ci : fix windows builds to use 2019

* Add option to set the SYCL architecture for all targets
* Convert GGML_SYCL_HIP_TARGET to the more generic GGML_SYCL_ARCH option
* Document that setting GGML_SYCL_ARCH can improve the performance

.github/workflows/bindings-ruby.yml

@@ -0,0 +1,75 @@
+name: Bindings Tests (Ruby)
+on:
+  push:
+    paths:
+      - bindings/ruby/**
+      - src/whisper.cpp
+      - include/whisper.h
+      - ggml/src/ggml.c
+      - ggml/src/ggml-impl.h
+      - ggml/src/ggml-aarch64.h
+      - ggml/src/ggml-aarch64.c
+      - ggml/src/ggml-alloc.c
+      - ggml/src/ggml-backend-impl.h
+      - ggml/src/ggml-backend.cpp
+      - ggml/src/ggml-common.h
+      - ggml/src/ggml-quants.h
+      - ggml/src/ggml-quants.c
+      - ggml/src/ggml-cpu-impl.h
+      - ggml/src/ggml-metal.m
+      - ggml/src/ggml-metal.metal
+      - ggml/src/ggml-blas.cpp
+      - ggml/include/ggml.h
+      - ggml/include/ggml-alloc.h
+      - ggml/include/ggml-backend.h
+      - ggml/include/ggml-cuda.h
+      - ggml/include/ggml-kompute.h
+      - ggml/include/ggml-metal.h
+      - ggml/include/ggml-sycl.h
+      - ggml/include/ggml-vulkan.h
+      - ggml/include/ggml-blas.h
+      - scripts/get-flags.mk
+      - examples/dr_wav.h
+  pull_request:
+    paths:
+      - bindings/ruby/**
+      - src/whisper.cpp
+      - include/whisper.h
+      - ggml/src/ggml.c
+      - ggml/src/ggml-impl.h
+      - ggml/src/ggml-aarch64.h
+      - ggml/src/ggml-aarch64.c
+      - ggml/src/ggml-alloc.c
+      - ggml/src/ggml-backend-impl.h
+      - ggml/src/ggml-backend.cpp
+      - ggml/src/ggml-common.h
+      - ggml/src/ggml-quants.h
+      - ggml/src/ggml-quants.c
+      - ggml/src/ggml-cpu-impl.h
+      - ggml/src/ggml-metal.m
+      - ggml/src/ggml-metal.metal
+      - ggml/src/ggml-blas.cpp
+      - ggml/include/ggml.h
+      - ggml/include/ggml-alloc.h
+      - ggml/include/ggml-backend.h
+      - ggml/include/ggml-cuda.h
+      - ggml/include/ggml-kompute.h
+      - ggml/include/ggml-metal.h
+      - ggml/include/ggml-sycl.h
+      - ggml/include/ggml-vulkan.h
+      - ggml/include/ggml-blas.h
+      - scripts/get-flags.mk
+      - examples/dr_wav.h
+
+jobs:
+  ubuntu-latest:
+    runs-on: ubuntu-latest
+    defaults:
+      run:
+        working-directory: bindings/ruby
+    steps:
+      - uses: ruby/setup-ruby@v1
+        with:
+          ruby-version: '3.0'
+      - uses: actions/checkout@v4
+      - run: rake test

.github/workflows/bindings-ruby.yml.disabled

@@ -1,23 +0,0 @@
-# TODO: fix this workflow file, disabled for now
-name: Bindings Tests (Ruby)
-on:
-  push:
-    paths:
-      - bindings/ruby/**
-      - whisper.h
-  pull_request:
-    paths:
-      - bindings/ruby/**
-      - whisper.h
-
-jobs:
-  ubuntu-latest:
-    runs-on: ubuntu-latest
-    steps:
-      - uses: ruby/setup-ruby@v1
-        with:
-          ruby-version: '3.0'
-      - uses: actions/checkout@v1
-      - run: |
-          cd bindings/ruby/ext
-          ruby extconf.rb && make

.github/workflows/build.yml

@@ -3,6 +3,7 @@ on: [push, pull_request]
 
 env:
   ubuntu_image: "ubuntu:22.04"
+  VCPKG_BINARY_SOURCES: "clear;x-gha,readwrite"
 
 jobs:
   ubuntu-latest:
@@ -308,11 +309,11 @@ jobs:
       - name: Build using CMake w/ OpenBLAS
         shell: msys2 {0}
         run: |
-            cmake -B build -DGGML_OPENBLAS=ON
+            cmake -B build -DGGML_BLAS=ON -DGGML_BLAS_VENDOR=OpenBLAS
             cmake --build build --config ${{ matrix.build }} -j $(nproc)
 
   windows:
-    runs-on: windows-latest
+    runs-on: windows-2019
 
     strategy:
       matrix:
@@ -372,7 +373,7 @@ jobs:
           path: build/bin/${{ matrix.build }}
 
   windows-blas:
-    runs-on: windows-latest
+    runs-on: windows-2019
 
     strategy:
       matrix:
@@ -382,10 +383,8 @@ jobs:
         sdl2: [ON]
         include:
           - arch: Win32
-            obzip: https://github.com/OpenMathLib/OpenBLAS/releases/download/v0.3.25/OpenBLAS-0.3.25-x86.zip
             s2arc: x86
           - arch: x64
-            obzip: https://github.com/OpenMathLib/OpenBLAS/releases/download/v0.3.25/OpenBLAS-0.3.25-x64.zip
             s2arc: x64
           - sdl2: ON
             s2ver: 2.28.5
@@ -394,17 +393,21 @@ jobs:
       - name: Clone
         uses: actions/checkout@v4
 
+      - name: Export GitHub Actions cache environment variables
+        uses: actions/github-script@v7
+        with:
+          script: |
+            core.exportVariable('ACTIONS_CACHE_URL', process.env.ACTIONS_CACHE_URL || '');
+            core.exportVariable('ACTIONS_RUNTIME_TOKEN', process.env.ACTIONS_RUNTIME_TOKEN || '');
+
       - name: Add msbuild to PATH
         uses: microsoft/setup-msbuild@v2
 
-      - name: Fetch OpenBLAS
+      - name: Install OpenBLAS and pkgconfiglite
         if: matrix.blas == 'ON'
         run: |
-          C:/msys64/usr/bin/wget.exe -qO blas.zip ${{ matrix.obzip }}
-          7z x blas.zip -oblas -y
-          copy blas/include/cblas.h .
-          copy blas/include/openblas_config.h .
-          echo "OPENBLAS_PATH=$env:GITHUB_WORKSPACE/blas" >> $env:GITHUB_ENV
+          vcpkg install --triplet=${{ matrix.s2arc }}-windows openblas
+          choco install pkgconfiglite
 
       - name: Fetch SDL2 and set SDL2_DIR
         if: matrix.sdl2 == 'ON'
@@ -416,9 +419,10 @@ jobs:
       - name: Configure
         run: >
           cmake -S . -B ./build -A ${{ matrix.arch }}
+          -DCMAKE_TOOLCHAIN_FILE="$env:VCPKG_INSTALLATION_ROOT/scripts/buildsystems/vcpkg.cmake"
           -DCMAKE_BUILD_TYPE=${{ matrix.build }}
-          -DGGML_OPENBLAS=${{ matrix.blas }}
-          -DCMAKE_LIBRARY_PATH="$env:OPENBLAS_PATH/lib"
+          -DGGML_BLAS=${{ matrix.blas }}
+          -DGGML_BLAS_VENDOR=OpenBLAS
           -DWHISPER_SDL2=${{ matrix.sdl2 }}
 
       - name: Build
@@ -426,9 +430,9 @@ jobs:
           cd ./build
           msbuild ALL_BUILD.vcxproj -t:build -p:configuration=${{ matrix.build }} -p:platform=${{ matrix.arch }}
 
-      - name: Copy libopenblas.dll
+      - name: Copy openblas.dll
         if: matrix.blas == 'ON'
-        run: copy "$env:OPENBLAS_PATH/bin/libopenblas.dll" build/bin/${{ matrix.build }}
+        run: copy "C:/vcpkg/packages/openblas_${{ matrix.s2arc }}-windows/bin/openblas.dll" build/bin/${{ matrix.build }}
 
       - name: Copy SDL2.dll
         if: matrix.sdl2 == 'ON'
@@ -560,12 +564,6 @@ jobs:
         with:
           path: whisper
 
-      - name: Clone
-        uses: actions/checkout@v4
-        with:
-          repository: ggerganov/ggml
-          path: ggml
-
       - name: Install Java
         uses: actions/setup-java@v4
         with:
@@ -584,7 +582,7 @@ jobs:
         run: |
           export PATH_TO_GGML=$PWD/ggml
           cd whisper/examples/whisper.android
-          ./gradlew assembleRelease --no-daemon -PGGML_HOME=$PATH_TO_GGML
+          ./gradlew assembleRelease --no-daemon
 
 # TODO: disable because of following fail: https://github.com/ggerganov/whisper.cpp/actions/runs/11019444420/job/30627193602
 #  android_java:

.github/workflows/docker.yml

@@ -45,7 +45,7 @@ jobs:
         with:
           context: .
           push: true
-          platforms: ${{ matrix.config.platforms }}
+          platforms: ${{ matrix.config.platform }}
           tags: "ghcr.io/${{ github.repository }}:${{ matrix.config.tag }}-${{ env.COMMIT_SHA }}"
           file: ${{ matrix.config.dockerfile }}
 
@@ -54,6 +54,6 @@ jobs:
         with:
           context: .
           push: ${{ github.event_name == 'push' }}
-          platforms: ${{ matrix.config.platforms }}
+          platforms: ${{ matrix.config.platform }}
           tags: "ghcr.io/${{ github.repository }}:${{ matrix.config.tag }}"
           file: ${{ matrix.config.dockerfile }}

.gitignore

@@ -1,5 +1,6 @@
 *.o
 *.a
+*.d
 .cache/
 .coreml/
 .test/
@@ -19,6 +20,9 @@ build-*/
 .swiftpm
 *.metallib
 
+ggml-metal-embed.metal
+ggml-metal-embed.metal.tmp
+
 /main
 /stream
 /command

CMakeLists.txt

@@ -1,6 +1,6 @@
 cmake_minimum_required(VERSION 3.5) # for add_link_options and implicit target directories.
 project("whisper.cpp" C CXX)
-project("whisper.cpp" VERSION 1.7.1)
+project("whisper.cpp" VERSION 1.7.2)
 include(CheckIncludeFileCXX)
 
 set(SOVERSION 1)

Makefile

@@ -134,6 +134,10 @@ ifdef GGML_RPC
 	BUILD_TARGETS += rpc-server
 endif
 
+ifdef GGML_VULKAN
+	BUILD_TARGETS += vulkan-shaders-gen
+endif
+
 ifeq ($(shell sdl2-config --cflags --libs 2>/dev/null),)
 else
 	BUILD_TARGETS += \
@@ -440,17 +444,17 @@ endif
 else
 	MK_CFLAGS   += -march=rv64gcv -mabi=lp64d
 	MK_CXXFLAGS += -march=rv64gcv -mabi=lp64d
-endif
+endif # RISCV
 
 ifndef GGML_NO_ACCELERATE
 	# Mac OS - include Accelerate framework.
 	# `-framework Accelerate` works both with Apple Silicon and Mac Intel
 	ifeq ($(UNAME_S),Darwin)
-		MK_CPPFLAGS += -DGGML_USE_ACCELERATE -DGGML_USE_BLAS
+		MK_CPPFLAGS += -DGGML_USE_ACCELERATE -DGGML_USE_BLAS -DGGML_BLAS_USE_ACCELERATE
 		MK_CPPFLAGS += -DACCELERATE_NEW_LAPACK
 		MK_CPPFLAGS += -DACCELERATE_LAPACK_ILP64
 		MK_LDFLAGS  += -framework Accelerate
-		OBJ_GGML    += ggml/src/ggml-blas.o
+		OBJ_GGML    += ggml/src/ggml-blas/ggml-blas.o
 	endif
 endif # GGML_NO_ACCELERATE
 
@@ -460,29 +464,38 @@ ifndef GGML_NO_OPENMP
 	MK_CXXFLAGS += -fopenmp
 endif # GGML_NO_OPENMP
 
+ifdef WHISPER_COREML
+	MK_CXXFLAGS += -DWHISPER_USE_COREML
+	LDFLAGS     += -framework Foundation -framework CoreML
+
+ifdef WHISPER_COREML_ALLOW_FALLBACK
+	MK_CXXFLAGS += -DWHISPER_COREML_ALLOW_FALLBACK
+endif
+endif # WHISPER_COREML
+
 ifdef GGML_OPENBLAS
 	MK_CPPFLAGS += -DGGML_USE_BLAS $(shell pkg-config --cflags-only-I openblas)
 	MK_CFLAGS   += $(shell pkg-config --cflags-only-other openblas)
 	MK_LDFLAGS  += $(shell pkg-config --libs openblas)
-	OBJ_GGML    += ggml/src/ggml-blas.o
+	OBJ_GGML    += ggml/src/ggml-blas/ggml-blas.o
 endif # GGML_OPENBLAS
 
 ifdef GGML_OPENBLAS64
 	MK_CPPFLAGS += -DGGML_USE_BLAS $(shell pkg-config --cflags-only-I openblas64)
 	MK_CFLAGS   += $(shell pkg-config --cflags-only-other openblas64)
 	MK_LDFLAGS  += $(shell pkg-config --libs openblas64)
-	OBJ_GGML    += ggml/src/ggml-blas.o
+	OBJ_GGML    += ggml/src/ggml-blas/ggml-blas.o
 endif # GGML_OPENBLAS64
 
 ifdef GGML_BLIS
 	MK_CPPFLAGS += -DGGML_USE_BLAS -I/usr/local/include/blis -I/usr/include/blis
 	MK_LDFLAGS  += -lblis -L/usr/local/lib
-	OBJ_GGML    += ggml/src/ggml-blas.o
+	OBJ_GGML    += ggml/src/ggml-blas/ggml-blas.o
 endif # GGML_BLIS
 
 ifdef GGML_RPC
 	MK_CPPFLAGS += -DGGML_USE_RPC
-	OBJ_GGML    += ggml/src/ggml-rpc.o
+	OBJ_GGML    += ggml/src/ggml-rpc/ggml-rpc.o
 endif # GGML_RPC
 
 OBJ_CUDA_TMPL      = $(patsubst %.cu,%.o,$(wildcard ggml/src/ggml-cuda/template-instances/fattn-wmma*.cu))
@@ -509,7 +522,7 @@ ifdef GGML_CUDA
 	MK_LDFLAGS   += -lcuda -lcublas -lculibos -lcudart -lcublasLt -lpthread -ldl -lrt -L$(CUDA_PATH)/lib64 -L/usr/lib64 -L$(CUDA_PATH)/targets/$(UNAME_M)-linux/lib -L$(CUDA_PATH)/lib64/stubs -L/usr/lib/wsl/lib
 	MK_NVCCFLAGS += -use_fast_math
 
-	OBJ_GGML += ggml/src/ggml-cuda.o
+	OBJ_GGML += ggml/src/ggml-cuda/ggml-cuda.o
 	OBJ_GGML += $(patsubst %.cu,%.o,$(wildcard ggml/src/ggml-cuda/*.cu))
 	OBJ_GGML += $(OBJ_CUDA_TMPL)
 ifdef WHISPER_FATAL_WARNINGS
@@ -611,11 +624,11 @@ ggml/src/ggml-cuda/%.o: \
 	ggml/src/ggml-cuda/common.cuh
 	$(NVCC_COMPILE)
 
-ggml/src/ggml-cuda.o: \
-	ggml/src/ggml-cuda.cu \
+ggml/src/ggml-cuda/ggml-cuda.o: \
+	ggml/src/ggml-cuda/ggml-cuda.cu \
+	ggml/include/ggml-cuda.h \
 	ggml/include/ggml.h \
 	ggml/include/ggml-backend.h \
-	ggml/include/ggml-cuda.h \
 	ggml/src/ggml-backend-impl.h \
 	ggml/src/ggml-common.h \
 	$(wildcard ggml/src/ggml-cuda/*.cuh)
@@ -624,8 +637,8 @@ endif # GGML_CUDA
 
 ifdef GGML_VULKAN
 	MK_CPPFLAGS += -DGGML_USE_VULKAN
-	MK_LDFLAGS  += -lvulkan
-	OBJ_GGML    += ggml/src/ggml-vulkan.o
+	MK_LDFLAGS  += $(shell pkg-config --libs vulkan)
+	OBJ_GGML    += ggml/src/ggml-vulkan.o ggml/src/ggml-vulkan-shaders.o
 
 ifdef GGML_VULKAN_CHECK_RESULTS
 	MK_CPPFLAGS  += -DGGML_VULKAN_CHECK_RESULTS
@@ -639,6 +652,10 @@ ifdef GGML_VULKAN_MEMORY_DEBUG
 	MK_CPPFLAGS  += -DGGML_VULKAN_MEMORY_DEBUG
 endif
 
+ifdef GGML_VULKAN_PERF
+	MK_CPPFLAGS  += -DGGML_VULKAN_PERF
+endif
+
 ifdef GGML_VULKAN_VALIDATE
 	MK_CPPFLAGS  += -DGGML_VULKAN_VALIDATE
 endif
@@ -647,10 +664,28 @@ ifdef GGML_VULKAN_RUN_TESTS
 	MK_CPPFLAGS  += -DGGML_VULKAN_RUN_TESTS
 endif
 
-ggml/src/ggml-vulkan.o: \
-	ggml/src/ggml-vulkan.cpp \
-	ggml/include/ggml-vulkan.h
-	$(CXX) $(CXXFLAGS) -c $< -o $@
+GLSLC_CMD  = glslc
+_ggml_vk_genshaders_cmd = $(shell pwd)/vulkan-shaders-gen
+_ggml_vk_header = ggml/src/ggml-vulkan-shaders.hpp
+_ggml_vk_source = ggml/src/ggml-vulkan-shaders.cpp
+_ggml_vk_input_dir = ggml/src/vulkan-shaders
+_ggml_vk_shader_deps = $(echo $(_ggml_vk_input_dir)/*.comp)
+
+ggml/src/ggml-vulkan.o: ggml/src/ggml-vulkan.cpp ggml/include/ggml-vulkan.h $(_ggml_vk_header) $(_ggml_vk_source)
+	$(CXX) $(CXXFLAGS) $(shell pkg-config --cflags vulkan) -c $< -o $@
+
+$(_ggml_vk_header): $(_ggml_vk_source)
+
+$(_ggml_vk_source): $(_ggml_vk_shader_deps) vulkan-shaders-gen
+	$(_ggml_vk_genshaders_cmd) \
+		--glslc      $(GLSLC_CMD) \
+		--input-dir  $(_ggml_vk_input_dir) \
+		--target-hpp $(_ggml_vk_header) \
+		--target-cpp $(_ggml_vk_source)
+
+vulkan-shaders-gen: ggml/src/vulkan-shaders/vulkan-shaders-gen.cpp
+	$(CXX) $(CXXFLAGS) -o $@ $(LDFLAGS) ggml/src/vulkan-shaders/vulkan-shaders-gen.cpp
+
 endif # GGML_VULKAN
 
 ifdef GGML_HIPBLAS
@@ -716,50 +751,43 @@ endif # GGML_HIPBLAS
 ifdef GGML_METAL
 	MK_CPPFLAGS += -DGGML_USE_METAL
 	MK_LDFLAGS  += -framework Foundation -framework Metal -framework MetalKit
-	OBJ_GGML	+= ggml/src/ggml-metal.o
+	OBJ_GGML	+= ggml/src/ggml-metal/ggml-metal.o
 ifdef GGML_METAL_NDEBUG
 	MK_CPPFLAGS += -DGGML_METAL_NDEBUG
 endif
 
 ifdef GGML_METAL_EMBED_LIBRARY
 	MK_CPPFLAGS += -DGGML_METAL_EMBED_LIBRARY
-	OBJ_GGML    += ggml/src/ggml-metal-embed.o
+	OBJ_GGML    += ggml/src/ggml-metal/ggml-metal-embed.o
 endif
 endif # GGML_METAL
 
-ifdef WHISPER_COREML
-	MK_CXXFLAGS += -DWHISPER_USE_COREML
-	LDFLAGS     += -framework Foundation -framework CoreML
-
-ifdef WHISPER_COREML_ALLOW_FALLBACK
-	MK_CXXFLAGS += -DWHISPER_COREML_ALLOW_FALLBACK
-endif
-endif
-
-# ===
-
 ifdef GGML_METAL
-ggml/src/ggml-metal.o: \
-	ggml/src/ggml-metal.m \
+ggml/src/ggml-metal/ggml-metal.o: \
+	ggml/src/ggml-metal/ggml-metal.m \
+	ggml/src/ggml-metal/ggml-metal-impl.h \
 	ggml/include/ggml-metal.h \
 	ggml/include/ggml.h
 	$(CC) $(CFLAGS) -c $< -o $@
 
 ifdef GGML_METAL_EMBED_LIBRARY
-ggml/src/ggml-metal-embed.o: \
-	ggml/src/ggml-metal.metal \
+ggml/src/ggml-metal/ggml-metal-embed.o: \
+	ggml/src/ggml-metal/ggml-metal.metal \
+	ggml/src/ggml-metal/ggml-metal-impl.h \
 	ggml/src/ggml-common.h
 	@echo "Embedding Metal library"
-	@sed -e '/#include "ggml-common.h"/r ggml/src/ggml-common.h' -e '/#include "ggml-common.h"/d' < ggml/src/ggml-metal.metal > ggml/src/ggml-metal-embed.metal
-	$(eval TEMP_ASSEMBLY=$(shell mktemp))
-	@echo ".section __DATA, __ggml_metallib"            >  $(TEMP_ASSEMBLY)
-	@echo ".globl _ggml_metallib_start"                 >> $(TEMP_ASSEMBLY)
-	@echo "_ggml_metallib_start:"                       >> $(TEMP_ASSEMBLY)
-	@echo ".incbin \"ggml/src/ggml-metal-embed.metal\"" >> $(TEMP_ASSEMBLY)
-	@echo ".globl _ggml_metallib_end"                   >> $(TEMP_ASSEMBLY)
-	@echo "_ggml_metallib_end:"                         >> $(TEMP_ASSEMBLY)
-	@$(AS) $(TEMP_ASSEMBLY) -o $@
-	@rm -f ${TEMP_ASSEMBLY}
+	@sed -e '/__embed_ggml-common.h__/r      ggml/src/ggml-common.h'                -e '/__embed_ggml-common.h__/d'      < ggml/src/ggml-metal/ggml-metal.metal           > ggml/src/ggml-metal/ggml-metal-embed.metal.tmp
+	@sed -e '/#include "ggml-metal-impl.h"/r ggml/src/ggml-metal/ggml-metal-impl.h' -e '/#include "ggml-metal-impl.h"/d' < ggml/src/ggml-metal/ggml-metal-embed.metal.tmp > ggml/src/ggml-metal/ggml-metal-embed.metal
+	$(eval TEMP_ASSEMBLY=$(shell mktemp -d))
+	@echo ".section __DATA, __ggml_metallib"                       >  $(TEMP_ASSEMBLY)/ggml-metal-embed.s
+	@echo ".globl _ggml_metallib_start"                            >> $(TEMP_ASSEMBLY)/ggml-metal-embed.s
+	@echo "_ggml_metallib_start:"                                  >> $(TEMP_ASSEMBLY)/ggml-metal-embed.s
+	@echo ".incbin \"ggml/src/ggml-metal/ggml-metal-embed.metal\"" >> $(TEMP_ASSEMBLY)/ggml-metal-embed.s
+	@echo ".globl _ggml_metallib_end"                              >> $(TEMP_ASSEMBLY)/ggml-metal-embed.s
+	@echo "_ggml_metallib_end:"                                    >> $(TEMP_ASSEMBLY)/ggml-metal-embed.s
+	$(CC) $(CFLAGS) -c $(TEMP_ASSEMBLY)/ggml-metal-embed.s -o $@
+	@rm -f ${TEMP_ASSEMBLY}/ggml-metal-embed.s
+	@rmdir ${TEMP_ASSEMBLY}
 endif
 endif # GGML_METAL
 
@@ -775,10 +803,17 @@ endif
 
 OBJ_GGML += \
 	ggml/src/ggml.o \
+	ggml/src/ggml-aarch64.o \
 	ggml/src/ggml-alloc.o \
 	ggml/src/ggml-backend.o \
+	ggml/src/ggml-backend-reg.o \
+	ggml/src/ggml-opt.o \
 	ggml/src/ggml-quants.o \
-	ggml/src/ggml-aarch64.o
+	ggml/src/ggml-threading.o \
+	ggml/src/ggml-cpu/ggml-cpu.o \
+	ggml/src/ggml-cpu/ggml-cpu-cpp.o \
+	ggml/src/ggml-cpu/ggml-cpu-aarch64.o \
+	ggml/src/ggml-cpu/ggml-cpu-quants.o
 
 OBJ_WHISPER += \
 	src/whisper.o
@@ -883,108 +918,64 @@ endif
 # Build libraries
 #
 
-# ggml
+LIB_GGML   = libggml.so
+LIB_GGML_S = libggml.a
 
-ggml/src/ggml.o: \
-	ggml/src/ggml.c \
-	ggml/include/ggml.h
-	$(CC)  $(CFLAGS)   -c $< -o $@
+LIB_LLAMA   = libllama.so
+LIB_LLAMA_S = libllama.a
 
-ggml/src/ggml-alloc.o: \
-	ggml/src/ggml-alloc.c \
-	ggml/include/ggml.h \
-	ggml/include/ggml-alloc.h
-	$(CC)  $(CFLAGS)   -c $< -o $@
+LIB_COMMON   = libcommon.so
+LIB_COMMON_S = libcommon.a
 
-ggml/src/ggml-backend.o: \
-	ggml/src/ggml-backend.cpp \
-	ggml/include/ggml.h \
-	ggml/include/ggml-backend.h
-	$(CXX) $(CXXFLAGS) -c $< -o $@
+LIB_COMMON_SDL   = libcommon-sdl.so
+LIB_COMMON_SDL_S = libcommon-sdl.a
 
-ggml/src/ggml-quants.o: \
-	ggml/src/ggml-quants.c \
-	ggml/include/ggml.h \
-	ggml/src/ggml-quants.h \
-	ggml/src/ggml-common.h
-	$(CC) $(CFLAGS)    -c $< -o $@
+# Targets
+BUILD_TARGETS += $(LIB_GGML) $(LIB_GGML_S) $(LIB_LLAMA) $(LIB_LLAMA_S) $(LIB_COMMON) $(LIB_COMMON_S)
 
-ggml/src/ggml-aarch64.o: \
-	ggml/src/ggml-aarch64.c \
-	ggml/include/ggml.h \
-	ggml/src/ggml-aarch64.h \
-	ggml/src/ggml-common.h
-	$(CC) $(CFLAGS)    -c $< -o $@
+# Dependency files
+DEP_FILES = $(OBJ_GGML:.o=.d) $(OBJ_LLAMA:.o=.d) $(OBJ_COMMON:.o=.d)
 
-ggml/src/ggml-blas.o: \
-	ggml/src/ggml-blas.cpp \
-	ggml/include/ggml-blas.h
-	$(CXX) $(CXXFLAGS) -c $< -o $@
+# Default target
+all: $(BUILD_TARGETS)
 
-ifdef GGML_LLAMAFILE
-ggml/src/sgemm.o: \
-	ggml/src/sgemm.cpp \
-	ggml/src/sgemm.h \
-	ggml/include/ggml.h
-	$(CXX) $(CXXFLAGS) -c $< -o $@
-endif # GGML_LLAMAFILE
-
-ifdef GGML_RPC
-ggml/src/ggml-rpc.o: \
-	ggml/src/ggml-rpc.cpp \
-	ggml/include/ggml-rpc.h
-	$(CXX) $(CXXFLAGS) -c $< -o $@
-endif # GGML_RPC
-
-$(LIB_GGML): \
-	$(OBJ_GGML)
-	$(CXX) $(CXXFLAGS) -shared -fPIC -o $@ $^ $(LDFLAGS)
-
-$(LIB_GGML_S): \
-	$(OBJ_GGML)
-	ar rcs $(LIB_GGML_S) $^
-
-# whisper
-
-src/whisper.o: \
-	src/whisper.cpp \
-	include/whisper.h \
+# Note: need this exception because `ggml-cpu.c` and `ggml-cpu.cpp` both produce the same obj/dep files
+#       g++ -M -I ./ggml/include/ -I ./ggml/src ggml/src/ggml-cpu/ggml-cpu.cpp | grep ggml
+ggml/src/ggml-cpu/ggml-cpu-cpp.o: \
+	ggml/src/ggml-cpu/ggml-cpu.cpp \
+	ggml/include/ggml-backend.h \
 	ggml/include/ggml.h \
 	ggml/include/ggml-alloc.h \
-	ggml/include/ggml-backend.h \
-	ggml/include/ggml-cuda.h \
-	ggml/include/ggml-metal.h
-	$(CXX) $(CXXFLAGS) -c $< -o $@
+	ggml/src/ggml-backend-impl.h \
+	ggml/include/ggml-cpu.h \
+	ggml/src/ggml-impl.h
+	$(CXX) $(CXXFLAGS)   -c $< -o $@
 
-$(LIB_WHISPER): \
-	$(OBJ_WHISPER) \
-	$(LIB_GGML)
+# Rules for building object files
+ggml/%.o: ggml/%.c
+	$(CC) $(CFLAGS) -MMD -c $< -o $@
+
+ggml/%.o: ggml/%.cpp
+	$(CXX) $(CXXFLAGS) -MMD -c $< -o $@
+
+src/%.o: src/%.cpp
+	$(CXX) $(CXXFLAGS) -MMD -c $< -o $@
+
+examples/%.o: examples/%.cpp
+	$(CXX) $(CXXFLAGS) -MMD -c $< -o $@
+
+# Rules for building libraries
+$(LIB_GGML): $(OBJ_GGML)
 	$(CXX) $(CXXFLAGS) -shared -fPIC -o $@ $^ $(LDFLAGS)
 
-$(LIB_WHISPER_S): \
-	$(OBJ_WHISPER) \
-	$(OBJ_GGML)
-	ar rcs $(LIB_WHISPER_S) $^
+$(LIB_GGML_S): $(OBJ_GGML)
+	ar rcs $(LIB_GGML_S) $^
 
-# common
-
-examples/common.o: \
-	examples/common.cpp \
-	examples/common.h
-	$(CXX) $(CXXFLAGS) -c $< -o $@
-
-examples/common-ggml.o: \
-	examples/common-ggml.cpp \
-	examples/common-ggml.h
-	$(CXX) $(CXXFLAGS) -c $< -o $@
-
-$(LIB_COMMON): \
-	$(OBJ_COMMON)
+$(LIB_LLAMA): $(OBJ_LLAMA) $(LIB_GGML)
 	$(CXX) $(CXXFLAGS) -shared -fPIC -o $@ $^ $(LDFLAGS)
 
-$(LIB_COMMON_S): \
-	$(OBJ_COMMON)
-	ar rcs $(LIB_COMMON_S) $^
+$(LIB_LLAMA_S): $(OBJ_LLAMA)
+	ar rcs $(LIB_LLAMA_S) $^
 
 # common-sdl
 
@@ -996,34 +987,21 @@ examples/common-sdl.o: \
 	examples/common-sdl.h
 	$(CXX) $(CXXFLAGS) $(CFLAGS_SDL) -c $< -o $@
 
-$(LIB_COMMON_SDL): \
-	$(OBJ_SDL)
-	$(CXX) $(CXXFLAGS) -shared -fPIC -o $@ $^ $(LDFLAGS) $(LDFLAGS_SDL)
+$(LIB_COMMON): $(OBJ_COMMON) $(LIB_LLAMA) $(LIB_GGML)
+	$(CXX) $(CXXFLAGS) -shared -fPIC -o $@ $^ $(LDFLAGS)
 
-$(LIB_COMMON_SDL_S): \
-	$(OBJ_SDL)
-	ar rcs $(LIB_COMMON_SDL_S) $^
+$(LIB_COMMON_S): $(OBJ_COMMON)
+	ar rcs $(LIB_COMMON_S) $^
 
+# Include dependency files
+-include $(DEP_FILES)
+
+# Clean rule
 clean:
-	rm -vrf *.dot $(BUILD_TARGETS) $(TEST_TARGETS)
-	rm -rvf src/*.o
-	rm -rvf src/coreml/*.o
-	rm -rvf tests/*.o
-	rm -rvf examples/*.o
-	rm -rvf *.a
-	rm -rvf *.dll
-	rm -rvf *.so
-	rm -rvf *.dot
-	rm -rvf ggml/*.a
-	rm -rvf ggml/*.dll
-	rm -rvf ggml/*.so
-	rm -vrf ggml/src/*.o
-	rm -vrf ggml/src/ggml-metal-embed.metal
-	rm -vrf ggml/src/ggml-cuda/*.o
-	rm -vrf ggml/src/ggml-cuda/template-instances/*.o
-	rm -rvf $(BUILD_TARGETS)
-	rm -rvf $(TEST_TARGETS)
-	find examples -type f -name "*.o" -delete
+	rm -vrf $(BUILD_TARGETS) $(TEST_TARGETS)
+	rm -rvf *.a *.dll *.so *.dot
+	find ggml src tests examples -type f -name "*.o" -delete
+	find ggml src tests examples -type f -name "*.d" -delete
 
 #
 # Examples

Package.swift

@@ -18,16 +18,17 @@ let package = Package(
             name: "whisper",
             path: ".",
             exclude: [
+               "build",
                "bindings",
                "cmake",
-               "coreml",
                "examples",
-               "extra",
+               "scripts",
                "models",
                "samples",
                "tests",
                "CMakeLists.txt",
-               "Makefile"
+               "Makefile",
+               "ggml/src/ggml-metal/ggml-metal-embed.metal"
             ],
             sources: [
                 "ggml/src/ggml.c",
@@ -35,15 +36,22 @@ let package = Package(
                 "ggml/src/ggml-aarch64.c",
                 "ggml/src/ggml-alloc.c",
                 "ggml/src/ggml-backend.cpp",
+                "ggml/src/ggml-backend-reg.cpp",
+                "ggml/src/ggml-cpu/ggml-cpu.c",
+                "ggml/src/ggml-cpu/ggml-cpu.cpp",
+                "ggml/src/ggml-cpu/ggml-cpu-aarch64.c",
+                "ggml/src/ggml-cpu/ggml-cpu-quants.c",
                 "ggml/src/ggml-quants.c",
-                "ggml/src/ggml-metal.m"
+                "ggml/src/ggml-threading.cpp",
+                "ggml/src/ggml-metal/ggml-metal.m"
             ],
-            resources: [.process("ggml-metal.metal")],
+            resources: [.process("ggml/src/ggml-metal/ggml-metal.metal")],
             publicHeadersPath: "spm-headers",
             cSettings: [
                 .unsafeFlags(["-Wno-shorten-64-to-32", "-O3", "-DNDEBUG"]),
-                .define("GGML_USE_ACCELERATE"),
                 .unsafeFlags(["-fno-objc-arc"]),
+                .headerSearchPath("ggml/src"),
+                .define("GGML_USE_ACCELERATE"),
                 .define("GGML_USE_METAL")
                 // NOTE: NEW_LAPACK will required iOS version 16.4+
                 // We should consider add this in the future when we drop support for iOS 14

README.md

@@ -7,21 +7,22 @@
 [![Conan Center](https://shields.io/conan/v/whisper-cpp)](https://conan.io/center/whisper-cpp)
 [![npm](https://img.shields.io/npm/v/whisper.cpp.svg)](https://www.npmjs.com/package/whisper.cpp/)
 
-Stable: [v1.7.1](https://github.com/ggerganov/whisper.cpp/releases/tag/v1.7.1) / [Roadmap | F.A.Q.](https://github.com/ggerganov/whisper.cpp/discussions/126)
+Stable: [v1.7.2](https://github.com/ggerganov/whisper.cpp/releases/tag/v1.7.2) / [Roadmap | F.A.Q.](https://github.com/ggerganov/whisper.cpp/discussions/126)
 
 High-performance inference of [OpenAI's Whisper](https://github.com/openai/whisper) automatic speech recognition (ASR) model:
 
 - Plain C/C++ implementation without dependencies
-- Apple Silicon first-class citizen - optimized via ARM NEON, Accelerate framework, Metal and [Core ML](https://github.com/ggerganov/whisper.cpp#core-ml-support)
+- Apple Silicon first-class citizen - optimized via ARM NEON, Accelerate framework, Metal and [Core ML](#core-ml-support)
 - AVX intrinsics support for x86 architectures
 - VSX intrinsics support for POWER architectures
 - Mixed F16 / F32 precision
-- [4-bit and 5-bit integer quantization support](https://github.com/ggerganov/whisper.cpp#quantization)
+- [Integer quantization support](#quantization)
 - Zero memory allocations at runtime
+- [Vulkan support](#vulkan-gpu-support)
 - Support for CPU-only inference
-- [Efficient GPU support for NVIDIA](https://github.com/ggerganov/whisper.cpp#nvidia-gpu-support-via-cublas)
-- [OpenVINO Support](https://github.com/ggerganov/whisper.cpp#openvino-support)
-- [Ascend NPU Support](https://github.com/ggerganov/whisper.cpp#ascend-npu-support)
+- [Efficient GPU support for NVIDIA](#nvidia-gpu-support)
+- [OpenVINO Support](#openvino-support)
+- [Ascend NPU Support](#ascend-npu-support)
 - [C-style API](https://github.com/ggerganov/whisper.cpp/blob/master/include/whisper.h)
 
 Supported platforms:
@@ -72,6 +73,12 @@ First clone the repository:
 git clone https://github.com/ggerganov/whisper.cpp.git
 ```
 
+Navigate into the directory:
+
+```
+cd whisper.cpp
+```
+
 Then, download one of the Whisper [models](models/README.md) converted in [`ggml` format](#ggml-format). For example:
 
 ```bash
@@ -82,7 +89,7 @@ Now build the [main](examples/main) example and transcribe an audio file like th
 
 ```bash
 # build the main example
-make
+make -j
 
 # transcribe an audio file
 ./main -f samples/jfk.wav
@@ -93,7 +100,7 @@ make
 For a quick demo, simply run `make base.en`:
 
 ```text
-$ make base.en
+$ make -j base.en
 
 cc  -I.              -O3 -std=c11   -pthread -DGGML_USE_ACCELERATE   -c ggml.c -o ggml.o
 c++ -I. -I./examples -O3 -std=c++11 -pthread -c whisper.cpp -o whisper.o
@@ -217,7 +224,7 @@ ffmpeg -i input.mp3 -ar 16000 -ac 1 -c:a pcm_s16le output.wav
 If you want some extra audio samples to play with, simply run:
 
 ```
-make samples
+make -j samples
 ```
 
 This will download a few more audio files from Wikipedia and convert them to 16-bit WAV format via `ffmpeg`.
@@ -225,18 +232,18 @@ This will download a few more audio files from Wikipedia and convert them to 16-
 You can download and run the other models as follows:
 
 ```
-make tiny.en
-make tiny
-make base.en
-make base
-make small.en
-make small
-make medium.en
-make medium
-make large-v1
-make large-v2
-make large-v3
-make large-v3-turbo
+make -j tiny.en
+make -j tiny
+make -j base.en
+make -j base
+make -j small.en
+make -j small
+make -j medium.en
+make -j medium
+make -j large-v1
+make -j large-v2
+make -j large-v3
+make -j large-v3-turbo
 ```
 
 ## Memory usage
@@ -258,7 +265,7 @@ Here are the steps for creating and using a quantized model:
 
 ```bash
 # quantize a model with Q5_0 method
-make quantize
+make -j quantize
 ./quantize models/ggml-base.en.bin models/ggml-base.en-q5_0.bin q5_0
 
 # run the examples as usual, specifying the quantized model file
@@ -423,6 +430,16 @@ make clean
 GGML_CUDA=1 make -j
 ```
 
+## Vulkan GPU support
+Cross-vendor solution which allows you to accelerate workload on your GPU.
+First, make sure your graphics card driver provides support for Vulkan API.
+
+Now build `whisper.cpp` with Vulkan support:
+```
+make clean
+make GGML_VULKAN=1 -j
+```
+
 ## BLAS CPU support via OpenBLAS
 
 Encoder processing can be accelerated on the CPU via OpenBLAS.
@@ -619,7 +636,7 @@ The [stream](examples/stream) tool samples the audio every half a second and run
 More info is available in [issue #10](https://github.com/ggerganov/whisper.cpp/issues/10).
 
 ```bash
-make stream
+make stream -j
 ./stream -m ./models/ggml-base.en.bin -t 8 --step 500 --length 5000
 ```
 

bindings/javascript/package.json

@@ -1,6 +1,6 @@
 {
   "name": "whisper.cpp",
-  "version": "1.7.1",
+  "version": "1.7.2",
   "description": "Whisper speech recognition",
   "main": "whisper.js",
   "scripts": {

bindings/ruby/.gitignore

@@ -0,0 +1,3 @@
+LICENSE
+pkg/
+lib/whisper.*

bindings/ruby/README.md

@@ -0,0 +1,169 @@
+whispercpp
+==========
+
+![whisper.cpp](https://user-images.githubusercontent.com/1991296/235238348-05d0f6a4-da44-4900-a1de-d0707e75b763.jpeg)
+
+Ruby bindings for [whisper.cpp][], an interface of automatic speech recognition model.
+
+Installation
+------------
+
+Install the gem and add to the application's Gemfile by executing:
+
+    $ bundle add whispercpp
+
+If bundler is not being used to manage dependencies, install the gem by executing:
+
+    $ gem install whispercpp
+
+Usage
+-----
+
+```ruby
+require "whisper"
+
+whisper = Whisper::Context.new("path/to/model.bin")
+
+params = Whisper::Params.new
+params.language = "en"
+params.offset = 10_000
+params.duration = 60_000
+params.max_text_tokens = 300
+params.translate = true
+params.print_timestamps = false
+params.initial_prompt = "Initial prompt here."
+
+whisper.transcribe("path/to/audio.wav", params) do |whole_text|
+  puts whole_text
+end
+
+```
+
+### Preparing model ###
+
+Use script to download model file(s):
+
+```bash
+git clone https://github.com/ggerganov/whisper.cpp.git
+cd whisper.cpp
+sh ./models/download-ggml-model.sh base.en
+```
+
+There are some types of models. See [models][] page for details.
+
+### Preparing audio file ###
+
+Currently, whisper.cpp accepts only 16-bit WAV files.
+
+### API ###
+
+Once `Whisper::Context#transcribe` called, you can retrieve segments by `#each_segment`:
+
+```ruby
+def format_time(time_ms)
+  sec, decimal_part = time_ms.divmod(1000)
+  min, sec = sec.divmod(60)
+  hour, min = min.divmod(60)
+  "%02d:%02d:%02d.%03d" % [hour, min, sec, decimal_part]
+end
+
+whisper.transcribe("path/to/audio.wav", params)
+
+whisper.each_segment.with_index do |segment, index|
+  line = "[%{nth}: %{st} --> %{ed}] %{text}" % {
+    nth: index + 1,
+    st: format_time(segment.start_time),
+    ed: format_time(segment.end_time),
+    text: segment.text
+  }
+  line << " (speaker turned)" if segment.speaker_next_turn?
+  puts line
+end
+
+```
+
+You can also add hook to params called on new segment:
+
+```ruby
+def format_time(time_ms)
+  sec, decimal_part = time_ms.divmod(1000)
+  min, sec = sec.divmod(60)
+  hour, min = min.divmod(60)
+  "%02d:%02d:%02d.%03d" % [hour, min, sec, decimal_part]
+end
+
+# Add hook before calling #transcribe
+params.on_new_segment do |segment|
+  line = "[%{st} --> %{ed}] %{text}" % {
+    st: format_time(segment.start_time),
+    ed: format_time(segment.end_time),
+    text: segment.text
+  }
+  line << " (speaker turned)" if segment.speaker_next_turn?
+  puts line
+end
+
+whisper.transcribe("path/to/audio.wav", params)
+
+```
+
+You can see model information:
+
+```ruby
+whisper = Whisper::Context.new("path/to/model.bin")
+model = whisper.model
+
+model.n_vocab # => 51864
+model.n_audio_ctx # => 1500
+model.n_audio_state # => 512
+model.n_audio_head # => 8
+model.n_audio_layer # => 6
+model.n_text_ctx # => 448
+model.n_text_state # => 512
+model.n_text_head # => 8
+model.n_text_layer # => 6
+model.n_mels # => 80
+model.ftype # => 1
+model.type # => "base"
+
+```
+
+You can set log callback:
+
+```ruby
+prefix = "[MyApp] "
+log_callback = ->(level, buffer, user_data) {
+  case level
+  when Whisper::LOG_LEVEL_NONE
+    puts "#{user_data}none: #{buffer}"
+  when Whisper::LOG_LEVEL_INFO
+    puts "#{user_data}info: #{buffer}"
+  when Whisper::LOG_LEVEL_WARN
+    puts "#{user_data}warn: #{buffer}"
+  when Whisper::LOG_LEVEL_ERROR
+    puts "#{user_data}error: #{buffer}"
+  when Whisper::LOG_LEVEL_DEBUG
+    puts "#{user_data}debug: #{buffer}"
+  when Whisper::LOG_LEVEL_CONT
+    puts "#{user_data}same to previous: #{buffer}"
+  end
+}
+Whisper.log_set log_callback, prefix
+```
+
+Using this feature, you are also able to suppress log:
+
+```ruby
+Whisper.log_set ->(level, buffer, user_data) {
+  # do nothing
+}, nil
+Whisper::Context.new(MODEL)
+```
+
+License
+-------
+
+The same to [whisper.cpp][].
+
+[whisper.cpp]: https://github.com/ggerganov/whisper.cpp
+[models]: https://github.com/ggerganov/whisper.cpp/tree/master/models

bindings/ruby/Rakefile

@@ -1,12 +1,68 @@
 require 'rake/clean'
-  require 'rubygems/package'
+require "bundler/gem_tasks"
+require "pathname"
+require "yaml"
+require "rake/testtask"
 
