files:
+        - any-glob-to-any-file:
+            - scripts/**
+android:
+    - changed-files:
+        - any-glob-to-any-file:
+            - examples/llama.android/**
+server:
+    - changed-files:
+        - any-glob-to-any-file:
+            - examples/server/**
+ggml:
+    - changed-files:
+        - any-glob-to-any-file:
+            - ggml-*.c
+            - ggml-*.h
+            - ggml-cuda/**
+nix:
+    - changed-files:
+        - any-glob-to-any-file:
+            - "**/*.nix"
+            - .github/workflows/nix-*.yml
+            - .devops/nix/nixpkgs-instances.nix

.github/workflows/build.yml

@@ -271,49 +271,15 @@ jobs:
           path: llama-${{ steps.tag.outputs.name }}-bin-ubuntu-x64.zip
           name: llama-bin-ubuntu-x64.zip
 
-#  ubuntu-latest-cmake-sanitizer:
-#    runs-on: ubuntu-latest
-#
-#    continue-on-error: true
-#
-#    strategy:
-#      matrix:
-#        sanitizer: [ADDRESS, THREAD, UNDEFINED]
-#        build_type: [Debug, Release]
-#
-#    steps:
-#      - name: Clone
-#        id: checkout
-#        uses: actions/checkout@v4
-#
-#      - name: Dependencies
-#        id: depends
-#        run: |
-#          sudo apt-get update
-#          sudo apt-get install build-essential
-#
-#      - name: Build
-#        id: cmake_build
-#        run: |
-#          mkdir build
-#          cd build
-#          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
-#        id: cmake_test
-#        run: |
-#          cd build
-#          ctest -L main --verbose --timeout 900
-
-  ubuntu-latest-cmake-mpi:
+  ubuntu-latest-cmake-sanitizer:
     runs-on: ubuntu-latest
 
     continue-on-error: true
 
     strategy:
       matrix:
-        mpi_library: [mpich, libopenmpi-dev]
+        sanitizer: [ADDRESS, THREAD, UNDEFINED]
+        build_type: [Debug, Release]
 
     steps:
       - name: Clone
@@ -324,21 +290,21 @@ jobs:
         id: depends
         run: |
           sudo apt-get update
-          sudo apt-get install build-essential ${{ matrix.mpi_library }}
+          sudo apt-get install build-essential
 
       - name: Build
         id: cmake_build
         run: |
           mkdir build
           cd build
-          cmake -DLLAMA_MPI=ON ..
-          cmake --build . --config Release -j $(nproc)
+          cmake .. -DLLAMA_FATAL_WARNINGS=ON -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON -DCMAKE_BUILD_TYPE=${{ matrix.build_type }}
+          cmake --build . --config ${{ matrix.build_type }} -j $(nproc)
 
       - name: Test
         id: cmake_test
         run: |
           cd build
-          ctest -L main --verbose
+          ctest -L main --verbose --timeout 900
 
   ubuntu-latest-cmake-rpc:
     runs-on: ubuntu-latest
@@ -392,6 +358,33 @@ jobs:
           cmake -DLLAMA_VULKAN=ON ..
           cmake --build . --config Release -j $(nproc)
 
+  ubuntu-22-cmake-hip:
+    runs-on: ubuntu-22.04
+    container: rocm/dev-ubuntu-22.04:6.0.2
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v3
+
+      - name: Dependencies
+        id: depends
+        run: |
+          sudo apt-get update
+          sudo apt-get install -y build-essential git cmake rocblas-dev hipblas-dev
+
+      - name: Build with native CMake HIP support
+        id: cmake_build
+        run: |
+          cmake -B build -S . -DCMAKE_HIP_COMPILER="$(hipconfig -l)/clang" -DLLAMA_HIPBLAS=ON
+          cmake --build build --config Release -j $(nproc)
+
+      - name: Build with legacy HIP support
+        id: cmake_build_legacy_hip
+        run: |
+          cmake -B build2 -S . -DCMAKE_C_COMPILER=hipcc -DCMAKE_CXX_COMPILER=hipcc -DLLAMA_HIPBLAS=ON
+          cmake --build build2 --config Release -j $(nproc)
+
   ubuntu-22-cmake-sycl:
     runs-on: ubuntu-22.04
 
@@ -989,6 +982,37 @@ jobs:
           path: llama-${{ steps.tag.outputs.name }}-bin-win-sycl-x64.zip
           name: llama-bin-win-sycl-x64.zip
 
+  windows-latest-cmake-hip:
+    runs-on: windows-latest
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v3
+
+      - name: Install
+        id: depends
+        run: |
+          $ErrorActionPreference = "Stop"
+          write-host "Downloading AMD HIP SDK Installer"
+          Invoke-WebRequest -Uri "https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-23.Q4-WinSvr2022-For-HIP.exe" -OutFile "${env:RUNNER_TEMP}\rocm-install.exe"
+          write-host "Installing AMD HIP SDK"
+          Start-Process "${env:RUNNER_TEMP}\rocm-install.exe" -ArgumentList '-install' -NoNewWindow -Wait
+          write-host "Completed AMD HIP SDK installation"
+
+      - name: Verify ROCm
+        id: verify
+        run: |
+          & 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' --version
+
+      - name: Build
+        id: cmake_build
+        run: |
+          $env:HIP_PATH=$(Resolve-Path 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' | split-path | split-path)
+          $env:CMAKE_PREFIX_PATH="${env:HIP_PATH}"
+          cmake -G "Unix Makefiles" -B build -S . -DCMAKE_C_COMPILER="${env:HIP_PATH}\bin\clang.exe" -DCMAKE_CXX_COMPILER="${env:HIP_PATH}\bin\clang++.exe" -DLLAMA_HIPBLAS=ON
+          cmake --build build --config Release
+
   ios-xcode-build:
     runs-on: macos-latest
 

.github/workflows/labeler.yml

@@ -0,0 +1,17 @@
+name: "Pull Request Labeler"
+on:
+- pull_request_target
+
+jobs:
+  labeler:
+    permissions:
+      contents: read
+      pull-requests: write
+    runs-on: ubuntu-latest
+    steps:
+    - uses: actions/checkout@v4
+      with:
+        repository: "ggerganov/llama.cpp"
+    - uses: actions/labeler@v5
+      with:
+        configuration-path: '.github/labeler.yml'

.github/workflows/server.yml

@@ -32,13 +32,14 @@ jobs:
 
     strategy:
       matrix:
-        # TODO: temporary disabled due to linux kernel issues
-        #sanitizer: [ADDRESS, THREAD, UNDEFINED]
-        sanitizer: [UNDEFINED]
+        sanitizer: [ADDRESS, THREAD, UNDEFINED]
         build_type: [Debug]
         include:
           - build_type: Release
             sanitizer: ""
+          - build_type: Debug
+            sanitizer: THREAD
+            disabled_on_pr: true
       fail-fast: false # While -DLLAMA_SANITIZE_THREAD=ON is broken
 
     steps:

CMakeLists.txt

@@ -1,4 +1,4 @@
-cmake_minimum_required(VERSION 3.14)  # for add_link_options and implicit target directories.
+cmake_minimum_required(VERSION 3.14) # for add_link_options and implicit target directories.
 project("llama.cpp" C CXX)
 include(CheckIncludeFileCXX)
 
@@ -122,7 +122,6 @@ set(LLAMA_METAL_MACOSX_VERSION_MIN "" CACHE STRING
                                              "llama: metal minimum macOS version")
 set(LLAMA_METAL_STD "" CACHE STRING          "llama: metal standard version (-std flag)")
 option(LLAMA_KOMPUTE                         "llama: use Kompute"                               OFF)
-option(LLAMA_MPI                             "llama: use MPI"                                   OFF)
 option(LLAMA_RPC                             "llama: use RPC"                                   OFF)
 option(LLAMA_QKK_64                          "llama: use super-block size of 64 for k-quants"   OFF)
 option(LLAMA_SYCL                            "llama: use SYCL"                                  OFF)
@@ -466,35 +465,6 @@ if (LLAMA_CUDA)
     endif()
 endif()
 
-if (LLAMA_MPI)
-    cmake_minimum_required(VERSION 3.10)
-    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)
-
-        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)
-            set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS}     ${MPI_CXX_LIBRARIES})
-        endif()
-    else()
-        message(WARNING "MPI not found")
-    endif()
-endif()
-
 if (LLAMA_RPC)
     add_compile_definitions(GGML_USE_RPC)
 
@@ -555,16 +525,37 @@ if (LLAMA_VULKAN)
 endif()
 
 if (LLAMA_HIPBLAS)
-    list(APPEND CMAKE_PREFIX_PATH /opt/rocm)
+    if ($ENV{ROCM_PATH})
+        set(ROCM_PATH $ENV{ROCM_PATH})
+    else()
+        set(ROCM_PATH /opt/rocm)
+    endif()
+    list(APPEND CMAKE_PREFIX_PATH ${ROCM_PATH})
 
-    if (NOT ${CMAKE_C_COMPILER_ID} MATCHES "Clang")
-        message(WARNING "Only LLVM is supported for HIP, hint: CC=/opt/rocm/llvm/bin/clang")
+    # CMake on Windows doesn't support the HIP language yet
+    if(WIN32)
+        set(CXX_IS_HIPCC TRUE)
+    else()
+        string(REGEX MATCH "hipcc(\.bat)?$" CXX_IS_HIPCC "${CMAKE_CXX_COMPILER}")
     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()
+    if(CXX_IS_HIPCC)
+        if(LINUX)
+            if (NOT ${CMAKE_CXX_COMPILER_ID} MATCHES "Clang")
+                message(WARNING "Only LLVM is supported for HIP, hint: CXX=/opt/rocm/llvm/bin/clang++")
+            endif()
 
+            message(WARNING "Setting hipcc as the C++ compiler is legacy behavior."
+                    " Prefer setting the HIP compiler directly. See README for details.")
+        endif()
+    else()
+        # Forward AMDGPU_TARGETS to CMAKE_HIP_ARCHITECTURES.
+        if(AMDGPU_TARGETS AND NOT CMAKE_HIP_ARCHITECTURES)
+            set(CMAKE_HIP_ARCHITECTURES ${AMDGPU_TARGETS})
+        endif()
+        cmake_minimum_required(VERSION 3.21)
+        enable_language(HIP)
+    endif()
     find_package(hip     REQUIRED)
     find_package(hipblas REQUIRED)
     find_package(rocblas REQUIRED)
@@ -598,13 +589,18 @@ if (LLAMA_HIPBLAS)
     add_compile_definitions(GGML_CUDA_MMV_Y=${LLAMA_CUDA_MMV_Y})
     add_compile_definitions(K_QUANTS_PER_ITERATION=${LLAMA_CUDA_KQUANTS_ITER})
 
-    set_source_files_properties(${GGML_SOURCES_ROCM} PROPERTIES LANGUAGE CXX)
+    if (CXX_IS_HIPCC)
+        set_source_files_properties(${GGML_SOURCES_ROCM} PROPERTIES LANGUAGE CXX)
+        set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} hip::device)
+    else()
+        set_source_files_properties(${GGML_SOURCES_ROCM} PROPERTIES LANGUAGE HIP)
+    endif()
 
     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)
+    set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} PUBLIC hip::host roc::rocblas roc::hipblas)
 endif()
 
 if (LLAMA_SYCL)
@@ -1192,7 +1188,6 @@ add_library(ggml OBJECT
             ${GGML_SOURCES_CUDA}      ${GGML_HEADERS_CUDA}
             ${GGML_SOURCES_OPENCL}    ${GGML_HEADERS_OPENCL}
             ${GGML_SOURCES_METAL}     ${GGML_HEADERS_METAL}
-            ${GGML_SOURCES_MPI}       ${GGML_HEADERS_MPI}
             ${GGML_SOURCES_RPC}       ${GGML_HEADERS_RPC}
             ${GGML_SOURCES_EXTRA}     ${GGML_HEADERS_EXTRA}
             ${GGML_SOURCES_SYCL}      ${GGML_HEADERS_SYCL}
@@ -1280,7 +1275,7 @@ install(FILES ${CMAKE_CURRENT_BINARY_DIR}/LlamaConfig.cmake
 
 set(GGML_PUBLIC_HEADERS "ggml.h" "ggml-alloc.h" "ggml-backend.h"
         "${GGML_HEADERS_CUDA}"  "${GGML_HEADERS_OPENCL}"
-        "${GGML_HEADERS_METAL}" "${GGML_HEADERS_MPI}" "${GGML_HEADERS_EXTRA}")
+        "${GGML_HEADERS_METAL}" "${GGML_HEADERS_EXTRA}")
 
 set_target_properties(ggml PROPERTIES PUBLIC_HEADER "${GGML_PUBLIC_HEADERS}")
 install(TARGETS ggml PUBLIC_HEADER)

Makefile

@@ -399,13 +399,6 @@ ifndef LLAMA_NO_ACCELERATE
 	endif
 endif # LLAMA_NO_ACCELERATE
 
-ifdef LLAMA_MPI
-	MK_CPPFLAGS += -DGGML_USE_MPI
-	MK_CFLAGS   += -Wno-cast-qual
-	MK_CXXFLAGS += -Wno-cast-qual
-	OBJS        += ggml-mpi.o
-endif # LLAMA_MPI
-
 ifdef LLAMA_OPENBLAS
 	MK_CPPFLAGS += -DGGML_USE_OPENBLAS $(shell pkg-config --cflags-only-I openblas)
 	MK_CFLAGS   += $(shell pkg-config --cflags-only-other openblas)
@@ -560,10 +553,10 @@ endif # LLAMA_VULKAN
 ifdef LLAMA_HIPBLAS
 	ifeq ($(wildcard /opt/rocm),)
 		ROCM_PATH	?= /usr
-		GPU_TARGETS ?= $(shell $(shell which amdgpu-arch))
+		AMDGPU_TARGETS ?= $(shell $(shell which amdgpu-arch))
 	else
 		ROCM_PATH	?= /opt/rocm
-		GPU_TARGETS ?= $(shell $(ROCM_PATH)/llvm/bin/amdgpu-arch)
+		AMDGPU_TARGETS ?= $(shell $(ROCM_PATH)/llvm/bin/amdgpu-arch)
 	endif
 	HIPCC                   ?= $(CCACHE) $(ROCM_PATH)/bin/hipcc
 	LLAMA_CUDA_DMMV_X       ?= 32
@@ -575,7 +568,7 @@ ifdef LLAMA_HIP_UMA
 endif # LLAMA_HIP_UMA
 	MK_LDFLAGS  += -L$(ROCM_PATH)/lib -Wl,-rpath=$(ROCM_PATH)/lib
 	MK_LDFLAGS	+= -lhipblas -lamdhip64 -lrocblas
-	HIPFLAGS    += $(addprefix --offload-arch=,$(GPU_TARGETS))
+	HIPFLAGS    += $(addprefix --offload-arch=,$(AMDGPU_TARGETS))
 	HIPFLAGS    += -DGGML_CUDA_DMMV_X=$(LLAMA_CUDA_DMMV_X)
 	HIPFLAGS    += -DGGML_CUDA_MMV_Y=$(LLAMA_CUDA_MMV_Y)
 	HIPFLAGS    += -DK_QUANTS_PER_ITERATION=$(LLAMA_CUDA_KQUANTS_ITER)
@@ -629,11 +622,6 @@ ggml-metal-embed.o: ggml-metal.metal ggml-common.h
 endif
 endif # LLAMA_METAL
 
-ifdef LLAMA_MPI
-ggml-mpi.o: ggml-mpi.c ggml-mpi.h
-	$(CC) $(CFLAGS) -c $< -o $@
-endif # LLAMA_MPI
-
 ifndef LLAMA_NO_LLAMAFILE
 sgemm.o: sgemm.cpp sgemm.h ggml.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@

README.md

@@ -382,45 +382,6 @@ To disable the Metal build at compile time use the `LLAMA_NO_METAL=1` flag or th
 When built with Metal support, you can explicitly disable GPU inference with the `--n-gpu-layers|-ngl 0` command-line
 argument.
 
-### MPI Build
-
-MPI lets you distribute the computation over a cluster of machines. Because of the serial nature of LLM prediction, this won't yield any end-to-end speed-ups, but it will let you run larger models than would otherwise fit into RAM on a single machine.
-
-First you will need MPI libraries installed on your system. The two most popular (only?) options are [MPICH](https://www.mpich.org) and [OpenMPI](https://www.open-mpi.org). Either can be installed with a package manager (`apt`, Homebrew, MacPorts, etc).
-
-Next you will need to build the project with `LLAMA_MPI` set to true on all machines; if you're building with `make`, you will also need to specify an MPI-capable compiler (when building with CMake, this is configured automatically):
-
-- Using `make`:
-
-  ```bash
-  make CC=mpicc CXX=mpicxx LLAMA_MPI=1
-  ```
-
-- Using `CMake`:
-
-  ```bash
-  cmake -S . -B build -DLLAMA_MPI=ON
-  ```
-
-Once the programs are built, download/convert the weights on all of the machines in your cluster. The paths to the weights and programs should be identical on all machines.
-
-Next, ensure password-less SSH access to each machine from the primary host, and create a `hostfile` with a list of the hostnames and their relative "weights" (slots). If you want to use localhost for computation, use its local subnet IP address rather than the loopback address or "localhost".
-
-Here is an example hostfile:
-
-```
-192.168.0.1:2
-malvolio.local:1
-```
-
-The above will distribute the computation across 2 processes on the first host and 1 process on the second host. Each process will use roughly an equal amount of RAM. Try to keep these numbers small, as inter-process (intra-host) communication is expensive.
-
-Finally, you're ready to run a computation using `mpirun`:
-
-```bash
-mpirun -hostfile hostfile -n 3 ./main -m ./models/7B/ggml-model-q4_0.gguf -n 128
-```
-
 ### BLAS Build
 
 Building the program with BLAS support may lead to some performance improvements in prompt processing using batch sizes higher than 32 (the default is 512). Support with CPU-only BLAS implementations doesn't affect the normal generation performance. We may see generation performance improvements with GPU-involved BLAS implementations, e.g. cuBLAS, hipBLAS and CLBlast. There are currently several different BLAS implementations available for build and use:
@@ -528,13 +489,28 @@ Building the program with BLAS support may lead to some performance improvements
     ```
   - Using `CMake` for Linux (assuming a gfx1030-compatible AMD GPU):
     ```bash
-    CC=/opt/rocm/llvm/bin/clang CXX=/opt/rocm/llvm/bin/clang++ \
-        cmake -B build -DLLAMA_HIPBLAS=ON -DAMDGPU_TARGETS=gfx1030 -DCMAKE_BUILD_TYPE=Release \
+    HIPCXX="$(hipconfig -l)/clang" HIP_PATH="$(hipconfig -R)" \
+        cmake -S . -B build -DLLAMA_HIPBLAS=ON -DAMDGPU_TARGETS=gfx1030 -DCMAKE_BUILD_TYPE=Release \
         && cmake --build build --config Release -- -j 16
     ```
     On Linux it is also possible to use unified memory architecture (UMA) to share main memory between the CPU and integrated GPU by setting `-DLLAMA_HIP_UMA=ON`.
     However, this hurts performance for non-integrated GPUs (but enables working with integrated GPUs).
 
+    Note that if you get the following error:
+    ```
+    clang: error: cannot find ROCm device library; provide its path via '--rocm-path' or '--rocm-device-lib-path', or pass '-nogpulib' to build without ROCm device library
+    ```
+    Try searching for a directory under `HIP_PATH` that contains the file
+    `oclc_abi_version_400.bc`. Then, add the following to the start of the
+    command: `HIP_DEVICE_LIB_PATH=<directory-you-just-found>`, so something
+    like:
+    ```bash
+    HIPCXX="$(hipconfig -l)/clang" HIP_PATH="$(hipconfig -p)" \
+    HIP_DEVICE_LIB_PATH=<directory-you-just-found> \
+        cmake -S . -B build -DLLAMA_HIPBLAS=ON -DAMDGPU_TARGETS=gfx1030 -DCMAKE_BUILD_TYPE=Release \
+        && cmake --build build -- -j 16
+    ```
+
   - Using `make` (example for target gfx1030, build with 16 CPU threads):
     ```bash
     make -j16 LLAMA_HIPBLAS=1 LLAMA_HIP_UMA=1 AMDGPU_TARGETS=gfx1030
@@ -543,10 +519,8 @@ Building the program with BLAS support may lead to some performance improvements
   - Using `CMake` for Windows (using x64 Native Tools Command Prompt for VS, and assuming a gfx1100-compatible AMD GPU):
     ```bash
     set PATH=%HIP_PATH%\bin;%PATH%
-    mkdir build
-    cd build
-    cmake -G Ninja -DAMDGPU_TARGETS=gfx1100 -DLLAMA_HIPBLAS=ON -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ -DCMAKE_BUILD_TYPE=Release ..
-    cmake --build .
+    cmake -S . -B build -G Ninja -DAMDGPU_TARGETS=gfx1100 -DLLAMA_HIPBLAS=ON -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ -DCMAKE_BUILD_TYPE=Release
+    cmake --build build
     ```
     Make sure that `AMDGPU_TARGETS` is set to the GPU arch you want to compile for. The above example uses `gfx1100` that corresponds to Radeon RX 7900XTX/XT/GRE. You can find a list of targets [here](https://llvm.org/docs/AMDGPUUsage.html#processors)
     Find your gpu version string by matching the most significant version information from `rocminfo | grep gfx | head -1 | awk '{print $2}'` with the list of processors, e.g. `gfx1035` maps to `gfx1030`.

convert-hf-to-gguf-update.py

@@ -72,6 +72,7 @@ models = [
     {"name": "mpt",            "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/mosaicml/mpt-7b", },
     {"name": "starcoder",      "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/bigcode/starcoder2-3b", },
     {"name": "gpt-2",          "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/openai-community/gpt2", },
+    {"name": "stablelm",       "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/stabilityai/stablelm-2-zephyr-1_6b", },
     {"name": "refact",         "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/smallcloudai/Refact-1_6-base", },
     {"name": "command-r",      "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/CohereForAI/c4ai-command-r-v01", },
     {"name": "qwen2",          "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/Qwen/Qwen1.5-7B", },

convert-hf-to-gguf.py

@@ -446,6 +446,9 @@ class Model:
         if chkhsh == "3ce83efda5659b07b1ad37ca97ca5797ea4285d9b9ab0dc679e4a720c9da7454":
             # ref: https://huggingface.co/openai-community/gpt2
             res = "gpt-2"
+        if chkhsh == "32d85c31273f8019248f2559fed492d929ea28b17e51d81d3bb36fff23ca72b3":
+            # ref: https://huggingface.co/stabilityai/stablelm-2-1_6b
+            res = "stablelm2"
         if chkhsh == "6221ad2852e85ce96f791f476e0b390cf9b474c9e3d1362f53a24a06dc8220ff":
             # ref: https://huggingface.co/smallcloudai/Refact-1_6-base
             res = "refact"
@@ -573,6 +576,10 @@ class Model:
 
         vocab_size = self.hparams.get('vocab_size', tokenizer.vocab_size())
 
+        tokens: list[bytes] = [f"[PAD{i}]".encode("utf-8") for i in range(vocab_size)]
+        scores: list[float] = [-10000.0] * vocab_size
+        toktypes: list[int] = [SentencePieceTokenTypes.UNKNOWN] * vocab_size
+
         for token_id in range(tokenizer.vocab_size()):
             piece = tokenizer.IdToPiece(token_id)
             text = piece.encode("utf-8")
@@ -588,21 +595,23 @@ class Model:
             elif tokenizer.IsByte(token_id):
                 toktype = SentencePieceTokenTypes.BYTE
 
-            tokens.append(text)
-            scores.append(score)
-            toktypes.append(toktype)
+            tokens[token_id] = text
+            scores[token_id] = score
+            toktypes[token_id] = toktype
 
         added_tokens_file = self.dir_model / 'added_tokens.json'
         if added_tokens_file.is_file():
             with open(added_tokens_file, "r", encoding="utf-8") as f:
                 added_tokens_json = json.load(f)
-
                 for key in added_tokens_json:
-                    key = key.encode("utf-8")
-                    if key not in tokens:
-                        tokens.append(key)
-                        scores.append(-1000.0)
-                        toktypes.append(SentencePieceTokenTypes.USER_DEFINED)
+                    token_id = added_tokens_json[key]
+                    if (token_id >= vocab_size):
+                        logger.warning(f'ignore token {token_id}: id is out of range, max={vocab_size - 1}')
+                        continue
+
+                    tokens[token_id] = key.encode("utf-8")
+                    scores[token_id] = -1000.0
+                    toktypes[token_id] = SentencePieceTokenTypes.USER_DEFINED
 
         if vocab_size > len(tokens):
             pad_count = vocab_size - len(tokens)
@@ -612,8 +621,6 @@ class Model:
                 scores.append(-1000.0)
                 toktypes.append(SentencePieceTokenTypes.UNUSED)
 
-        assert len(tokens) == vocab_size
-
         self.gguf_writer.add_tokenizer_model("llama")
         self.gguf_writer.add_tokenizer_pre("default")
         self.gguf_writer.add_token_list(tokens)

examples/llama.android/app/src/main/cpp/CMakeLists.txt

@@ -12,17 +12,20 @@ cmake_minimum_required(VERSION 3.22.1)
 # build script scope).
 project("llama-android")
 
+## Fetch latest llama.cpp from GitHub
 #include(FetchContent)
 #FetchContent_Declare(
 #        llama
 #        GIT_REPOSITORY https://github.com/ggerganov/llama.cpp
-#        GIT_TAG        ci-android
+#        GIT_TAG        master
 #)
 #
 ## Also provides "common"
 #FetchContent_MakeAvailable(llama)
 
-add_subdirectory(../../../../../../ please-work)
+# llama.cpp CI uses the code from the current branch
+# ref: https://github.com/ggerganov/llama.cpp/pull/7341#issuecomment-2117617700
+add_subdirectory(../../../../../../ build-llama)
 
 # Creates and names a library, sets it as either STATIC
 # or SHARED, and provides the relative paths to its source code.

examples/perplexity/perplexity.cpp

@@ -1425,7 +1425,7 @@ static void multiple_choice_score(llama_context * ctx, const gpt_params & params
         // Use all tasks
         tasks.resize(n_task);
         printf("%s: reading tasks", __func__);
-        int n_dot = n_task/100;
+        int n_dot = std::max((int) n_task/100, 1);
         int i = 0;
         for (auto& task : tasks) {
             ++i;
@@ -1675,7 +1675,7 @@ static void multiple_choice_score(llama_context * ctx, const gpt_params & params
 
     llama_batch_free(batch);
 
-    if (n_done < 100) return;
+    if (n_done < 100 && (params.multiple_choice_tasks != 0 && params.multiple_choice_tasks < (size_t)n_task)) return;
 
     float p = 1.f*n_correct/n_done;
     float sigma = sqrt(p*(1-p)/(n_done-1));

examples/quantize/quantize.cpp

@@ -284,7 +284,7 @@ int main(int argc, char ** argv) {
             } else {
                 usage(argv[0]);
             }
-        } else if (strcmp(argv[arg_idx], "--keep-split")) {
+        } else if (strcmp(argv[arg_idx], "--keep-split") == 0) {
             params.keep_split = true;
         } else {
             usage(argv[0]);

examples/rpc/rpc-server.cpp

@@ -56,6 +56,10 @@ static bool rpc_server_params_parse(int argc, char ** argv, rpc_server_params &
         } else if (arg == "-h" || arg == "--help") {
             print_usage(argc, argv, params);
             exit(0);
+        } else {
+            fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
+            print_usage(argc, argv, params);
+            exit(0);
         }
     }
     return true;

examples/server/README.md

@@ -18,8 +18,8 @@ The project is under active development, and we are [looking for feedback and co
 **Command line options:**
 
 - `-v`, `--verbose`: Enable verbose server output. When using the `/completion` endpoint, this includes the tokenized prompt, the full request and the full response.
-- `-t N`, `--threads N`: Set the number of threads to use during generation. Not used if model layers are offloaded to GPU. The server is using batching. This parameter is used only if one token is to be processed on CPU backend.
-- `-tb N, --threads-batch N`: Set the number of threads to use during batch and prompt processing. If not specified, the number of threads will be set to the number of threads used for generation. Not used if model layers are offloaded to GPU.
+- `-t N`, `--threads N`: Set the number of threads to use by CPU layers during generation. Not used by model layers that are offloaded to GPU. This option has no effect when using the maximum number of GPU layers. Default: `std::thread::hardware_concurrency()` (number of CPU cores).
+- `-tb N, --threads-batch N`: Set the number of threads to use by CPU layers during batch and prompt processing (>= 32 tokens). This option has no effect if a GPU is available. Default: `--threads`.
 - `--threads-http N`: Number of threads in the http server pool to process requests. Default: `max(std::thread::hardware_concurrency() - 1, --parallel N + 2)`
 - `-m FNAME`, `--model FNAME`: Specify the path to the LLaMA model file (e.g., `models/7B/ggml-model.gguf`).
 - `-mu MODEL_URL --model-url MODEL_URL`: Specify a remote http url to download the file. Default: unused
@@ -48,7 +48,7 @@ The project is under active development, and we are [looking for feedback and co
 - `--api-key`: Set an api key for request authorization. By default, the server responds to every request. With an api key set, the requests must have the Authorization header set with the api key as Bearer token. May be used multiple times to enable multiple valid keys.
 - `--api-key-file`: Path to file containing api keys delimited by new lines. If set, requests must include one of the keys for access. May be used in conjunction with `--api-key`s.
 - `--embeddings`: Enable embedding vector output and the OAI compatible endpoint /v1/embeddings. Physical batch size (`--ubatch-size`) must be carefully defined. Default: disabled
-- `-np N`, `--parallel N`: Set the number of slots for process requests. Default: `1`
+- `-np N`, `--parallel N`: Set the number of slots for process requests. Default: `1`. Values > 1 will allow for higher throughput with multiple parallel requests but the results will **not** be deterministic due to differences in rounding error.
 - `-cb`, `--cont-batching`: Enable continuous batching (a.k.a dynamic batching).  Default: disabled
 - `-spf FNAME`, `--system-prompt-file FNAME` Set a file to load a system prompt (initial prompt of all slots). This is useful for chat applications. [See more](#change-system-prompt-on-runtime)
 - `--mmproj MMPROJ_FILE`: Path to a multimodal projector file for LLaVA.

examples/server/server.cpp

@@ -102,7 +102,6 @@ struct slot_params {
     bool stream       = true;
     bool cache_prompt = false; // remember the prompt to avoid reprocessing all prompt
 
-    uint32_t seed      = -1; // RNG seed
     int32_t  n_keep    =  0; // number of tokens to keep from initial prompt
     int32_t  n_discard =  0; // number of tokens after n_keep that may be discarded when shifting context, 0 defaults to half
     int32_t  n_predict = -1; // new tokens to predict
@@ -1264,7 +1263,7 @@ struct server_context {
             {"n_ctx",                     slot.n_ctx},
             {"n_predict",                 slot.n_predict},
             {"model",                     params.model_alias},
-            {"seed",                      slot.params.seed},
+            {"seed",                      slot.sparams.seed},
             {"temperature",               slot.sparams.temp},
             {"dynatemp_range",            slot.sparams.dynatemp_range},
             {"dynatemp_exponent",         slot.sparams.dynatemp_exponent},

examples/server/tests/features/results.feature

@@ -70,12 +70,48 @@ Feature: Results
     Then all predictions are equal
     Examples:
       | n_parallel | temp |
-      |  1         | 0.0  |
-      |  2         | 0.0  |
-      |  4         | 0.0  |
-      |  1         | 1.0  |
-      # FIXME: These tests fail on master. The problem seems to be the unified KV cache.
+      | 1          | 0.0  |
+      | 2          | 0.0  |
+      | 4          | 0.0  |
+      | 1          | 1.0  |
+      # FIXME: These tests fail on master.
+      # Problems: unified KV cache (except for CPU backend with LLAMA_NO_LLAMAFILE=1), SIMD nondeterminism.
       # See https://github.com/ggerganov/whisper.cpp/issues/1941#issuecomment-1986923227
-      # and https://github.com/ggerganov/llama.cpp/pull/6122#discussion_r1531405574 .
-      # |  2         | 1.0  |
-      # |  4         | 1.0  |
+      # and https://github.com/ggerganov/llama.cpp/pull/6122#discussion_r1531405574
+      # and https://github.com/ggerganov/llama.cpp/pull/7347 .
+      # | 2          | 1.0  |
+      # | 4          | 1.0  |
+
+  Scenario Outline: consistent token probs with same seed and prompt
+    Given <n_slots> slots
+    And   <n_kv> KV cache size
+    And   1.0 temperature
+    And   <n_predict> max tokens to predict
+    Then  the server is starting
+    Then  the server is healthy
+
+    Given 1 prompts "The meaning of life is" with seed 42
+    And   concurrent completion requests
+    # Then the server is busy # Not all slots will be utilized.
+    Then  the server is idle
+    And   all slots are idle
+
+    Given <n_parallel> prompts "The meaning of life is" with seed 42
+    And   concurrent completion requests
+    # Then the server is busy # Not all slots will be utilized.
+    Then the server is idle
+    And  all slots are idle
+
+    Then all token probabilities are equal
+    Examples:
+      | n_slots | n_kv | n_predict | n_parallel |
+      | 4       | 1024 | 1         | 1          |
+      | 4       | 1024 | 1         | 4          |
+      # FIXME: These tests fail on master.
+      # Problems: unified KV cache (except for CPU backend with LLAMA_NO_LLAMAFILE=1), SIMD nondeterminism.
+      # See https://github.com/ggerganov/whisper.cpp/issues/1941#issuecomment-1986923227
+      # and https://github.com/ggerganov/llama.cpp/pull/6122#discussion_r1531405574
+      # and https://github.com/ggerganov/llama.cpp/pull/7347 .
+      # | 4       | 1024 | 100       | 1          |
+      # This test still fails even the above patches; the first token probabilities are already different.
+      # | 4       | 1024 | 100       | 4          |

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

@@ -23,6 +23,7 @@ from prometheus_client import parser
 def step_server_config(context, server_fqdn, server_port):
     context.server_fqdn = server_fqdn
     context.server_port = int(server_port)
+    context.n_threads = None
     context.n_gpu_layer = None
     if 'PORT' in os.environ:
         context.server_port = int(os.environ['PORT'])
@@ -109,6 +110,11 @@ def step_n_gpu_layer(context, ngl):
     context.n_gpu_layer = ngl
 
 
+@step('{n_threads:d} threads')
+def step_n_threads(context, n_threads):
+    context.n_thread = n_threads
+
+
 @step('{draft:d} as draft')
 def step_draft(context, draft):
     context.draft = draft
@@ -274,13 +280,22 @@ async def step_predictions_equal(context):
 
 @step('all predictions are different')
 @async_run_until_complete
-async def step_predictions_equal(context):
+async def step_predictions_different(context):
     n_completions = await gather_tasks_results(context)
     assert n_completions >= 2, "need at least 2 completions"
     assert_all_predictions_different(context.tasks_result)
     context.tasks_result = []
 
 
+@step('all token probabilities are equal')
+@async_run_until_complete
+async def step_token_probabilities_equal(context):
+    n_completions = await gather_tasks_results(context)
+    assert n_completions >= 2, "need at least 2 completions"
+    assert_all_token_probabilities_equal(context.tasks_result)
+    context.tasks_result = []
+
+
 @step('the completion is  truncated')
 def step_assert_completion_truncated(context):
     step_assert_completion_truncated(context, '')
@@ -869,6 +884,7 @@ async def request_completion(prompt,
                                     "id_slot": id_slot,
                                     "seed": seed if seed is not None else 42,
                                     "temperature": temperature if temperature is not None else "0.8f",
+                                    "n_probs": 2,
                                 },
                                 headers=headers,
                                 timeout=3600) as response:
@@ -1123,6 +1139,23 @@ def assert_all_predictions_different(completion_responses):
         assert content_i != content_j, "contents not different"
 
 
+def assert_all_token_probabilities_equal(completion_responses):
+    n_predict = len(completion_responses[0]['completion_probabilities'])
+    if 'DEBUG' in os.environ and os.environ['DEBUG'] == 'ON':
+        for pos in range(n_predict):
+            for i, response_i in enumerate(completion_responses):
+                probs_i = response_i['completion_probabilities'][pos]['probs']
+                print(f"pos {pos}, probs {i}: {probs_i}")
+    for pos in range(n_predict):
+        for i, response_i in enumerate(completion_responses):
+            probs_i = response_i['completion_probabilities'][pos]['probs']
+            for j, response_j in enumerate(completion_responses):
+                if i == j:
+                    continue
+                probs_j = response_j['completion_probabilities'][pos]['probs']
+            assert probs_i == probs_j, "contents not equal"
+
+
 async def gather_tasks_results(context):
     n_tasks = len(context.concurrent_tasks)
     if context.debug:
@@ -1261,6 +1294,8 @@ def start_server_background(context):
         server_args.extend(['--batch-size', context.n_batch])
     if context.n_ubatch:
         server_args.extend(['--ubatch-size', context.n_ubatch])
+    if context.n_threads:
+        server_args.extend(['--threads', context.threads])
     if context.n_gpu_layer:
         server_args.extend(['--n-gpu-layers', context.n_gpu_layer])
     if context.draft is not None:

ggml-cuda.cu

@@ -43,19 +43,59 @@
 #include <mutex>
 #include <stdint.h>
 #include <stdio.h>
+#include <stdarg.h>
+#include <stdlib.h>
 #include <string>
 #include <vector>
 
 static_assert(sizeof(half) == sizeof(ggml_fp16_t), "wrong fp16 size");
 