-desc 'Build gem'
-task :package do
-  spec_source = File.read File.join(File.dirname(__FILE__),'whispercpp.gemspec')
-  spec = nil
-  # see: http://gist.github.com/16215
-  Thread.new { spec = eval("#{spec_source}") }.join
-  spec.validate
-  Gem::Package.build(spec)
+extsources = YAML.load_file("extsources.yaml")
+SOURCES = FileList[]
+extsources.each do |src|
+  basename = src.pathmap("%f")
+  dest = basename == "LICENSE" ? basename : basename.pathmap("ext/%f")
+  file src
+  file dest => src do |t|
+    cp t.source, t.name
+  end
+  SOURCES.include dest
+end
+CLEAN.include SOURCES
+CLEAN.include FileList[
+                "ext/*.o",
+                "ext/*.metal",
+                "ext/whisper.{so,bundle,dll}",
+                "ext/depend"
+              ]
+
+task build: FileList[
+       "ext/Makefile",
+       "ext/ruby_whisper.h",
+       "ext/ruby_whisper.cpp",
+       "whispercpp.gemspec",
+     ]
+
+directory "pkg"
+CLOBBER.include "pkg"
+
+TEST_MODEL = "../../models/ggml-base.en.bin"
+LIB_NAME = "whisper".ext(RbConfig::CONFIG["DLEXT"])
+SO_FILE = File.join("ext", LIB_NAME)
+LIB_FILE = File.join("lib", LIB_NAME)
+
+file "ext/Makefile" => ["ext/extconf.rb", "ext/ruby_whisper.h", "ext/ruby_whisper.cpp"] + SOURCES do |t|
+  Dir.chdir "ext" do
+    ruby "extconf.rb"
+  end
+end
+
+file SO_FILE => "ext/Makefile" do |t|
+  Dir.chdir "ext" do
+    sh "make"
+  end
+end
+CLEAN.include LIB_FILE
+
+directory "lib"
+file LIB_FILE => [SO_FILE, "lib"] do |t|
+  copy t.source, t.name
+end
+
+Rake::TestTask.new do |t|
+  t.test_files = FileList["tests/test_*.rb"]
+end
+task test: [TEST_MODEL, LIB_FILE]
+
+file TEST_MODEL do
+  Dir.chdir "../.." do
+    sh "./models/download-ggml-model.sh base.en"
+  end
 end

bindings/ruby/ext/.gitignore

@@ -3,7 +3,33 @@ ggml.c
 ggml.h
 ggml-alloc.c
 ggml-alloc.h
-whisper.bundle
+ggml-aarch64.c
+ggml-aarch64.h
+ggml-backend.cpp
+ggml-backend-impl.h
+ggml-backend.c
+ggml-backend.h
+ggml-common.h
+ggml-cpu-impl.h
+ggml-metal.m
+ggml-metal.metal
+ggml-metal-embed.metal
+ggml-blas.cpp
+ggml-cuda.h
+ggml-impl.h
+ggml-kompute.h
+ggml-metal.h
+ggml-opencl.h
+ggml-quants.c
+ggml-quants.h
+ggml-sycl.h
+ggml-vulkan.h
+ggml-blas.h
+get-flags.mk
 whisper.cpp
 whisper.h
 dr_wav.h
+depend
+whisper.bundle
+whisper.so
+whisper.dll

bindings/ruby/ext/extconf.rb

@@ -1,24 +1,10 @@
 require 'mkmf'
-system("cp #{File.join(File.dirname(__FILE__),'..','..','..','whisper.cpp')} .")
-system("cp #{File.join(File.dirname(__FILE__),'..','..','..','whisper.h')} .")
-system("cp #{File.join(File.dirname(__FILE__),'..','..','..','ggml.h')} .")
-system("cp #{File.join(File.dirname(__FILE__),'..','..','..','ggml.c')} .")
-system("cp #{File.join(File.dirname(__FILE__),'..','..','..','ggml-impl.h')} .")
-system("cp #{File.join(File.dirname(__FILE__),'..','..','..','ggml-aarch64.h')} .")
-system("cp #{File.join(File.dirname(__FILE__),'..','..','..','ggml-aarch64.c')} .")
-system("cp #{File.join(File.dirname(__FILE__),'..','..','..','ggml-alloc.h')} .")
-system("cp #{File.join(File.dirname(__FILE__),'..','..','..','ggml-alloc.c')} .")
-system("cp #{File.join(File.dirname(__FILE__),'..','..','..','ggml-backend-impl.h')} .")
-system("cp #{File.join(File.dirname(__FILE__),'..','..','..','ggml-backend.h')} .")
-system("cp #{File.join(File.dirname(__FILE__),'..','..','..','ggml-backend.cpp')} .")
-system("cp #{File.join(File.dirname(__FILE__),'..','..','..','ggml-common.h')} .")
-system("cp #{File.join(File.dirname(__FILE__),'..','..','..','ggml-quants.h')} .")
-system("cp #{File.join(File.dirname(__FILE__),'..','..','..','ggml-quants.c')} .")
-system("cp #{File.join(File.dirname(__FILE__),'..','..','..','examples','dr_wav.h')} .")
-
 
 # need to use c++ compiler flags
 $CXXFLAGS << ' -std=c++11'
+
+$LDFLAGS << ' -lstdc++'
+
 # Set to true when building binary gems
 if enable_config('static-stdlib', false)
   $LDFLAGS << ' -static-libgcc -static-libstdc++'
@@ -29,4 +15,180 @@ if enable_config('march-tune-native', false)
   $CXXFLAGS << ' -march=native -mtune=native'
 end
 
+if ENV['WHISPER_METAL']
+  $GGML_METAL ||= true
+  $DEPRECATE_WARNING ||= true
+end
+
+$UNAME_S = `uname -s`.chomp
+$UNAME_P = `uname -p`.chomp
+$UNAME_M = `uname -m`.chomp
+
+if $UNAME_S == 'Darwin'
+  unless ENV['GGML_NO_METAL']
+    $GGML_METAL ||= true
+  end
+  $GGML_NO_OPENMP ||= true
+end
+
+if $GGML_METAL
+  $GGML_METAL_EMBED_LIBRARY = true
+end
+
+$MK_CPPFLAGS = ''
+$MK_CFLAGS   = '-std=c11   -fPIC'
+$MK_CXXFLAGS = '-std=c++11 -fPIC'
+$MK_NVCCFLAGS = '-std=c++11'
+$MK_LDFLAGS = ''
+
+$OBJ_GGML = []
+$OBJ_WHISPER = []
+$OBJ_COMMON = []
+$OBJ_SDL = []
+
+$MK_CPPFLAGS << ' -D_XOPEN_SOURCE=600'
+
+if $UNAME_S == 'Linux'
+  $MK_CPPFLAGS << ' -D_GNU_SOURCE'
+end
+
+if $UNAME_S == 'Darwin'
+  $MK_CPPFLAGS << ' -D_DARWIN_C_SOURCE'
+end
+
+if ENV['WHISPER_DEBUG']
+  $MK_CFLAGS    << ' -O0 -g'
+  $MK_CXXFLAGS  << ' -O0 -g'
+  $MK_LDFLAGS   << ' -g'
+  $MK_NVCCFLAGS << ' -O0 -g'
+else
+  $MK_CPPFLAGS   << ' -DNDEBUG'
+  $MK_CFLAGS     << ' -O3'
+  $MK_CXXFLAGS   << ' -O3'
+  $MK_NVCCFLAGS  << ' -O3'
+end
+
+$WARN_FLAGS =
+  ' -Wall' <<
+  ' -Wextra' <<
+  ' -Wpedantic' <<
+  ' -Wcast-qual' <<
+  ' -Wno-unused-function'
+
+$MK_CFLAGS <<
+  $WARN_FLAGS <<
+  ' -Wshadow' <<
+  ' -Wstrict-prototypes' <<
+  ' -Wpointer-arith' <<
+  ' -Wmissing-prototypes' <<
+  ' -Werror=implicit-int' <<
+  ' -Werror=implicit-function-declaration'
+
+$MK_CXXFLAGS <<
+  $WARN_FLAGS <<
+  ' -Wmissing-declarations' <<
+  ' -Wmissing-noreturn'
+
+unless `#{cc_command} #{$LDFLAGS} -Wl,-v 2>&1`.chomp.include? 'dyld-1015.7'
+  $MK_CPPFLAGS << ' -DHAVE_BUGGY_APPLE_LINKER'
+end
+
+if %w[Linux Darwin FreeBSD NetBSD OpenBSD Haiku].include? $UNAME_S
+  $MK_CFLAGS   << ' -pthread'
+  $MK_CXXFLAGS << ' -pthread'
+end
+
+unless $_WIN32
+  $DSO_EXT = '.so'
+else
+  $DSO_EXT = '.dll'
+end
+
+unless ENV['RISCV']
+  if %w[x86_64 i686 amd64].include? $UNAME_M
+    $HOST_CXXFLAGS ||= ''
+
+    $MK_CFLAGS     << ' -march=native -mtune=native'
+    $HOST_CXXFLAGS << ' -march=native -mtune=native'
+  end
+
+  if $UNAME_M.match? /aarch64.*/
+    $MK_CFLAGS   << ' -mcpu=native'
+    $MK_CXXFLAGS << ' -mcpu=native'
+  end
+else
+  $MK_CFLAGS   << ' -march=rv64gcv -mabi=lp64d'
+  $MK_CXXFLAGS << ' -march=rv64gcv -mabi=lp64d'
+end
+
+unless ENV['GGML_NO_ACCELERATE']
+  if $UNAME_S == 'Darwin'
+    $MK_CPPFLAGS << ' -DGGML_USE_ACCELERATE -DGGML_USE_BLAS'
+    $MK_CPPFLAGS << ' -DACCELERATE_NEW_LAPACK'
+    $MK_CPPFLAGS << ' -DACCELERATE_LAPACK_ILP64'
+    $MK_LDFLAGS  << ' -framework Accelerate'
+    $OBJ_GGML    << 'ggml-blas.o'
+  end
+end
+
+if ENV['GGML_OPENBLAS']
+  $MK_CPPFLAGS << " -DGGML_USE_BLAS #{`pkg-config --cflags-only-I openblas`.chomp}"
+  $MK_CFLAGS   << " #{`pkg-config --cflags-only-other openblas)`.chomp}"
+  $MK_LDFLAGS  << " #{`pkg-config --libs openblas`}"
+  $OBJ_GGML    << 'ggml-blas.o'
+end
+
+if ENV['GGML_OPENBLAS64']
+  $MK_CPPFLAGS << " -DGGML_USE_BLAS #{`pkg-config --cflags-only-I openblas64`.chomp}"
+  $MK_CFLAGS   << " #{`pkg-config --cflags-only-other openblas64)`.chomp}"
+  $MK_LDFLAGS  << " #{`pkg-config --libs openblas64`}"
+  $OBJ_GGML    << 'ggml-blas.o'
+end
+
+if $GGML_METAL
+  $MK_CPPFLAGS << ' -DGGML_USE_METAL'
+  $MK_LDFLAGS  << ' -framework Foundation -framework Metal -framework MetalKit'
+  $OBJ_GGML    << 'ggml-metal.o'
+
+  if ENV['GGML_METAL_NDEBUG']
+    $MK_CPPFLAGS << ' -DGGML_METAL_NDEBUG'
+  end
+
+  if $GGML_METAL_EMBED_LIBRARY
+    $MK_CPPFLAGS << ' -DGGML_METAL_EMBED_LIBRARY'
+    $OBJ_GGML    << 'ggml-metal-embed.o'
+  end
+end
+
+$OBJ_GGML <<
+  'ggml.o' <<
+  'ggml-cpu.o' <<
+  'ggml-alloc.o' <<
+  'ggml-backend.o' <<
+  'ggml-quants.o' <<
+  'ggml-aarch64.o'
+
+$OBJ_WHISPER <<
+  'whisper.o'
+
+$objs = $OBJ_GGML + $OBJ_WHISPER + $OBJ_COMMON + $OBJ_SDL
+$objs << "ruby_whisper.o"
+
+$CPPFLAGS  = "#{$MK_CPPFLAGS} #{$CPPFLAGS}"
+$CFLAGS    = "#{$CPPFLAGS} #{$MK_CFLAGS} #{$GF_CFLAGS} #{$CFLAGS}"
+$BASE_CXXFLAGS = "#{$MK_CXXFLAGS} #{$CXXFLAGS}"
+$CXXFLAGS  = "#{$BASE_CXXFLAGS} #{$HOST_CXXFLAGS} #{$GF_CXXFLAGS} #{$CPPFLAGS}"
+$NVCCFLAGS = "#{$MK_NVCCFLAGS} #{$NVCCFLAGS}"
+$LDFLAGS   = "#{$MK_LDFLAGS} #{$LDFLAGS}"
+
 create_makefile('whisper')
+
+File.open 'Makefile', 'a' do |file|
+  file.puts 'include get-flags.mk'
+
+  if $GGML_METAL
+    if $GGML_METAL_EMBED_LIBRARY
+      file.puts 'include metal-embed.mk'
+    end
+  end
+end

bindings/ruby/ext/ggml-backend-impl.h

@@ -1,141 +0,0 @@
-#pragma once
-
-// ggml-backend internal header
-
-#include "ggml-backend.h"
-
-#ifdef  __cplusplus
-extern "C" {
-#endif
-
-    //
-    // Backend buffer
-    //
-
-    // buffer type
-    typedef void * ggml_backend_buffer_type_context_t;
-
-    struct ggml_backend_buffer_type_i {
-        const char *          (*GGML_CALL get_name)        (ggml_backend_buffer_type_t buft);
-        ggml_backend_buffer_t (*GGML_CALL alloc_buffer)    (ggml_backend_buffer_type_t buft, size_t size);
-        size_t                (*GGML_CALL get_alignment)   (ggml_backend_buffer_type_t buft); // tensor alignment
-        size_t                (*GGML_CALL get_max_size)    (ggml_backend_buffer_type_t buft); // allocation max size
-        size_t                (*GGML_CALL get_alloc_size)  (ggml_backend_buffer_type_t buft, const struct ggml_tensor * tensor); // data size needed to allocate the tensor, including padding
-        bool                  (*GGML_CALL supports_backend)(ggml_backend_buffer_type_t buft, ggml_backend_t backend); // check if the buffer type is usable by the backend
-        // check if tensor data is in host memory
-        // should be equivalent to supports_backend(buft, ggml_backend_cpu_init())
-        bool                  (*GGML_CALL is_host)         (ggml_backend_buffer_type_t buft);
-    };
-
-    struct ggml_backend_buffer_type {
-        struct ggml_backend_buffer_type_i  iface;
-        ggml_backend_buffer_type_context_t context;
-    };
-
-    // buffer
-    typedef void * ggml_backend_buffer_context_t;
-
-    struct ggml_backend_buffer_i {
-        const char * (*GGML_CALL get_name)   (ggml_backend_buffer_t buffer);
-        void         (*GGML_CALL free_buffer)(ggml_backend_buffer_t buffer);
-        void *       (*GGML_CALL get_base)   (ggml_backend_buffer_t buffer);
-        void         (*GGML_CALL init_tensor)(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
-        void         (*GGML_CALL set_tensor) (ggml_backend_buffer_t buffer,       struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
-        void         (*GGML_CALL get_tensor) (ggml_backend_buffer_t buffer, const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
-        bool         (*GGML_CALL cpy_tensor) (ggml_backend_buffer_t buffer, const struct ggml_tensor * src, struct ggml_tensor * dst); // dst is in the buffer, src may be in any buffer
-        void         (*GGML_CALL clear)      (ggml_backend_buffer_t buffer, uint8_t value);
-        void         (*GGML_CALL reset)      (ggml_backend_buffer_t buffer); // reset any internal state due to tensor initialization, such as tensor extras
-    };
-
-    struct ggml_backend_buffer {
-        struct ggml_backend_buffer_i  iface;
-        ggml_backend_buffer_type_t    buft;
-        ggml_backend_buffer_context_t context;
-        size_t size;
-        enum ggml_backend_buffer_usage usage;
-    };
-
-    GGML_CALL ggml_backend_buffer_t ggml_backend_buffer_init(
-                   ggml_backend_buffer_type_t      buft,
-            struct ggml_backend_buffer_i           iface,
-                   ggml_backend_buffer_context_t   context,
-                   size_t                          size);
-
-    // do not use directly, use ggml_backend_tensor_copy instead
-    bool ggml_backend_buffer_copy_tensor(const struct ggml_tensor * src, struct ggml_tensor * dst);
-
-    // buffer that contains a collection of buffers
-    GGML_CALL ggml_backend_buffer_t ggml_backend_multi_buffer_alloc_buffer(ggml_backend_buffer_t * buffers, size_t n_buffers);
-    GGML_CALL bool                  ggml_backend_buffer_is_multi_buffer(ggml_backend_buffer_t buffer);
-    GGML_CALL void                  ggml_backend_multi_buffer_set_usage(ggml_backend_buffer_t buffer, enum ggml_backend_buffer_usage usage);
-
-    //
-    // Backend
-    //
-
-    typedef void * ggml_backend_context_t;
-
-    struct ggml_backend_i {
-        const char * (*GGML_CALL get_name)(ggml_backend_t backend);
-
-        void (*GGML_CALL free)(ggml_backend_t backend);
-
-        // buffer allocation
-        ggml_backend_buffer_type_t (*GGML_CALL get_default_buffer_type)(ggml_backend_t backend);
-
-        // (optional) asynchronous tensor data access
-        void (*GGML_CALL set_tensor_async)(ggml_backend_t backend,       struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
-        void (*GGML_CALL get_tensor_async)(ggml_backend_t backend, const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
-        bool (*GGML_CALL cpy_tensor_async)(ggml_backend_t backend_src, ggml_backend_t backend_dst, const struct ggml_tensor * src, struct ggml_tensor * dst);
-
-        // (optional) complete all pending operations
-        void (*GGML_CALL synchronize)(ggml_backend_t backend);
-
-        // compute graph with a plan (not used currently)
-        ggml_backend_graph_plan_t (*GGML_CALL graph_plan_create) (ggml_backend_t backend, const struct ggml_cgraph * cgraph);
-        void                      (*GGML_CALL graph_plan_free)   (ggml_backend_t backend, ggml_backend_graph_plan_t plan);
-
-        // compute graph with a plan
-        enum ggml_status (*GGML_CALL graph_plan_compute)(ggml_backend_t backend, ggml_backend_graph_plan_t plan);
-        // compute graph without a plan (async)
-        enum ggml_status (*GGML_CALL graph_compute)     (ggml_backend_t backend, struct ggml_cgraph * cgraph);
-
-        // check if the backend supports an operation
-        bool (*GGML_CALL supports_op)(ggml_backend_t backend, const struct ggml_tensor * op);
-
-        // check if the backend wants to run an operation, even if the weights are allocated in a CPU buffer
-        // these should be expensive operations with large batch sizes that may benefit from running on this backend
-        // even if the weight has to be copied from the CPU temporarily
-        bool (*GGML_CALL offload_op)(ggml_backend_t backend, const struct ggml_tensor * op);
-
-        // (optional) event synchronization
-        ggml_backend_event_t (*GGML_CALL event_new)         (ggml_backend_t backend);
-        void                 (*GGML_CALL event_free)        (ggml_backend_event_t event);
-        void                 (*GGML_CALL event_record)      (ggml_backend_event_t event);
-        void                 (*GGML_CALL event_wait)        (ggml_backend_t backend, ggml_backend_event_t event);
-        void                 (*GGML_CALL event_synchronize) (ggml_backend_event_t event);
-    };
-
-    struct ggml_backend {
-        ggml_guid_t guid;
-
-        struct ggml_backend_i iface;
-        ggml_backend_context_t context;
-    };
-
-    struct ggml_backend_event {
-        ggml_backend_t backend;
-        void * context;
-    };
-
-    //
-    // Backend registry
-    //
-
-    typedef ggml_backend_t (*GGML_CALL ggml_backend_init_fn)(const char * params, void * user_data);
-
-    GGML_CALL void ggml_backend_register(const char * name, ggml_backend_init_fn init_fn, ggml_backend_buffer_type_t default_buffer_type, void * user_data);
-
-#ifdef  __cplusplus
-}
-#endif

bindings/ruby/ext/ggml-backend.h

@@ -1,233 +0,0 @@
-#pragma once
-
-#include "ggml.h"
-#include "ggml-alloc.h"
-
-#ifdef  __cplusplus
-extern "C" {
-#endif
-
-    typedef struct ggml_backend_buffer_type * ggml_backend_buffer_type_t;
-    typedef struct ggml_backend_buffer * ggml_backend_buffer_t;
-    typedef struct ggml_backend_event * ggml_backend_event_t;
-    typedef struct ggml_backend * ggml_backend_t;
-    typedef void * ggml_backend_graph_plan_t;
-
-    //
-    // Backend buffer
-    //
-
-    // buffer type
-    GGML_API           const char *          ggml_backend_buft_name            (ggml_backend_buffer_type_t buft);
-    GGML_API GGML_CALL ggml_backend_buffer_t ggml_backend_buft_alloc_buffer    (ggml_backend_buffer_type_t buft, size_t size);
-    GGML_API           size_t                ggml_backend_buft_get_alignment   (ggml_backend_buffer_type_t buft);
-    GGML_API           size_t                ggml_backend_buft_get_max_size    (ggml_backend_buffer_type_t buft);
-    GGML_API GGML_CALL size_t                ggml_backend_buft_get_alloc_size  (ggml_backend_buffer_type_t buft, struct ggml_tensor * tensor);
-    GGML_API           bool                  ggml_backend_buft_supports_backend(ggml_backend_buffer_type_t buft, ggml_backend_t backend);
-    GGML_API           bool                  ggml_backend_buft_is_host         (ggml_backend_buffer_type_t buft);
-
-    // buffer
-    enum ggml_backend_buffer_usage {
-        GGML_BACKEND_BUFFER_USAGE_ANY = 0,
-        GGML_BACKEND_BUFFER_USAGE_WEIGHTS = 1,
-    };
-
-    GGML_API           const char *               ggml_backend_buffer_name          (ggml_backend_buffer_t buffer);
-    GGML_API           void                       ggml_backend_buffer_free          (ggml_backend_buffer_t buffer);
-    GGML_API           void *                     ggml_backend_buffer_get_base      (ggml_backend_buffer_t buffer);
-    GGML_API           size_t                     ggml_backend_buffer_get_size      (ggml_backend_buffer_t buffer);
-    GGML_API GGML_CALL void                       ggml_backend_buffer_init_tensor   (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
-    GGML_API           size_t                     ggml_backend_buffer_get_alignment (ggml_backend_buffer_t buffer);
-    GGML_API           size_t                     ggml_backend_buffer_get_max_size  (ggml_backend_buffer_t buffer);
-    GGML_API           size_t                     ggml_backend_buffer_get_alloc_size(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
-    GGML_API           void                       ggml_backend_buffer_clear         (ggml_backend_buffer_t buffer, uint8_t value);
-    GGML_API           bool                       ggml_backend_buffer_is_host       (ggml_backend_buffer_t buffer);
-    GGML_API           void                       ggml_backend_buffer_set_usage     (ggml_backend_buffer_t buffer, enum ggml_backend_buffer_usage usage);
-    GGML_API           ggml_backend_buffer_type_t ggml_backend_buffer_get_type      (ggml_backend_buffer_t buffer);
-    GGML_API           void                       ggml_backend_buffer_reset         (ggml_backend_buffer_t buffer);
-
-    //
-    // Backend
-    //
-
-    GGML_API ggml_guid_t  ggml_backend_guid(ggml_backend_t backend);
-    GGML_API const char * ggml_backend_name(ggml_backend_t backend);
-    GGML_API void         ggml_backend_free(ggml_backend_t backend);
-
-    GGML_API ggml_backend_buffer_type_t ggml_backend_get_default_buffer_type(ggml_backend_t backend);
-    GGML_API ggml_backend_buffer_t      ggml_backend_alloc_buffer(ggml_backend_t backend, size_t size);
-    GGML_API size_t                     ggml_backend_get_alignment(ggml_backend_t backend);
-    GGML_API size_t                     ggml_backend_get_max_size(ggml_backend_t backend);
-
-    GGML_API void ggml_backend_tensor_set_async(ggml_backend_t backend,       struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
-    GGML_API void ggml_backend_tensor_get_async(ggml_backend_t backend, const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
-
-    GGML_API GGML_CALL void ggml_backend_tensor_set(      struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
-    GGML_API GGML_CALL void ggml_backend_tensor_get(const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
-
-    GGML_API void ggml_backend_synchronize(ggml_backend_t backend);
-
-    GGML_API ggml_backend_graph_plan_t ggml_backend_graph_plan_create(ggml_backend_t backend, struct ggml_cgraph * cgraph);
-    GGML_API void                      ggml_backend_graph_plan_free  (ggml_backend_t backend, ggml_backend_graph_plan_t plan);
-
-    GGML_API enum ggml_status ggml_backend_graph_plan_compute (ggml_backend_t backend, ggml_backend_graph_plan_t plan);
-    GGML_API enum ggml_status ggml_backend_graph_compute      (ggml_backend_t backend, struct ggml_cgraph * cgraph);
-    GGML_API enum ggml_status ggml_backend_graph_compute_async(ggml_backend_t backend, struct ggml_cgraph * cgraph);
-    GGML_API bool ggml_backend_supports_op(ggml_backend_t backend, const struct ggml_tensor * op);
-    GGML_API bool ggml_backend_offload_op(ggml_backend_t backend, const struct ggml_tensor * op);
-
-    // tensor copy between different backends
-    GGML_API void ggml_backend_tensor_copy(struct ggml_tensor * src, struct ggml_tensor * dst);
-
-    // asynchronous copy
-    // the copy is performed after all the currently queued operations in backend_src
-    // backend_dst will wait for the copy to complete before performing other operations
-    // automatic fallback to sync copy if async is not supported
-    GGML_API void ggml_backend_tensor_copy_async(ggml_backend_t backend_src, ggml_backend_t backend_dst, struct ggml_tensor * src, struct ggml_tensor * dst);
-
-    // events
-    GGML_API ggml_backend_event_t   ggml_backend_event_new        (ggml_backend_t backend);
-    GGML_API void                   ggml_backend_event_free       (ggml_backend_event_t event);
-    GGML_API void                   ggml_backend_event_record     (ggml_backend_event_t event);
-    GGML_API void                   ggml_backend_event_synchronize(ggml_backend_event_t event);
-    GGML_API void                   ggml_backend_event_wait       (ggml_backend_t backend, ggml_backend_event_t event); // wait async on event
-
-    //
-    // CPU backend
-    //
-
-    GGML_API ggml_backend_t ggml_backend_cpu_init(void);
-
-    GGML_API GGML_CALL bool ggml_backend_is_cpu                (ggml_backend_t backend);
-    GGML_API           void ggml_backend_cpu_set_n_threads     (ggml_backend_t backend_cpu, int n_threads);
-    GGML_API           void ggml_backend_cpu_set_abort_callback(ggml_backend_t backend_cpu, ggml_abort_callback abort_callback, void * abort_callback_data);
-
-    // Create a backend buffer from an existing pointer
-    GGML_API GGML_CALL ggml_backend_buffer_t ggml_backend_cpu_buffer_from_ptr(void * ptr, size_t size);
-
-    GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_cpu_buffer_type(void);
-
-#ifdef GGML_USE_CPU_HBM
-    GGML_API ggml_backend_buffer_type_t ggml_backend_cpu_hbm_buffer_type(void);
-#endif
-
-    //
-    // Backend registry
-    //
-
-    // The backend registry is a registry of all the available backends, and allows initializing backends in a generic way
-
-    GGML_API size_t                     ggml_backend_reg_get_count(void);
-    GGML_API size_t                     ggml_backend_reg_find_by_name(const char * name);
-    GGML_API ggml_backend_t             ggml_backend_reg_init_backend_from_str(const char * backend_str); // str is name[:params]
-    GGML_API const char *               ggml_backend_reg_get_name(size_t i);
-    GGML_API ggml_backend_t             ggml_backend_reg_init_backend(size_t i, const char * params); // params is backend-specific
-    GGML_API ggml_backend_buffer_type_t ggml_backend_reg_get_default_buffer_type(size_t i);
-    GGML_API ggml_backend_buffer_t      ggml_backend_reg_alloc_buffer(size_t i, size_t size);
-
-    //
-    // Backend scheduler
-    //
-
-    // The backend scheduler allows for multiple backends to be used together
-    // Handles compute buffer allocation, assignment of tensors to backends, and copying of tensors between backends
-    // The backends are selected based on:
-    // - the backend that supports the operation
-    // - the location of the pre-allocated tensors (e.g. the weights)
-    /*
-      Example usage:
-
-        // operations that use tensors allocated in a buffer with USAGE_WEIGHTS will be assigned
-        // preferrably to run on the same backend as the buffer
-        ggml_backend_buffer_set_usage(buf_weights, GGML_BACKEND_BUFFER_USAGE_WEIGHTS);
-
-        sched = ggml_backend_sched_new({backend_gpu, backend_gpu2, backend_cpu}, NULL, num_backends, GGML_DEFAULT_GRAPH_SIZE, false);
-
-        // initialize buffers from a max size graph (optional)
-        reserve_graph = build_graph(sched, max_batch_size);
-
-        // manually assign nodes to a backend (optional, should not be needed in most cases)
-        struct ggml_tensor * node = ggml_mul_mat(ctx, ...);
-        ggml_backend_sched_set_tensor_backend(sched, node, backend_gpu);
-
-        ggml_backend_sched_reserve(sched, reserve_graph);
-
-        // compute
-        graph = build_graph(sched);
-        ggml_backend_sched_graph_compute(sched, graph);
-
-        // if there are graph inputs:
-        ggml_backend_sched_reset(sched);
-        ggml_backend_sched_alloc_graph(sched, graph);
-        ggml_backend_tensor_set(input_tensor, ...);
-        ggml_backend_sched_graph_compute(sched, graph);
-    }
-    */
-
-    struct ggml_backend_sched;
-    typedef struct ggml_backend_sched * ggml_backend_sched_t;
-
-    // when ask == true, the scheduler wants to know if the user wants to observe this node
-    // this allows the scheduler to batch nodes together in order to evaluate them in a single call
-    //
-    // when ask == false, the scheduler is passing the node tensor to the user for observation
-    // if the user returns false, the scheduler will cancel the graph compute
-    //
-    typedef bool (*ggml_backend_sched_eval_callback)(struct ggml_tensor * t, bool ask, void * user_data);
-
-    // Initialize a backend scheduler
-    GGML_API ggml_backend_sched_t ggml_backend_sched_new(ggml_backend_t * backends, ggml_backend_buffer_type_t * bufts, int n_backends, size_t graph_size, bool parallel);
-    GGML_API void                 ggml_backend_sched_free(ggml_backend_sched_t sched);
-
-    // Initialize backend buffers from a measure graph
-    GGML_API bool                 ggml_backend_sched_reserve(ggml_backend_sched_t sched, struct ggml_cgraph * measure_graph);
-
-    // Get the number of splits of the last graph
-    GGML_API int                  ggml_backend_sched_get_n_splits(ggml_backend_sched_t sched);
-    GGML_API int                  ggml_backend_sched_get_n_copies(ggml_backend_sched_t sched);
-
-    GGML_API size_t               ggml_backend_sched_get_buffer_size(ggml_backend_sched_t sched, ggml_backend_t backend);
-
-    GGML_API void                 ggml_backend_sched_set_tensor_backend(ggml_backend_sched_t sched, struct ggml_tensor * node, ggml_backend_t backend);
-    GGML_API ggml_backend_t       ggml_backend_sched_get_tensor_backend(ggml_backend_sched_t sched, struct ggml_tensor * node);
-
-    // Allocate and compute graph on the backend scheduler
-    GGML_API bool                 ggml_backend_sched_alloc_graph(ggml_backend_sched_t sched, struct ggml_cgraph * graph);
-    GGML_API enum ggml_status     ggml_backend_sched_graph_compute(ggml_backend_sched_t sched, struct ggml_cgraph * graph);
-    GGML_API enum ggml_status     ggml_backend_sched_graph_compute_async(ggml_backend_sched_t sched, struct ggml_cgraph * graph);
-    GGML_API void                 ggml_backend_sched_synchronize(ggml_backend_sched_t sched);
-
-    // Reset all assignments and allocators - must be called before changing the node backends
-    GGML_API void                 ggml_backend_sched_reset(ggml_backend_sched_t sched);
-
-    // Set a callback to be called for each resulting node during graph compute
-    GGML_API void                 ggml_backend_sched_set_eval_callback(ggml_backend_sched_t sched, ggml_backend_sched_eval_callback callback, void * user_data);
-
-    //
-    // Utils
-    //
-
-    struct ggml_backend_graph_copy {
-        ggml_backend_buffer_t buffer;
-        struct ggml_context * ctx_allocated;
-        struct ggml_context * ctx_unallocated;
-        struct ggml_cgraph * graph;
-    };
-
-    // Copy a graph to a different backend
-    GGML_API struct ggml_backend_graph_copy ggml_backend_graph_copy(ggml_backend_t backend, struct ggml_cgraph * graph);
-    GGML_API void                           ggml_backend_graph_copy_free(struct ggml_backend_graph_copy copy);
-
-    typedef bool (*GGML_CALL ggml_backend_eval_callback)(int node_index, struct ggml_tensor * t1, struct ggml_tensor * t2, void * user_data);
-
-    // Compare the output of two backends
-    GGML_API bool ggml_backend_compare_graph_backend(ggml_backend_t backend1, ggml_backend_t backend2, struct ggml_cgraph * graph, ggml_backend_eval_callback callback, void * user_data);
-
-    // Tensor initialization
-    GGML_API void ggml_backend_tensor_alloc(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, void * addr);
-    GGML_API void ggml_backend_view_init(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
-
-
-#ifdef  __cplusplus
-}
-#endif

bindings/ruby/ext/ggml-cuda.h

@@ -1,43 +0,0 @@
-#pragma once
-
-#include "ggml.h"
-#include "ggml-backend.h"
-
-#ifdef GGML_USE_HIPBLAS
-#define GGML_CUDA_NAME "ROCm"
-#define GGML_CUBLAS_NAME "hipBLAS"
-#else
-#define GGML_CUDA_NAME "CUDA"
-#define GGML_CUBLAS_NAME "cuBLAS"
-#endif
-
-#ifdef  __cplusplus
-extern "C" {
-#endif
-
-#define GGML_CUDA_MAX_DEVICES       16
-
-// backend API
-GGML_API GGML_CALL ggml_backend_t ggml_backend_cuda_init(int device);
-
-GGML_API GGML_CALL bool ggml_backend_is_cuda(ggml_backend_t backend);
-
-// device buffer
-GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_cuda_buffer_type(int device);
-
-// split tensor buffer that splits matrices by rows across multiple devices
-GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_cuda_split_buffer_type(const float * tensor_split);
-
-// pinned host buffer for use with the CPU backend for faster copies between CPU and GPU
-GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_cuda_host_buffer_type(void);
-
-GGML_API GGML_CALL int  ggml_backend_cuda_get_device_count(void);
-GGML_API GGML_CALL void ggml_backend_cuda_get_device_description(int device, char * description, size_t description_size);
-GGML_API GGML_CALL void ggml_backend_cuda_get_device_memory(int device, size_t * free, size_t * total);
-
-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);
-
-#ifdef  __cplusplus
-}
-#endif

bindings/ruby/ext/ggml-impl.h

@@ -1,272 +0,0 @@
-#pragma once
-
-#include "ggml.h"
-
-// GGML internal header
-
-#include <assert.h>
-#include <stdlib.h> // load `stdlib.h` before other headers to work around MinGW bug: https://sourceforge.net/p/mingw-w64/bugs/192/
-#include <stddef.h>
-#include <stdbool.h>
-#include <string.h> // memcpy
-#include <math.h>   // fabsf
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-// static_assert should be a #define, but if it's not,
-// fall back to the _Static_assert C11 keyword.
-// if C99 - static_assert is noop
-// ref: https://stackoverflow.com/a/53923785/4039976
-#ifndef __cplusplus
-#ifndef static_assert
-#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201100L)
-#define static_assert(cond, msg) _Static_assert(cond, msg)
-#else
-#define static_assert(cond, msg) struct global_scope_noop_trick
-#endif
-#endif
-#endif
-
-// __FMA__ and __F16C__ are not defined in MSVC, however they are implied with AVX2/AVX512
-#if defined(_MSC_VER) && (defined(__AVX2__) || defined(__AVX512F__))
-#ifndef __FMA__
-#define __FMA__
-#endif
-#ifndef __F16C__
-#define __F16C__
-#endif
-#endif
-
-// __SSE3__ and __SSSE3__ are not defined in MSVC, but SSE3/SSSE3 are present when AVX/AVX2/AVX512 are available
-#if defined(_MSC_VER) && (defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__))
-#ifndef __SSE3__
-#define __SSE3__
-#endif
-#ifndef __SSSE3__
-#define __SSSE3__
-#endif
-#endif
-
-// 16-bit float
-// on Arm, we use __fp16
-// on x86, we use uint16_t
-#if defined(__ARM_NEON) && !defined(_MSC_VER)
-
-// if YCM cannot find <arm_neon.h>, make a symbolic link to it, for example:
-//
-//   $ ln -sfn /Library/Developer/CommandLineTools/usr/lib/clang/13.1.6/include/arm_neon.h ./src/
-//
-#include <arm_neon.h>
-
-typedef __fp16 ggml_fp16_internal_t;
-
-#define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
-#define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x)
-
-#define GGML_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
-
-static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
-    ggml_fp16_internal_t tmp;
-    memcpy(&tmp, &h, sizeof(ggml_fp16_t));
-    return (float)tmp;
-}
-
-static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
-    ggml_fp16_t res;
-    ggml_fp16_internal_t tmp = f;
-    memcpy(&res, &tmp, sizeof(ggml_fp16_t));
-    return res;
-}
-
-#else
-
-typedef uint16_t ggml_fp16_internal_t;
-
-#ifdef __wasm_simd128__
-#include <wasm_simd128.h>
-#else
-#ifdef __POWER9_VECTOR__
-#include <altivec.h>
-#undef bool
-#define bool _Bool
-#else
-#if defined(_MSC_VER) || defined(__MINGW32__)
-#include <intrin.h>
-#else
-#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__) || defined(__SSSE3__) || defined(__SSE3__)
-#if !defined(__riscv)
-#include <immintrin.h>
-#endif
-#endif
-#endif
-#endif
-#endif
-
-#ifdef __riscv_v_intrinsic
-#include <riscv_vector.h>
-#endif
-
-#ifdef __F16C__
-
-#ifdef _MSC_VER
-#define GGML_COMPUTE_FP16_TO_FP32(x) _mm_cvtss_f32(_mm_cvtph_ps(_mm_cvtsi32_si128(x)))
-#define GGML_COMPUTE_FP32_TO_FP16(x) _mm_extract_epi16(_mm_cvtps_ph(_mm_set_ss(x), 0), 0)
-#else
-#define GGML_COMPUTE_FP16_TO_FP32(x) _cvtsh_ss(x)
-#define GGML_COMPUTE_FP32_TO_FP16(x) _cvtss_sh(x, 0)
-#endif
-
-#elif defined(__POWER9_VECTOR__)
-
-#define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
-#define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x)
-/* the inline asm below is about 12% faster than the lookup method */
-#define GGML_FP16_TO_FP32(x) GGML_COMPUTE_FP16_TO_FP32(x)
-#define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x)
-
-static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
-    register float f;
-    register double d;
-    __asm__(
-        "mtfprd %0,%2\n"
-        "xscvhpdp %0,%0\n"
-        "frsp %1,%0\n" :
-        /* temp */ "=d"(d),
-        /* out */  "=f"(f):
-        /* in */   "r"(h));
-    return f;
-}
-
-static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
-    register double d;
-    register ggml_fp16_t r;
-    __asm__( /* xscvdphp can work on double or single precision */
-        "xscvdphp %0,%2\n"
-        "mffprd %1,%0\n" :
-        /* temp */ "=d"(d),
-        /* out */  "=r"(r):
-        /* in */   "f"(f));
-    return r;
-}
-
-#else
-
-// FP16 <-> FP32
-// ref: https://github.com/Maratyszcza/FP16
-
-static inline float fp32_from_bits(uint32_t w) {
-    union {
-        uint32_t as_bits;
-        float as_value;
-    } fp32;
-    fp32.as_bits = w;
-    return fp32.as_value;
-}
-
-static inline uint32_t fp32_to_bits(float f) {
-    union {
-        float as_value;
-        uint32_t as_bits;
-    } fp32;
-    fp32.as_value = f;
-    return fp32.as_bits;
-}
-
-static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
-    const uint32_t w = (uint32_t) h << 16;
-    const uint32_t sign = w & UINT32_C(0x80000000);
-    const uint32_t two_w = w + w;
-
-    const uint32_t exp_offset = UINT32_C(0xE0) << 23;
-#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) || defined(__GNUC__) && !defined(__STRICT_ANSI__)
-    const float exp_scale = 0x1.0p-112f;
-#else
-    const float exp_scale = fp32_from_bits(UINT32_C(0x7800000));
-#endif
-    const float normalized_value = fp32_from_bits((two_w >> 4) + exp_offset) * exp_scale;
-
-    const uint32_t magic_mask = UINT32_C(126) << 23;
-    const float magic_bias = 0.5f;
-    const float denormalized_value = fp32_from_bits((two_w >> 17) | magic_mask) - magic_bias;
-
-    const uint32_t denormalized_cutoff = UINT32_C(1) << 27;
-    const uint32_t result = sign |
-        (two_w < denormalized_cutoff ? fp32_to_bits(denormalized_value) : fp32_to_bits(normalized_value));
-    return fp32_from_bits(result);
-}
-
-static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
-#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) || defined(__GNUC__) && !defined(__STRICT_ANSI__)
-    const float scale_to_inf = 0x1.0p+112f;
-    const float scale_to_zero = 0x1.0p-110f;
-#else
-    const float scale_to_inf = fp32_from_bits(UINT32_C(0x77800000));
-    const float scale_to_zero = fp32_from_bits(UINT32_C(0x08800000));
-#endif
-    float base = (fabsf(f) * scale_to_inf) * scale_to_zero;
-
-    const uint32_t w = fp32_to_bits(f);
-    const uint32_t shl1_w = w + w;
-    const uint32_t sign = w & UINT32_C(0x80000000);
-    uint32_t bias = shl1_w & UINT32_C(0xFF000000);
-    if (bias < UINT32_C(0x71000000)) {
-        bias = UINT32_C(0x71000000);
-    }
-
-    base = fp32_from_bits((bias >> 1) + UINT32_C(0x07800000)) + base;
-    const uint32_t bits = fp32_to_bits(base);
-    const uint32_t exp_bits = (bits >> 13) & UINT32_C(0x00007C00);
-    const uint32_t mantissa_bits = bits & UINT32_C(0x00000FFF);
-    const uint32_t nonsign = exp_bits + mantissa_bits;
-    return (sign >> 16) | (shl1_w > UINT32_C(0xFF000000) ? UINT16_C(0x7E00) : nonsign);
-}
-
-#define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
-#define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x)
-
-#endif // __F16C__
-
-#endif // __ARM_NEON
-
-// precomputed f32 table for f16 (256 KB)
-// defined in ggml.c, initialized in ggml_init()
-extern float ggml_table_f32_f16[1 << 16];
-
-// On ARM NEON, it's quicker to directly convert x -> x instead of calling into ggml_lookup_fp16_to_fp32,
-// so we define GGML_FP16_TO_FP32 and GGML_FP32_TO_FP16 elsewhere for NEON.
-// This is also true for POWER9.
-#if !defined(GGML_FP16_TO_FP32)
-inline static float ggml_lookup_fp16_to_fp32(ggml_fp16_t f) {
-    uint16_t s;
-    memcpy(&s, &f, sizeof(uint16_t));
-    return ggml_table_f32_f16[s];
-}
-
-#define GGML_FP16_TO_FP32(x) ggml_lookup_fp16_to_fp32(x)
-#endif
-
-#if !defined(GGML_FP32_TO_FP16)
-#define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x)
-#endif
-
-#define GGML_HASHTABLE_FULL ((size_t)-1)
-#define GGML_HASHTABLE_ALREADY_EXISTS ((size_t)-2)
-
-struct ggml_hash_set ggml_hash_set_new(size_t size);
-
-bool   ggml_hash_contains      (const struct ggml_hash_set hash_set, struct ggml_tensor * key);
-
-// returns GGML_HASHTABLE_FULL if table is full, otherwise the current index of the key or where it should be inserted
-size_t ggml_hash_find          (const struct ggml_hash_set hash_set, struct ggml_tensor * key);
-
-// returns GGML_HASHTABLE_ALREADY_EXISTS if key already exists, index otherwise, asserts if table is full
-size_t ggml_hash_insert        (      struct ggml_hash_set hash_set, struct ggml_tensor * key);
-
-// return index, asserts if table is full
-size_t ggml_hash_find_or_insert(      struct ggml_hash_set hash_set, struct ggml_tensor * key);
-
-#ifdef __cplusplus
-}
-#endif