+static void ggml_cuda_default_log_callback(enum ggml_log_level level, const char * msg, void * user_data) {
+    GGML_UNUSED(level);
+    GGML_UNUSED(user_data);
+    fprintf(stderr, "%s", msg);
+}
+
+ggml_log_callback ggml_cuda_log_callback = ggml_cuda_default_log_callback;
+void * ggml_cuda_log_user_data = NULL;
+
+GGML_API void ggml_backend_cuda_log_set_callback(ggml_log_callback log_callback, void * user_data) {
+    ggml_cuda_log_callback = log_callback;
+    ggml_cuda_log_user_data = user_data;
+}
+
+#define GGML_CUDA_LOG_INFO(...) ggml_cuda_log(GGML_LOG_LEVEL_INFO, __VA_ARGS__)
+#define GGML_CUDA_LOG_WARN(...) ggml_cuda_log(GGML_LOG_LEVEL_WARN, __VA_ARGS__)
+#define GGML_CUDA_LOG_ERROR(...) ggml_cuda_log(GGML_LOG_LEVEL_ERROR, __VA_ARGS__)
+
+GGML_ATTRIBUTE_FORMAT(2, 3)
+static void ggml_cuda_log(enum ggml_log_level level, const char * format, ...) {
+    if (ggml_cuda_log_callback != NULL) {
+        va_list args;
+        va_start(args, format);
+        char buffer[128];
+        int len = vsnprintf(buffer, 128, format, args);
+        if (len < 128) {
+            ggml_cuda_log_callback(level, buffer, ggml_cuda_log_user_data);
+        } else {
+            std::vector<char> buffer2(len + 1);  // vsnprintf adds a null terminator
+            va_end(args);
+            va_start(args, format);
+            vsnprintf(&buffer2[0], buffer2.size(), format, args);
+            ggml_cuda_log_callback(level, buffer2.data(), ggml_cuda_log_user_data);
+        }
+        va_end(args);
+    }
+}
+
 [[noreturn]]
 void ggml_cuda_error(const char * stmt, const char * func, const char * file, int line, const char * msg) {
     int id = -1; // in case cudaGetDevice fails
     cudaGetDevice(&id);
 
-    fprintf(stderr, "CUDA error: %s\n", msg);
-    fprintf(stderr, "  current device: %d, in function %s at %s:%d\n", id, func, file, line);
-    fprintf(stderr, "  %s\n", stmt);
+    GGML_CUDA_LOG_ERROR("CUDA error: %s\n", msg);
+    GGML_CUDA_LOG_ERROR("  current device: %d, in function %s at %s:%d\n", id, func, file, line);
+    GGML_CUDA_LOG_ERROR("  %s\n", stmt);
     // abort with GGML_ASSERT to get a stack trace
     GGML_ASSERT(!"CUDA error");
 }
@@ -91,7 +131,7 @@ static ggml_cuda_device_info ggml_cuda_init() {
 
     cudaError_t err = cudaGetDeviceCount(&info.device_count);
     if (err != cudaSuccess) {
-        fprintf(stderr, "%s: failed to initialize " GGML_CUDA_NAME ": %s\n", __func__, cudaGetErrorString(err));
+        GGML_CUDA_LOG_ERROR("%s: failed to initialize " GGML_CUDA_NAME ": %s\n", __func__, cudaGetErrorString(err));
         return info;
     }
 
@@ -99,16 +139,16 @@ static ggml_cuda_device_info ggml_cuda_init() {
 
     int64_t total_vram = 0;
 #if defined(GGML_CUDA_FORCE_MMQ)
-    fprintf(stderr, "%s: GGML_CUDA_FORCE_MMQ:   yes\n", __func__);
+    GGML_CUDA_LOG_INFO("%s: GGML_CUDA_FORCE_MMQ:   yes\n", __func__);
 #else
-    fprintf(stderr, "%s: GGML_CUDA_FORCE_MMQ:   no\n", __func__);
+    GGML_CUDA_LOG_INFO("%s: GGML_CUDA_FORCE_MMQ:   no\n", __func__);
 #endif
 #if defined(CUDA_USE_TENSOR_CORES)
-    fprintf(stderr, "%s: CUDA_USE_TENSOR_CORES: yes\n", __func__);
+    GGML_CUDA_LOG_INFO("%s: CUDA_USE_TENSOR_CORES: yes\n", __func__);
 #else
-    fprintf(stderr, "%s: CUDA_USE_TENSOR_CORES: no\n", __func__);
+    GGML_CUDA_LOG_INFO("%s: CUDA_USE_TENSOR_CORES: no\n", __func__);
 #endif
-    fprintf(stderr, "%s: found %d " GGML_CUDA_NAME " devices:\n", __func__, info.device_count);
+    GGML_CUDA_LOG_INFO("%s: found %d " GGML_CUDA_NAME " devices:\n", __func__, info.device_count);
     for (int id = 0; id < info.device_count; ++id) {
         int device_vmm = 0;
 
@@ -129,7 +169,7 @@ static ggml_cuda_device_info ggml_cuda_init() {
 
         cudaDeviceProp prop;
         CUDA_CHECK(cudaGetDeviceProperties(&prop, id));
-        fprintf(stderr, "  Device %d: %s, compute capability %d.%d, VMM: %s\n", id, prop.name, prop.major, prop.minor, device_vmm ? "yes" : "no");
+        GGML_CUDA_LOG_INFO("  Device %d: %s, compute capability %d.%d, VMM: %s\n", id, prop.name, prop.major, prop.minor, device_vmm ? "yes" : "no");
 
         info.default_tensor_split[id] = total_vram;
         total_vram += prop.totalGlobalMem;
@@ -235,8 +275,8 @@ struct ggml_cuda_pool_leg : public ggml_cuda_pool {
         *actual_size = look_ahead_size;
         pool_size += look_ahead_size;
 #ifdef DEBUG_CUDA_MALLOC
-        fprintf(stderr, "%s[%d]: %d buffers, max_size = %u MB, pool_size = %u MB, requested %u MB\n", __func__, device, nnz,
-                (uint32_t)(max_size/1024/1024), (uint32_t)(pool_size/1024/1024), (uint32_t)(size/1024/1024));
+        GGML_CUDA_LOG_INFO("%s[%d]: %d buffers, max_size = %u MB, pool_size = %u MB, requested %u MB\n", __func__, device, nnz,
+                           (uint32_t)(max_size / 1024 / 1024), (uint32_t)(pool_size / 1024 / 1024), (uint32_t)(size / 1024 / 1024));
 #endif
         return ptr;
     }
@@ -250,7 +290,7 @@ struct ggml_cuda_pool_leg : public ggml_cuda_pool {
                 return;
             }
         }
-        fprintf(stderr, "WARNING: cuda buffer pool full, increase MAX_CUDA_BUFFERS\n");
+        GGML_CUDA_LOG_WARN("Cuda buffer pool full, increase MAX_CUDA_BUFFERS\n");
         ggml_cuda_set_device(device);
         CUDA_CHECK(cudaFree(ptr));
         pool_size -= size;
@@ -499,7 +539,9 @@ GGML_CALL static ggml_backend_buffer_t ggml_backend_cuda_buffer_type_alloc_buffe
     void * dev_ptr;
     cudaError_t err = cudaMalloc(&dev_ptr, size);
     if (err != cudaSuccess) {
-        fprintf(stderr, "%s: allocating %.2f MiB on device %d: cudaMalloc failed: %s\n", __func__, size/1024.0/1024.0, buft_ctx->device, cudaGetErrorString(err));
+        // clear the error
+        cudaGetLastError();
+        GGML_CUDA_LOG_ERROR("%s: allocating %.2f MiB on device %d: cudaMalloc failed: %s\n", __func__, size / 1024.0 / 1024.0, buft_ctx->device, cudaGetErrorString(err));
         return nullptr;
     }
 
@@ -1002,8 +1044,8 @@ static void * ggml_cuda_host_malloc(size_t size) {
     if (err != cudaSuccess) {
         // clear the error
         cudaGetLastError();
-        fprintf(stderr, "%s: warning: failed to allocate %.2f MiB of pinned memory: %s\n", __func__,
-            size/1024.0/1024.0, cudaGetErrorString(err));
+        GGML_CUDA_LOG_WARN("%s: failed to allocate %.2f MiB of pinned memory: %s\n", __func__,
+                           size / 1024.0 / 1024.0, cudaGetErrorString(err));
         return nullptr;
     }
 
@@ -2246,7 +2288,7 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
             break;
         case GGML_OP_MUL_MAT:
             if (dst->src[0]->ne[3] != dst->src[1]->ne[3]) {
-                fprintf(stderr, "%s: cannot compute %s: src0->ne[3] = %" PRId64 ", src1->ne[3] = %" PRId64 " - fallback to CPU\n", __func__, dst->name, dst->src[0]->ne[3], dst->src[1]->ne[3]);
+                GGML_CUDA_LOG_ERROR("%s: cannot compute %s: src0->ne[3] = %" PRId64 ", src1->ne[3] = %" PRId64 " - fallback to CPU\n", __func__, dst->name, dst->src[0]->ne[3], dst->src[1]->ne[3]);
                 return false;
             } else {
                 ggml_cuda_mul_mat(ctx, dst->src[0], dst->src[1], dst);
@@ -2300,7 +2342,7 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
 
     cudaError_t err = cudaGetLastError();
     if (err != cudaSuccess) {
-        fprintf(stderr, "%s: %s failed\n", __func__, ggml_op_desc(dst));
+        GGML_CUDA_LOG_ERROR("%s: %s failed\n", __func__, ggml_op_desc(dst));
         CUDA_CHECK(err);
     }
 
@@ -2476,7 +2518,7 @@ GGML_CALL static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t
         if (ggml_cuda_info().devices[cuda_ctx->device].cc < CC_AMPERE) {
             cuda_ctx->cuda_graph->disable_due_to_gpu_arch = true;
 #ifndef NDEBUG
-            fprintf(stderr, "%s: disabling CUDA graphs due to GPU architecture\n", __func__);
+            GGML_CUDA_LOG_WARN("%s: disabling CUDA graphs due to GPU architecture\n", __func__);
 #endif
         }
     }
@@ -2523,14 +2565,14 @@ GGML_CALL static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t
             if (node->src[0] && ggml_backend_buffer_is_cuda_split(node->src[0]->buffer)) {
                 use_cuda_graph = false; // Split buffers are not supported by CUDA graph capture
 #ifndef NDEBUG
-                fprintf(stderr, "%s: disabling CUDA graphs due to split buffer\n", __func__);
+                GGML_CUDA_LOG_WARN("%s: disabling CUDA graphs due to split buffer\n", __func__);
 #endif
             }
 
             if (node->op == GGML_OP_MUL_MAT_ID) {
                 use_cuda_graph = false; // This node type is not supported by CUDA graph capture
 #ifndef NDEBUG
-                fprintf(stderr, "%s: disabling CUDA graphs due to mul_mat_id\n", __func__);
+                GGML_CUDA_LOG_WARN("%s: disabling CUDA graphs due to mul_mat_id\n", __func__);
 #endif
             }
 
@@ -2539,7 +2581,7 @@ GGML_CALL static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t
                 // Changes in batch size or context size can cause changes to the grid size of some kernels.
                 use_cuda_graph = false;
 #ifndef NDEBUG
-                fprintf(stderr, "%s: disabling CUDA graphs due to batch size > 1 [%s] [%ld %ld %ld %ld]\n", __func__, node->name, node->ne[0], node->ne[1], node->ne[2], node->ne[3]);
+                GGML_CUDA_LOG_WARN("%s: disabling CUDA graphs due to batch size > 1 [%s] [%ld %ld %ld %ld]\n", __func__, node->name, node->ne[0], node->ne[1], node->ne[2], node->ne[3]);
 #endif
             }
 
@@ -2567,7 +2609,7 @@ GGML_CALL static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t
         if (cuda_ctx->cuda_graph->number_consecutive_updates >= 4) {
             cuda_ctx->cuda_graph->disable_due_to_too_many_updates = true;
 #ifndef NDEBUG
-            fprintf(stderr, "%s: disabling CUDA graphs due to too many consecutive updates\n", __func__);
+            GGML_CUDA_LOG_WARN("%s: disabling CUDA graphs due to too many consecutive updates\n", __func__);
 #endif
         }
     }
@@ -2605,7 +2647,7 @@ GGML_CALL static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t
 
                 bool ok = ggml_cuda_compute_forward(*cuda_ctx, node);
                 if (!ok) {
-                    fprintf(stderr, "%s: error: op not supported %s (%s)\n", __func__, node->name, ggml_op_name(node->op));
+                    GGML_CUDA_LOG_ERROR("%s: op not supported %s (%s)\n", __func__, node->name, ggml_op_name(node->op));
                 }
                 GGML_ASSERT(ok);
             }
@@ -2624,7 +2666,7 @@ GGML_CALL static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t
                 use_cuda_graph = false;
                 cuda_ctx->cuda_graph->disable_due_to_failed_graph_capture = true;
 #ifndef NDEBUG
-                fprintf(stderr, "%s: disabling CUDA graphs due to failed graph capture\n", __func__);
+                GGML_CUDA_LOG_WARN("%s: disabling CUDA graphs due to failed graph capture\n", __func__);
 #endif
             } else {
                 graph_evaluated_or_captured = true; // CUDA graph has been captured
@@ -2691,7 +2733,7 @@ GGML_CALL static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t
         cudaError_t stat = cudaGraphExecUpdate(cuda_ctx->cuda_graph->instance, cuda_ctx->cuda_graph->graph, &result_info);
         if (stat == cudaErrorGraphExecUpdateFailure) {
 #ifndef NDEBUG
-            fprintf(stderr, "%s: CUDA graph update failed\n", __func__);
+            GGML_CUDA_LOG_ERROR("%s: CUDA graph update failed\n", __func__);
 #endif
             // The pre-existing graph exec cannot be updated due to violated constraints
             // so instead clear error and re-instantiate
@@ -2948,13 +2990,13 @@ static ggml_guid_t ggml_backend_cuda_guid() {
 
 GGML_CALL ggml_backend_t ggml_backend_cuda_init(int device) {
     if (device < 0 || device >= ggml_backend_cuda_get_device_count()) {
-        fprintf(stderr, "%s: error: invalid device %d\n", __func__, device);
+        GGML_CUDA_LOG_ERROR("%s: invalid device %d\n", __func__, device);
         return nullptr;
     }
 
     ggml_backend_cuda_context * ctx = new ggml_backend_cuda_context(device);
     if (ctx == nullptr) {
-        fprintf(stderr, "%s: error: failed to allocate context\n", __func__);
+        GGML_CUDA_LOG_ERROR("%s: failed to allocate context\n", __func__);
         return nullptr;
     }
 
@@ -2998,8 +3040,8 @@ GGML_CALL bool ggml_backend_cuda_register_host_buffer(void * buffer, size_t size
         // clear the error
         cudaGetLastError();
 
-        fprintf(stderr, "%s: warning: failed to register %.2f MiB of pinned memory: %s\n", __func__,
-                size/1024.0/1024.0, cudaGetErrorString(err));
+        GGML_CUDA_LOG_WARN("%s: failed to register %.2f MiB of pinned memory: %s\n", __func__,
+                           size / 1024.0 / 1024.0, cudaGetErrorString(err));
         return false;
     }
     return true;

ggml-cuda.h

@@ -38,6 +38,7 @@ GGML_API GGML_CALL void ggml_backend_cuda_get_device_memory(int device, size_t *
 GGML_API GGML_CALL bool ggml_backend_cuda_register_host_buffer(void * buffer, size_t size);
 GGML_API GGML_CALL void ggml_backend_cuda_unregister_host_buffer(void * buffer);
 
+GGML_API void ggml_backend_cuda_log_set_callback(ggml_log_callback log_callback, void * user_data);
 #ifdef  __cplusplus
 }
 #endif

ggml-cuda/common.cuh

@@ -315,6 +315,20 @@ static __device__ __forceinline__ int __dp4a(const int a, const int b, int c) {
 #endif
     return c;
 }
+
+#if defined(__HIP_PLATFORM_AMD__) && HIP_VERSION < 50600000
+// __shfl_xor() for half2 was added in ROCm 5.6
+static __device__ __forceinline__ half2 __shfl_xor(half2 var, int laneMask, int width) {
+    typedef union half2_b32 {
+        half2 val;
+        int   b32;
+    } half2_b32_t;
+    half2_b32_t tmp;
+    tmp.val = var;
+    tmp.b32 = __shfl_xor(tmp.b32, laneMask, width);
+    return tmp.val;
+}
+#endif // defined(__HIP_PLATFORM_AMD__) && HIP_VERSION < 50600000
 #endif // defined(GGML_USE_HIPBLAS)
 
 #define FP16_AVAILABLE (defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) || __CUDA_ARCH__ >= CC_PASCAL
@@ -463,6 +477,17 @@ static const __device__ int8_t kvalues_iq4nl[16] = {-127, -104, -83, -65, -49, -
 
 typedef void (*dequantize_kernel_t)(const void * vx, const int64_t ib, const int iqs, dfloat2 & v);
 
+static __device__ __forceinline__ float get_alibi_slope(
+    const float max_bias, const uint32_t h, const uint32_t n_head_log2, const float m0, const float m1
+) {
+    if (max_bias <= 0.0f) {
+        return 1.0f;
+    }
+    const float base = h < n_head_log2 ? m0 : m1;
+    const int   exph = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1;
+
+    return powf(base, exph);
+}
 
 //////////////////////
 

ggml-cuda/fattn-common.cuh

@@ -1,7 +1,44 @@
+#include "common.cuh"
+
+#include <cstdint>
+
 #define FATTN_KQ_STRIDE       256
 #define HALF_MAX_HALF         __float2half(65504.0f/2) // Use neg. of this instead of -INFINITY to initialize KQ max vals to avoid NaN upon subtraction.
 #define SOFTMAX_FTZ_THRESHOLD -20.0f                   // Softmax exp. of values smaller than this are flushed to zero to avoid NaNs.
 
+typedef void (* fattn_kernel_t)(
+        const char * __restrict__ Q,
+        const char * __restrict__ K,
+        const char * __restrict__ V,
+        const char * __restrict__ mask,
+        float      * __restrict__ dst,
+        float2     * __restrict__ dst_meta,
+        const float scale,
+        const float max_bias,
+        const float m0,
+        const float m1,
+        const uint32_t n_head_log2,
+        const int ne00,
+        const int ne01,
+        const int ne02,
+        const int ne03,
+        const int ne10,
+        const int ne11,
+        const int ne12,
+        const int ne13,
+        const int ne31,
+        const int nb31,
+        const int nb01,
+        const int nb02,
+        const int nb03,
+        const int nb11,
+        const int nb12,
+        const int nb13,
+        const int ne0,
+        const int ne1,
+        const int ne2,
+        const int ne3);
+
 template<int D, int parallel_blocks> // D == head size
 #if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
 __launch_bounds__(D, 1)
@@ -45,3 +82,81 @@ static __global__ void flash_attn_combine_results(
 
     dst[blockIdx.y*D + tid] = VKQ_numerator / VKQ_denominator;
 }
+
+template <int D, int parallel_blocks>
+void launch_fattn(ggml_backend_cuda_context & ctx, ggml_tensor * dst, fattn_kernel_t fattn_kernel, int nwarps, int cols_per_block) {
+    const ggml_tensor * Q = dst->src[0];
+    const ggml_tensor * K = dst->src[1];
+    const ggml_tensor * V = dst->src[2];
+
+    const ggml_tensor * mask = dst->src[3];
+
+    ggml_tensor * KQV = dst;
+
+    GGML_ASSERT(Q->type == GGML_TYPE_F32);
+    GGML_ASSERT(K->type == GGML_TYPE_F16);
+    GGML_ASSERT(V->type == GGML_TYPE_F16);
+    GGML_ASSERT(KQV->type == GGML_TYPE_F32);
+
+    GGML_ASSERT(!mask || mask->type == GGML_TYPE_F16);
+    GGML_ASSERT(!mask || mask->ne[1] >= GGML_PAD(Q->ne[1], 16) &&
+                                "the Flash-Attention CUDA kernel requires the mask to be padded to 16 and at least n_queries big");
+
+    GGML_ASSERT(K->ne[1] % FATTN_KQ_STRIDE == 0 && "Incorrect KV cache padding.");
+
+    ggml_cuda_pool & pool = ctx.pool();
+    cudaStream_t main_stream = ctx.stream();
+
+    ggml_cuda_pool_alloc<float>  dst_tmp(pool);
+    ggml_cuda_pool_alloc<float2> dst_tmp_meta(pool);
+
+    if (parallel_blocks > 1) {
+        dst_tmp.alloc(parallel_blocks*ggml_nelements(KQV));
+        dst_tmp_meta.alloc(parallel_blocks*ggml_nrows(KQV));
+    }
+
+    const dim3 block_dim(WARP_SIZE, nwarps, 1);
+    const dim3 blocks_num(parallel_blocks*((Q->ne[1] + cols_per_block - 1) / cols_per_block), Q->ne[2], Q->ne[3]);
+    const int  shmem = 0;
+
+    float scale    = 1.0f;
+    float max_bias = 0.0f;
+
+    memcpy(&scale,    (float *) KQV->op_params + 0, sizeof(float));
+    memcpy(&max_bias, (float *) KQV->op_params + 1, sizeof(float));
+
+    const uint32_t n_head      = Q->ne[2];
+    const uint32_t n_head_log2 = 1u << (uint32_t) floorf(log2f((float) n_head));
+
+    const float m0 = powf(2.0f, -(max_bias       ) / n_head_log2);
+    const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2);
+
+    fattn_kernel<<<blocks_num, block_dim, shmem, main_stream>>>(
+        (const char *) Q->data,
+        (const char *) K->data,
+        (const char *) V->data,
+        mask ? ((const char *) mask->data) : nullptr,
+        (parallel_blocks) == 1 ? (float *) KQV->data : dst_tmp.ptr, dst_tmp_meta.ptr,
+        scale, max_bias, m0, m1, n_head_log2,
+        Q->ne[0], Q->ne[1], Q->ne[2], Q->ne[3],
+        K->ne[0], K->ne[1], K->ne[2], K->ne[3],
+        mask ? mask->ne[1] : 0, mask ?  mask->nb[1] : 0,
+        Q->nb[1], Q->nb[2], Q->nb[3],
+        K->nb[1], K->nb[2], K->nb[3],
+        KQV->ne[0], KQV->ne[1], KQV->ne[2], KQV->ne[3]
+    );
+    CUDA_CHECK(cudaGetLastError());
+
+    if ((parallel_blocks) == 1) {
+        return;
+    }
+
+    const dim3 block_dim_combine(D, 1, 1);
+    const dim3 blocks_num_combine(Q->ne[1], blocks_num.y, blocks_num.z);
+    const int  shmem_combine = 0;
+
+    flash_attn_combine_results<D, parallel_blocks>
+        <<<blocks_num_combine, block_dim_combine, shmem_combine, main_stream>>>
+        (dst_tmp.ptr, dst_tmp_meta.ptr, (float *) KQV->data);
+    CUDA_CHECK(cudaGetLastError());
+}

ggml-cuda/fattn-tile-f16.cu

@@ -0,0 +1,312 @@
+#include "common.cuh"
+#include "fattn-common.cuh"
+#include "fattn-tile-f16.cuh"
+
+#define FATTN_KQ_STRIDE_TILE_F16 64
+
+template<int D, int ncols, int nwarps, int parallel_blocks> // D == head size
+#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+__launch_bounds__(nwarps*WARP_SIZE, 1)
+#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+static __global__ void flash_attn_tile_ext_f16(
+        const char * __restrict__ Q,
+        const char * __restrict__ K,
+        const char * __restrict__ V,
+        const char * __restrict__ mask,
+        float      * __restrict__ dst,
+        float2     * __restrict__ dst_meta,
+        const float scale,
+        const float max_bias,
+        const float m0,
+        const float m1,
+        const uint32_t n_head_log2,
+        const int ne00,
+        const int ne01,
+        const int ne02,
+        const int ne03,
+        const int ne10,
+        const int ne11,
+        const int ne12,
+        const int ne13,
+        const int ne31,
+        const int nb31,
+        const int nb01,
+        const int nb02,
+        const int nb03,
+        const int nb11,
+        const int nb12,
+        const int nb13,
+        const int ne0,
+        const int ne1,
+        const int ne2,
+        const int ne3) {
+#if FP16_AVAILABLE
+    //In this kernel Q, K, V are matrices while i, j, k are matrix indices.
+
+    const int ic0 = (blockIdx.x / parallel_blocks) * ncols; // Index of the Q/QKV column to work on.
+    const int ip  =  blockIdx.x % parallel_blocks; // Index in group of blocks running for the same column in parallel.
+
+    const int gqa_ratio = ne02 / ne12; // With grouped query attention there are > 1 Q matrices per K, V matrix.
+    const float2 * Q_f2  = (const float2 *) (Q    + nb02* blockIdx.y              + nb01*ic0);
+    const half2  * K_h2  = (const half2  *) (K    + nb12*(blockIdx.y / gqa_ratio));
+    const half2  * V_h2  = (const half2  *) (V    + nb12*(blockIdx.y / gqa_ratio)); // K and V have same shape
+    const half   * maskh = (const half   *)  mask + ne11*ic0;
+
+    const int stride_KV2 = nb11 / sizeof(half2);
+
+    const float slopef = get_alibi_slope(max_bias, blockIdx.y, n_head_log2, m0, m1);
+    const half  slopeh = __float2half(slopef);
+
+    static_assert(D % (2*WARP_SIZE) == 0, "D not divisible by 2*WARP_SIZE == 64.");
+
+    __shared__ half KQ[ncols*FATTN_KQ_STRIDE_TILE_F16];
+    half2 * KQ2 = (half2 *) KQ;
+
+    __shared__ half2 KV_tmp[FATTN_KQ_STRIDE_TILE_F16][D/2 + 1]; // Pad D to avoid memory bank conflicts.
+
+    half kqmax[ncols/nwarps];
+#pragma unroll
+    for (int j0 = 0; j0 < ncols; j0 += nwarps) {
+        kqmax[j0/nwarps] = -HALF_MAX_HALF;
+    }
+    half2 kqsum[ncols/nwarps] = {{0.0f, 0.0f}};
+
+    half2 VKQ[ncols/nwarps][(D/2)/WARP_SIZE] = {{{0.0f, 0.0f}}};
+
+    // Convert Q to half2 and store in registers:
+    __shared__ half2 Q_h2[ncols][D/2];
+#pragma unroll
+    for (int j0 = 0; j0 < ncols; j0 += nwarps) {
+        const int j = j0 + threadIdx.y;
+
+#pragma unroll
+        for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) {
+            const int i = i0 + threadIdx.x;
+
+            const float2 tmp = Q_f2[j*(nb01/sizeof(float2)) + i];
+            Q_h2[j][i] = make_half2(scale, scale) * make_half2(tmp.x, tmp.y);
+        }
+    }
+
+    __syncthreads();
+
+    const int k_start = parallel_blocks == 1 ? 0 : ip*FATTN_KQ_STRIDE_TILE_F16;
+    for (int k_VKQ_0 = k_start; k_VKQ_0 < ne11; k_VKQ_0 += parallel_blocks*FATTN_KQ_STRIDE_TILE_F16) {
+        // Calculate KQ tile and keep track of new maximum KQ values:
+
+        half kqmax_new[ncols/nwarps];
+#pragma unroll
+        for (int j = 0; j < ncols/nwarps; ++j) {
+            kqmax_new[j] = kqmax[j];
+        }
+
+#pragma unroll
+        for (int i_KQ_0 = 0; i_KQ_0 < FATTN_KQ_STRIDE_TILE_F16; i_KQ_0 += nwarps) {
+            const int i_KQ = i_KQ_0 + threadIdx.y;
+
+#pragma unroll
+            for (int k_KQ_0 = 0; k_KQ_0 < D/2; k_KQ_0 += WARP_SIZE) {
+                const int k_KQ = k_KQ_0 + threadIdx.x;
+
+                KV_tmp[i_KQ][k_KQ] = K_h2[(k_VKQ_0 + i_KQ)*stride_KV2 + k_KQ];
+            }
+        }
+
+        __syncthreads();
+
+        half2 sum2[FATTN_KQ_STRIDE_TILE_F16/WARP_SIZE][ncols/nwarps] = {{{0.0f, 0.0f}}};
+
+#pragma unroll
+        for (int k_KQ = 0; k_KQ < D/2; ++k_KQ) {
+            half2 K_k[FATTN_KQ_STRIDE_TILE_F16/WARP_SIZE];
+            half2 Q_k[ncols/nwarps];
+
+#pragma unroll
+            for (int i_KQ_0 = 0; i_KQ_0 < FATTN_KQ_STRIDE_TILE_F16; i_KQ_0 += WARP_SIZE) {
+                const int i_KQ = i_KQ_0 + threadIdx.x;
+
+                K_k[i_KQ_0/WARP_SIZE] = KV_tmp[i_KQ][k_KQ];
+            }
+#pragma unroll
+            for (int j_KQ_0 = 0; j_KQ_0 < ncols; j_KQ_0 += nwarps) {
+                const int j_KQ = j_KQ_0 + threadIdx.y;
+
+                Q_k[j_KQ_0/nwarps] = Q_h2[j_KQ][k_KQ];
+            }
+
+#pragma unroll
+            for (int i_KQ_0 = 0; i_KQ_0 < FATTN_KQ_STRIDE_TILE_F16; i_KQ_0 += WARP_SIZE) {
+#pragma unroll
+                for (int j_KQ_0 = 0; j_KQ_0 < ncols; j_KQ_0 += nwarps) {
+                    sum2[i_KQ_0/WARP_SIZE][j_KQ_0/nwarps] += K_k[i_KQ_0/WARP_SIZE]*Q_k[j_KQ_0/nwarps];
+                }
+            }
+        }
+
+#pragma unroll
+        for (int i_KQ_0 = 0; i_KQ_0 < FATTN_KQ_STRIDE_TILE_F16; i_KQ_0 += WARP_SIZE) {
+            const int i_KQ = i_KQ_0 + threadIdx.x;
+
+#pragma unroll
+            for (int j_KQ_0 = 0; j_KQ_0 < ncols; j_KQ_0 += nwarps) {
+                const int j_KQ = j_KQ_0 + threadIdx.y;
+
+                half sum = __low2half(sum2[i_KQ_0/WARP_SIZE][j_KQ_0/nwarps]) + __high2half(sum2[i_KQ_0/WARP_SIZE][j_KQ_0/nwarps]);
+                sum += mask ? slopeh*maskh[j_KQ*ne11 + k_VKQ_0 + i_KQ] : __float2half(0.0f);
+
+                kqmax_new[j_KQ_0/nwarps] = ggml_cuda_hmax(kqmax_new[j_KQ_0/nwarps], sum);
+
+                KQ[j_KQ*FATTN_KQ_STRIDE_TILE_F16 + i_KQ] = sum;
+            }
+        }
+
+        __syncthreads();
+
+#pragma unroll
+        for (int j0 = 0; j0 < ncols; j0 += nwarps) {
+            const int j = j0 + threadIdx.y;
+
+            kqmax_new[j0/nwarps] = warp_reduce_max(kqmax_new[j0/nwarps]);
+            const half2 KQ_max_scale = __half2half2(hexp(kqmax[j0/nwarps] - kqmax_new[j0/nwarps]));
+            kqmax[j0/nwarps] = kqmax_new[j0/nwarps];
+
+#pragma unroll
+            for (int i0 = 0; i0 < FATTN_KQ_STRIDE_TILE_F16/2; i0 += WARP_SIZE) {
+                const int i = i0 + threadIdx.x;
+
+                const half2 diff = KQ2[j*(FATTN_KQ_STRIDE_TILE_F16/2) + i] - __half2half2(kqmax[j0/nwarps]);
+                const half2 val = h2exp(diff);
+                kqsum[j0/nwarps] = kqsum[j0/nwarps]*KQ_max_scale + val;
+                KQ2[j*(FATTN_KQ_STRIDE_TILE_F16/2) + i] = val;
+            }
+
+#pragma unroll
+            for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) {
+                VKQ[j0/nwarps][i0/WARP_SIZE] *= KQ_max_scale;
+            }
+        }
+
+        __syncthreads();
+
+#pragma unroll
+        for (int k0 = 0; k0 < FATTN_KQ_STRIDE_TILE_F16; k0 += nwarps) {
+            const int k = k0 + threadIdx.y;
+
+#pragma unroll
+            for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) {
+                const int i = i0 + threadIdx.x;
+
+                KV_tmp[k][i] = V_h2[(k_VKQ_0 + k)*stride_KV2 + i];
+            }
+        }
+
+        __syncthreads();
+
+#pragma unroll
+        for (int k0 = 0; k0 < FATTN_KQ_STRIDE_TILE_F16; k0 += 2) {
+            half2  V_k[(D/2)/WARP_SIZE][2];
+            half2 KQ_k[ncols/nwarps];
+
+#pragma unroll
+            for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) {
+                const int i = i0 + threadIdx.x;
+
+                V_k[i0/WARP_SIZE][0] = KV_tmp[k0 + 0][i];
+                V_k[i0/WARP_SIZE][1] = KV_tmp[k0 + 1][i];
+            }
+#pragma unroll
+            for (int j0 = 0; j0 < ncols; j0 += nwarps) {
+                const int j = j0 + threadIdx.y;
+
+                KQ_k[j0/nwarps] = KQ2[j*(FATTN_KQ_STRIDE_TILE_F16/2) + k0/2];
+            }
+
+#pragma unroll
+            for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) {
+#pragma unroll
+                for (int j0 = 0; j0 < ncols; j0 += nwarps) {
+                    VKQ[j0/nwarps][i0/WARP_SIZE] += V_k[i0/WARP_SIZE][0]* __low2half2(KQ_k[j0/nwarps]);
+                    VKQ[j0/nwarps][i0/WARP_SIZE] += V_k[i0/WARP_SIZE][1]*__high2half2(KQ_k[j0/nwarps]);
+                }
+            }
+        }
+
+        __syncthreads();
+    }
+
+#pragma unroll
+    for (int j_VKQ_0 = 0; j_VKQ_0 < ncols; j_VKQ_0 += nwarps) {
+        const int j_VKQ = j_VKQ_0 + threadIdx.y;
+
+        half kqsum_j = __low2half(kqsum[j_VKQ_0/nwarps]) + __high2half(kqsum[j_VKQ_0/nwarps]);
+        kqsum_j = warp_reduce_sum(kqsum_j);
+
+#pragma unroll
+        for (int i00 = 0; i00 < D; i00 += 2*WARP_SIZE) {
+            const int i0 = i00 + 2*threadIdx.x;
+
+            half2 dst_val = VKQ[j_VKQ_0/nwarps][i0/(2*WARP_SIZE)];
+            if (parallel_blocks == 1) {
+                dst_val /= __half2half2(kqsum_j);
+            }
+            const int j_dst = (ic0 + j_VKQ)*parallel_blocks + ip;
+            dst[j_dst*D*gridDim.y + D*blockIdx.y + i0 + 0] =  __low2float(dst_val);
+            dst[j_dst*D*gridDim.y + D*blockIdx.y + i0 + 1] = __high2float(dst_val);
+        }
+
+        if (parallel_blocks != 1 && threadIdx.x == 0) {
+            dst_meta[(ic0 + j_VKQ)*gridDim.y*parallel_blocks + blockIdx.y*parallel_blocks + ip] = make_float2(kqmax[j_VKQ_0/nwarps], kqsum_j);
+        }
+    }
+#else
+   NO_DEVICE_CODE;
+#endif // FP16_AVAILABLE
+}
+
+template <int cols_per_block, int parallel_blocks>
+void launch_fattn_tile_f16_64_128(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * Q = dst->src[0];
+    switch (Q->ne[0]) {
+        case  64: {
+            constexpr int      D = 64;
+            constexpr int nwarps = 8;
+            fattn_kernel_t fattn_kernel = flash_attn_tile_ext_f16<D, cols_per_block, nwarps, parallel_blocks>;
+            launch_fattn<D, parallel_blocks>(ctx, dst, fattn_kernel, nwarps, cols_per_block);
+        } break;
+        case 128: {
+            constexpr int      D = 128;
+            constexpr int nwarps = 8;
+            fattn_kernel_t fattn_kernel = flash_attn_tile_ext_f16<D, cols_per_block, nwarps, parallel_blocks>;
+            launch_fattn<D, parallel_blocks>(ctx, dst, fattn_kernel, nwarps, cols_per_block);
+        } break;
+        default: {
+            GGML_ASSERT(false && "FlashAttention without tensor cores only supports head sizes 64 and 128.");
+        } break;
+    }
+}
+
+void ggml_cuda_flash_attn_ext_tile_f16(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * KQV = dst;
+    const ggml_tensor * Q   = dst->src[0];
+
+    const int32_t precision = KQV->op_params[2];
+    GGML_ASSERT(precision == GGML_PREC_DEFAULT);
+
+    if (Q->ne[1] <= 16) {
+        constexpr int cols_per_block = 16;
+        constexpr int parallel_blocks = 4;
+        launch_fattn_tile_f16_64_128<cols_per_block, parallel_blocks>(ctx, dst);
+        return;
+    }
+
+    if (Q->ne[1] <= 32) {
+        constexpr int cols_per_block = 32;
+        constexpr int parallel_blocks = 4;
+        launch_fattn_tile_f16_64_128<cols_per_block, parallel_blocks>(ctx, dst);
+        return;
+    }
+
+    constexpr int cols_per_block = 32;
+    constexpr int parallel_blocks = 1;
+    launch_fattn_tile_f16_64_128<cols_per_block, parallel_blocks>(ctx, dst);
+}

ggml-cuda/fattn-tile-f16.cuh

@@ -0,0 +1,3 @@
+#include "common.cuh"
+
+void ggml_cuda_flash_attn_ext_tile_f16(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

ggml-cuda/fattn-tile-f32.cu

@@ -0,0 +1,309 @@
+#include "common.cuh"
+#include "fattn-common.cuh"
+#include "fattn-tile-f32.cuh"
+
+#define FATTN_KQ_STRIDE_TILE_F32 32
+
+template<int D, int ncols, int nwarps, int parallel_blocks> // D == head size
+#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+__launch_bounds__(nwarps*WARP_SIZE, 1)
+#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+static __global__ void flash_attn_tile_ext_f32(
+        const char * __restrict__ Q,
+        const char * __restrict__ K,
+        const char * __restrict__ V,
+        const char * __restrict__ mask,
+        float      * __restrict__ dst,
+        float2     * __restrict__ dst_meta,
+        const float scale,
+        const float max_bias,
+        const float m0,
+        const float m1,
+        const uint32_t n_head_log2,
+        const int ne00,
+        const int ne01,
+        const int ne02,
+        const int ne03,
+        const int ne10,
+        const int ne11,
+        const int ne12,
+        const int ne13,
+        const int ne31,
+        const int nb31,
+        const int nb01,
+        const int nb02,
+        const int nb03,
+        const int nb11,
+        const int nb12,
+        const int nb13,
+        const int ne0,
+        const int ne1,
+        const int ne2,
+        const int ne3) {
+    //In this kernel Q, K, V are matrices while i, j, k are matrix indices.
+
+    const int ic0 = (blockIdx.x / parallel_blocks) * ncols; // Index of the Q/QKV column to work on.
+    const int ip  =  blockIdx.x % parallel_blocks; // Index in group of blocks running for the same column in parallel.
+
+    const int gqa_ratio = ne02 / ne12; // With grouped query attention there are > 1 Q matrices per K, V matrix.
+    const float2 * Q_f2  = (const float2 *) (Q    + nb02* blockIdx.y              + nb01*ic0);
+    const half2  * K_h2  = (const half2  *) (K    + nb12*(blockIdx.y / gqa_ratio));
+    const half2  * V_h2  = (const half2  *) (V    + nb12*(blockIdx.y / gqa_ratio)); // K and V have same shape
+    const half   * maskh = (const half   *)  mask + ne11*ic0;
+
+    const int stride_KV2 = nb11 / sizeof(half2);
+
+    const float slope = get_alibi_slope(max_bias, blockIdx.y, n_head_log2, m0, m1);
+
+    static_assert(D % (2*WARP_SIZE) == 0, "D not divisible by 2*WARP_SIZE == 64.");
+
+    __shared__ float KQ[ncols*FATTN_KQ_STRIDE_TILE_F32];
+
+    __shared__ float KV_tmp[FATTN_KQ_STRIDE_TILE_F32][D + 1]; // Pad D to avoid memory bank conflicts.
+    float2 * KV_tmp2 = (float2 *) KV_tmp;
+
+    float kqmax[ncols/nwarps];
+#pragma unroll
+    for (int j0 = 0; j0 < ncols; j0 += nwarps) {
+        kqmax[j0/nwarps] = -FLT_MAX/2.0f;
+    }
+    float kqsum[ncols/nwarps] = {0.0f};
+
+    float2 VKQ[ncols/nwarps][(D/2)/WARP_SIZE] = {{{0.0f, 0.0f}}};
+
+    // Convert Q to half2 and store in registers:
+    __shared__ float Q_f[ncols][D];
+#pragma unroll
+    for (int j0 = 0; j0 < ncols; j0 += nwarps) {
+        const int j = j0 + threadIdx.y;
+
+#pragma unroll
+        for (int i0 = 0; i0 < D; i0 += 2*WARP_SIZE) {
+            float2 tmp = Q_f2[j*(nb01/sizeof(float2)) + i0/2 + threadIdx.x];
+            Q_f[j][i0 + 0*WARP_SIZE + threadIdx.x] = tmp.x * scale;
+            Q_f[j][i0 + 1*WARP_SIZE + threadIdx.x] = tmp.y * scale;
+        }
+    }
+
+    __syncthreads();
+
+    const int k_start = parallel_blocks == 1 ? 0 : ip*FATTN_KQ_STRIDE_TILE_F32;
+    for (int k_VKQ_0 = k_start; k_VKQ_0 < ne11; k_VKQ_0 += parallel_blocks*FATTN_KQ_STRIDE_TILE_F32) {
+        // Calculate KQ tile and keep track of new maximum KQ values:
+
+        float kqmax_new[ncols/nwarps];
+#pragma unroll
+        for (int j = 0; j < ncols/nwarps; ++j) {
+            kqmax_new[j] = kqmax[j];
+        }
+
+#pragma unroll
+        for (int i_KQ_0 = 0; i_KQ_0 < FATTN_KQ_STRIDE_TILE_F32; i_KQ_0 += nwarps) {
+            const int i_KQ = i_KQ_0 + threadIdx.y;
+
+#pragma unroll
+            for (int k_KQ_0 = 0; k_KQ_0 < D; k_KQ_0 += 2*WARP_SIZE) {
+                const half2 tmp = K_h2[(k_VKQ_0 + i_KQ)*stride_KV2 + k_KQ_0/2 + threadIdx.x];
+                KV_tmp[i_KQ][k_KQ_0 + 0*WARP_SIZE + threadIdx.x] =  __low2float(tmp);
+                KV_tmp[i_KQ][k_KQ_0 + 1*WARP_SIZE + threadIdx.x] = __high2float(tmp);
+            }
+        }
+
+        __syncthreads();
+
+        float sum[FATTN_KQ_STRIDE_TILE_F32/WARP_SIZE][ncols/nwarps] = {{0.0f}};
+
+#pragma unroll
+        for (int k_KQ = 0; k_KQ < D; ++k_KQ) {
+            float K_k[FATTN_KQ_STRIDE_TILE_F32/WARP_SIZE];
+            float Q_k[ncols/nwarps];
+
+#pragma unroll
+            for (int i_KQ_0 = 0; i_KQ_0 < FATTN_KQ_STRIDE_TILE_F32; i_KQ_0 += WARP_SIZE) {
+                const int i_KQ = i_KQ_0 + threadIdx.x;
+
+                K_k[i_KQ_0/WARP_SIZE] = KV_tmp[i_KQ][k_KQ];
+            }
+#pragma unroll
+            for (int j_KQ_0 = 0; j_KQ_0 < ncols; j_KQ_0 += nwarps) {
+                const int j_KQ = j_KQ_0 + threadIdx.y;
+
+                Q_k[j_KQ_0/nwarps] = Q_f[j_KQ][k_KQ];
+            }
+
+#pragma unroll
+            for (int i_KQ_0 = 0; i_KQ_0 < FATTN_KQ_STRIDE_TILE_F32; i_KQ_0 += WARP_SIZE) {
+#pragma unroll
+                for (int j_KQ_0 = 0; j_KQ_0 < ncols; j_KQ_0 += nwarps) {
+                    sum[i_KQ_0/WARP_SIZE][j_KQ_0/nwarps] += K_k[i_KQ_0/WARP_SIZE] * Q_k[j_KQ_0/nwarps];
+                }
+            }
+        }
+
+#pragma unroll
+        for (int i_KQ_0 = 0; i_KQ_0 < FATTN_KQ_STRIDE_TILE_F32; i_KQ_0 += WARP_SIZE) {
+            const int i_KQ = i_KQ_0 + threadIdx.x;
+
+#pragma unroll
+            for (int j_KQ_0 = 0; j_KQ_0 < ncols; j_KQ_0 += nwarps) {
+                const int j_KQ = j_KQ_0 + threadIdx.y;
+
+                sum[i_KQ_0/WARP_SIZE][j_KQ_0/nwarps] += mask ? slope*__half2float(maskh[j_KQ*ne11 + k_VKQ_0 + i_KQ]) : 0.0f;
+
+                kqmax_new[j_KQ_0/nwarps] = fmaxf(kqmax_new[j_KQ_0/nwarps], sum[i_KQ_0/WARP_SIZE][j_KQ_0/nwarps]);
+
+                KQ[j_KQ*FATTN_KQ_STRIDE_TILE_F32 + i_KQ] = sum[i_KQ_0/WARP_SIZE][j_KQ_0/nwarps];
+            }
+        }
+
+        __syncthreads();
+
+#pragma unroll
+        for (int j0 = 0; j0 < ncols; j0 += nwarps) {
+            const int j = j0 + threadIdx.y;
+
+            kqmax_new[j0/nwarps] = warp_reduce_max(kqmax_new[j0/nwarps]);
+            const float KQ_max_scale = expf(kqmax[j0/nwarps] - kqmax_new[j0/nwarps]);
+            kqmax[j0/nwarps] = kqmax_new[j0/nwarps];
+
+            float kqsum_add = 0.0f;
+#pragma unroll
+            for (int i0 = 0; i0 < FATTN_KQ_STRIDE_TILE_F32; i0 += WARP_SIZE) {
+                const int i = i0 + threadIdx.x;
+
+                const float diff = KQ[j*FATTN_KQ_STRIDE_TILE_F32 + i] - kqmax[j0/nwarps];
+                const float val = expf(diff);
+                kqsum_add += val;
+                KQ[j*FATTN_KQ_STRIDE_TILE_F32 + i] = val;
+            }
+            kqsum[j0/nwarps] = kqsum[j0/nwarps]*KQ_max_scale + kqsum_add;
+
+#pragma unroll
+            for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) {
+                VKQ[j0/nwarps][i0/WARP_SIZE].x *= KQ_max_scale;
+                VKQ[j0/nwarps][i0/WARP_SIZE].y *= KQ_max_scale;
+            }
+        }
+
+        __syncthreads();
+
+#pragma unroll
+        for (int k0 = 0; k0 < FATTN_KQ_STRIDE_TILE_F32; k0 += nwarps) {
+            const int k = k0 + threadIdx.y;
+
+#pragma unroll
+            for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) {
+                const int i = i0 + threadIdx.x;
+
+                KV_tmp2[k*(D/2) + i].x =  __low2float(V_h2[(k_VKQ_0 + k)*stride_KV2 + i]);
+                KV_tmp2[k*(D/2) + i].y = __high2float(V_h2[(k_VKQ_0 + k)*stride_KV2 + i]);
+            }
+        }
+
+        __syncthreads();
+
+#pragma unroll
+        for (int k = 0; k < FATTN_KQ_STRIDE_TILE_F32; ++k) {
+            float2 V_k[(D/2)/WARP_SIZE];
+            float  KQ_k[ncols/nwarps];
+
+#pragma unroll
+            for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) {
+                const int i = i0 + threadIdx.x;
+
+                V_k[i0/WARP_SIZE] = KV_tmp2[k*(D/2) + i];
+            }
+#pragma unroll
+            for (int j0 = 0; j0 < ncols; j0 += nwarps) {
+                const int j = j0 + threadIdx.y;
+
+                KQ_k[j0/nwarps] = KQ[j*FATTN_KQ_STRIDE_TILE_F32 + k];
+            }
+
+#pragma unroll
+            for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) {
+#pragma unroll
+                for (int j0 = 0; j0 < ncols; j0 += nwarps) {
+                    VKQ[j0/nwarps][i0/WARP_SIZE].x += V_k[i0/WARP_SIZE].x*KQ_k[j0/nwarps];
+                    VKQ[j0/nwarps][i0/WARP_SIZE].y += V_k[i0/WARP_SIZE].y*KQ_k[j0/nwarps];
+                }
+            }
+        }
+
+        __syncthreads();
+    }
+
+#pragma unroll
+    for (int j_VKQ_0 = 0; j_VKQ_0 < ncols; j_VKQ_0 += nwarps) {
+        const int j_VKQ = j_VKQ_0 + threadIdx.y;
+
+        float kqsum_j = kqsum[j_VKQ_0/nwarps];
+        kqsum_j = warp_reduce_sum(kqsum_j);
+
+#pragma unroll
+        for (int i00 = 0; i00 < D; i00 += 2*WARP_SIZE) {
+            const int i0 = i00 + 2*threadIdx.x;
+
+            float2 dst_val = VKQ[j_VKQ_0/nwarps][i0/(2*WARP_SIZE)];
+            if (parallel_blocks == 1) {
+                dst_val.x /= kqsum_j;
+                dst_val.y /= kqsum_j;
+            }
+            const int j_dst = (ic0 + j_VKQ)*parallel_blocks + ip;
+            dst[j_dst*D*gridDim.y + D*blockIdx.y + i0 + 0] = dst_val.x;
+            dst[j_dst*D*gridDim.y + D*blockIdx.y + i0 + 1] = dst_val.y;
+        }
+
+        if (parallel_blocks != 1 && threadIdx.x == 0) {
+            dst_meta[(ic0 + j_VKQ)*gridDim.y*parallel_blocks + blockIdx.y*parallel_blocks + ip] = make_float2(kqmax[j_VKQ_0/nwarps], kqsum_j);
+        }
+    }
+}
+
+template <int cols_per_block, int parallel_blocks>
+void launch_fattn_tile_f32_64_128(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * Q = dst->src[0];
+    switch (Q->ne[0]) {
+        case  64: {
+            constexpr int      D = 64;
+            constexpr int nwarps = 8;
+            fattn_kernel_t fattn_kernel = flash_attn_tile_ext_f32<D, cols_per_block, nwarps, parallel_blocks>;
+            launch_fattn<D, parallel_blocks>(ctx, dst, fattn_kernel, nwarps, cols_per_block);
+        } break;
+        case 128: {
+            constexpr int      D = 128;
+            constexpr int nwarps = 8;
+            fattn_kernel_t fattn_kernel = flash_attn_tile_ext_f32<D, cols_per_block, nwarps, parallel_blocks>;
+            launch_fattn<D, parallel_blocks>(ctx, dst, fattn_kernel, nwarps, cols_per_block);
+        } break;
+        default: {
+            GGML_ASSERT(false && "FlashAttention without tensor cores only supports head sizes 64 and 128.");
+        } break;
+    }
+}
+
+void ggml_cuda_flash_attn_ext_tile_f32(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * KQV = dst;
+    const ggml_tensor * Q   = dst->src[0];
+
+    const int32_t precision = KQV->op_params[2];
+    GGML_ASSERT(precision == GGML_PREC_DEFAULT);
+
+    if (Q->ne[1] <= 16) {
+        constexpr int cols_per_block = 16;
+        constexpr int parallel_blocks = 4;
+        launch_fattn_tile_f32_64_128<cols_per_block, parallel_blocks>(ctx, dst);
+        return;
+    }
+
+    if (Q->ne[1] <= 32) {
+        constexpr int cols_per_block = 32;
+        constexpr int parallel_blocks = 4;
+        launch_fattn_tile_f32_64_128<cols_per_block, parallel_blocks>(ctx, dst);
+        return;
+    }
+
+    constexpr int cols_per_block = 32;
+    constexpr int parallel_blocks = 1;
+    launch_fattn_tile_f32_64_128<cols_per_block, parallel_blocks>(ctx, dst);
+}

ggml-cuda/fattn-tile-f32.cuh

@@ -0,0 +1,3 @@
+#include "common.cuh"
+
+void ggml_cuda_flash_attn_ext_tile_f32(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

ggml-cuda/fattn-vec-f16.cu

@@ -53,17 +53,8 @@ static __global__ void flash_attn_vec_ext_f16(
     const int stride_KV  = nb11 / sizeof(half);
     const int stride_KV2 = nb11 / sizeof(half2);
 
-    half  slopeh = __float2half(1.0f);
-
-    // ALiBi
-    if (max_bias > 0.0f) {
-        const int h = blockIdx.y;
-
-        const float base = h < n_head_log2 ? m0 : m1;
-        const int   exph = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1;
-
-        slopeh = __float2half(powf(base, exph));
-    }
+    const float slopef = get_alibi_slope(max_bias, blockIdx.y, n_head_log2, m0, m1);
+    const half  slopeh = __float2half(slopef);
 
     static_assert(D % (2*WARP_SIZE) == 0, "D not divisible by 2*WARP_SIZE == 64.");
     constexpr int nwarps = D / WARP_SIZE;
@@ -232,199 +223,104 @@ static __global__ void flash_attn_vec_ext_f16(
         dst[j_dst*D*gridDim.y + D*blockIdx.y + tid] = dst_val;
     }
 
-    if (parallel_blocks != 1 && tid != 0) {
-#pragma unroll
-        for (int j = 0; j < ncols; ++j) {
-            dst_meta[(ic0 + j)*gridDim.y*parallel_blocks + blockIdx.y*parallel_blocks + ip] = make_float2(kqmax[j], kqsum[j]);
-        }
+    if (parallel_blocks != 1 && tid < ncols) {
+        dst_meta[(ic0 + tid)*gridDim.y*parallel_blocks + blockIdx.y*parallel_blocks + ip] = make_float2(kqmax[tid], kqsum[tid]);
     }
 #else
    NO_DEVICE_CODE;
 #endif // FP16_AVAILABLE
 }
 
-template <int D, int cols_per_block, int parallel_blocks> void launch_fattn_vec_f16(
-        const ggml_tensor * Q, const ggml_tensor * K, const ggml_tensor * V, ggml_tensor * KQV, const ggml_tensor * mask,
-        ggml_cuda_pool & pool, cudaStream_t main_stream
-) {
-    ggml_cuda_pool_alloc<float>  dst_tmp(pool);
-    ggml_cuda_pool_alloc<float2> dst_tmp_meta(pool);
-
-    if (parallel_blocks > 1) {
-        dst_tmp.alloc(parallel_blocks*ggml_nelements(KQV));
-        dst_tmp_meta.alloc(parallel_blocks*ggml_nrows(KQV));
-    }
-
-    constexpr int  nwarps = (D + WARP_SIZE - 1) / WARP_SIZE;
-    const     dim3 block_dim(WARP_SIZE, nwarps, 1);
-    const     dim3 blocks_num(parallel_blocks*((Q->ne[1] + cols_per_block - 1) / cols_per_block), Q->ne[2], Q->ne[3]);
-    const     int  shmem = 0;
-
-    float scale    = 1.0f;
-    float max_bias = 0.0f;
-
-    memcpy(&scale,    (float *) KQV->op_params + 0, sizeof(float));
-    memcpy(&max_bias, (float *) KQV->op_params + 1, sizeof(float));
-
-    const uint32_t n_head      = Q->ne[2];
-    const uint32_t n_head_log2 = 1u << (uint32_t) floorf(log2f((float) n_head));
-
-    const float m0 = powf(2.0f, -(max_bias       ) / n_head_log2);
-    const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2);
-
-    flash_attn_vec_ext_f16<D, cols_per_block, parallel_blocks>
-        <<<blocks_num, block_dim, shmem, main_stream>>> (
-                (const char *) Q->data,
-                (const char *) K->data,
-                (const char *) V->data,
-                mask ? ((const char *) mask->data) : nullptr,
-                parallel_blocks == 1 ? (float *) KQV->data : dst_tmp.ptr, dst_tmp_meta.ptr,
-                scale, max_bias, m0, m1, n_head_log2,
-                Q->ne[0], Q->ne[1], Q->ne[2], Q->ne[3],
-                K->ne[0], K->ne[1], K->ne[2], K->ne[3],
-                mask ? mask->ne[1] : 0, mask ?  mask->nb[1] : 0,
-                Q->nb[1], Q->nb[2], Q->nb[3],
-                K->nb[1], K->nb[2], K->nb[3],
-                KQV->ne[0], KQV->ne[1], KQV->ne[2], KQV->ne[3]
-                );
-    CUDA_CHECK(cudaGetLastError());
-
-    if (parallel_blocks == 1) {
-        return;
-    }
-
-    const dim3 block_dim_combine(D, 1, 1);
-    const dim3 blocks_num_combine(Q->ne[1], blocks_num.y, blocks_num.z);
-    const int  shmem_combine = 0;
-
-    flash_attn_combine_results<D, parallel_blocks>
-        <<<blocks_num_combine, block_dim_combine, shmem_combine, main_stream>>>
-        (dst_tmp.ptr, dst_tmp_meta.ptr, (float *) KQV->data);
-    CUDA_CHECK(cudaGetLastError());
-}
-
 void ggml_cuda_flash_attn_ext_vec_f16(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
-    const ggml_tensor * Q = dst->src[0];
-    const ggml_tensor * K = dst->src[1];
-    const ggml_tensor * V = dst->src[2];
-
-    const ggml_tensor * mask = dst->src[3];
-
     ggml_tensor * KQV = dst;
+    ggml_tensor * Q   = dst->src[0];
 
     const int32_t precision = KQV->op_params[2];
     GGML_ASSERT(precision == GGML_PREC_DEFAULT);
 
-    constexpr int cols_per_block = 1;
+    constexpr int cols_per_block  = 1;
     constexpr int parallel_blocks = 4;
     switch (Q->ne[0]) {
-        case 64:
-            launch_fattn_vec_f16< 64, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
-            break;
-        case 128:
-            launch_fattn_vec_f16<128, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
-            break;
-        case 256:
-            launch_fattn_vec_f16<256, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
-            break;
+        case  64: {
+            constexpr int      D = 64;
+            constexpr int nwarps = D/WARP_SIZE;
+            fattn_kernel_t fattn_kernel = flash_attn_vec_ext_f16<D, cols_per_block, parallel_blocks>;
+            launch_fattn<D, parallel_blocks>(ctx, dst, fattn_kernel, nwarps, cols_per_block);
+        } break;
+        case 128: {
+            constexpr int      D = 128;
+            constexpr int nwarps = D/WARP_SIZE;
+            fattn_kernel_t fattn_kernel = flash_attn_vec_ext_f16<D, cols_per_block, parallel_blocks>;
+            launch_fattn<D, parallel_blocks>(ctx, dst, fattn_kernel, nwarps, cols_per_block);
+        } break;
+        case 256: {
+            constexpr int      D = 256;
+            constexpr int nwarps = D/WARP_SIZE;
+            fattn_kernel_t fattn_kernel = flash_attn_vec_ext_f16<D, cols_per_block, parallel_blocks>;
+            launch_fattn<D, parallel_blocks>(ctx, dst, fattn_kernel, nwarps, cols_per_block);
+        } break;
         default:
             GGML_ASSERT(false);
             break;
     }
 }
 
+template <int cols_per_block, int parallel_blocks>
+void launch_fattn_vec_f16_64_128(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * Q = dst->src[0];
+    switch (Q->ne[0]) {
+        case  64: {
+            constexpr int      D = 64;
+            constexpr int nwarps = D/WARP_SIZE;
+            fattn_kernel_t fattn_kernel = flash_attn_vec_ext_f16<D, cols_per_block, parallel_blocks>;
+            launch_fattn<D, parallel_blocks>(ctx, dst, fattn_kernel, nwarps, cols_per_block);
+        } break;
+        case 128: {
+            constexpr int      D = 128;
+            constexpr int nwarps = D/WARP_SIZE;
+            fattn_kernel_t fattn_kernel = flash_attn_vec_ext_f16<D, cols_per_block, parallel_blocks>;
+            launch_fattn<D, parallel_blocks>(ctx, dst, fattn_kernel, nwarps, cols_per_block);
+        } break;
+        default: {
+            GGML_ASSERT(false && "FlashAttention without tensor cores only supports head sizes 64 and 128.");
+        } break;
+    }
+}
+
 void ggml_cuda_flash_attn_ext_vec_f16_no_mma(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
-    const ggml_tensor * Q = dst->src[0];
-    const ggml_tensor * K = dst->src[1];
-    const ggml_tensor * V = dst->src[2];
-
-    const ggml_tensor * mask = dst->src[3];
-
-    ggml_tensor * KQV = dst;
+    const ggml_tensor * KQV = dst;
+    const ggml_tensor * Q   = dst->src[0];
 
     const int32_t precision = KQV->op_params[2];
     GGML_ASSERT(precision == GGML_PREC_DEFAULT);
-    GGML_ASSERT(Q->ne[0] == 64 || Q->ne[0] == 128 && "FlashAttention without tensor cores only supports head sizes 64 and 128.");
 
     if (Q->ne[1] == 1) {
-        constexpr int cols_per_block = 1;
-        constexpr int parallel_blocks = 4;
-        switch (Q->ne[0]) {
-            case 64:
-                launch_fattn_vec_f16< 64, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
-                break;
-            case 128:
-                launch_fattn_vec_f16<128, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
-                break;
-            default:
-                GGML_ASSERT(false);
-                break;
-        }
+        ggml_cuda_flash_attn_ext_vec_f16(ctx, dst);
         return;
     }
 
     if (Q->ne[1] == 2) {
-        constexpr int cols_per_block = 2;
+        constexpr int cols_per_block  = 2;
         constexpr int parallel_blocks = 4;
-        switch (Q->ne[0]) {
-            case 64:
-                launch_fattn_vec_f16< 64, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
-                break;
-            case 128:
-                launch_fattn_vec_f16<128, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
-                break;
-            default:
-                GGML_ASSERT(false);
-                break;
-        }
+        launch_fattn_vec_f16_64_128<cols_per_block, parallel_blocks>(ctx, dst);
         return;
     }
 
     if (Q->ne[1] <= 4) {
-        constexpr int cols_per_block = 4;
+        constexpr int cols_per_block  = 4;
         constexpr int parallel_blocks = 4;
-        switch (Q->ne[0]) {
-            case 64:
-                launch_fattn_vec_f16< 64, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
-                break;
-            case 128:
-                launch_fattn_vec_f16<128, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
-                break;
-            default:
-                GGML_ASSERT(false);
-                break;
-        }
+        launch_fattn_vec_f16_64_128<cols_per_block, parallel_blocks>(ctx, dst);
         return;
     }
 
     if (Q->ne[1] <= 8) {
-        constexpr int cols_per_block = 8;
+        constexpr int cols_per_block  = 8;
         constexpr int parallel_blocks = 4;
-        switch (Q->ne[0]) {
-            case 64:
-                launch_fattn_vec_f16< 64, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
-                break;
-            case 128:
-                launch_fattn_vec_f16<128, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
-                break;
-            default:
-                GGML_ASSERT(false);
-                break;
-        }
+        launch_fattn_vec_f16_64_128<cols_per_block, parallel_blocks>(ctx, dst);
         return;
     }
 
-    constexpr int cols_per_block = 8;
+    constexpr int cols_per_block  = 8;
     constexpr int parallel_blocks = 1;
-    switch (Q->ne[0]) {
-        case 64:
-            launch_fattn_vec_f16< 64, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
-            break;
-        case 128:
-            launch_fattn_vec_f16<128, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
-            break;
-        default:
-            GGML_ASSERT(false);
-            break;
-    }
+    launch_fattn_vec_f16_64_128<cols_per_block, parallel_blocks>(ctx, dst);
 }

ggml-cuda/fattn-vec-f32.cu

@@ -52,17 +52,7 @@ static __global__ void flash_attn_vec_ext_f32(
     const int stride_KV  = nb11 / sizeof(half);
     const int stride_KV2 = nb11 / sizeof(half2);
 
-    float slope = 1.0f;
-
-    // ALiBi
-    if (max_bias > 0.0f) {
-        const int h = blockIdx.y;
-
-        const float base = h < n_head_log2 ? m0 : m1;
-        const int   exph = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1;
-
-        slope = powf(base, exph);
-    }
+    const float slope = get_alibi_slope(max_bias, blockIdx.y, n_head_log2, m0, m1);
 
     static_assert(D % (2*WARP_SIZE) == 0, "D not divisible by 2*WARP_SIZE == 64.");
     constexpr int nwarps = D / WARP_SIZE;
@@ -221,164 +211,65 @@ static __global__ void flash_attn_vec_ext_f32(
         dst[j_dst*D*gridDim.y + D*blockIdx.y + tid] = dst_val;
     }
 
-    if (parallel_blocks != 1 && tid != 0) {
-#pragma unroll
-        for (int j = 0; j < ncols; ++j) {
-            dst_meta[(ic0 + j)*gridDim.y*parallel_blocks + blockIdx.y*parallel_blocks + ip] = make_float2(kqmax[j], kqsum[j]);
-        }
+    if (parallel_blocks != 1 && tid < ncols) {
+        dst_meta[(ic0 + tid)*gridDim.y*parallel_blocks + blockIdx.y*parallel_blocks + ip] = make_float2(kqmax[tid], kqsum[tid]);
     }
 }
 
-template <int D, int cols_per_block, int parallel_blocks> void launch_fattn_vec_f32(
-        const ggml_tensor * Q, const ggml_tensor * K, const ggml_tensor * V, ggml_tensor * KQV, const ggml_tensor * mask,
-        ggml_cuda_pool & pool, cudaStream_t main_stream
-) {
-    ggml_cuda_pool_alloc<float>  dst_tmp(pool);
-    ggml_cuda_pool_alloc<float2> dst_tmp_meta(pool);
-
-    if (parallel_blocks > 1) {
-        dst_tmp.alloc(parallel_blocks*ggml_nelements(KQV));
-        dst_tmp_meta.alloc(parallel_blocks*ggml_nrows(KQV));
+template <int cols_per_block, int parallel_blocks>
+void launch_fattn_vec_f32_64_128(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * Q = dst->src[0];
+    switch (Q->ne[0]) {
+        case  64: {
+            constexpr int      D = 64;
+            constexpr int nwarps = D/WARP_SIZE;
+            fattn_kernel_t fattn_kernel = flash_attn_vec_ext_f32<D, cols_per_block, parallel_blocks>;
+            launch_fattn<D, parallel_blocks>(ctx, dst, fattn_kernel, nwarps, cols_per_block);
+        } break;
+        case 128: {
+            constexpr int      D = 128;
+            constexpr int nwarps = D/WARP_SIZE;
+            fattn_kernel_t fattn_kernel = flash_attn_vec_ext_f32<D, cols_per_block, parallel_blocks>;
+            launch_fattn<D, parallel_blocks>(ctx, dst, fattn_kernel, nwarps, cols_per_block);
+        } break;
+        default: {
+            GGML_ASSERT(false && "FlashAttention without tensor cores only supports head sizes 64 and 128.");
+        } break;
     }
-
-    constexpr int  nwarps = (D + WARP_SIZE - 1) / WARP_SIZE;
-    const     dim3 block_dim(WARP_SIZE, nwarps, 1);
-    const     dim3 blocks_num(parallel_blocks*((Q->ne[1] + cols_per_block - 1) / cols_per_block), Q->ne[2], Q->ne[3]);
-    const     int  shmem = 0;
-
-    float scale    = 1.0f;
-    float max_bias = 0.0f;
-
-    memcpy(&scale,    (float *) KQV->op_params + 0, sizeof(float));
-    memcpy(&max_bias, (float *) KQV->op_params + 1, sizeof(float));
-
-    const uint32_t n_head      = Q->ne[2];
-    const uint32_t n_head_log2 = 1u << (uint32_t) floorf(log2f((float) n_head));
-
-    const float m0 = powf(2.0f, -(max_bias       ) / n_head_log2);
-    const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2);
-
-    flash_attn_vec_ext_f32<D, cols_per_block, parallel_blocks>
-        <<<blocks_num, block_dim, shmem, main_stream>>> (
-                (const char *) Q->data,
-                (const char *) K->data,
-                (const char *) V->data,
-                mask ? ((const char *) mask->data) : nullptr,
-                parallel_blocks == 1 ? (float *) KQV->data : dst_tmp.ptr, dst_tmp_meta.ptr,
-                scale, max_bias, m0, m1, n_head_log2,
-                Q->ne[0], Q->ne[1], Q->ne[2], Q->ne[3],
-                K->ne[0], K->ne[1], K->ne[2], K->ne[3],
-                mask ? mask->ne[1] : 0, mask ?  mask->nb[1] : 0,
-                Q->nb[1], Q->nb[2], Q->nb[3],
-                K->nb[1], K->nb[2], K->nb[3],
-                KQV->ne[0], KQV->ne[1], KQV->ne[2], KQV->ne[3]
-                );
-    CUDA_CHECK(cudaGetLastError());
-
-    if (parallel_blocks == 1) {
-        return;
-    }
-
-    const dim3 block_dim_combine(D, 1, 1);
-    const dim3 blocks_num_combine(Q->ne[1], blocks_num.y, blocks_num.z);
-    const int  shmem_combine = 0;
-
-    flash_attn_combine_results<D, parallel_blocks>
-        <<<blocks_num_combine, block_dim_combine, shmem_combine, main_stream>>>
-        (dst_tmp.ptr, dst_tmp_meta.ptr, (float *) KQV->data);
-    CUDA_CHECK(cudaGetLastError());
 }
 
 void ggml_cuda_flash_attn_ext_vec_f32(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     const ggml_tensor * Q = dst->src[0];
-    const ggml_tensor * K = dst->src[1];
-    const ggml_tensor * V = dst->src[2];
-
-    const ggml_tensor * mask = dst->src[3];
-
-    ggml_tensor * KQV = dst;
-
-    GGML_ASSERT(Q->ne[0] == 64 || Q->ne[0] == 128 && "FlashAttention without tensor cores only supports head sizes 64 and 128.");
 
     if (Q->ne[1] == 1) {
-        constexpr int cols_per_block = 1;
+        constexpr int cols_per_block  = 1;
         constexpr int parallel_blocks = 4;
-        switch (Q->ne[0]) {
-            case 64:
-                launch_fattn_vec_f32< 64, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
-                break;
-            case 128:
-                launch_fattn_vec_f32<128, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
-                break;
-            default:
-                GGML_ASSERT(false);
-                break;
-        }
+        launch_fattn_vec_f32_64_128<cols_per_block, parallel_blocks>(ctx, dst);
         return;
     }
 
     if (Q->ne[1] == 2) {
-        constexpr int cols_per_block = 2;
+        constexpr int cols_per_block  = 2;
         constexpr int parallel_blocks = 4;
-        switch (Q->ne[0]) {
-            case 64:
-                launch_fattn_vec_f32< 64, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
-                break;
-            case 128:
-                launch_fattn_vec_f32<128, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
-                break;
-            default:
-                GGML_ASSERT(false);
-                break;
-        }
+        launch_fattn_vec_f32_64_128<cols_per_block, parallel_blocks>(ctx, dst);
         return;
     }
 
     if (Q->ne[1] <= 4) {
-        constexpr int cols_per_block = 4;
+        constexpr int cols_per_block  = 4;
         constexpr int parallel_blocks = 4;
-        switch (Q->ne[0]) {
-            case 64:
-                launch_fattn_vec_f32< 64, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
-                break;
-            case 128:
-                launch_fattn_vec_f32<128, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
-                break;
-            default:
-                GGML_ASSERT(false);
-                break;
-        }
+        launch_fattn_vec_f32_64_128<cols_per_block, parallel_blocks>(ctx, dst);
         return;
     }
 
     if (Q->ne[1] <= 8) {
-        constexpr int cols_per_block = 8;
+        constexpr int cols_per_block  = 8;
         constexpr int parallel_blocks = 4;
-        switch (Q->ne[0]) {
-            case 64:
-                launch_fattn_vec_f32< 64, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
-                break;
-            case 128:
-                launch_fattn_vec_f32<128, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
-                break;
-            default:
-                GGML_ASSERT(false);
-                break;
-        }
+        launch_fattn_vec_f32_64_128<cols_per_block, parallel_blocks>(ctx, dst);
         return;
     }
 
-    constexpr int cols_per_block = 8;
+    constexpr int cols_per_block  = 8;
     constexpr int parallel_blocks = 1;
-    switch (Q->ne[0]) {
-        case 64:
-            launch_fattn_vec_f32< 64, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
-            break;
-        case 128:
-            launch_fattn_vec_f32<128, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
-            break;
-        default:
-            GGML_ASSERT(false);
-            break;
-    }
+    launch_fattn_vec_f32_64_128<cols_per_block, parallel_blocks>(ctx, dst);
 }

ggml-cuda/fattn.cu

@@ -1,5 +1,7 @@
 #include "common.cuh"
 #include "fattn-common.cuh"
+#include "fattn-tile-f16.cuh"
+#include "fattn-tile-f32.cuh"
 #include "fattn-vec-f16.cuh"
 #include "fattn-vec-f32.cuh"
 #include "fattn.cuh"
@@ -83,19 +85,9 @@ static __global__ void flash_attn_ext_f16(
     const int stride_Q  = nb01 / sizeof(float);
     const int stride_KV = nb11 / sizeof(half);
 
-    half  slopeh = __float2half(1.0f);
-    half2 slope2 = make_half2(1.0f, 1.0f);
-
-    // ALiBi
-    if (max_bias > 0.0f) {
-        const int h = blockIdx.y;
-
-        const float base = h < n_head_log2 ? m0 : m1;
-        const int   exph = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1;
-
-        slopeh = __float2half(powf(base, exph));
-        slope2 = make_half2(slopeh, slopeh);
-    }
+    const float slopef = get_alibi_slope(max_bias, blockIdx.y, n_head_log2, m0, m1);
+    const half  slopeh = __float2half(slopef);
+    const half2 slope2 = make_half2(slopef, slopef);
 
     frag_b Q_b[D/16][ncols/frag_n];
 
@@ -437,117 +429,64 @@ static_assert(get_VKQ_stride( 80, 1, 16) ==  16, "Test failed.");
 static_assert(get_VKQ_stride( 80, 2, 16) ==  16, "Test failed.");
 static_assert(get_VKQ_stride( 80, 4, 16) ==  16, "Test failed.");
 
-template <int D, int cols_per_block, int nwarps, int parallel_blocks, typename KQ_acc_t> void launch_fattn_f16_impl(
-        const ggml_tensor * Q, const ggml_tensor * K, const ggml_tensor * V, ggml_tensor * KQV, const ggml_tensor * mask,
-        ggml_cuda_pool & pool, cudaStream_t main_stream
-) {
-    ggml_cuda_pool_alloc<float>  dst_tmp(pool);
-    ggml_cuda_pool_alloc<float2> dst_tmp_meta(pool);
+template <int D, int cols_per_block, int nwarps, typename KQ_acc_t>
+void launch_fattn_f16(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * Q = dst->src[0];
 