bindings/ruby/ext/ggml-kompute.h

@@ -1,46 +0,0 @@
-#pragma once
-
-#include "ggml.h"
-#include "ggml-backend.h"
-
-#include <stdbool.h>
-#include <stddef.h>
-#include <stdint.h>
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-struct ggml_vk_device {
-    int index;
-    int type; // same as VkPhysicalDeviceType
-    size_t heapSize;
-    const char * name;
-    const char * vendor;
-    int subgroupSize;
-    uint64_t bufferAlignment;
-    uint64_t maxAlloc;
-};
-
-struct ggml_vk_device * ggml_vk_available_devices(size_t memoryRequired, size_t * count);
-bool ggml_vk_get_device(struct ggml_vk_device * device, size_t memoryRequired, const char * name);
-bool ggml_vk_has_vulkan(void);
-bool ggml_vk_has_device(void);
-struct ggml_vk_device ggml_vk_current_device(void);
-
-//
-// backend API
-//
-
-// forward declaration
-typedef struct ggml_backend * ggml_backend_t;
-
-GGML_API ggml_backend_t ggml_backend_kompute_init(int device);
-
-GGML_API bool ggml_backend_is_kompute(ggml_backend_t backend);
-
-GGML_API ggml_backend_buffer_type_t ggml_backend_kompute_buffer_type(int device);
-
-#ifdef __cplusplus
-}
-#endif

bindings/ruby/ext/ggml-metal.h

@@ -1,66 +0,0 @@
-// An interface allowing to compute ggml_cgraph with Metal
-//
-// This is a fully functional interface that extends ggml with GPU support for Apple devices.
-// A similar interface can be created for other GPU backends (e.g. Vulkan, CUDA, OpenCL, etc.)
-//
-// How it works?
-//
-// As long as your program can create and evaluate a ggml_cgraph on the CPU, you can use this
-// interface to evaluate the same graph on the GPU. Instead of using ggml_graph_compute(), you
-// use ggml_metal_graph_compute() (or ggml_vulkan_graph_compute(), etc.)
-//
-// You only need to make sure that all memory buffers that you used during the graph creation
-// are mapped to the device memory with the ggml_metal_add_buffer() function. This mapping is
-// used during the graph evaluation to determine the arguments of the compute kernels.
-//
-// Synchronization between device and host memory (for example for input and output tensors)
-// is done with the ggml_metal_set_tensor() and ggml_metal_get_tensor() functions.
-//
-
-#pragma once
-
-#include "ggml.h"
-#include "ggml-backend.h"
-
-#include <stddef.h>
-#include <stdbool.h>
-
-// max memory buffers that can be mapped to the device
-#define GGML_METAL_MAX_BUFFERS 64
-
-struct ggml_tensor;
-struct ggml_cgraph;
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-//
-// backend API
-// user-code should use only these functions
-//
-
-GGML_API void ggml_backend_metal_log_set_callback(ggml_log_callback log_callback, void * user_data);
-
-GGML_API ggml_backend_t ggml_backend_metal_init(void);
-
-GGML_API bool ggml_backend_is_metal(ggml_backend_t backend);
-
-GGML_API GGML_CALL ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void * data, size_t size, size_t max_size);
-
-GGML_API void ggml_backend_metal_set_n_cb(ggml_backend_t backend, int n_cb);
-
-GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_metal_buffer_type(void);
-
-// helper to check if the device supports a specific family
-// ideally, the user code should be doing these checks
-// ref: https://developer.apple.com/metal/Metal-Feature-Set-Tables.pdf
-GGML_API bool ggml_backend_metal_supports_family(ggml_backend_t backend, int family);
-
-// capture all command buffers committed the next time `ggml_backend_graph_compute` is called
-GGML_API void ggml_backend_metal_capture_next_compute(ggml_backend_t backend);
-
-#ifdef __cplusplus
-}
-#endif
-

bindings/ruby/ext/ggml-opencl.h

@@ -1,36 +0,0 @@
-#pragma once
-
-#include "ggml.h"
-#include "ggml-backend.h"
-
-#ifdef  __cplusplus
-extern "C" {
-#endif
-
-GGML_API void ggml_cl_init(void);
-
-GGML_API void   ggml_cl_mul(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
-GGML_API void   ggml_cl_add(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
-GGML_API bool   ggml_cl_can_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, const struct ggml_tensor * dst);
-GGML_API size_t ggml_cl_mul_mat_get_wsize(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
-GGML_API void   ggml_cl_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst, void * wdata, size_t wsize);
-
-// GGML_API void * ggml_cl_host_malloc(size_t size);
-// GGML_API void   ggml_cl_host_free(void * ptr);
-
-GGML_API void ggml_cl_free_data(const struct ggml_tensor* tensor);
-
-GGML_API void ggml_cl_transform_tensor(void * data, struct ggml_tensor * tensor);
-
-// backend API
-
-// GGML_API ggml_backend_t ggml_backend_opencl_init(void);
-
-// GGML_API bool ggml_backend_is_opencl(ggml_backend_t backend);
-
-GGML_API ggml_backend_buffer_type_t ggml_backend_opencl_buffer_type(void);
-// GGML_API ggml_backend_buffer_type_t ggml_backend_opencl_host_buffer_type(void);
-
-#ifdef  __cplusplus
-}
-#endif

bindings/ruby/ext/ggml-quants.h

@@ -1,133 +0,0 @@
-#pragma once
-
-#define GGML_COMMON_DECL_C
-#include "ggml-common.h"
-
-#include "ggml.h"
-
-// GGML internal header
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-// Quantization
-void quantize_row_q4_0_reference(const float * GGML_RESTRICT x, block_q4_0 * GGML_RESTRICT y, int64_t k);
-void quantize_row_q4_1_reference(const float * GGML_RESTRICT x, block_q4_1 * GGML_RESTRICT y, int64_t k);
-void quantize_row_q5_0_reference(const float * GGML_RESTRICT x, block_q5_0 * GGML_RESTRICT y, int64_t k);
-void quantize_row_q5_1_reference(const float * GGML_RESTRICT x, block_q5_1 * GGML_RESTRICT y, int64_t k);
-void quantize_row_q8_0_reference(const float * GGML_RESTRICT x, block_q8_0 * GGML_RESTRICT y, int64_t k);
-void quantize_row_q8_1_reference(const float * GGML_RESTRICT x, block_q8_1 * GGML_RESTRICT y, int64_t k);
-
-void quantize_row_q2_K_reference(const float * GGML_RESTRICT x, block_q2_K * GGML_RESTRICT y, int64_t k);
-void quantize_row_q3_K_reference(const float * GGML_RESTRICT x, block_q3_K * GGML_RESTRICT y, int64_t k);
-void quantize_row_q4_K_reference(const float * GGML_RESTRICT x, block_q4_K * GGML_RESTRICT y, int64_t k);
-void quantize_row_q5_K_reference(const float * GGML_RESTRICT x, block_q5_K * GGML_RESTRICT y, int64_t k);
-void quantize_row_q6_K_reference(const float * GGML_RESTRICT x, block_q6_K * GGML_RESTRICT y, int64_t k);
-void quantize_row_q8_K_reference(const float * GGML_RESTRICT x, block_q8_K * GGML_RESTRICT y, int64_t k);
-
-void quantize_row_iq3_xxs_reference(const float * GGML_RESTRICT x, block_iq3_xxs * GGML_RESTRICT y, int64_t k);
-void quantize_row_iq4_nl_reference (const float * GGML_RESTRICT x, block_iq4_nl  * GGML_RESTRICT y, int64_t k);
-void quantize_row_iq4_xs_reference (const float * GGML_RESTRICT x, block_iq4_xs  * GGML_RESTRICT y, int64_t k);
-void quantize_row_iq3_s_reference  (const float * GGML_RESTRICT x, block_iq3_s   * GGML_RESTRICT y, int64_t k);
-void quantize_row_iq2_s_reference  (const float * GGML_RESTRICT x, block_iq2_s   * GGML_RESTRICT y, int64_t k);
-
-void quantize_row_q4_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_row_q4_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_row_q5_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_row_q5_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_row_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-
-void quantize_row_q2_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_row_q3_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_row_q4_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_row_q5_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_row_q6_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_row_q8_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-
-void quantize_row_iq3_xxs(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_row_iq4_nl (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_row_iq4_xs (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_row_iq3_s  (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_row_iq2_s  (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-
-// Dequantization
-void dequantize_row_q4_0(const block_q4_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-void dequantize_row_q4_1(const block_q4_1 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-void dequantize_row_q5_0(const block_q5_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-void dequantize_row_q5_1(const block_q5_1 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-void dequantize_row_q8_0(const block_q8_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-//void dequantize_row_q8_1(const block_q8_1 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-
-void dequantize_row_q2_K(const block_q2_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-void dequantize_row_q3_K(const block_q3_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-void dequantize_row_q4_K(const block_q4_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-void dequantize_row_q5_K(const block_q5_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-void dequantize_row_q6_K(const block_q6_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-void dequantize_row_q8_K(const block_q8_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-
-void dequantize_row_iq2_xxs(const block_iq2_xxs * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-void dequantize_row_iq2_xs (const block_iq2_xs  * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-void dequantize_row_iq2_s  (const block_iq2_s   * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-void dequantize_row_iq3_xxs(const block_iq3_xxs * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-void dequantize_row_iq1_s  (const block_iq1_s   * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-void dequantize_row_iq1_m  (const block_iq1_m   * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-void dequantize_row_iq4_nl (const block_iq4_nl  * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-void dequantize_row_iq4_xs (const block_iq4_xs  * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-void dequantize_row_iq3_s  (const block_iq3_s   * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-
-// Dot product
-void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-
-void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-
-void ggml_vec_dot_iq2_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_iq2_xs_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_iq2_s_q8_K  (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_iq1_s_q8_K  (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_iq1_m_q8_K  (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_iq4_nl_q8_0 (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_iq4_xs_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_iq3_s_q8_K  (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-
-// Quantization utilizing an importance matrix (a.k.a. "Activation aWare Quantization")
-size_t quantize_iq2_xxs(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_iq2_xs (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_iq2_s  (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_iq3_xxs(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_iq1_s  (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_iq1_m  (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_iq4_nl (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_iq4_xs (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_iq3_s  (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-
-size_t quantize_q2_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_q3_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_q4_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_q5_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_q6_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_q4_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_q4_1(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_q5_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_q5_1(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_q8_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-
-void iq2xs_init_impl(enum ggml_type type);
-void iq2xs_free_impl(enum ggml_type type);
-void iq3xs_init_impl(int grid_size);
-void iq3xs_free_impl(int grid_size);
-
-#ifdef __cplusplus
-}
-#endif
-

bindings/ruby/ext/ggml-sycl.h

@@ -1,49 +0,0 @@
-//
-//  MIT license
-//  Copyright (C) 2024 Intel Corporation
-//  SPDX-License-Identifier: MIT
-//
-
-#pragma once
-
-#include "ggml.h"
-#include "ggml-backend.h"
-
-#ifdef  __cplusplus
-extern "C" {
-#endif
-
-#define GGML_SYCL_MAX_DEVICES       48
-#define GGML_SYCL_NAME "SYCL"
-
-// backend API
-GGML_API ggml_backend_t ggml_backend_sycl_init(int device);
-
-// devide buffer
-GGML_API ggml_backend_buffer_type_t ggml_backend_sycl_buffer_type(int device);
-
-// split tensor buffer that splits matrices by rows across multiple devices
-GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_sycl_split_buffer_type(const float * tensor_split);
-
-// pinned host buffer for use with the CPU backend for faster copies between CPU and GPU
-GGML_API ggml_backend_buffer_type_t ggml_backend_sycl_host_buffer_type(void);
-
-GGML_API void   ggml_backend_sycl_print_sycl_devices(void);
-GGML_API GGML_CALL void   ggml_sycl_get_gpu_list(int *id_list, int max_len);
-GGML_API GGML_CALL void   ggml_sycl_get_device_description(int device, char *description, size_t description_size);
-GGML_API GGML_CALL int   ggml_backend_sycl_get_device_count();
-GGML_API GGML_CALL void ggml_backend_sycl_get_device_memory(int device, size_t *free, size_t *total);
-GGML_API GGML_CALL int ggml_backend_sycl_get_device_index(int device_id);
-
-// TODO: these are temporary
-//       ref: https://github.com/ggerganov/llama.cpp/pull/6022#issuecomment-1992615670
-GGML_API GGML_CALL int ggml_backend_sycl_get_device_id(int device_index);
-GGML_API GGML_CALL void ggml_backend_sycl_set_single_device_mode(int main_gpu_id);
-GGML_API GGML_CALL void ggml_backend_sycl_set_mul_device_mode();
-
-// SYCL doesn't support registering host memory, keep here for reference
-// GGML_API GGML_CALL bool ggml_backend_sycl_register_host_buffer(void * buffer, size_t size);
-// GGML_API GGML_CALL void ggml_backend_sycl_unregister_host_buffer(void * buffer);
-#ifdef  __cplusplus
-}
-#endif

bindings/ruby/ext/ggml-vulkan.h

@@ -1,29 +0,0 @@
-#pragma once
-
-#include "ggml.h"
-#include "ggml-backend.h"
-
-#ifdef  __cplusplus
-extern "C" {
-#endif
-
-#define GGML_VK_NAME "Vulkan"
-#define GGML_VK_MAX_DEVICES 16
-
-GGML_API void ggml_vk_instance_init(void);
-
-// backend API
-GGML_API GGML_CALL ggml_backend_t ggml_backend_vk_init(size_t dev_num);
-
-GGML_API GGML_CALL bool ggml_backend_is_vk(ggml_backend_t backend);
-GGML_API GGML_CALL int  ggml_backend_vk_get_device_count(void);
-GGML_API GGML_CALL void ggml_backend_vk_get_device_description(int device, char * description, size_t description_size);
-GGML_API GGML_CALL void ggml_backend_vk_get_device_memory(int device, size_t * free, size_t * total);
-
-GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_vk_buffer_type(size_t dev_num);
-// pinned host buffer for use with the CPU backend for faster copies between CPU and GPU
-GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_vk_host_buffer_type(void);
-
-#ifdef  __cplusplus
-}
-#endif

bindings/ruby/ext/metal-embed.mk

@@ -0,0 +1,14 @@
+ggml-metal-embed.o: \
+	ggml-metal.metal \
+	ggml-common.h
+	@echo "Embedding Metal library"
+	@sed -e '/#include "ggml-common.h"/r ggml-common.h' -e '/#include "ggml-common.h"/d' < ggml-metal.metal > ggml-metal-embed.metal
+	$(eval TEMP_ASSEMBLY=$(shell mktemp))
+	@echo ".section __DATA, __ggml_metallib"            >  $(TEMP_ASSEMBLY)
+	@echo ".globl _ggml_metallib_start"                 >> $(TEMP_ASSEMBLY)
+	@echo "_ggml_metallib_start:"                       >> $(TEMP_ASSEMBLY)
+	@echo ".incbin \"ggml-metal-embed.metal\""          >> $(TEMP_ASSEMBLY)
+	@echo ".globl _ggml_metallib_end"                   >> $(TEMP_ASSEMBLY)
+	@echo "_ggml_metallib_end:"                         >> $(TEMP_ASSEMBLY)
+	@$(AS) $(TEMP_ASSEMBLY) -o $@
+	@rm -f ${TEMP_ASSEMBLY}

bindings/ruby/ext/ruby_whisper.h

@@ -3,6 +3,13 @@
 
 #include "whisper.h"
 
+typedef struct {
+  VALUE *context;
+  VALUE user_data;
+  VALUE callback;
+  VALUE callbacks;
+} ruby_whisper_callback_container;
+
 typedef struct {
   struct whisper_context *context;
 } ruby_whisper;
@@ -10,6 +17,9 @@ typedef struct {
 typedef struct {
   struct whisper_full_params params;
   bool diarize;
+  ruby_whisper_callback_container *new_segment_callback_container;
+  ruby_whisper_callback_container *progress_callback_container;
+  ruby_whisper_callback_container *abort_callback_container;
 } ruby_whisper_params;
 
 #endif

bindings/ruby/extsources.yaml

@@ -0,0 +1,31 @@
+---
+- ../../src/whisper.cpp
+- ../../include/whisper.h
+- ../../ggml/src/ggml.c
+- ../../ggml/src/ggml-cpu.c
+- ../../ggml/src/ggml-impl.h
+- ../../ggml/src/ggml-aarch64.h
+- ../../ggml/src/ggml-aarch64.c
+- ../../ggml/src/ggml-alloc.c
+- ../../ggml/src/ggml-backend-impl.h
+- ../../ggml/src/ggml-backend.cpp
+- ../../ggml/src/ggml-common.h
+- ../../ggml/src/ggml-quants.h
+- ../../ggml/src/ggml-quants.c
+- ../../ggml/src/ggml-cpu-impl.h
+- ../../ggml/src/ggml-metal.m
+- ../../ggml/src/ggml-metal.metal
+- ../../ggml/src/ggml-blas.cpp
+- ../../ggml/include/ggml.h
+- ../../ggml/include/ggml-alloc.h
+- ../../ggml/include/ggml-backend.h
+- ../../ggml/include/ggml-cpu.h
+- ../../ggml/include/ggml-cuda.h
+- ../../ggml/include/ggml-kompute.h
+- ../../ggml/include/ggml-metal.h
+- ../../ggml/include/ggml-sycl.h
+- ../../ggml/include/ggml-vulkan.h
+- ../../ggml/include/ggml-blas.h
+- ../../scripts/get-flags.mk
+- ../../examples/dr_wav.h
+- ../../LICENSE

bindings/ruby/tests/helper.rb

@@ -0,0 +1,7 @@
+require "test/unit"
+require "whisper"
+
+class TestBase < Test::Unit::TestCase
+  MODEL = File.join(__dir__, "..", "..", "..", "models", "ggml-base.en.bin")
+  AUDIO = File.join(__dir__, "..", "..", "..", "samples", "jfk.wav")
+end

bindings/ruby/tests/test_callback.rb

@@ -0,0 +1,163 @@
+require "test/unit"
+require "whisper"
+
+class TestCallback < Test::Unit::TestCase
+  TOPDIR = File.expand_path(File.join(File.dirname(__FILE__), '..'))
+
+  def setup
+    GC.start
+    @params = Whisper::Params.new
+    @whisper = Whisper::Context.new(File.join(TOPDIR, '..', '..', 'models', 'ggml-base.en.bin'))
+    @audio = File.join(TOPDIR, '..', '..', 'samples', 'jfk.wav')
+  end
+
+  def test_new_segment_callback
+    @params.new_segment_callback = ->(context, state, n_new, user_data) {
+      assert_kind_of Integer, n_new
+      assert n_new > 0
+      assert_same @whisper, context
+
+      n_segments = context.full_n_segments
+      n_new.times do |i|
+        i_segment = n_segments - 1 + i
+        start_time = context.full_get_segment_t0(i_segment) * 10
+        end_time = context.full_get_segment_t1(i_segment) * 10
+        text = context.full_get_segment_text(i_segment)
+
+        assert_kind_of Integer, start_time
+        assert start_time >= 0
+        assert_kind_of Integer, end_time
+        assert end_time > 0
+        assert_match /ask not what your country can do for you, ask what you can do for your country/, text if i_segment == 0
+      end
+    }
+
+    @whisper.transcribe(@audio, @params)
+  end
+
+  def test_new_segment_callback_closure
+    search_word = "what"
+    @params.new_segment_callback = ->(context, state, n_new, user_data) {
+      n_segments = context.full_n_segments
+      n_new.times do |i|
+        i_segment = n_segments - 1 + i
+        text = context.full_get_segment_text(i_segment)
+        if text.include?(search_word)
+          t0 = context.full_get_segment_t0(i_segment)
+          t1 = context.full_get_segment_t1(i_segment)
+          raise "search word '#{search_word}' found at between #{t0} and #{t1}"
+        end
+      end
+    }
+
+    assert_raise RuntimeError do
+      @whisper.transcribe(@audio, @params)
+    end
+  end
+
+  def test_new_segment_callback_user_data
+    udata = Object.new
+    @params.new_segment_callback_user_data = udata
+    @params.new_segment_callback = ->(context, state, n_new, user_data) {
+      assert_same udata, user_data
+    }
+
+    @whisper.transcribe(@audio, @params)
+  end
+
+  def test_new_segment_callback_user_data_gc
+    @params.new_segment_callback_user_data = "My user data"
+    @params.new_segment_callback = ->(context, state, n_new, user_data) {
+      assert_equal "My user data", user_data
+    }
+    GC.start
+
+    assert_same @whisper, @whisper.transcribe(@audio, @params)
+  end
+
+  def test_progress_callback
+    first = nil
+    last = nil
+    @params.progress_callback = ->(context, state, progress, user_data) {
+      assert_kind_of Integer, progress
+      assert 0 <= progress && progress <= 100
+      assert_same @whisper, context
+      first = progress if first.nil?
+      last = progress
+    }
+    @whisper.transcribe(@audio, @params)
+    assert_equal 0, first
+    assert_equal 100, last
+  end
+
+  def test_progress_callback_user_data
+    udata = Object.new
+    @params.progress_callback_user_data = udata
+    @params.progress_callback = ->(context, state, n_new, user_data) {
+      assert_same udata, user_data
+    }
+
+    @whisper.transcribe(@audio, @params)
+  end
+
+  def test_on_progress
+    first = nil
+    last = nil
+    @params.on_progress do |progress|
+      assert_kind_of Integer, progress
+      assert 0 <= progress && progress <= 100
+      first = progress if first.nil?
+      last = progress
+    end
+    @whisper.transcribe(@audio, @params)
+    assert_equal 0, first
+    assert_equal 100, last
+  end
+
+  def test_abort_callback
+    i = 0
+    @params.abort_callback = ->(user_data) {
+      assert_nil user_data
+      i += 1
+      return false
+    }
+    @whisper.transcribe(@audio, @params)
+    assert i > 0
+  end
+
+  def test_abort_callback_abort
+    i = 0
+    @params.abort_callback = ->(user_data) {
+      i += 1
+      return i == 3
+    }
+    @whisper.transcribe(@audio, @params)
+    assert_equal 3, i
+  end
+
+  def test_abort_callback_user_data
+    udata = Object.new
+    @params.abort_callback_user_data = udata
+    yielded = nil
+    @params.abort_callback = ->(user_data) {
+      yielded = user_data
+    }
+    @whisper.transcribe(@audio, @params)
+    assert_same udata, yielded
+  end
+
+  def test_abort_on
+    do_abort = false
+    aborted_from_callback = false
+    @params.on_new_segment do |segment|
+      do_abort = true if segment.text.match? /ask/
+    end
+    i = 0
+    @params.abort_on do
+      i += 1
+      do_abort
+    end
+    @whisper.transcribe(@audio, @params)
+    assert i > 0
+  end
+end

bindings/ruby/tests/test_model.rb

@@ -0,0 +1,44 @@
+require_relative "helper"
+
+class TestModel < TestBase
+  def test_model
+    whisper = Whisper::Context.new(MODEL)
+    assert_instance_of Whisper::Model, whisper.model
+  end
+
+  def test_attributes
+    whisper = Whisper::Context.new(MODEL)
+    model = whisper.model
+
+    assert_equal 51864, model.n_vocab
+    assert_equal 1500, model.n_audio_ctx
+    assert_equal 512, model.n_audio_state
+    assert_equal 8, model.n_audio_head
+    assert_equal 6, model.n_audio_layer
+    assert_equal 448, model.n_text_ctx
+    assert_equal 512, model.n_text_state
+    assert_equal 8, model.n_text_head
+    assert_equal 6, model.n_text_layer
+    assert_equal 80, model.n_mels
+    assert_equal 1, model.ftype
+    assert_equal "base", model.type
+  end
+
+  def test_gc
+    model = Whisper::Context.new(MODEL).model
+    GC.start
+
+    assert_equal 51864, model.n_vocab
+    assert_equal 1500, model.n_audio_ctx
+    assert_equal 512, model.n_audio_state
+    assert_equal 8, model.n_audio_head
+    assert_equal 6, model.n_audio_layer
+    assert_equal 448, model.n_text_ctx
+    assert_equal 512, model.n_text_state
+    assert_equal 8, model.n_text_head
+    assert_equal 6, model.n_text_layer
+    assert_equal 80, model.n_mels
+    assert_equal 1, model.ftype
+    assert_equal "base", model.type
+  end
+end

bindings/ruby/tests/test_package.rb

@@ -0,0 +1,31 @@
+require_relative "helper"
+require 'tempfile'
+require 'tmpdir'
+require 'shellwords'
+
+class TestPackage < TestBase
+  def test_build
+    Tempfile.create do |file|
+      assert system("gem", "build", "whispercpp.gemspec", "--output", file.to_path.shellescape, exception: true)
+      assert file.size > 0
+      assert_path_exist file.to_path
+    end
+  end
+
+  sub_test_case "Building binary on installation" do
+    def setup
+      system "rake", "build", exception: true
+    end
+
+    def test_install
+      match_data = `rake -Tbuild`.match(/(whispercpp-(.+)\.gem)/)
+      filename = match_data[1]
+      version = match_data[2]
+      basename = "whisper.#{RbConfig::CONFIG["DLEXT"]}"
+      Dir.mktmpdir do |dir|
+        system "gem", "install", "--install-dir", dir.shellescape, "pkg/#{filename.shellescape}", exception: true
+        assert_path_exist File.join(dir, "gems/whispercpp-#{version}/lib", basename)
+      end
+    end
+  end
+end

bindings/ruby/tests/test_params.rb

@@ -0,0 +1,154 @@
+require_relative "helper"
+
+class TestParams < TestBase
+  def setup
+    @params  = Whisper::Params.new
+  end
+
+  def test_language
+    @params.language = "en"
+    assert_equal @params.language, "en"
+    @params.language = "auto"
+    assert_equal @params.language, "auto"
+  end
+
+  def test_offset
+    @params.offset = 10_000
+    assert_equal @params.offset, 10_000
+    @params.offset = 0
+    assert_equal @params.offset, 0
+  end
+
+  def test_duration
+    @params.duration = 60_000
+    assert_equal @params.duration, 60_000
+    @params.duration = 0
+    assert_equal @params.duration, 0
+  end
+
+  def test_max_text_tokens
+    @params.max_text_tokens = 300
+    assert_equal @params.max_text_tokens, 300
+    @params.max_text_tokens = 0
+    assert_equal @params.max_text_tokens, 0
+  end
+
+  def test_translate
+    @params.translate = true
+    assert @params.translate
+    @params.translate = false
+    assert !@params.translate
+  end
+
+  def test_no_context
+    @params.no_context = true
+    assert @params.no_context
+    @params.no_context = false
+    assert !@params.no_context
+  end
+
+  def test_single_segment
+    @params.single_segment = true
+    assert @params.single_segment
+    @params.single_segment = false
+    assert !@params.single_segment
+  end
+
+  def test_print_special
+    @params.print_special = true
+    assert @params.print_special
+    @params.print_special = false
+    assert !@params.print_special
+  end
+
+  def test_print_progress
+    @params.print_progress = true
+    assert @params.print_progress
+    @params.print_progress = false
+    assert !@params.print_progress
+  end
+
+  def test_print_realtime
+    @params.print_realtime = true
+    assert @params.print_realtime
+    @params.print_realtime = false
+    assert !@params.print_realtime
+  end
+
+  def test_print_timestamps
+    @params.print_timestamps = true
+    assert @params.print_timestamps
+    @params.print_timestamps = false
+    assert !@params.print_timestamps
+  end
+
+  def test_suppress_blank
+    @params.suppress_blank = true
+    assert @params.suppress_blank
+    @params.suppress_blank = false
+    assert !@params.suppress_blank
+  end
+
+  def test_suppress_non_speech_tokens
+    @params.suppress_non_speech_tokens = true
+    assert @params.suppress_non_speech_tokens
+    @params.suppress_non_speech_tokens = false
+    assert !@params.suppress_non_speech_tokens
+  end
+
+  def test_token_timestamps
+    @params.token_timestamps = true
+    assert @params.token_timestamps
+    @params.token_timestamps = false
+    assert !@params.token_timestamps
+  end
+
+  def test_split_on_word
+    @params.split_on_word = true
+    assert @params.split_on_word
+    @params.split_on_word = false
+    assert !@params.split_on_word
+  end
+
+  def test_initial_prompt
+    assert_nil @params.initial_prompt
+    @params.initial_prompt = "You are a polite person."
+    assert_equal "You are a polite person.", @params.initial_prompt
+  end
+
+  def test_temperature
+    assert_equal 0.0, @params.temperature
+    @params.temperature = 0.5
+    assert_equal 0.5, @params.temperature
+  end
+
+  def test_max_initial_ts
+    assert_equal 1.0, @params.max_initial_ts
+    @params.max_initial_ts = 600.0
+    assert_equal 600.0, @params.max_initial_ts
+  end
+
+  def test_length_penalty
+    assert_equal -1.0, @params.length_penalty
+    @params.length_penalty = 0.5
+    assert_equal 0.5, @params.length_penalty
+  end
+
+  def test_temperature_inc
+    assert_in_delta 0.2, @params.temperature_inc
+    @params.temperature_inc = 0.5
+    assert_in_delta 0.5, @params.temperature_inc
+  end
+
+  def test_entropy_thold
+    assert_in_delta 2.4, @params.entropy_thold
+    @params.entropy_thold = 3.0
+    assert_in_delta 3.0, @params.entropy_thold
+  end
+
+  def test_logprob_thold
+    assert_in_delta -1.0, @params.logprob_thold
+    @params.logprob_thold = -0.5
+    assert_in_delta -0.5, @params.logprob_thold
+  end
+end

bindings/ruby/tests/test_segment.rb

@@ -0,0 +1,83 @@
+require_relative "helper"
+
+class TestSegment < TestBase
+  class << self
+    attr_reader :whisper
+
+    def startup
+      @whisper = Whisper::Context.new(TestBase::MODEL)
+      params = Whisper::Params.new
+      params.print_timestamps = false
+      @whisper.transcribe(TestBase::AUDIO, params)
+    end
+  end
+
+  def test_iteration
+    whisper.each_segment do |segment|
+      assert_instance_of Whisper::Segment, segment
+    end
+  end
+
+  def test_enumerator
+    enum = whisper.each_segment
+    assert_instance_of Enumerator, enum
+    enum.to_a.each_with_index do |segment, index|
+      assert_instance_of Whisper::Segment, segment
+      assert_kind_of Integer, index
+    end
+  end
+
+  def test_start_time
+    i = 0
+    whisper.each_segment do |segment|
+      assert_equal 0, segment.start_time if i == 0
+      i += 1
+    end
+  end
+
+  def test_end_time
+    i = 0
+    whisper.each_segment do |segment|
+      assert_equal whisper.full_get_segment_t1(i) * 10, segment.end_time
+      i += 1
+    end
+  end
+
+  def test_on_new_segment
+    params = Whisper::Params.new
+    seg = nil
+    index = 0
+    params.on_new_segment do |segment|
+      assert_instance_of Whisper::Segment, segment
+      if index == 0
+        seg = segment
+        assert_equal 0, segment.start_time
+        assert_match /ask not what your country can do for you, ask what you can do for your country/, segment.text
+      end
+      index += 1
+    end
+    whisper.transcribe(AUDIO, params)
+    assert_equal 0, seg.start_time
+    assert_match /ask not what your country can do for you, ask what you can do for your country/, seg.text
+  end
+
+  def test_on_new_segment_twice
+    params = Whisper::Params.new
+    seg = nil
+    params.on_new_segment do |segment|
+      seg = segment
+      return
+    end
+    params.on_new_segment do |segment|
+      assert_same seg, segment
+      return
+    end
+    whisper.transcribe(AUDIO, params)
+  end
+
+  private
+
+  def whisper
+    self.class.whisper
+  end
+end

bindings/ruby/tests/test_whisper.rb

@@ -1,131 +1,127 @@
-TOPDIR = File.expand_path(File.join(File.dirname(__FILE__), '..'))
-EXTDIR = File.join(TOPDIR, 'ext')
-#$LIBDIR = File.join(TOPDIR, 'lib')
-#$:.unshift(LIBDIR)
-$:.unshift(EXTDIR)
+require_relative "helper"
+require "stringio"
 
-require 'whisper'
-require 'test/unit'
+# Exists to detect memory-related bug
+Whisper.log_set ->(level, buffer, user_data) {}, nil
 
-class TestWhisper < Test::Unit::TestCase
+class TestWhisper < TestBase
   def setup
     @params  = Whisper::Params.new
   end
 
-  def test_language
-    @params.language = "en"
-    assert_equal @params.language, "en"
-    @params.language = "auto"
-    assert_equal @params.language, "auto"
-  end
-
-  def test_offset
-    @params.offset = 10_000
-    assert_equal @params.offset, 10_000
-    @params.offset = 0
-    assert_equal @params.offset, 0
-  end
-
-  def test_duration
-    @params.duration = 60_000
-    assert_equal @params.duration, 60_000
-    @params.duration = 0
-    assert_equal @params.duration, 0
-  end
-
-  def test_max_text_tokens
-    @params.max_text_tokens = 300
-    assert_equal @params.max_text_tokens, 300
-    @params.max_text_tokens = 0
-    assert_equal @params.max_text_tokens, 0
-  end
-
-  def test_translate
-    @params.translate = true
-    assert @params.translate
-    @params.translate = false
-    assert !@params.translate
-  end
-
-  def test_no_context
-    @params.no_context = true
-    assert @params.no_context
-    @params.no_context = false
-    assert !@params.no_context
-  end
-
-  def test_single_segment
-    @params.single_segment = true
-    assert @params.single_segment
-    @params.single_segment = false
-    assert !@params.single_segment
-  end
-
-  def test_print_special
-    @params.print_special = true
-    assert @params.print_special
-    @params.print_special = false
-    assert !@params.print_special
-  end
-
-  def test_print_progress
-    @params.print_progress = true
-    assert @params.print_progress
-    @params.print_progress = false
-    assert !@params.print_progress
-  end
-
-  def test_print_realtime
-    @params.print_realtime = true
-    assert @params.print_realtime
-    @params.print_realtime = false
-    assert !@params.print_realtime
-  end
-
-  def test_print_timestamps
-    @params.print_timestamps = true
-    assert @params.print_timestamps
-    @params.print_timestamps = false
-    assert !@params.print_timestamps
-  end
-
-  def test_suppress_blank
-    @params.suppress_blank = true
-    assert @params.suppress_blank
-    @params.suppress_blank = false
-    assert !@params.suppress_blank
-  end
-
-  def test_suppress_non_speech_tokens
-    @params.suppress_non_speech_tokens = true
-    assert @params.suppress_non_speech_tokens
-    @params.suppress_non_speech_tokens = false
-    assert !@params.suppress_non_speech_tokens
-  end
-
-  def test_token_timestamps
-    @params.token_timestamps = true
-    assert @params.token_timestamps
-    @params.token_timestamps = false
-    assert !@params.token_timestamps
-  end
-
-  def test_split_on_word
-    @params.split_on_word = true
-    assert @params.split_on_word
-    @params.split_on_word = false
-    assert !@params.split_on_word
-  end
-
   def test_whisper
-    @whisper = Whisper::Context.new(File.join(TOPDIR, '..', '..', 'models', 'ggml-base.en.bin'))
+    @whisper = Whisper::Context.new(MODEL)
     params  = Whisper::Params.new
     params.print_timestamps = false
 
-    jfk = File.join(TOPDIR, '..', '..', 'samples', 'jfk.wav')
-    @whisper.transcribe(jfk, params) {|text|
+    @whisper.transcribe(AUDIO, params) {|text|
       assert_match /ask not what your country can do for you, ask what you can do for your country/, text
     }
   end
 
+  sub_test_case "After transcription" do
+    class << self
+      attr_reader :whisper
+
+      def startup
+        @whisper = Whisper::Context.new(TestBase::MODEL)
+        params = Whisper::Params.new
+        params.print_timestamps = false
+        @whisper.transcribe(TestBase::AUDIO, params)
+      end
+    end
+
+    def whisper
+      self.class.whisper
+    end
+
+    def test_full_n_segments
+      assert_equal 1, whisper.full_n_segments
+    end
+
+    def test_full_lang_id
+      assert_equal 0, whisper.full_lang_id
+    end
+
+    def test_full_get_segment_t0
+      assert_equal 0, whisper.full_get_segment_t0(0)
+      assert_raise IndexError do
+        whisper.full_get_segment_t0(whisper.full_n_segments)
+      end
+      assert_raise IndexError do
+        whisper.full_get_segment_t0(-1)
+      end
+    end
+
+    def test_full_get_segment_t1
+      t1 = whisper.full_get_segment_t1(0)
+      assert_kind_of Integer, t1
+      assert t1 > 0
+      assert_raise IndexError do
+        whisper.full_get_segment_t1(whisper.full_n_segments)
+      end
+    end
+
+    def test_full_get_segment_speaker_turn_next
+      assert_false whisper.full_get_segment_speaker_turn_next(0)
+    end
+
+    def test_full_get_segment_text
+      assert_match /ask not what your country can do for you, ask what you can do for your country/, whisper.full_get_segment_text(0)
+    end
+  end
+
+  def test_lang_max_id
+    assert_kind_of Integer, Whisper.lang_max_id
+  end
+
+  def test_lang_id
+    assert_equal 0, Whisper.lang_id("en")
+    assert_raise ArgumentError do
+      Whisper.lang_id("non existing language")
+    end
+  end
+
+  def test_lang_str
+    assert_equal "en", Whisper.lang_str(0)
+    assert_raise IndexError do
+      Whisper.lang_str(Whisper.lang_max_id + 1)
+    end
+  end
+
+  def test_lang_str_full
+    assert_equal "english", Whisper.lang_str_full(0)
+    assert_raise IndexError do
+      Whisper.lang_str_full(Whisper.lang_max_id + 1)
+    end
+  end
+
+  def test_log_set
+    user_data = Object.new
+    logs = []
+    log_callback = ->(level, buffer, udata) {
+      logs << [level, buffer, udata]
+    }
+    Whisper.log_set log_callback, user_data
+    Whisper::Context.new(MODEL)
+
+    assert logs.length > 30
+    logs.each do |log|
+      assert_equal Whisper::LOG_LEVEL_INFO, log[0]
+      assert_same user_data, log[2]
+    end
+  end
+
+  def test_log_suppress
+    stderr = $stderr
+    Whisper.log_set ->(level, buffer, user_data) {
+      # do nothing
+    }, nil
+    dev = StringIO.new("")
+    $stderr = dev
+    Whisper::Context.new(MODEL)
+    assert_empty dev.string
+  ensure
+    $stderr = stderr
+  end
 end

bindings/ruby/whispercpp.gemspec

@@ -1,3 +1,5 @@
+require "yaml"
+
 Gem::Specification.new do |s|
   s.name    = "whispercpp"
   s.authors = ["Georgi Gerganov", "Todd A. Fisher"]
@@ -7,10 +9,16 @@ Gem::Specification.new do |s|
   s.email   = 'todd.fisher@gmail.com'
   s.extra_rdoc_files = ['LICENSE', 'README.md']
   
-  s.files = ["LICENSE", "README.md", "Rakefile", "ext/extconf.rb", "ext/ggml.c", "ext/ruby_whisper.cpp", "ext/whisper.cpp", "ext/dr_wav.h", "ext/ggml.h", "ext/ruby_whisper.h", "ext/whisper.h"]
+  s.files = `git ls-files . -z`.split("\x0") +
+              YAML.load_file("extsources.yaml").collect {|file|
+                basename = File.basename(file)
+                if s.extra_rdoc_files.include?(basename)
+                  basename
+                else
+                  File.join("ext", basename)
+                end
+              }
 
-  #### Load-time details
-  s.require_paths = ['lib','ext']
   s.summary = %q{Ruby whisper.cpp bindings}
   s.test_files = ["tests/test_whisper.rb"]
   

examples/CMakeLists.txt

@@ -137,7 +137,7 @@ if (WHISPER_SDL2)
     set_target_properties(lsp PROPERTIES FOLDER "examples")
     if (GGML_SYCL)
         add_subdirectory(sycl)
-        set_target_properties(sycl PROPERTIES FOLDER "examples")
+        set_target_properties(ls-sycl-device PROPERTIES FOLDER "examples")
     endif()
 endif (WHISPER_SDL2)
 endif()

examples/ffmpeg-transcode.cpp

@@ -204,8 +204,6 @@ static int decode_audio(struct audio_buffer *audio_buf, s16 **data, int *size)
     const size_t errbuffsize = 1024;
     char errbuff[errbuffsize];
 
-    av_register_all(); // from avformat. Still a must-have call for ffmpeg v3! (can be skipped for later versions)
-
     fmt_ctx = avformat_alloc_context();
     avio_ctx_buffer = (u8*)av_malloc(AVIO_CTX_BUF_SZ);
     LOG("Creating an avio context: AVIO_CTX_BUF_SZ=%d\n", AVIO_CTX_BUF_SZ);

examples/main/main.cpp

@@ -997,6 +997,7 @@ int main(int argc, char ** argv) {
         if (params.dtw == "large.v1")  cparams.dtw_aheads_preset = WHISPER_AHEADS_LARGE_V1;
         if (params.dtw == "large.v2")  cparams.dtw_aheads_preset = WHISPER_AHEADS_LARGE_V2;
         if (params.dtw == "large.v3")  cparams.dtw_aheads_preset = WHISPER_AHEADS_LARGE_V3;
+        if (params.dtw == "large.v3.turbo")  cparams.dtw_aheads_preset = WHISPER_AHEADS_LARGE_V3_TURBO;
 
         if (cparams.dtw_aheads_preset == WHISPER_AHEADS_NONE) {
             fprintf(stderr, "error: unknown DTW preset '%s'\n", params.dtw.c_str());

examples/sycl/CMakeLists.txt

@@ -5,5 +5,5 @@
 set(TARGET ls-sycl-device)
 add_executable(${TARGET} ls-sycl-device.cpp)
 install(TARGETS ${TARGET} RUNTIME)
-target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
-target_compile_features(${TARGET} PRIVATE cxx_std_17)
+target_link_libraries(${TARGET} PRIVATE common whisper ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_17)

examples/sycl/build.sh

@@ -7,13 +7,16 @@ cd build
 source /opt/intel/oneapi/setvars.sh
 
 #for FP16
-#cmake .. -DWHISPER_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DWHISPER_SYCL_F16=ON # faster for long-prompt inference
+#cmake .. -DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DWHISPER_SYCL_F16=ON # faster for long-prompt inference
 
 #for FP32
-cmake .. -DWHISPER_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
+cmake .. -DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
+
+#for other features from the examples, e.g. stream and talk link with SDL2:
+#cmake .. -DGGML_SYCL=ON -DWHISPER_SDL2=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
 
 #build example/main only
 #cmake --build . --config Release --target main
 
 #build all binary
-cmake --build . --config Release -v
+cmake --build . --config Release -v

examples/talk-llama/llama-sampling.cpp

@@ -63,6 +63,30 @@ static void llama_log_softmax(float * array, size_t size) {
 }
 */
 
+static void llama_sampler_temp_impl(llama_token_data_array * cur_p, float temp) {
+    if (temp <= 0.0f) {
+        // find the token with the highest logit and set the rest to -inf
+        size_t max_i = 0;
+        float  max_l = cur_p->data[0].logit;
+
+        for (size_t i = 1; i < cur_p->size; ++i) {
+            if (cur_p->data[i    ].logit > max_l) {
+                cur_p->data[max_i].logit = -INFINITY;
+                max_i = i;
+                max_l = cur_p->data[i].logit;
+            } else {
+                cur_p->data[i].logit = -INFINITY;
+            }
+        }
+
+        return;
+    }
+
+    for (size_t i = 0; i < cur_p->size; ++i) {
+        cur_p->data[i].logit /= temp;
+    }
+}
+
 static void llama_sampler_softmax_impl(llama_token_data_array * cur_p) {
     GGML_ASSERT(cur_p->size > 0);
 
@@ -89,7 +113,7 @@ static void llama_sampler_softmax_impl(llama_token_data_array * cur_p) {
 }
 
 static void llama_sampler_top_k_impl(llama_token_data_array * cur_p, int32_t k) {
-    // TODO: move bucket sort to separate function so that top_p/tail_free/typical/softmax first is equally fast
+    // TODO: move bucket sort to separate function so that top_p/typical/softmax first is equally fast
     // if (k >= (int32_t)cur_p->size) {
     //     return;
     // }
@@ -427,6 +451,9 @@ static const char * llama_sampler_dist_name(const struct llama_sampler * /*smpl*
 
 static void llama_sampler_dist_apply(struct llama_sampler * smpl, llama_token_data_array * cur_p) {
     auto * ctx = (llama_sampler_dist *) smpl->ctx;
+
+    llama_sampler_softmax_impl(cur_p);
+
     cur_p->selected = llama_sample_dist(cur_p, ctx->rng);
 }
 
@@ -706,101 +733,6 @@ struct llama_sampler * llama_sampler_init_min_p(float p, size_t min_keep) {
     };
 }
 
-// tail-free
-
-struct llama_sampler_tail_free {
-    const float  z;
-    const size_t min_keep;
-};
-
-static const char * llama_sampler_tail_free_name(const struct llama_sampler * /*smpl*/) {
-    return "tail-free";
-}
-
-static void llama_sampler_tail_free_apply(struct llama_sampler * smpl, llama_token_data_array * cur_p) {
-    const auto * ctx = (llama_sampler_tail_free *) smpl->ctx;
-
-    if (ctx->z >= 1.0f || cur_p->size <= 2) {
-        return;
-    }
-
-    llama_sampler_softmax_impl(cur_p);
-
-    // Compute the first and second derivatives
-    std::vector<float> first_derivatives(cur_p->size - 1);
-    std::vector<float> second_derivatives(cur_p->size - 2);
-
-    for (size_t i = 0; i < first_derivatives.size(); ++i) {
-        first_derivatives[i] = cur_p->data[i].p - cur_p->data[i + 1].p;
-    }
-    for (size_t i = 0; i < second_derivatives.size(); ++i) {
-        second_derivatives[i] = first_derivatives[i] - first_derivatives[i + 1];
-    }
-
-    // Calculate absolute value of second derivatives
-    for (size_t i = 0; i < second_derivatives.size(); ++i) {
-        second_derivatives[i] = std::abs(second_derivatives[i]);
-    }
-
-    // Normalize the second derivatives
-    {
-        const float second_derivatives_sum = std::accumulate(second_derivatives.begin(), second_derivatives.end(), 0.0f);
-
-        if (second_derivatives_sum > 1e-6f) {
-            for (float & value : second_derivatives) {
-                value /= second_derivatives_sum;
-            }
-        } else {
-            for (float & value : second_derivatives) {
-                value = 1.0f / second_derivatives.size();
-            }
-        }
-    }
-
-    float cum_sum = 0.0f;
-    size_t last_idx = cur_p->size;
-    for (size_t i = 0; i < second_derivatives.size(); ++i) {
-        cum_sum += second_derivatives[i];
-
-        // Check if the running sum is greater than z or if we have kept at least min_keep tokens
-        if (cum_sum > ctx->z && i >= ctx->min_keep) {
-            last_idx = i;
-            break;
-        }
-    }
-
-    // Resize the output vector to keep only the tokens above the tail location
-    cur_p->size = last_idx;
-}
-
-static struct llama_sampler * llama_sampler_tail_free_clone(const struct llama_sampler * smpl) {
-    const auto * ctx = (const llama_sampler_tail_free *) smpl->ctx;
-    return llama_sampler_init_tail_free(ctx->z, ctx->min_keep);
-}
-
-static void llama_sampler_tail_free_free(struct llama_sampler * smpl) {
-    delete (llama_sampler_tail_free *) smpl->ctx;
-}
-
-static struct llama_sampler_i llama_sampler_tail_free_i = {
-    /* .name   = */ llama_sampler_tail_free_name,
-    /* .accept = */ nullptr,
-    /* .apply  = */ llama_sampler_tail_free_apply,
-    /* .reset  = */ nullptr,
-    /* .clone  = */ llama_sampler_tail_free_clone,
-    /* .free   = */ llama_sampler_tail_free_free,
-};
-
-struct llama_sampler * llama_sampler_init_tail_free(float z, size_t min_keep) {
-    return new llama_sampler {
-        /* .iface = */ &llama_sampler_tail_free_i,
-        /* .ctx   = */ new llama_sampler_tail_free {
-            /* .z        = */ z,
-            /*. min_keep = */ min_keep,
-        },
-    };
-}
-
 // typical
 
 struct llama_sampler_typical {
@@ -912,9 +844,8 @@ static const char * llama_sampler_temp_name(const struct llama_sampler * /*smpl*
 
 static void llama_sampler_temp_apply(struct llama_sampler * smpl, llama_token_data_array * cur_p) {
     const auto * ctx = (llama_sampler_temp *) smpl->ctx;
-    for (size_t i = 0; i < cur_p->size; ++i) {
-        cur_p->data[i].logit /= ctx->temp;
-    }
+
+    llama_sampler_temp_impl(cur_p, ctx->temp);
 }
 
 static struct llama_sampler * llama_sampler_temp_clone(const struct llama_sampler * smpl) {
@@ -961,6 +892,7 @@ static void llama_sampler_temp_ext_apply(struct llama_sampler * smpl, llama_toke
     if (ctx->delta > 0) {
         const float min_temp = std::max(0.0f, ctx->temp - ctx->delta);
         const float max_temp = ctx->temp + ctx->delta;
+
         float exponent_val = ctx->exponent;
 
         // no need to do anything if there is only one (or zero) candidates
@@ -998,9 +930,7 @@ static void llama_sampler_temp_ext_apply(struct llama_sampler * smpl, llama_toke
     #endif
 
         // Apply the dynamically calculated temperature scaling
-        for (size_t i = 0; i < cur_p->size; ++i) {
-            cur_p->data[i].logit /= dyn_temp;
-        }
+        llama_sampler_temp_impl(cur_p, dyn_temp);
 
         // Re-compute softmax probabilities after scaling logits with dynamic temperature
         const double max_l_double = cur_p->data[0].logit;
@@ -1024,9 +954,7 @@ static void llama_sampler_temp_ext_apply(struct llama_sampler * smpl, llama_toke
         }
     #endif
     } else {
-        for (size_t i = 0; i < cur_p->size; ++i) {
-            cur_p->data[i].logit /= ctx->temp;
-        }
+        llama_sampler_temp_impl(cur_p, ctx->temp);
     }
 }
 
@@ -1059,6 +987,101 @@ struct llama_sampler * llama_sampler_init_temp_ext(float temp, float delta, floa
     };
 }
 
+// xtc
+
+struct llama_sampler_xtc {
+    const float    probability;
+    const float    threshold;
+    const size_t   min_keep;
+
+    const uint32_t seed;
+    uint32_t       seed_cur;
+
+    std::mt19937   rng;
+};
+
+static const char * llama_sampler_xtc_name(const struct llama_sampler * /*smpl*/) {
+    return "xtc";
+}
+
+static void llama_sample_xtc_apply(struct llama_sampler * smpl, llama_token_data_array * cur_p) {
+    auto * ctx = (llama_sampler_xtc *) smpl->ctx;
+
+    if (ctx->probability <= 0.0f
+        || ctx->threshold > 0.5f
+        || cur_p->size < 2) {
+        return;
+    }
+
+    std::uniform_real_distribution<float> distribution(0.0f, 1.0f);
+    float chance = distribution(ctx->rng);
+    if (chance > ctx->probability) return;
+
+    // in case it's not sorted/recalculated yet
+    llama_sampler_softmax_impl(cur_p);
+
+    int pos_last = 0;
+
+    for (size_t i = 0; i < cur_p->size; ++i) {
+        if (cur_p->data[i].p >= ctx->threshold) {
+            pos_last = i;
+        } else break;
+    }
+
+    if (cur_p->size - pos_last >= ctx->min_keep && pos_last > 0) {
+        cur_p->data += pos_last;
+        cur_p->size -= pos_last;
+    }
+}
+
+static struct llama_sampler * llama_sampler_xtc_clone(const struct llama_sampler * smpl) {
+    const auto * ctx = (const llama_sampler_xtc *) smpl->ctx;
+    auto * result = llama_sampler_init_xtc(ctx->probability, ctx->threshold, ctx->min_keep, ctx->seed);
+
+    // copy the state
+    {
+        auto * result_ctx = (llama_sampler_xtc *) result->ctx;
+
+        result_ctx->rng = ctx->rng;
+    }
+
+    return result;
+}
+
+static void llama_sampler_xtc_free(struct llama_sampler * smpl) {
+    delete (llama_sampler_xtc *) smpl->ctx;
+}
+
+static void llama_sampler_xtc_reset(struct llama_sampler * smpl) {
+    auto * ctx = (llama_sampler_xtc *) smpl->ctx;
+    ctx->seed_cur = get_rng_seed(ctx->seed);
+    ctx->rng.seed(ctx->seed_cur);
+}
+
+static struct llama_sampler_i llama_sampler_xtc_i = {
+    /* .name   = */ llama_sampler_xtc_name,
+    /* .accept = */ nullptr,
+    /* .apply  = */ llama_sample_xtc_apply,
+    /* .reset  = */ llama_sampler_xtc_reset,
+    /* .clone  = */ llama_sampler_xtc_clone,
+    /* .free   = */ llama_sampler_xtc_free,
+};
+
+struct llama_sampler * llama_sampler_init_xtc(float p, float t, size_t min_keep, uint32_t seed) {
+    auto seed_cur = get_rng_seed(seed);
+    return new llama_sampler {
+        /* .iface = */ &llama_sampler_xtc_i,
+        /* .ctx   = */ new llama_sampler_xtc {
+            /* .probability   = */ p,
+            /* .threshold     = */ t,
+            /* .min_keep      = */ min_keep,
+            /* .seed          = */ seed,
+            /* .seed_cur      = */ seed_cur,
+            /* .rng           = */ std::mt19937(seed_cur),
+        },
+    };
+}
+
 // mirostat
 
 struct llama_sampler_mirostat {
@@ -1565,6 +1588,400 @@ struct llama_sampler * llama_sampler_init_penalties(
     };
 }
 
+// DRY
+
+struct llama_sampler_dry {
+    int32_t total_context_size;
+
+    const float   dry_multiplier;
+    const float   dry_base;
+    const int32_t dry_allowed_length;
+    const int32_t dry_penalty_last_n;
+
+    std::unordered_multimap<llama_token, std::vector<llama_token>> dry_processed_breakers;
+    std::vector<int> dry_repeat_count;
+    std::unordered_map<llama_token, int> dry_max_token_repeat;
+    ring_buffer<llama_token> last_tokens;
+};
+
+// Ported from Koboldcpp, original PR: https://github.com/LostRuins/koboldcpp/pull/982 (Original author: pi6am)
+static void get_overlapping_token_sequences(const llama_vocab & vocab, const std::string& str, std::unordered_multimap<llama_token, std::vector<llama_token>>& token_sequences, int max_tail_len = -1) {
+    for (llama_token token_id = 0; token_id < (llama_token)vocab.n_vocab; token_id++) {
+        std::string word = llama_detokenize(vocab, {token_id}, true);
+        if (word.find(str) != std::string::npos) {
+            token_sequences.emplace(token_id, std::vector<llama_token>());
+        } else {
+            size_t word_len = word.size(), str_len = str.size();
+            size_t pos = -1;
+            while ((pos = word.find(str[0], pos + 1)) != std::string::npos) {
+                bool match = true;
+                size_t i;
+                for (i = 1; i < str_len && i + pos < word_len; ++i) {
+                    if (word[pos + i] != str[i]) {
+                        match = false;
+                        break;
+                    }
+                }
+                if (match) {
+                    std::vector<llama_token> tokenization = llama_tokenize_internal(vocab, str.substr(i), false, false);
+                    if (max_tail_len >= 0 && tokenization.size() > (size_t)max_tail_len) {
+                        tokenization.resize(max_tail_len);
+                    }
+
+                    // Ensure we don't already have a duplicate matching tokenization
+                    auto its = token_sequences.equal_range(token_id);
+                    bool found = false;
+                    for (auto it = its.first; it != its.second; ++it) {
+                        if (tokenization == it->second) {
+                            found = true;
+                            break;
+                        }
+                    }
+                    if (!found) {
+                        token_sequences.emplace(token_id, tokenization);
+                    }
+                }
+            }
+        }
+    }
+}
+
+static const char * llama_sampler_dry_name(const struct llama_sampler * /*smpl*/) {
+    return "dry";
+}
+
+static void llama_sampler_dry_accept(struct llama_sampler * smpl, llama_token token) {
+    auto * ctx = (llama_sampler_dry *) smpl->ctx;
+    if (ctx->dry_multiplier == 0.0f || ctx->dry_base < 1.0f || ctx->dry_penalty_last_n == 0) {
+        return;
+    }
+
+    ctx->last_tokens.push_back(token);
+}
+
+// Ported from Koboldcpp, original PR: https://github.com/LostRuins/koboldcpp/pull/982 (Original author: pi6am)
+static void llama_sampler_dry_apply(struct llama_sampler * smpl, llama_token_data_array * cur_p) {
+    auto * ctx = (llama_sampler_dry *) smpl->ctx;
+
+    if (ctx->dry_multiplier == 0.0f || ctx->dry_base < 1.0f || ctx->dry_penalty_last_n == 0) {
+        return;
+    }
+
+    int32_t effective_dry_penalty_last_n = (ctx->dry_penalty_last_n == -1) ? ctx->total_context_size : std::max(ctx->dry_penalty_last_n, 0);
+    int last_n_repeat = std::min(std::min((int)ctx->last_tokens.size(), effective_dry_penalty_last_n), ctx->total_context_size);
+
+    if (last_n_repeat <= ctx->dry_allowed_length) {
+        return;
+    }
+
+    ctx->dry_repeat_count.assign(last_n_repeat, 0);
+    ctx->dry_max_token_repeat.clear();
+
+    // Step 1: Look for restart sequences to limit the maximum repetition length.
+    // Work backwards through the context looking for any token that begins a restart sequence.
+    //
+    // The collection `restart_sequences` is a mapping from a "head" token to all "tail"
+    // sequences that together comprise a restart sequence. This allows us to quickly check
+    // whether each token is the head of a complete sequence. Most restart sequences are actually
+    // a single token, and for these the "tail" is an empty vector.
+    //
+    // If the token is a "head", test all restart sequences that begin with this token
+    // (there will often only be one sequence for each token, but if sequences like 'aaaq1' and
+    // 'aaa1' are used as restart strings, both could start with 'aaa' when tokenized). The
+    // longest matching sequence (if any) is used to limit the maximum repetition length.
+    //
+    // Note that in the case case of a short sequence contained in a longer one, this might fail to
+    // find the smallest value for `rep_limit`. For example, if 'amniotic' and 'ni' are both used as
+    // restart sequences, 'ni' will be found first, and since it's shorter it will fail to suppress
+    // 'otic'. This is a minor issue since fully contained restart sequences are likely to be rare.
+    //
+    // This is theoretically worst-case O(N^2) for arbitrary restart sequences, which is why we
+    // have already clamped the maximum tail sequence length when generating `restart_sequences`.
+    // With clamping, this scan is O(N) in the context length.
+
+    int rep_limit = last_n_repeat;
+    for (int i = 0; i < last_n_repeat; ++i) {
+        llama_token token = ctx->last_tokens.rat(i);
+        auto its = ctx->dry_processed_breakers.equal_range(token);
+        if (its.first == ctx->dry_processed_breakers.end()) {
+            continue;
+        }
+        int longest_match = -1;
+        for (auto it = its.first; it != its.second; ++it) {
+            // Note that (*it) does not contain the head character, so seq_len will be
+            // the restart sequence length minus 1.
+            // In the common case of a single-token restart sequence, (*it) will be empty
+            // and we will trivially match.
+            int seq_len = (int)it->second.size();
+            if (seq_len > longest_match && seq_len <= (int)i) {
+                bool match = true;
+                for (int offset = 0; offset < seq_len; ++offset) {
+                    // The -1 when indexing `last_tokens` is because we already matched the head.
+                    if (it->second[offset] != ctx->last_tokens.rat(i - offset - 1)) {
+                        match = false;
+                        break;
+                    }
+                }
+                if (match) {
+                    longest_match = seq_len;
+                }
+            }
+        }
+        if (longest_match >= 0) {
+            // We found a restart sequence starting `i` tokens from the end and continuing for
+            // `longest_match` tokens.
+            rep_limit = i - longest_match;
+            break;
+        }
+    }
+    if (rep_limit < ctx->dry_allowed_length) {
+        return;
+    }
+
+    // Step 2: Iterate in reverse over the last N tokens of the context, using the "Z-algorithm" (in
+    // the reverse direction) to efficiently compute the positions and lengths of suffixes appearing
+    // elsewhere in the context. We limit the suffix length to `rep_limit` to respect restart sequences.
+    //
+    // This algorithm is not currently documented on Wikipedia, but there is a clear description here:
+    // https://ivanyu.me/blog/2014/10/15/z-algorithm/
+    //
+    // The code below is adapted from the public domain implementation by the same author here:
+    // https://github.com/ivanyu/string-algorithms/blob/master/z_algorithm.py
+    //
+    // Example:
+    // Last N tokens: a b c c b c y a b c
+    // Repeat counts: 0 0 3 1 0 2 0 0 0 0
+    //                    ^
+    //   This `3` means that the last three tokens of the context (a b c) also appear here.
+    //
+    // This step is worst case O(N) since the Z-algorithm is linear, despite the appearance of nested
+    // for/while loops. This can be seen by observing that the `lt` and `rt` bounds are set after each
+    // repeated suffix is detected (i.e. after each while loop when n > 0). These bound variables
+    // ensure that the inner while loops only examine each token in the context once as the outer
+    // for loop iterates over the context.
+
+    {
+        const int last = last_n_repeat - 1;
+        int rt = 0, lt = 0;
+
+        for (int k = 1; k < last_n_repeat; ++k) {
+            if (k > rt) {
+                // If k is outside the current Z-box, do naive computation.
+                int n = 0;
+                while (n + k < last_n_repeat && ctx->last_tokens.rat(n) == ctx->last_tokens.rat(n+k)) {
+                    ++n;
+                }
+                ctx->dry_repeat_count[last - k] = std::min(n, rep_limit);
+                if (n > 0) {
+                    lt = k;
+                    rt = k+n-1;
+                }
+            } else {
+                // If k is inside the current Z-box, consider two cases.
+
+                int p = k - lt; // Pair index.
+                int right_part_len = rt - k + 1;
+
+                if (ctx->dry_repeat_count[last - p] < right_part_len) {
+                    int n = std::min(ctx->dry_repeat_count[last - p], rep_limit);
+                    ctx->dry_repeat_count[last - k] = n;
+                } else {
+                    int i = rt + 1;
+                    while (i < last_n_repeat && ctx->last_tokens.rat(i) == ctx->last_tokens.rat(i - k)) {
+                        i += 1;
+                    }
+
+                    int n = std::min(i - k, rep_limit);
+                    ctx->dry_repeat_count[last - k] = n;
+                    lt = k;
+                    rt = i - 1;
+                }
+            }
+        }
+    }
+
+    // Step 3: Iterate over dry_repeat_count and last_tokens, examining the maximum repeat length
+    // that would be generated by emitting each new token that would extend a sequence.
+    //
+    // Following the same example as above:
+    // Last N tokens: a b c c b c y a b c
+    // Repeat counts: 0 0 3 1 0 2 0 0 0 0
+    //
+    // For each non-zero, look ahead one token. This token, if emitted, would extend the repetition.
+    // c: 3 -> 4 (from `a b c` to `a b c c`)
+    // b: 1 -> 2 (from `c` to `c b`)
+    // y: 2 -> 3 (from `b c` to `b c y`)
+
+    for (int i = 0; i < last_n_repeat - 1; ++i) {
+        int repeat_len = ctx->dry_repeat_count[i];
+        if (repeat_len >= ctx->dry_allowed_length) {
+            // This token ends a repeat, so the next token would continue one.
+            // By convention, the value of `repeat_len` only includes the tokens currently
+            // in the context, not the new token that would be added.
+            llama_token token = ctx->last_tokens.rat(last_n_repeat - 2 - i);
+            // Track the maximum sequence ending in this token.
+            const auto& it = ctx->dry_max_token_repeat.find(token);
+            if (it == ctx->dry_max_token_repeat.end() || it->second < repeat_len) {
+                ctx->dry_max_token_repeat[token] = repeat_len;
+            }
+        }
+    }
+
+    // Step 4: Apply logit penalties based on the maximum repeat length for relevant tokens.
+
+    // Prevent floating point overflow in `pow(penalty_base, exponent)` by clamping to `max_exponent`.
+    // Compute it from `penalty_base` and the approximate log of `std::numeric_limits<float>::max()`
+    const float FLOAT_MAX_LOG = 88.7228391f;
+    int max_exponent = 0;
+    if (ctx->dry_base > 1.000001f) {
+        max_exponent = FLOAT_MAX_LOG / std::log(ctx->dry_base);
+    }
+
+    for (size_t i = 0; i < cur_p->size; ++i) {
+        const auto& af_kvp = ctx->dry_max_token_repeat.find(cur_p->data[i].id);
+        if (af_kvp != ctx->dry_max_token_repeat.end()) {
+            // Check all sequence breakers starting with this token
+            auto range = ctx->dry_processed_breakers.equal_range(cur_p->data[i].id);
+            bool is_single_token_breaker = false;
+
+            for (auto it = range.first; it != range.second; ++it) {
+                if (it->second.empty()) {
+                    is_single_token_breaker = true;
+                    break;
+                }
+            }
+
+            // Apply penalty only if it's not a single-token sequence breaker
+            if (!is_single_token_breaker) {
+                int repeat_exp = af_kvp->second - ctx->dry_allowed_length;
+                if (max_exponent > 0 && repeat_exp > max_exponent) {
+                    repeat_exp = max_exponent;
+                }
+                float penalty = ctx->dry_multiplier * std::pow(ctx->dry_base, repeat_exp);
+                cur_p->data[i].logit -= penalty;
+            }
+        }
+    }
+
+    cur_p->sorted = false;
+}
+
+static void llama_sampler_dry_reset(struct llama_sampler * smpl) {
+    auto * ctx = (llama_sampler_dry *) smpl->ctx;
+    ctx->last_tokens.clear();
+    ctx->dry_repeat_count.clear();
+    ctx->dry_max_token_repeat.clear();
+}
+
+static struct llama_sampler * llama_sampler_dry_clone(const struct llama_sampler * smpl) {
+    const auto * ctx = (llama_sampler_dry *) smpl->ctx;
+
+    llama_vocab dummy_vocab;
+
+    // dummy vocab is passed because it is only needed for raw sequence breaker processing, which we have already done and will simply be copying
+    auto * result = llama_sampler_init_dry_impl(dummy_vocab, ctx->total_context_size, ctx->dry_multiplier, ctx->dry_base, ctx->dry_allowed_length, ctx->dry_penalty_last_n, NULL, 0);
+
+    // Copy the state, including the processed breakers
+    {
+        auto * result_ctx = (llama_sampler_dry *) result->ctx;
+        result_ctx->dry_processed_breakers = ctx->dry_processed_breakers;
+        result_ctx->dry_repeat_count = ctx->dry_repeat_count;
+        result_ctx->dry_max_token_repeat = ctx->dry_max_token_repeat;
+        result_ctx->last_tokens = ctx->last_tokens;
+    }
+
+    return result;
+}
+
+static void llama_sampler_dry_free(struct llama_sampler * smpl) {
+    delete (llama_sampler_dry *) smpl->ctx;
+}
+
+static struct llama_sampler_i llama_sampler_dry_i = {
+    /* .name   = */ llama_sampler_dry_name,
+    /* .accept = */ llama_sampler_dry_accept,
+    /* .apply  = */ llama_sampler_dry_apply,
+    /* .reset  = */ llama_sampler_dry_reset,
+    /* .clone  = */ llama_sampler_dry_clone,
+    /* .free   = */ llama_sampler_dry_free,
+};
+
+struct llama_sampler * llama_sampler_init_dry_impl(const struct llama_vocab & vocab, int32_t context_size, float dry_multiplier, float dry_base, int32_t dry_allowed_length, int32_t dry_penalty_last_n, const char** seq_breakers, size_t num_breakers) {
+    int32_t effective_dry_penalty_last_n = (dry_penalty_last_n == -1) ? context_size : std::max(dry_penalty_last_n, 0);
+    std::unordered_multimap<llama_token, std::vector<llama_token>> processed_breakers;
+    const int MAX_CHAR_LEN = 40;
+    const int MAX_SEQ_LEN = 20;
+
+    const bool dry_enabled = (dry_multiplier != 0.0f && dry_base >= 1.0f && dry_penalty_last_n != 0);
+
+    if (dry_enabled && seq_breakers != nullptr && num_breakers > 0) {
+        // Process sequence breakers
+        for (size_t i = 0; i < num_breakers; ++i) {
+            if (seq_breakers[i] == nullptr || std::strlen(seq_breakers[i]) == 0) {
+                LLAMA_LOG_WARN("skipping null or empty DRY sequence breaker at index %zu\n", i);
+                continue;
+            }
+
+            std::string sequence_break(seq_breakers[i]);
+            if (sequence_break.empty()) {
+                LLAMA_LOG_WARN("skipping empty DRY sequence breaker\n");
+                continue;
+            }
+
+            if (sequence_break.size() > MAX_CHAR_LEN) {
+                LLAMA_LOG_WARN("truncating DRY sequence breaker to %d characters\n", MAX_CHAR_LEN);
+                sequence_break.resize(MAX_CHAR_LEN);
+            }
+
+            get_overlapping_token_sequences(vocab, sequence_break, processed_breakers, MAX_SEQ_LEN);
+        }
+    }
+
+    return new llama_sampler {
+        /* .iface = */ &llama_sampler_dry_i,
+        /* .ctx   = */ new llama_sampler_dry {
+            /* .total_context_size     = */ context_size,
+            /* .dry_multiplier         = */ dry_multiplier,
+            /* .dry_base               = */ dry_base,
+            /* .dry_allowed_length     = */ dry_allowed_length,
+            /* .dry_penalty_last_n     = */ dry_penalty_last_n,
+            /* .dry_processed_breakers = */ std::move(processed_breakers),
+            /* .dry_repeat_count       = */ dry_enabled ? std::vector<int>(effective_dry_penalty_last_n, 0) : std::vector<int>{},
+            /* .dry_max_token_repeat   = */ {},
+            /* .last_tokens            = */ dry_enabled ? ring_buffer<llama_token>(effective_dry_penalty_last_n) : ring_buffer<llama_token>(0),
+        },
+    };
+}
+
+// wrapper for test-sampling.cpp
+struct llama_sampler * llama_sampler_init_dry_testing(int32_t context_size, float dry_multiplier, float dry_base, int32_t dry_allowed_length, int32_t dry_penalty_last_n, const std::vector<std::vector<llama_token>>& seq_breakers) {
+    llama_vocab dummy_vocab;
+    auto * result = llama_sampler_init_dry_impl(dummy_vocab, context_size, dry_multiplier, dry_base, dry_allowed_length, dry_penalty_last_n, NULL, 0);
+    auto * ctx = (llama_sampler_dry *) result->ctx;
+
+    // Process the token-based sequence breakers
+    ctx->dry_processed_breakers.clear();
+    if (seq_breakers.empty()) {
+        LLAMA_LOG_WARN("empty DRY sequence breakers list in llama_sampler_init_dry_testing\n");
+    } else {
+        for (const auto& breaker : seq_breakers) {
+            if (breaker.empty()) {
+                LLAMA_LOG_WARN("skipping DRY empty sequence breaker\n");
+                continue;
+            }
+            llama_token head_token = breaker[0];
+            std::vector<llama_token> tail_tokens(breaker.begin() + 1, breaker.end());
+            ctx->dry_processed_breakers.emplace(head_token, std::move(tail_tokens));
+        }
+
+        if (ctx->dry_processed_breakers.empty()) {
+            LLAMA_LOG_WARN("no valid DRY sequence breakers processed in llama_sampler_init_dry_testing\n");
+        }
+    }
+
+    return result;
+}
+
 // logit-bias
 
 struct llama_sampler_logit_bias {
@@ -1644,6 +2061,229 @@ struct llama_sampler * llama_sampler_init_logit_bias(
     };
 }
 
+// infill
+
+//#define GGML_DEBUG_SAMPLER_INFILL
+
+struct llama_sampler_infill {
+    const struct llama_vocab * vocab;
+
+    std::vector<char> buf0;
+    std::vector<char> buf1;
+};
+
+static const char * llama_sampler_infill_name(const struct llama_sampler * /*smpl*/) {
+    return "infill";
+}
+
+static void llama_sampler_infill_apply(struct llama_sampler * smpl, llama_token_data_array * cur_p) {
+    auto * ctx = (llama_sampler_infill *) smpl->ctx;
+
+    llama_sampler_softmax_impl(cur_p);
+
+#if defined(GGML_DEBUG_SAMPLER_INFILL)
+#define LOG_DBG_CUR LLAMA_LOG_DEBUG
+#else
+#define LOG_DBG_CUR(...)
+#endif
+
+    for (size_t i = 0; i < cur_p->size; ++i) {
+        LOG_DBG_CUR("%s: cur_p[%3zu] = { id: %6d, p: %.6f, logit: %6.3f }\n", __func__, i, cur_p->data[i].id, cur_p->data[i].p, cur_p->data[i].logit);
+    }
+
+    float p_txt_sum = 0.0f;
+    float p_eog_sum = 0.0f;
+
+    for (size_t i = 0; i < cur_p->size; ++i) {
+        if (llama_token_is_eog_impl(*ctx->vocab, cur_p->data[i].id)) {
+            p_eog_sum += cur_p->data[i].p;
+        } else {
+            p_txt_sum += cur_p->data[i].p;
+        }
+    }
+
+    const float rat = p_eog_sum == 0.0 ? INFINITY : p_txt_sum / p_eog_sum; GGML_UNUSED(rat);
+
+    LOG_DBG_CUR("%s: p_txt_sum = %.2f, p_eog_sum = %.2f, rat = %.2f, n = %zu\n", __func__, p_txt_sum, p_eog_sum, rat, cur_p->size);
+
+    if (3*p_eog_sum*cur_p->size > p_txt_sum) {
+        LOG_DBG_CUR("%s: the ratio p_txt/p_eog = %.2f is too low -> sampling EOG\n", __func__, p_txt_sum/p_eog_sum);
+
+        // keep just the EOG tokens
+        const auto size_org = cur_p->size;
+
+        cur_p->size = 0;
+
+        float p_sum = 0.0f;
+
+        for (size_t i = 0; i < size_org; ++i) {
+            if (llama_token_is_eog_impl(*ctx->vocab, cur_p->data[i].id)) {
+                p_sum += cur_p->data[i].p;
+
+                cur_p->data[cur_p->size++] = cur_p->data[i];
+            }
+        }
+
+        // normalize probs
+        for (size_t i = 0; i < cur_p->size; ++i) {
+            cur_p->data[i].p /= p_sum;
+        }
+
+        return;
+    }
+
+    size_t n_combined = 0; GGML_UNUSED(n_combined);
+
+    // combine tokens with common prefix
+    for (size_t i0 = 0; i0 < cur_p->size; ++i0) {
+        for (size_t i1 = 0; i1 < cur_p->size; ++i1) {
+            if (cur_p->data[i0].logit == -INFINITY) {
+                break;
+            }
+
+            if (i0 == i1 || cur_p->data[i1].logit == -INFINITY) {
+                continue;
+            }
+
+            int len0 = llama_token_to_piece_impl(*ctx->vocab, cur_p->data[i0].id, ctx->buf0.data(), ctx->buf0.size(), 0, false);
+            if (len0 < 0) {
+                ctx->buf0.resize(len0);
+                len0 = llama_token_to_piece_impl(*ctx->vocab, cur_p->data[i0].id, ctx->buf0.data(), ctx->buf0.size(), 0, false);
+                assert(len0 > 0);
+            }
+
+            int len1 = llama_token_to_piece_impl(*ctx->vocab, cur_p->data[i1].id, ctx->buf1.data(), ctx->buf1.size(), 0, false);
+            if (len1 < 0) {
+                ctx->buf1.resize(len1);
+                len1 = llama_token_to_piece_impl(*ctx->vocab, cur_p->data[i1].id, ctx->buf1.data(), ctx->buf1.size(), 0, false);
+                assert(len1 > 0);
+            }
+
+            // token i0 is a prefix of token i1
+            if (len0 > 0 && len0 <= len1 && memcmp(ctx->buf0.data(), ctx->buf1.data(), len0) == 0) {
+                int dst = i0;
+                int src = i1;
+
+                // merge into the token with higher probability
+                if (cur_p->data[i1].p > cur_p->data[i0].p) {
+                    std::swap(dst, src);
+                }
+
+                cur_p->data[dst].p += cur_p->data[src].p;
+                cur_p->data[src].logit = -INFINITY;
+                cur_p->data[src].p     = 0.0f;
+
+                n_combined++;
+            }
+        }
+    }
+
+    size_t n_non_eog = 0;
+
+    size_t size_org = cur_p->size;
+
+    float p_sum = 0.0f;
+    float thold = 0.2f;
+
+    cur_p->size = 0;
+
+    LOG_DBG_CUR("%s: n_combined = %zu, applying thold = %.3f\n", __func__, n_combined, thold);
+
+    for (size_t i = 0; i < size_org; ++i) {
+        const bool is_eog = llama_token_is_eog_impl(*ctx->vocab, cur_p->data[i].id);
+
+        if (cur_p->data[i].p < thold && !is_eog) {
+            continue;
+        }
+
+        if (!is_eog) {
+            ++n_non_eog;
+        }
+
+        p_sum += cur_p->data[i].p;
+
+        // keep this token
+        cur_p->data[cur_p->size++] = cur_p->data[i];
+    }
+
+    LOG_DBG_CUR("%s: n_non_eog = %zu\n", __func__, n_non_eog);
+
+    // if no non-EOG tokens are left -> reduce cur_p to single EOT token
+    if (n_non_eog == 0) {
+        cur_p->size = 1;
+        cur_p->data[0].id = llama_token_eot_impl(*ctx->vocab);
+        cur_p->data[0].logit = 1.0f;
+
+        return;
+    }
+
+    // normalize probs
+    for (size_t i = 0; i < cur_p->size; ++i) {
+        cur_p->data[i].p /= p_sum;
+
+        LOG_DBG_CUR("%s: cur_p[%3zu] = { id: %6d, p: %.6f, logit: %6.3f }\n", __func__, i, cur_p->data[i].id, cur_p->data[i].p, cur_p->data[i].logit);
+    }
+
+    size_org = cur_p->size;
+    p_sum = 0.0f;
+    thold = 1.0/(n_non_eog + 1);
+
+    cur_p->size = 0;
+
+    LOG_DBG_CUR("%s: applying thold = %.3f\n", __func__, thold);
+
+    for (size_t i = 0; i < size_org; ++i) {
+        const bool is_eog = llama_token_is_eog_impl(*ctx->vocab, cur_p->data[i].id);
+
+        if (cur_p->data[i].p < thold && !is_eog) {
+            continue;
+        }
+
+        p_sum += cur_p->data[i].p;
+
+        cur_p->data[cur_p->size++] = cur_p->data[i];
+    }
+
+    // normalize probs
+    for (size_t i = 0; i < cur_p->size; ++i) {
+        cur_p->data[i].p /= p_sum;
+
+        LOG_DBG_CUR("%s: cur_p[%3zu] = { id: %6d, p: %.6f, logit: %6.3f }\n", __func__, i, cur_p->data[i].id, cur_p->data[i].p, cur_p->data[i].logit);
+    }
+
+#undef LOG_DBG_CUR
+}
+
+static struct llama_sampler * llama_sampler_infill_clone(const struct llama_sampler * smpl) {
+    const auto * ctx = (const llama_sampler_infill *) smpl->ctx;
+    return llama_sampler_init_infill_impl(*ctx->vocab);
+}
+
+static void llama_sampler_infill_free(struct llama_sampler * smpl) {
+    delete (llama_sampler_infill *) smpl->ctx;
+}
+
+static struct llama_sampler_i llama_sampler_infill_i = {
+    /* .name   = */ llama_sampler_infill_name,
+    /* .accept = */ nullptr,
+    /* .apply  = */ llama_sampler_infill_apply,
+    /* .reset  = */ nullptr,
+    /* .clone  = */ llama_sampler_infill_clone,
+    /* .free   = */ llama_sampler_infill_free,
+};
+
+struct llama_sampler * llama_sampler_init_infill_impl(
+        const struct llama_vocab & vocab) {
+    return new llama_sampler {
+        /* .iface = */ &llama_sampler_infill_i,
+        /* .ctx   = */ new llama_sampler_infill {
+            /* .vocab = */ &vocab,
+            /* .buf0 = */ std::vector<char>(512),
+            /* .buf1 = */ std::vector<char>(512),
+        },
+    };
+}
+
 // utils
 
 uint32_t llama_sampler_get_seed(const struct llama_sampler * smpl) {

examples/talk-llama/llama-sampling.h

@@ -4,8 +4,6 @@
 
 #include "llama-grammar.h"
 
-#include <unordered_map>
-
 struct llama_vocab;
 struct llama_grammar;
 
@@ -27,3 +25,24 @@ struct llama_sampler * llama_sampler_init_grammar_impl(
         const struct llama_vocab & vocab,
                       const char * grammar_str,
                       const char * grammar_root);
+
+struct llama_sampler * llama_sampler_init_infill_impl(
+        const struct llama_vocab & vocab);
+
+struct llama_sampler * llama_sampler_init_dry_impl(
+        const struct llama_vocab &  vocab,
+                         int32_t    context_size,
+                           float    dry_multiplier,
+                           float    dry_base,
+                         int32_t    dry_allowed_length,
+                         int32_t    dry_penalty_last_n,
+                      const char ** seq_breakers,
+                          size_t    num_breakers);
+
+struct llama_sampler * llama_sampler_init_dry_testing(
+                         int32_t   context_size,
+                           float   dry_multiplier,
+                           float   dry_base,
+                         int32_t   dry_allowed_length,
+                         int32_t   dry_penalty_last_n,
+  const std::vector<std::vector<llama_token>>& seq_breakers);

examples/talk-llama/llama-vocab.cpp

@@ -221,7 +221,7 @@ struct llm_tokenizer_spm_session {
         }
 
         // seed the work queue with all possible 2-character tokens.
-        for (size_t i = 1; i < symbols.size(); ++i) {
+        for (int i = 1; i < (int) symbols.size(); ++i) {
             try_add_bigram(i - 1, i);
         }
 
@@ -563,7 +563,7 @@ struct llm_tokenizer_bpe_session {
                 index++;
                 symbols.emplace_back(sym);
             }
-            for (size_t i = 1; i < symbols.size(); ++i) {
+            for (int i = 1; i < (int) symbols.size(); ++i) {
                 add_new_bigram(i - 1, i);
             }
 
@@ -1663,6 +1663,14 @@ llama_token llama_token_eos_impl(const struct llama_vocab & vocab) {
     return vocab.special_eos_id;
 }
 
+llama_token llama_token_eot_impl(const struct llama_vocab & vocab) {
+    return vocab.special_eot_id;
+}
+
+llama_token llama_token_eom_impl(const struct llama_vocab & vocab) {
+    return vocab.special_eom_id;
+}
+
 llama_token llama_token_cls_impl(const struct llama_vocab & vocab) {
     return vocab.special_cls_id;
 }
@@ -1688,23 +1696,39 @@ bool llama_add_eos_token_impl(const struct llama_vocab & vocab) {
 }
 
 llama_token llama_token_prefix_impl(const struct llama_vocab & vocab) {
-    return vocab.special_prefix_id;
+    return vocab.special_fim_pre_id;
 }
 
 llama_token llama_token_middle_impl(const struct llama_vocab & vocab) {
-    return vocab.special_middle_id;
+    return vocab.special_fim_mid_id;
 }
 
 llama_token llama_token_suffix_impl(const struct llama_vocab & vocab) {
-    return vocab.special_suffix_id;
+    return vocab.special_fim_suf_id;
 }
 
-llama_token llama_token_eot_impl(const struct llama_vocab & vocab) {
-    return vocab.special_eot_id;
+llama_token llama_token_fim_pre_impl(const struct llama_vocab & vocab) {
+    return vocab.special_fim_pre_id;
 }
 
-llama_token llama_token_eom_impl(const struct llama_vocab & vocab) {
-    return vocab.special_eom_id;
+llama_token llama_token_fim_suf_impl(const struct llama_vocab & vocab) {
+    return vocab.special_fim_suf_id;
+}
+
+llama_token llama_token_fim_mid_impl(const struct llama_vocab & vocab) {
+    return vocab.special_fim_mid_id;
+}
+
+llama_token llama_token_fim_pad_impl(const struct llama_vocab & vocab) {
+    return vocab.special_fim_pad_id;
+}
+
+llama_token llama_token_fim_rep_impl(const struct llama_vocab & vocab) {
+    return vocab.special_fim_rep_id;
+}
+
+llama_token llama_token_fim_sep_impl(const struct llama_vocab & vocab) {
+    return vocab.special_fim_sep_id;
 }
 
 int32_t llama_tokenize_impl(
@@ -1942,3 +1966,19 @@ int32_t llama_detokenize_impl(
 
     return total <= text_len_max ? total : -total;
 }
+
+std::string llama_detokenize(const struct llama_vocab & vocab, const std::vector<llama_token> & tokens, bool special) {
+    std::string text;
+    text.resize(std::max(text.capacity(), tokens.size()));
+    int32_t n_chars = llama_detokenize_impl(vocab, tokens.data(), (int32_t)tokens.size(), &text[0], (int32_t)text.size(), false, special);
+    if (n_chars < 0) {
+        text.resize(-n_chars);
+        n_chars = llama_detokenize_impl(vocab, tokens.data(), (int32_t)tokens.size(), &text[0], (int32_t)text.size(), false, special);
+        GGML_ASSERT(n_chars <= (int32_t)text.size());  // whitespace trimming is performed after per-token detokenization
+    }
+
+    text.resize(n_chars);
+
+    // NOTE: the original tokenizer decodes bytes after collecting the pieces.
+    return text;
+}

examples/talk-llama/llama-vocab.h

@@ -37,20 +37,26 @@ struct llama_vocab {
     std::map<std::pair<std::string, std::string>, int> bpe_ranks;
 
     // default LLaMA special tokens
+    // TODO: should we set all of these to LLAMA_TOKEN_NULL?
     id special_bos_id  = 1;
     id special_eos_id  = 2;
+    id special_eot_id  = LLAMA_TOKEN_NULL;
+    id special_eom_id  = LLAMA_TOKEN_NULL;
     id special_unk_id  = 0;
-    id special_sep_id  = -1;
-    id special_pad_id  = -1;
-    id special_cls_id  = -1;
-    id special_mask_id = -1;
+    id special_sep_id  = LLAMA_TOKEN_NULL;
+    id special_pad_id  = LLAMA_TOKEN_NULL;
+    id special_cls_id  = LLAMA_TOKEN_NULL;
+    id special_mask_id = LLAMA_TOKEN_NULL;
 
-    id linefeed_id       = 13;
-    id special_prefix_id = -1;
-    id special_suffix_id = -1;
-    id special_middle_id = -1;
-    id special_eot_id    = -1; // TODO: move above after "eos_id", and here add "file separator" token
-    id special_eom_id    = -1;
+    id linefeed_id = 13;
+
+    // fim tokens
+    id special_fim_pre_id = LLAMA_TOKEN_NULL;
+    id special_fim_suf_id = LLAMA_TOKEN_NULL;
+    id special_fim_mid_id = LLAMA_TOKEN_NULL;
+    id special_fim_pad_id = LLAMA_TOKEN_NULL;
+    id special_fim_rep_id = LLAMA_TOKEN_NULL; // repo
+    id special_fim_sep_id = LLAMA_TOKEN_NULL; // file separator
 
     // set of all tokens that cause "end of generation"
     std::set<id> special_eog_ids;
@@ -104,19 +110,26 @@ bool llama_token_is_control_impl(const struct llama_vocab & vocab, llama_token t
 
 llama_token llama_token_bos_impl(const struct llama_vocab & vocab);
 llama_token llama_token_eos_impl(const struct llama_vocab & vocab);
+llama_token llama_token_eot_impl(const struct llama_vocab & vocab);
+llama_token llama_token_eom_impl(const struct llama_vocab & vocab);
 llama_token llama_token_cls_impl(const struct llama_vocab & vocab);
 llama_token llama_token_sep_impl(const struct llama_vocab & vocab);
 llama_token llama_token_nl_impl (const struct llama_vocab & vocab);
 llama_token llama_token_pad_impl(const struct llama_vocab & vocab);
 
-bool llama_add_bos_token_impl(const struct llama_vocab & vocab);
-bool llama_add_eos_token_impl(const struct llama_vocab & vocab);
-
 llama_token llama_token_prefix_impl(const struct llama_vocab & vocab);
 llama_token llama_token_middle_impl(const struct llama_vocab & vocab);
 llama_token llama_token_suffix_impl(const struct llama_vocab & vocab);
-llama_token llama_token_eot_impl   (const struct llama_vocab & vocab);
-llama_token llama_token_eom_impl   (const struct llama_vocab & vocab);
+
+llama_token llama_token_fim_pre_impl(const struct llama_vocab & vocab);
+llama_token llama_token_fim_suf_impl(const struct llama_vocab & vocab);
+llama_token llama_token_fim_mid_impl(const struct llama_vocab & vocab);
+llama_token llama_token_fim_pad_impl(const struct llama_vocab & vocab);
+llama_token llama_token_fim_rep_impl(const struct llama_vocab & vocab);
+llama_token llama_token_fim_sep_impl(const struct llama_vocab & vocab);
+
+bool llama_add_bos_token_impl(const struct llama_vocab & vocab);
+bool llama_add_eos_token_impl(const struct llama_vocab & vocab);
 
 int32_t llama_tokenize_impl(
         const struct llama_vocab & vocab,
@@ -136,6 +149,12 @@ int32_t llama_token_to_piece_impl(
                          int32_t   lstrip,
                             bool   special);
 
+// check if token0 is contained as a prefix in token1
+bool llama_token_is_prefix_impl(
+        const struct llama_vocab & vocab,
+                     llama_token   token0,
+                     llama_token   token1);
+
 int32_t llama_detokenize_impl(
         const struct llama_vocab & vocab,
                const llama_token * tokens,
@@ -144,3 +163,8 @@ int32_t llama_detokenize_impl(
                          int32_t   text_len_max,
                             bool   remove_special,
                             bool   unparse_special);
+
+std::string llama_detokenize(
+        const struct llama_vocab & vocab,
+  const std::vector<llama_token> & tokens,
+                            bool   special);

examples/talk-llama/llama.h

@@ -2,6 +2,7 @@
 #define LLAMA_H
 
 #include "ggml.h"
+#include "ggml-cpu.h"
 #include "ggml-backend.h"
 