-    if (parallel_blocks > 1) {
-        dst_tmp.alloc(parallel_blocks*ggml_nelements(KQV));
-        dst_tmp_meta.alloc(parallel_blocks*ggml_nrows(KQV));
-    }
-
-    constexpr int  frag_m = (cols_per_block) == 8 && (D) % 32 == 0 ? 32 : 16;
-    const     dim3 block_dim(WARP_SIZE, nwarps, 1);
-    const     dim3 blocks_num(parallel_blocks*(Q->ne[1] + cols_per_block - 1) / cols_per_block, Q->ne[2], Q->ne[3]);
-    const     int  shmem = 0;
-
-    float scale    = 1.0f;
-    float max_bias = 0.0f;
-
-    memcpy(&scale,    (float *) KQV->op_params + 0, sizeof(float));
-    memcpy(&max_bias, (float *) KQV->op_params + 1, sizeof(float));
-
-    const uint32_t n_head      = Q->ne[2];
-    const uint32_t n_head_log2 = 1u << (uint32_t) floorf(log2f((float) n_head));
-
-    const float m0 = powf(2.0f, -(max_bias       ) / n_head_log2);
-    const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2);
-
-    flash_attn_ext_f16<D, cols_per_block, nwarps, get_VKQ_stride(D, nwarps, frag_m), parallel_blocks, KQ_acc_t>
-        <<<blocks_num, block_dim, shmem, main_stream>>> (
-                (const char *) Q->data,
-                (const char *) K->data,
-                (const char *) V->data,
-                mask ? ((const char *) mask->data) : nullptr,
-                (parallel_blocks) == 1 ? (float *) KQV->data : dst_tmp.ptr, dst_tmp_meta.ptr,
-                scale, max_bias, m0, m1, n_head_log2,
-                Q->ne[0], Q->ne[1], Q->ne[2], Q->ne[3],
-                K->ne[0], K->ne[1], K->ne[2], K->ne[3],
-                mask ? mask->ne[1] : 0, mask ?  mask->nb[1] : 0,
-                Q->nb[1], Q->nb[2], Q->nb[3],
-                K->nb[1], K->nb[2], K->nb[3],
-                KQV->ne[0], KQV->ne[1], KQV->ne[2], KQV->ne[3]
-                );
-    CUDA_CHECK(cudaGetLastError());
-
-    if ((parallel_blocks) == 1) {
-        return;
-    }
-
-    const dim3 block_dim_combine(D, 1, 1);
-    const dim3 blocks_num_combine(Q->ne[1], blocks_num.y, blocks_num.z);
-    const int  shmem_combine = 0;
-
-    flash_attn_combine_results<D, parallel_blocks>
-        <<<blocks_num_combine, block_dim_combine, shmem_combine, main_stream>>>
-        (dst_tmp.ptr, dst_tmp_meta.ptr, (float *) KQV->data);
-    CUDA_CHECK(cudaGetLastError());
-}
-
-template <int D, int cols_per_block, int nwarps, typename KQ_acc_t> void launch_fattn_f16(
-        const ggml_tensor * Q, const ggml_tensor * K, const ggml_tensor * V, ggml_tensor * KQV, const ggml_tensor * mask,
-        const int nsm, ggml_cuda_pool & pool, cudaStream_t main_stream
-) {
+    constexpr int frag_m = cols_per_block == 8 && D % 32 == 0 ? 32 : 16;
     const int blocks_num_pb1 = ((Q->ne[1] + cols_per_block - 1) / cols_per_block)*Q->ne[2]*Q->ne[3];
+    const int nsm = ggml_cuda_info().devices[ggml_cuda_get_device()].nsm;
 
     if (4*blocks_num_pb1 < 2*nsm) {
-        launch_fattn_f16_impl<D, cols_per_block, nwarps, 4, KQ_acc_t>(Q, K, V, KQV, mask, pool, main_stream);
+        constexpr int parallel_blocks = 4;
+        fattn_kernel_t fattn_kernel = flash_attn_ext_f16<D, cols_per_block, nwarps, get_VKQ_stride(D, nwarps, frag_m), parallel_blocks, KQ_acc_t>;
+        launch_fattn<D, parallel_blocks>(ctx, dst, fattn_kernel, nwarps, cols_per_block);
         return;
     }
     if (2*blocks_num_pb1 < 2*nsm) {
-        launch_fattn_f16_impl<D, cols_per_block, nwarps, 2, KQ_acc_t>(Q, K, V, KQV, mask, pool, main_stream);
+        constexpr int parallel_blocks = 2;
+        fattn_kernel_t fattn_kernel = flash_attn_ext_f16<D, cols_per_block, nwarps, get_VKQ_stride(D, nwarps, frag_m), parallel_blocks, KQ_acc_t>;
+        launch_fattn<D, parallel_blocks>(ctx, dst, fattn_kernel, nwarps, cols_per_block);
         return;
     }
-    launch_fattn_f16_impl<D, cols_per_block, nwarps, 1, KQ_acc_t>(Q, K, V, KQV, mask, pool, main_stream);
+    constexpr int parallel_blocks = 1;
+    fattn_kernel_t fattn_kernel = flash_attn_ext_f16<D, cols_per_block, nwarps, get_VKQ_stride(D, nwarps, frag_m), parallel_blocks, KQ_acc_t>;
+    launch_fattn<D, parallel_blocks>(ctx, dst, fattn_kernel, nwarps, cols_per_block);
 }
 
 void ggml_cuda_flash_attn_ext(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
-    const ggml_tensor * Q = dst->src[0];
-    const ggml_tensor * K = dst->src[1];
-    const ggml_tensor * V = dst->src[2];
-
-    const ggml_tensor * mask = dst->src[3];
-
-    ggml_tensor * KQV = dst;
-
-    GGML_ASSERT(Q->type == GGML_TYPE_F32);
-    GGML_ASSERT(K->type == GGML_TYPE_F16);
-    GGML_ASSERT(V->type == GGML_TYPE_F16);
-    GGML_ASSERT(KQV->type == GGML_TYPE_F32);
-
-    GGML_ASSERT(!mask || mask->type == GGML_TYPE_F16);
-    GGML_ASSERT(!mask || mask->ne[1] >= GGML_PAD(Q->ne[1], 16) &&
-                                "the Flash-Attention CUDA kernel requires the mask to be padded to 16 and at least n_queries big");
-
-    GGML_ASSERT(K->ne[1] % FATTN_KQ_STRIDE == 0 && "Incorrect KV cache padding.");
+    const ggml_tensor * KQV = dst;
+    const ggml_tensor * Q   = dst->src[0];
 
     ggml_cuda_set_device(ctx.device);
-
-    const int cc  = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
-    const int nsm = ggml_cuda_info().devices[ggml_cuda_get_device()].nsm;
-
+    const int cc = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
     const int32_t precision = KQV->op_params[2];
 
+    // On AMD the tile kernels perform poorly, use the vec kernel instead:
+    if (cc >= CC_OFFSET_AMD) {
+        if (precision == GGML_PREC_DEFAULT) {
+            ggml_cuda_flash_attn_ext_vec_f16_no_mma(ctx, dst);
+        } else {
+            ggml_cuda_flash_attn_ext_vec_f32(ctx, dst);
+        }
+        return;
+    }
+
     if (!fast_fp16_available(cc)) {
-        ggml_cuda_flash_attn_ext_vec_f32(ctx, dst);
+        if (Q->ne[1] <= 8) {
+            ggml_cuda_flash_attn_ext_vec_f32(ctx, dst);
+        } else {
+            ggml_cuda_flash_attn_ext_tile_f32(ctx, dst);
+        }
         return;
     }
 
     if (!fp16_mma_available(cc)) {
-        ggml_cuda_flash_attn_ext_vec_f16_no_mma(ctx, dst);
+        if (Q->ne[1] <= 8) {
+            ggml_cuda_flash_attn_ext_vec_f16_no_mma(ctx, dst);
+        } else {
+            ggml_cuda_flash_attn_ext_tile_f16(ctx, dst);
+        }
         return;
     }
 
@@ -562,22 +501,22 @@ void ggml_cuda_flash_attn_ext(ggml_backend_cuda_context & ctx, ggml_tensor * dst
             constexpr int nwarps         =  4;
             switch (Q->ne[0]) {
                 case 64:
-                    launch_fattn_f16< 64, cols_per_block, nwarps, float>(Q, K, V, KQV, mask, nsm, ctx.pool(), ctx.stream());
+                    launch_fattn_f16< 64, cols_per_block, nwarps, float>(ctx, dst);
                     break;
                 case 80:
-                    launch_fattn_f16< 80, cols_per_block, nwarps, float>(Q, K, V, KQV, mask, nsm, ctx.pool(), ctx.stream());
+                    launch_fattn_f16< 80, cols_per_block, nwarps, float>(ctx, dst);
                     break;
                 case 96:
-                    launch_fattn_f16< 96, cols_per_block, nwarps, float>(Q, K, V, KQV, mask, nsm, ctx.pool(), ctx.stream());
+                    launch_fattn_f16< 96, cols_per_block, nwarps, float>(ctx, dst);
                     break;
                 case 112:
-                    launch_fattn_f16<112, cols_per_block, nwarps, float>(Q, K, V, KQV, mask, nsm, ctx.pool(), ctx.stream());
+                    launch_fattn_f16<112, cols_per_block, nwarps, float>(ctx, dst);
                     break;
                 case 128:
-                    launch_fattn_f16<128, cols_per_block, nwarps, float>(Q, K, V, KQV, mask, nsm, ctx.pool(), ctx.stream());
+                    launch_fattn_f16<128, cols_per_block, nwarps, float>(ctx, dst);
                     break;
                 case 256:
-                    launch_fattn_f16<256, cols_per_block, nwarps, float>(Q, K, V, KQV, mask, nsm, ctx.pool(), ctx.stream());
+                    launch_fattn_f16<256, cols_per_block, nwarps, float>(ctx, dst);
                     break;
                 default:
                     GGML_ASSERT(false);
@@ -588,22 +527,22 @@ void ggml_cuda_flash_attn_ext(ggml_backend_cuda_context & ctx, ggml_tensor * dst
             constexpr int nwarps         =  4;
             switch (Q->ne[0]) {
                 case 64:
-                    launch_fattn_f16< 64, cols_per_block, nwarps, float>(Q, K, V, KQV, mask, nsm, ctx.pool(), ctx.stream());
+                    launch_fattn_f16< 64, cols_per_block, nwarps, float>(ctx, dst);
                     break;
                 case 80:
-                    launch_fattn_f16< 80, cols_per_block, nwarps, float>(Q, K, V, KQV, mask, nsm, ctx.pool(), ctx.stream());
+                    launch_fattn_f16< 80, cols_per_block, nwarps, float>(ctx, dst);
                     break;
                 case 96:
-                    launch_fattn_f16< 96, cols_per_block, nwarps, float>(Q, K, V, KQV, mask, nsm, ctx.pool(), ctx.stream());
+                    launch_fattn_f16< 96, cols_per_block, nwarps, float>(ctx, dst);
                     break;
                 case 112:
-                    launch_fattn_f16<112, cols_per_block, nwarps, float>(Q, K, V, KQV, mask, nsm, ctx.pool(), ctx.stream());
+                    launch_fattn_f16<112, cols_per_block, nwarps, float>(ctx, dst);
                     break;
                 case 128:
-                    launch_fattn_f16<128, cols_per_block, nwarps, float>(Q, K, V, KQV, mask, nsm, ctx.pool(), ctx.stream());
+                    launch_fattn_f16<128, cols_per_block, nwarps, float>(ctx, dst);
                     break;
                 // case 256:
-                //     launch_fattn_f16<256, cols_per_block, nwarps, float>(Q, K, V, KQV, mask, nsm, ctx.pool(), ctx.stream());
+                //     launch_fattn_f16<256, cols_per_block, nwarps, float>(ctx, dst);
                 //     break;
                 default:
                     GGML_ASSERT(false);
@@ -623,16 +562,16 @@ void ggml_cuda_flash_attn_ext(ggml_backend_cuda_context & ctx, ggml_tensor * dst
         constexpr int nwarps         = 4;
         switch (Q->ne[0]) {
             case 64:
-                launch_fattn_f16< 64, cols_per_block, nwarps, half>(Q, K, V, KQV, mask, nsm, ctx.pool(), ctx.stream());
+                launch_fattn_f16< 64, cols_per_block, nwarps, half>(ctx, dst);
                 break;
             case 96:
-                launch_fattn_f16< 96, cols_per_block, nwarps, half>(Q, K, V, KQV, mask, nsm, ctx.pool(), ctx.stream());
+                launch_fattn_f16< 96, cols_per_block, nwarps, half>(ctx, dst);
                 break;
             case 128:
-                launch_fattn_f16<128, cols_per_block, nwarps, half>(Q, K, V, KQV, mask, nsm, ctx.pool(), ctx.stream());
+                launch_fattn_f16<128, cols_per_block, nwarps, half>(ctx, dst);
                 break;
             case 256:
-                launch_fattn_f16<256, cols_per_block, nwarps, half>(Q, K, V, KQV, mask, nsm, ctx.pool(), ctx.stream());
+                launch_fattn_f16<256, cols_per_block, nwarps, half>(ctx, dst);
                 break;
             default:
                 GGML_ASSERT(false);
@@ -646,22 +585,22 @@ void ggml_cuda_flash_attn_ext(ggml_backend_cuda_context & ctx, ggml_tensor * dst
         constexpr int nwarps         =  4;
         switch (Q->ne[0]) {
             case 64:
-                launch_fattn_f16< 64, cols_per_block, nwarps, half>(Q, K, V, KQV, mask, nsm, ctx.pool(), ctx.stream());
+                launch_fattn_f16< 64, cols_per_block, nwarps, half>(ctx, dst);
                 break;
             case 80:
-                launch_fattn_f16< 80, cols_per_block, nwarps, half>(Q, K, V, KQV, mask, nsm, ctx.pool(), ctx.stream());
+                launch_fattn_f16< 80, cols_per_block, nwarps, half>(ctx, dst);
                 break;
             case 96:
-                launch_fattn_f16< 96, cols_per_block, nwarps, half>(Q, K, V, KQV, mask, nsm, ctx.pool(), ctx.stream());
+                launch_fattn_f16< 96, cols_per_block, nwarps, half>(ctx, dst);
                 break;
             case 112:
-                launch_fattn_f16<112, cols_per_block, nwarps, half>(Q, K, V, KQV, mask, nsm, ctx.pool(), ctx.stream());
+                launch_fattn_f16<112, cols_per_block, nwarps, half>(ctx, dst);
                 break;
             case 128:
-                launch_fattn_f16<128, cols_per_block, nwarps, half>(Q, K, V, KQV, mask, nsm, ctx.pool(), ctx.stream());
+                launch_fattn_f16<128, cols_per_block, nwarps, half>(ctx, dst);
                 break;
             case 256:
-                launch_fattn_f16<256, cols_per_block, nwarps, half>(Q, K, V, KQV, mask, nsm, ctx.pool(), ctx.stream());
+                launch_fattn_f16<256, cols_per_block, nwarps, half>(ctx, dst);
                 break;
             default:
                 GGML_ASSERT(false);
@@ -674,22 +613,22 @@ void ggml_cuda_flash_attn_ext(ggml_backend_cuda_context & ctx, ggml_tensor * dst
     constexpr int nwarps         =  4;
     switch (Q->ne[0]) {
         case 64:
-            launch_fattn_f16< 64, cols_per_block, nwarps, half>(Q, K, V, KQV, mask, nsm, ctx.pool(), ctx.stream());
+            launch_fattn_f16< 64, cols_per_block, nwarps, half>(ctx, dst);
             break;
         case 80:
-            launch_fattn_f16< 80, cols_per_block, nwarps, half>(Q, K, V, KQV, mask, nsm, ctx.pool(), ctx.stream());
+            launch_fattn_f16< 80, cols_per_block, nwarps, half>(ctx, dst);
             break;
         case 96:
-            launch_fattn_f16< 96, cols_per_block, nwarps, half>(Q, K, V, KQV, mask, nsm, ctx.pool(), ctx.stream());
+            launch_fattn_f16< 96, cols_per_block, nwarps, half>(ctx, dst);
             break;
         case 112:
-            launch_fattn_f16<112, cols_per_block, nwarps, half>(Q, K, V, KQV, mask, nsm, ctx.pool(), ctx.stream());
+            launch_fattn_f16<112, cols_per_block, nwarps, half>(ctx, dst);
             break;
         case 128:
-            launch_fattn_f16<128, cols_per_block, nwarps, half>(Q, K, V, KQV, mask, nsm, ctx.pool(), ctx.stream());
+            launch_fattn_f16<128, cols_per_block, nwarps, half>(ctx, dst);
             break;
         case 256:
-            launch_fattn_f16<256, cols_per_block, nwarps, half>(Q, K, V, KQV, mask, nsm, ctx.pool(), ctx.stream());
+            launch_fattn_f16<256, cols_per_block, nwarps, half>(ctx, dst);
             break;
         default:
             GGML_ASSERT(false);

ggml-cuda/softmax.cu

@@ -1,3 +1,4 @@
+#include "common.cuh"
 #include "softmax.cuh"
 
 template <typename T>
@@ -23,17 +24,7 @@ static __global__ void soft_max_f32(const float * x, const T * mask, float * dst
     const int warp_id = threadIdx.x / WARP_SIZE;
     const int lane_id = threadIdx.x % WARP_SIZE;
 
-    float slope = 1.0f;
-
-    // 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   exph = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1;
-
-        slope = powf(base, exph);
-    }
+    const float slope = get_alibi_slope(max_bias, rowx/nrows_y, n_head_log2, m0, m1);
 
     extern __shared__ float data_soft_max_f32[];
     float * buf_iw = data_soft_max_f32; // shared memory buffer for inter-warp communication

ggml-mpi.c

@@ -1,216 +0,0 @@
-#include "ggml-mpi.h"
-
-#include "ggml.h"
-
-#include <mpi.h>
-
-#include <stdio.h>
-#include <stdlib.h>
-
-#define MIN(a, b) ((a) < (b) ? (a) : (b))
-
-#define UNUSED GGML_UNUSED
-
-struct ggml_mpi_context {
-    int rank;
-    int size;
-};
-
-void ggml_mpi_backend_init(void) {
-    MPI_Init(NULL, NULL);
-}
-
-void ggml_mpi_backend_free(void) {
-    MPI_Finalize();
-}
-
-struct ggml_mpi_context * ggml_mpi_init(void) {
-    struct ggml_mpi_context * ctx = calloc(1, sizeof(struct ggml_mpi_context));
-
-    MPI_Comm_rank(MPI_COMM_WORLD, &ctx->rank);
-    MPI_Comm_size(MPI_COMM_WORLD, &ctx->size);
-
-    return ctx;
-}
-
-void ggml_mpi_free(struct ggml_mpi_context * ctx) {
-    free(ctx);
-}
-
-int ggml_mpi_rank(struct ggml_mpi_context * ctx) {
-    return ctx->rank;
-}
-
-void ggml_mpi_eval_init(
-        struct ggml_mpi_context * ctx_mpi,
-                            int * n_tokens,
-                            int * n_past,
-                            int * n_threads) {
-    UNUSED(ctx_mpi);
-
-    // synchronize the worker node parameters with the root node
-    MPI_Barrier(MPI_COMM_WORLD);
-
-    MPI_Bcast(n_tokens,  1, MPI_INT, 0, MPI_COMM_WORLD);
-    MPI_Bcast(n_past,    1, MPI_INT, 0, MPI_COMM_WORLD);
-    MPI_Bcast(n_threads, 1, MPI_INT, 0, MPI_COMM_WORLD);
-}
-
-static int ggml_graph_get_node_idx(struct ggml_cgraph * gf, const char * name) {
-    struct ggml_tensor * t = ggml_graph_get_tensor(gf, name);
-    if (t == NULL) {
-        fprintf(stderr, "%s: tensor %s not found\n", __func__, name);
-        return -1;
-    }
-
-    for (int i = 0; i < gf->n_nodes; i++) {
-        if (gf->nodes[i] == t) {
-            return i;
-        }
-    }
-
-    fprintf(stderr, "%s: tensor %s not found in graph (should not happen)\n", __func__, name);
-    return -1;
-}
-
-static void ggml_mpi_tensor_send(struct ggml_tensor * t, int mpi_rank_dst) {
-    MPI_Datatype mpi_type;
-
-    switch (t->type) {
-        case GGML_TYPE_I32: mpi_type = MPI_INT32_T; break;
-        case GGML_TYPE_F32: mpi_type = MPI_FLOAT;   break;
-        default: GGML_ASSERT(false && "not implemented");
-    }
-
-    const int retval = MPI_Send(t->data, ggml_nelements(t), mpi_type, mpi_rank_dst, 0, MPI_COMM_WORLD);
-    GGML_ASSERT(retval == MPI_SUCCESS);
-}
-
-static void ggml_mpi_tensor_recv(struct ggml_tensor * t, int mpi_rank_src) {
-    MPI_Datatype mpi_type;
-
-    switch (t->type) {
-        case GGML_TYPE_I32: mpi_type = MPI_INT32_T; break;
-        case GGML_TYPE_F32: mpi_type = MPI_FLOAT;   break;
-        default: GGML_ASSERT(false && "not implemented");
-    }
-
-    MPI_Status status; UNUSED(status);
-
-    const int retval = MPI_Recv(t->data, ggml_nelements(t), mpi_type, mpi_rank_src, MPI_ANY_TAG, MPI_COMM_WORLD, &status);
-    GGML_ASSERT(retval == MPI_SUCCESS);
-}
-
-// TODO: there are many improvements that can be done to this implementation
-void ggml_mpi_graph_compute_pre(
-        struct ggml_mpi_context * ctx_mpi,
-             struct ggml_cgraph * gf,
-                            int   n_layers) {
-    const int mpi_rank = ctx_mpi->rank;
-    const int mpi_size = ctx_mpi->size;
-
-    struct ggml_tensor * inp_tokens = ggml_graph_get_tensor(gf, "inp_tokens");
-    if (inp_tokens == NULL) {
-        fprintf(stderr, "%s: tensor 'inp_tokens' not found\n", __func__);
-        return;
-    }
-
-    struct ggml_tensor * inp0 = ggml_graph_get_tensor(gf, "layer_inp_0");
-    if (inp0 == NULL) {
-        fprintf(stderr, "%s: tensor 'inp0' not found\n", __func__);
-        return;
-    }
-
-    GGML_ASSERT(inp0 == gf->nodes[0]);
-
-    // distribute the compute graph into slices across the MPI nodes
-    //
-    // the main node (0) processes the last layers + the remainder of the compute graph
-    // and is responsible to pass the input tokens to the first node (1)
-    //
-    // node 1:   [(  0) * n_per_node, (  1) * n_per_node)
-    // node 2:   [(  1) * n_per_node, (  2) * n_per_node)
-    // ...
-    // node n-1: [(n-2) * n_per_node, (n-1) * n_per_node)
-    // node 0:   [(n-1) * n_per_node,            n_nodes)
-    //
-    if (mpi_rank > 0) {
-        if (mpi_rank == 1) {
-            // the first node (1) receives the input tokens from the main node (0)
-            ggml_mpi_tensor_recv(inp_tokens, 0);
-        } else {
-            // recv input data for each node into the "inp0" tensor (i.e. the first node in the compute graph)
-            ggml_mpi_tensor_recv(inp0, mpi_rank - 1);
-        }
-    } else if (mpi_size > 1) {
-        // node 0 sends the input tokens to node 1
-        ggml_mpi_tensor_send(inp_tokens, 1);
-
-        // recv the output data from the last node
-        ggml_mpi_tensor_recv(inp0, mpi_size - 1);
-    }
-
-    {
-        const int n_per_node = (n_layers + (mpi_size - 1)) / mpi_size;
-
-        const int mpi_idx = mpi_rank > 0 ? mpi_rank - 1 : mpi_size - 1;
-
-        const int il0 =               (mpi_idx + 0) * n_per_node;
-        const int il1 = MIN(n_layers, (mpi_idx + 1) * n_per_node);
-
-        char name_l0[GGML_MAX_NAME];
-        char name_l1[GGML_MAX_NAME];
-
-        snprintf(name_l0, sizeof(name_l0), "layer_inp_%d", il0);
-        snprintf(name_l1, sizeof(name_l1), "layer_inp_%d", il1);
-
-        const int idx_l0 =                ggml_graph_get_node_idx(gf, name_l0);
-        const int idx_l1 = mpi_rank > 0 ? ggml_graph_get_node_idx(gf, name_l1) + 1 : gf->n_nodes;
-
-        if (idx_l0 < 0 || idx_l1 < 0) {
-            fprintf(stderr, "%s: layer input nodes not found\n", __func__);
-            return;
-        }
-
-        // attach the input data to all nodes that need it
-        // TODO: not great - should be able to do this without modifying the compute graph (see next TODO below)
-        for (int i = idx_l0; i < idx_l1; i++) {
-            if (gf->nodes[i]->src[0] == gf->nodes[idx_l0]) {
-                gf->nodes[i]->src[0] =  inp0;
-            }
-            if (gf->nodes[i]->src[1] == gf->nodes[idx_l0]) {
-                gf->nodes[i]->src[1] =  inp0;
-            }
-        }
-
-        // TODO: instead of rearranging the nodes, we should be able to execute a subset of the compute graph
-        for (int i = 1; i < idx_l1 - idx_l0; i++) {
-            gf->nodes[i] = gf->nodes[idx_l0 + i];
-            gf->grads[i] = gf->grads[idx_l0 + i];
-        }
-
-        // the first node performs the "get_rows" operation, the rest of the nodes get the data from the previous node
-        if (mpi_idx != 0) {
-            gf->nodes[0]->op = GGML_OP_NONE;
-        }
-
-        gf->n_nodes = idx_l1 - idx_l0;
-
-        //fprintf(stderr, "%s: node %d: processing %d nodes [%d, %d)\n", __func__, mpi_rank, gf->n_nodes, il0, il1);
-    }
-}
-
-void ggml_mpi_graph_compute_post(
-        struct ggml_mpi_context * ctx_mpi,
-             struct ggml_cgraph * gf,
-                            int   n_layers) {
-    UNUSED(n_layers);
-
-    const int mpi_rank = ctx_mpi->rank;
-    const int mpi_size = ctx_mpi->size;
-
-    // send the output data to the next node
-    if (mpi_rank > 0) {
-        ggml_mpi_tensor_send(gf->nodes[gf->n_nodes - 1], (mpi_rank + 1) % mpi_size);
-    }
-}

ggml-mpi.h

@@ -1,39 +0,0 @@
-#pragma once
-
-struct ggml_context;
-struct ggml_tensor;
-struct ggml_cgraph;
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-struct ggml_mpi_context;
-
-void ggml_mpi_backend_init(void);
-void ggml_mpi_backend_free(void);
-
-struct ggml_mpi_context * ggml_mpi_init(void);
-void ggml_mpi_free(struct ggml_mpi_context * ctx);
-
-int ggml_mpi_rank(struct ggml_mpi_context * ctx);
-
-void ggml_mpi_eval_init(
-        struct ggml_mpi_context * ctx_mpi,
-                            int * n_tokens,
-                            int * n_past,
-                            int * n_threads);
-
-void ggml_mpi_graph_compute_pre(
-        struct ggml_mpi_context * ctx_mpi,
-             struct ggml_cgraph * gf,
-                            int   n_layers);
-
-void ggml_mpi_graph_compute_post(
-        struct ggml_mpi_context * ctx_mpi,
-             struct ggml_cgraph * gf,
-                            int   n_layers);
-
-#ifdef __cplusplus
-}
-#endif

ggml-quants.c

@@ -14,6 +14,12 @@
 #include <stdlib.h> // for qsort
 #include <stdio.h>  // for GGML_ASSERT
 
+#define GROUP_MAX_EPS 1e-15f
+#define GROUP_MAX_EPS_IQ3_XXS 1e-8f
+#define GROUP_MAX_EPS_IQ2_S 1e-8f
+#define GROUP_MAX_EPS_IQ1_M 1e-7f
+#define GROUP_MAX_EPS_IQ1_S 1e-12f
+
 #if defined(_MSC_VER)
 // disable "possible loss of data" to avoid warnings for hundreds of casts
 // we should just be careful :)
@@ -1109,7 +1115,7 @@ static float make_qx_quants(int n, int nmax, const float * restrict x, int8_t *
         float ax = fabsf(x[i]);
         if (ax > amax) { amax = ax; max = x[i]; }
     }
-    if (amax < 1e-30f) { // all zero
+    if (amax < GROUP_MAX_EPS) { // all zero
         for (int i = 0; i < n; ++i) {
             L[i] = 0;
         }
@@ -1143,7 +1149,7 @@ static float make_qx_quants(int n, int nmax, const float * restrict x, int8_t *
         sumlx += w*x[i]*l;
         suml2 += w*l*l;
     }
-    float scale = sumlx/suml2;
+    float scale = suml2 ? sumlx/suml2 : 0.0f;
     if (return_early) return suml2 > 0 ? 0.5f*(scale + 1/iscale) : 1/iscale;
     float best = scale * sumlx;
     for (int is = -9; is <= 9; ++is) {
@@ -1177,7 +1183,7 @@ static float make_q3_quants(int n, int nmax, const float * restrict x, int8_t *
         float ax = fabsf(x[i]);
         if (ax > amax) { amax = ax; max = x[i]; }
     }
-    if (!amax) { // all zero
+    if (amax < GROUP_MAX_EPS) { // all zero
         for (int i = 0; i < n; ++i) { L[i] = 0; }
         return 0.f;
     }
@@ -1646,7 +1652,7 @@ static float make_qp_quants(int n, int nmax, const float * restrict x, uint8_t *
             break;
         }
     }
-    return sumlx / suml2;
+    return sumlx/suml2;
 }
 
 static void quantize_row_q2_K_impl(const float * restrict x, block_q2_K * restrict y, int k, const float * restrict quant_weights) {
@@ -2653,7 +2659,7 @@ void quantize_row_q6_K_reference(const float * restrict x, block_q6_K * restrict
 
         }
 
-        if (!max_abs_scale) {
+        if (max_abs_scale < GROUP_MAX_EPS) {
             memset(&y[i], 0, sizeof(block_q6_K));
             y[i].d = GGML_FP32_TO_FP16(0.f);
             x += QK_K;
@@ -2805,7 +2811,7 @@ static void quantize_row_q6_K_impl(const float * restrict x, block_q6_K * restri
 
         }
 
-        if (!max_abs_scale) {
+        if (max_abs_scale < GROUP_MAX_EPS) {
             memset(&y[i], 0, sizeof(block_q6_K));
             y[i].d = GGML_FP32_TO_FP16(0.f);
             x += QK_K;
@@ -12599,7 +12605,7 @@ static void quantize_row_iq2_xxs_impl(const float * restrict x, void * restrict
             }
             float max = xval[0];
             for (int i = 1; i < 32; ++i) max = MAX(max, xval[i]);
-            if (!max) {
+            if (max < GROUP_MAX_EPS) {
                 scales[ib] = 0;
                 memset(L, 0, 32);
                 continue;
@@ -12775,7 +12781,7 @@ static void quantize_row_iq2_xs_impl(const float * restrict x, void * restrict v
             }
             float max = xval[0];
             for (int i = 1; i < 16; ++i) max = MAX(max, xval[i]);
-            if (!max) {
+            if (max < GROUP_MAX_EPS) {
                 scales[ib] = 0;
                 memset(L, 0, 16);
                 continue;
@@ -13216,7 +13222,7 @@ static void quantize_row_iq3_xxs_impl(int grid_size, const float * restrict x, v
             }
             float max = xval[0];
             for (int i = 1; i < 32; ++i) max = MAX(max, xval[i]);
-            if (!max) {
+            if (max < GROUP_MAX_EPS_IQ3_XXS) {
                 scales[ib] = 0;
                 memset(L, 0, 32);
                 continue;
@@ -13756,7 +13762,7 @@ static void quantize_row_iq1_s_impl(const float * restrict x, void * restrict vy
             for (int i = 0; i < block_size; ++i) weight[i] = qw[i] * sqrtf(sigma2 + xb[i]*xb[i]);
             float max = fabsf(xb[0]);
             for (int i = 1; i < block_size; ++i) max = MAX(max, fabsf(xb[i]));
-            if (!max) {
+            if (max < GROUP_MAX_EPS_IQ1_S) {
                 scales[ib] = 0;
                 memset(L, 1, block_size);
                 continue;
@@ -13944,7 +13950,7 @@ static void quantize_row_iq1_m_impl(const float * restrict x, void * restrict vy
             }
             float max = fabsf(xb[0]);
             for (int i = 1; i < block_size; ++i) max = MAX(max, fabsf(xb[i]));
-            if (!max) {
+            if (max < GROUP_MAX_EPS_IQ1_M) {
                 scales[ib] = 0;
                 memset(L, 1, block_size);
                 continue;
@@ -14208,7 +14214,7 @@ static void quantize_row_iq4_nl_impl(const int super_block_size, const int block
                 amax = ax; max = xb[j];
             }
         }
-        if (!amax) {
+        if (amax < GROUP_MAX_EPS) {
             scales[ib] = 0;
             continue;
         }
@@ -14429,7 +14435,7 @@ static void quantize_row_iq2_s_impl(const float * restrict x, void * restrict vy
             }
             float max = xval[0];
             for (int i = 1; i < 16; ++i) max = MAX(max, xval[i]);
-            if (!max) {
+            if (max < GROUP_MAX_EPS_IQ2_S) {
                 scales[ib] = 0;
                 continue;
             }

ggml-rpc.cpp

@@ -134,7 +134,13 @@ static bool set_no_delay(sockfd_t sockfd) {
     int flag = 1;
     // set TCP_NODELAY to disable Nagle's algorithm
     int ret = setsockopt(sockfd, IPPROTO_TCP, TCP_NODELAY, (char *)&flag, sizeof(int));
-    return ret >= 0;
+    return ret == 0;
+}
+
+static bool set_reuse_addr(sockfd_t sockfd) {
+    int flag = 1;
+    int ret = setsockopt(sockfd, SOL_SOCKET, SO_REUSEADDR, (char *)&flag, sizeof(int));
+    return ret == 0;
 }
 
 static std::shared_ptr<socket_t> socket_connect(const char * host, int port) {
@@ -181,7 +187,10 @@ static std::shared_ptr<socket_t> create_server_socket(const char * host, int por
     if (sock == nullptr) {
         return nullptr;
     }
-
+    if (!set_reuse_addr(sockfd)) {
+        fprintf(stderr, "Failed to set SO_REUSEADDR\n");
+        return nullptr;
+    }
     struct sockaddr_in serv_addr;
     serv_addr.sin_family = AF_INET;
     serv_addr.sin_addr.s_addr = inet_addr(host);

ggml-vulkan.cpp

@@ -114,6 +114,7 @@ struct vk_device {
     size_t idx;
 
     vk_matmul_pipeline pipeline_matmul_f32;
+    vk_matmul_pipeline pipeline_matmul_f32_f16;
     vk_matmul_pipeline pipeline_matmul_f16;
     vk_matmul_pipeline pipeline_matmul_f16_f32;
     vk_pipeline pipeline_matmul_split_k_reduce;
@@ -294,7 +295,6 @@ struct vk_op_rope_neox_push_constants {
 struct vk_op_soft_max_push_constants {
     uint32_t KX;
     uint32_t KY;
-    uint32_t KZ;
     float scale;
     float max_bias;
     float m0;
@@ -304,7 +304,8 @@ struct vk_op_soft_max_push_constants {
 
 struct vk_op_argsort_push_constants {
     uint32_t ncols;
-    bool ascending;
+    uint32_t ncols_pad;
+    int32_t order;
 };
 
 // Allow pre-recording command buffers
@@ -375,13 +376,12 @@ struct ggml_backend_vk_context {
     vk_context * compute_ctx;
     vk_context * transfer_ctx;
 
-    bool disable;
     bool initialized;
 
     size_t idx;
 };
 
-struct vk_instance {
+struct vk_instance_t {
     vk::Instance instance;
 
     std::vector<size_t> device_indices;
@@ -423,7 +423,7 @@ static void ggml_vk_check_results_1(ggml_backend_vk_context * ctx, ggml_compute_
 typedef void (*ggml_vk_func_t)(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst);
 
 static bool vk_instance_initialized = false;
-static vk_instance vk_instance;
+static vk_instance_t vk_instance;
 
 GGML_CALL static void ggml_backend_vk_free(ggml_backend_t backend);
 
@@ -1013,6 +1013,7 @@ static void ggml_vk_load_shaders(ggml_backend_vk_context * ctx) {
     uint32_t s_align =  32;
 
     ctx->device->pipeline_matmul_f32 = std::make_shared<vk_matmul_pipeline_struct>();
+    ctx->device->pipeline_matmul_f32_f16 = std::make_shared<vk_matmul_pipeline_struct>();
     ctx->device->pipeline_matmul_f16_f32 = std::make_shared<vk_matmul_pipeline_struct>();
     ctx->device->pipeline_matmul_f16 = std::make_shared<vk_matmul_pipeline_struct>();
     ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0] = std::make_shared<vk_matmul_pipeline_struct>();
@@ -1048,6 +1049,13 @@ static void ggml_vk_load_shaders(ggml_backend_vk_context * ctx) {
         ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f32->a_m, "matmul_f32_aligned_m", matmul_f32_aligned_len, matmul_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, m_align);
         ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f32->a_s, "matmul_f32_aligned_s", matmul_f32_aligned_len, matmul_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, s_align);
 
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f32_f16->l, "matmul_f32_f16_l", matmul_f32_f16_len, matmul_f32_f16_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, 1);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f32_f16->m, "matmul_f32_f16_m", matmul_f32_f16_len, matmul_f32_f16_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, 1);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f32_f16->s, "matmul_f32_f16_s", matmul_f32_f16_len, matmul_f32_f16_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, 1);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f32_f16->a_l, "matmul_f32_f16_aligned_l", matmul_f32_f16_aligned_len, matmul_f32_f16_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, l_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f32_f16->a_m, "matmul_f32_f16_aligned_m", matmul_f32_f16_aligned_len, matmul_f32_f16_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, m_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f32_f16->a_s, "matmul_f32_f16_aligned_s", matmul_f32_f16_aligned_len, matmul_f32_f16_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, s_align);
+
         ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f16->l, "matmul_f16_l", matmul_f16_len, matmul_f16_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, 1);
         ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f16->m, "matmul_f16_m", matmul_f16_len, matmul_f16_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, 1);
         ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f16->s, "matmul_f16_s", matmul_f16_len, matmul_f16_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, 1);
@@ -1230,6 +1238,13 @@ static void ggml_vk_load_shaders(ggml_backend_vk_context * ctx) {
         ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f32->a_m, "matmul_f32_aligned_m", matmul_f32_aligned_fp32_len, matmul_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, m_align);
         ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f32->a_s, "matmul_f32_aligned_s", matmul_f32_aligned_fp32_len, matmul_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, s_align);
 