 #include <stddef.h>
@@ -205,7 +206,7 @@ extern "C" {
     enum llama_split_mode {
         LLAMA_SPLIT_MODE_NONE  = 0, // single GPU
         LLAMA_SPLIT_MODE_LAYER = 1, // split layers and KV across GPUs
-        LLAMA_SPLIT_MODE_ROW   = 2, // split rows across GPUs
+        LLAMA_SPLIT_MODE_ROW   = 2, // split layers and KV across GPUs, use tensor parallelism if supported
     };
 
     // TODO: simplify (https://github.com/ggerganov/llama.cpp/pull/9294#pullrequestreview-2286561979)
@@ -217,6 +218,7 @@ extern "C" {
 
     typedef struct llama_token_data_array {
         // TODO: consider SoA
+        // NOTE: this pointer can be modified by the samplers
         llama_token_data * data;
         size_t size;
         int64_t selected; // this is the index in the data array (i.e. not the token id)
@@ -232,8 +234,11 @@ extern "C" {
     // - token  : the token ids of the input (used when embd is NULL)
     // - embd   : token embeddings (i.e. float vector of size n_embd) (used when token is NULL)
     // - pos    : the positions of the respective token in the sequence
+    //            (if set to NULL, the token position will be tracked automatically by llama_decode)
     // - seq_id : the sequence to which the respective token belongs
+    //            (if set to NULL, the sequence ID will be assumed to be 0)
     // - logits : if zero, the logits (and/or the embeddings) for the respective token will not be output
+    //            (if set to NULL, only the logits for last token will be returned)
     //
     typedef struct llama_batch {
         int32_t n_tokens;
@@ -244,15 +249,6 @@ extern "C" {
         int32_t      *  n_seq_id;
         llama_seq_id ** seq_id;
         int8_t       *  logits; // TODO: rename this to "output"
-
-        // NOTE: helpers for smooth API transition - can be deprecated in the future
-        //       for future-proof code, use the above fields instead and ignore everything below
-        //
-        // pos[i] = all_pos_0 + i*all_pos_1
-        //
-        llama_pos    all_pos_0;  // used if pos == NULL
-        llama_pos    all_pos_1;  // used if pos == NULL
-        llama_seq_id all_seq_id; // used if seq_id == NULL
     } llama_batch;
 
     enum llama_model_kv_override_type {
@@ -279,10 +275,7 @@ extern "C" {
         int32_t n_gpu_layers; // number of layers to store in VRAM
         enum llama_split_mode split_mode; // how to split the model across multiple GPUs
 
-        // main_gpu interpretation depends on split_mode:
-        // LLAMA_SPLIT_MODE_NONE: the GPU that is used for the entire model
-        // LLAMA_SPLIT_MODE_ROW: the GPU that is used for small tensors and intermediate results
-        // LLAMA_SPLIT_MODE_LAYER: ignored
+        // the GPU that is used for the entire model when split_mode is LLAMA_SPLIT_MODE_NONE
         int32_t main_gpu;
 
         // proportion of the model (layers or rows) to offload to each GPU, size: llama_max_devices()
@@ -433,6 +426,7 @@ extern "C" {
     LLAMA_API bool llama_supports_mmap       (void);
     LLAMA_API bool llama_supports_mlock      (void);
     LLAMA_API bool llama_supports_gpu_offload(void);
+    LLAMA_API bool llama_supports_rpc        (void);
 
     LLAMA_API uint32_t llama_n_ctx      (const struct llama_context * ctx);
     LLAMA_API uint32_t llama_n_batch    (const struct llama_context * ctx);
@@ -673,6 +667,9 @@ extern "C" {
     // Apply the KV cache updates (such as K-shifts, defragmentation, etc.)
     LLAMA_API void llama_kv_cache_update(struct llama_context * ctx);
 
+    // Check if the context supports KV cache shifting
+    LLAMA_API bool llama_kv_cache_can_shift(struct llama_context * ctx);
+
     //
     // State / sessions
     //
@@ -775,15 +772,15 @@ extern "C" {
     // Decoding
     //
 
-    // Return batch for single sequence of tokens starting at pos_0
+    // Return batch for single sequence of tokens
+    // The sequence ID will be fixed to 0
+    // The position of the tokens will be tracked automatically by llama_decode
     //
     // NOTE: this is a helper function to facilitate transition to the new batch API - avoid using it
     //
     LLAMA_API struct llama_batch llama_batch_get_one(
                   llama_token * tokens,
-                      int32_t   n_tokens,
-                    llama_pos   pos_0,
-                 llama_seq_id   seq_id);
+                      int32_t   n_tokens);
 
     // Allocates a batch of tokens on the heap that can hold a maximum of n_tokens
     // Each token can be assigned up to n_seq_max sequence ids
@@ -803,7 +800,7 @@ extern "C" {
     // Processes a batch of tokens with the ecoder part of the encoder-decoder model.
     // Stores the encoder output internally for later use by the decoder cross-attention layers.
     //   0 - success
-    // < 0 - error
+    // < 0 - error. the KV cache state is restored to the state before this call
     LLAMA_API int32_t llama_encode(
             struct llama_context * ctx,
               struct llama_batch   batch);
@@ -811,7 +808,7 @@ extern "C" {
     // Positive return values does not mean a fatal error, but rather a warning.
     //   0 - success
     //   1 - could not find a KV slot for the batch (try reducing the size of the batch or increase the context)
-    // < 0 - error
+    // < 0 - error. the KV cache state is restored to the state before this call
     LLAMA_API int32_t llama_decode(
             struct llama_context * ctx,
               struct llama_batch   batch);
@@ -896,6 +893,7 @@ extern "C" {
     // Special tokens
     LLAMA_API llama_token llama_token_bos(const struct llama_model * model); // beginning-of-sentence
     LLAMA_API llama_token llama_token_eos(const struct llama_model * model); // end-of-sentence
+    LLAMA_API llama_token llama_token_eot(const struct llama_model * model); // end-of-turn
     LLAMA_API llama_token llama_token_cls(const struct llama_model * model); // classification
     LLAMA_API llama_token llama_token_sep(const struct llama_model * model); // sentence separator
     LLAMA_API llama_token llama_token_nl (const struct llama_model * model); // next-line
@@ -904,11 +902,17 @@ extern "C" {
     LLAMA_API bool llama_add_bos_token(const struct llama_model * model);
     LLAMA_API bool llama_add_eos_token(const struct llama_model * model);
 
-    // Codellama infill tokens
-    LLAMA_API llama_token llama_token_prefix(const struct llama_model * model); // Beginning of infill prefix
-    LLAMA_API llama_token llama_token_middle(const struct llama_model * model); // Beginning of infill middle
-    LLAMA_API llama_token llama_token_suffix(const struct llama_model * model); // Beginning of infill suffix
-    LLAMA_API llama_token llama_token_eot   (const struct llama_model * model); // End of infill middle
+    // infill tokens
+    DEPRECATED(LLAMA_API llama_token llama_token_prefix(const struct llama_model * model), "use llama_token_fim_pre instead");
+    DEPRECATED(LLAMA_API llama_token llama_token_middle(const struct llama_model * model), "use llama_token_fim_mid instead");
+    DEPRECATED(LLAMA_API llama_token llama_token_suffix(const struct llama_model * model), "use llama_token_fim_suf instead");
+
+    LLAMA_API llama_token llama_token_fim_pre(const struct llama_model * model);
+    LLAMA_API llama_token llama_token_fim_suf(const struct llama_model * model);
+    LLAMA_API llama_token llama_token_fim_mid(const struct llama_model * model);
+    LLAMA_API llama_token llama_token_fim_pad(const struct llama_model * model);
+    LLAMA_API llama_token llama_token_fim_rep(const struct llama_model * model);
+    LLAMA_API llama_token llama_token_fim_sep(const struct llama_model * model);
 
     //
     // Tokenization
@@ -1067,12 +1071,13 @@ extern "C" {
 
     // available samplers:
 
-    LLAMA_API struct llama_sampler * llama_sampler_init_greedy     (void);
-    LLAMA_API struct llama_sampler * llama_sampler_init_dist       (uint32_t seed);
+    LLAMA_API struct llama_sampler * llama_sampler_init_greedy(void);
+    LLAMA_API struct llama_sampler * llama_sampler_init_dist  (uint32_t seed);
 
     /// @details Sorts candidate tokens by their logits in descending order and calculate probabilities based on logits.
     /// NOTE: Avoid using on the full vocabulary as the sorting can become slow. For example, apply top-k or top-p sampling first.
-    LLAMA_API struct llama_sampler * llama_sampler_init_softmax    (void);
+    DEPRECATED(LLAMA_API struct llama_sampler * llama_sampler_init_softmax    (void),
+        "will be removed in the future (see https://github.com/ggerganov/llama.cpp/pull/9896#discussion_r1800920915)");
 
     /// @details Top-K sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751
     LLAMA_API struct llama_sampler * llama_sampler_init_top_k      (int32_t k);
@@ -1083,16 +1088,18 @@ extern "C" {
     /// @details Minimum P sampling as described in https://github.com/ggerganov/llama.cpp/pull/3841
     LLAMA_API struct llama_sampler * llama_sampler_init_min_p      (float   p, size_t min_keep);
 
-    /// @details Tail Free Sampling described in https://www.trentonbricken.com/Tail-Free-Sampling/.
-    LLAMA_API struct llama_sampler * llama_sampler_init_tail_free  (float   z, size_t min_keep);
-
     /// @details Locally Typical Sampling implementation described in the paper https://arxiv.org/abs/2202.00666.
     LLAMA_API struct llama_sampler * llama_sampler_init_typical    (float   p, size_t min_keep);
+
+    /// #details Updates the logits l_i` = l_i/t. When t <= 0.0f, the maximum logit is kept at it's original value, the rest are set to -inf
     LLAMA_API struct llama_sampler * llama_sampler_init_temp       (float   t);
 
     /// @details Dynamic temperature implementation (a.k.a. entropy) described in the paper https://arxiv.org/abs/2309.02772.
     LLAMA_API struct llama_sampler * llama_sampler_init_temp_ext   (float   t, float   delta, float exponent);
 
+    /// @details XTC sampler as described in https://github.com/oobabooga/text-generation-webui/pull/6335
+    LLAMA_API struct llama_sampler * llama_sampler_init_xtc        (float   p, float   t,     size_t min_keep, uint32_t seed);
+
     /// @details Mirostat 1.0 algorithm described in the paper https://arxiv.org/abs/2007.14966. Uses tokens instead of words.
     /// @param candidates A vector of `llama_token_data` containing the candidate tokens, their probabilities (p), and log-odds (logit) for the current position in the generated text.
     /// @param tau  The target cross-entropy (or surprise) value you want to achieve for the generated text. A higher value corresponds to more surprising or less predictable text, while a lower value corresponds to less surprising or more predictable text.
@@ -1132,11 +1139,43 @@ extern "C" {
                                 bool   penalize_nl,     // consider newlines as a repeatable token
                                 bool   ignore_eos);     // ignore the end-of-sequence token
 
+    ///  @details DRY sampler, designed by p-e-w, as described in: https://github.com/oobabooga/text-generation-webui/pull/5677, porting Koboldcpp implementation authored by pi6am: https://github.com/LostRuins/koboldcpp/pull/982
+    LLAMA_API struct llama_sampler *    llama_sampler_init_dry(
+            const struct llama_model *  model,
+                               float    dry_multiplier,
+                               float    dry_base,
+                             int32_t    dry_allowed_length,
+                             int32_t    dry_penalty_last_n,
+                          const char ** seq_breakers,
+                              size_t    num_breakers);
+
     LLAMA_API struct llama_sampler * llama_sampler_init_logit_bias(
                              int32_t   n_vocab,
                              int32_t   n_logit_bias,
               const llama_logit_bias * logit_bias);
 
+    // this sampler is meant to be used for fill-in-the-middle infilling
+    // it's supposed to be used after top_k + top_p sampling
+    //
+    // 1. if the sum of the EOG probs times the number of candidates is higher than the sum of the other probs -> pick EOG
+    // 2. combine probs of tokens that have the same prefix
+    //
+    // example:
+    //
+    // - before:
+    //   "hel":   0.5
+    //   "hell":  0.2
+    //   "hello": 0.1
+    //   "dummy": 0.1
+    //
+    // - after:
+    //   "hel":   0.8
+    //   "dummy": 0.1
+    //
+    // 3. discard non-EOG tokens with low prob
+    // 4. if no tokens are left -> pick EOT
+    //
+    LLAMA_API struct llama_sampler * llama_sampler_init_infill(const struct llama_model * model);
 
     // Returns the seed used by the sampler if applicable, LLAMA_DEFAULT_SEED otherwise
     LLAMA_API uint32_t llama_sampler_get_seed(const struct llama_sampler * smpl);
@@ -1208,8 +1247,6 @@ extern "C" {
     LLAMA_API void                           llama_perf_sampler_print(const struct llama_sampler * chain);
     LLAMA_API void                           llama_perf_sampler_reset(      struct llama_sampler * chain);
 
-    LLAMA_API void llama_perf_dump_yaml(FILE * stream, const struct llama_context * ctx);
-
 #ifdef __cplusplus
 }
 #endif

examples/talk-llama/unicode-data.cpp

@@ -2311,7 +2311,7 @@ const std::unordered_set<uint32_t> unicode_set_whitespace = {
 0x003000,
 };
 
-// list is always in ascending order, to enable binary searh
+// list is always in ascending order, to enable binary search
 const std::initializer_list<std::pair<uint32_t, uint32_t>> unicode_map_lowercase = {
 {0x000041, 0x000061},
 {0x000042, 0x000062},
@@ -3748,7 +3748,7 @@ const std::initializer_list<std::pair<uint32_t, uint32_t>> unicode_map_lowercase
 {0x01E921, 0x01E943},
 };
 
-// list is always in ascending order, to enable binary searh
+// list is always in ascending order, to enable binary search
 const std::initializer_list<std::pair<uint32_t, uint32_t>> unicode_map_uppercase = {
 {0x000061, 0x000041},
 {0x000062, 0x000042},

examples/whisper.android/lib/src/main/jni/whisper/CMakeLists.txt

@@ -22,7 +22,13 @@ if (NOT GGML_HOME)
         ${WHISPER_LIB_DIR}/ggml/src/ggml-aarch64.c
         ${WHISPER_LIB_DIR}/ggml/src/ggml-alloc.c
         ${WHISPER_LIB_DIR}/ggml/src/ggml-backend.cpp
+        ${WHISPER_LIB_DIR}/ggml/src/ggml-backend-reg.cpp
         ${WHISPER_LIB_DIR}/ggml/src/ggml-quants.c
+        ${WHISPER_LIB_DIR}/ggml/src/ggml-threading.cpp
+        ${WHISPER_LIB_DIR}/ggml/src/ggml-cpu/ggml-cpu.c
+        ${WHISPER_LIB_DIR}/ggml/src/ggml-cpu/ggml-cpu.cpp
+        ${WHISPER_LIB_DIR}/ggml/src/ggml-cpu/ggml-cpu-aarch64.c
+        ${WHISPER_LIB_DIR}/ggml/src/ggml-cpu/ggml-cpu-quants.c
         )
 endif()
 

examples/whisper.objc/whisper.objc.xcodeproj/project.pbxproj

@@ -24,6 +24,12 @@
 		18A2760B2C2A9B43001C8D37 /* ggml-metal.metal in Resources */ = {isa = PBXBuildFile; fileRef = 1844471D2AB2195F007D6BFE /* ggml-metal.metal */; };
 		18ABE15A2AF556340044A204 /* ggml-backend.cpp in Sources */ = {isa = PBXBuildFile; fileRef = 18ABE1572AF556340044A204 /* ggml-backend.cpp */; };
 		18ABE15B2AF556340044A204 /* ggml-quants.c in Sources */ = {isa = PBXBuildFile; fileRef = 18ABE1592AF556340044A204 /* ggml-quants.c */; };
+		18E864A92CE73C1E0094B8B3 /* ggml-cpu.c in Sources */ = {isa = PBXBuildFile; fileRef = 18E864A82CE73C1E0094B8B3 /* ggml-cpu.c */; };
+		18F8C0BC2CEDF4DC00CAD607 /* ggml-threading.cpp in Sources */ = {isa = PBXBuildFile; fileRef = 18F8C0BB2CEDF4DC00CAD607 /* ggml-threading.cpp */; };
+		18F8C0BE2CEDF50700CAD607 /* ggml-cpu.cpp in Sources */ = {isa = PBXBuildFile; fileRef = 18F8C0BD2CEDF50700CAD607 /* ggml-cpu.cpp */; };
+		18F8C0C42CEDF52700CAD607 /* ggml-cpu-aarch64.c in Sources */ = {isa = PBXBuildFile; fileRef = 18F8C0C02CEDF52700CAD607 /* ggml-cpu-aarch64.c */; };
+		18F8C0C52CEDF52700CAD607 /* ggml-cpu-quants.c in Sources */ = {isa = PBXBuildFile; fileRef = 18F8C0C32CEDF52700CAD607 /* ggml-cpu-quants.c */; };
+		18F8C0C72CEDF7AB00CAD607 /* ggml-backend-reg.cpp in Sources */ = {isa = PBXBuildFile; fileRef = 18F8C0C62CEDF7AB00CAD607 /* ggml-backend-reg.cpp */; };
 		7FE3424B2A0C3FA20015A058 /* whisper-encoder-impl.m in Sources */ = {isa = PBXBuildFile; fileRef = 7FE342452A0C3FA20015A058 /* whisper-encoder-impl.m */; };
 		7FE3424C2A0C3FA20015A058 /* whisper-encoder.mm in Sources */ = {isa = PBXBuildFile; fileRef = 7FE342472A0C3FA20015A058 /* whisper-encoder.mm */; };
 		7FE3424D2A0C3FA20015A058 /* whisper-decoder-impl.m in Sources */ = {isa = PBXBuildFile; fileRef = 7FE3424A2A0C3FA20015A058 /* whisper-decoder-impl.m */; };
@@ -49,8 +55,8 @@
 		18133C7F2C64E342005CEAAC /* ggml-aarch64.c */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.c; name = "ggml-aarch64.c"; path = "../../../ggml/src/ggml-aarch64.c"; sourceTree = "<group>"; };
 		184447182AB211A2007D6BFE /* ggml-alloc.c */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.c; name = "ggml-alloc.c"; path = "../../../ggml/src/ggml-alloc.c"; sourceTree = "<group>"; };
 		184447192AB211A2007D6BFE /* ggml-alloc.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; name = "ggml-alloc.h"; path = "../../../ggml/include/ggml-alloc.h"; sourceTree = "<group>"; };
-		1844471B2AB21655007D6BFE /* ggml-metal.m */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.objc; name = "ggml-metal.m"; path = "../../../ggml/src/ggml-metal.m"; sourceTree = "<group>"; };
-		1844471D2AB2195F007D6BFE /* ggml-metal.metal */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.metal; name = "ggml-metal.metal"; path = "../../../ggml/src/ggml-metal.metal"; sourceTree = "<group>"; };
+		1844471B2AB21655007D6BFE /* ggml-metal.m */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.objc; name = "ggml-metal.m"; path = "../../../ggml/src/ggml-metal/ggml-metal.m"; sourceTree = "<group>"; };
+		1844471D2AB2195F007D6BFE /* ggml-metal.metal */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.metal; name = "ggml-metal.metal"; path = "../../../ggml/src/ggml-metal/ggml-metal.metal"; sourceTree = "<group>"; };
 		18627C7629052BDF00BD2A04 /* whisper.objc.app */ = {isa = PBXFileReference; explicitFileType = wrapper.application; includeInIndex = 0; path = whisper.objc.app; sourceTree = BUILT_PRODUCTS_DIR; };
 		18627C7929052BDF00BD2A04 /* AppDelegate.h */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.c.h; path = AppDelegate.h; sourceTree = "<group>"; };
 		18627C7A29052BDF00BD2A04 /* AppDelegate.m */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.c.objc; path = AppDelegate.m; sourceTree = "<group>"; };
@@ -76,6 +82,17 @@
 		18ABE1572AF556340044A204 /* ggml-backend.cpp */ = {isa = PBXFileReference; explicitFileType = sourcecode.cpp.cpp; fileEncoding = 4; name = "ggml-backend.cpp"; path = "../../../ggml/src/ggml-backend.cpp"; sourceTree = "<group>"; };
 		18ABE1582AF556340044A204 /* ggml-impl.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; name = "ggml-impl.h"; path = "../../../ggml/src/ggml-impl.h"; sourceTree = "<group>"; };
 		18ABE1592AF556340044A204 /* ggml-quants.c */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.c; name = "ggml-quants.c"; path = "../../../ggml/src/ggml-quants.c"; sourceTree = "<group>"; };
+		18E864A82CE73C1E0094B8B3 /* ggml-cpu.c */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.c.c; name = "ggml-cpu.c"; path = "../../../ggml/src/ggml-cpu/ggml-cpu.c"; sourceTree = "<group>"; };
+		18E864AA2CE73C580094B8B3 /* ggml-cpu.h */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.c.h; name = "ggml-cpu.h"; path = "../../../ggml/include/ggml-cpu.h"; sourceTree = "<group>"; };
+		18F8C0BA2CEDF4DC00CAD607 /* ggml-threading.h */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.c.h; name = "ggml-threading.h"; path = "../../../ggml/src/ggml-threading.h"; sourceTree = "<group>"; };
+		18F8C0BB2CEDF4DC00CAD607 /* ggml-threading.cpp */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.cpp.cpp; name = "ggml-threading.cpp"; path = "../../../ggml/src/ggml-threading.cpp"; sourceTree = "<group>"; };
+		18F8C0BD2CEDF50700CAD607 /* ggml-cpu.cpp */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.cpp.cpp; name = "ggml-cpu.cpp"; path = "../../../ggml/src/ggml-cpu/ggml-cpu.cpp"; sourceTree = "<group>"; };
+		18F8C0BF2CEDF52700CAD607 /* ggml-cpu-aarch64.h */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.c.h; name = "ggml-cpu-aarch64.h"; path = "../../../ggml/src/ggml-cpu/ggml-cpu-aarch64.h"; sourceTree = "<group>"; };
+		18F8C0C02CEDF52700CAD607 /* ggml-cpu-aarch64.c */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.c.c; name = "ggml-cpu-aarch64.c"; path = "../../../ggml/src/ggml-cpu/ggml-cpu-aarch64.c"; sourceTree = "<group>"; };
+		18F8C0C12CEDF52700CAD607 /* ggml-cpu-impl.h */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.c.h; name = "ggml-cpu-impl.h"; path = "../../../ggml/src/ggml-cpu/ggml-cpu-impl.h"; sourceTree = "<group>"; };
+		18F8C0C22CEDF52700CAD607 /* ggml-cpu-quants.h */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.c.h; name = "ggml-cpu-quants.h"; path = "../../../ggml/src/ggml-cpu/ggml-cpu-quants.h"; sourceTree = "<group>"; };
+		18F8C0C32CEDF52700CAD607 /* ggml-cpu-quants.c */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.c.c; name = "ggml-cpu-quants.c"; path = "../../../ggml/src/ggml-cpu/ggml-cpu-quants.c"; sourceTree = "<group>"; };
+		18F8C0C62CEDF7AB00CAD607 /* ggml-backend-reg.cpp */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.cpp.cpp; name = "ggml-backend-reg.cpp"; path = "../../../ggml/src/ggml-backend-reg.cpp"; sourceTree = "<group>"; };
 		7FE342452A0C3FA20015A058 /* whisper-encoder-impl.m */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.objc; path = "whisper-encoder-impl.m"; sourceTree = "<group>"; };
 		7FE342462A0C3FA20015A058 /* whisper-encoder.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; path = "whisper-encoder.h"; sourceTree = "<group>"; };
 		7FE342472A0C3FA20015A058 /* whisper-encoder.mm */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.cpp.objcpp; path = "whisper-encoder.mm"; sourceTree = "<group>"; };
@@ -115,6 +132,17 @@
 		18627C7829052BDF00BD2A04 /* whisper.objc */ = {
 			isa = PBXGroup;
 			children = (
+				18F8C0C62CEDF7AB00CAD607 /* ggml-backend-reg.cpp */,
+				18F8C0BF2CEDF52700CAD607 /* ggml-cpu-aarch64.h */,
+				18F8C0C02CEDF52700CAD607 /* ggml-cpu-aarch64.c */,
+				18F8C0C12CEDF52700CAD607 /* ggml-cpu-impl.h */,
+				18F8C0C22CEDF52700CAD607 /* ggml-cpu-quants.h */,
+				18F8C0C32CEDF52700CAD607 /* ggml-cpu-quants.c */,
+				18F8C0BD2CEDF50700CAD607 /* ggml-cpu.cpp */,
+				18F8C0BA2CEDF4DC00CAD607 /* ggml-threading.h */,
+				18F8C0BB2CEDF4DC00CAD607 /* ggml-threading.cpp */,
+				18E864AA2CE73C580094B8B3 /* ggml-cpu.h */,
+				18E864A82CE73C1E0094B8B3 /* ggml-cpu.c */,
 				18133C7F2C64E342005CEAAC /* ggml-aarch64.c */,
 				18133C7E2C64E342005CEAAC /* ggml-aarch64.h */,
 				18A275FF2C2A9563001C8D37 /* ggml-common.h */,
@@ -247,10 +275,16 @@
 				18627C9629052C5800BD2A04 /* ggml.c in Sources */,
 				18627C7B29052BDF00BD2A04 /* AppDelegate.m in Sources */,
 				7FE3424D2A0C3FA20015A058 /* whisper-decoder-impl.m in Sources */,
+				18F8C0C72CEDF7AB00CAD607 /* ggml-backend-reg.cpp in Sources */,
+				18F8C0BE2CEDF50700CAD607 /* ggml-cpu.cpp in Sources */,
 				1844471A2AB211A2007D6BFE /* ggml-alloc.c in Sources */,
+				18F8C0C42CEDF52700CAD607 /* ggml-cpu-aarch64.c in Sources */,
+				18F8C0C52CEDF52700CAD607 /* ggml-cpu-quants.c in Sources */,
+				18E864A92CE73C1E0094B8B3 /* ggml-cpu.c in Sources */,
 				18ABE15A2AF556340044A204 /* ggml-backend.cpp in Sources */,
 				18627C8C29052BE000BD2A04 /* main.m in Sources */,
 				18627C7E29052BDF00BD2A04 /* SceneDelegate.m in Sources */,
+				18F8C0BC2CEDF4DC00CAD607 /* ggml-threading.cpp in Sources */,
 				1844471C2AB21655007D6BFE /* ggml-metal.m in Sources */,
 				7FE3424B2A0C3FA20015A058 /* whisper-encoder-impl.m in Sources */,
 			);
@@ -329,6 +363,7 @@
 				GCC_WARN_UNINITIALIZED_AUTOS = YES_AGGRESSIVE;
 				GCC_WARN_UNUSED_FUNCTION = YES;
 				GCC_WARN_UNUSED_VARIABLE = YES;
+				HEADER_SEARCH_PATHS = "";
 				IPHONEOS_DEPLOYMENT_TARGET = 16.0;
 				MTL_ENABLE_DEBUG_INFO = INCLUDE_SOURCE;
 				MTL_FAST_MATH = YES;
@@ -382,6 +417,7 @@
 				GCC_WARN_UNINITIALIZED_AUTOS = YES_AGGRESSIVE;
 				GCC_WARN_UNUSED_FUNCTION = YES;
 				GCC_WARN_UNUSED_VARIABLE = YES;
+				HEADER_SEARCH_PATHS = "";
 				IPHONEOS_DEPLOYMENT_TARGET = 16.0;
 				MTL_ENABLE_DEBUG_INFO = NO;
 				MTL_FAST_MATH = YES;
@@ -404,6 +440,7 @@
 				DEVELOPMENT_TEAM = P8JZH34X63;
 				GCC_WARN_64_TO_32_BIT_CONVERSION = NO;
 				GENERATE_INFOPLIST_FILE = YES;
+				HEADER_SEARCH_PATHS = ../../../ggml/src/;
 				INFOPLIST_FILE = whisper.objc/Info.plist;
 				INFOPLIST_KEY_UIApplicationSupportsIndirectInputEvents = YES;
 				INFOPLIST_KEY_UILaunchStoryboardName = LaunchScreen;
@@ -433,6 +470,7 @@
 				DEVELOPMENT_TEAM = P8JZH34X63;
 				GCC_WARN_64_TO_32_BIT_CONVERSION = NO;
 				GENERATE_INFOPLIST_FILE = YES;
+				HEADER_SEARCH_PATHS = ../../../ggml/src/;
 				INFOPLIST_FILE = whisper.objc/Info.plist;
 				INFOPLIST_KEY_UIApplicationSupportsIndirectInputEvents = YES;
 				INFOPLIST_KEY_UILaunchStoryboardName = LaunchScreen;

examples/whisper.swiftui/whisper.cpp.swift/LibWhisper.swift

@@ -1,4 +1,5 @@
 import Foundation
+import UIKit
 import whisper
 
 enum WhisperError: Error {
@@ -55,11 +56,91 @@ actor WhisperContext {
         return transcription
     }
 
+    static func benchMemcpy(nThreads: Int32) async -> String {
+        return String.init(cString: whisper_bench_memcpy_str(nThreads))
+    }
+
+    static func benchGgmlMulMat(nThreads: Int32) async -> String {
+        return String.init(cString: whisper_bench_ggml_mul_mat_str(nThreads))
+    }
+
+    private func systemInfo() -> String {
+        var info = ""
+        if (ggml_cpu_has_neon() != 0) { info += "NEON " }
+        return String(info.dropLast())
+    }
+
+    func benchFull(modelName: String, nThreads: Int32) async -> String {
+        let nMels = whisper_model_n_mels(context)
+        if (whisper_set_mel(context, nil, 0, nMels) != 0) {
+            return "error: failed to set mel"
+        }
+        
+        // heat encoder
+        if (whisper_encode(context, 0, nThreads) != 0) {
+            return "error: failed to encode"
+        }
+        
+        var tokens = [whisper_token](repeating: 0, count: 512)
+        
+        // prompt heat
+        if (whisper_decode(context, &tokens, 256, 0, nThreads) != 0) {
+            return "error: failed to decode"
+        }
+        
+        // text-generation heat
+        if (whisper_decode(context, &tokens, 1, 256, nThreads) != 0) {
+            return "error: failed to decode"
+        }
+        
+        whisper_reset_timings(context)
+        
+        // actual run
+        if (whisper_encode(context, 0, nThreads) != 0) {
+            return "error: failed to encode"
+        }
+        
+        // text-generation
+        for i in 0..<256 {
+            if (whisper_decode(context, &tokens, 1, Int32(i), nThreads) != 0) {
+                return "error: failed to decode"
+            }
+        }
+        
+        // batched decoding
+        for _ in 0..<64 {
+            if (whisper_decode(context, &tokens, 5, 0, nThreads) != 0) {
+                return "error: failed to decode"
+            }
+        }
+        
+        // prompt processing
+        for _ in 0..<16 {
+            if (whisper_decode(context, &tokens, 256, 0, nThreads) != 0) {
+                return "error: failed to decode"
+            }
+        }
+
+        whisper_print_timings(context)
+
+        let deviceModel = await UIDevice.current.model
+        let systemName = await UIDevice.current.systemName
+        let systemInfo = self.systemInfo()
+        let timings: whisper_timings = whisper_get_timings(context).pointee
+        let encodeMs = String(format: "%.2f", timings.encode_ms)
+        let decodeMs = String(format: "%.2f", timings.decode_ms)
+        let batchdMs = String(format: "%.2f", timings.batchd_ms)
+        let promptMs = String(format: "%.2f", timings.prompt_ms)
+        return "| \(deviceModel) | \(systemName) | \(systemInfo) | \(modelName) | \(nThreads) | 1 | \(encodeMs) | \(decodeMs) | \(batchdMs) | \(promptMs) | <todo> |"
+    }
+
     static func createContext(path: String) throws -> WhisperContext {
         var params = whisper_context_default_params()
 #if targetEnvironment(simulator)
         params.use_gpu = false
         print("Running on the simulator, using CPU")
+#else
+        params.flash_attn = true // Enabled by default for Metal
 #endif
         let context = whisper_init_from_file_with_params(path, params)
         if let context {

examples/whisper.swiftui/whisper.swiftui.demo/Models/Model.swift

@@ -0,0 +1,17 @@
+import Foundation
+
+struct Model: Identifiable {
+    var id = UUID()
+    var name: String
+    var info: String
+    var url: String
+
+    var filename: String
+    var fileURL: URL {
+        FileManager.default.urls(for: .documentDirectory, in: .userDomainMask)[0].appendingPathComponent(filename)
+    }
+
+    func fileExists() -> Bool {
+        FileManager.default.fileExists(atPath: fileURL.path)
+    }
+}

examples/whisper.swiftui/whisper.swiftui.demo/Models/WhisperState.swift

@@ -14,7 +14,7 @@ class WhisperState: NSObject, ObservableObject, AVAudioRecorderDelegate {
     private var recordedFile: URL? = nil
     private var audioPlayer: AVAudioPlayer?
     
-    private var modelUrl: URL? {
+    private var builtInModelUrl: URL? {
         Bundle.main.url(forResource: "ggml-base.en", withExtension: "bin", subdirectory: "models")
     }
     
@@ -28,23 +28,59 @@ class WhisperState: NSObject, ObservableObject, AVAudioRecorderDelegate {
     
     override init() {
         super.init()
+        loadModel()
+    }
+    
+    func loadModel(path: URL? = nil, log: Bool = true) {
         do {
-            try loadModel()
+            whisperContext = nil
+            if (log) { messageLog += "Loading model...\n" }
+            let modelUrl = path ?? builtInModelUrl
+            if let modelUrl {
+                whisperContext = try WhisperContext.createContext(path: modelUrl.path())
+                if (log) { messageLog += "Loaded model \(modelUrl.lastPathComponent)\n" }
+            } else {
+                if (log) { messageLog += "Could not locate model\n" }
+            }
             canTranscribe = true
         } catch {
             print(error.localizedDescription)
-            messageLog += "\(error.localizedDescription)\n"
+            if (log) { messageLog += "\(error.localizedDescription)\n" }
         }
     }
-    
-    private func loadModel() throws {
-        messageLog += "Loading model...\n"
-        if let modelUrl {
-            whisperContext = try WhisperContext.createContext(path: modelUrl.path())
-            messageLog += "Loaded model \(modelUrl.lastPathComponent)\n"
-        } else {
-            messageLog += "Could not locate model\n"
+
+    func benchCurrentModel() async {
+        if whisperContext == nil {
+            messageLog += "Cannot bench without loaded model\n"
+            return
         }
+        messageLog += "Running benchmark for loaded model\n"
+        let result = await whisperContext?.benchFull(modelName: "<current>", nThreads: Int32(min(4, cpuCount())))
+        if (result != nil) { messageLog += result! + "\n" }
+    }
+
+    func bench(models: [Model]) async {
+        let nThreads = Int32(min(4, cpuCount()))
+
+//        messageLog += "Running memcpy benchmark\n"
+//        messageLog += await WhisperContext.benchMemcpy(nThreads: nThreads) + "\n"
+//
+//        messageLog += "Running ggml_mul_mat benchmark with \(nThreads) threads\n"
+//        messageLog += await WhisperContext.benchGgmlMulMat(nThreads: nThreads) + "\n"
+
+        messageLog += "Running benchmark for all downloaded models\n"
+        messageLog += "| CPU | OS | Config | Model | Th | FA | Enc. | Dec. | Bch5 | PP | Commit |\n"
+        messageLog += "| --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |\n"
+        for model in models {
+            loadModel(path: model.fileURL, log: false)
+            if whisperContext == nil {
+                messageLog += "Cannot bench without loaded model\n"
+                break
+            }
+            let result = await whisperContext?.benchFull(modelName: model.name, nThreads: nThreads)
+            if (result != nil) { messageLog += result! + "\n" }
+        }
+        messageLog += "Benchmarking completed\n"
     }
     
     func transcribeSample() async {
@@ -160,3 +196,8 @@ class WhisperState: NSObject, ObservableObject, AVAudioRecorderDelegate {
         isRecording = false
     }
 }
+
+
+fileprivate func cpuCount() -> Int {
+    ProcessInfo.processInfo.processorCount
+}

examples/whisper.swiftui/whisper.swiftui.demo/UI/ContentView.swift

@@ -1,5 +1,6 @@
 import SwiftUI
 import AVFoundation
+import Foundation
 
 struct ContentView: View {
     @StateObject var whisperState = WhisperState()
@@ -29,15 +30,125 @@ struct ContentView: View {
                     Text(verbatim: whisperState.messageLog)
                         .frame(maxWidth: .infinity, alignment: .leading)
                 }
+                .font(.footnote)
+                .padding()
+                .background(Color.gray.opacity(0.1))
+                .cornerRadius(10)
+
+                HStack {
+                    Button("Clear Logs", action: {
+                        whisperState.messageLog = ""
+                    })
+                    .font(.footnote)
+                    .buttonStyle(.bordered)
+
+                    Button("Copy Logs", action: {
+                        UIPasteboard.general.string = whisperState.messageLog
+                    })
+                    .font(.footnote)
+                    .buttonStyle(.bordered)
+
+                    Button("Bench", action: {
+                        Task {
+                            await whisperState.benchCurrentModel()
+                        }
+                    })
+                    .font(.footnote)
+                    .buttonStyle(.bordered)
+                    .disabled(!whisperState.canTranscribe)
+
+                    Button("Bench All", action: {
+                        Task {
+                            await whisperState.bench(models: ModelsView.getDownloadedModels())
+                        }
+                    })
+                    .font(.footnote)
+                    .buttonStyle(.bordered)
+                    .disabled(!whisperState.canTranscribe)
+                }
+
+                NavigationLink(destination: ModelsView(whisperState: whisperState)) {
+                    Text("View Models")
+                }
+                .font(.footnote)
+                .padding()
             }
             .navigationTitle("Whisper SwiftUI Demo")
             .padding()
         }
     }
-}
 
-struct ContentView_Previews: PreviewProvider {
-    static var previews: some View {
-        ContentView()
+    struct ModelsView: View {
+        @ObservedObject var whisperState: WhisperState
+        @Environment(\.dismiss) var dismiss
+        
+        private static let models: [Model] = [
+            Model(name: "tiny", info: "(F16, 75 MiB)", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-tiny.bin", filename: "tiny.bin"),
+            Model(name: "tiny-q5_1", info: "(31 MiB)", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-tiny-q5_1.bin", filename: "tiny-q5_1.bin"),
+            Model(name: "tiny-q8_0", info: "(42 MiB)", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-tiny-q8_0.bin", filename: "tiny-q8_0.bin"),
+            Model(name: "tiny.en", info: "(F16, 75 MiB)", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-tiny.en.bin", filename: "tiny.en.bin"),
+            Model(name: "tiny.en-q5_1", info: "(31 MiB)", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-tiny.en-q5_1.bin", filename: "tiny.en-q5_1.bin"),
+            Model(name: "tiny.en-q8_0", info: "(42 MiB)", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-tiny.en-q8_0.bin", filename: "tiny.en-q8_0.bin"),
+            Model(name: "base", info: "(F16, 142 MiB)", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-base.bin", filename: "base.bin"),
+            Model(name: "base-q5_1", info: "(57 MiB)", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-base-q5_1.bin", filename: "base-q5_1.bin"),
+            Model(name: "base-q8_0", info: "(78 MiB)", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-base-q8_0.bin", filename: "base-q8_0.bin"),
+            Model(name: "base.en", info: "(F16, 142 MiB)", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-base.en.bin", filename: "base.en.bin"),
+            Model(name: "base.en-q5_1", info: "(57 MiB)", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-base.en-q5_1.bin", filename: "base.en-q5_1.bin"),
+            Model(name: "base.en-q8_0", info: "(78 MiB)", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-base.en-q8_0.bin", filename: "base.en-q8_0.bin"),
+            Model(name: "small", info: "(F16, 466 MiB)", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-small.bin", filename: "small.bin"),
+            Model(name: "small-q5_1", info: "(181 MiB)", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-small-q5_1.bin", filename: "small-q5_1.bin"),
+            Model(name: "small-q8_0", info: "(252 MiB)", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-small-q8_0.bin", filename: "small-q8_0.bin"),
+            Model(name: "small.en", info: "(F16, 466 MiB)", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-small.en.bin", filename: "small.en.bin"),
+            Model(name: "small.en-q5_1", info: "(181 MiB)", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-small.en-q5_1.bin", filename: "small.en-q5_1.bin"),
+            Model(name: "small.en-q8_0", info: "(252 MiB)", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-small.en-q8_0.bin", filename: "small.en-q8_0.bin"),
+            Model(name: "medium", info: "(F16, 1.5 GiB)", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-medium.bin", filename: "medium.bin"),
+            Model(name: "medium-q5_0", info: "(514 MiB)", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-medium-q5_0.bin", filename: "medium-q5_0.bin"),
+            Model(name: "medium-q8_0", info: "(785 MiB)", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-medium-q8_0.bin", filename: "medium-q8_0.bin"),
+            Model(name: "medium.en", info: "(F16, 1.5 GiB)", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-medium.en.bin", filename: "medium.en.bin"),
+            Model(name: "medium.en-q5_0", info: "(514 MiB)", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-medium.en-q5_0.bin", filename: "medium.en-q5_0.bin"),
+            Model(name: "medium.en-q8_0", info: "(785 MiB)", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-medium.en-q8_0.bin", filename: "medium.en-q8_0.bin"),
+            Model(name: "large-v1", info: "(F16, 2.9 GiB)", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-large.bin", filename: "large.bin"),
+            Model(name: "large-v2", info: "(F16, 2.9 GiB)", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-large-v2.bin", filename: "large-v2.bin"),
+            Model(name: "large-v2-q5_0", info: "(1.1 GiB)", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-large-v2-q5_0.bin", filename: "large-v2-q5_0.bin"),
+            Model(name: "large-v2-q8_0", info: "(1.5 GiB)", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-large-v2-q8_0.bin", filename: "large-v2-q8_0.bin"),
+            Model(name: "large-v3", info: "(F16, 2.9 GiB)", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-large-v3.bin", filename: "large-v3.bin"),
+            Model(name: "large-v3-q5_0", info: "(1.1 GiB)", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-large-v3-q5_0.bin", filename: "large-v3-q5_0.bin"),
+            Model(name: "large-v3-turbo", info: "(F16, 1.5 GiB)", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-large-v3-turbo.bin", filename: "large-v3-turbo.bin"),
+            Model(name: "large-v3-turbo-q5_0", info: "(547 MiB)", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-large-v3-turbo-q5_0.bin", filename: "large-v3-turbo-q5_0.bin"),
+            Model(name: "large-v3-turbo-q8_0", info: "(834 MiB)", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-large-v3-turbo-q8_0.bin", filename: "large-v3-turbo-q8_0.bin"),
+        ]
+
+        static func getDownloadedModels() -> [Model] {
+            // Filter models that have been downloaded
+            return models.filter {
+                FileManager.default.fileExists(atPath: $0.fileURL.path())
+            }
+        }
+
+        func loadModel(model: Model) {
+            Task {
+                dismiss()
+                whisperState.loadModel(path: model.fileURL)
+            }
+        }
+
+        var body: some View {
+            List {
+                Section(header: Text("Models")) {
+                    ForEach(ModelsView.models) { model in
+                        DownloadButton(model: model)
+                            .onLoad(perform: loadModel)
+                    }
+                }
+            }
+            .listStyle(GroupedListStyle())
+            .navigationBarTitle("Models", displayMode: .inline).toolbar {}
+        }
     }
 }
+
+//struct ContentView_Previews: PreviewProvider {
+//    static var previews: some View {
+//        ContentView()
+//    }
+//}

examples/whisper.swiftui/whisper.swiftui.demo/UI/DownloadButton.swift

@@ -0,0 +1,102 @@
+import SwiftUI
+
+struct DownloadButton: View {
+    private var model: Model
+
+    @State private var status: String
+
+    @State private var downloadTask: URLSessionDownloadTask?
+    @State private var progress = 0.0
+    @State private var observation: NSKeyValueObservation?
+
+    private var onLoad: ((_ model: Model) -> Void)?
+
+    init(model: Model) {
+        self.model = model
+        status = model.fileExists() ? "downloaded" : "download"
+    }
+
+    func onLoad(perform action: @escaping (_ model: Model) -> Void) -> DownloadButton {
+        var button = self
+        button.onLoad = action
+        return button
+    }
+
+    private func download() {
+        status = "downloading"
+        print("Downloading model \(model.name) from \(model.url)")
+        guard let url = URL(string: model.url) else { return }
+
+        downloadTask = URLSession.shared.downloadTask(with: url) { temporaryURL, response, error in
+            if let error = error {
+                print("Error: \(error.localizedDescription)")
+                return
+            }
+
+            guard let response = response as? HTTPURLResponse, (200...299).contains(response.statusCode) else {
+                print("Server error!")
+                return
+            }
+
+            do {
+                if let temporaryURL = temporaryURL {
+                    try FileManager.default.copyItem(at: temporaryURL, to: model.fileURL)
+                    print("Writing to \(model.filename) completed")
+                    status = "downloaded"
+                }
+            } catch let err {
+                print("Error: \(err.localizedDescription)")
+            }
+        }
+
+        observation = downloadTask?.progress.observe(\.fractionCompleted) { progress, _ in
+            self.progress = progress.fractionCompleted
+        }
+
+        downloadTask?.resume()
+    }
+
+    var body: some View {
+        VStack {
+            Button(action: {
+                if (status == "download") {
+                    download()
+                } else if (status == "downloading") {
+                    downloadTask?.cancel()
+                    status = "download"
+                } else if (status == "downloaded") {
+                    if !model.fileExists() {
+                        download()
+                    }
+                    onLoad?(model)
+                }
+            }) {
+                let title = "\(model.name) \(model.info)"
+                if (status == "download") {
+                    Text("Download \(title)")
+                } else if (status == "downloading") {
+                    Text("\(title) (Downloading \(Int(progress * 100))%)")
+                } else if (status == "downloaded") {
+                    Text("Load \(title)")
+                } else {
+                    Text("Unknown status")
+                }
+            }.swipeActions {
+                if (status == "downloaded") {
+                    Button("Delete") {
+                        do {
+                            try FileManager.default.removeItem(at: model.fileURL)
+                        } catch {
+                            print("Error deleting file: \(error)")
+                        }
+                        status = "download"
+                    }
+                    .tint(.red)
+                }
+            }
+        }
+        .onDisappear() {
+            downloadTask?.cancel()
+        }
+    }
+}

examples/whisper.swiftui/whisper.swiftui.xcodeproj/project.pbxproj

@@ -17,6 +17,8 @@
 		0AAC5D9F29539CD0003032C3 /* Assets.xcassets in Resources */ = {isa = PBXBuildFile; fileRef = 0AAC5D9E29539CD0003032C3 /* Assets.xcassets */; };
 		0AAC5DCE2953A05C003032C3 /* WhisperState.swift in Sources */ = {isa = PBXBuildFile; fileRef = 0AAC5DCD2953A05C003032C3 /* WhisperState.swift */; };
 		0AAC5DD12953A394003032C3 /* LibWhisper.swift in Sources */ = {isa = PBXBuildFile; fileRef = 0AAC5DD02953A394003032C3 /* LibWhisper.swift */; };
+		7F79E0EE2CE0A78000ACD7BF /* DownloadButton.swift in Sources */ = {isa = PBXBuildFile; fileRef = 7F79E0ED2CE0A78000ACD7BF /* DownloadButton.swift */; };
+		7F79E0F02CE0C6F700ACD7BF /* Model.swift in Sources */ = {isa = PBXBuildFile; fileRef = 7F79E0EF2CE0C6F700ACD7BF /* Model.swift */; };
 		E3F92DC52AFA8E3800A6A9D4 /* whisper in Frameworks */ = {isa = PBXBuildFile; productRef = E3F92DC42AFA8E3800A6A9D4 /* whisper */; };
 /* End PBXBuildFile section */
 