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f32_f16->l, "matmul_f32_f16_l", matmul_f32_f16_fp32_len, matmul_f32_f16_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, 1);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f32_f16->m, "matmul_f32_f16_m", matmul_f32_f16_fp32_len, matmul_f32_f16_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, 1);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f32_f16->s, "matmul_f32_f16_s", matmul_f32_f16_fp32_len, matmul_f32_f16_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, 1);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f32_f16->a_l, "matmul_f32_f16_aligned_l", matmul_f32_f16_aligned_fp32_len, matmul_f32_f16_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, l_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f32_f16->a_m, "matmul_f32_f16_aligned_m", matmul_f32_f16_aligned_fp32_len, matmul_f32_f16_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, m_align);
+        ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f32_f16->a_s, "matmul_f32_f16_aligned_s", matmul_f32_f16_aligned_fp32_len, matmul_f32_f16_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, s_align);
+
         ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f16->l, "matmul_f16_l", matmul_f16_fp32_len, matmul_f16_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, 1);
         ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f16->m, "matmul_f16_m", matmul_f16_fp32_len, matmul_f16_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, 1);
         ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f16->s, "matmul_f16_s", matmul_f16_fp32_len, matmul_f16_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, 1);
@@ -1501,8 +1516,8 @@ static void ggml_vk_load_shaders(ggml_backend_vk_context * ctx) {
 
     ggml_vk_create_pipeline(ctx, ctx->device->pipeline_diag_mask_inf_f32, "diag_mask_inf_f32", diag_mask_inf_f32_len, diag_mask_inf_f32_data, "main", 2, sizeof(vk_op_diag_mask_push_constants), {512, 1, 1}, {}, 1);
 
-    ggml_vk_create_pipeline(ctx, ctx->device->pipeline_soft_max_f32, "soft_max_f32", soft_max_f32_len, soft_max_f32_data, "main", 4, sizeof(vk_op_soft_max_push_constants), {1, 1, 1}, {}, 1);
-    ggml_vk_create_pipeline(ctx, ctx->device->pipeline_soft_max_f32_f16, "soft_max_f32_f16", soft_max_f32_f16_len, soft_max_f32_f16_data, "main", 4, sizeof(vk_op_soft_max_push_constants), {1, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(ctx, ctx->device->pipeline_soft_max_f32, "soft_max_f32", soft_max_f32_len, soft_max_f32_data, "main", 3, sizeof(vk_op_soft_max_push_constants), {1, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(ctx, ctx->device->pipeline_soft_max_f32_f16, "soft_max_f32_f16", soft_max_f32_f16_len, soft_max_f32_f16_data, "main", 3, sizeof(vk_op_soft_max_push_constants), {1, 1, 1}, {}, 1);
 
     ggml_vk_create_pipeline(ctx, ctx->device->pipeline_rope_f32, "rope_f32", rope_f32_len, rope_f32_data, "main", 3, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1);
     ggml_vk_create_pipeline(ctx, ctx->device->pipeline_rope_f16, "rope_f16", rope_f16_len, rope_f16_data, "main", 3, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1);
@@ -1859,7 +1874,6 @@ static void ggml_vk_init(ggml_backend_vk_context * ctx, size_t idx) {
     ctx->compute_ctx = nullptr;
     ctx->transfer_ctx = nullptr;
 
-    ctx->disable = false;
     ctx->initialized = true;
 
     ctx->idx = idx;
@@ -1903,6 +1917,9 @@ static vk_matmul_pipeline ggml_vk_get_mul_mat_mat_pipeline(ggml_backend_vk_conte
     if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_F32) {
         return ctx->device->pipeline_matmul_f32;
     }
+    if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_F16) {
+        return ctx->device->pipeline_matmul_f32_f16;
+    }
     if (src0_type == GGML_TYPE_F16 && src1_type == GGML_TYPE_F32) {
         return ctx->device->pipeline_matmul_f16_f32;
     }
@@ -2722,7 +2739,7 @@ static void ggml_vk_matmul(
         uint32_t batch_stride_a, uint32_t batch_stride_b, uint32_t batch_stride_d,
         uint32_t expert_stride_b, uint32_t expert_stride_d, uint32_t idx, uint32_t nbi1, uint32_t n_as) {
 #ifdef GGML_VULKAN_DEBUG
-    std::cerr << "ggml_vk_matmul(a: (" << a.buffer->buffer << ", " << a.offset << ", " << a.size << "), b: (" << b.buffer->buffer << ", " << b.offset << ", " << b.size << "), c: (" << d.buffer->buffer << ", " << d.offset << ", " << d.size << "), split_k: (" << split_k_buffer.buffer->buffer << ", " << split_k_buffer.offset << ", " << split_k_buffer.size << "), m: " << m << ", n: " << n << ", k: " << k << ", stride_a: " << stride_a << ", stride_b: " << stride_b << ", stride_d: " << stride_d << ", split_k: " << split_k << ", batch: " << batch << ", ne02: " << ne02 << ", ne12: " << ne12 << ", broadcast2: " << broadcast2 << ", broadcast3: " << broadcast3 << ", batch_stride_a: " << batch_stride_a << ", batch_stride_b: " << batch_stride_b << ", batch_stride_d: " << batch_stride_d << ")" << std::endl;
+    std::cerr << "ggml_vk_matmul(a: (" << a.buffer->buffer << ", " << a.offset << ", " << a.size << "), b: (" << b.buffer->buffer << ", " << b.offset << ", " << b.size << "), c: (" << d.buffer->buffer << ", " << d.offset << ", " << d.size << "), split_k: (" << (split_k_buffer.buffer != nullptr ? split_k_buffer.buffer->buffer : VK_NULL_HANDLE) << ", " << split_k_buffer.offset << ", " << split_k_buffer.size << "), m: " << m << ", n: " << n << ", k: " << k << ", stride_a: " << stride_a << ", stride_b: " << stride_b << ", stride_d: " << stride_d << ", split_k: " << split_k << ", batch: " << batch << ", ne02: " << ne02 << ", ne12: " << ne12 << ", broadcast2: " << broadcast2 << ", broadcast3: " << broadcast3 << ", batch_stride_a: " << batch_stride_a << ", batch_stride_b: " << batch_stride_b << ", batch_stride_d: " << batch_stride_d << ")" << std::endl;
 #endif
     ggml_vk_sync_buffers(subctx);
     if (split_k == 1) {
@@ -2792,7 +2809,7 @@ static vk_pipeline ggml_vk_get_cpy_pipeline(ggml_backend_vk_context * ctx, ggml_
 
 static void ggml_vk_cpy_to_contiguous(ggml_backend_vk_context * ctx, vk_context * subctx, vk_pipeline pipeline, const ggml_tensor * tensor, vk_subbuffer&& in, vk_subbuffer&& out) {
 #ifdef GGML_VULKAN_DEBUG
-    std::cerr << "ggml_vk_cpy_to_contiguous((" << tensor << ", type=" << tensor->type << ", backend=" << tensor->backend << ", ne0=" << tensor->ne[0] << ", ne1=" << tensor->ne[1] << ", ne2=" << tensor->ne[2] << ", ne3=" << tensor->ne[3] << ", nb0=" << tensor->nb[0] << ", nb1=" << tensor->nb[1] << ", nb2=" << tensor->nb[2] << ", nb3=" << tensor->nb[3] << "), ";
+    std::cerr << "ggml_vk_cpy_to_contiguous((" << tensor << ", type=" << tensor->type << ", ne0=" << tensor->ne[0] << ", ne1=" << tensor->ne[1] << ", ne2=" << tensor->ne[2] << ", ne3=" << tensor->ne[3] << ", nb0=" << tensor->nb[0] << ", nb1=" << tensor->nb[1] << ", nb2=" << tensor->nb[2] << ", nb3=" << tensor->nb[3] << "), ";
     std::cerr << "buffer in size=" << in.buffer->size << ", buffer out size=" << out.buffer->size << ")" << std::endl;
 #endif
     const int tensor_type_size = ggml_type_size(tensor->type);
@@ -2812,9 +2829,9 @@ static void ggml_vk_cpy_to_contiguous(ggml_backend_vk_context * ctx, vk_context
 
 static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
 #ifdef GGML_VULKAN_DEBUG
-    std::cerr << "ggml_vk_mul_mat_q_f16((" << src0 << ", name=" << src0->name << ", type=" << src0->type << ", backend=" << src0->backend << ", ne0=" << src0->ne[0] << ", ne1=" << src0->ne[1] << ", ne2=" << src0->ne[2] << ", ne3=" << src0->ne[3] << ", nb0=" << src0->nb[0] << ", nb1=" << src0->nb[1] << ", nb2=" << src0->nb[2] << ", nb3=" << src0->nb[3];
-    std::cerr << "), (" << src1 << ", name=" << src1->name << ", type=" << src1->type << ", backend=" << src1->backend << ", ne0=" << src1->ne[0] << ", ne1=" << src1->ne[1] << ", ne2=" << src1->ne[2] << ", ne3=" << src1->ne[3] << ", nb0=" << src1->nb[0] << ", nb1=" << src1->nb[1] << ", nb2=" << src1->nb[2] << ", nb3=" << src1->nb[3];
-    std::cerr << "), (" << dst << ", name=" << dst->name << ", type=" << dst->type << ", backend=" << dst->backend << ", ne0=" << dst->ne[0] << ", ne1=" << dst->ne[1] << ", ne2=" << dst->ne[2] << ", ne3=" << dst->ne[3] << ", nb0=" << dst->nb[0] << ", nb1=" << dst->nb[1] << ", nb2=" << dst->nb[2] << ", nb3=" << dst->nb[3] << "),)" << std::endl;
+    std::cerr << "ggml_vk_mul_mat_q_f16((" << src0 << ", name=" << src0->name << ", type=" << src0->type << ", ne0=" << src0->ne[0] << ", ne1=" << src0->ne[1] << ", ne2=" << src0->ne[2] << ", ne3=" << src0->ne[3] << ", nb0=" << src0->nb[0] << ", nb1=" << src0->nb[1] << ", nb2=" << src0->nb[2] << ", nb3=" << src0->nb[3];
+    std::cerr << "), (" << src1 << ", name=" << src1->name << ", type=" << src1->type << ", ne0=" << src1->ne[0] << ", ne1=" << src1->ne[1] << ", ne2=" << src1->ne[2] << ", ne3=" << src1->ne[3] << ", nb0=" << src1->nb[0] << ", nb1=" << src1->nb[1] << ", nb2=" << src1->nb[2] << ", nb3=" << src1->nb[3];
+    std::cerr << "), (" << dst << ", name=" << dst->name << ", type=" << dst->type << ", ne0=" << dst->ne[0] << ", ne1=" << dst->ne[1] << ", ne2=" << dst->ne[2] << ", ne3=" << dst->ne[3] << ", nb0=" << dst->nb[0] << ", nb1=" << dst->nb[1] << ", nb2=" << dst->nb[2] << ", nb3=" << dst->nb[3] << "),)" << std::endl;
 #endif
     GGML_ASSERT(ggml_vk_dim01_contiguous(src0) || src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16);  // NOLINT
     GGML_ASSERT(ggml_vk_dim01_contiguous(src1) || src1->type == GGML_TYPE_F32 || src1->type == GGML_TYPE_F16);  // NOLINT
@@ -2982,19 +2999,13 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context * su
         ne01, ne11, ne10, ne10, ne10, ne01, split_k, ne12*ne13, ne02, ne12, r2, r3, stride_batch_x, stride_batch_y, ne20*ne21,
         0, 0, 0, 0, 1
     );  // NOLINT
-
-    if (dst->backend == GGML_BACKEND_TYPE_CPU) {
-        // copy dst to host
-        float * d = (float *) ((char *) dst->data);
-        ggml_vk_buffer_read_async(ctx, subctx, d_D, 0, d, sizeof(float) * d_ne * ne12 * ne13);
-    }
 }
 
 static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
 #ifdef GGML_VULKAN_DEBUG
-    std::cerr << "ggml_vk_mul_mat_vec_q_f16((" << src0 << ", name=" << src0->name << ", type=" << src0->type << ",  backend=" << src0->backend << ", ne0=" << src0->ne[0] << ", ne1=" << src0->ne[1] << ", ne2=" << src0->ne[2] << ", ne3=" << src0->ne[3] << ", nb0=" << src0->nb[0] << ", nb1=" << src0->nb[1] << ", nb2=" << src0->nb[2] << ", nb3=" << src0->nb[3];
-    std::cerr << "), (" << src1 << ", name=" << src1->name << ", type=" << src1->type << ",  backend=" << src1->backend << ", ne0=" << src1->ne[0] << ", ne1=" << src1->ne[1] << ", ne2=" << src1->ne[2] << ", ne3=" << src1->ne[3] << ", nb0=" << src1->nb[0] << ", nb1=" << src1->nb[1] << ", nb2=" << src1->nb[2] << ", nb3=" << src1->nb[3];
-    std::cerr << "), (" << dst << ", name=" << dst->name << ", type=" << dst->type << ",  backend=" << dst->backend << ", ne0=" << dst->ne[0] << ", ne1=" << dst->ne[1] << ", ne2=" << dst->ne[2] << ", ne3=" << dst->ne[3] << ", nb0=" << dst->nb[0] << ", nb1=" << dst->nb[1] << ", nb2=" << dst->nb[2] << ", nb3=" << dst->nb[3] << "),)" << std::endl;
+    std::cerr << "ggml_vk_mul_mat_vec_q_f16((" << src0 << ", name=" << src0->name << ", type=" << src0->type << ", ne0=" << src0->ne[0] << ", ne1=" << src0->ne[1] << ", ne2=" << src0->ne[2] << ", ne3=" << src0->ne[3] << ", nb0=" << src0->nb[0] << ", nb1=" << src0->nb[1] << ", nb2=" << src0->nb[2] << ", nb3=" << src0->nb[3];
+    std::cerr << "), (" << src1 << ", name=" << src1->name << ", type=" << src1->type << ", ne0=" << src1->ne[0] << ", ne1=" << src1->ne[1] << ", ne2=" << src1->ne[2] << ", ne3=" << src1->ne[3] << ", nb0=" << src1->nb[0] << ", nb1=" << src1->nb[1] << ", nb2=" << src1->nb[2] << ", nb3=" << src1->nb[3];
+    std::cerr << "), (" << dst << ", name=" << dst->name << ", type=" << dst->type << ", ne0=" << dst->ne[0] << ", ne1=" << dst->ne[1] << ", ne2=" << dst->ne[2] << ", ne3=" << dst->ne[3] << ", nb0=" << dst->nb[0] << ", nb1=" << dst->nb[1] << ", nb2=" << dst->nb[2] << ", nb3=" << dst->nb[3] << "),)" << std::endl;
 #endif
     GGML_ASSERT(ggml_vk_dim01_contiguous(src0) || src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16);  // NOLINT
     GGML_ASSERT(ggml_vk_dim01_contiguous(src1) || src1->type == GGML_TYPE_F32 || src1->type == GGML_TYPE_F16);  // NOLINT
@@ -3147,12 +3158,11 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context
 
 static void ggml_vk_mul_mat_vec_p021_f16_f32(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
 #ifdef GGML_VULKAN_DEBUG
-    std::cerr << "ggml_vk_mul_mat_p021_f16_f32((" << src0 << ", name=" << src0->name << ", type=" << src0->type << ",  backend=" << src0->backend << ", ne0=" << src0->ne[0] << ", ne1=" << src0->ne[1] << ", ne2=" << src0->ne[2] << ", ne3=" << src0->ne[3] << ", nb0=" << src0->nb[0] << ", nb1=" << src0->nb[1] << ", nb2=" << src0->nb[2] << ", nb3=" << src0->nb[3];
-    std::cerr << "), (" << src1 << ", name=" << src1->name << ", type=" << src1->type << ",  backend=" << src1->backend << ", ne0=" << src1->ne[0] << ", ne1=" << src1->ne[1] << ", ne2=" << src1->ne[2] << ", ne3=" << src1->ne[3] << ", nb0=" << src1->nb[0] << ", nb1=" << src1->nb[1] << ", nb2=" << src1->nb[2] << ", nb3=" << src1->nb[3];
-    std::cerr << "), (" << dst << ", name=" << dst->name << ", type=" << dst->type << ",  backend=" << dst->backend << ", ne0=" << dst->ne[0] << ", ne1=" << dst->ne[1] << ", ne2=" << dst->ne[2] << ", ne3=" << dst->ne[3] << ", nb0=" << dst->nb[0] << ", nb1=" << dst->nb[1] << ", nb2=" << dst->nb[2] << ", nb3=" << dst->nb[3] << "),)" << std::endl;
+    std::cerr << "ggml_vk_mul_mat_p021_f16_f32((" << src0 << ", name=" << src0->name << ", type=" << src0->type << ", ne0=" << src0->ne[0] << ", ne1=" << src0->ne[1] << ", ne2=" << src0->ne[2] << ", ne3=" << src0->ne[3] << ", nb0=" << src0->nb[0] << ", nb1=" << src0->nb[1] << ", nb2=" << src0->nb[2] << ", nb3=" << src0->nb[3];
+    std::cerr << "), (" << src1 << ", name=" << src1->name << ", type=" << src1->type << ", ne0=" << src1->ne[0] << ", ne1=" << src1->ne[1] << ", ne2=" << src1->ne[2] << ", ne3=" << src1->ne[3] << ", nb0=" << src1->nb[0] << ", nb1=" << src1->nb[1] << ", nb2=" << src1->nb[2] << ", nb3=" << src1->nb[3];
+    std::cerr << "), (" << dst << ", name=" << dst->name << ", type=" << dst->type << ", ne0=" << dst->ne[0] << ", ne1=" << dst->ne[1] << ", ne2=" << dst->ne[2] << ", ne3=" << dst->ne[3] << ", nb0=" << dst->nb[0] << ", nb1=" << dst->nb[1] << ", nb2=" << dst->nb[2] << ", nb3=" << dst->nb[3] << "),)" << std::endl;
 #endif
     GGML_ASSERT(ggml_is_permuted(src0) && ggml_is_permuted(src1));
-    GGML_ASSERT(src0->backend == GGML_BACKEND_TYPE_GPU);
     GGML_ASSERT(src0->nb[0] <= src0->nb[1] && src0->nb[2] <= src0->nb[3]);  // NOLINT
     GGML_ASSERT(src1->nb[0] <= src1->nb[1] && src1->nb[2] <= src1->nb[3]);  // NOLINT
     GGML_ASSERT(src0->type == GGML_TYPE_F16);
@@ -3217,25 +3227,17 @@ static void ggml_vk_mul_mat_vec_p021_f16_f32(ggml_backend_vk_context * ctx, vk_c
     const std::array<uint32_t, 6> pc = { (uint32_t)ne00, (uint32_t)ne01, (uint32_t)ne02, (uint32_t)ne12, (uint32_t)(qy_shader_offset / ggml_type_size(src1->type)), (uint32_t)(d_shader_offset / ggml_type_size(dst->type)) };
     ggml_vk_sync_buffers(subctx);
     ggml_vk_dispatch_pipeline(ctx, subctx, ctx->device->pipeline_mul_mat_vec_p021_f16_f32, { { d_Qx, qx_buf_offset, qx_sz }, { d_Qy, qy_buffer_offset, qy_sz + qy_shader_offset }, { d_D, d_buffer_offset, d_sz + d_shader_offset } }, 6 * sizeof(uint32_t), &pc, { 1, (uint32_t)ne01, (uint32_t)ne12 });
-
-    if (dst->backend == GGML_BACKEND_TYPE_CPU) {
-        // copy dst to host
-        float * d = (float *) dst->data;
-        ggml_vk_sync_buffers(subctx);
-        ggml_vk_buffer_read_async(ctx, subctx, d_D, d_buf_offset, d, sizeof(float) * d_ne);
-    }
 }
 
 static void ggml_vk_mul_mat_vec_nc_f16_f32(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
 #ifdef GGML_VULKAN_DEBUG
-    std::cerr << "ggml_vk_mul_mat_nc_f16_f32((" << src0 << ", name=" << src0->name << ", type=" << src0->type << ",  backend=" << src0->backend << ", ne0=" << src0->ne[0] << ", ne1=" << src0->ne[1] << ", ne2=" << src0->ne[2] << ", ne3=" << src0->ne[3] << ", nb0=" << src0->nb[0] << ", nb1=" << src0->nb[1] << ", nb2=" << src0->nb[2] << ", nb3=" << src0->nb[3];
-    std::cerr << "), (" << src1 << ", name=" << src1->name << ", type=" << src1->type << ",  backend=" << src1->backend << ", ne0=" << src1->ne[0] << ", ne1=" << src1->ne[1] << ", ne2=" << src1->ne[2] << ", ne3=" << src1->ne[3] << ", nb0=" << src1->nb[0] << ", nb1=" << src1->nb[1] << ", nb2=" << src1->nb[2] << ", nb3=" << src1->nb[3];
-    std::cerr << "), (" << dst << ", name=" << dst->name << ", type=" << dst->type << ",  backend=" << dst->backend << ", ne0=" << dst->ne[0] << ", ne1=" << dst->ne[1] << ", ne2=" << dst->ne[2] << ", ne3=" << dst->ne[3] << ", nb0=" << dst->nb[0] << ", nb1=" << dst->nb[1] << ", nb2=" << dst->nb[2] << ", nb3=" << dst->nb[3] << "),)" << std::endl;
+    std::cerr << "ggml_vk_mul_mat_nc_f16_f32((" << src0 << ", name=" << src0->name << ", type=" << src0->type << ", ne0=" << src0->ne[0] << ", ne1=" << src0->ne[1] << ", ne2=" << src0->ne[2] << ", ne3=" << src0->ne[3] << ", nb0=" << src0->nb[0] << ", nb1=" << src0->nb[1] << ", nb2=" << src0->nb[2] << ", nb3=" << src0->nb[3];
+    std::cerr << "), (" << src1 << ", name=" << src1->name << ", type=" << src1->type << ", ne0=" << src1->ne[0] << ", ne1=" << src1->ne[1] << ", ne2=" << src1->ne[2] << ", ne3=" << src1->ne[3] << ", nb0=" << src1->nb[0] << ", nb1=" << src1->nb[1] << ", nb2=" << src1->nb[2] << ", nb3=" << src1->nb[3];
+    std::cerr << "), (" << dst << ", name=" << dst->name << ", type=" << dst->type << ", ne0=" << dst->ne[0] << ", ne1=" << dst->ne[1] << ", ne2=" << dst->ne[2] << ", ne3=" << dst->ne[3] << ", nb0=" << dst->nb[0] << ", nb1=" << dst->nb[1] << ", nb2=" << dst->nb[2] << ", nb3=" << dst->nb[3] << "),)" << std::endl;
 #endif
     GGML_ASSERT(!ggml_is_transposed(src0));
     GGML_ASSERT(!ggml_is_transposed(src1));
     GGML_ASSERT(!ggml_is_permuted(src0));
-    GGML_ASSERT(src0->backend == GGML_BACKEND_TYPE_GPU);
     GGML_ASSERT(src0->type == GGML_TYPE_F16);
     GGML_ASSERT(src1->type == GGML_TYPE_F32);
 
@@ -3302,26 +3304,6 @@ static void ggml_vk_mul_mat_vec_nc_f16_f32(ggml_backend_vk_context * ctx, vk_con
     const std::array<uint32_t, 7> pc = { (uint32_t)ne00, (uint32_t)ne01, row_stride_x, channel_stride_x, (uint32_t)(ne12 / ne02), (uint32_t)(qy_shader_offset / ggml_type_size(src1->type)), (uint32_t)(d_shader_offset / ggml_type_size(dst->type)) };
     ggml_vk_sync_buffers(subctx);
     ggml_vk_dispatch_pipeline(ctx, subctx, ctx->device->pipeline_mul_mat_vec_nc_f16_f32, { { d_Qx, qx_buf_offset, qx_sz }, { d_Qy, qy_buffer_offset, qy_sz + qy_shader_offset }, { d_D, d_buffer_offset, d_sz + d_shader_offset } }, 7 * sizeof(uint32_t), &pc, { 1, (uint32_t)ne01, (uint32_t)ne12 });
-
-    if (dst->backend == GGML_BACKEND_TYPE_CPU) {
-        // copy dst to host
-        float * d = (float *) dst->data;
-        ggml_vk_sync_buffers(subctx);
-        ggml_vk_buffer_read_async(ctx, subctx, d_D, d_buf_offset, d, sizeof(float) * d_ne);
-    }
-}
-
-static bool ggml_vk_can_mul_mat(const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * dst) {
-    const uint64_t ne10 = src1->ne[0];
-
-    const uint64_t ne0 = dst->ne[0];
-    const uint64_t ne1 = dst->ne[1];
-
-    // TODO: find the optimal values for these
-    return (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 || ggml_is_quantized(src0->type)) &&
-           (src1->type == GGML_TYPE_F32 || src1->type == GGML_TYPE_F16 || ggml_is_quantized(src1->type)) &&
-           dst->type == GGML_TYPE_F32 &&
-           ((ne0 >= 32 && ne1 >= 32 && ne10 >= 32) || src0->backend == GGML_BACKEND_TYPE_GPU);
 }
 
 static void ggml_vk_mul_mat(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
@@ -3711,8 +3693,6 @@ static void ggml_vk_op_repeat(ggml_backend_vk_context * ctx, vk_context * subctx
     // TODO: support for transposed / permuted tensors
     GGML_ASSERT(nb0  == sizeof(float));
     GGML_ASSERT(nb00 == sizeof(float));
-    GGML_ASSERT(src0->backend == GGML_BACKEND_TYPE_GPU);
-    GGML_ASSERT(dst->backend == GGML_BACKEND_TYPE_GPU);
 
     ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) dst->extra;
     ggml_tensor_extra_gpu * extra_src0 = (ggml_tensor_extra_gpu *) src0->extra;
@@ -3752,7 +3732,7 @@ static void ggml_vk_op_repeat(ggml_backend_vk_context * ctx, vk_context * subctx
 }
 
 
-static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, ggml_tensor * dst, ggml_op op) {
+static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, ggml_op op) {
     switch (op) {
     case GGML_OP_ADD:
         if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
@@ -3834,7 +3814,7 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
         if (src0->type == GGML_TYPE_F32 && (src1 == nullptr || src1->type == GGML_TYPE_F32) && dst->type == GGML_TYPE_F32) {
             return ctx->device->pipeline_soft_max_f32;
         }
-        if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F16 && src2->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F32) {
+        if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F32) {
             return ctx->device->pipeline_soft_max_f32_f16;
         }
         return nullptr;
@@ -3900,16 +3880,13 @@ static bool ggml_vk_op_supports_incontiguous(ggml_op op) {
 }
 
 template<typename PC>
-static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, ggml_tensor * dst, ggml_op op, const PC&& pc) {
+static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, ggml_op op, const PC&& pc) {
 #ifdef GGML_VULKAN_DEBUG
-    std::cerr << "ggml_vk_op_f32((" << src0 << ", name=" << src0->name << ", type=" << src0->type << ", backend=" << src0->backend << ", ne0=" << src0->ne[0] << ", ne1=" << src0->ne[1] << ", ne2=" << src0->ne[2] << ", ne3=" << src0->ne[3] << ", nb0=" << src0->nb[0] << ", nb1=" << src0->nb[1] << ", nb2=" << src0->nb[2] << ", nb3=" << src0->nb[3];
+    std::cerr << "ggml_vk_op_f32((" << src0 << ", name=" << src0->name << ", type=" << src0->type << ", ne0=" << src0->ne[0] << ", ne1=" << src0->ne[1] << ", ne2=" << src0->ne[2] << ", ne3=" << src0->ne[3] << ", nb0=" << src0->nb[0] << ", nb1=" << src0->nb[1] << ", nb2=" << src0->nb[2] << ", nb3=" << src0->nb[3];
     if (src1 != nullptr) {
-        std::cerr << "), (" << src1 << ", name=" << src1->name << ", type=" << src1->type << ", backend=" << src1->backend << ", ne0=" << src1->ne[0] << ", ne1=" << src1->ne[1] << ", ne2=" << src1->ne[2] << ", ne3=" << src1->ne[3] << ", nb0=" << src1->nb[0] << ", nb1=" << src1->nb[1] << ", nb2=" << src1->nb[2] << ", nb3=" << src1->nb[3];
+        std::cerr << "), (" << src1 << ", name=" << src1->name << ", type=" << src1->type << ", ne0=" << src1->ne[0] << ", ne1=" << src1->ne[1] << ", ne2=" << src1->ne[2] << ", ne3=" << src1->ne[3] << ", nb0=" << src1->nb[0] << ", nb1=" << src1->nb[1] << ", nb2=" << src1->nb[2] << ", nb3=" << src1->nb[3];
     }
-    if (src2 != nullptr) {
-        std::cerr << "), (" << src2 << ", name=" << src2->name << ", type=" << src2->type << ", backend=" << src2->backend << ", ne0=" << src2->ne[0] << ", ne1=" << src2->ne[1] << ", ne2=" << src2->ne[2] << ", ne3=" << src2->ne[3] << ", nb0=" << src2->nb[0] << ", nb1=" << src2->nb[1] << ", nb2=" << src2->nb[2] << ", nb3=" << src2->nb[3];
-    }
-    std::cerr << "), (" << dst << ", name=" << dst->name << ", type=" << dst->type << ", backend=" << dst->backend << ", ne0=" << dst->ne[0] << ", ne1=" << dst->ne[1] << ", ne2=" << dst->ne[2] << ", ne3=" << dst->ne[3] << ", nb0=" << dst->nb[0] << ", nb1=" << dst->nb[1] << ", nb2=" << dst->nb[2] << ", nb3=" << dst->nb[3] << "), " << ggml_op_name(op) << ")" << std::endl;
+    std::cerr << "), (" << dst << ", name=" << dst->name << ", type=" << dst->type << ", ne0=" << dst->ne[0] << ", ne1=" << dst->ne[1] << ", ne2=" << dst->ne[2] << ", ne3=" << dst->ne[3] << ", nb0=" << dst->nb[0] << ", nb1=" << dst->nb[1] << ", nb2=" << dst->nb[2] << ", nb3=" << dst->nb[3] << "), " << ggml_op_name(op) << ")" << std::endl;
 #endif
     GGML_ASSERT(op == GGML_OP_GET_ROWS || (!ggml_is_quantized(src0->type) && (src1 == nullptr || !ggml_is_quantized(src1->type))));  // NOLINT
     GGML_ASSERT(op == GGML_OP_CPY || ggml_vk_dim01_contiguous(src0));  // NOLINT
@@ -3926,13 +3903,8 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context * subctx, c
     const uint64_t ne13 = use_src1 ? src1->ne[3] : 0;
     const uint64_t ne1 = ne10 * ne11;
     // const uint64_t nb10 = use_src1 ? src1->nb[0] : 0;
-    const uint64_t nb2  = dst->nb[2];
-    const uint64_t nb3  = dst->nb[3];
 
-    const bool use_src2 = src2 != nullptr;
-    const uint64_t ne2 = use_src2 ? src2->ne[0] * src2->ne[1] : 0;
-
-    vk_pipeline pipeline = ggml_vk_op_get_pipeline(ctx, src0, src1, src2, dst, op);
+    vk_pipeline pipeline = ggml_vk_op_get_pipeline(ctx, src0, src1, dst, op);
     ggml_vk_func_t op_func;
 
     if (pipeline == nullptr) {
@@ -3955,18 +3927,15 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context * subctx, c
     ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) dst->extra;
     ggml_tensor_extra_gpu * extra_src0 = (ggml_tensor_extra_gpu *) src0->extra;
     ggml_tensor_extra_gpu * extra_src1 = use_src1 ? (ggml_tensor_extra_gpu *) src1->extra : nullptr;
-    ggml_tensor_extra_gpu * extra_src2 = use_src2 ? (ggml_tensor_extra_gpu *) src2->extra : nullptr;
 
     vk_buffer d_X = nullptr;
     size_t x_buf_offset = 0;
     vk_buffer d_Y = nullptr;
     size_t y_buf_offset = 0;
     vk_buffer d_Z = nullptr;
-    size_t z_buf_offset = 0;
 
     bool src0_uma = false;
     bool src1_uma = false;
-    bool src2_uma = false;
 
     if (ctx->device->uma) {
         ggml_vk_host_get(ctx, src0->data, d_X, x_buf_offset);
@@ -3975,21 +3944,16 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context * subctx, c
             ggml_vk_host_get(ctx, src1->data, d_Y, y_buf_offset);
             src1_uma = d_Y != nullptr;
         }
-        if (use_src2) {
-            ggml_vk_host_get(ctx, src1->data, d_Z, z_buf_offset);
-            src2_uma = d_Z != nullptr;
-        }
     }
 
     uint64_t x_sz = ggml_vk_align_size(ggml_type_size(src0->type)/ggml_blck_size(src0->type) * ne0, ctx->device->properties.limits.minStorageBufferOffsetAlignment);
     uint64_t y_sz = use_src1 ? ggml_vk_align_size(ggml_type_size(src1->type) * ne1, ctx->device->properties.limits.minStorageBufferOffsetAlignment) : 0;
-    uint64_t z_sz = use_src2 ? ggml_vk_align_size(ggml_type_size(src2->type) * ne2, ctx->device->properties.limits.minStorageBufferOffsetAlignment) : 0;
     uint64_t d_sz = ggml_type_size(dst->type) * ne0;
 
     vk_buffer d_D = extra->buffer_gpu.lock();
 
     // Workaround for tiny tensor inputs on ROPE
-    if (use_src1 && src1->backend == GGML_BACKEND_TYPE_GPU && y_sz > d_D->size) {
+    if (use_src1 && y_sz > d_D->size) {
         y_sz = VK_WHOLE_SIZE;
     }
 
@@ -4007,12 +3971,6 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context * subctx, c
         GGML_ASSERT(d_Y != nullptr);
     }
 
-    if (use_src2 && !src2_uma) {
-        d_Z = extra_src2->buffer_gpu.lock();
-        z_buf_offset = extra_src2->offset;
-        GGML_ASSERT(d_Z != nullptr);
-    }
-
     if (op_supports_incontiguous) {
         x_sz = ggml_nbytes(src0);
         y_sz = use_src1 ? ggml_nbytes(src1) : 0;
@@ -4046,7 +4004,10 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context * subctx, c
             elements = { (uint32_t)ggml_nrows(src0), (uint32_t)ne00, 1 };
             break;
         case GGML_OP_GET_ROWS:
-            elements = {  (uint32_t)ne00, (uint32_t)ne10, (uint32_t)(ne11 * ne12) };
+            elements = { (uint32_t)ne00, (uint32_t)ne10, (uint32_t)(ne11 * ne12) };
+            break;
+        case GGML_OP_ARGSORT:
+            elements = { (uint32_t)ne00, (uint32_t)ggml_nrows(src0), 1 };
             break;
         default:
             elements = { (uint32_t)ggml_nelements(src0), 1, 1 };
@@ -4066,7 +4027,7 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context * subctx, c
         }
 
         if (op == GGML_OP_SOFT_MAX) {
-            // Empty src1 and src2 are possible on soft_max, but the shader needs buffers
+            // Empty src1 is possible on soft_max, but the shader needs a buffer
             vk_subbuffer subbuf_y;
             if (use_src1) {
                 subbuf_y = { d_Y, y_buf_offset, y_sz };
@@ -4074,15 +4035,8 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context * subctx, c
                 subbuf_y = { d_X, 0, d_X->size };
             }
 
-            vk_subbuffer subbuf_z;
-            if (use_src2) {
-                subbuf_z = { d_Z, z_buf_offset, z_sz };
-            } else {
-                subbuf_z = { d_X, 0, d_X->size };
-            }
-
             ggml_vk_sync_buffers(subctx);
-            ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { { d_X, x_buf_offset, x_sz }, subbuf_y, subbuf_z, { d_D, d_buf_offset, d_sz } }, sizeof(PC), &pc, elements);
+            ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { { d_X, x_buf_offset, x_sz }, subbuf_y, { d_D, d_buf_offset, d_sz } }, sizeof(PC), &pc, elements);
         } else if (use_src1) {
             ggml_vk_sync_buffers(subctx);
             ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { { d_X, x_buf_offset, x_sz }, { d_Y, y_buf_offset, y_sz }, { d_D, d_buf_offset, d_sz } }, sizeof(PC), &pc, elements);
@@ -4090,22 +4044,15 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context * subctx, c
             ggml_vk_sync_buffers(subctx);
             ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { { d_X, x_buf_offset, x_sz }, { d_D, d_buf_offset, d_sz } }, sizeof(PC), &pc, elements);
         }
-        if (dst->backend == GGML_BACKEND_TYPE_CPU && op == GGML_OP_CPY) {
-            ggml_vk_d2h_tensor_2d(ctx, subctx, d_D, 0, dst);
-        } else if(dst->backend == GGML_BACKEND_TYPE_CPU) {
-            // copy dst to host
-            float * d = (float *) dst->data;
-            ggml_vk_buffer_read_async(ctx, subctx, d_D, 0, d, d_sz);
-        }
     } else {
         GGML_ASSERT(op != GGML_OP_SOFT_MAX);
+        GGML_ASSERT(op != GGML_OP_ARGSORT);
 
         ggml_pipeline_allocate_descriptor_sets(ctx, pipeline, ne02 * ne03);
 
         switch (dst->op) {
         case GGML_OP_NORM:
         case GGML_OP_RMS_NORM:
-        case GGML_OP_SOFT_MAX:
             elements = { (uint32_t)ne01, 1, 1 };
             break;
         case GGML_OP_DIAG_MASK_INF:
@@ -4135,17 +4082,13 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context * subctx, c
                     ggml_vk_sync_buffers(subctx);
                     ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { { d_X, x_buf_offset + x_offset, x_sz }, { d_D, d_buf_offset + d_offset, d_sz } }, sizeof(PC), &pc, elements);
                 }
-                if (dst->backend == GGML_BACKEND_TYPE_CPU) {
-                    // copy dst to host
-                    ggml_vk_buffer_read_async(ctx, subctx, d_D, d_buf_offset + d_offset, (char *) dst->data + i02*nb2 + i03*nb3, d_sz);
-                }
             }
         }
     }
 }
 
 static void ggml_vk_repeat(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
-    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_REPEAT, { (uint32_t)ggml_nelements(src0), (uint32_t)ggml_nelements(src1), 0.0f, 0.0f });
+    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, src1, dst, GGML_OP_REPEAT, { (uint32_t)ggml_nelements(src0), (uint32_t)ggml_nelements(src1), 0.0f, 0.0f });
 }
 
 static void ggml_vk_get_rows(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
@@ -4153,7 +4096,7 @@ static void ggml_vk_get_rows(ggml_backend_vk_context * ctx, vk_context * subctx,
     const uint32_t src1_type_size = ggml_type_size(src1->type);
     const uint32_t dst_type_size = ggml_type_size(dst->type);
 
-    ggml_vk_op_f32<vk_op_binary_push_constants>(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_GET_ROWS, {
+    ggml_vk_op_f32<vk_op_binary_push_constants>(ctx, subctx, src0, src1, dst, GGML_OP_GET_ROWS, {
         (uint32_t)ggml_nelements(src0),
         (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2],(uint32_t)src0->ne[3], (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)src0->nb[1] / src0_type_size, (uint32_t)src0->nb[2] / src0_type_size, (uint32_t)src0->nb[3] / src0_type_size,
         (uint32_t)src1->ne[0], (uint32_t)src1->ne[1], (uint32_t)src1->ne[2],(uint32_t)src1->ne[3], (uint32_t)src1->nb[0] / src1_type_size, (uint32_t)src1->nb[1] / src1_type_size, (uint32_t)src1->nb[2] / src1_type_size, (uint32_t)src1->nb[3] / src1_type_size,
@@ -4168,7 +4111,7 @@ static void ggml_vk_add(ggml_backend_vk_context * ctx, vk_context * subctx, cons
     const uint32_t src1_type_size = ggml_type_size(src1->type);
     const uint32_t dst_type_size = ggml_type_size(dst->type);
 
-    ggml_vk_op_f32<vk_op_binary_push_constants>(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_ADD, {
+    ggml_vk_op_f32<vk_op_binary_push_constants>(ctx, subctx, src0, src1, dst, GGML_OP_ADD, {
         (uint32_t)ggml_nelements(src0),
         (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2],(uint32_t)src0->ne[3], (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)src0->nb[1] / src0_type_size, (uint32_t)src0->nb[2] / src0_type_size, (uint32_t)src0->nb[3] / src0_type_size,
         (uint32_t)src1->ne[0], (uint32_t)src1->ne[1], (uint32_t)src1->ne[2],(uint32_t)src1->ne[3], (uint32_t)src1->nb[0] / src1_type_size, (uint32_t)src1->nb[1] / src1_type_size, (uint32_t)src1->nb[2] / src1_type_size, (uint32_t)src1->nb[3] / src1_type_size,
@@ -4183,7 +4126,7 @@ static void ggml_vk_mul(ggml_backend_vk_context * ctx, vk_context * subctx, cons
     const uint32_t src1_type_size = ggml_type_size(src1->type);
     const uint32_t dst_type_size = ggml_type_size(dst->type);
 
-    ggml_vk_op_f32<vk_op_binary_push_constants>(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_MUL, {
+    ggml_vk_op_f32<vk_op_binary_push_constants>(ctx, subctx, src0, src1, dst, GGML_OP_MUL, {
         (uint32_t)ggml_nelements(src0),
         (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2],(uint32_t)src0->ne[3], (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)src0->nb[1] / src0_type_size, (uint32_t)src0->nb[2] / src0_type_size, (uint32_t)src0->nb[3] / src0_type_size,
         (uint32_t)src1->ne[0], (uint32_t)src1->ne[1], (uint32_t)src1->ne[2],(uint32_t)src1->ne[3], (uint32_t)src1->nb[0] / src1_type_size, (uint32_t)src1->nb[1] / src1_type_size, (uint32_t)src1->nb[2] / src1_type_size, (uint32_t)src1->nb[3] / src1_type_size,
@@ -4198,7 +4141,7 @@ static void ggml_vk_scale(ggml_backend_vk_context * ctx, vk_context * subctx, co
     const uint32_t src0_type_size = ggml_type_size(src0->type);
     const uint32_t dst_type_size = ggml_type_size(dst->type);
 
-    ggml_vk_op_f32<vk_op_unary_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_SCALE, {
+    ggml_vk_op_f32<vk_op_unary_push_constants>(ctx, subctx, src0, nullptr, dst, GGML_OP_SCALE, {
         (uint32_t)ggml_nelements(src0),
         (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2], (uint32_t)src0->ne[3], (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)src0->nb[1] / src0_type_size, (uint32_t)src0->nb[2] / src0_type_size, (uint32_t)src0->nb[3] / src0_type_size,
         (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2], (uint32_t) dst->ne[3], (uint32_t) dst->nb[0] /  dst_type_size, (uint32_t) dst->nb[1] /  dst_type_size, (uint32_t) dst->nb[2] /  dst_type_size, (uint32_t) dst->nb[3] /  dst_type_size,
@@ -4211,7 +4154,7 @@ static void ggml_vk_sqr(ggml_backend_vk_context * ctx, vk_context * subctx, cons
     const uint32_t src0_type_size = ggml_type_size(src0->type);
     const uint32_t dst_type_size = ggml_type_size(dst->type);
 
-    ggml_vk_op_f32<vk_op_unary_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_SQR, {
+    ggml_vk_op_f32<vk_op_unary_push_constants>(ctx, subctx, src0, nullptr, dst, GGML_OP_SQR, {
         (uint32_t)ggml_nelements(src0),
         (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2], (uint32_t)src0->ne[3], (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)src0->nb[1] / src0_type_size, (uint32_t)src0->nb[2] / src0_type_size, (uint32_t)src0->nb[3] / src0_type_size,
         (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2], (uint32_t) dst->ne[3], (uint32_t) dst->nb[0] /  dst_type_size, (uint32_t) dst->nb[1] /  dst_type_size, (uint32_t) dst->nb[2] /  dst_type_size, (uint32_t) dst->nb[3] /  dst_type_size,
@@ -4225,7 +4168,7 @@ static void ggml_vk_clamp(ggml_backend_vk_context * ctx, vk_context * subctx, co
     const uint32_t src0_type_size = ggml_type_size(src0->type);
     const uint32_t dst_type_size = ggml_type_size(dst->type);
 
-    ggml_vk_op_f32<vk_op_unary_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_CLAMP, {
+    ggml_vk_op_f32<vk_op_unary_push_constants>(ctx, subctx, src0, nullptr, dst, GGML_OP_CLAMP, {
         (uint32_t)ggml_nelements(src0),
         (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2], (uint32_t)src0->ne[3], (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)src0->nb[1] / src0_type_size, (uint32_t)src0->nb[2] / src0_type_size, (uint32_t)src0->nb[3] / src0_type_size,
         (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2], (uint32_t) dst->ne[3], (uint32_t) dst->nb[0] /  dst_type_size, (uint32_t) dst->nb[1] /  dst_type_size, (uint32_t) dst->nb[2] /  dst_type_size, (uint32_t) dst->nb[3] /  dst_type_size,
@@ -4240,7 +4183,7 @@ static void ggml_vk_cpy(ggml_backend_vk_context * ctx, vk_context * subctx, cons
     const uint32_t dst_type_size = ggml_type_size(dst->type);
     const uint32_t d_offset = (extra->offset % ctx->device->properties.limits.minStorageBufferOffsetAlignment) / dst_type_size;
 
-    ggml_vk_op_f32<vk_op_unary_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_CPY, {
+    ggml_vk_op_f32<vk_op_unary_push_constants>(ctx, subctx, src0, nullptr, dst, GGML_OP_CPY, {
         (uint32_t)ggml_nelements(src0),
         (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2], (uint32_t)src0->ne[3], (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)src0->nb[1] / src0_type_size, (uint32_t)src0->nb[2] / src0_type_size, (uint32_t)src0->nb[3] / src0_type_size,
         (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2], (uint32_t) dst->ne[3], (uint32_t) dst->nb[0] /  dst_type_size, (uint32_t) dst->nb[1] /  dst_type_size, (uint32_t) dst->nb[2] /  dst_type_size, (uint32_t) dst->nb[3] /  dst_type_size,
@@ -4252,24 +4195,24 @@ static void ggml_vk_cpy(ggml_backend_vk_context * ctx, vk_context * subctx, cons
 static void ggml_vk_norm(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, ggml_tensor * dst) {
     float * op_params = (float *)dst->op_params;
 
-    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_NORM, { (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], op_params[0], 0.0f });
+    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, dst, GGML_OP_NORM, { (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], op_params[0], 0.0f });
 }
 
 static void ggml_vk_rms_norm(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, ggml_tensor * dst) {
     float * op_params = (float *)dst->op_params;
-    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_RMS_NORM, { (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], op_params[0], 0.0f });
+    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, dst, GGML_OP_RMS_NORM, { (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], op_params[0], 0.0f });
 }
 
 static void ggml_vk_unary(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, ggml_tensor * dst) {
-    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_UNARY, { (uint32_t)ggml_nelements(src0), 0, 0.0f, 0.0f });
+    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, dst, GGML_OP_UNARY, { (uint32_t)ggml_nelements(src0), 0, 0.0f, 0.0f });
 }
 
 static void ggml_vk_diag_mask_inf(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, ggml_tensor * dst) {
     int32_t * op_params = (int32_t *)dst->op_params;
-    ggml_vk_op_f32<vk_op_diag_mask_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_DIAG_MASK_INF, { (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], op_params[0] });
+    ggml_vk_op_f32<vk_op_diag_mask_push_constants>(ctx, subctx, src0, nullptr, dst, GGML_OP_DIAG_MASK_INF, { (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], op_params[0] });
 }
 
-static void ggml_vk_soft_max(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, ggml_tensor * dst) {
+static void ggml_vk_soft_max(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     float * op_params = (float *)dst->op_params;
 
     float scale = op_params[0];
@@ -4285,13 +4228,9 @@ static void ggml_vk_soft_max(ggml_backend_vk_context * ctx, vk_context * subctx,
     const float m0 = powf(2.0f, -(max_bias       ) / n_head_log2);
     const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2);
 
-#pragma message("TODO: src2 is no longer used in soft_max - should be removed and ALiBi calculation should be updated")
-#pragma message("ref:  https://github.com/ggerganov/llama.cpp/pull/7192")
-
-    ggml_vk_op_f32<vk_op_soft_max_push_constants>(ctx, subctx, src0, src1, src2, dst, GGML_OP_SOFT_MAX, {
+    ggml_vk_op_f32<vk_op_soft_max_push_constants>(ctx, subctx, src0, src1, dst, GGML_OP_SOFT_MAX, {
         ncols,
         src1 != nullptr ? nrows_y : (uint32_t)0,
-        src2 != nullptr ? (uint32_t)1 : (uint32_t)0,
         scale, max_bias,
         m0, m1,
         n_head_log2,
@@ -4321,15 +4260,39 @@ static void ggml_vk_rope(ggml_backend_vk_context * ctx, vk_context * subctx, con
     if (is_neox) {
         const float theta_scale = powf(freq_base, -2.0f/n_dims);
         const float inv_ndims = -1.0f / n_dims;
-        ggml_vk_op_f32<vk_op_rope_neox_push_constants>(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_ROPE, { (uint32_t)src0->ne[0], (uint32_t)n_dims, freq_scale, (uint32_t)src0->ne[1], freq_base, ext_factor, attn_factor, {corr_dims[0], corr_dims[1], 0.0f, 0.0f}, theta_scale, inv_ndims });
+        ggml_vk_op_f32<vk_op_rope_neox_push_constants>(ctx, subctx, src0, src1, dst, GGML_OP_ROPE, {
+            (uint32_t)src0->ne[0], (uint32_t)n_dims, freq_scale, (uint32_t)src0->ne[1],
+            freq_base, ext_factor, attn_factor, {corr_dims[0], corr_dims[1], 0.0f, 0.0f}, theta_scale, inv_ndims
+        });
     } else {
-        ggml_vk_op_f32<vk_op_rope_push_constants>(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_ROPE, { (uint32_t)src0->ne[0], freq_scale, (uint32_t)src0->ne[1], freq_base, ext_factor, attn_factor, {corr_dims[0], corr_dims[1], 0.0f, 0.0f} });
+        ggml_vk_op_f32<vk_op_rope_push_constants>(ctx, subctx, src0, src1, dst, GGML_OP_ROPE, {
+            (uint32_t)src0->ne[0], freq_scale, (uint32_t)src0->ne[1],
+            freq_base, ext_factor, attn_factor, {corr_dims[0], corr_dims[1], 0.0f, 0.0f}
+        });
     }
 }
 
 static void ggml_vk_argsort(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, ggml_tensor * dst) {
     int32_t * op_params = (int32_t *)dst->op_params;
-    ggml_vk_op_f32<vk_op_argsort_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_ARGSORT, { (uint32_t)src0->ne[0], ((ggml_sort_order) op_params[0]) == GGML_SORT_ORDER_ASC });
+
+    uint32_t ncols = src0->ne[0];
+
+    uint32_t ncols_pad = 1;
+    while (ncols_pad < ncols) {
+        ncols_pad *= 2;
+    }
+
+    GGML_ASSERT(ncols_pad <= 1024);
+
+    std::cerr << "ncols=" << ncols << " ncols_pad=" << ncols_pad << " ascending=" << op_params[0] << std::endl;
+
+    std::cerr << ((ggml_sort_order) op_params[0]) << " " << GGML_SORT_ORDER_ASC << std::endl;
+
+    ggml_vk_op_f32<vk_op_argsort_push_constants>(ctx, subctx, src0, nullptr, dst, GGML_OP_ARGSORT, {
+        ncols,
+        ncols_pad,
+        op_params[0],
+    });
 }
 
 #ifdef GGML_VULKAN_RUN_TESTS
@@ -4381,6 +4344,9 @@ static void ggml_vk_test_matmul(ggml_backend_vk_context * ctx, size_t m, size_t
         if (std::is_same<float, X_TYPE>() && std::is_same<float, Y_TYPE>()) {
             p = ctx->device->pipeline_matmul_f32->a_s;
             shname = "F32_ALIGNED_S";
+        } else if (std::is_same<float, X_TYPE>() && std::is_same<ggml_fp16_t, Y_TYPE>()) {
+            p = ctx->device->pipeline_matmul_f32_f16->a_s;
+            shname = "F32_F16_ALIGNED_S";
         } else if (std::is_same<ggml_fp16_t, X_TYPE>() && std::is_same<float, Y_TYPE>()) {
             p = ctx->device->pipeline_matmul_f16_f32->a_s;
             shname = "F16_F32_ALIGNED_S";
@@ -4394,6 +4360,9 @@ static void ggml_vk_test_matmul(ggml_backend_vk_context * ctx, size_t m, size_t
         if (std::is_same<float, X_TYPE>() && std::is_same<float, Y_TYPE>()) {
             p = ctx->device->pipeline_matmul_f32->a_m;
             shname = "F32_ALIGNED_M";
+        } else if (std::is_same<float, X_TYPE>() && std::is_same<ggml_fp16_t, Y_TYPE>()) {
+            p = ctx->device->pipeline_matmul_f32_f16->a_m;
+            shname = "F32_F16_ALIGNED_M";
         } else if (std::is_same<ggml_fp16_t, X_TYPE>() && std::is_same<float, Y_TYPE>()) {
             p = ctx->device->pipeline_matmul_f16_f32->a_m;
             shname = "F16_F32_ALIGNED_M";
@@ -4407,6 +4376,9 @@ static void ggml_vk_test_matmul(ggml_backend_vk_context * ctx, size_t m, size_t
         if (std::is_same<float, X_TYPE>() && std::is_same<float, Y_TYPE>()) {
             p = ctx->device->pipeline_matmul_f32->a_l;
             shname = "F32_ALIGNED_L";
+        } else if (std::is_same<float, X_TYPE>() && std::is_same<ggml_fp16_t, Y_TYPE>()) {
+            p = ctx->device->pipeline_matmul_f32_f16->a_l;
+            shname = "F32_F16_ALIGNED_L";
         } else if (std::is_same<ggml_fp16_t, X_TYPE>() && std::is_same<float, Y_TYPE>()) {
             p = ctx->device->pipeline_matmul_f16_f32->a_l;
             shname = "F16_F32_ALIGNED_L";
@@ -4427,6 +4399,9 @@ static void ggml_vk_test_matmul(ggml_backend_vk_context * ctx, size_t m, size_t
             if (std::is_same<float, X_TYPE>() && std::is_same<float, Y_TYPE>()) {
                 p = ctx->device->pipeline_matmul_f32->s;
                 shname = "F32_S";
+            } else if (std::is_same<float, X_TYPE>() && std::is_same<ggml_fp16_t, Y_TYPE>()) {
+                p = ctx->device->pipeline_matmul_f32_f16->s;
+                shname = "F32_F16_S";
             } else if (std::is_same<ggml_fp16_t, X_TYPE>() && std::is_same<float, Y_TYPE>()) {
                 p = ctx->device->pipeline_matmul_f16_f32->s;
                 shname = "F16_F32_S";
@@ -4438,6 +4413,9 @@ static void ggml_vk_test_matmul(ggml_backend_vk_context * ctx, size_t m, size_t
             if (std::is_same<float, X_TYPE>() && std::is_same<float, Y_TYPE>()) {
                 p = ctx->device->pipeline_matmul_f32->m;
                 shname = "F32_M";
+            } else if (std::is_same<float, X_TYPE>() && std::is_same<ggml_fp16_t, Y_TYPE>()) {
+                p = ctx->device->pipeline_matmul_f32_f16->m;
+                shname = "F32_F16_M";
             } else if (std::is_same<ggml_fp16_t, X_TYPE>() && std::is_same<float, Y_TYPE>()) {
                 p = ctx->device->pipeline_matmul_f16_f32->m;
                 shname = "F16_F32_M";
@@ -4449,6 +4427,9 @@ static void ggml_vk_test_matmul(ggml_backend_vk_context * ctx, size_t m, size_t
             if (std::is_same<float, X_TYPE>() && std::is_same<float, Y_TYPE>()) {
                 p = ctx->device->pipeline_matmul_f32->l;
                 shname = "F32_L";
+            } else if (std::is_same<float, X_TYPE>() && std::is_same<ggml_fp16_t, Y_TYPE>()) {
+                p = ctx->device->pipeline_matmul_f32_f16->l;
+                shname = "F32_F16_L";
             } else if (std::is_same<ggml_fp16_t, X_TYPE>() && std::is_same<float, Y_TYPE>()) {
                 p = ctx->device->pipeline_matmul_f16_f32->l;
                 shname = "F16_F32_L";
@@ -4561,15 +4542,11 @@ static void ggml_vk_test_matmul(ggml_backend_vk_context * ctx, size_t m, size_t
     src1_ggml->data = y;
     tensor_ggml->data = d_chk;
 
-    ctx->disable = true;
-
     ggml_cgraph * cgraph = ggml_new_graph(ggml_ctx);
     ggml_build_forward_expand(cgraph, tensor_ggml);
 
     ggml_graph_compute_with_ctx(ggml_ctx, cgraph, 1);
 
-    ctx->disable = false;
-
     ggml_free(ggml_ctx);
 
     double avg_err = 0.0;
@@ -5049,15 +5026,11 @@ static void ggml_vk_test_dequant_matmul(ggml_backend_vk_context * ctx, size_t m,
     src1_ggml->data = y;
     tensor_ggml->data = d_chk;
 
-    ctx->disable = true;
-
     ggml_cgraph * cgraph = ggml_new_graph(ggml_ctx);
     ggml_build_forward_expand(cgraph, tensor_ggml);
 
     ggml_graph_compute_with_ctx(ggml_ctx, cgraph, 1);
 
-    ctx->disable = false;
-
     ggml_free(ggml_ctx);
 
     double avg_err = 0.0;
@@ -5134,12 +5107,12 @@ static void ggml_vk_preallocate_buffers_graph(ggml_backend_vk_context * ctx, ggm
 #ifdef GGML_VULKAN_DEBUG
     std::cerr << "ggml_vk_preallocate_buffers_graph(" << node << ")" << std::endl;
 #endif
-    if (ctx->disable || node->backend != GGML_BACKEND_TYPE_GPU) {
+    ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) node->extra;
+
+    if (extra == nullptr) {
         return;
     }
 
-    ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) node->extra;
-
     ggml_tensor * src0 = node->src[0];
     ggml_tensor * src1 = node->src[1];
 
@@ -5244,9 +5217,6 @@ static void ggml_vk_preallocate_buffers_graph(ggml_backend_vk_context * ctx, ggm
 }
 
 static void ggml_vk_preallocate_buffers(ggml_backend_vk_context * ctx) {
-    if (ctx->disable) {
-        return;
-    }
 #ifdef GGML_VULKAN_DEBUG
     std::cerr << "ggml_vk_preallocate_buffers(x_size: " << ctx->prealloc_size_x << " y_size: " << ctx->prealloc_size_y << " split_k_size: " << ctx->prealloc_size_split_k << ")" << std::endl;
 #endif
@@ -5420,7 +5390,9 @@ static void ggml_vk_preallocate_buffers(ggml_backend_vk_context * ctx) {
 }
 
 static void ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * node, bool last_node){
-    if (ctx->disable || node->backend != GGML_BACKEND_TYPE_GPU || ggml_is_empty(node)) {
+    ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) node->extra;
+
+    if (ggml_is_empty(node) || extra == nullptr) {
         return;
     }
 
@@ -5432,9 +5404,6 @@ static void ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod
 
     const ggml_tensor * src0 = node->src[0];
     const ggml_tensor * src1 = node->src[1];
-    const ggml_tensor * src2 = node->src[2];
-
-    ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) node->extra;
 
     switch (node->op) {
     case GGML_OP_UNARY:
@@ -5547,7 +5516,7 @@ static void ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod
 
         break;
     case GGML_OP_SOFT_MAX:
-        ggml_vk_soft_max(ctx, ctx->compute_ctx, src0, src1, src2, node);
+        ggml_vk_soft_max(ctx, ctx->compute_ctx, src0, src1, node);
 
         break;
     case GGML_OP_ROPE:
@@ -5580,7 +5549,7 @@ static void ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod
     last_node = true;
 #endif
 
-    if (node->backend == GGML_BACKEND_TYPE_CPU || last_node) {
+    if (last_node) {
         ggml_vk_ctx_end(ctx->compute_ctx);
         ctx->compute_ctx->exit_tensor = node;
         ctx->compute_ctx = nullptr;
@@ -5588,10 +5557,6 @@ static void ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod
 }
 
 static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_compute_params * params, ggml_tensor * tensor){
-    if (ctx->disable) {
-        return false;
-    }
-
     ggml_tensor_extra_gpu * extra = nullptr;
 
     switch (tensor->op) {
@@ -5650,7 +5615,7 @@ static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_compute_
     }
 
 #ifdef GGML_VULKAN_DEBUG
-    std::cerr << "ggml_vk_compute_forward(" << tensor << ", name=" << tensor->name << ", op=" << ggml_op_name(tensor->op) << ", type=" << tensor->type << ", backend=" << tensor->backend << ", ne0=" << tensor->ne[0] << ", ne1=" << tensor->ne[1] << ", ne2=" << tensor->ne[2] << ", ne3=" << tensor->ne[3] << ", nb0=" << tensor->nb[0] << ", nb1=" << tensor->nb[1] << ", nb2=" << tensor->nb[2] << ", nb3=" << tensor->nb[3] << ", view_src=" << tensor->view_src << ", view_offs=" << tensor->view_offs << ")" << std::endl;
+    std::cerr << "ggml_vk_compute_forward(" << tensor << ", name=" << tensor->name << ", op=" << ggml_op_name(tensor->op) << ", type=" << tensor->type << ", ne0=" << tensor->ne[0] << ", ne1=" << tensor->ne[1] << ", ne2=" << tensor->ne[2] << ", ne3=" << tensor->ne[3] << ", nb0=" << tensor->nb[0] << ", nb1=" << tensor->nb[1] << ", nb2=" << tensor->nb[2] << ", nb3=" << tensor->nb[3] << ", view_src=" << tensor->view_src << ", view_offs=" << tensor->view_offs << ")" << std::endl;
 #endif
 
 #ifdef GGML_VULKAN_CHECK_RESULTS
@@ -5690,9 +5655,6 @@ static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_compute_
 
 // Clean up after graph processing is done
 static void ggml_vk_graph_cleanup(ggml_backend_vk_context * ctx) {
-    if (ctx->disable) {
-        return;
-    }
 #ifdef GGML_VULKAN_DEBUG
     std::cerr << "ggml_vk_graph_cleanup()" << std::endl;
 #endif
@@ -5865,7 +5827,6 @@ GGML_CALL static void ggml_backend_vk_buffer_init_tensor(ggml_backend_buffer_t b
         extra->offset = (uint8_t *) tensor->data - (uint8_t *) vk_ptr_base;
     }
 
-    tensor->backend = GGML_BACKEND_TYPE_GPU;
     tensor->extra = extra;
 }
 
@@ -5873,8 +5834,6 @@ GGML_CALL static void ggml_backend_vk_buffer_set_tensor(ggml_backend_buffer_t bu
 #ifdef GGML_VULKAN_DEBUG
     std::cerr << "ggml_backend_vk_buffer_set_tensor(" << buffer << ", " << tensor << ", " << data << ", " << offset << ", " << size << ")" << std::endl;
 #endif
-    GGML_ASSERT(tensor->backend == GGML_BACKEND_TYPE_GPU);
-
     ggml_backend_vk_buffer_context * ctx = (ggml_backend_vk_buffer_context *)buffer->context;
 
     ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) tensor->extra;
@@ -5888,8 +5847,6 @@ GGML_CALL static void ggml_backend_vk_buffer_get_tensor(ggml_backend_buffer_t bu
 #ifdef GGML_VULKAN_DEBUG
     std::cerr << "ggml_backend_vk_buffer_get_tensor(" << buffer << ", " << tensor << ", " << data << ", " << offset << ", " << size << ")" << std::endl;
 #endif
-    GGML_ASSERT(tensor->backend == GGML_BACKEND_TYPE_GPU);
-
     ggml_backend_vk_buffer_context * ctx = (ggml_backend_vk_buffer_context *)buffer->context;
 
     ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) tensor->extra;
@@ -6032,6 +5989,7 @@ GGML_CALL static ggml_backend_buffer_t ggml_backend_vk_host_buffer_type_alloc_bu
 #ifdef GGML_VULKAN_DEBUG
     std::cerr << "ggml_backend_vk_host_buffer_type_alloc_buffer(" << size << ")" << std::endl;
 #endif
+    size += 32;  // Behave like the CPU buffer type
     void * ptr = nullptr;
     try {
         ptr = ggml_vk_host_malloc(&vk_instance.contexts[0], size);
@@ -6119,7 +6077,6 @@ GGML_CALL static void ggml_backend_vk_set_tensor_async(ggml_backend_t backend, g
 #endif
     ggml_backend_vk_context * ctx = (ggml_backend_vk_context *)backend->context;
     GGML_ASSERT((tensor->buffer->buft == ggml_backend_vk_buffer_type(ctx->idx) || tensor->buffer->buft == ggml_backend_vk_host_buffer_type()) && "unsupported buffer type");
-    GGML_ASSERT(tensor->backend == GGML_BACKEND_TYPE_GPU);
 
     ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) tensor->extra;
 
@@ -6140,7 +6097,6 @@ GGML_CALL static void ggml_backend_vk_get_tensor_async(ggml_backend_t backend, c
 #endif
     ggml_backend_vk_context * ctx = (ggml_backend_vk_context *)backend->context;
     GGML_ASSERT((tensor->buffer->buft == ggml_backend_vk_buffer_type(ctx->idx) || tensor->buffer->buft == ggml_backend_vk_host_buffer_type()) && "unsupported buffer type");
-    GGML_ASSERT(tensor->backend == GGML_BACKEND_TYPE_GPU);
 
     ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) tensor->extra;
 
@@ -6206,6 +6162,10 @@ GGML_CALL static void ggml_backend_vk_synchronize(ggml_backend_t backend) {
     ctx->transfer_ctx = nullptr;
 }
 
+static bool ggml_vk_is_empty(ggml_tensor * node) {
+    return ggml_is_empty(node) || node->op == GGML_OP_NONE || node->op == GGML_OP_RESHAPE || node->op == GGML_OP_TRANSPOSE || node->op == GGML_OP_VIEW || node->op == GGML_OP_PERMUTE;
+}
+
 GGML_CALL static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cgraph * cgraph) {
 #ifdef GGML_VULKAN_DEBUG
     std::cerr << "ggml_backend_vk_graph_compute(" << cgraph->n_nodes << " nodes)" << std::endl;
@@ -6220,7 +6180,7 @@ GGML_CALL static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backen
     int last_node = cgraph->n_nodes - 1;
 
     // If the last op in the cgraph isn't backend GPU, the command buffer doesn't get closed properly
-    while (last_node > 0 && (cgraph->nodes[last_node]->backend != GGML_BACKEND_TYPE_GPU || ggml_is_empty(cgraph->nodes[last_node]))) {
+    while (last_node > 0 && ggml_vk_is_empty(cgraph->nodes[last_node])) {
         last_node -= 1;
     }
 
@@ -6234,7 +6194,7 @@ GGML_CALL static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backen
     for (int i = 0; i < cgraph->n_nodes; i++) {
         ggml_tensor * node = cgraph->nodes[i];
 
-        if (ggml_is_empty(node) || node->op == GGML_OP_RESHAPE || node->op == GGML_OP_TRANSPOSE || node->op == GGML_OP_VIEW || node->op == GGML_OP_PERMUTE || node->op == GGML_OP_NONE) {
+        if (ggml_vk_is_empty(node)) {
             continue;
         }
 
@@ -6548,7 +6508,7 @@ static void ggml_vk_print_graph_origin(const ggml_tensor * tensor, std::vector<c
 }
 
 static void ggml_vk_print_tensor_area(const ggml_tensor * tensor, const void * data, int i0, int i1, int i2, int i3) {
-    if (tensor->type != GGML_TYPE_F32 && tensor->type != GGML_TYPE_F16) {
+    if (tensor->type != GGML_TYPE_F32 && tensor->type != GGML_TYPE_F16 && tensor->type != GGML_TYPE_I32) {
         return;
     }
     i0 = std::max(i0, 5);
@@ -6569,6 +6529,8 @@ static void ggml_vk_print_tensor_area(const ggml_tensor * tensor, const void * d
                     val = *(const float *) ((const char *) data + i3*tensor->nb[3] + i2*tensor->nb[2] + idx1*tensor->nb[1] + idx0*tensor->nb[0]);
                 } else if (tensor->type == GGML_TYPE_F16) {
                     val = ggml_fp16_to_fp32(*(const ggml_fp16_t *) ((const char *) data + i3*tensor->nb[3] + i2*tensor->nb[2] + idx1*tensor->nb[1] + idx0*tensor->nb[0]));
+                } else if (tensor->type == GGML_TYPE_I32) {
+                    val = *(const int32_t *) ((const char *) data + i3*tensor->nb[3] + i2*tensor->nb[2] + idx1*tensor->nb[1] + idx0*tensor->nb[0]);
                 } else {
                     GGML_ASSERT(false);
                 }
@@ -6671,7 +6633,6 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_compute_
 
     ggml_tensor * src0 = tensor->src[0];
     ggml_tensor * src1 = tensor->src[1];
-    ggml_tensor * src2 = tensor->src[2];
 
     struct ggml_init_params iparams = {
         /*.mem_size   =*/ 1024*1024*1024,
@@ -6798,66 +6759,6 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_compute_
 
         ggml_vk_check_tensor(std::string(ggml_op_name(tensor->op)) + "->src1", src1_clone);
     }
-    if (src2 != nullptr) {
-        src2_clone = ggml_dup_tensor(ggml_ctx, src2);
-
-        src2_size = ggml_nbytes(src2);
-
-        src2_buffer = malloc(src2_size);
-        src2_clone->data = src2_buffer;
-        if (src2->backend == GGML_BACKEND_TYPE_CPU) {
-            memcpy(src2_clone->data, src2->data, src2_size);
-            memcpy(src2_clone->nb, src2->nb, sizeof(size_t) * GGML_MAX_DIMS);
-        } else if (src2->backend == GGML_BACKEND_TYPE_GPU) {
-            ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) src2->extra;
-            vk_buffer buf = extra->buffer_gpu.lock();
-            uint64_t offset = extra->offset;
-            if (!ggml_is_contiguous(src2) && ggml_vk_dim01_contiguous(src2)) {
-                for (int i3 = 0; i3 < src2->ne[3]; i3++) {
-                    for (int i2 = 0; i2 < src2->ne[2]; i2++) {
-                        const int idx = i3*src2->ne[2] + i2;
-                        ggml_vk_buffer_read(ctx, buf, offset + idx * src2->nb[2], ((char *)src2_clone->data + idx * src2_clone->nb[2]), src2->ne[1] * src2->nb[1]);
-                    }
-                }
-
-                src2_clone->nb[0] = src2->nb[0];
-                src2_clone->nb[1] = src2->nb[1];
-                for (int i = 2; i < GGML_MAX_DIMS; i++) {
-                    src2_clone->nb[i] = src2_clone->nb[i - 1]*src2_clone->ne[i - 1];
-                }
-            } else {
-                if (offset + src2_size >= buf->size) {
-                    src2_size = buf->size - offset;
-                }
-                ggml_vk_buffer_read(ctx, buf, offset, src2_clone->data, src2_size);
-                memcpy(src2_clone->nb, src2->nb, sizeof(size_t) * GGML_MAX_DIMS);
-            }
-        } else {
-            GGML_ASSERT(false);
-        }
-
-        if (vk_output_tensor > 0 && vk_output_tensor == check_counter) {
-            ggml_vk_print_tensor(ctx, src2, "src2");
-            std::cerr << "TENSOR CHECK: " << ggml_op_name(src2_clone->op) << " (check " << check_counter << ")" << std::endl;
-            std::cerr << "src2_clone=" << tensor << " src2_clone->backend: " << src2_clone->backend << " src2_clone->type: " << ggml_type_name(src2_clone->type) << " ne0=" << src2_clone->ne[0] << " nb0=" << src2_clone->nb[0] << " ne1=" << src2_clone->ne[1] << " nb1=" << src2_clone->nb[1] << " ne2=" << src2_clone->ne[2] << " nb2=" << src2_clone->nb[2] << " ne3=" << src2_clone->ne[3] << " nb3=" << src2_clone->nb[3] << std::endl;
-            if (src2->src[0] != nullptr) {
-                std::cerr << "src2->src[0]=" << src2->src[0] << " op=" << ggml_op_name(src2->src[0]->op) << " type=" << ggml_type_name(src2->src[0]->type) << " backend=" << src2->src[0]->backend << " ne0=" << src2->src[0]->ne[0] << " nb0=" << src2->src[0]->nb[0] << " ne1=" << src2->src[0]->ne[1] << " nb1=" << src2->src[0]->nb[1] << " ne2=" << src2->src[0]->ne[2] << " nb2=" << src2->src[0]->nb[2] << " ne3=" << src2->src[0]->ne[3] << " nb3=" << src2->src[0]->nb[3] << std::endl;
-            }
-            if (src2->src[1] != nullptr) {
-                std::cerr << "src2->src[1]=" << src2->src[1] << " op=" << ggml_op_name(src2->src[1]->op) << " type=" << ggml_type_name(src2->src[1]->type) << " backend=" << src2->src[1]->backend << " ne0=" << src2->src[1]->ne[0] << " nb0=" << src2->src[1]->nb[0] << " ne1=" << src2->src[1]->ne[1] << " nb1=" << src2->src[1]->nb[1] << " ne2=" << src2->src[1]->ne[2] << " nb2=" << src2->src[1]->nb[2] << " ne3=" << src2->src[1]->ne[3] << " nb3=" << src2->src[1]->nb[3] << std::endl;
-            }
-            std::cerr << std::endl << "Result:" << std::endl;
-            ggml_vk_print_tensor_area(src2_clone, src2_clone->data, 5, 5, 0, 0);
-            std::cerr << std::endl;
-            std::cerr << std::endl << "Result:" << std::endl;
-            ggml_vk_print_tensor_area(src2_clone, src2_clone->data, 5, 5, 1, 0);
-            std::cerr << std::endl;
-            std::vector<const ggml_tensor *> done;
-            ggml_vk_print_graph_origin(src2_clone, done);
-        }
-
-        ggml_vk_check_tensor(std::string(ggml_op_name(tensor->op)) + "->src2", src2_clone);
-    }
 
     if (tensor->op == GGML_OP_MUL_MAT) {
         tensor_clone = ggml_mul_mat(ggml_ctx, src0_clone, src1_clone);
@@ -6877,7 +6778,7 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_compute_
         tensor_clone = ggml_rms_norm(ggml_ctx, src0_clone, *(float *)tensor->op_params);
     } else if (tensor->op == GGML_OP_SOFT_MAX) {
         if (src1 != nullptr) {
-            tensor_clone = ggml_soft_max_ext(ggml_ctx, src0_clone, src1_clone, src2_clone, ((float *)tensor->op_params)[0], ((float *)tensor->op_params)[1]);
+            tensor_clone = ggml_soft_max_ext(ggml_ctx, src0_clone, src1_clone, ((float *)tensor->op_params)[0], ((float *)tensor->op_params)[1]);
         } else {
             tensor_clone = ggml_soft_max(ggml_ctx, src0_clone);
         }
@@ -6937,16 +6838,11 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_compute_
         GGML_ASSERT(false);
     }
 
-    // Disable vulkan here to avoid the hooks in ggml.c
-    ctx->disable = true;
-
     ggml_cgraph * cgraph = ggml_new_graph(ggml_ctx);
     ggml_build_forward_expand(cgraph, tensor_clone);
 
     ggml_graph_compute_with_ctx(ggml_ctx, cgraph, 8);
 