@@ -33,6 +35,8 @@
 		0AAC5DA029539CD0003032C3 /* WhisperCppDemo.entitlements */ = {isa = PBXFileReference; lastKnownFileType = text.plist.entitlements; path = WhisperCppDemo.entitlements; sourceTree = "<group>"; };
 		0AAC5DCD2953A05C003032C3 /* WhisperState.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = WhisperState.swift; sourceTree = "<group>"; };
 		0AAC5DD02953A394003032C3 /* LibWhisper.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = LibWhisper.swift; sourceTree = "<group>"; };
+		7F79E0ED2CE0A78000ACD7BF /* DownloadButton.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = DownloadButton.swift; sourceTree = "<group>"; };
+		7F79E0EF2CE0C6F700ACD7BF /* Model.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = Model.swift; sourceTree = "<group>"; };
 		E3F92DC22AFA8DD800A6A9D4 /* whisper.cpp */ = {isa = PBXFileReference; lastKnownFileType = wrapper; name = whisper.cpp; path = ../..; sourceTree = "<group>"; };
 /* End PBXFileReference section */
 
@@ -52,6 +56,7 @@
 			isa = PBXGroup;
 			children = (
 				0AAC5DCD2953A05C003032C3 /* WhisperState.swift */,
+				7F79E0EF2CE0C6F700ACD7BF /* Model.swift */,
 			);
 			path = Models;
 			sourceTree = "<group>";
@@ -119,6 +124,7 @@
 			isa = PBXGroup;
 			children = (
 				0AAC5D9C29539CCF003032C3 /* ContentView.swift */,
+				7F79E0ED2CE0A78000ACD7BF /* DownloadButton.swift */,
 			);
 			path = UI;
 			sourceTree = "<group>";
@@ -220,7 +226,9 @@
 				0AAC5DCE2953A05C003032C3 /* WhisperState.swift in Sources */,
 				0AAC5DD12953A394003032C3 /* LibWhisper.swift in Sources */,
 				0AA7514C2953B569001EE061 /* RiffWaveUtils.swift in Sources */,
+				7F79E0EE2CE0A78000ACD7BF /* DownloadButton.swift in Sources */,
 				0AA7514E2953D958001EE061 /* Recorder.swift in Sources */,
+				7F79E0F02CE0C6F700ACD7BF /* Model.swift in Sources */,
 			);
 			runOnlyForDeploymentPostprocessing = 0;
 		};
@@ -370,7 +378,9 @@
 				PRODUCT_BUNDLE_IDENTIFIER = com.whispercppdemo.WhisperCppDemo;
 				PRODUCT_NAME = "$(TARGET_NAME)";
 				SDKROOT = auto;
-				SUPPORTED_PLATFORMS = "iphoneos iphonesimulator macosx";
+				SUPPORTED_PLATFORMS = "iphoneos iphonesimulator";
+				SUPPORTS_MACCATALYST = NO;
+				SUPPORTS_MAC_DESIGNED_FOR_IPHONE_IPAD = YES;
 				SWIFT_EMIT_LOC_STRINGS = YES;
 				SWIFT_OPTIMIZATION_LEVEL = "-Onone";
 				SWIFT_VERSION = 5.0;
@@ -415,7 +425,9 @@
 				PRODUCT_BUNDLE_IDENTIFIER = com.whispercppdemo.WhisperCppDemo;
 				PRODUCT_NAME = "$(TARGET_NAME)";
 				SDKROOT = auto;
-				SUPPORTED_PLATFORMS = "iphoneos iphonesimulator macosx";
+				SUPPORTED_PLATFORMS = "iphoneos iphonesimulator";
+				SUPPORTS_MACCATALYST = NO;
+				SUPPORTS_MAC_DESIGNED_FOR_IPHONE_IPAD = YES;
 				SWIFT_EMIT_LOC_STRINGS = YES;
 				SWIFT_VERSION = 5.0;
 				TARGETED_DEVICE_FAMILY = "1,2";

ggml/CMakeLists.txt

@@ -92,6 +92,7 @@ else()
 endif()
 
 option(GGML_CPU_HBM     "ggml: use memkind for CPU HBM" OFF)
+option(GGML_CPU_AARCH64 "ggml: use runtime weight conversion of Q4_0 to Q4_X_X" ON)
 
 option(GGML_AVX         "ggml: enable AVX"              ${INS_ENB})
 option(GGML_AVX2        "ggml: enable AVX2"             ${INS_ENB})
@@ -99,6 +100,9 @@ option(GGML_AVX512      "ggml: enable AVX512"           OFF)
 option(GGML_AVX512_VBMI "ggml: enable AVX512-VBMI"      OFF)
 option(GGML_AVX512_VNNI "ggml: enable AVX512-VNNI"      OFF)
 option(GGML_AVX512_BF16 "ggml: enable AVX512-BF16"      OFF)
+option(GGML_AMX_TILE    "ggml: enable AMX-TILE"         OFF)
+option(GGML_AMX_INT8    "ggml: enable AMX-INT8"         OFF)
+option(GGML_AMX_BF16    "ggml: enable AMX-BF16"         OFF)
 option(GGML_FMA         "ggml: enable FMA"              ${INS_ENB})
 if (NOT MSVC)
     option(GGML_F16C    "ggml: enable F16C"             ${INS_ENB}) # in MSVC F16C is implied with AVX2/AVX512
@@ -113,6 +117,7 @@ endif()
 
 # ggml core
 set(GGML_SCHED_MAX_COPIES  "4" CACHE STRING "ggml: max input copies for pipeline parallelism")
+option(GGML_CPU                             "ggml: enable CPU backend"                        ON)
 
 # 3rd party libs / backends
 option(GGML_ACCELERATE                      "ggml: enable Accelerate framework"               ON)
@@ -123,14 +128,9 @@ option(GGML_LLAMAFILE                       "ggml: use LLAMAFILE"
 
 option(GGML_CUDA                            "ggml: use CUDA"                                  OFF)
 option(GGML_MUSA                            "ggml: use MUSA"                                  OFF)
-option(GGML_CUDA_FORCE_DMMV                 "ggml: use dmmv instead of mmvq CUDA kernels"     OFF)
 option(GGML_CUDA_FORCE_MMQ                  "ggml: use mmq kernels instead of cuBLAS"         OFF)
 option(GGML_CUDA_FORCE_CUBLAS               "ggml: always use cuBLAS instead of mmq kernels"  OFF)
-set   (GGML_CUDA_DMMV_X   "32" CACHE STRING "ggml: x stride for dmmv CUDA kernels")
-set   (GGML_CUDA_MMV_Y     "1" CACHE STRING "ggml: y block size for mmv CUDA kernels")
 option(GGML_CUDA_F16                        "ggml: use 16 bit floats for some calculations"   OFF)
-set   (GGML_CUDA_KQUANTS_ITER "2" CACHE STRING
-                                            "ggml: iters./thread per block for Q2_K/Q6_K")
 set   (GGML_CUDA_PEER_MAX_BATCH_SIZE "128" CACHE STRING
                                             "ggml: max. batch size for using peer access")
 option(GGML_CUDA_NO_PEER_COPY               "ggml: do not use peer to peer copies"            OFF)
@@ -138,7 +138,7 @@ option(GGML_CUDA_NO_VMM                     "ggml: do not try to use CUDA VMM"
 option(GGML_CUDA_FA_ALL_QUANTS              "ggml: compile all quants for FlashAttention"     OFF)
 option(GGML_CUDA_GRAPHS                     "ggml: use CUDA graphs (llama.cpp only)"          ${GGML_CUDA_GRAPHS_DEFAULT})
 
-option(GGML_HIPBLAS                         "ggml: use hipBLAS"                               OFF)
+option(GGML_HIP                             "ggml: use HIP"                                   OFF)
 option(GGML_HIP_UMA                         "ggml: use HIP unified memory architecture"       OFF)
 option(GGML_VULKAN                          "ggml: use Vulkan"                                OFF)
 option(GGML_VULKAN_CHECK_RESULTS            "ggml: run Vulkan op checks"                      OFF)
@@ -150,6 +150,7 @@ option(GGML_VULKAN_VALIDATE                 "ggml: enable Vulkan validation"
 option(GGML_VULKAN_RUN_TESTS                "ggml: run Vulkan tests"                          OFF)
 option(GGML_KOMPUTE                         "ggml: use Kompute"                               OFF)
 option(GGML_METAL                           "ggml: use Metal"                                 ${GGML_METAL_DEFAULT})
+option(GGML_METAL_USE_BF16                  "ggml: use bfloat if available"                   OFF)
 option(GGML_METAL_NDEBUG                    "ggml: disable Metal debugging"                   OFF)
 option(GGML_METAL_SHADER_DEBUG              "ggml: compile Metal with -fno-fast-math"         OFF)
 option(GGML_METAL_EMBED_LIBRARY             "ggml: embed Metal library"                       ${GGML_METAL})
@@ -158,10 +159,13 @@ set   (GGML_METAL_MACOSX_VERSION_MIN "" CACHE STRING
 set   (GGML_METAL_STD "" CACHE STRING       "ggml: metal standard version (-std flag)")
 option(GGML_OPENMP                          "ggml: use OpenMP"                                ON)
 option(GGML_RPC                             "ggml: use RPC"                                   OFF)
+option(GGML_AMX                             "ggml: use AMX"                                   OFF)
 option(GGML_SYCL                            "ggml: use SYCL"                                  OFF)
 option(GGML_SYCL_F16                        "ggml: use 16 bit floats for sycl calculations"   OFF)
 set   (GGML_SYCL_TARGET "INTEL" CACHE STRING
                                             "ggml: sycl target device")
+set   (GGML_SYCL_DEVICE_ARCH "" CACHE STRING
+                                            "ggml: sycl device architecture")
 
 # extra artifacts
 option(GGML_BUILD_TESTS    "ggml: build tests"    ${GGML_STANDALONE})
@@ -214,13 +218,14 @@ include(CMakePackageConfigHelpers)
 # all public headers
 set(GGML_PUBLIC_HEADERS
     include/ggml.h
+    include/ggml-cpu.h
     include/ggml-alloc.h
     include/ggml-backend.h
     include/ggml-blas.h
     include/ggml-cann.h
     include/ggml-cuda.h
-    include/ggml.h
     include/ggml-kompute.h
+    include/ggml-opt.h
     include/ggml-metal.h
     include/ggml-rpc.h
     include/ggml-sycl.h
@@ -230,15 +235,14 @@ set_target_properties(ggml PROPERTIES PUBLIC_HEADER "${GGML_PUBLIC_HEADERS}")
 #if (GGML_METAL)
 #    set_target_properties(ggml PROPERTIES RESOURCE "${CMAKE_CURRENT_SOURCE_DIR}/src/ggml-metal.metal")
 #endif()
-install(TARGETS ggml PUBLIC_HEADER)
-
-if (BUILD_SHARED_LIBS)
-    install(TARGETS ggml LIBRARY)
-endif()
+install(TARGETS ggml LIBRARY PUBLIC_HEADER)
+install(TARGETS ggml-base LIBRARY)
 
+# FIXME: this should be done in the backend cmake files
 if (GGML_METAL)
+    # FIXME: does this need to be installed with GGML_METAL_EMBED_LIBRARY?
     install(
-        FILES src/ggml-metal.metal
+        FILES src/ggml-metal/ggml-metal.metal
         PERMISSIONS
             OWNER_READ
             OWNER_WRITE

ggml/ggml_vk_generate_shaders.py

@@ -1,220 +0,0 @@
-#!/usr/bin/env python
-
-import logging
-import argparse
-import asyncio
-import os
-from tempfile import gettempdir
-
-logger = logging.getLogger("ggml-vk-generate-shaders")
-
-GLSLC = "glslc"
-
-type_names = [
-    "f32",
-    "f16",
-    "q4_0",
-    "q4_1",
-    "q5_0",
-    "q5_1",
-    "q8_0",
-    "q2_k",
-    "q3_k",
-    "q4_k",
-    "q5_k",
-    "q6_k",
-]
-
-ASYNCIO_CONCURRENCY = 64
-
-input_dir = "vulkan-shaders"
-output_dir = gettempdir()
-
-lock = asyncio.Lock()
-shader_fnames = []
-
-
-async def string_to_spv(name, in_fname, defines, fp16=True):
-    name = f"{name}{'_fp32' if not fp16 else ''}"
-    out_fname = os.path.join(output_dir, f"{name}.spv")
-
-    in_path = os.path.join(input_dir, in_fname)
-
-    cmd = [GLSLC, "-fshader-stage=compute", "--target-env=vulkan1.2", "-O", in_path, "-o", out_fname]
-
-    cmd.extend([f"-D{key}={value}" for key, value in defines.items()])
-
-    proc = await asyncio.create_subprocess_exec(*cmd, stdout=asyncio.subprocess.PIPE, stderr=asyncio.subprocess.PIPE)
-
-    stdout, stderr = await proc.communicate()
-
-    stdout = stdout.decode()
-    error = stderr.decode()
-
-    if proc.returncode:
-        cmd = " ".join(cmd)
-        logger.error(f"cannot compile {name}\n\n{cmd}\n\n{error}")
-        return
-
-    async with lock:
-        shader_fnames.append((name, out_fname))
-
-
-def matmul_shaders(tasks, fp16, matmul_id):
-    if fp16:
-        load_vec = "8"
-        aligned_b_type_f32 = "mat2x4"
-        aligned_b_type_f16 = "f16mat2x4"
-    else:
-        load_vec = "4"
-        aligned_b_type_f32 = "vec4"
-        aligned_b_type_f16 = "f16vec4"
-
-    base_dict = {"FLOAT_TYPE": "float" if not fp16 else "float16_t"}
-    shader_name = "matmul"
-
-    if matmul_id:
-        base_dict["MUL_MAT_ID"] = "1"
-        shader_name = "matmul_id"
-
-    if fp16:
-        base_dict["FLOAT16"] = "1"
-
-    # Shaders with f16 B_TYPE
-    tasks.append(string_to_spv(f"{shader_name}_f32_f16", "mul_mm.comp", base_dict | {"DATA_A_F32": "1", "B_TYPE": "float16_t", "D_TYPE": "float"}, fp16))
-    tasks.append(string_to_spv(f"{shader_name}_f32_f16_aligned", "mul_mm.comp", base_dict | {"DATA_A_F32": "1", "LOAD_VEC_A": load_vec, "LOAD_VEC_B": load_vec, "B_TYPE": aligned_b_type_f16, "D_TYPE": "float"}, fp16))
-
-    tasks.append(string_to_spv(f"{shader_name}_f16", "mul_mm.comp", base_dict | {"DATA_A_F16": "1", "B_TYPE": "float16_t", "D_TYPE": "float"}, fp16))
-    tasks.append(string_to_spv(f"{shader_name}_f16_aligned", "mul_mm.comp", base_dict | {"DATA_A_F16": "1", "LOAD_VEC_A": load_vec, "LOAD_VEC_B": load_vec, "B_TYPE": aligned_b_type_f16, "D_TYPE": "float"}, fp16))
-
-    for tname in type_names:
-        data_a_key = f"DATA_A_{tname.upper()}"
-        load_vec_a = load_vec if tname in ("f32", "f16") else "2"
-        tasks.append(string_to_spv(f"{shader_name}_{tname}_f32", "mul_mm.comp", base_dict | {data_a_key: "1", "B_TYPE": "float", "D_TYPE": "float"}, fp16))
-        tasks.append(string_to_spv(f"{shader_name}_{tname}_f32_aligned", "mul_mm.comp", base_dict | {data_a_key: "2", "LOAD_VEC_A": load_vec_a, "LOAD_VEC_B": load_vec, "B_TYPE": aligned_b_type_f32, "D_TYPE": "float"}, fp16))
-
-
-async def main():
-    logger.info("ggml_vulkan: Generating and compiling shaders to SPIR-V")
-
-    tasks = []
-
-    for fp16 in (False, True):
-        # MUL_MAT
-        matmul_shaders(tasks, fp16, False)
-        # MUL_MAT_ID
-        matmul_shaders(tasks, fp16, True)
-
-    for tname in type_names:
-        base_dict = {"FLOAT_TYPE": "float"}
-
-        # mul mat vec
-        data_a_key = f"DATA_A_{tname.upper()}"
-        shader = f"mul_mat_vec_{tname}.comp" if tname.endswith("_k") else "mul_mat_vec.comp"
-
-        tasks.append(string_to_spv(f"mul_mat_vec_{tname}_f32_f32", shader, base_dict | {data_a_key: "1", "B_TYPE": "float", "D_TYPE": "float"}))
-        tasks.append(string_to_spv(f"mul_mat_vec_{tname}_f16_f32", shader, base_dict | {data_a_key: "1", "B_TYPE": "float16_t", "D_TYPE": "float"}))
-
-        tasks.append(string_to_spv(f"mul_mat_vec_id_{tname}_f32", shader, base_dict | {"MUL_MAT_ID": "1", data_a_key: "1", "B_TYPE": "float", "D_TYPE": "float"}))
-
-        # Dequant shaders
-        if tname != "f16":
-            tasks.append(string_to_spv(f"dequant_{tname}", f"dequant_{tname}.comp", base_dict | {data_a_key: "1", "D_TYPE": "float16_t"}))
-
-        # get_rows
-        if not tname.endswith("_k"):
-            shader = "get_rows.comp" if tname in ("f32", "f16") else "get_rows_quant.comp"
-
-            if tname == "f16":
-                tasks.append(string_to_spv(f"get_rows_{tname}", shader, {data_a_key: "1", "B_TYPE": "int", "D_TYPE": "float16_t", "OPTIMIZATION_ERROR_WORKAROUND": "1"}))
-            else:
-                tasks.append(string_to_spv(f"get_rows_{tname}", shader, {data_a_key: "1", "B_TYPE": "int", "D_TYPE": "float16_t"}))
-            tasks.append(string_to_spv(f"get_rows_{tname}_f32", shader, {data_a_key: "1", "B_TYPE": "int", "D_TYPE": "float"}))
-
-    tasks.append(string_to_spv("mul_mat_vec_p021_f16_f32", "mul_mat_vec_p021.comp", {"A_TYPE": "float16_t", "B_TYPE": "float", "D_TYPE": "float"}))
-    tasks.append(string_to_spv("mul_mat_vec_nc_f16_f32", "mul_mat_vec_nc.comp", {"A_TYPE": "float16_t", "B_TYPE": "float", "D_TYPE": "float"}))
-
-    # Norms
-    tasks.append(string_to_spv("norm_f32", "norm.comp", base_dict | {"A_TYPE": "float", "D_TYPE": "float"}))
-    tasks.append(string_to_spv("rms_norm_f32", "rms_norm.comp", base_dict | {"A_TYPE": "float", "D_TYPE": "float"}))
-
-    tasks.append(string_to_spv("cpy_f32_f32", "copy.comp", {"A_TYPE": "float", "D_TYPE": "float"}))
-    tasks.append(string_to_spv("cpy_f32_f16", "copy.comp", {"A_TYPE": "float", "D_TYPE": "float16_t"}))
-    tasks.append(string_to_spv("cpy_f16_f16", "copy.comp", {"A_TYPE": "float16_t", "D_TYPE": "float16_t", "OPTIMIZATION_ERROR_WORKAROUND": "1"}))
-
-    tasks.append(string_to_spv("add_f32", "add.comp", {"A_TYPE": "float", "B_TYPE": "float", "D_TYPE": "float", "FLOAT_TYPE": "float"}))
-
-    tasks.append(string_to_spv("split_k_reduce", "mul_mat_split_k_reduce.comp", {}))
-
-    tasks.append(string_to_spv("mul_f32", "mul.comp", {"A_TYPE": "float", "B_TYPE": "float", "D_TYPE": "float", "FLOAT_TYPE": "float"}))
-
-    tasks.append(string_to_spv("div_f32", "div.comp", {"A_TYPE": "float", "B_TYPE": "float", "D_TYPE": "float", "FLOAT_TYPE": "float"}))
-
-    tasks.append(string_to_spv("scale_f32", "scale.comp", {"A_TYPE": "float", "D_TYPE": "float", "FLOAT_TYPE": "float"}))
-
-    tasks.append(string_to_spv("sqr_f32", "square.comp", {"A_TYPE": "float", "D_TYPE": "float", "FLOAT_TYPE": "float"}))
-
-    tasks.append(string_to_spv("clamp_f32", "clamp.comp", {"A_TYPE": "float", "D_TYPE": "float", "FLOAT_TYPE": "float"}))
-
-    tasks.append(string_to_spv("gelu_f32", "gelu.comp", {"A_TYPE": "float", "D_TYPE": "float"}))
-    tasks.append(string_to_spv("silu_f32", "silu.comp", {"A_TYPE": "float", "D_TYPE": "float"}))
-    tasks.append(string_to_spv("relu_f32", "relu.comp", {"A_TYPE": "float", "D_TYPE": "float"}))
-
-    tasks.append(string_to_spv("diag_mask_inf_f32", "diag_mask_inf.comp", {"A_TYPE": "float", "D_TYPE": "float"}))
-
-    tasks.append(string_to_spv("soft_max_f32", "soft_max.comp", base_dict | {"A_TYPE": "float", "B_TYPE": "float", "D_TYPE": "float"}))
-    tasks.append(string_to_spv("soft_max_f32_f16", "soft_max.comp", base_dict | {"A_TYPE": "float", "B_TYPE": "float16_t", "D_TYPE": "float"}))
-
-    tasks.append(string_to_spv("rope_norm_f32", "rope_norm.comp", {"A_TYPE": "float", "D_TYPE": "float"}))
-    tasks.append(string_to_spv("rope_norm_f16", "rope_norm.comp", {"A_TYPE": "float16_t", "D_TYPE": "float16_t"}))
-
-    tasks.append(string_to_spv("rope_neox_f32", "rope_neox.comp", {"A_TYPE": "float", "D_TYPE": "float"}))
-    tasks.append(string_to_spv("rope_neox_f16", "rope_neox.comp", {"A_TYPE": "float16_t", "D_TYPE": "float16_t"}))
-
-    tasks.append(string_to_spv("argsort_f32", "argsort.comp", {"A_TYPE": "float"}))
-
-    tasks.append(string_to_spv("sum_rows_f32", "sum_rows.comp", base_dict | {"A_TYPE": "float", "D_TYPE": "float"}))
-
-    # Helper to decorate tasks with semaphore acquisition.
-    async def withSemaphore(sem, task):
-        async with sem:
-            return await task
-
-    # Run tasks concurrently guarded by a concurrency limit.
-    sem = asyncio.Semaphore(ASYNCIO_CONCURRENCY)
-    await asyncio.gather(*(withSemaphore(sem, task) for task in tasks))
-
-    with open("ggml-vulkan-shaders.hpp", "w") as f:
-        f.write("#include <cstdint>\n\n")
-        for name, path in sorted(shader_fnames):
-
-            with open(path, "rb") as spv:
-                counter = 0
-                newline_counter = 0
-                f.write(f"unsigned char {name}_data[] = {{\n")
-                for val in spv.read():
-                    f.write(f"0x{val:02x},")
-                    newline_counter += 1
-                    counter += 1
-                    if newline_counter >= 12:
-                        newline_counter = 0
-                        f.write("\n")
-            f.write("\n};\n")
-            f.write(f"const uint64_t {name}_len = {counter};\n\n")
-            os.remove(path)
-
-
-if __name__ == "__main__":
-    parser = argparse.ArgumentParser(description="GGML Vulkan Shader Generator")
-
-    parser.add_argument("--glslc", help="Path to glslc")
-    parser.add_argument("--verbose", action="store_true", help="increase output verbosity")
-
-    args = parser.parse_args()
-
-    logging.basicConfig(level=logging.DEBUG if args.verbose else logging.INFO)
-
-    if args.glslc:
-        GLSLC = args.glslc
-
-    asyncio.run(main())

ggml/include/ggml-amx.h

@@ -0,0 +1,25 @@
+#pragma once
+
+#include "ggml.h"
+#include "ggml-backend.h"
+
+
+#ifdef  __cplusplus
+extern "C" {
+#endif
+
+// buffer_type API
+GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_amx_buffer_type(void);
+
+GGML_BACKEND_API bool ggml_backend_is_amx(ggml_backend_t backend);
+
+// backend API
+GGML_BACKEND_API ggml_backend_t ggml_backend_amx_init(void);
+
+GGML_BACKEND_API void ggml_backend_amx_set_n_threads(ggml_backend_t backend_amx, int n_threads);
+
+GGML_BACKEND_API ggml_backend_reg_t ggml_backend_amx_reg(void);
+
+#ifdef  __cplusplus
+}
+#endif

ggml/include/ggml-backend.h

@@ -3,6 +3,20 @@
 #include "ggml.h"
 #include "ggml-alloc.h"
 
+#ifdef GGML_BACKEND_SHARED
+#    if defined(_WIN32) && !defined(__MINGW32__)
+#        ifdef GGML_BACKEND_BUILD
+#            define GGML_BACKEND_API __declspec(dllexport) extern
+#        else
+#            define GGML_BACKEND_API __declspec(dllimport) extern
+#        endif
+#    else
+#        define GGML_BACKEND_API __attribute__ ((visibility ("default"))) extern
+#    endif
+#else
+#    define GGML_BACKEND_API extern
+#endif
+
 #ifdef  __cplusplus
 extern "C" {
 #endif
@@ -72,7 +86,7 @@ extern "C" {
     GGML_API void ggml_backend_tensor_set_async(ggml_backend_t backend,       struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
     GGML_API void ggml_backend_tensor_get_async(ggml_backend_t backend, const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
 
-    // "offset" refers to the offset of the tensor data for setting/getting data
+    // "offset" refers to the offset in tensor->data for setting/getting data
     GGML_API void ggml_backend_tensor_set(      struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
     GGML_API void ggml_backend_tensor_get(const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
     GGML_API void ggml_backend_tensor_memset(   struct ggml_tensor * tensor,     uint8_t value, size_t offset, size_t size);
@@ -114,11 +128,12 @@ extern "C" {
     //
 
     enum ggml_backend_dev_type {
+        // CPU device using system memory
         GGML_BACKEND_DEVICE_TYPE_CPU,
+        // GPU device using dedicated memory
         GGML_BACKEND_DEVICE_TYPE_GPU,
-        // devices with full capabilities (excludes backends such as BLAS that only support matrix multiplication)
-        GGML_BACKEND_DEVICE_TYPE_CPU_FULL,
-        GGML_BACKEND_DEVICE_TYPE_GPU_FULL
+        // accelerator devices intended to be used together with the CPU backend (e.g. BLAS or AMX)
+        GGML_BACKEND_DEVICE_TYPE_ACCEL
     };
 
     // functionality supported by the device
@@ -127,6 +142,8 @@ extern "C" {
         bool async;
         // pinned host buffer
         bool host_buffer;
+        // creating buffers from host ptr
+        bool buffer_from_host_ptr;
         // event synchronization
         bool events;
     };
@@ -165,9 +182,14 @@ extern "C" {
     GGML_API ggml_backend_dev_t ggml_backend_reg_dev_get(ggml_backend_reg_t reg, size_t index);
     GGML_API void *             ggml_backend_reg_get_proc_address(ggml_backend_reg_t reg, const char * name);
 
+    // Common functions that may be obtained using ggml_backend_reg_get_proc_address
 
-    // Functions that may be obtained using ggml_backend_reg_get_proc_address
-    typedef ggml_backend_buffer_type_t (*ggml_backend_split_buffer_type_t)(const float *);
+    // Split buffer type for tensor parallelism
+    typedef ggml_backend_buffer_type_t   (*ggml_backend_split_buffer_type_t)(int main_device, const float * tensor_split);
+    // Set the number of threads for the backend
+    typedef void                         (*ggml_backend_set_n_threads_t)(ggml_backend_t backend, int n_threads);
+    // Get additional buffer types provided by the device (returns a NULL-terminated array)
+    typedef ggml_backend_buffer_type_t * (*ggml_backend_dev_get_extra_bufts_t)(ggml_backend_dev_t device);
 
     //
     // Backend registry
@@ -189,7 +211,7 @@ extern "C" {
     GGML_API ggml_backend_t ggml_backend_init_by_name(const char * name, const char * params);
     // = ggml_backend_dev_init(ggml_backend_dev_by_type(type), params)
     GGML_API ggml_backend_t ggml_backend_init_by_type(enum ggml_backend_dev_type type, const char * params);
-    // = ggml_backend_dev_init(ggml_backend_dev_by_type(GPU_FULL) OR ggml_backend_dev_by_type(CPU_FULL), NULL)
+    // = ggml_backend_dev_init(ggml_backend_dev_by_type(GPU) OR ggml_backend_dev_by_type(CPU), NULL)
     GGML_API ggml_backend_t ggml_backend_init_best(void);
 
     //
@@ -220,14 +242,20 @@ extern "C" {
         ggml_backend_sched_reserve(sched, reserve_graph);
 
         // compute
-        graph = build_graph(sched);
-        ggml_backend_sched_graph_compute(sched, graph);
+        graph = build_graph(sched); // the graph and its tensors are single-use in terms of allocation, multi-use in terms of computation
+        for (int i = 0; i < 10; ++i) {
+            ggml_backend_sched_graph_compute(sched, graph); // on the first iteration the graph is allocated automatically
+        }
 
         // if there are graph inputs:
-        ggml_backend_sched_reset(sched);
-        ggml_backend_sched_alloc_graph(sched, graph);
-        ggml_backend_tensor_set(input_tensor, ...);
-        ggml_backend_sched_graph_compute(sched, graph);
+        graph = build_graph(sched); // get a new graph that is not allocated (the metadata for the old graph is freed once ggml_free is called)
+        ggml_backend_sched_reset(sched); // clear the allocation of the previous graph
+        ggml_backend_sched_alloc_graph(sched, graph); // explicitly allocate the new graph but do not execute it
+        ggml_backend_tensor_set(input_tensor, ...); // copy data to the newly allocated graph tensors
+        ggml_backend_sched_graph_compute(sched, graph); // execute the graph
+
+        // as an alternative to the above it is also possible to assign the inputs to a dedicated context and
+        // allocate them statically via ggml_backend_alloc_ctx_tensors
     }
     */
 
@@ -242,7 +270,7 @@ extern "C" {
     //
     typedef bool (*ggml_backend_sched_eval_callback)(struct ggml_tensor * t, bool ask, void * user_data);
 
-    // Initialize a backend scheduler
+    // Initialize a backend scheduler, backends with low index are given priority over backends with high index
     GGML_API ggml_backend_sched_t ggml_backend_sched_new(ggml_backend_t * backends, ggml_backend_buffer_type_t * bufts, int n_backends, size_t graph_size, bool parallel);
     GGML_API void                 ggml_backend_sched_free(ggml_backend_sched_t sched);
 
@@ -267,7 +295,9 @@ extern "C" {
     GGML_API enum ggml_status     ggml_backend_sched_graph_compute_async(ggml_backend_sched_t sched, struct ggml_cgraph * graph);
     GGML_API void                 ggml_backend_sched_synchronize(ggml_backend_sched_t sched);
 
-    // Reset all assignments and allocators - must be called before changing the node backends
+    // Reset all assignments and allocators - must be called before changing the node backends or allocating a new graph.
+    // This in effect deallocates all tensors that were previously allocated and leaves them with dangling pointers.
+    // The correct way to use this API is to discard the deallocated tensors and create new ones.
     GGML_API void                 ggml_backend_sched_reset(ggml_backend_sched_t sched);
 
     // Set a callback to be called for each resulting node during graph compute
@@ -297,27 +327,10 @@ extern "C" {
     GGML_API void ggml_backend_tensor_alloc(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, void * addr);
     GGML_API void ggml_backend_view_init(struct ggml_tensor * tensor);
 
-    //
-    // CPU backend
-    //
-
-    GGML_API ggml_backend_t ggml_backend_cpu_init(void);
-
-    GGML_API bool ggml_backend_is_cpu                (ggml_backend_t backend);
-    GGML_API void ggml_backend_cpu_set_n_threads     (ggml_backend_t backend_cpu, int n_threads);
-    GGML_API void ggml_backend_cpu_set_threadpool    (ggml_backend_t backend_cpu, ggml_threadpool_t threadpool);
-    GGML_API void ggml_backend_cpu_set_abort_callback(ggml_backend_t backend_cpu, ggml_abort_callback abort_callback, void * abort_callback_data);
-
-    // Create a backend buffer from an existing pointer
+    // CPU buffer types are always available
     GGML_API ggml_backend_buffer_t      ggml_backend_cpu_buffer_from_ptr(void * ptr, size_t size);
     GGML_API ggml_backend_buffer_type_t ggml_backend_cpu_buffer_type(void);
 
-    GGML_API ggml_backend_reg_t ggml_backend_cpu_reg(void);
-
-#ifdef GGML_USE_CPU_HBM
-    GGML_API ggml_backend_buffer_type_t ggml_backend_cpu_hbm_buffer_type(void);
-#endif
-
 #ifdef  __cplusplus
 }
 #endif

ggml/include/ggml-blas.h

@@ -9,13 +9,15 @@ extern "C" {
 #endif
 
 // backend API
-GGML_API ggml_backend_t ggml_backend_blas_init(void);
+GGML_BACKEND_API ggml_backend_t ggml_backend_blas_init(void);
 
-GGML_API bool ggml_backend_is_blas(ggml_backend_t backend);
+GGML_BACKEND_API bool ggml_backend_is_blas(ggml_backend_t backend);
 
 // number of threads used for conversion to float
 // for openblas and blis, this will also set the number of threads used for blas operations
-GGML_API void ggml_backend_blas_set_n_threads(ggml_backend_t backend_blas, int n_threads);
+GGML_BACKEND_API void ggml_backend_blas_set_n_threads(ggml_backend_t backend_blas, int n_threads);
+
+GGML_BACKEND_API ggml_backend_reg_t ggml_backend_blas_reg(void);
 
 
 #ifdef  __cplusplus

ggml/include/ggml-cann.h

@@ -34,6 +34,8 @@ extern "C" {
  */
 #define GGML_CANN_MAX_DEVICES 16
 
+GGML_BACKEND_API ggml_backend_reg_t ggml_backend_cann_reg(void);
+
 /**
  * @brief Initializes the CANN backend for a specified device.
  *
@@ -44,7 +46,7 @@ extern "C" {
  * @param device The index of the device to initialize.
  * @return A pointer to the initialized backend instance, or nullptr on failure.
  */
-GGML_API ggml_backend_t ggml_backend_cann_init(int32_t device);
+GGML_BACKEND_API ggml_backend_t ggml_backend_cann_init(int32_t device);
 
 /**
  * @brief Checks if a given backend is a CANN backend.
@@ -55,7 +57,7 @@ GGML_API ggml_backend_t ggml_backend_cann_init(int32_t device);
  * @param backend The backend instance to check.
  * @return True if the backend is a CANN backend, false otherwise.
  */
-GGML_API bool ggml_backend_is_cann(ggml_backend_t backend);
+GGML_BACKEND_API bool ggml_backend_is_cann(ggml_backend_t backend);
 
 /**
  * @brief Retrieves the CANN buffer type for a specified device.
@@ -67,7 +69,7 @@ GGML_API bool ggml_backend_is_cann(ggml_backend_t backend);
  * @return A pointer to the buffer type interface for the specified device, or
  * nullptr if the device index is out of range.
  */
-GGML_API ggml_backend_buffer_type_t
+GGML_BACKEND_API ggml_backend_buffer_type_t
 ggml_backend_cann_buffer_type(int32_t device);
 
 /**
@@ -78,14 +80,14 @@ ggml_backend_cann_buffer_type(int32_t device);
  *
  * @return The number of CANN devices available.
  */
-GGML_API int32_t ggml_backend_cann_get_device_count(void);
+GGML_BACKEND_API int32_t ggml_backend_cann_get_device_count(void);
 
 /**
  * @brief pinned host buffer for use with the CPU backend for faster copies between CPU and NPU.
  *
  * @return A pointer to the host buffer type interface.
  */
-GGML_API ggml_backend_buffer_type_t ggml_backend_cann_host_buffer_type(void);
+GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_cann_host_buffer_type(void);
 
 /**
  * @brief Retrieves the description of a specific CANN device.
@@ -97,7 +99,7 @@ GGML_API ggml_backend_buffer_type_t ggml_backend_cann_host_buffer_type(void);
  * @param description Pointer to a buffer where the description will be written.
  * @param description_size Size of the description buffer.
  */
-GGML_API void ggml_backend_cann_get_device_description(
+GGML_BACKEND_API void ggml_backend_cann_get_device_description(
     int32_t device, char* description, size_t description_size);
 
 /**
@@ -112,7 +114,7 @@ GGML_API void ggml_backend_cann_get_device_description(
  * @param total Pointer to a variable where the total memory size will be
  * stored.
  */
-GGML_API void ggml_backend_cann_get_device_memory(int32_t device,
+GGML_BACKEND_API void ggml_backend_cann_get_device_memory(int32_t device,
                                                   size_t* free,
                                                   size_t* total);
 

ggml/include/ggml-cpp.h

@@ -0,0 +1,38 @@
+#pragma once
+
+#ifndef __cplusplus
+#error "This header is for C++ only"
+#endif
+
+#include "ggml.h"
+#include "ggml-alloc.h"
+#include "ggml-backend.h"
+#include <memory>
+
+// Smart pointers for ggml types
+
+// ggml
+
+struct ggml_context_deleter { void operator()(ggml_context * ctx) { ggml_free(ctx); } };
+struct gguf_context_deleter { void operator()(gguf_context * ctx) { gguf_free(ctx); } };
+
+typedef std::unique_ptr<ggml_context, ggml_context_deleter> ggml_context_ptr;
+typedef std::unique_ptr<gguf_context, gguf_context_deleter> gguf_context_ptr;
+
+// ggml-alloc
+
+struct ggml_gallocr_deleter { void operator()(ggml_gallocr_t galloc) { ggml_gallocr_free(galloc); } };
+
+typedef std::unique_ptr<ggml_gallocr_t, ggml_gallocr_deleter> ggml_gallocr_ptr;
+
+// ggml-backend
+
+struct ggml_backend_deleter        { void operator()(ggml_backend_t backend)       { ggml_backend_free(backend); } };
+struct ggml_backend_buffer_deleter { void operator()(ggml_backend_buffer_t buffer) { ggml_backend_buffer_free(buffer); } };
+struct ggml_backend_event_deleter  { void operator()(ggml_backend_event_t event)   { ggml_backend_event_free(event); } };
+struct ggml_backend_sched_deleter  { void operator()(ggml_backend_sched_t sched)   { ggml_backend_sched_free(sched); } };
+
+typedef std::unique_ptr<ggml_backend,        ggml_backend_deleter>        ggml_backend_ptr;
+typedef std::unique_ptr<ggml_backend_buffer, ggml_backend_buffer_deleter> ggml_backend_buffer_ptr;
+typedef std::unique_ptr<ggml_backend_event,  ggml_backend_event_deleter>  ggml_backend_event_ptr;
+typedef std::unique_ptr<ggml_backend_sched,  ggml_backend_sched_deleter>  ggml_backend_sched_ptr;

ggml/include/ggml-cpu.h

@@ -0,0 +1,177 @@
+#pragma once
+
+#include "ggml.h"
+#include "ggml-backend.h"
+
+#ifdef  __cplusplus
+extern "C" {
+#endif
+
+    // Scheduling priorities
+    enum ggml_sched_priority {
+        GGML_SCHED_PRIO_NORMAL,
+        GGML_SCHED_PRIO_MEDIUM,
+        GGML_SCHED_PRIO_HIGH,
+        GGML_SCHED_PRIO_REALTIME
+    };
+
+    // Threadpool params
+    // Use ggml_threadpool_params_default() or ggml_threadpool_params_init() to populate the defaults
+    struct ggml_threadpool_params {
+        bool                cpumask[GGML_MAX_N_THREADS]; // mask of cpu cores (all-zeros means use default affinity settings)
+        int                 n_threads;                   // number of threads
+        enum ggml_sched_priority prio;                   // thread priority
+        uint32_t            poll;                        // polling level (0 - no polling, 100 - aggressive polling)
+        bool                strict_cpu;                  // strict cpu placement
+        bool                paused;                      // start in paused state
+    };
+
+    struct ggml_threadpool;     // forward declaration, see ggml.c
+
+    typedef struct ggml_threadpool * ggml_threadpool_t;
+
+    // the compute plan that needs to be prepared for ggml_graph_compute()
+    // since https://github.com/ggerganov/ggml/issues/287
+    struct ggml_cplan {
+        size_t    work_size; // size of work buffer, calculated by `ggml_graph_plan()`
+        uint8_t * work_data; // work buffer, to be allocated by caller before calling to `ggml_graph_compute()`
+
+        int n_threads;
+        struct ggml_threadpool * threadpool;
+
+        // abort ggml_graph_compute when true
+        ggml_abort_callback abort_callback;
+        void *              abort_callback_data;
+    };
+
+    // numa strategies
+    enum ggml_numa_strategy {
+        GGML_NUMA_STRATEGY_DISABLED   = 0,
+        GGML_NUMA_STRATEGY_DISTRIBUTE = 1,
+        GGML_NUMA_STRATEGY_ISOLATE    = 2,
+        GGML_NUMA_STRATEGY_NUMACTL    = 3,
+        GGML_NUMA_STRATEGY_MIRROR     = 4,
+        GGML_NUMA_STRATEGY_COUNT
+    };
+
+    GGML_BACKEND_API void    ggml_numa_init(enum ggml_numa_strategy numa); // call once for better performance on NUMA systems
+    GGML_BACKEND_API bool    ggml_is_numa(void); // true if init detected that system has >1 NUMA node
+
+    GGML_BACKEND_API struct ggml_tensor * ggml_new_i32(struct ggml_context * ctx, int32_t value);
+    GGML_BACKEND_API struct ggml_tensor * ggml_new_f32(struct ggml_context * ctx, float value);
+
+    GGML_BACKEND_API struct ggml_tensor * ggml_set_i32 (struct ggml_tensor * tensor, int32_t value);
+    GGML_BACKEND_API struct ggml_tensor * ggml_set_f32 (struct ggml_tensor * tensor, float value);
+
+    GGML_BACKEND_API int32_t ggml_get_i32_1d(const struct ggml_tensor * tensor, int i);
+    GGML_BACKEND_API void    ggml_set_i32_1d(const struct ggml_tensor * tensor, int i, int32_t value);
+
+    GGML_BACKEND_API int32_t ggml_get_i32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3);
+    GGML_BACKEND_API void    ggml_set_i32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3, int32_t value);
+
+    GGML_BACKEND_API float   ggml_get_f32_1d(const struct ggml_tensor * tensor, int i);
+    GGML_BACKEND_API void    ggml_set_f32_1d(const struct ggml_tensor * tensor, int i, float value);
+
+    GGML_BACKEND_API float   ggml_get_f32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3);
+    GGML_BACKEND_API void    ggml_set_f32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3, float value);
+
+    GGML_BACKEND_API struct ggml_threadpool_params ggml_threadpool_params_default(int n_threads);
+    GGML_BACKEND_API void                          ggml_threadpool_params_init   (struct ggml_threadpool_params * p, int n_threads);
+    GGML_BACKEND_API bool                          ggml_threadpool_params_match  (const struct ggml_threadpool_params * p0, const struct ggml_threadpool_params * p1);
+    GGML_BACKEND_API struct ggml_threadpool *      ggml_threadpool_new          (struct ggml_threadpool_params  * params);
+    GGML_BACKEND_API void                          ggml_threadpool_free         (struct ggml_threadpool * threadpool);
+    GGML_BACKEND_API int                           ggml_threadpool_get_n_threads(struct ggml_threadpool * threadpool);
+    GGML_BACKEND_API void                          ggml_threadpool_pause        (struct ggml_threadpool * threadpool);
+    GGML_BACKEND_API void                          ggml_threadpool_resume       (struct ggml_threadpool * threadpool);
+
+    // ggml_graph_plan() has to be called before ggml_graph_compute()
+    // when plan.work_size > 0, caller must allocate memory for plan.work_data
+    GGML_BACKEND_API struct ggml_cplan ggml_graph_plan(
+                  const struct ggml_cgraph * cgraph,
+                                       int   n_threads, /* = GGML_DEFAULT_N_THREADS */
+                    struct ggml_threadpool * threadpool /* = NULL */ );
+    GGML_BACKEND_API enum ggml_status  ggml_graph_compute(struct ggml_cgraph * cgraph, struct ggml_cplan * cplan);
+
+    // same as ggml_graph_compute() but the work data is allocated as a part of the context
+    // note: the drawback of this API is that you must have ensured that the context has enough memory for the work data
+    GGML_BACKEND_API enum ggml_status  ggml_graph_compute_with_ctx(struct ggml_context * ctx, struct ggml_cgraph * cgraph, int n_threads);
+
+    //
+    // system info
+    //
+
+    // x86
+    GGML_BACKEND_API int ggml_cpu_has_sse3       (void);
+    GGML_BACKEND_API int ggml_cpu_has_ssse3      (void);
+    GGML_BACKEND_API int ggml_cpu_has_avx        (void);
+    GGML_BACKEND_API int ggml_cpu_has_avx2       (void);
+    GGML_BACKEND_API int ggml_cpu_has_f16c       (void);
+    GGML_BACKEND_API int ggml_cpu_has_fma        (void);
+    GGML_BACKEND_API int ggml_cpu_has_avx_vnni   (void);
+    GGML_BACKEND_API int ggml_cpu_has_avx512     (void);
+    GGML_BACKEND_API int ggml_cpu_has_avx512_vbmi(void);
+    GGML_BACKEND_API int ggml_cpu_has_avx512_vnni(void);
+    GGML_BACKEND_API int ggml_cpu_has_avx512_bf16(void);
+    GGML_BACKEND_API int ggml_cpu_has_amx_int8   (void);
+    // ARM
+    GGML_BACKEND_API int ggml_cpu_has_neon       (void);
+    GGML_BACKEND_API int ggml_cpu_has_arm_fma    (void);
+    GGML_BACKEND_API int ggml_cpu_has_fp16_va    (void);
+    GGML_BACKEND_API int ggml_cpu_has_matmul_int8(void);
+    GGML_BACKEND_API int ggml_cpu_has_sve        (void);
+    GGML_BACKEND_API int ggml_cpu_get_sve_cnt    (void);  // sve vector length in bytes
+    // other
+    GGML_BACKEND_API int ggml_cpu_has_riscv_v    (void);
+    GGML_BACKEND_API int ggml_cpu_has_vsx        (void);
+    GGML_BACKEND_API int ggml_cpu_has_wasm_simd  (void);
+    GGML_BACKEND_API int ggml_cpu_has_llamafile  (void);
+
+    // Internal types and functions exposed for tests and benchmarks
+
+    typedef void (*ggml_from_float_to_mat_t)
+                                     (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t nr, int64_t k, int64_t bs);
+    typedef void (*ggml_vec_dot_t)  (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT x, size_t bx,
+                                       const void * GGML_RESTRICT y, size_t by, int nrc);
+    typedef void (*ggml_gemv_t)     (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT x,
+                                       const void * GGML_RESTRICT y, int nr, int nc);
+    typedef void (*ggml_gemm_t)     (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT x,
+                                       const void * GGML_RESTRICT y, int nr, int nc);
+
+    struct ggml_type_traits_cpu {
+        ggml_from_float_t        from_float;
+        ggml_from_float_to_mat_t from_float_to_mat;
+        ggml_vec_dot_t           vec_dot;
+        enum ggml_type           vec_dot_type;
+        int64_t                  nrows; // number of rows to process simultaneously
+        int64_t                  ncols; // number of columns to process simultaneously
+        ggml_gemv_t              gemv;
+        ggml_gemm_t              gemm;
+    };
+
+    GGML_BACKEND_API const struct ggml_type_traits_cpu * ggml_get_type_traits_cpu(enum ggml_type type);
+
+    GGML_BACKEND_API void ggml_cpu_init(void);
+
+    //
+    // CPU backend
+    //
+
+    GGML_BACKEND_API ggml_backend_t ggml_backend_cpu_init(void);
+
+    GGML_BACKEND_API bool ggml_backend_is_cpu                (ggml_backend_t backend);
+    GGML_BACKEND_API void ggml_backend_cpu_set_n_threads     (ggml_backend_t backend_cpu, int n_threads);
+    GGML_BACKEND_API void ggml_backend_cpu_set_threadpool    (ggml_backend_t backend_cpu, ggml_threadpool_t threadpool);
+    GGML_BACKEND_API void ggml_backend_cpu_set_abort_callback(ggml_backend_t backend_cpu, ggml_abort_callback abort_callback, void * abort_callback_data);
+
+    GGML_BACKEND_API ggml_backend_reg_t ggml_backend_cpu_reg(void);
+
+#ifdef GGML_USE_CPU_HBM
+    GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_cpu_hbm_buffer_type(void);
+#endif
+
+    GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_cpu_aarch64_buffer_type(void);
+    GGML_BACKEND_API bool ggml_backend_cpu_buft_is_aarch64(ggml_backend_buffer_type_t buft);
+
+#ifdef __cplusplus
+}
+#endif

ggml/include/ggml-cuda.h

@@ -7,7 +7,7 @@
 extern "C" {
 #endif
 
-#ifdef GGML_USE_HIPBLAS
+#ifdef GGML_USE_HIP
 #define GGML_CUDA_NAME "ROCm"
 #define GGML_CUBLAS_NAME "hipBLAS"
 #elif defined(GGML_USE_MUSA)
@@ -20,27 +20,27 @@ extern "C" {
 #define GGML_CUDA_MAX_DEVICES       16
 