-    ctx->disable = false;
-
     ggml_vk_check_tensor(ggml_op_name(tensor->op), tensor_clone);
     if (vk_output_tensor > 0 && vk_output_tensor == check_counter) {
         ggml_vk_print_tensor(ctx, tensor_clone, "tensor_clone");
@@ -6964,9 +6860,6 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_compute_
     if (src1 != nullptr) {
         free(src1_buffer);
     }
-    if (src2 != nullptr) {
-        free(src2_buffer);
-    }
 
     ggml_free(ggml_ctx);
 }
@@ -7026,8 +6919,11 @@ static void ggml_vk_check_results_1(ggml_backend_vk_context * ctx, ggml_compute_
                         } else if (tensor->type == GGML_TYPE_F16) {
                             correct = ggml_fp16_to_fp32(*(ggml_fp16_t *) ((char *) comp_result + i3*comp_nb[3] + i2*comp_nb[2] + i1*comp_nb[1] + i0*comp_nb[0]));
                             result  = ggml_fp16_to_fp32(*(ggml_fp16_t *) ((char *) tensor_data + i3*tensor->nb[3] + i2*tensor->nb[2] + i1*tensor->nb[1] + i0*tensor->nb[0]));
+                        } else if (tensor->type == GGML_TYPE_I32) {
+                            correct = *(int32_t *) ((char *) comp_result + i3*comp_nb[3] + i2*comp_nb[2] + i1*comp_nb[1] + i0*comp_nb[0]);
+                            result  = *(int32_t *) ((char *) tensor_data + i3*tensor->nb[3] + i2*tensor->nb[2] + i1*tensor->nb[1] + i0*tensor->nb[0]);
                         } else {
-                            std::cerr << "comp_size=" << comp_size << " but required is " << (i3*comp_nb[3] + i2*comp_nb[2] + i1*comp_nb[1] + i0*comp_nb[0]) << std::endl;
+                            std::cerr << "Results check not implemented for type " << ggml_type_name(tensor->type) << std::endl;
                         }
                     } else {
                         std::cerr << "Missing debug code for type " << ggml_type_name(tensor->type) << std::endl;

ggml.c

@@ -15882,9 +15882,10 @@ static void ggml_compute_forward_flash_attn_ext_f16(
     GGML_ASSERT(ne0 == D);
     GGML_ASSERT(ne2 == N);
 
-    GGML_ASSERT(nbq0 == sizeof(float));
-    GGML_ASSERT(nbk0 == sizeof(ggml_fp16_t));
-    GGML_ASSERT(nbv0 == sizeof(ggml_fp16_t));
+    // input tensor rows must be contiguous
+    GGML_ASSERT(nbq0 == ggml_type_size(q->type));
+    GGML_ASSERT(nbk0 == ggml_type_size(k->type));
+    GGML_ASSERT(nbv0 == ggml_type_size(v->type));
 
     GGML_ASSERT(neq0 == D);
     GGML_ASSERT(nek0 == D);
@@ -15938,6 +15939,11 @@ static void ggml_compute_forward_flash_attn_ext_f16(
     const float m0 = powf(2.0f, -(max_bias       ) / n_head_log2);
     const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2);
 
+    enum ggml_type    const k_vec_dot_type = type_traits[k->type].vec_dot_type;
+    ggml_from_float_t const q_to_vec_dot   = type_traits[k_vec_dot_type].from_float;
+    ggml_vec_dot_t    const kq_vec_dot     = type_traits[k->type].vec_dot;
+    ggml_to_float_t   const v_to_float     = type_traits[v->type].to_float;
+
     // loop over n_batch and n_head
     for (int ir = ir0; ir < ir1; ++ir) {
         // q indices
@@ -15945,17 +15951,22 @@ static void ggml_compute_forward_flash_attn_ext_f16(
         const int iq2 = (ir - iq3*neq2*neq1)/neq1;
         const int iq1 = (ir - iq3*neq2*neq1 - iq2*neq1);
 
-        const uint32_t h = iq2; // head
+        const uint32_t h = iq2; // head index
         const float slope = (max_bias > 0.0f) ? h < n_head_log2 ? powf(m0, h + 1) : powf(m1, 2*(h - n_head_log2) + 1) : 1.0f;
 
-        float S = 0.0f;
-        float M = -INFINITY;
+        float S = 0.0f;      // sum
+        float M = -INFINITY; // maximum KQ value
 
-        float       * V32 = (float       *) params->wdata + ith*(2*D + CACHE_LINE_SIZE_F32);
-        ggml_fp16_t * Q16 = (ggml_fp16_t *) (V32); // reuse memory
-        ggml_fp16_t * V16 = (ggml_fp16_t *) (V32 + D);
+        float       * VKQ32 = (float       *) params->wdata + ith*(3*D + CACHE_LINE_SIZE_F32); // FP32 VKQ accumulator
+        float       * V32   =                 (VKQ32 + 1*D); // (temporary) FP32 V buffer
+        ggml_fp16_t * VKQ16 = (ggml_fp16_t *) (VKQ32 + 1*D); // (temporary) FP16 VKQ accumulator
+        ggml_fp16_t * Q_q   = (ggml_fp16_t *) (VKQ32 + 2*D); // (temporary) buffer for Q converted to quantized/FP16
 
-        memset(V16, 0, D*sizeof(ggml_fp16_t));
+        if (v->type == GGML_TYPE_F16) {
+            memset(VKQ16, 0, D*sizeof(ggml_fp16_t));
+        } else {
+            memset(VKQ32, 0, D*sizeof(float));
+        }
 
         const ggml_fp16_t * mp = mask ? (ggml_fp16_t *)((char *) mask->data + iq1*mask->nb[1]) : NULL;
 
@@ -15967,6 +15978,9 @@ static void ggml_compute_forward_flash_attn_ext_f16(
         const int iv3 = iq3 / rv3;
         const int iv2 = iq2 / rv2;
 
+        const float * pq = (const float *) ((char *) q->data + (iq1*nbq1 + iq2*nbq2 + iq3*nbq3));
+        q_to_vec_dot(pq, Q_q, D);
+
         // online softmax / attention
         // loop over n_kv and n_head_kv
         // ref: https://arxiv.org/pdf/2112.05682.pdf
@@ -15976,51 +15990,66 @@ static void ggml_compute_forward_flash_attn_ext_f16(
                 continue;
             }
 
-            float s;
+            float s; // KQ value
 
-            // convert Q to F16 in V32
-            {
-                const float * pq = (const float *) ((char *) q->data + (iq1*nbq1 + iq2*nbq2 + iq3*nbq3));
+            const char * k_data = (const char *) k->data + ( ic*nbk1 + ik2*nbk2 + ik3*nbk3);
+            kq_vec_dot(D, &s, 0, k_data, 0, Q_q, 0, 1);
 
-                for (int64_t d = 0; d < D; ++d) {
-                    Q16[d] = GGML_FP32_TO_FP16(pq[d]);
-                }
-            }
-
-            ggml_vec_dot_f16(D,
-                    &s, 0,
-                    (ggml_fp16_t *) ((char *) k->data + ( ic*nbk1 + ik2*nbk2 + ik3*nbk3)), 0,
-                    Q16, 0, 1);
-
-            s = s*scale + mv;
+            s = s*scale + mv; // scale KQ value and apply mask
 
             const float Mold = M;
 
-            float ms = 1.0f;
-            float vs = 1.0f;
+            float ms = 1.0f; // upon new higher max val, scale VKQ and KQ sum with this value
+            float vs = 1.0f; // post-softmax KQ value, expf(s - M)
 
-            if (s > M) {
-                M = s;
-                ms = expf(Mold - M);
+            const char * v_data = ((const char *) v->data + (ic*nbv1 + iv2*nbv2 + iv3*nbv3));
 
-                // V = V*expf(Mold - M)
-                ggml_vec_scale_f16(D, V16, ms);
+            if (v->type== GGML_TYPE_F16) {
+                if (s > M) {
+                    // s is new maximum, ms < 1.0f, vs == expf(s - s) == 1.0f
+                    M = s;
+                    ms = expf(Mold - M);
+
+                    // V = V*expf(Mold - M)
+                    ggml_vec_scale_f16(D, VKQ16, ms);
+                } else {
+                    // no new maximum, ms == 1.0f, vs != 1.0f
+                    vs = expf(s - M);
+                }
+
+                // V += v*expf(s - M)
+                ggml_vec_mad_f16(D, VKQ16, (const ggml_fp16_t *) v_data, vs);
             } else {
-                vs = expf(s - M);
+                if (s > M) {
+                    // s is new maximum, ms < 1.0f, vs == expf(s - s) == 1.0f
+                    M = s;
+                    ms = expf(Mold - M);
+
+                    // V = V*expf(Mold - M)
+                    ggml_vec_scale_f32(D, VKQ32, ms);
+                } else {
+                    // no new maximum, ms == 1.0f, vs != 1.0f
+                    vs = expf(s - M);
+                }
+
+                v_to_float(v_data, V32, D);
+
+                // V += v*expf(s - M)
+                ggml_vec_mad_f32(D, VKQ32, V32, vs);
             }
 
-            const ggml_fp16_t * v16 = (const ggml_fp16_t *) ((char *) v->data + (ic*nbv1 + iv2*nbv2 + iv3*nbv3));
+            S = S*ms + vs; // scale and increment sum with partial sum
+        }
 
-            // V += v*expf(s - M)
-            ggml_vec_mad_f16(D, V16, v16, vs);
-
-            S = S*ms + vs;
+        if (v->type == GGML_TYPE_F16) {
+            for (int64_t d = 0; d < D; ++d) {
+                VKQ32[d] = GGML_FP16_TO_FP32(VKQ16[d]);
+            }
         }
 
         // V /= S
-        for (int64_t d = 0; d < D; ++d) {
-            V32[d] = GGML_FP16_TO_FP32(V16[d])/S;
-        }
+        const float S_inv = 1.0f/S;
+        ggml_vec_scale_f32(D, VKQ32, S_inv);
 
         // dst indices
         const int i1 = iq1;
@@ -16031,7 +16060,7 @@ static void ggml_compute_forward_flash_attn_ext_f16(
         //memcpy((char *) dst->data + (i1*nb1 + i2*nb2 + i3*nb3), V, nev0*sizeof(float));
 
         // permute(0, 2, 1, 3)
-        memcpy((char *) dst->data + (i3*ne2*ne1 + i2 + i1*ne1)*nb1, V32, nb1);
+        memcpy((char *) dst->data + (i3*ne2*ne1 + i2 + i1*ne1)*nb1, VKQ32, nb1);
     }
 }
 
@@ -19972,7 +20001,7 @@ struct ggml_cplan ggml_graph_plan(const struct ggml_cgraph * cgraph, int n_threa
                 {
                     const int64_t ne00 = node->src[0]->ne[0]; // D
 
-                    cur = 2*sizeof(float)*ne00*n_tasks; // 2x head size
+                    cur = 3*sizeof(float)*ne00*n_tasks; // 3x head size/thread
                 } break;
             case GGML_OP_FLASH_FF:
                 {

ggml_vk_generate_shaders.py

@@ -2432,7 +2432,6 @@ layout (push_constant) uniform parameter
 {
     uint KX;
     uint KY;
-    uint KZ;
     float scale;
     float max_bias;
     float m0;
@@ -2449,8 +2448,7 @@ layout(local_size_x = BLOCK_SIZE, local_size_y = 1, local_size_z = 1) in;
 
 layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
 layout (binding = 1) readonly buffer Y {B_TYPE data_b[];};
-layout (binding = 2) readonly buffer Z {C_TYPE data_c[];};
-layout (binding = 3) buffer D {D_TYPE data_d[];};
+layout (binding = 2) buffer D {D_TYPE data_d[];};
 
 shared FLOAT_TYPE vals[BLOCK_SIZE];
 
@@ -2459,7 +2457,7 @@ void main() {
     const uint rowx = gl_WorkGroupID.x;
     const uint rowy = rowx % p.KY;
 
-    float slope = 0.0f;
+    float slope = 1.0f;
 
     // ALiBi
     if (p.max_bias > 0.0f) {
@@ -2472,11 +2470,18 @@ void main() {
     }
 
     // Find max
-    vals[tid] = uintBitsToFloat(0xFF800000);
+    FLOAT_TYPE max_val = uintBitsToFloat(0xFF800000);
 
-    [[unroll]] for (uint col = tid; col < p.KX; col += BLOCK_SIZE) {
-        vals[tid] = max(vals[tid], FLOAT_TYPE(data_a[rowx * p.KX + col]) * p.scale + (p.KY > 0 ? FLOAT_TYPE(data_b[rowy * p.KX + col]) : FLOAT_TYPE(0.0f)) + (p.KZ > 0 ? slope * FLOAT_TYPE(data_c[col]) : 0.0f));
+    [[unroll]] for (uint col0 = 0; col0 < p.KX; col0 += BLOCK_SIZE) {
+        const uint col = col0 + tid;
+
+        if (col >= p.KX) {
+            break;
+        }
+
+        max_val = max(max_val, FLOAT_TYPE(data_a[rowx * p.KX + col]) * p.scale + (p.KY > 0 ? slope * FLOAT_TYPE(data_b[rowy * p.KX + col]) : FLOAT_TYPE(0.0f)));
     }
+    vals[tid] = max_val;
 
     barrier();
     [[unroll]] for (int s = BLOCK_SIZE / 2; s > 0; s >>= 1) {
@@ -2486,15 +2491,21 @@ void main() {
         barrier();
     }
 
-    const FLOAT_TYPE max_val = vals[0];
+    max_val = vals[0];
     barrier();
 
     // Sum up values
     vals[tid] = FLOAT_TYPE(0.0f);
 
-    [[unroll]] for (uint col = tid; col < p.KX; col += BLOCK_SIZE) {
+    [[unroll]] for (uint col0 = 0; col0 < p.KX; col0 += BLOCK_SIZE) {
+        const uint col = col0 + tid;
+
+        if (col >= p.KX) {
+            break;
+        }
+
         const uint i = rowx * p.KX + col;
-        const FLOAT_TYPE val = exp(FLOAT_TYPE(data_a[i]) * p.scale + (p.KY > 0 ? FLOAT_TYPE(data_b[rowy * p.KX + col]) : FLOAT_TYPE(0.0f)) - max_val);
+        const FLOAT_TYPE val = exp(FLOAT_TYPE(data_a[i]) * p.scale + (p.KY > 0 ? slope * FLOAT_TYPE(data_b[rowy * p.KX + col]) : FLOAT_TYPE(0.0f)) - max_val);
         vals[tid] += val;
         data_d[i] = D_TYPE(val);
     }
@@ -2509,7 +2520,13 @@ void main() {
 
     const D_TYPE divisor = D_TYPE(vals[0]);
 
-    [[unroll]] for (uint col = tid; col < p.KX; col += BLOCK_SIZE) {
+    [[unroll]] for (uint col0 = 0; col0 < p.KX; col0 += BLOCK_SIZE) {
+        const uint col = col0 + tid;
+
+        if (col >= p.KX) {
+            break;
+        }
+
         data_d[rowx*p.KX + col] /= divisor;
     }
 }
@@ -2672,20 +2689,26 @@ argsort_src = """
 
 #extension GL_EXT_shader_16bit_storage : require
 
-layout(local_size_x = 1024, local_size_y = 1, local_size_z = 1) in;
+#define BLOCK_SIZE 1024
+#define ASC 0
+
+layout(local_size_x = BLOCK_SIZE, local_size_y = 1, local_size_z = 1) in;
 
 layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
 layout (binding = 1)          buffer D {int data_d[];};
 
 layout (push_constant) uniform parameter {
     uint ncols;
-    bool ascending;
+    uint ncols_pad;
+    uint order;
 } p;
 
+shared int dst_row[BLOCK_SIZE];
+
 void swap(uint idx0, uint idx1) {
-    int tmp = data_d[idx0];
-    data_d[idx0] = data_d[idx1];
-    data_d[idx1] = tmp;
+    int tmp = dst_row[idx0];
+    dst_row[idx0] = dst_row[idx1];
+    dst_row[idx1] = tmp;
 }
 
 void main() {
@@ -2693,36 +2716,45 @@ void main() {
     const int col = int(gl_LocalInvocationID.x);
     const uint row = gl_WorkGroupID.y;
 
-    if (col >= p.ncols) {
+    if (col >= p.ncols_pad) {
         return;
     }
 
-    const uint a_idx = row * p.ncols;
-    const uint d_idx = row * p.ncols;
+    const uint row_offset = row * p.ncols;
 
     // initialize indices
-    if (col < p.ncols) {
-        data_d[col] = col;
-    }
+    dst_row[col] = col;
     barrier();
 
-    for (uint k = 2; k <= p.ncols; k *= 2) {
+    for (uint k = 2; k <= p.ncols_pad; k *= 2) {
         for (uint j = k / 2; j > 0; j /= 2) {
             const uint ixj = col ^ j;
             if (ixj > col) {
                 if ((col & k) == 0) {
-                    if (p.ascending ? data_a[a_idx + data_d[d_idx + col]] > data_a[a_idx + data_d[d_idx + ixj]] : data_a[a_idx + data_d[d_idx + col]] < data_a[a_idx + data_d[d_idx + ixj]]) {
-                        swap(d_idx + col, d_idx + ixj);
+                    if (dst_row[col] >= p.ncols ||
+                        (dst_row[ixj] < p.ncols && (p.order == ASC ?
+                            data_a[row_offset + dst_row[col]] > data_a[row_offset + dst_row[ixj]] :
+                            data_a[row_offset + dst_row[col]] < data_a[row_offset + dst_row[ixj]]))
+                    ) {
+                        swap(col, ixj);
                     }
                 } else {
-                    if (p.ascending ? data_a[a_idx + data_d[d_idx + col]] < data_a[a_idx + data_d[d_idx + ixj]] : data_a[a_idx + data_d[d_idx + col]] > data_a[a_idx + data_d[d_idx + ixj]]) {
-                        swap(d_idx + col, d_idx + ixj);
+                    if (dst_row[ixj] >= p.ncols ||
+                        (dst_row[col] < p.ncols && (p.order == ASC ?
+                            data_a[row_offset + dst_row[col]] < data_a[row_offset + dst_row[ixj]] :
+                            data_a[row_offset + dst_row[col]] > data_a[row_offset + dst_row[ixj]]))
+                    ) {
+                        swap(col, ixj);
                     }
                 }
             }
             barrier();
         }
     }
+
+    if (col < p.ncols) {
+        data_d[row_offset + col] = dst_row[col];
+    }
 }
 """
 
@@ -2845,13 +2877,16 @@ async def main():
         stream.clear()
         stream.extend((mulmat_head, shader_float_type, mulmat_body1, mulmat_load_scalar, mulmat_body2))
         tasks.append(string_to_spv("matmul_f32", "".join(stream), {"A_TYPE": "float", "B_TYPE": "float", "D_TYPE": "float"}, fp16))
-        tasks.append(string_to_spv("matmul_f32_aligned", "".join(stream), {"LOAD_VEC_A": 1, "LOAD_VEC_B": load_vec, "A_TYPE": "float", "B_TYPE": vec_type, "D_TYPE": "float"}, fp16))
+        tasks.append(string_to_spv("matmul_f32_aligned", "".join(stream), {"LOAD_VEC_A": load_vec, "LOAD_VEC_B": load_vec, "A_TYPE": vec_type, "B_TYPE": vec_type, "D_TYPE": "float"}, fp16))
+
+        tasks.append(string_to_spv("matmul_f32_f16", "".join(stream), {"A_TYPE": "float", "B_TYPE": "float16_t", "D_TYPE": "float"}, fp16))
+        tasks.append(string_to_spv("matmul_f32_f16_aligned", "".join(stream), {"LOAD_VEC_A": load_vec, "LOAD_VEC_B": load_vec, "A_TYPE": vec_type, "B_TYPE": vec_type_f16, "D_TYPE": "float"}, fp16))
 
         tasks.append(string_to_spv("matmul_f16", "".join(stream), {"A_TYPE": "float16_t", "B_TYPE": "float16_t", "D_TYPE": "float"}, fp16))
-        tasks.append(string_to_spv("matmul_f16_aligned", "".join(stream), {"LOAD_VEC_A": 1, "LOAD_VEC_B": load_vec, "A_TYPE": "float16_t", "B_TYPE": vec_type_f16, "D_TYPE": "float"}, fp16))
+        tasks.append(string_to_spv("matmul_f16_aligned", "".join(stream), {"LOAD_VEC_A": load_vec, "LOAD_VEC_B": load_vec, "A_TYPE": vec_type_f16, "B_TYPE": vec_type_f16, "D_TYPE": "float"}, fp16))
 
         tasks.append(string_to_spv("matmul_f16_f32", "".join(stream), {"A_TYPE": "float16_t", "B_TYPE": "float", "D_TYPE": "float"}, fp16))
-        tasks.append(string_to_spv("matmul_f16_f32_aligned", "".join(stream), {"LOAD_VEC_A": 1, "LOAD_VEC_B": load_vec, "A_TYPE": "float16_t", "B_TYPE": vec_type, "D_TYPE": "float"}, fp16))
+        tasks.append(string_to_spv("matmul_f16_f32_aligned", "".join(stream), {"LOAD_VEC_A": load_vec, "LOAD_VEC_B": load_vec, "A_TYPE": vec_type_f16, "B_TYPE": vec_type, "D_TYPE": "float"}, fp16))
 
         stream.clear()
         stream.extend((mulmat_head, shader_int8_ext, shader_float_type, shader_q4_0_defines, mulmat_body1, mulmat_load_q4_0, mulmat_body2))

llama.cpp

@@ -26,9 +26,6 @@
 #ifdef GGML_USE_METAL
 #  include "ggml-metal.h"
 #endif
-#ifdef GGML_USE_MPI
-#  include "ggml-mpi.h"
-#endif
 #ifndef QK_K
 #  ifdef GGML_QKK_64
 #    define QK_K 64
@@ -1697,6 +1694,8 @@ struct llama_state {
     llama_state() {
 #ifdef GGML_USE_METAL
         ggml_backend_metal_log_set_callback(log_callback, log_callback_user_data);
+#elif defined(GGML_USE_CUDA)
+        ggml_backend_cuda_log_set_callback(log_callback, log_callback_user_data);
 #endif
     }
 
@@ -2268,10 +2267,6 @@ struct llama_context {
 
     // control vectors
     struct llama_control_vector cvec;
-
-#ifdef GGML_USE_MPI
-    ggml_mpi_context * ctx_mpi = NULL;
-#endif
 };
 
 static ggml_backend_buffer_type_t llama_default_buffer_type_offload(const llama_model & model, int gpu) {
@@ -4461,6 +4456,9 @@ static void llm_load_vocab(
             } else if (
                 tokenizer_pre == "qwen2") {
                 vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_QWEN2;
+            } else if (
+                tokenizer_pre == "stablelm2") {
+                vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_STABLELM2;
             } else if (
                 tokenizer_pre == "olmo") {
                 vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_OLMO;
@@ -5188,7 +5186,14 @@ static bool llm_load_tensors(
                     {
                         model.output_norm   = ml.create_tensor(ctx_output,       tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd});
                         model.output_norm_b = ml.create_tensor(ctx_output,       tn(LLM_TENSOR_OUTPUT_NORM, "bias"),   {n_embd});
-                        model.output        = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab});
+                        model.output        = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, false);
+                        if (!model.output) {
+                            // needs to be on GPU
+                            model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});
+                            ml.n_created--; // artificial tensor
+                            ml.size_data += ggml_nbytes(model.output);
+                        }
+
                     }
 
                     for (int i = 0; i < n_layer; ++i) {
@@ -6324,10 +6329,7 @@ static struct ggml_tensor * llm_build_inp_embd(
 
         inpL = ggml_get_rows(ctx, tok_embd, lctx.inp_tokens);
     } else {
-#ifdef GGML_USE_MPI
-        GGML_ASSERT(false && "not implemented");
-#endif
-        lctx.inp_embd = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, batch.n_tokens);
+       lctx.inp_embd = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, batch.n_tokens);
         inpL = lctx.inp_embd;
         ggml_set_input(lctx.inp_embd);
     }
@@ -11339,11 +11341,6 @@ static void llama_graph_compute(
         llama_context & lctx,
           ggml_cgraph * gf,
                   int   n_threads) {
-#ifdef GGML_USE_MPI
-    const int64_t n_layer = lctx.model.hparams.n_layer;
-    ggml_mpi_graph_compute_pre(lctx.ctx_mpi, gf, n_layer);
-#endif
-
 #ifdef GGML_USE_METAL
     if (ggml_backend_is_metal(lctx.backend_metal)) {
         ggml_backend_metal_set_n_cb(lctx.backend_metal, n_threads);
@@ -11358,10 +11355,6 @@ static void llama_graph_compute(
     ggml_backend_sched_graph_compute_async(lctx.sched, gf);
 
     // fprintf(stderr, "splits: %d\n", ggml_backend_sched_get_n_splits(lctx.sched));
-
-#ifdef GGML_USE_MPI
-    ggml_mpi_graph_compute_post(lctx.ctx_mpi, gf, n_layer);
-#endif
 }
 
 // decode a batch of tokens by evaluating the transformer
@@ -11399,12 +11392,6 @@ static int llama_decode_internal(
     }
     lctx.n_queued_tokens += n_tokens_all;
 
-#ifdef GGML_USE_MPI
-    // TODO: needs fix after #3228
-    GGML_ASSERT(false && "not implemented");
-    //ggml_mpi_eval_init(lctx.ctx_mpi, &n_tokens, &n_past, &n_threads);
-#endif
-
     auto & kv_self = lctx.kv_self;
 
     const int64_t n_embd  = hparams.n_embd;
@@ -12354,6 +12341,7 @@ struct llm_tokenizer_bpe {
                             "'s|'t|'re|'ve|'m|'ll|'d| ?\\p{L}+| ?\\p{N}+| ?[^\\s\\p{L}\\p{N}]+|\\s+(?!\\S)",
                         });
                         break;
+                    case LLAMA_VOCAB_PRE_TYPE_STABLELM2:
                     case LLAMA_VOCAB_PRE_TYPE_QWEN2:
                         word_collection = unicode_regex_split(text, {
                             // original regex from tokenizer.json
@@ -12576,16 +12564,16 @@ struct llm_tokenizer_wpm {
         // to lowercase, pad chinese characters, pad punctuation
         std::string new_str = "";
         for (uint32_t code : cpts_nfd) {
-            int type = unicode_cpt_type(code);
-            if (type == CODEPOINT_TYPE_ACCENT_MARK || type == CODEPOINT_TYPE_CONTROL) {
+            const codepoint_flags flags = unicode_cpt_flags(code);
+            if (flags.is_accent_mark || flags.is_control) {
                 continue;
             }
             code = unicode_tolower(code);
-            if (type == CODEPOINT_TYPE_SEPARATOR) {
+            if (flags.is_separator || flags.is_whitespace) {  //####FIXME: is_separator ?
                 code = ' ';
             }
             std::string s = unicode_cpt_to_utf8(code);
-            if (type == CODEPOINT_TYPE_PUNCTUATION || is_ascii_punct(code) || is_chinese_char(code)) {
+            if (flags.is_punctuation || is_ascii_punct(code) || is_chinese_char(code)) {
                 new_str += " ";
                 new_str += s;
                 new_str += " ";
@@ -15533,10 +15521,6 @@ void llama_backend_init(void) {
         struct ggml_context * ctx = ggml_init(params);
         ggml_free(ctx);
     }
-
-#ifdef GGML_USE_MPI
-    ggml_mpi_backend_init();
-#endif
 }
 
 void llama_numa_init(enum ggml_numa_strategy numa) {
@@ -15546,9 +15530,6 @@ void llama_numa_init(enum ggml_numa_strategy numa) {
 }
 
 void llama_backend_free(void) {
-#ifdef GGML_USE_MPI
-    ggml_mpi_backend_free();
-#endif
     ggml_quantize_free();
 }
 
@@ -15949,20 +15930,6 @@ struct llama_context * llama_new_context_with_model(
         }
     }
 
-#ifdef GGML_USE_MPI
-    ctx->ctx_mpi = ggml_mpi_init();
-
-    if (ggml_mpi_rank(ctx->ctx_mpi) > 0) {
-        // Enter a blocking eval loop with dummy input, letting rank=0 drive the process
-        // TODO: needs fix after #3228
-        GGML_ASSERT(false && "not implemented");
-        //const std::vector<llama_token> tmp(ctx->model.hparams.n_ctx, llama_token_bos(ctx));
-        //while (!llama_eval(ctx, tmp.data(), tmp.size(), 0, 0)) {};
-        llama_backend_free();
-        exit(1);
-    }
-#endif
-
     return ctx;
 }
 
@@ -18167,6 +18134,8 @@ void llama_log_set(ggml_log_callback log_callback, void * user_data) {
     g_state.log_callback_user_data = user_data;
 #ifdef GGML_USE_METAL
     ggml_backend_metal_log_set_callback(g_state.log_callback, g_state.log_callback_user_data);
+#elif defined(GGML_USE_CUDA)
+    ggml_backend_cuda_log_set_callback(g_state.log_callback, g_state.log_callback_user_data);
 #endif
 }
 

llama.h

@@ -81,9 +81,10 @@ extern "C" {
         LLAMA_VOCAB_PRE_TYPE_GPT2           = 7,
         LLAMA_VOCAB_PRE_TYPE_REFACT         = 8,
         LLAMA_VOCAB_PRE_TYPE_COMMAND_R      = 9,
-        LLAMA_VOCAB_PRE_TYPE_QWEN2          = 10,
-        LLAMA_VOCAB_PRE_TYPE_OLMO           = 11,
-        LLAMA_VOCAB_PRE_TYPE_DBRX           = 12,
+        LLAMA_VOCAB_PRE_TYPE_STABLELM2      = 10,
+        LLAMA_VOCAB_PRE_TYPE_QWEN2          = 11,
+        LLAMA_VOCAB_PRE_TYPE_OLMO           = 12,
+        LLAMA_VOCAB_PRE_TYPE_DBRX           = 13,
     };
 
     // note: these values should be synchronized with ggml_rope

scripts/LlamaConfig.cmake.in

@@ -5,7 +5,6 @@ set(LLAMA_SHARED_LIB @BUILD_SHARED_LIBS@)
 set(LLAMA_BLAS @LLAMA_BLAS@)
 set(LLAMA_CUDA @LLAMA_CUDA@)
 set(LLAMA_METAL @LLAMA_METAL@)
-set(LLAMA_MPI @LLAMA_MPI@)
 set(LLAMA_CLBLAST @LLAMA_CLBLAST@)
 set(LLAMA_HIPBLAS @LLAMA_HIPBLAS@)
 set(LLAMA_ACCELERATE @LLAMA_ACCELERATE@)
@@ -37,10 +36,6 @@ if (LLAMA_METAL)
     find_library(METALKIT_FRAMEWORK MetalKit REQUIRED)
 endif()
 
-if (LLAMA_MPI)
-    find_package(MPI REQUIRED)
-endif()
-
 if (LLAMA_CLBLAST)
     find_package(CLBlast REQUIRED)
 endif()

scripts/gen-unicode-data.py

@@ -1,64 +1,134 @@
 import regex
+import ctypes
+import unicodedata
 
 
-def get_matches(regex_expr):
-    regex_expr_compiled = regex.compile(regex_expr)
-    unicode_ranges = []
-    current_range = None
-
-    for codepoint in range(0x110000):
-        char = chr(codepoint)
-        if regex_expr_compiled.match(char):
-            if current_range is None:
-                current_range = [codepoint, codepoint]
-            else:
-                current_range[1] = codepoint
-        elif current_range is not None:
-            unicode_ranges.append(tuple(current_range))
-            current_range = None
-
-    if current_range is not None:
-        unicode_ranges.append(tuple(current_range))
-
-    return unicode_ranges
+class CoodepointFlags (ctypes.Structure):
+    _fields_ = [  # see definition in unicode.h
+        ("is_undefined",   ctypes.c_uint16, 1),
+        ("is_number",      ctypes.c_uint16, 1),  # regex: \p{N}
+        ("is_letter",      ctypes.c_uint16, 1),  # regex: \p{L}
+        ("is_separator",   ctypes.c_uint16, 1),  # regex: \p{Z}
+        ("is_accent_mark", ctypes.c_uint16, 1),  # regex: \p{M}
+        ("is_punctuation", ctypes.c_uint16, 1),  # regex: \p{P}
+        ("is_symbol",      ctypes.c_uint16, 1),  # regex: \p{S}
+        ("is_control",     ctypes.c_uint16, 1),  # regex: \p{C}
+    ]
 
 
-def print_cat(mode, cat, ranges):
-    if mode == "range":
-        print("const std::vector<std::pair<uint32_t, uint32_t>> unicode_ranges_{} = {{".format(cat)) # noqa: NP100
-    if mode == "map":
-        print("const std::map<uint32_t, uint32_t> unicode_map_{} = {{".format(cat)) # noqa: NP100
-    for i, values in enumerate(ranges):
-        end = ",\n" if (i % 4 == 3 or i + 1 == len(ranges)) else ", "
-        values = ["0x%08X" % value for value in values]
-        print("{" + ", ".join(values) + "}", end=end) # noqa: NP100
-    print("};") # noqa: NP100
-    print("") # noqa: NP100
+assert (ctypes.sizeof(CoodepointFlags) == 2)
 
 
-print_cat("range", "number",      get_matches(r'\p{N}'))
-print_cat("range", "letter",      get_matches(r'\p{L}'))
-print_cat("range", "separator",   get_matches(r'\p{Z}'))
-print_cat("range", "accent_mark", get_matches(r'\p{M}'))
-print_cat("range", "punctuation", get_matches(r'\p{P}'))
-print_cat("range", "symbol",      get_matches(r'\p{S}'))
-print_cat("range", "control",     get_matches(r'\p{C}'))
+MAX_CODEPOINTS = 0x110000
 
-print_cat("range", "whitespace",  get_matches(r'\s'))
+regex_number      = regex.compile(r'\p{N}')
+regex_letter      = regex.compile(r'\p{L}')
+regex_separator   = regex.compile(r'\p{Z}')
+regex_accent_mark = regex.compile(r'\p{M}')
+regex_punctuation = regex.compile(r'\p{P}')
+regex_symbol      = regex.compile(r'\p{S}')
+regex_control     = regex.compile(r'\p{C}')
+regex_whitespace  = regex.compile(r'\s')
 
+codepoint_flags = (CoodepointFlags * MAX_CODEPOINTS)()
+table_whitespace = []
+table_lowercase = []
+table_uppercase = []
+table_nfd = []
 
-map_lowercase = []
-map_uppercase = []
-for codepoint in range(0x110000):
+for codepoint in range(MAX_CODEPOINTS):
+    # convert codepoint to unicode character
     char = chr(codepoint)
+
+    # regex categories
+    flags = codepoint_flags[codepoint]
+    flags.is_number      = bool(regex_number.match(char))
+    flags.is_letter      = bool(regex_letter.match(char))
+    flags.is_separator   = bool(regex_separator.match(char))
+    flags.is_accent_mark = bool(regex_accent_mark.match(char))
+    flags.is_punctuation = bool(regex_punctuation.match(char))
+    flags.is_symbol      = bool(regex_symbol.match(char))
+    flags.is_control     = bool(regex_control.match(char))
+    flags.is_undefined   = bytes(flags)[0] == 0
+    assert (not flags.is_undefined)
+
+    # whitespaces
+    if bool(regex_whitespace.match(char)):
+        table_whitespace.append(codepoint)
+
+    # lowercase conversion
     lower = ord(char.lower()[0])
-    upper = ord(char.upper()[0])
     if codepoint != lower:
-        map_lowercase.append((codepoint, lower))
+        table_lowercase.append((codepoint, lower))
+
+    # uppercase conversion
+    upper = ord(char.upper()[0])
     if codepoint != upper:
-        map_uppercase.append((codepoint, upper))
-print_cat("map", "lowercase", map_lowercase)
-print_cat("map", "uppercase", map_uppercase)
+        table_uppercase.append((codepoint, upper))
+
+    # NFD normalization
+    norm = ord(unicodedata.normalize('NFD', char)[0])
+    if codepoint != norm:
+        table_nfd.append((codepoint, norm))
 
 
-# TODO: generate unicode_map_nfd
+# group ranges with same flags
+ranges_flags = [(0, codepoint_flags[0])]  # start, flags
+for codepoint, flags in enumerate(codepoint_flags):
+    if bytes(flags) != bytes(ranges_flags[-1][1]):
+        ranges_flags.append((codepoint, flags))
+ranges_flags.append((MAX_CODEPOINTS, CoodepointFlags()))
+
+
+# group ranges with same nfd
+ranges_nfd = [(0, 0, 0)]  # start, last, nfd
+for codepoint, norm in table_nfd:
+    start = ranges_nfd[-1][0]
+    if ranges_nfd[-1] != (start, codepoint - 1, norm):
+        ranges_nfd.append(None)
+        start = codepoint
+    ranges_nfd[-1] = (start, codepoint, norm)
+
+
+# Generate 'unicode-data.cpp'
+
+
+def out(line=""):
+    print(line, end='\n')  # noqa
+
+
+out("""\
+// generated with scripts/gen-unicode-data.py
+
+#include "unicode-data.h"
+
+#include <cstdint>
+#include <vector>
+#include <unordered_map>
+#include <unordered_set>
+""")
+
+out("const std::vector<std::pair<uint32_t, uint16_t>> unicode_ranges_flags = {  // start, flags // last=next_start-1")
+for codepoint, flags in ranges_flags:
+    flags = int.from_bytes(bytes(flags), "little")
+    out("{0x%06X, 0x%04X}," % (codepoint, flags))
+out("};\n")
+
+out("const std::unordered_set<uint32_t> unicode_set_whitespace = {")
+out(", ".join("0x%06X" % cpt for cpt in table_whitespace))
+out("};\n")
+
+out("const std::unordered_map<uint32_t, uint32_t> unicode_map_lowercase = {")
+for tuple in table_lowercase:
+    out("{0x%06X, 0x%06X}," % tuple)
+out("};\n")
+
+out("const std::unordered_map<uint32_t, uint32_t> unicode_map_uppercase = {")
+for tuple in table_uppercase:
+    out("{0x%06X, 0x%06X}," % tuple)
+out("};\n")
+
+out("const std::vector<range_nfd> unicode_ranges_nfd = {  // start, last, nfd")
+for triple in ranges_nfd:
+    out("{0x%06X, 0x%06X, 0x%06X}," % triple)
+out("};\n")

tests/test-backend-ops.cpp

@@ -16,6 +16,7 @@
 #include <thread>
 #include <vector>
 
+
 static void init_tensor_uniform(ggml_tensor * tensor, float min = -1.0f, float max = 1.0f) {
     // static RNG initialization (revisit if n_threads stops being constant)
     static const size_t n_threads = std::thread::hardware_concurrency();
@@ -49,6 +50,22 @@ static void init_tensor_uniform(ggml_tensor * tensor, float min = -1.0f, float m
         t.join();
     }
 
+#if 0
+    const char * val_str = getenv("GGML_TEST_EPS");
+    float val = 1e-9f;
+    if (val_str != nullptr) {
+        val = std::stof(val_str);
+        printf("GGML_TEST_EPS=%e\n", val);
+    }
+
+    // test quantization with very small values that may result in nan scales due to division by zero
+    if (ggml_is_quantized(tensor->type)) {
+        for (int i = 0; i < 256; i++) {
+            data[i] = val;
+        }
+    }
+#endif
+
     if (tensor->type == GGML_TYPE_F32 || tensor->type == GGML_TYPE_I32) {
         ggml_backend_tensor_set(tensor, data.data(), 0, size * sizeof(float));
     } else if (ggml_is_quantized(tensor->type) || tensor->type == GGML_TYPE_F16 || tensor->type == GGML_TYPE_BF16) {
@@ -64,6 +81,7 @@ static void init_tensor_uniform(ggml_tensor * tensor, float min = -1.0f, float m
             }
         }
         ggml_quantize_chunk(tensor->type, data.data(), dataq.data(), 0, size/tensor->ne[0], tensor->ne[0], im);
+        GGML_ASSERT(ggml_validate_row_data(tensor->type, dataq.data(), dataq.size()));
         ggml_backend_tensor_set(tensor, dataq.data(), 0, dataq.size());
     } else if (tensor->type == GGML_TYPE_I8 || tensor->type == GGML_TYPE_I16 || tensor->type == GGML_TYPE_I32) {
         // This is going to create some weird integers though.

tests/test-tokenizer-random.py

@@ -1,5 +1,5 @@
 # Test libllama tokenizer == AutoTokenizer.
-# Brute force random tokens/text generation.
+# Brute force random words/text generation.
 #
 # Sample usage:
 #
@@ -12,10 +12,10 @@ import argparse
 import subprocess
 import random
 
-from typing import Iterator
+from typing import Callable, Iterator
 
 import cffi
-from transformers import AutoTokenizer, PreTrainedTokenizerBase
+from transformers import AutoTokenizer
 
 logger = logging.getLogger("test-tokenizer-random-bpe")
 
@@ -145,28 +145,35 @@ def generator_custom_text() -> Iterator[str]:
 def generator_custom_text_edge_cases() -> Iterator[str]:
     """Edge cases found while debugging"""
     yield from [
-        '\x1f-a',   # unicode_ranges_control, {0x00001C, 0x00001F}
-        '¼-a',      # unicode_ranges_digit, 0x00BC
-        '½-a',      # unicode_ranges_digit, 0x00BD
-        '¾-a',      # unicode_ranges_digit, 0x00BE
-        'a 〇b',    # unicode_ranges_digit, 0x3007
-        'Ⅵ-a',     # unicode_ranges_digit, {0x00002150, 0x0000218F} // Number Forms
-        '\uFEFF//', # unicode_ranges_control, 0xFEFF (BOM)
-        '<s>a'      # TODO: Phi-3 fail
+        '\x1f-a',     # unicode_ranges_control, {0x00001C, 0x00001F}
+        '¼-a',        # unicode_ranges_digit, 0x00BC
+        '½-a',        # unicode_ranges_digit, 0x00BD
+        '¾-a',        # unicode_ranges_digit, 0x00BE
+        'a 〇b',      # unicode_ranges_digit, 0x3007
+        'Ⅵ-a',       # unicode_ranges_digit, {0x00002150, 0x0000218F} // Number Forms
+        '\uFEFF//',   # unicode_ranges_control, 0xFEFF (BOM)
+        'Cửa Việt',   # llama-3, ignore_merges = true
+        '<s>a',       # TODO: Phi-3 fail
+        'a\na',       # TODO: Bert fail
     ]
 
 
-def generator_random_chars(iterations = 100) -> Iterator[str]:
+def generator_vocab_words(vocab: list[str]) -> Iterator[str]:
+    """Brute force check all vocab words"""
+    yield from vocab
+
+
+def generator_random_chars(iterations=100) -> Iterator[str]:
     """Brute force random text with simple characters"""
 
     WHITESPACES = list(" " * 20 + "\n" * 5 + "\r\n" * 5 + "\t" * 5)
-    CHARS = list(set("""
+    CHARS = list(sorted(set("""
         ABCDEFGHIJKLMNOPQRSTUVWXYZ
         abcdefghijklmnopqrstuvwxyz
         ÁÉÍÓÚÀÈÌÒÙÂÊÎÔÛÄËÏÖÜ
         áéíóúàèìòùâêîôûäëïöü
         .-,*/-+ª!"·$%&/()=?¿[]{}<>\\|@#~½¬~;:_
-    """))
+    """)))
 
     rand = random.Random()
     for m in range(iterations):
@@ -181,13 +188,13 @@ def generator_random_chars(iterations = 100) -> Iterator[str]:
         yield "".join(text)
 
 
-def generator_random_vocab_chars(tokenizer: PreTrainedTokenizerBase, iterations = 100) -> Iterator[str]:
+def generator_random_vocab_chars(vocab: list[str], iterations=100) -> Iterator[str]:
     """Brute force random text with vocab characters"""
 
-    vocab_ids = list(tokenizer.vocab.values())
-    vocab_text = tokenizer.decode(vocab_ids, skip_special_tokens=True)
-    vocab_chars = list(set(vocab_text))
-    del vocab_ids, vocab_text
+    vocab_chars = set()
+    for word in vocab:
+        vocab_chars.update(word)
+    vocab_chars = list(sorted(vocab_chars))
 
     rand = random.Random()
     for m in range(iterations):
@@ -196,19 +203,11 @@ def generator_random_vocab_chars(tokenizer: PreTrainedTokenizerBase, iterations
         yield "".join(text)
 
 
-def generator_random_vocab_tokens(tokenizer: PreTrainedTokenizerBase, iterations = 100) -> Iterator[str]:
-    """Brute force random text from vocab tokens"""
+def generator_random_vocab_words(vocab: list[str], iterations=100) -> Iterator[str]:
+    """Brute force random text from vocab words"""
 
-    space_id = tokenizer.encode(" ", add_special_tokens=False)[0]
-    vocab_ids = list(tokenizer.vocab.values())
-    vocab_ids = list(sorted(vocab_ids + vocab_ids))
-    for i in range(1, len(vocab_ids), 2):
-        vocab_ids[i] = space_id
-    vocab_tokens = tokenizer.decode(vocab_ids, skip_special_tokens=True)
-    vocab_tokens = vocab_tokens.split(" ")
-    del vocab_ids
-
-    yield from vocab_tokens
+    vocab = [w.strip() for w in vocab]
+    yield from vocab
 
     rand = random.Random()
     for m in range(iterations):
@@ -217,14 +216,13 @@ def generator_random_vocab_tokens(tokenizer: PreTrainedTokenizerBase, iterations
         num_words = rand.randint(300, 400)
         for i in range(num_words):
             k = rand.randint(1, 3)
-            tokens = rand.choices(vocab_tokens, k=k)
-            tokens = [t.strip(" \n\r\t") for t in tokens]
+            words = rand.choices(vocab, k=k)
             sep = rand.choice("     \n\r\t")
-            text.append("".join(tokens) + sep)
+            text.append("".join(words) + sep)
         yield "".join(text)
 
 
-def generator_random_bytes(iterations = 100) -> Iterator[str]:
+def generator_random_bytes(iterations=100) -> Iterator[str]:
     """Brute force random bytes"""
 
     WHITESPACES = list(" " * 20 + "\n" * 5 + "\r\n" * 5 + "\t" * 5)
@@ -242,10 +240,10 @@ def generator_random_bytes(iterations = 100) -> Iterator[str]:
         yield "".join(text)
 
 
-def test_compare_tokenizer(model: LibLlamaModel, tokenizer: PreTrainedTokenizerBase, generator: Iterator[str]):
+def test_compare_tokenizer(func_tokenize1: Callable, func_tokenize2: Callable, generator: Iterator[str]):
 
     def find_first_mismatch(ids1: list[int], ids2: list[int]):
-        for i, (a,b) in enumerate(zip(ids1, ids2)):
+        for i, (a, b) in enumerate(zip(ids1, ids2)):
             if a != b:
                 return i
         if len(ids1) == len(ids2):
@@ -255,15 +253,12 @@ def test_compare_tokenizer(model: LibLlamaModel, tokenizer: PreTrainedTokenizerB
     t0 = time.perf_counter()
     logger.info("%s: %s" % (generator.__name__, "ini"))
     for text in generator:
-        ids1 = model.tokenize(text, add_special=False, parse_special=False)
-        ids2 = tokenizer.encode(text, add_special_tokens=False)
+        ids1 = func_tokenize1(text)
+        ids2 = func_tokenize2(text)
         if ids1 != ids2:
             i = find_first_mismatch(ids1, ids2)
             ids1 = list(ids1)[max(0, i - 2) : i + 2 + 1]
             ids2 = list(ids2)[max(0, i - 2) : i + 2 + 1]
-            text2 = tokenizer.decode(ids2, skip_special_tokens=True)
-            assert (text2 in text)
-            logger.info(" Text:     " + repr(text2))
             logger.info(" TokenIDs: " + str(ids1))
             logger.info(" Expected: " + str(ids2))
             raise Exception()
@@ -271,25 +266,37 @@ def test_compare_tokenizer(model: LibLlamaModel, tokenizer: PreTrainedTokenizerB
     logger.info("%s: end, time: %.3f secs" % (generator.__name__, t1 - t0))
 
 
-if __name__ == "__main__":
-
+def main(argv: list[str] = None):
     parser = argparse.ArgumentParser()
     parser.add_argument("vocab_file", help="path to vocab 'gguf' file")
     parser.add_argument("dir_tokenizer", help="directory containing 'tokenizer.model' file")
     parser.add_argument("--verbose", action="store_true", help="increase output verbosity")
-    args = parser.parse_args()
+    args = parser.parse_args(argv)
 
     logging.basicConfig(level=logging.DEBUG if args.verbose else logging.INFO)
 
-    model = LibLlamaModel(LibLlama(), args.vocab_file, mparams=dict(vocab_only=True), cparams=dict(n_ctx=2048))
-
+    model = LibLlamaModel(LibLlama(), args.vocab_file, mparams=dict(vocab_only=True), cparams=dict(n_ctx=4096))
     tokenizer = AutoTokenizer.from_pretrained(args.dir_tokenizer)
 
-    test_compare_tokenizer(model, tokenizer, generator_custom_text())
-    test_compare_tokenizer(model, tokenizer, generator_custom_text_edge_cases())
-    test_compare_tokenizer(model, tokenizer, generator_random_chars(10_000))
-    test_compare_tokenizer(model, tokenizer, generator_random_vocab_chars(tokenizer, 10_000))
-    test_compare_tokenizer(model, tokenizer, generator_random_vocab_tokens(tokenizer, 10_000))
-    # test_compare_tokenizer(model, tokenizer, generator_random_bytes(10_000)) # FAIL
+    def func_tokenize2(text: str):
+        return tokenizer.encode(text, add_special_tokens=False)
+
+    parse_special = all(len(func_tokenize2(t)) == 1 for t in tokenizer.all_special_tokens)
+
+    def func_tokenize1(text: str):
+        return model.tokenize(text, add_special=False, parse_special=parse_special)
+
+    vocab = list(sorted(tokenizer.batch_decode(list(tokenizer.get_vocab().values()), skip_special_tokens=True)))
+    test_compare_tokenizer(func_tokenize1, func_tokenize2, generator_custom_text())
+    test_compare_tokenizer(func_tokenize1, func_tokenize2, generator_custom_text_edge_cases())
+    test_compare_tokenizer(func_tokenize1, func_tokenize2, generator_vocab_words(vocab))
+    test_compare_tokenizer(func_tokenize1, func_tokenize2, generator_random_chars(10_000))
+    test_compare_tokenizer(func_tokenize1, func_tokenize2, generator_random_vocab_chars(vocab, 10_000))
+    test_compare_tokenizer(func_tokenize1, func_tokenize2, generator_random_vocab_words(vocab, 10_000))
+    # test_compare_tokenizer(func_tokenize1, func_tokenize2, generator_random_bytes(10_000)) # FAIL
 
     model.free()
+
+
+if __name__ == "__main__":
+    main()

unicode-data.h

@@ -1,17 +1,20 @@
 #pragma once
 
 #include <cstdint>
-#include <map>
-#include <utility>
 #include <vector>
+#include <unordered_map>
+#include <unordered_set>
 
-extern const std::vector<std::pair<uint32_t, uint32_t>> unicode_ranges_number;
-extern const std::vector<std::pair<uint32_t, uint32_t>> unicode_ranges_letter;
-extern const std::vector<std::pair<uint32_t, uint32_t>> unicode_ranges_separator;
-extern const std::vector<std::pair<uint32_t, uint32_t>> unicode_ranges_whitespace;
-extern const std::vector<std::pair<uint32_t, uint32_t>> unicode_ranges_accent_mark;
-extern const std::vector<std::pair<uint32_t, uint32_t>> unicode_ranges_punctuation;
-extern const std::vector<std::pair<uint32_t, uint32_t>> unicode_ranges_symbol;
-extern const std::vector<std::pair<uint32_t, uint32_t>> unicode_ranges_control;
-extern const std::multimap<uint32_t, uint32_t>          unicode_map_nfd;
-extern const std::map<char32_t, char32_t>               unicode_map_lowercase;
+struct range_nfd {
+    uint32_t first;
+    uint32_t last;
+    uint32_t nfd;
+};
+
+static const uint32_t MAX_CODEPOINTS = 0x110000;
+
+extern const std::vector<std::pair<uint32_t, uint16_t>> unicode_ranges_flags;
+extern const std::unordered_set<uint32_t> unicode_set_whitespace;
+extern const std::unordered_map<uint32_t, uint32_t> unicode_map_lowercase;
+extern const std::unordered_map<uint32_t, uint32_t> unicode_map_uppercase;
+extern const std::vector<range_nfd> unicode_ranges_nfd;

unicode.cpp

@@ -1,4 +1,4 @@
-#include "unicode.h"
+#include "unicode.h"
 #include "unicode-data.h"
 
 #include <cassert>
@@ -109,57 +109,49 @@ static uint32_t unicode_cpt_from_utf8(const std::string & utf8, size_t & offset)
 //    return result;
 //}
 
-static std::unordered_map<uint32_t, int> unicode_cpt_type_map() {
-    std::unordered_map<uint32_t, int> cpt_types;
-    for (auto p : unicode_ranges_number) {
-        for (auto i = p.first; i <= p.second; ++i) {
-            cpt_types[i] = CODEPOINT_TYPE_NUMBER;
+static std::vector<codepoint_flags> unicode_cpt_flags_array() {
+    std::vector<codepoint_flags> cpt_flags(MAX_CODEPOINTS, codepoint_flags::UNDEFINED);
+
+    assert (unicode_ranges_flags.front().first == 0);
+    assert (unicode_ranges_flags.back().first == MAX_CODEPOINTS);
+    for (size_t i = 1; i < unicode_ranges_flags.size(); ++i) {
+        const auto range_ini = unicode_ranges_flags[i-1];  // codepoint_ini, flags
+        const auto range_end = unicode_ranges_flags[i];    // codepoint_end, flags
+        for (uint32_t cpt = range_ini.first; cpt < range_end.first; ++cpt) {
+            cpt_flags[cpt] = range_ini.second;
         }
     }
-    for (auto p : unicode_ranges_letter) {
-        for (auto i = p.first; i <= p.second; ++i) {
-            cpt_types[i] = CODEPOINT_TYPE_LETTER;
-        }
+
+    for (auto cpt : unicode_set_whitespace) {
+        cpt_flags[cpt].is_whitespace = true;
     }
-    for (auto p : unicode_ranges_separator) {
-        for (auto i = p.first; i <= p.second; ++i) {
-            cpt_types[i] = CODEPOINT_TYPE_SEPARATOR;
-        }
+
+    for (auto p : unicode_map_lowercase) {
+        cpt_flags[p.second].is_lowercase = true;
     }
-    for (auto p : unicode_ranges_accent_mark) {
-        for (auto i = p.first; i <= p.second; ++i) {
-            cpt_types[i] = CODEPOINT_TYPE_ACCENT_MARK;
-        }
+
+    for (auto p : unicode_map_uppercase) {
+        cpt_flags[p.second].is_uppercase = true;
     }
-    for (auto p : unicode_ranges_punctuation) {
-        for (auto i = p.first; i <= p.second; ++i) {
-            cpt_types[i] = CODEPOINT_TYPE_PUNCTUATION;
-        }
+
+    for (auto &range : unicode_ranges_nfd) {  // start, last, nfd
+        cpt_flags[range.nfd].is_nfd = true;
     }
-    for  (auto p : unicode_ranges_symbol) {
-        for (auto i = p.first; i <= p.second; ++i) {
-            cpt_types[i] = CODEPOINT_TYPE_SYMBOL;
-        }
-    }
-    for (auto p : unicode_ranges_control) {
-        for (auto i = p.first; i <= p.second; ++i) {
-            cpt_types[i] = CODEPOINT_TYPE_CONTROL;
-        }
-    }
-    return cpt_types;
+
+    return cpt_flags;
 }
 
 static std::unordered_map<uint8_t, std::string> unicode_byte_to_utf8_map() {
     std::unordered_map<uint8_t, std::string> map;
-    for (int ch = u'!'; ch <= u'~'; ++ch) {
+    for (int ch = 0x21; ch <= 0x7E; ++ch) {  // u'!' to u'~'
         assert(0 <= ch && ch < 256);
         map[ch] = unicode_cpt_to_utf8(ch);
     }
-    for (int ch = u'¡'; ch <= u'¬'; ++ch) {
+    for (int ch = 0xA1; ch <= 0xAC; ++ch) {  // u'¡' to u'¬'
         assert(0 <= ch && ch < 256);
         map[ch] = unicode_cpt_to_utf8(ch);
     }
-    for (int ch = u'®'; ch <= u'ÿ'; ++ch) {
+    for (int ch = 0xAE; ch <= 0xFF; ++ch) {  // u'®' to u'ÿ'
         assert(0 <= ch && ch < 256);
         map[ch] = unicode_cpt_to_utf8(ch);
     }
@@ -175,15 +167,15 @@ static std::unordered_map<uint8_t, std::string> unicode_byte_to_utf8_map() {
 
 static std::unordered_map<std::string, uint8_t> unicode_utf8_to_byte_map() {
     std::unordered_map<std::string, uint8_t> map;
-    for (int ch = u'!'; ch <= u'~'; ++ch) {
+    for (int ch = 0x21; ch <= 0x7E; ++ch) {  // u'!' to u'~'
         assert(0 <= ch && ch < 256);
         map[unicode_cpt_to_utf8(ch)] = ch;
     }
-    for (int ch = u'¡'; ch <= u'¬'; ++ch) {
+    for (int ch = 0xA1; ch <= 0xAC; ++ch) {  // u'¡' to u'¬'
         assert(0 <= ch && ch < 256);
         map[unicode_cpt_to_utf8(ch)] = ch;
     }
-    for (int ch = u'®'; ch <= u'ÿ'; ++ch) {
+    for (int ch = 0xAE; ch <= 0xFF; ++ch) {  // u'®' to u'ÿ'
         assert(0 <= ch && ch < 256);
         map[unicode_cpt_to_utf8(ch)] = ch;
     }
@@ -238,8 +230,9 @@ static std::vector<size_t> unicode_regex_split_custom_gpt2(const std::string & t
             return (offset_ini <= pos && pos < offset_end) ? cpts[pos] : 0;
         };
 
-        auto _get_cpt_type = [&] (const size_t pos) -> int {
-            return (offset_ini <= pos && pos < offset_end) ? unicode_cpt_type(cpts[pos]) : CODEPOINT_TYPE_UNIDENTIFIED;
+        auto _get_flags = [&] (const size_t pos) -> codepoint_flags {
+            static const codepoint_flags undef(codepoint_flags::UNDEFINED);
+            return (offset_ini <= pos && pos < offset_end) ? unicode_cpt_flags(cpts[pos]) : undef;
         };
 
         size_t _prev_end = offset_ini;
@@ -261,7 +254,7 @@ static std::vector<size_t> unicode_regex_split_custom_gpt2(const std::string & t
 
         for (size_t pos = offset_ini; pos < offset_end; /*pos++*/ ) {
             const char32_t cpt = _get_cpt(pos);
-            const int cpt_type = _get_cpt_type(pos);
+            const auto flags = _get_flags(pos);
 
             // regex: 's|'t|'re|'ve|'m|'ll|'d
             if (cpt == '\'' && pos+1 < offset_end) {
@@ -281,39 +274,37 @@ static std::vector<size_t> unicode_regex_split_custom_gpt2(const std::string & t
                 }
             }
 
-            char32_t cpt2 = (cpt == ' ' ? _get_cpt(pos+1) : cpt);
-            int cpt2_type = (cpt == ' ' ? _get_cpt_type(pos+1) : cpt_type);
+            auto flags2 = (cpt == ' ' ? _get_flags(pos+1) : flags);
             // regex: <space>?\p{L}+
-            if (cpt2_type == CODEPOINT_TYPE_LETTER) {
+            if (flags2.is_letter) {
                 pos += (cpt == ' ');
-                while (cpt2_type == CODEPOINT_TYPE_LETTER) {
-                    cpt2_type = _get_cpt_type(++pos);
+                while (flags2.is_letter) {
+                    flags2 = _get_flags(++pos);
                 }
                 _add_token(pos);
                 continue;
             }
             // regex: <space>?\p{N}+
-            if (cpt2_type == CODEPOINT_TYPE_NUMBER) {
+            if (flags2.is_number) {
                 pos += (cpt == ' ');
-                while (cpt2_type == CODEPOINT_TYPE_NUMBER) {
-                    cpt2_type = _get_cpt_type(++pos);
+                while (flags2.is_number) {
+                    flags2 = _get_flags(++pos);
                 }
                 _add_token(pos);
                 continue;
             }
             // regex: <space>?[^\s\p{L}\p{N}]+
-            if (!unicode_cpt_is_whitespace(cpt2) && cpt2_type != CODEPOINT_TYPE_LETTER && cpt2_type != CODEPOINT_TYPE_NUMBER && cpt2_type != CODEPOINT_TYPE_UNIDENTIFIED) {
+            if (!(flags2.is_whitespace || flags2.is_letter || flags2.is_number || flags2.is_undefined)) {
                 pos += (cpt == ' ');
-                while (!unicode_cpt_is_whitespace(cpt2) && cpt2_type != CODEPOINT_TYPE_LETTER && cpt2_type != CODEPOINT_TYPE_NUMBER && cpt2_type != CODEPOINT_TYPE_UNIDENTIFIED) {
-                    cpt2_type = _get_cpt_type(++pos);
-                    cpt2 = _get_cpt(pos);
+                while (!(flags2.is_whitespace || flags2.is_letter || flags2.is_number || flags2.is_undefined)) {
+                    flags2 = _get_flags(++pos);
                 }
                 _add_token(pos);
                 continue;
             }
 
             size_t num_whitespaces = 0;
-            while (unicode_cpt_is_whitespace(_get_cpt(pos+num_whitespaces))) {
+            while (_get_flags(pos+num_whitespaces).is_whitespace) {
                 num_whitespaces++;
             }
 
@@ -357,8 +348,9 @@ static std::vector<size_t> unicode_regex_split_custom_llama3(const std::string &
             return (offset_ini <= pos && pos < offset_end) ? cpts[pos] : 0;
         };
 
-        auto _get_cpt_type = [&] (const size_t pos) -> int {
-            return (offset_ini <= pos && pos < offset_end) ? unicode_cpt_type(cpts[pos]) : CODEPOINT_TYPE_UNIDENTIFIED;
+        auto _get_flags = [&] (const size_t pos) -> codepoint_flags {
+            static const codepoint_flags undef(codepoint_flags::UNDEFINED);
+            return (offset_ini <= pos && pos < offset_end) ? unicode_cpt_flags(cpts[pos]) : undef;
         };
 
         size_t _prev_end = offset_ini;
@@ -380,7 +372,7 @@ static std::vector<size_t> unicode_regex_split_custom_llama3(const std::string &
 
         for (size_t pos = offset_ini; pos < offset_end; /*pos++*/ ) {
             const char32_t cpt = _get_cpt(pos);
-            const int cpt_type = _get_cpt_type(pos);
+            const auto flags = _get_flags(pos);
 
             // regex: (?i:'s|'t|'re|'ve|'m|'ll|'d) // case insensitive
             if (cpt == '\'' && pos+1 < offset_end) {
@@ -401,10 +393,10 @@ static std::vector<size_t> unicode_regex_split_custom_llama3(const std::string &
             }
 
             // regex: [^\r\n\p{L}\p{N}]?\p{L}+  //####FIXME: the first \p{L} is correct?
-            if (cpt != '\r' && cpt != '\n' && /*cpt_type != CODEPOINT_TYPE_LETTER &&*/ cpt_type != CODEPOINT_TYPE_NUMBER) {
-                if (cpt_type == CODEPOINT_TYPE_LETTER || _get_cpt_type(pos+1) == CODEPOINT_TYPE_LETTER) {  // one or more letters
+            if (!(cpt == '\r' || cpt == '\n' || /*flags.is_letter |*/ flags.is_number)) {
+                if (flags.is_letter || _get_flags(pos+1).is_letter) {  // one or more letters
                     pos++;
-                    while (_get_cpt_type(pos) == CODEPOINT_TYPE_LETTER) {
+                    while (_get_flags(pos).is_letter) {
                         pos++;
                     }
                     _add_token(pos);
@@ -413,9 +405,9 @@ static std::vector<size_t> unicode_regex_split_custom_llama3(const std::string &
             }
 
             // regex: \p{N}{1,3}
-            if (cpt_type == CODEPOINT_TYPE_NUMBER) {
+            if (flags.is_number) {
                 size_t ini = pos;
-                while (_get_cpt_type(pos) == CODEPOINT_TYPE_NUMBER) {
+                while (_get_flags(pos).is_number) {
                     if (++pos - ini >= 3 ) {
                         _add_token(pos);
                         ini = pos;
@@ -426,14 +418,13 @@ static std::vector<size_t> unicode_regex_split_custom_llama3(const std::string &
             }
 
             // regex: <space>?[^\s\p{L}\p{N}]+[\r\n]*
-            char32_t cpt2 = (cpt == ' ' ? _get_cpt(pos+1) : cpt);
-            int cpt2_type = (cpt == ' ' ? _get_cpt_type(pos+1) : cpt_type);
-            if (!unicode_cpt_is_whitespace(cpt2) && cpt2_type != CODEPOINT_TYPE_LETTER && cpt2_type != CODEPOINT_TYPE_NUMBER && cpt2_type != CODEPOINT_TYPE_UNIDENTIFIED) {
+            auto flags2 = (cpt == ' ' ? _get_flags(pos+1) : flags);
+            if (!(flags2.is_whitespace || flags2.is_letter || flags2.is_number || flags2.is_undefined)) {
                 pos += (cpt == ' ');
-                while (!unicode_cpt_is_whitespace(cpt2) && cpt2_type != CODEPOINT_TYPE_LETTER && cpt2_type != CODEPOINT_TYPE_NUMBER && cpt2_type != CODEPOINT_TYPE_UNIDENTIFIED) {
-                    cpt2_type = _get_cpt_type(++pos);
-                    cpt2 = _get_cpt(pos);
+                while (!(flags2.is_whitespace || flags2.is_letter || flags2.is_number || flags2.is_undefined)) {
+                    flags2 = _get_flags(++pos);
                 }
+                char32_t cpt2 = _get_cpt(pos);
                 while (cpt2 == '\r' || cpt2 == '\n') {
                     cpt2 = _get_cpt(++pos);
                 }
@@ -443,7 +434,7 @@ static std::vector<size_t> unicode_regex_split_custom_llama3(const std::string &
 
             size_t num_whitespaces = 0;
             size_t last_end_r_or_n = 0;
-            while (unicode_cpt_is_whitespace(_get_cpt(pos+num_whitespaces))) {
+            while (_get_flags(pos+num_whitespaces).is_whitespace) {
                 char32_t cpt2 = _get_cpt(pos+num_whitespaces);
                 if (cpt2 == '\r' || cpt2 == '\n') {
                     last_end_r_or_n = pos + num_whitespaces + 1;
@@ -589,15 +580,14 @@ std::string unicode_cpt_to_utf8(uint32_t cp) {
 }
 
 std::vector<uint32_t> unicode_cpts_normalize_nfd(const std::vector<uint32_t> & cpts) {
-    std::vector<uint32_t> result;
-    result.reserve(cpts.size());
+    auto comp = [] (const uint32_t cpt, const range_nfd & range) {
+        return cpt < range.first;
+    };
+    std::vector<uint32_t> result(cpts.size());
     for (size_t i = 0; i < cpts.size(); ++i) {
-        auto it = unicode_map_nfd.find(cpts[i]);
-        if (it == unicode_map_nfd.end()) {
-            result.push_back(cpts[i]);
-        } else {
-            result.push_back(it->second);
-        }
+        const uint32_t cpt = cpts[i];
+        auto it = std::upper_bound(unicode_ranges_nfd.cbegin(), unicode_ranges_nfd.cend(), cpt, comp) - 1;
+        result[i] = (it->first <= cpt && cpt <= it->last) ? it->nfd : cpt;
     }
     return result;
 }
@@ -611,31 +601,19 @@ std::vector<uint32_t> unicode_cpts_from_utf8(const std::string & utf8) {
     return result;
 }
 
-int unicode_cpt_type(uint32_t cp) {
-    static std::unordered_map<uint32_t, int> cpt_types = unicode_cpt_type_map();
-    const auto it = cpt_types.find(cp);
-    return it == cpt_types.end() ? CODEPOINT_TYPE_UNIDENTIFIED : it->second;
+codepoint_flags unicode_cpt_flags(const uint32_t cp) {
+    static const codepoint_flags undef(codepoint_flags::UNDEFINED);
+    static const auto cpt_flags = unicode_cpt_flags_array();
+    return cp < cpt_flags.size() ? cpt_flags[cp] : undef;
 }
 
-int unicode_cpt_type(const std::string & utf8) {
-    if (utf8.length() == 0) {
-        return CODEPOINT_TYPE_UNIDENTIFIED;
+codepoint_flags unicode_cpt_flags(const std::string & utf8) {
+    static const codepoint_flags undef(codepoint_flags::UNDEFINED);
+    if (utf8.empty()) {
+        return undef;  // undefined
     }
     size_t offset = 0;
-    return unicode_cpt_type(unicode_cpt_from_utf8(utf8, offset));
-}
-
-bool unicode_cpt_is_whitespace(uint32_t cp) {
-    static const std::unordered_set<uint32_t> is_whitespace = [] {
-        std::unordered_set<uint32_t> is_whitespace;
-        for (auto p : unicode_ranges_whitespace) {
-            for (auto i = p.first; i <= p.second; ++i) {
-                is_whitespace.insert(i);
-            }
-        }
-        return is_whitespace;
-    }();
-    return (bool)is_whitespace.count(cp);
+    return unicode_cpt_flags(unicode_cpt_from_utf8(utf8, offset));
 }
 
 std::string unicode_byte_to_utf8(uint8_t byte) {
@@ -656,21 +634,21 @@ char32_t unicode_tolower(char32_t cp) {
 std::vector<std::string> unicode_regex_split(const std::string & text, const std::vector<std::string> & regex_exprs) {
     // unicode categories
     static const std::map<std::string, int> k_ucat_enum = {
-        { "\\p{N}", CODEPOINT_TYPE_NUMBER },
-        { "\\p{L}", CODEPOINT_TYPE_LETTER },
-        { "\\p{P}", CODEPOINT_TYPE_PUNCTUATION },
+        { "\\p{N}", codepoint_flags::NUMBER },
+        { "\\p{L}", codepoint_flags::LETTER },
+        { "\\p{P}", codepoint_flags::PUNCTUATION },
     };
 
     static const std::map<int, int> k_ucat_cpt = {
-        { CODEPOINT_TYPE_NUMBER,        0xD1 },
-        { CODEPOINT_TYPE_LETTER,        0xD2 },
-        { CODEPOINT_TYPE_PUNCTUATION,   0xD3 },
+        { codepoint_flags::NUMBER,        0xD1 },
+        { codepoint_flags::LETTER,        0xD2 },
+        { codepoint_flags::PUNCTUATION,   0xD3 },
     };
 
     static const std::map<int, std::string> k_ucat_map = {
-        { CODEPOINT_TYPE_NUMBER,        "\x30-\x39" }, // 0-9
-        { CODEPOINT_TYPE_LETTER,        "\x41-\x5A\x61-\x7A" }, // A-Za-z
-        { CODEPOINT_TYPE_PUNCTUATION,   "\x21-\x23\x25-\x2A\x2C-\x2F\x3A-\x3B\x3F-\x40\\\x5B-\\\x5D\x5F\\\x7B\\\x7D" }, // !-#%-*,-/:-;?-@\[-\]_\{\}
+        { codepoint_flags::NUMBER,        "\x30-\x39" }, // 0-9
+        { codepoint_flags::LETTER,        "\x41-\x5A\x61-\x7A" }, // A-Za-z
+        { codepoint_flags::PUNCTUATION,   "\x21-\x23\x25-\x2A\x2C-\x2F\x3A-\x3B\x3F-\x40\\\x5B-\\\x5D\x5F\\\x7B\\\x7D" }, // !-#%-*,-/:-;?-@\[-\]_\{\}
     };
 
     // compute collapsed codepoints only if needed by at least one regex
@@ -701,10 +679,10 @@ std::vector<std::string> unicode_regex_split(const std::string & text, const std
                 continue;
             }
 
-            const int cpt_type = unicode_cpt_type(cpts[i]);
+            const int cpt_flag = unicode_cpt_flags(cpts[i]).category_flag();
 
-            if (k_ucat_cpt.find(cpt_type) != k_ucat_cpt.end()) {
-                text_collapsed[i] = k_ucat_cpt.at(cpt_type);
+            if (k_ucat_cpt.find(cpt_flag) != k_ucat_cpt.end()) {
+                text_collapsed[i] = k_ucat_cpt.at(cpt_flag);
             } else {
                 text_collapsed[i] = (char) 0xD0; // fallback
             }

unicode.h

@@ -4,24 +4,56 @@
 #include <string>
 #include <vector>
 
-#define CODEPOINT_TYPE_UNIDENTIFIED 0
-#define CODEPOINT_TYPE_NUMBER       1
-#define CODEPOINT_TYPE_LETTER       2
-#define CODEPOINT_TYPE_SEPARATOR    3
-#define CODEPOINT_TYPE_ACCENT_MARK  4
-#define CODEPOINT_TYPE_PUNCTUATION  5
-#define CODEPOINT_TYPE_SYMBOL       6
-#define CODEPOINT_TYPE_CONTROL      7
+struct codepoint_flags {
+    enum {
+        UNDEFINED       = 0x0001,
+        NUMBER          = 0x0002,  // regex: \p{N}
+        LETTER          = 0x0004,  // regex: \p{L}
+        SEPARATOR       = 0x0008,  // regex: \p{Z}
+        ACCENT_MARK     = 0x0010,  // regex: \p{M}
+        PUNCTUATION     = 0x0020,  // regex: \p{P}
+        SYMBOL          = 0x0040,  // regex: \p{S}
+        CONTROL         = 0x0080,  // regex: \p{C}
+        MASK_CATEGORIES = 0x00FF,
+    };
+
+    // codepoint type
+    uint16_t is_undefined   : 1;
+    uint16_t is_number      : 1;  // regex: \p{N}
+    uint16_t is_letter      : 1;  // regex: \p{L}
+    uint16_t is_separator   : 1;  // regex: \p{Z}
+    uint16_t is_accent_mark : 1;  // regex: \p{M}
+    uint16_t is_punctuation : 1;  // regex: \p{P}
+    uint16_t is_symbol      : 1;  // regex: \p{S}
+    uint16_t is_control     : 1;  // regex: \p{C}
+    // helper flags
+    uint16_t is_whitespace  : 1;  // regex: \s
+    uint16_t is_lowercase   : 1;
+    uint16_t is_uppercase   : 1;
+    uint16_t is_nfd         : 1;
+
+    // decode from uint16
+    inline codepoint_flags(const uint16_t flags=0) {
+        *reinterpret_cast<uint16_t*>(this) = flags;
+    }
+
+    inline uint16_t as_uint() const {
+        return *reinterpret_cast<const uint16_t*>(this);
+    }
+
+    inline uint16_t category_flag() const {
+        return this->as_uint() & MASK_CATEGORIES;
+    }
+};
+
 
 std::string unicode_cpt_to_utf8(uint32_t cp);
 std::vector<uint32_t> unicode_cpts_from_utf8(const std::string & utf8);
 
 std::vector<uint32_t> unicode_cpts_normalize_nfd(const std::vector<uint32_t> & cpts);
 
-int unicode_cpt_type(uint32_t cp);
-int unicode_cpt_type(const std::string & utf8);
-
-bool unicode_cpt_is_whitespace(uint32_t cp);
+codepoint_flags unicode_cpt_flags(const uint32_t cp);
+codepoint_flags unicode_cpt_flags(const std::string & utf8);
 
 std::string unicode_byte_to_utf8(uint8_t byte);
 uint8_t unicode_utf8_to_byte(const std::string & utf8);