 // backend API
-GGML_API ggml_backend_t ggml_backend_cuda_init(int device);
+GGML_BACKEND_API ggml_backend_t ggml_backend_cuda_init(int device);
 
-GGML_API bool ggml_backend_is_cuda(ggml_backend_t backend);
+GGML_BACKEND_API bool ggml_backend_is_cuda(ggml_backend_t backend);
 
 // device buffer
-GGML_API ggml_backend_buffer_type_t ggml_backend_cuda_buffer_type(int device);
+GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_cuda_buffer_type(int device);
 
 // split tensor buffer that splits matrices by rows across multiple devices
-GGML_API ggml_backend_buffer_type_t ggml_backend_cuda_split_buffer_type(const float * tensor_split);
+GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_cuda_split_buffer_type(int main_device, const float * tensor_split);
 
 // pinned host buffer for use with the CPU backend for faster copies between CPU and GPU
-GGML_API ggml_backend_buffer_type_t ggml_backend_cuda_host_buffer_type(void);
+GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_cuda_host_buffer_type(void);
 
-GGML_API int  ggml_backend_cuda_get_device_count(void);
-GGML_API void ggml_backend_cuda_get_device_description(int device, char * description, size_t description_size);
-GGML_API void ggml_backend_cuda_get_device_memory(int device, size_t * free, size_t * total);
+GGML_BACKEND_API int  ggml_backend_cuda_get_device_count(void);
+GGML_BACKEND_API void ggml_backend_cuda_get_device_description(int device, char * description, size_t description_size);
+GGML_BACKEND_API void ggml_backend_cuda_get_device_memory(int device, size_t * free, size_t * total);
 
-GGML_API bool ggml_backend_cuda_register_host_buffer(void * buffer, size_t size);
-GGML_API void ggml_backend_cuda_unregister_host_buffer(void * buffer);
+GGML_BACKEND_API bool ggml_backend_cuda_register_host_buffer(void * buffer, size_t size);
+GGML_BACKEND_API void ggml_backend_cuda_unregister_host_buffer(void * buffer);
 
-GGML_API ggml_backend_reg_t ggml_backend_cuda_reg(void);
+GGML_BACKEND_API ggml_backend_reg_t ggml_backend_cuda_reg(void);
 
 #ifdef  __cplusplus
 }

ggml/include/ggml-kompute.h

@@ -11,6 +11,8 @@
 extern "C" {
 #endif
 
+#define GGML_KOMPUTE_MAX_DEVICES 16
+
 struct ggml_vk_device {
     int index;
     int type; // same as VkPhysicalDeviceType
@@ -35,11 +37,13 @@ struct ggml_vk_device ggml_vk_current_device(void);
 // forward declaration
 typedef struct ggml_backend * ggml_backend_t;
 
-GGML_API ggml_backend_t ggml_backend_kompute_init(int device);
+GGML_BACKEND_API ggml_backend_t ggml_backend_kompute_init(int device);
 
-GGML_API bool ggml_backend_is_kompute(ggml_backend_t backend);
+GGML_BACKEND_API bool ggml_backend_is_kompute(ggml_backend_t backend);
 
-GGML_API ggml_backend_buffer_type_t ggml_backend_kompute_buffer_type(int device);
+GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_kompute_buffer_type(int device);
+
+GGML_BACKEND_API ggml_backend_reg_t ggml_backend_kompute_reg(void);
 
 #ifdef __cplusplus
 }

ggml/include/ggml-metal.h

@@ -39,23 +39,27 @@ extern "C" {
 // user-code should use only these functions
 //
 
-GGML_API ggml_backend_t ggml_backend_metal_init(void);
+GGML_BACKEND_API ggml_backend_t ggml_backend_metal_init(void);
 
-GGML_API bool ggml_backend_is_metal(ggml_backend_t backend);
+GGML_BACKEND_API bool ggml_backend_is_metal(ggml_backend_t backend);
 
-GGML_API ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void * data, size_t size, size_t max_size);
+GGML_DEPRECATED(
+        GGML_BACKEND_API ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void * data, size_t size, size_t max_size),
+        "obsoleted by the new device interface - https://github.com/ggerganov/llama.cpp/pull/9713");
 
-GGML_API void ggml_backend_metal_set_abort_callback(ggml_backend_t backend, ggml_abort_callback abort_callback, void * user_data);
+GGML_BACKEND_API void ggml_backend_metal_set_abort_callback(ggml_backend_t backend, ggml_abort_callback abort_callback, void * user_data);
 
-GGML_API ggml_backend_buffer_type_t ggml_backend_metal_buffer_type(void);
+GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_metal_buffer_type(void);
 
 // helper to check if the device supports a specific family
 // ideally, the user code should be doing these checks
 // ref: https://developer.apple.com/metal/Metal-Feature-Set-Tables.pdf
-GGML_API bool ggml_backend_metal_supports_family(ggml_backend_t backend, int family);
+GGML_BACKEND_API bool ggml_backend_metal_supports_family(ggml_backend_t backend, int family);
 
 // capture all command buffers committed the next time `ggml_backend_graph_compute` is called
-GGML_API void ggml_backend_metal_capture_next_compute(ggml_backend_t backend);
+GGML_BACKEND_API void ggml_backend_metal_capture_next_compute(ggml_backend_t backend);
+
+GGML_BACKEND_API ggml_backend_reg_t ggml_backend_metal_reg(void);
 
 #ifdef __cplusplus
 }

ggml/include/ggml-opt.h

@@ -0,0 +1,216 @@
+// This file contains functionality for training models using GGML.
+// It is not strictly needed vs. just vanilla GGML but it provides a more high-level interface for common needs such as datasets.
+// At the bottom of this file especially there are relatively high-level functions that are suitable use or adaptation in user code.
+//
+// Module maintainer: Johannes Gäßler (@JohannesGaessler, johannesg@5d6.de)
+
+#pragma once
+
+#include "ggml.h"
+#include "ggml-backend.h"
+
+#include <stdint.h>
+
+#ifdef  __cplusplus
+extern "C" {
+#endif
+
+    struct ggml_opt_dataset;
+    struct ggml_opt_context;
+    struct ggml_opt_result;
+
+    typedef struct ggml_opt_dataset * ggml_opt_dataset_t;
+    typedef struct ggml_opt_context * ggml_opt_context_t;
+    typedef struct ggml_opt_result  * ggml_opt_result_t;
+
+    // ====== Loss ======
+
+    // built-in loss types, i.e. the built-in quantities minimized by the optimizer
+    // custom loss types can be defined via mean or sum which simply reduce the outputs for all datapoints to a single value
+    enum ggml_opt_loss_type {
+        GGML_OPT_LOSS_TYPE_MEAN,
+        GGML_OPT_LOSS_TYPE_SUM,
+        GGML_OPT_LOSS_TYPE_CROSS_ENTROPY,
+        GGML_OPT_LOSS_TYPE_MEAN_SQUARED_ERROR,
+    };
+
+    // ====== Dataset ======
+
+    GGML_API ggml_opt_dataset_t ggml_opt_dataset_init(
+            int64_t ne_datapoint, // number of elements per datapoint
+            int64_t ne_label,     // number of elements per label
+            int64_t ndata,        // total number of datapoints/labels
+            int64_t ndata_shard); // number of datapoints/labels per shard (unit at which the dataset is shuffled/copied)
+    GGML_API void ggml_opt_dataset_free(ggml_opt_dataset_t dataset);
+
+    // get underlying tensors that store the data
+    GGML_API struct ggml_tensor * ggml_opt_dataset_data  (ggml_opt_dataset_t dataset); // shape = [ne_datapoint, ndata]
+    GGML_API struct ggml_tensor * ggml_opt_dataset_labels(ggml_opt_dataset_t dataset); // shape = [nd_label,     ndata]
+
+    // shuffle idata first datapoints from dataset with RNG from opt_ctx, shuffle all datapoints if idata is negative
+    GGML_API void ggml_opt_dataset_shuffle(ggml_opt_context_t opt_ctx, ggml_opt_dataset_t dataset, int64_t idata);
+
+    // get batch at position ibatch from dataset and copy the data to data_batch and labels_batch
+    GGML_API void ggml_opt_dataset_get_batch(
+            ggml_opt_dataset_t   dataset,
+            struct ggml_tensor * data_batch,   // shape = [ne_datapoint, ndata_batch]
+            struct ggml_tensor * labels_batch, // shape = [ne_label,     ndata_batch]
+            int64_t              ibatch);
+
+    // ====== Model / Context ======
+
+    enum ggml_opt_build_type {
+        GGML_OPT_BUILD_TYPE_FORWARD,
+        GGML_OPT_BUILD_TYPE_GRAD,
+        GGML_OPT_BUILD_TYPE_OPT,
+    };
+
+    // parameters that control which optimizer is used and how said optimizer tries to find the minimal loss
+    struct ggml_opt_optimizer_params {
+        // AdamW optimizer parameters
+        struct {
+            float alpha; // learning rate
+            float beta1;
+            float beta2;
+            float eps;   // epsilon for numerical stability
+            float wd;    // weight decay for AdamW, use 0.0f to disable
+        } adamw;
+    };
+
+    // callback to calculate optimizer parameters prior to a backward pass
+    // userdata can be used to pass arbitrary data
+    typedef struct ggml_opt_optimizer_params (*ggml_opt_get_optimizer_params)(void * userdata);
+
+    // returns the default optimizer params (constant)
+    // userdata is not used
+    GGML_API struct ggml_opt_optimizer_params ggml_opt_get_default_optimizer_params(void * userdata);
+
+    // parameters for initializing a new optimization context
+    struct ggml_opt_params {
+        ggml_backend_sched_t backend_sched; // defines which backends are used to construct the compute graphs
+
+        struct ggml_context * ctx_compute; // created in user code, holds non-static tensors
+
+        // the forward graph is defined by inputs and outputs
+        // those tensors and all tensors inbetween are not intended to be reusable between multiple optimization contexts
+        struct ggml_tensor * inputs;
+        struct ggml_tensor * outputs;
+
+        enum ggml_opt_loss_type  loss_type;
+        enum ggml_opt_build_type build_type;
+
+        int32_t opt_period; // after how many gradient accumulation steps an optimizer step should be done
+
+        ggml_opt_get_optimizer_params get_opt_pars; // callback for calculating optimizer parameters
+        void * get_opt_pars_ud;                     // userdata for calculating optimizer parameters
+    };
+
+    // get parameters for an optimization context with defaults set where possible
+    // parameters for which no sensible defaults exist are supplied as arguments to this function
+    GGML_API ggml_opt_params ggml_opt_default_params(
+            ggml_backend_sched_t      backend_sched,
+            struct ggml_context     * ctx_compute,
+            struct ggml_tensor      * inputs,
+            struct ggml_tensor      * outputs,
+            enum ggml_opt_loss_type   loss_type);
+
+    GGML_API ggml_opt_context_t ggml_opt_init(struct ggml_opt_params params);
+    GGML_API void ggml_opt_free(ggml_opt_context_t opt_ctx);
+
+    // set gradients to zero, initilize loss, and optionally reset the optimizer
+    GGML_API void ggml_opt_reset(ggml_opt_context_t opt_ctx, bool optimizer);
+
+    // get underlying tensors that store data
+    GGML_API struct ggml_tensor * ggml_opt_inputs(  ggml_opt_context_t opt_ctx); // forward graph input tensor
+    GGML_API struct ggml_tensor * ggml_opt_outputs( ggml_opt_context_t opt_ctx); // forward graph output tensor
+    GGML_API struct ggml_tensor * ggml_opt_labels(  ggml_opt_context_t opt_ctx); // labels to compare outputs against
+    GGML_API struct ggml_tensor * ggml_opt_loss(    ggml_opt_context_t opt_ctx); // scalar tensor that contains the loss
+    GGML_API struct ggml_tensor * ggml_opt_pred(    ggml_opt_context_t opt_ctx); // predictions made by outputs
+    GGML_API struct ggml_tensor * ggml_opt_ncorrect(ggml_opt_context_t opt_ctx); // number of matching predictions between outputs and labels
+
+    GGML_API struct ggml_tensor * ggml_opt_grad_acc(ggml_opt_context_t opt_ctx, struct ggml_tensor * node);
+
+    // ====== Optimization Result ======
+
+    GGML_API ggml_opt_result_t ggml_opt_result_init();
+    GGML_API void ggml_opt_result_free(ggml_opt_result_t result);
+    GGML_API void ggml_opt_result_reset(ggml_opt_result_t result);
+
+    // get data from result, uncertainties are optional and can be ignored by passing NULL
+    GGML_API void ggml_opt_result_ndata(   ggml_opt_result_t result, int64_t * ndata);                  // writes 1 value, number of datapoints
+    GGML_API void ggml_opt_result_loss(    ggml_opt_result_t result, double  * loss,     double * unc); // writes 1 value
+    GGML_API void ggml_opt_result_pred(    ggml_opt_result_t result, int32_t * pred);                   // writes ndata values
+    GGML_API void ggml_opt_result_accuracy(ggml_opt_result_t result, double  * accuracy, double * unc); // writes 1 value
+
+    // ====== Computation ======
+
+    // do forward pass, increment result if not NULL
+    GGML_API void ggml_opt_forward(ggml_opt_context_t opt_ctx, ggml_opt_result_t result);
+
+    // do forward pass, increment result if not NULL, do backward pass
+    GGML_API void ggml_opt_forward_backward(ggml_opt_context_t opt_ctx, ggml_opt_result_t result);
+
+    // ############################################################################
+    // ## The high-level functions start here. They do not depend on any private ##
+    // ## functions or structs and can be copied to and adapted for user code.   ##
+    // ############################################################################
+
+    // ====== Intended Usage ======
+    //
+    // 1. Select the appropriate loss for your problem.
+    // 2. Create a dataset and set the data for the "data" tensor. Also set the "labels" tensor if your loss needs them.
+    //    Setting the shard size to 1 will be fine, it's the granularity with which data is shuffled/loaded (bigger values are faster).
+    // 3. Create a GGML graph for your model with no_alloc == true. Use two separate contexts for the tensors.
+    //    The first context should contain the model parameters and inputs and be allocated statically in user code.
+    //    The second context should contain all other tensors and will be (re)allocated automatically.
+    //    Due to this automated allocation the data of the second context is not defined when accessed in user code.
+    //    Note that the second dimension of the inputs/outputs are interpreted as the number of datapoints in those tensors.
+    // 4. Call ggml_opt_fit. If you need more control you can use ggml_opt_epoch instead.
+
+    // signature for a callback while evaluating opt_ctx on dataset, called after an evaluation
+    typedef void (*ggml_opt_epoch_callback)(
+            bool               train,       // true after training evaluation, false after validation evaluation
+            ggml_opt_context_t opt_ctx,
+            ggml_opt_dataset_t dataset,
+            ggml_opt_result_t  result,      // result associated with the dataset subsection
+            int64_t            ibatch,      // number of batches that have been evaluated so far
+            int64_t            ibatch_max,  // total number of batches in this dataset subsection
+            int64_t            t_start_us); // time at which the evaluation on the dataset subsection was started
+
+    // do training on front of dataset, do evaluation only on back of dataset
+    GGML_API void ggml_opt_epoch(
+            ggml_opt_context_t      opt_ctx,
+            ggml_opt_dataset_t      dataset,
+            ggml_opt_result_t       result_train,   // result to increment during training, ignored if NULL
+            ggml_opt_result_t       result_eval,    // result to increment during evaluation, ignored if NULL
+            int64_t                 idata_split,    // data index at which to split training and evaluation
+            ggml_opt_epoch_callback callback_train,
+            ggml_opt_epoch_callback callback_eval);
+
+    // callback that prints a progress bar on stderr
+    GGML_API void ggml_opt_epoch_callback_progress_bar(
+            bool               train,
+            ggml_opt_context_t opt_ctx,
+            ggml_opt_dataset_t dataset,
+            ggml_opt_result_t  result,
+            int64_t            ibatch,
+            int64_t            ibatch_max,
+            int64_t            t_start_us);
+
+    // fit model defined by inputs and outputs to dataset
+    GGML_API void ggml_opt_fit(
+            ggml_backend_sched_t            backend_sched,  // backend scheduler for constructing the compute graphs
+            ggml_context                  * ctx_compute,    // context with temporarily allocated tensors to calculate the outputs
+            ggml_tensor                   * inputs,         // input tensor with shape [ne_datapoint, ndata_batch]
+            ggml_tensor                   * outputs,        // output tensor, must have shape [ne_label, ndata_batch] if labels are used
+            ggml_opt_dataset_t              dataset,        // dataset with data and optionally also labels
+            enum ggml_opt_loss_type         loss_type,      // loss to minimize
+            ggml_opt_get_optimizer_params   get_opt_pars,   // callback to get optimizer params, userdata is pointer to epoch (of type int64_t)
+            int64_t                         nepoch,         // how many times the dataset should be iterated over
+            int64_t                         nbatch_logical, // datapoints optimizer step, must be a multiple of ndata_batch in inputs/outputs
+            float                           val_split,      // fraction of the dataset to use for validation, must be in [0.0f, 1.0f)
+            bool                            silent);        // whether or not info prints to stderr should be suppressed
+
+#ifdef  __cplusplus
+}
+#endif

ggml/include/ggml-rpc.h

@@ -10,14 +10,18 @@ extern "C" {
 #define GGML_RPC_MAX_SERVERS       16
 
 // backend API
-GGML_API ggml_backend_t ggml_backend_rpc_init(const char * endpoint);
-GGML_API bool ggml_backend_is_rpc(ggml_backend_t backend);
+GGML_BACKEND_API ggml_backend_t ggml_backend_rpc_init(const char * endpoint);
+GGML_BACKEND_API bool ggml_backend_is_rpc(ggml_backend_t backend);
 
-GGML_API ggml_backend_buffer_type_t ggml_backend_rpc_buffer_type(const char * endpoint);
+GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_rpc_buffer_type(const char * endpoint);
 
-GGML_API void ggml_backend_rpc_get_device_memory(const char * endpoint, size_t * free, size_t * total);
+GGML_BACKEND_API void ggml_backend_rpc_get_device_memory(const char * endpoint, size_t * free, size_t * total);
 
-GGML_API void start_rpc_server(ggml_backend_t backend, const char * endpoint, size_t free_mem, size_t total_mem);
+GGML_BACKEND_API void ggml_backend_rpc_start_server(ggml_backend_t backend, const char * endpoint, size_t free_mem, size_t total_mem);
+
+GGML_BACKEND_API ggml_backend_reg_t ggml_backend_rpc_reg(void);
+
+GGML_BACKEND_API ggml_backend_dev_t ggml_backend_rpc_add_device(const char * endpoint);
 
 #ifdef  __cplusplus
 }

ggml/include/ggml-sycl.h

@@ -17,26 +17,33 @@ extern "C" {
 #endif
 
 // backend API
-GGML_API ggml_backend_t ggml_backend_sycl_init(int device);
+GGML_BACKEND_API ggml_backend_t ggml_backend_sycl_init(int device);
+
+GGML_BACKEND_API bool ggml_backend_is_sycl(ggml_backend_t backend);
 
 // devide buffer
-GGML_API ggml_backend_buffer_type_t ggml_backend_sycl_buffer_type(int device);
+GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_sycl_buffer_type(int device);
 
 // split tensor buffer that splits matrices by rows across multiple devices
-GGML_API ggml_backend_buffer_type_t ggml_backend_sycl_split_buffer_type(const float * tensor_split);
+GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_sycl_split_buffer_type(const float * tensor_split);
 
 // pinned host buffer for use with the CPU backend for faster copies between CPU and GPU
-GGML_API ggml_backend_buffer_type_t ggml_backend_sycl_host_buffer_type(void);
+GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_sycl_host_buffer_type(void);
 
-GGML_API void ggml_backend_sycl_print_sycl_devices(void);
-GGML_API void ggml_sycl_get_gpu_list(int *id_list, int max_len);
-GGML_API void ggml_sycl_get_device_description(int device, char *description, size_t description_size);
-GGML_API int  ggml_backend_sycl_get_device_count();
-GGML_API void ggml_backend_sycl_get_device_memory(int device, size_t *free, size_t *total);
+GGML_BACKEND_API void ggml_backend_sycl_print_sycl_devices(void);
+GGML_BACKEND_API void ggml_backend_sycl_get_gpu_list(int *id_list, int max_len);
+GGML_BACKEND_API void ggml_backend_sycl_get_device_description(int device,
+                                                       char *description,
+                                                       size_t description_size);
+GGML_BACKEND_API int  ggml_backend_sycl_get_device_count();
+GGML_BACKEND_API void ggml_backend_sycl_get_device_memory(int device, size_t *free, size_t *total);
 
 // SYCL doesn't support registering host memory, keep here for reference
-// GGML_API bool ggml_backend_sycl_register_host_buffer(void * buffer, size_t size);
-// GGML_API void ggml_backend_sycl_unregister_host_buffer(void * buffer);
+// GGML_BACKEND_API bool ggml_backend_sycl_register_host_buffer(void * buffer, size_t size);
+// GGML_BACKEND_API void ggml_backend_sycl_unregister_host_buffer(void * buffer);
+
+GGML_BACKEND_API ggml_backend_reg_t ggml_backend_sycl_reg(void);
+
 #ifdef  __cplusplus
 }
 #endif

ggml/include/ggml-vulkan.h

@@ -10,19 +10,21 @@ extern "C" {
 #define GGML_VK_NAME "Vulkan"
 #define GGML_VK_MAX_DEVICES 16
 
-GGML_API void ggml_vk_instance_init(void);
+GGML_BACKEND_API void ggml_vk_instance_init(void);
 
 // backend API
-GGML_API ggml_backend_t ggml_backend_vk_init(size_t dev_num);
+GGML_BACKEND_API ggml_backend_t ggml_backend_vk_init(size_t dev_num);
 
-GGML_API bool ggml_backend_is_vk(ggml_backend_t backend);
-GGML_API int  ggml_backend_vk_get_device_count(void);
-GGML_API void ggml_backend_vk_get_device_description(int device, char * description, size_t description_size);
-GGML_API void ggml_backend_vk_get_device_memory(int device, size_t * free, size_t * total);
+GGML_BACKEND_API bool ggml_backend_is_vk(ggml_backend_t backend);
+GGML_BACKEND_API int  ggml_backend_vk_get_device_count(void);
+GGML_BACKEND_API void ggml_backend_vk_get_device_description(int device, char * description, size_t description_size);
+GGML_BACKEND_API void ggml_backend_vk_get_device_memory(int device, size_t * free, size_t * total);
 
-GGML_API ggml_backend_buffer_type_t ggml_backend_vk_buffer_type(size_t dev_num);
+GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_vk_buffer_type(size_t dev_num);
 // pinned host buffer for use with the CPU backend for faster copies between CPU and GPU
-GGML_API ggml_backend_buffer_type_t ggml_backend_vk_host_buffer_type(void);
+GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_vk_host_buffer_type(void);
+
+GGML_BACKEND_API ggml_backend_reg_t ggml_backend_vk_reg(void);
 
 #ifdef  __cplusplus
 }

ggml/include/ggml.h

@@ -176,15 +176,15 @@
 #ifdef GGML_SHARED
 #    if defined(_WIN32) && !defined(__MINGW32__)
 #        ifdef GGML_BUILD
-#            define GGML_API __declspec(dllexport)
+#            define GGML_API __declspec(dllexport) extern
 #        else
-#            define GGML_API __declspec(dllimport)
+#            define GGML_API __declspec(dllimport) extern
 #        endif
 #    else
-#        define GGML_API __attribute__ ((visibility ("default")))
+#        define GGML_API __attribute__ ((visibility ("default"))) extern
 #    endif
 #else
-#    define GGML_API
+#    define GGML_API extern
 #endif
 
 // TODO: support for clang
@@ -217,7 +217,6 @@
 
 #define GGML_MAX_DIMS           4
 #define GGML_MAX_PARAMS         2048
-#define GGML_MAX_CONTEXTS       64
 #define GGML_MAX_SRC            10
 #define GGML_MAX_N_THREADS      512
 #define GGML_MAX_OP_PARAMS      64
@@ -510,7 +509,7 @@ extern "C" {
         GGML_OP_WIN_UNPART,
         GGML_OP_GET_REL_POS,
         GGML_OP_ADD_REL_POS,
-        GGML_OP_RWKV_WKV,
+        GGML_OP_RWKV_WKV6,
 
         GGML_OP_UNARY,
 
@@ -559,10 +558,10 @@ extern "C" {
 
     enum ggml_log_level {
         GGML_LOG_LEVEL_NONE  = 0,
-        GGML_LOG_LEVEL_INFO  = 1,
-        GGML_LOG_LEVEL_WARN  = 2,
-        GGML_LOG_LEVEL_ERROR = 3,
-        GGML_LOG_LEVEL_DEBUG = 4,
+        GGML_LOG_LEVEL_DEBUG = 1,
+        GGML_LOG_LEVEL_INFO  = 2,
+        GGML_LOG_LEVEL_WARN  = 3,
+        GGML_LOG_LEVEL_ERROR = 4,
         GGML_LOG_LEVEL_CONT  = 5, // continue previous log
     };
 
@@ -574,6 +573,13 @@ extern "C" {
         GGML_TENSOR_FLAG_LOSS   =  8, // ...defines loss for numerical optimization (multiple loss tensors add up)
     };
 
+    struct ggml_init_params {
+        // memory pool
+        size_t mem_size;   // bytes
+        void * mem_buffer; // if NULL, memory will be allocated internally
+        bool   no_alloc;   // don't allocate memory for the tensor data
+    };
+
     // n-dimensional tensor
     struct ggml_tensor {
         enum ggml_type type;
@@ -596,7 +602,6 @@ extern "C" {
 
         int32_t flags;
 
-        struct ggml_tensor * grad;
         struct ggml_tensor * src[GGML_MAX_SRC];
 
         // source tensor and offset for views
@@ -609,7 +614,7 @@ extern "C" {
 
         void * extra; // extra things e.g. for ggml-cuda.cu
 
-        // char padding[4];
+        char padding[8];
     };
 
     static const size_t GGML_TENSOR_SIZE = sizeof(struct ggml_tensor);
@@ -619,66 +624,6 @@ extern "C" {
     // If it returns true, the computation is aborted
     typedef bool (*ggml_abort_callback)(void * data);
 
-    // Scheduling priorities
-    enum ggml_sched_priority {
-        GGML_SCHED_PRIO_NORMAL,
-        GGML_SCHED_PRIO_MEDIUM,
-        GGML_SCHED_PRIO_HIGH,
-        GGML_SCHED_PRIO_REALTIME
-    };
-
-    // Threadpool params
-    // Use ggml_threadpool_params_default() or ggml_threadpool_params_init() to populate the defaults
-    struct ggml_threadpool_params {
-        bool                cpumask[GGML_MAX_N_THREADS]; // mask of cpu cores (all-zeros means use default affinity settings)
-        int                 n_threads;                   // number of threads
-        enum ggml_sched_priority prio;                   // thread priority
-        uint32_t            poll;                        // polling level (0 - no polling, 100 - aggressive polling)
-        bool                strict_cpu;                  // strict cpu placement
-        bool                paused;                      // start in paused state
-    };
-
-    struct ggml_threadpool;     // forward declaration, see ggml.c
-
-    typedef struct ggml_threadpool * ggml_threadpool_t;
-
-    // the compute plan that needs to be prepared for ggml_graph_compute()
-    // since https://github.com/ggerganov/ggml/issues/287
-    struct ggml_cplan {
-        size_t    work_size; // size of work buffer, calculated by `ggml_graph_plan()`
-        uint8_t * work_data; // work buffer, to be allocated by caller before calling to `ggml_graph_compute()`
-
-        int n_threads;
-        struct ggml_threadpool * threadpool;
-
-        // abort ggml_graph_compute when true
-        ggml_abort_callback abort_callback;
-        void *              abort_callback_data;
-    };
-
-    // scratch buffer
-    struct ggml_scratch {
-        size_t offs;
-        size_t size;
-        void * data;
-    };
-
-    struct ggml_init_params {
-        // memory pool
-        size_t mem_size;   // bytes
-        void * mem_buffer; // if NULL, memory will be allocated internally
-        bool   no_alloc;   // don't allocate memory for the tensor data
-    };
-
-    // numa strategies
-    enum ggml_numa_strategy {
-        GGML_NUMA_STRATEGY_DISABLED   = 0,
-        GGML_NUMA_STRATEGY_DISTRIBUTE = 1,
-        GGML_NUMA_STRATEGY_ISOLATE    = 2,
-        GGML_NUMA_STRATEGY_NUMACTL    = 3,
-        GGML_NUMA_STRATEGY_MIRROR     = 4,
-        GGML_NUMA_STRATEGY_COUNT
-    };
 
     //
     // GUID
@@ -701,9 +646,6 @@ extern "C" {
     // accepts a UTF-8 path, even on Windows
     GGML_API FILE *  ggml_fopen(const char * fname, const char * mode);
 
-    GGML_API void    ggml_numa_init(enum ggml_numa_strategy numa); // call once for better performance on NUMA systems
-    GGML_API bool    ggml_is_numa(void); // true if init detected that system has >1 NUMA node
-
     GGML_API void    ggml_print_object (const struct ggml_object * obj);
     GGML_API void    ggml_print_objects(const struct ggml_context * ctx);
 
@@ -760,12 +702,12 @@ extern "C" {
 
     // main
 
-    GGML_API struct ggml_context * ggml_init(struct ggml_init_params params);
-    GGML_API void                  ggml_free(struct ggml_context * ctx);
+    GGML_API struct ggml_context * ggml_init (struct ggml_init_params params);
+    GGML_API void                  ggml_reset(struct ggml_context * ctx);
+    GGML_API void                  ggml_free (struct ggml_context * ctx);
 
     GGML_API size_t  ggml_used_mem(const struct ggml_context * ctx);
 
-    GGML_API size_t  ggml_set_scratch (struct ggml_context * ctx, struct ggml_scratch scratch);
     GGML_API bool    ggml_get_no_alloc(struct ggml_context * ctx);
     GGML_API void    ggml_set_no_alloc(struct ggml_context * ctx, bool no_alloc);
 
@@ -805,8 +747,7 @@ extern "C" {
             int64_t ne2,
             int64_t ne3);
 
-    GGML_API struct ggml_tensor * ggml_new_i32(struct ggml_context * ctx, int32_t value);
-    GGML_API struct ggml_tensor * ggml_new_f32(struct ggml_context * ctx, float value);
+    GGML_API void * ggml_new_buffer(struct ggml_context * ctx, size_t nbytes);
 
     GGML_API struct ggml_tensor * ggml_dup_tensor (struct ggml_context * ctx, const struct ggml_tensor * src);
     GGML_API struct ggml_tensor * ggml_view_tensor(struct ggml_context * ctx, struct ggml_tensor * src);
@@ -816,35 +757,25 @@ extern "C" {
     GGML_API struct ggml_tensor * ggml_get_next_tensor (const struct ggml_context * ctx, struct ggml_tensor * tensor);
     GGML_API struct ggml_tensor * ggml_get_tensor(struct ggml_context * ctx, const char * name);
 
-    GGML_API struct ggml_tensor * ggml_set_zero(struct ggml_tensor * tensor);
-    GGML_API struct ggml_tensor * ggml_set_i32 (struct ggml_tensor * tensor, int32_t value);
-    GGML_API struct ggml_tensor * ggml_set_f32 (struct ggml_tensor * tensor, float value);
-
     // Converts a flat index into coordinates
-    GGML_API void    ggml_unravel_index(const struct ggml_tensor * tensor, int64_t i, int64_t * i0, int64_t * i1, int64_t * i2, int64_t * i3);
+    GGML_API void ggml_unravel_index(const struct ggml_tensor * tensor, int64_t i, int64_t * i0, int64_t * i1, int64_t * i2, int64_t * i3);
 
-    GGML_API int32_t ggml_get_i32_1d(const struct ggml_tensor * tensor, int i);
-    GGML_API void    ggml_set_i32_1d(const struct ggml_tensor * tensor, int i, int32_t value);
-
-    GGML_API int32_t ggml_get_i32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3);
-    GGML_API void    ggml_set_i32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3, int32_t value);
-
-    GGML_API float   ggml_get_f32_1d(const struct ggml_tensor * tensor, int i);
-    GGML_API void    ggml_set_f32_1d(const struct ggml_tensor * tensor, int i, float value);
-
-    GGML_API float   ggml_get_f32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3);
-    GGML_API void    ggml_set_f32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3, float value);
+    GGML_API enum ggml_unary_op ggml_get_unary_op(const struct ggml_tensor * tensor);
 
     GGML_API void *  ggml_get_data    (const struct ggml_tensor * tensor);
     GGML_API float * ggml_get_data_f32(const struct ggml_tensor * tensor);
 
-    GGML_API enum ggml_unary_op ggml_get_unary_op(const struct ggml_tensor * tensor);
-
     GGML_API const char *         ggml_get_name   (const struct ggml_tensor * tensor);
     GGML_API struct ggml_tensor * ggml_set_name   (      struct ggml_tensor * tensor, const char * name);
     GGML_ATTRIBUTE_FORMAT(2, 3)
     GGML_API struct ggml_tensor * ggml_format_name(      struct ggml_tensor * tensor, const char * fmt, ...);
 
+    // Tensor flags
+    GGML_API void ggml_set_input(struct ggml_tensor * tensor);
+    GGML_API void ggml_set_output(struct ggml_tensor * tensor);
+    GGML_API void ggml_set_param(struct ggml_context * ctx, struct ggml_tensor * tensor);
+    GGML_API void ggml_set_loss(struct ggml_tensor * tensor);
+
     //
     // operations on tensors with backpropagation
     //
@@ -1558,7 +1489,7 @@ extern "C" {
         "use ggml_rope_ext_inplace instead");
 
     // compute correction dims for YaRN RoPE scaling
-    void ggml_rope_yarn_corr_dims(
+    GGML_API void ggml_rope_yarn_corr_dims(
         int n_dims, int n_ctx_orig, float freq_base, float beta_fast, float beta_slow, float dims[2]);
 
     // rotary position embedding backward, i.e compute dx from dy
@@ -1814,6 +1745,9 @@ extern "C" {
             struct ggml_tensor * a,
             enum ggml_prec       prec);
 
+    GGML_API enum ggml_prec ggml_flash_attn_ext_get_prec(
+            const struct ggml_tensor * a);
+
     // TODO: needs to be adapted to ggml_flash_attn_ext
     GGML_API struct ggml_tensor * ggml_flash_attn_back(
            struct ggml_context * ctx,
@@ -1887,7 +1821,7 @@ extern "C" {
             struct ggml_tensor  * pw,
             struct ggml_tensor  * ph);
 
-    GGML_API struct ggml_tensor * ggml_rwkv_wkv(
+    GGML_API struct ggml_tensor * ggml_rwkv_wkv6(
             struct ggml_context * ctx,
             struct ggml_tensor  * k,
             struct ggml_tensor  * v,
@@ -2050,31 +1984,20 @@ extern "C" {
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
             struct ggml_tensor  * grad,
-            float                 alpha,
-            float                 beta1,
-            float                 beta2,
-            float                 eps,
-            float                 wd); // weight decay
+            struct ggml_tensor  * m,
+            struct ggml_tensor  * v,
+            struct ggml_tensor  * adamw_params); // parameters such a the learning rate
 
     //
     // automatic differentiation
     //
 
-    GGML_API void ggml_set_param(struct ggml_context * ctx, struct ggml_tensor * tensor);
-    GGML_API void ggml_set_loss(struct ggml_tensor * tensor);
-
-    GGML_API void ggml_build_forward_expand (struct ggml_cgraph * cgraph, struct ggml_tensor * tensor);
-    GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool accumulate);
-
-    GGML_API void ggml_build_opt_adamw(
-            struct ggml_context * ctx,
-            struct ggml_cgraph  * gf,
-            struct ggml_cgraph  * gb,
-            float                 alpha,
-            float                 beta1,
-            float                 beta2,
-            float                 eps,
-            float                 wd); // weight decay
+    GGML_API void ggml_build_forward_expand(struct ggml_cgraph * cgraph, struct ggml_tensor * tensor);
+    GGML_API void ggml_build_backward_expand(
+        struct ggml_context * ctx_static,  // context for static gradients (loss + gradient accumulation)
+        struct ggml_context * ctx_compute, // context for gradient computation
+        struct ggml_cgraph  * cgraph,
+        bool                  accumulate); // whether or not gradients should be accumulated, requires static allocation of tensors in ctx_static
 
     // graph allocation in a context
     GGML_API struct ggml_cgraph * ggml_new_graph       (struct ggml_context * ctx); // size = GGML_DEFAULT_GRAPH_SIZE, grads = false
@@ -2094,28 +2017,9 @@ extern "C" {
     GGML_API size_t ggml_graph_overhead(void);
     GGML_API size_t ggml_graph_overhead_custom(size_t size, bool grads);
 
-    GGML_API struct ggml_threadpool_params ggml_threadpool_params_default(int n_threads);
-    GGML_API void                          ggml_threadpool_params_init   (struct ggml_threadpool_params * p, int n_threads);
-    GGML_API bool                          ggml_threadpool_params_match  (const struct ggml_threadpool_params * p0, const struct ggml_threadpool_params * p1);
-    GGML_API struct ggml_threadpool *      ggml_threadpool_new          (struct ggml_threadpool_params  * params);
-    GGML_API void                          ggml_threadpool_free         (struct ggml_threadpool * threadpool);
-    GGML_API int                           ggml_threadpool_get_n_threads(struct ggml_threadpool * threadpool);
-    GGML_API void                          ggml_threadpool_pause        (struct ggml_threadpool * threadpool);
-    GGML_API void                          ggml_threadpool_resume       (struct ggml_threadpool * threadpool);
-
-    // ggml_graph_plan() has to be called before ggml_graph_compute()
-    // when plan.work_size > 0, caller must allocate memory for plan.work_data
-    GGML_API struct ggml_cplan ggml_graph_plan(
-                  const struct ggml_cgraph * cgraph,
-                                       int   n_threads, /* = GGML_DEFAULT_N_THREADS */
-                    struct ggml_threadpool * threadpool /* = NULL */ );
-    GGML_API enum ggml_status  ggml_graph_compute(struct ggml_cgraph * cgraph, struct ggml_cplan * cplan);
-
-    // same as ggml_graph_compute() but the work data is allocated as a part of the context
-    // note: the drawback of this API is that you must have ensured that the context has enough memory for the work data
-    GGML_API enum ggml_status  ggml_graph_compute_with_ctx(struct ggml_context * ctx, struct ggml_cgraph * cgraph, int n_threads);
-
-    GGML_API struct ggml_tensor * ggml_graph_get_tensor(struct ggml_cgraph * cgraph, const char * name);
+    GGML_API struct ggml_tensor * ggml_graph_get_tensor  (const struct ggml_cgraph * cgraph, const char * name);
+    GGML_API struct ggml_tensor * ggml_graph_get_grad    (const struct ggml_cgraph * cgraph, const struct ggml_tensor * node);
+    GGML_API struct ggml_tensor * ggml_graph_get_grad_acc(const struct ggml_cgraph * cgraph, const struct ggml_tensor * node);
 
     GGML_API void                 ggml_graph_export(const struct ggml_cgraph * cgraph, const char * fname);
     GGML_API struct ggml_cgraph * ggml_graph_import(const char * fname, struct ggml_context ** ctx_data, struct ggml_context ** ctx_eval);
@@ -2126,201 +2030,14 @@ extern "C" {
     // dump the graph into a file using the dot format
     GGML_API void ggml_graph_dump_dot(const struct ggml_cgraph * gb, const struct ggml_cgraph * gf, const char * filename);
 
-    // build gradient checkpointing backward graph gb for gf using provided checkpoints
-    // gb_tmp will contain original backward graph with rewritten backward process nodes,
-    // but without the second forward pass nodes.
-    GGML_API void ggml_build_backward_gradient_checkpointing(
-            struct ggml_context   * ctx,
-            struct ggml_cgraph    * gf,
-            struct ggml_cgraph    * gb,
-            struct ggml_cgraph    * gb_tmp,
-            struct ggml_tensor  * * checkpoints,
-            int                     n_checkpoints);
-    //
-    // optimization
-    //
-
-    // optimization methods
-    enum ggml_opt_type {
-        GGML_OPT_TYPE_ADAM,
-        GGML_OPT_TYPE_LBFGS,
-    };
-
-    // linesearch methods
-    enum ggml_linesearch {
-        GGML_LINESEARCH_DEFAULT = 1,
-
-        GGML_LINESEARCH_BACKTRACKING_ARMIJO       = 0,
-        GGML_LINESEARCH_BACKTRACKING_WOLFE        = 1,
-        GGML_LINESEARCH_BACKTRACKING_STRONG_WOLFE = 2,
-    };
-
-    // optimization return values
-    enum ggml_opt_result {
-        GGML_OPT_RESULT_OK = 0,
-        GGML_OPT_RESULT_DID_NOT_CONVERGE,
-        GGML_OPT_RESULT_NO_CONTEXT,
-        GGML_OPT_RESULT_INVALID_WOLFE,
-        GGML_OPT_RESULT_FAIL,
-        GGML_OPT_RESULT_CANCEL,
-
-        GGML_LINESEARCH_FAIL = -128,
-        GGML_LINESEARCH_MINIMUM_STEP,
-        GGML_LINESEARCH_MAXIMUM_STEP,
-        GGML_LINESEARCH_MAXIMUM_ITERATIONS,
-        GGML_LINESEARCH_INVALID_PARAMETERS,
-    };
-
-    typedef void (*ggml_opt_callback)(void * data, int accum_step, float * sched, bool * cancel);
+    // TODO these functions were sandwiched in the old optimization interface, is there a better place for them?
     typedef void (*ggml_log_callback)(enum ggml_log_level level, const char * text, void * user_data);
 
     // Set callback for all future logging events.
     // If this is not called, or NULL is supplied, everything is output on stderr.
     GGML_API void ggml_log_set(ggml_log_callback log_callback, void * user_data);
 
-    // optimization parameters
-    //
-    //   see ggml.c (ggml_opt_default_params) for default values
-    //
-    struct ggml_opt_params {
-        enum ggml_opt_type type;
-
-        size_t graph_size;
-
-        int n_threads;
-
-        // delta-based convergence test
-        //
-        //   if past == 0 - disabled
-        //   if past > 0:
-        //     stop if |f(x) - f(x_past)| < delta * max(1, |f(x)|)
-        //
-        int past;
-        float delta;
-
-        // maximum number of iterations without improvement
-        //
-        //   if 0 - disabled
-        //   if > 0:
-        //     assume convergence if no cost improvement in this number of iterations
-        //
-        int max_no_improvement;
-
-        bool print_forward_graph;
-        bool print_backward_graph;
-
-        int n_gradient_accumulation;
-
-        // ADAM parameters
-        struct {
-            int n_iter;
-
-            float sched; // schedule multiplier (fixed, decay or warmup)
-            float decay; // weight decay for AdamW, use 0.0f to disable
-            int   decay_min_ndim; // minimum number of tensor dimension to apply weight decay
-            float alpha; // learning rate
-            float beta1;
-            float beta2;
-            float eps;   // epsilon for numerical stability
-            float eps_f; // epsilon for convergence test
-            float eps_g; // epsilon for convergence test
-            float gclip; // gradient clipping
-        } adam;
-
-        // LBFGS parameters
-        struct {
-            int m; // number of corrections to approximate the inv. Hessian
-            int n_iter;
-            int max_linesearch;
-
-            float eps;      // convergence tolerance
-            float ftol;     // line search tolerance
-            float wolfe;
-            float min_step;
-            float max_step;
-
-            enum ggml_linesearch linesearch;
-        } lbfgs;
-    };
-
-    struct ggml_opt_context {
-        struct ggml_context * ctx;
-        struct ggml_opt_params params;
-
-        int iter;
-        int64_t nx; // number of parameter elements
-
-        bool just_initialized;
-
-        float loss_before;
-        float loss_after;
-
-        struct {
-            struct ggml_tensor * g;  // current gradient
-            struct ggml_tensor * m;  // first moment
-            struct ggml_tensor * v;  // second moment
-            struct ggml_tensor * pf; // past function values
-            float fx_best;
-            float fx_prev;
-            int n_no_improvement;
-        } adam;
-
-        struct {
-            struct ggml_tensor * x;    // current parameters
-            struct ggml_tensor * xp;   // previous parameters
-            struct ggml_tensor * g;    // current gradient
-            struct ggml_tensor * gp;   // previous gradient
-            struct ggml_tensor * d;    // search direction
-            struct ggml_tensor * pf;   // past function values
-            struct ggml_tensor * lmal; // the L-BFGS memory alpha
-            struct ggml_tensor * lmys; // the L-BFGS memory ys
-            struct ggml_tensor * lms;  // the L-BFGS memory s
-            struct ggml_tensor * lmy;  // the L-BFGS memory y
-            float fx_best;
-            float step;
-            int j;
-            int k;
-            int end;
-            int n_no_improvement;
-        } lbfgs;
-    };
-
-    GGML_API struct ggml_opt_params ggml_opt_default_params(enum ggml_opt_type type);
-
-    // optimize the function defined by the tensor f
-    GGML_API enum ggml_opt_result ggml_opt(
-            struct ggml_context * ctx,
-            struct ggml_opt_params params,
-            struct ggml_tensor * f);
-
-    // initialize optimizer context
-    GGML_API void ggml_opt_init(
-            struct ggml_context     * ctx,
-            struct ggml_opt_context * opt,
-            struct ggml_opt_params    params,
-            int64_t                   nx);
-
-    // continue optimizing the function defined by the tensor f
-    GGML_API enum ggml_opt_result ggml_opt_resume(
-            struct ggml_context * ctx,
-            struct ggml_opt_context * opt,
-            struct ggml_tensor * f);
-
-    // continue optimizing the function defined by the tensor f
-    GGML_API enum ggml_opt_result ggml_opt_resume_g(
-            struct ggml_context * ctx,
-            struct ggml_opt_context * opt,
-            struct ggml_tensor * f,
-            struct ggml_cgraph * gf,
-            struct ggml_cgraph * gb,
-            ggml_opt_callback callback,
-            void * callback_data);
-
-    //
-    // tensor flags
-    //
-    GGML_API void ggml_set_input(struct ggml_tensor * tensor);
-    GGML_API void ggml_set_output(struct ggml_tensor * tensor);
+    GGML_API struct ggml_tensor * ggml_set_zero(struct ggml_tensor * tensor);
 
     //
     // quantization
@@ -2477,47 +2194,6 @@ extern "C" {
     GGML_API size_t gguf_get_meta_size(const struct gguf_context * ctx);
     GGML_API void   gguf_get_meta_data(const struct gguf_context * ctx, void * data);
 
-    //
-    // system info
-    //
-
-    GGML_API int ggml_cpu_has_avx        (void);
-    GGML_API int ggml_cpu_has_avx_vnni   (void);
-    GGML_API int ggml_cpu_has_avx2       (void);
-    GGML_API int ggml_cpu_has_avx512     (void);
-    GGML_API int ggml_cpu_has_avx512_vbmi(void);
-    GGML_API int ggml_cpu_has_avx512_vnni(void);
-    GGML_API int ggml_cpu_has_avx512_bf16(void);
-    GGML_API int ggml_cpu_has_fma        (void);
-    GGML_API int ggml_cpu_has_neon       (void);
-    GGML_API int ggml_cpu_has_sve        (void);
-    GGML_API int ggml_cpu_has_arm_fma    (void);
-    GGML_API int ggml_cpu_has_metal      (void);
-    GGML_API int ggml_cpu_has_f16c       (void);
-    GGML_API int ggml_cpu_has_fp16_va    (void);
-    GGML_API int ggml_cpu_has_wasm_simd  (void);
-    GGML_API int ggml_cpu_has_blas       (void);
-    GGML_API int ggml_cpu_has_cuda       (void);
-    GGML_API int ggml_cpu_has_vulkan     (void);
-    GGML_API int ggml_cpu_has_kompute    (void);
-    GGML_API int ggml_cpu_has_gpublas    (void);
-    GGML_API int ggml_cpu_has_sse3       (void);
-    GGML_API int ggml_cpu_has_ssse3      (void);
-    GGML_API int ggml_cpu_has_riscv_v    (void);
-    GGML_API int ggml_cpu_has_sycl       (void);
-    GGML_API int ggml_cpu_has_rpc        (void);
-    GGML_API int ggml_cpu_has_vsx        (void);
-    GGML_API int ggml_cpu_has_matmul_int8(void);
-    GGML_API int ggml_cpu_has_cann       (void);
-    GGML_API int ggml_cpu_has_llamafile  (void);
-
-    // get the sve vector length in bytes
-    GGML_API int ggml_cpu_get_sve_cnt(void);
-
-    //
-    // Internal types and functions exposed for tests and benchmarks
-    //
-
 #ifdef  __cplusplus
 // restrict not standard in C++
 #define GGML_RESTRICT
@@ -2526,34 +2202,18 @@ extern "C" {
 #endif
     typedef void (*ggml_to_float_t)  (const void  * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
     typedef void (*ggml_from_float_t)(const float * GGML_RESTRICT x, void  * GGML_RESTRICT y, int64_t k);
-    typedef void (*ggml_from_float_to_mat_t)
-                                     (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t nr, int64_t k, int64_t bs);
-    typedef void (*ggml_vec_dot_t)  (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT x, size_t bx,
-                                       const void * GGML_RESTRICT y, size_t by, int nrc);
-    typedef void (*ggml_gemv_t)     (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT x,
-                                       const void * GGML_RESTRICT y, int nr, int nc);
-    typedef void (*ggml_gemm_t)     (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT x,
-                                       const void * GGML_RESTRICT y, int nr, int nc);
 
-    typedef struct {
+    struct ggml_type_traits {
         const char             * type_name;
         int64_t                  blck_size;
         int64_t                  blck_size_interleave; // interleave elements in blocks
         size_t                   type_size;
         bool                     is_quantized;
         ggml_to_float_t          to_float;
-        ggml_from_float_t        from_float;
         ggml_from_float_t        from_float_ref;
-        ggml_from_float_to_mat_t from_float_to_mat;
-        ggml_vec_dot_t           vec_dot;
-        enum ggml_type           vec_dot_type;
-        int64_t                  nrows; // number of rows to process simultaneously
-        int64_t                  ncols; // number of columns to process simultaneously
-        ggml_gemv_t              gemv;
-        ggml_gemm_t              gemm;
-    } ggml_type_traits_t;
+    };
 
-    GGML_API ggml_type_traits_t ggml_internal_get_type_traits(enum ggml_type type);
+    GGML_API const struct ggml_type_traits * ggml_get_type_traits(enum ggml_type type);
 
 #ifdef  __cplusplus
 }

ggml/src/ggml-aarch64.h

@@ -1,9 +1,5 @@
-// SPDX-FileCopyrightText: Copyright 2024 Arm Ltd.
 #pragma once
 
-#define GGML_COMMON_DECL_C
-#include "ggml-common.h"
-
 #include "ggml.h"
 
 // GGML internal header
@@ -12,27 +8,11 @@
 extern "C" {
 #endif
 
-// Quantization
-void quantize_q8_0_4x4(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_q8_0_4x8(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-
-void quantize_mat_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t nrows, int64_t n_per_row, int64_t blck_size_interleave);
-
 // Quantization utilizing an importance matrix (a.k.a. "Activation aWare Quantization")
 size_t quantize_q4_0_4x4(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
 size_t quantize_q4_0_4x8(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
 size_t quantize_q4_0_8x8(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
 
-// GEMV
-void ggml_gemv_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
-void ggml_gemv_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
-void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
-
-// GEMM
-void ggml_gemm_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
-void ggml_gemm_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
-void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
-
 #ifdef __cplusplus
 }
 #endif

ggml/src/ggml-alloc.c

@@ -14,7 +14,7 @@
 
 //#define GGML_ALLOCATOR_DEBUG
 
-//#define AT_PRINTF(...) fprintf(stderr, __VA_ARGS__)
+//#define AT_PRINTF(...) GGML_LOG_DEBUG(__VA_ARGS__)
 #define AT_PRINTF(...)
 
 
@@ -89,7 +89,7 @@ void ggml_tallocr_alloc(struct ggml_tallocr * talloc, struct ggml_tensor * tenso
     size = GGML_PAD(size, talloc->alignment);
 
     if (talloc->offset + size > ggml_backend_buffer_get_size(talloc->buffer)) {
-        fprintf(stderr, "%s: not enough space in the buffer to allocate %s (needed %zu, available %zu)\n",
+        GGML_LOG_ERROR("%s: not enough space in the buffer to allocate %s (needed %zu, available %zu)\n",
                 __func__, tensor->name, size, ggml_backend_buffer_get_size(talloc->buffer) - talloc->offset);
         GGML_ABORT("not enough space in the buffer");
     }
@@ -172,7 +172,7 @@ static size_t ggml_dyn_tallocr_alloc(struct ggml_dyn_tallocr * alloc, size_t siz
             best_fit_block = alloc->n_free_blocks - 1;
         } else {
             // this should never happen
-            fprintf(stderr, "%s: not enough space in the buffer to allocate %zu bytes, largest block available %zu bytes\n",
+            GGML_LOG_ERROR("%s: not enough space in the buffer to allocate %zu bytes, largest block available %zu bytes\n",
                     __func__, size, max_avail);
             GGML_ABORT("not enough space in the buffer");
         }
@@ -209,16 +209,16 @@ static size_t ggml_dyn_tallocr_alloc(struct ggml_dyn_tallocr * alloc, size_t siz
                 }
             }
         }
-        fprintf(stderr, "max_size = %.2f MB: tensors: ", cur_max / 1024.0 / 1024.0);
+        GGML_LOG_DEBUG("max_size = %.2f MB: tensors: ", cur_max / 1024.0 / 1024.0);
         for (int i = 0; i < 1024; i++) {
             if (alloc->allocated_tensors[i].tensor) {
-                fprintf(stderr, "%s [%zx-%zx] (%.2f MB) ", alloc->allocated_tensors[i].tensor->name,
+                GGML_LOG_DEBUG("%s [%zx-%zx] (%.2f MB) ", alloc->allocated_tensors[i].tensor->name,
                     alloc->allocated_tensors[i].offset,
                     alloc->allocated_tensors[i].offset + ggml_nbytes(alloc->allocated_tensors[i].tensor),
                     ggml_nbytes(alloc->allocated_tensors[i].tensor) / 1024.0 / 1024.0);
             }
         }
-        fprintf(stderr, "\n");
+        GGML_LOG_DEBUG("\n");
     }
 #endif
 
@@ -348,7 +348,6 @@ struct tensor_alloc {
 };
 
 struct leaf_alloc {
-    int buffer_id;
     struct tensor_alloc leaf;
 };
 
@@ -467,18 +466,12 @@ static bool ggml_gallocr_is_own(ggml_gallocr_t galloc, struct ggml_tensor * t) {
     return ggml_gallocr_hash_get(galloc, t)->allocated;
 }
 
-static void ggml_gallocr_set_node_offset(ggml_gallocr_t galloc, struct ggml_tensor * node, int buffer_id, size_t offset) {
-    struct hash_node * hn = ggml_gallocr_hash_get(galloc, node);
-    hn->buffer_id = buffer_id;
-    hn->offset = offset;
-    hn->allocated = true;
-}
-
 static bool ggml_gallocr_is_allocated(ggml_gallocr_t galloc, struct ggml_tensor * t) {
     return t->data != NULL || ggml_gallocr_hash_get(galloc, t)->allocated;
 }
 
 static void ggml_gallocr_allocate_node(ggml_gallocr_t galloc, struct ggml_tensor * node, int buffer_id) {
+    GGML_ASSERT(buffer_id >= 0);
     struct hash_node * hn = ggml_gallocr_hash_get(galloc, node);
 
     if (!ggml_gallocr_is_allocated(galloc, node) && !ggml_is_view(node)) {
@@ -740,7 +733,6 @@ bool ggml_gallocr_reserve_n(ggml_gallocr_t galloc, struct ggml_cgraph * graph, c
     for (int i = 0; i < graph->n_leafs; i++) {
         struct ggml_tensor * leaf = graph->leafs[i];
         struct hash_node * hn = ggml_gallocr_hash_get(galloc, leaf);
-        galloc->leaf_allocs[i].buffer_id = hn->buffer_id;
         if (leaf->view_src || leaf->data) {
             galloc->leaf_allocs[i].leaf.buffer_id = -1;
             galloc->leaf_allocs[i].leaf.offset = SIZE_MAX;
@@ -768,13 +760,13 @@ bool ggml_gallocr_reserve_n(ggml_gallocr_t galloc, struct ggml_cgraph * graph, c
         // even if there are no tensors allocated in this buffer, we still need to allocate it to initialize views
         if (new_size > cur_size || galloc->buffers[i] == NULL) {
 #ifndef NDEBUG
-            fprintf(stderr, "%s: reallocating %s buffer from size %.02f MiB to %.02f MiB\n", __func__, ggml_backend_buft_name(galloc->bufts[i]), cur_size / 1024.0 / 1024.0, new_size / 1024.0 / 1024.0);
+            GGML_LOG_DEBUG("%s: reallocating %s buffer from size %.02f MiB to %.02f MiB\n", __func__, ggml_backend_buft_name(galloc->bufts[i]), cur_size / 1024.0 / 1024.0, new_size / 1024.0 / 1024.0);
 #endif
 
             ggml_backend_buffer_free(galloc->buffers[i]);
             galloc->buffers[i] = ggml_backend_buft_alloc_buffer(galloc->bufts[i], new_size);
             if (galloc->buffers[i] == NULL) {
-                fprintf(stderr, "%s: failed to allocate %s buffer of size %zu\n", __func__, ggml_backend_buft_name(galloc->bufts[i]), new_size);
+                GGML_LOG_ERROR("%s: failed to allocate %s buffer of size %zu\n", __func__, ggml_backend_buft_name(galloc->bufts[i]), new_size);
                 return false;
             }
             ggml_backend_buffer_set_usage(galloc->buffers[i], GGML_BACKEND_BUFFER_USAGE_COMPUTE);
@@ -818,21 +810,25 @@ static void ggml_gallocr_init_tensor(ggml_gallocr_t galloc, struct ggml_tensor *
 }
 
 static bool ggml_gallocr_node_needs_realloc(ggml_gallocr_t galloc, struct ggml_tensor * node, struct tensor_alloc * talloc) {
-    size_t node_size = (node->data || node->view_src) ? 0 : ggml_backend_buft_get_alloc_size(galloc->bufts[talloc->buffer_id], node);
+    size_t node_size = 0;
+    if (!node->data && !node->view_src) {
+        GGML_ASSERT(talloc->buffer_id >= 0); // prevent segfault when misusing the API
+        node_size = ggml_backend_buft_get_alloc_size(galloc->bufts[talloc->buffer_id], node);
+    }
     return talloc->size_max >= node_size;
 }
 
 static bool ggml_gallocr_needs_realloc(ggml_gallocr_t galloc, struct ggml_cgraph * graph) {
     if (galloc->n_nodes != graph->n_nodes) {
 #ifndef NDEBUG
-        fprintf(stderr, "%s: graph has different number of nodes\n", __func__);
+        GGML_LOG_DEBUG("%s: graph has different number of nodes\n", __func__);
 #endif
         return true;
     }
 
     if (galloc->n_leafs != graph->n_leafs) {
 #ifndef NDEBUG
-        fprintf(stderr, "%s: graph has different number of leafs\n", __func__);
+        GGML_LOG_DEBUG("%s: graph has different number of leafs\n", __func__);
 #endif
         return true;
     }
@@ -843,7 +839,7 @@ static bool ggml_gallocr_needs_realloc(ggml_gallocr_t galloc, struct ggml_cgraph
 
         if (!ggml_gallocr_node_needs_realloc(galloc, node, &node_alloc->dst)) {
 #ifndef NDEBUG
-            fprintf(stderr, "%s: node %s is not valid\n", __func__, node->name);
+            GGML_LOG_DEBUG("%s: node %s is not valid\n", __func__, node->name);
 #endif
             return true;
         }
@@ -855,7 +851,7 @@ static bool ggml_gallocr_needs_realloc(ggml_gallocr_t galloc, struct ggml_cgraph
             }
             if (!ggml_gallocr_node_needs_realloc(galloc, src, &node_alloc->src[j])) {
 #ifndef NDEBUG
-                fprintf(stderr, "%s: src %d (%s) of node %s is not valid\n", __func__, j, src->name, node->name);
+                GGML_LOG_DEBUG("%s: src %d (%s) of node %s is not valid\n", __func__, j, src->name, node->name);
 #endif
                 return true;
             }
@@ -869,14 +865,14 @@ bool ggml_gallocr_alloc_graph(ggml_gallocr_t galloc, struct ggml_cgraph * graph)
     if (ggml_gallocr_needs_realloc(galloc, graph)) {
         if (galloc->n_buffers == 1) {
 #ifndef NDEBUG
-            fprintf(stderr, "%s: reallocating buffers automatically\n", __func__);
+            GGML_LOG_DEBUG("%s: reallocating buffers automatically\n", __func__);
 #endif
             if (!ggml_gallocr_reserve(galloc, graph)) {
                 return false;
             }
         } else {
 #ifndef NDEBUG
-            fprintf(stderr, "%s: cannot reallocate multi buffer graph automatically, call reserve\n", __func__);
+            GGML_LOG_DEBUG("%s: cannot reallocate multi buffer graph automatically, call reserve\n", __func__);
 #endif
             return false;
         }
@@ -940,7 +936,7 @@ static bool alloc_tensor_range(struct ggml_context * ctx,
     ggml_backend_buffer_t buffer = ggml_backend_buft_alloc_buffer(buft, size);
     if (buffer == NULL) {
 #ifndef NDEBUG
-        fprintf(stderr, "%s: failed to allocate %s buffer of size %zu\n", __func__, ggml_backend_buft_name(buft), size);
+        GGML_LOG_DEBUG("%s: failed to allocate %s buffer of size %zu\n", __func__, ggml_backend_buft_name(buft), size);
 #endif
         for (size_t i = 0; i < *n_buffers; i++) {
             ggml_backend_buffer_free((*buffers)[i]);
@@ -990,7 +986,7 @@ ggml_backend_buffer_t ggml_backend_alloc_ctx_tensors_from_buft(struct ggml_conte
         }
 
         if (this_size > max_size) {
-            fprintf(stderr, "%s: tensor %s is too large to fit in a %s buffer (tensor size: %zu, max buffer size: %zu)\n",
+            GGML_LOG_ERROR("%s: tensor %s is too large to fit in a %s buffer (tensor size: %zu, max buffer size: %zu)\n",
                     __func__, t->name,
                     ggml_backend_buft_name(buft),
                     this_size, max_size);
@@ -1022,7 +1018,7 @@ ggml_backend_buffer_t ggml_backend_alloc_ctx_tensors_from_buft(struct ggml_conte
 
     if (n_buffers == 0) {
 #ifndef NDEBUG
-        fprintf(stderr, "%s: all tensors in the context are already allocated\n", __func__);
+        GGML_LOG_DEBUG("%s: all tensors in the context are already allocated\n", __func__);
 #endif
         return NULL;
     }

ggml/src/ggml-amx/CMakeLists.txt

@@ -0,0 +1,107 @@
+if (CMAKE_OSX_ARCHITECTURES STREQUAL "x86_64" OR CMAKE_GENERATOR_PLATFORM_LWR MATCHES "^(x86_64|i686|amd64|x64|win32)$" OR
+        (NOT CMAKE_OSX_ARCHITECTURES AND NOT CMAKE_GENERATOR_PLATFORM_LWR AND
+         CMAKE_SYSTEM_PROCESSOR MATCHES "^(x86_64|i686|AMD64)$") AND
+        CMAKE_COMPILER_IS_GNUCC AND CMAKE_CXX_COMPILER_VERSION VERSION_GREATER 11.0)
+    message(STATUS "Using AMX")
+
+    file(GLOB   GGML_HEADERS_AMX "*.h")
+    list(APPEND GGML_HEADERS_AMX "../../include/ggml-amx.h")
+
+    file(GLOB   GGML_SOURCES_AMX "*.cpp")
+
+    add_library(ggml-amx
+                ${GGML_HEADERS_AMX}
+                ${GGML_SOURCES_AMX})
+
+    target_link_libraries(ggml-amx PRIVATE ggml-base)
+    target_include_directories(ggml-amx PRIVATE . ..)
+
+    # this is duplicated from the CPU backend, since the AMX backend also depends on the architecture flags
+    # TODO: integrate AMX backend into the CPU backend
+    if (MSVC)
+        # instruction set detection for MSVC only
+        if (GGML_NATIVE)
+            # TODO: improve, should not reference files from the parent folder
+            include(../ggml-cpu/cmake/FindSIMD.cmake)
+        endif ()
+        if (GGML_AVX512)
+            list(APPEND ARCH_FLAGS /arch:AVX512)
+            # MSVC has no compile-time flags enabling specific
+            # AVX512 extensions, neither it defines the
+            # macros corresponding to the extensions.
+            # Do it manually.
+            if (GGML_AVX512_VBMI)
+                add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AVX512VBMI__>)
+                add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AVX512VBMI__>)
+            endif()
+            if (GGML_AVX512_VNNI)
+                add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AVX512VNNI__>)
+                add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AVX512VNNI__>)
+            endif()
+            if (GGML_AVX512_BF16)
+                add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AVX512BF16__>)
+                add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AVX512BF16__>)
+            endif()
+            if (GGML_AMX_TILE)
+                add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AMX_TILE__>)
+                add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AMX_TILE__>)
+            endif()
+            if (GGML_AMX_INT8)
+                add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AMX_INT8__>)
+                add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AMX_INT8__>)
+            endif()
+            if (GGML_AMX_BF16)
+                add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AMX_BF16__>)
+                add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AMX_BF16__>)
+            endif()
+        elseif (GGML_AVX2)
+            list(APPEND ARCH_FLAGS /arch:AVX2)
+        elseif (GGML_AVX)
+            list(APPEND ARCH_FLAGS /arch:AVX)
+        endif()
+    else()
+        if (GGML_NATIVE)
+            list(APPEND ARCH_FLAGS -march=native)
+        endif()
+        if (GGML_F16C)
+            list(APPEND ARCH_FLAGS -mf16c)
+        endif()
+        if (GGML_FMA)
+            list(APPEND ARCH_FLAGS -mfma)
+        endif()
+        if (GGML_AVX)
+            list(APPEND ARCH_FLAGS -mavx)
+        endif()
+        if (GGML_AVX2)
+            list(APPEND ARCH_FLAGS -mavx2)
+        endif()
+        if (GGML_AVX512)
+            list(APPEND ARCH_FLAGS -mavx512f)
+            list(APPEND ARCH_FLAGS -mavx512dq)
+            list(APPEND ARCH_FLAGS -mavx512bw)
+        endif()
+        if (GGML_AVX512_VBMI)
+            list(APPEND ARCH_FLAGS -mavx512vbmi)
+        endif()
+        if (GGML_AVX512_VNNI)
+            list(APPEND ARCH_FLAGS -mavx512vnni)
+        endif()
+        if (GGML_AVX512_BF16)
+            list(APPEND ARCH_FLAGS -mavx512bf16)
+        endif()
+        if (GGML_AMX_TILE)
+            list(APPEND ARCH_FLAGS -mamx-tile)
+        endif()
+        if (GGML_AMX_INT8)
+            list(APPEND ARCH_FLAGS -mamx-int8)
+        endif()
+        if (GGML_AMX_BF16)
+            list(APPEND ARCH_FLAGS -mamx-bf16)
+        endif()
+    endif()
+
+    target_compile_options(ggml-amx PRIVATE ${ARCH_FLAGS})
+else()
+    set(GGML_AMX OFF PARENT_SCOPE)
+    message(WARNING "AMX requires x86 and gcc version > 11.0. Turning off GGML_AMX.")
+endif()

ggml/src/ggml-amx/common.h

@@ -0,0 +1,94 @@
+#pragma once
+
+#include "ggml.h"
+// hack until AMX is moved into the CPU backend
+#include "../ggml-cpu/ggml-cpu-impl.h" // <immintrin.h>
+
+#include <algorithm>
+#include <memory>
+#include <type_traits>
+
+#if defined(_OPENMP)
+#include <omp.h>
+#endif
+
+#define TILE_M 16
+#define TILE_N 16
+#define TILE_K 32
+#define VNNI_BLK 4
+
+#define AMX_BLK_SIZE 32
+
+#define TMM0 0
+#define TMM1 1
+#define TMM2 2
+#define TMM3 3
+#define TMM4 4
+#define TMM5 5
+#define TMM6 6
+#define TMM7 7
+
+// parallel routines
+template <typename T, typename std::enable_if<std::is_integral<T>::value, int>::type = 0>
+inline T div_up(T x, T y) { return (x + y - 1) / y; }
+
+template <typename T>
+inline void balance211(T n, T nth, T ith, T& n_start, T& n_end) {
+#if 0
+    // onednn partition pattern
+    T& n_my = n_end;
+    if (nth <= 1 || n == 0) {
+        n_start = 0;
+        n_my = n;
+    } else {
+        T n1 = div_up(n, nth);
+        T n2 = n1 - 1;
+        T T1 = n - n2 * nth;
+        n_my = ith < T1 ? n1 : n2;
+        n_start = ith <= T1 ? ith*n1 : T1 * n1 + (ith - T1) * n2;
+    }
+    n_end += n_start;
+#else
+    // pytorch aten partition pattern
+    T n_my = div_up(n, nth);
+    n_start = ith * n_my;
+    n_end = std::min(n_start + n_my, n);
+#endif
+}
+
+template <typename func_t>
+inline void parallel_for(int nth, int n, const func_t& f) {
+#if defined(_OPENMP)
+#pragma omp parallel num_threads(nth)
+{
+    //int nth = omp_get_num_threads();
+    int ith = omp_get_thread_num();
+    int tbegin, tend;
+    balance211(n, nth, ith, tbegin, tend);
+    f(tbegin, tend);
+}
+#else
+    f(0, n);
+
+    GGML_UNUSED(nth);
+#endif
+}
+
+// quantized types that have AMX support
+inline bool qtype_has_amx_kernels(const enum ggml_type type) {
+    // TODO: fix padding for vnni format
+    return (type == GGML_TYPE_Q4_0) ||
+        (type == GGML_TYPE_Q4_1);
+        //(type == GGML_TYPE_Q8_0) ||
+        //(type == GGML_TYPE_Q4_K) ||
+        //(type == GGML_TYPE_Q5_K) ||
+        //(type == GGML_TYPE_Q6_K) ||
+        //(type == GGML_TYPE_IQ4_XS);
+}
+
+// ggml backend context
+struct ggml_backend_amx_context {
+    int n_threads = GGML_DEFAULT_N_THREADS;
+    std::unique_ptr<char[]> work_data;
+    size_t work_size = 0;
+};

ggml/src/ggml-amx/ggml-amx.cpp

@@ -0,0 +1,446 @@
+#include "ggml-amx.h"
+#include "ggml-amx/common.h"
+#include "ggml-amx/mmq.h"
+#include "ggml-backend-impl.h"
+#include "ggml-impl.h"
+
+#if defined(__gnu_linux__)
+#include <sys/syscall.h>
+#include <unistd.h>
+#endif
+
+#include <cstdlib>
+#include <cstring>
+#include <memory>
+
+#if defined(__AMX_INT8__)
+
+// AMX buffer interface
+static void ggml_backend_amx_buffer_free_buffer(ggml_backend_buffer_t buffer) {
+    free(buffer->context);
+}
+
+static void * ggml_backend_amx_buffer_get_base(ggml_backend_buffer_t buffer) {
+    return (void *)(buffer->context);
+}
+
+static void ggml_backend_amx_buffer_memset_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, uint8_t value, size_t offset, size_t size) {
+    memset((char *)tensor->data + offset, value, size);
+
+    GGML_UNUSED(buffer);
+}
+
+static void ggml_backend_amx_buffer_set_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
+    if (qtype_has_amx_kernels(tensor->type)) {
+        ggml_backend_amx_convert_weight(tensor, data, offset, size);
+    } else {
+        memcpy((char *)tensor->data + offset, data, size);
+    }
+
+    GGML_UNUSED(buffer);
+}
+
+static void ggml_backend_amx_buffer_get_tensor(ggml_backend_buffer_t buffer, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) {
+    GGML_ASSERT(!qtype_has_amx_kernels(tensor->type));
+    memcpy(data, (const char *)tensor->data + offset, size);
+
+    GGML_UNUSED(buffer);
+}
+
+static bool ggml_backend_amx_buffer_cpy_tensor(ggml_backend_buffer_t buffer, const struct ggml_tensor * src, struct ggml_tensor * dst) {
+    if (ggml_backend_buffer_is_host(src->buffer)) {
+        if (qtype_has_amx_kernels(src->type)) {
+            ggml_backend_amx_convert_weight(dst, src->data, 0, ggml_backend_amx_get_alloc_size(dst));
+        } else {
+            memcpy(dst->data, src->data, ggml_nbytes(src));
+        }
+        return true;
+    }
+    return false;
+
+    GGML_UNUSED(buffer);
+}
+
+static void ggml_backend_amx_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
+    memset(buffer->context, value, buffer->size);
+}
+
+static ggml_backend_buffer_i ggml_backend_amx_buffer_interface = {
+    /* .free_buffer     = */ ggml_backend_amx_buffer_free_buffer,
+    /* .get_base        = */ ggml_backend_amx_buffer_get_base,
+    /* .init_tensor     = */ NULL, // no initialization required
+    /* .memset_tensor   = */ ggml_backend_amx_buffer_memset_tensor,
+    /* .set_tensor      = */ ggml_backend_amx_buffer_set_tensor,
+    /* .get_tensor      = */ ggml_backend_amx_buffer_get_tensor,
+    /* .cpy_tensor      = */ ggml_backend_amx_buffer_cpy_tensor,
+    /* .clear           = */ ggml_backend_amx_buffer_clear,
+    /* .reset           = */ NULL,
+};
+
+static const char * ggml_backend_amx_buffer_type_get_name(ggml_backend_buffer_type_t buft) {
+    return "AMX";
+
+    GGML_UNUSED(buft);
+}
+
+static ggml_backend_buffer_t ggml_backend_amx_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
+    void * data = aligned_alloc(TENSOR_ALIGNMENT, size);
+    if (data == NULL) {
+        fprintf(stderr, "%s: failed to allocate buffer of size %zu\n", __func__, size);
+        return NULL;
+    }
+
+    return ggml_backend_buffer_init(buft, ggml_backend_amx_buffer_interface, data, size);
+}
+
+static size_t ggml_backend_amx_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
+    return TENSOR_ALIGNMENT;
+
+    GGML_UNUSED(buft);
+}
+
+static size_t ggml_backend_amx_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft, const ggml_tensor* tensor) {
+    return ggml_backend_amx_get_alloc_size(tensor);
+
+    GGML_UNUSED(buft);
+}
+
+static bool ggml_backend_amx_buffer_type_is_host(ggml_backend_buffer_type_t buft) {
+    return false;
+
+    GGML_UNUSED(buft);
+}
+
+ggml_backend_buffer_type_t ggml_backend_amx_buffer_type() {
+    static struct ggml_backend_buffer_type ggml_backend_buffer_type_amx = {
+        /* .iface = */ {
+            /* .get_name         = */ ggml_backend_amx_buffer_type_get_name,
+            /* .alloc_buffer     = */ ggml_backend_amx_buffer_type_alloc_buffer,
+            /* .get_alignment    = */ ggml_backend_amx_buffer_type_get_alignment,
+            /* .get_max_size     = */ NULL, // defaults to SIZE_MAX
+            /* .get_alloc_size   = */ ggml_backend_amx_buffer_type_get_alloc_size,
+            /* .is_host          = */ ggml_backend_amx_buffer_type_is_host,
+        },
+        /* .device  = */ ggml_backend_reg_dev_get(ggml_backend_amx_reg(), 0),
+        /* .context = */ NULL,
+    };
+
+    return &ggml_backend_buffer_type_amx;
+}
+
+// backend interface
+
+static const char * ggml_backend_amx_name(ggml_backend_t backend) {
+    return "AMX";
+
+    GGML_UNUSED(backend);
+}
+
+static void ggml_backend_amx_free(ggml_backend_t backend) {
+    ggml_backend_amx_context * ctx = (ggml_backend_amx_context *)backend->context;
+    delete ctx;
+    delete backend;
+}
+
+static enum ggml_status ggml_backend_amx_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
+    ggml_backend_amx_context * ctx = (ggml_backend_amx_context *)backend->context;
+
+    for (int i = 0; i < cgraph->n_nodes; i++) {
+        struct ggml_tensor * node = cgraph->nodes[i];
+
+        switch (node->op) {
+        case GGML_OP_MUL_MAT:
+            ggml_backend_amx_mul_mat(ctx, node);
+            break;
+
+        case GGML_OP_NONE:
+        case GGML_OP_RESHAPE:
+        case GGML_OP_VIEW:
+        case GGML_OP_PERMUTE:
+        case GGML_OP_TRANSPOSE:
+            break;
+
+        default:
+            fprintf(stderr, "%s: unsupported op %s\n", __func__, ggml_op_desc(node));
+            GGML_ASSERT(false);
+        }
+    }
+
+    return GGML_STATUS_SUCCESS;
+
+    GGML_UNUSED(backend);
+}
+
+static struct ggml_backend_i ggml_backend_amx_i = {
+    /* .get_name                = */ ggml_backend_amx_name,
+    /* .free                    = */ ggml_backend_amx_free,
+    /* .set_tensor_async        = */ NULL,
+    /* .get_tensor_async        = */ NULL,
+    /* .cpy_tensor_async        = */ NULL,
+    /* .synchronize             = */ NULL,
+    /* .graph_plan_create       = */ NULL,
+    /* .graph_plan_free         = */ NULL,
+    /* .graph_plan_update       = */ NULL,
+    /* .graph_plan_compute      = */ NULL,
+    /* .graph_compute           = */ ggml_backend_amx_graph_compute,
+    /* .event_record            = */ NULL,
+    /* .event_wait              = */ NULL,
+};
+
+static ggml_guid_t ggml_backend_amx_guid() {
+    static ggml_guid guid = { 0x13, 0xb8, 0xa4, 0xc4, 0xba, 0xfe, 0x51, 0x67, 0x87, 0x44, 0x55, 0x15, 0xb2, 0x35, 0x62, 0x3e };
+    return &guid;
+}
+
+#define ARCH_GET_XCOMP_PERM     0x1022
+#define ARCH_REQ_XCOMP_PERM     0x1023
+#define XFEATURE_XTILECFG       17
+#define XFEATURE_XTILEDATA      18
+
+static bool ggml_amx_init() {
+#if defined(__gnu_linux__)
+    if (syscall(SYS_arch_prctl, ARCH_REQ_XCOMP_PERM, XFEATURE_XTILEDATA)) {
+        fprintf(stderr, "AMX is not ready to be used!\n");
+        return false;
+    }
+    return true;
+#elif defined(_WIN32)
+    return true;
+#endif
+}
+
+ggml_backend_t ggml_backend_amx_init() {
+
+    // invoke a Linux system call to request access to AMX features
+    ggml_amx_init();
+
+    // backend context
+    ggml_backend_amx_context * ctx = new ggml_backend_amx_context;
+
+    // ggml amx backend
+    ggml_backend_t backend = new ggml_backend {
+        /* .guid      = */ ggml_backend_amx_guid(),
+        /* .interface = */ ggml_backend_amx_i,
+        /* .device    = */ ggml_backend_reg_dev_get(ggml_backend_amx_reg(), 0),
+        /* .context   = */ ctx,
+    };
+
+    return backend;
+}
+
+bool ggml_backend_is_amx(ggml_backend_t backend) {
+    return backend != NULL && ggml_guid_matches(backend->guid, ggml_backend_amx_guid());
+}
+
+void ggml_backend_amx_set_n_threads(ggml_backend_t backend_amx, int n_threads) {
+    GGML_ASSERT(ggml_backend_is_amx(backend_amx));
+
+    ggml_backend_amx_context * ctx = (ggml_backend_amx_context *)backend_amx->context;
+    ctx->n_threads = n_threads;
+}
+
+// device interface
+
+static const char * ggml_backend_amx_device_get_name(ggml_backend_dev_t dev) {
+    return "AMX";
+
+    GGML_UNUSED(dev);
+}
+
+static const char * ggml_backend_amx_device_get_description(ggml_backend_dev_t dev) {
+    return "Intel Advanced Matrix Extensions";
+
+    GGML_UNUSED(dev);
+}
+
+static void ggml_backend_amx_device_get_memory(ggml_backend_dev_t dev, size_t * free, size_t * total) {
+    // TODO
+    *free = 0;
+    *total = 0;
+
+    GGML_UNUSED(dev);
+}
+
+static enum ggml_backend_dev_type ggml_backend_amx_device_get_type(ggml_backend_dev_t dev) {
+    return GGML_BACKEND_DEVICE_TYPE_ACCEL;
+
+    GGML_UNUSED(dev);
+}
+
+static void ggml_backend_amx_device_get_props(ggml_backend_dev_t dev, struct ggml_backend_dev_props * props) {
+    props->name        = ggml_backend_amx_device_get_name(dev);
+    props->description = ggml_backend_amx_device_get_description(dev);
+    props->type        = ggml_backend_amx_device_get_type(dev);
+    ggml_backend_amx_device_get_memory(dev, &props->memory_free, &props->memory_total);
+
+    // `buffer_from_host_ptr` is intended to be used in mmap, when memory layout unchanged
+    props->caps = {
+        /* .async                 = */ false,
+        /* .host_buffer           = */ false,
+        /* .buffer_from_host_ptr  = */ false,
+        /* .events                = */ false,
+    };
+}
+
+static ggml_backend_t ggml_backend_amx_device_init(ggml_backend_dev_t dev, const char * params) {
+    return ggml_backend_amx_init();
+
+    GGML_UNUSED(dev);
+    GGML_UNUSED(params);
+}
+
+static ggml_backend_buffer_type_t ggml_backend_amx_device_get_buffer_type(ggml_backend_dev_t dev) {
+    return ggml_backend_amx_buffer_type();
+
+    GGML_UNUSED(dev);
+}
+
+static bool ggml_backend_amx_device_supports_op(ggml_backend_dev_t dev, const struct ggml_tensor * op) {
+
+    // handle only 2d gemm for now
+    auto is_contiguous_2d = [](const struct ggml_tensor * t) {
+        return ggml_is_contiguous(t) && t->ne[3] == 1 && t->ne[2] == 1;
+    };
+
+    switch (op->op) {
+        case GGML_OP_NONE:
+        case GGML_OP_RESHAPE:
+        case GGML_OP_VIEW:
+        case GGML_OP_PERMUTE:
+        case GGML_OP_TRANSPOSE:
+            return true;
+
+        case GGML_OP_MUL_MAT: {
+            const struct ggml_tensor * src0 = op->src[0];
+            const struct ggml_tensor * src1 = op->src[1];
+
+            const enum ggml_type type = src0->type;
+            const int64_t ne0 = op->ne[0];
+
+            // amx kernels enables for Q4_0, Q4_1, Q8_0, F16
+            // Q4_K, Q5_K, Q6_K, IQ4_XS enabled for QK_K = 256
+            bool has_amx_kernels = qtype_has_amx_kernels(type) || (type == GGML_TYPE_F16);
+
+            bool can_use_amx =
+                is_contiguous_2d(src0) &&       // src0 must be contiguous
+                is_contiguous_2d(src1) &&       // src1 must be contiguous
+                src1->type == GGML_TYPE_F32 &&  // src1 must be float32
+                has_amx_kernels &&              // with amx kernel impls
+                ne0 % (TILE_N * 2) == 0;        // out_features is 32x
+
+            return can_use_amx;
+        }
+        default:
+            return false;
+    }
+
+    GGML_UNUSED(dev);
+}
+
+static bool ggml_backend_amx_device_supports_buft(ggml_backend_dev_t dev, ggml_backend_buffer_type_t buft) {
+    return buft->iface.get_name == ggml_backend_amx_buffer_type_get_name;
+
+    GGML_UNUSED(dev);
+}
+
+static const struct ggml_backend_device_i ggml_backend_amx_device_i = {
+    /* .get_name             = */ ggml_backend_amx_device_get_name,
+    /* .get_description      = */ ggml_backend_amx_device_get_description,
+    /* .get_memory           = */ ggml_backend_amx_device_get_memory,
+    /* .get_type             = */ ggml_backend_amx_device_get_type,
+    /* .get_props            = */ ggml_backend_amx_device_get_props,
+    /* .init_backend         = */ ggml_backend_amx_device_init,
+    /* .get_buffer_type      = */ ggml_backend_amx_device_get_buffer_type,
+    /* .get_host_buffer_type = */ NULL,
+    /* .buffer_from_host_ptr = */ NULL,
+    /* .supports_op          = */ ggml_backend_amx_device_supports_op,
+    /* .supports_buft        = */ ggml_backend_amx_device_supports_buft,
+    /* .offload_op           = */ NULL,
+    /* .event_new            = */ NULL,
+    /* .event_free           = */ NULL,
+    /* .event_synchronize    = */ NULL,
+};
+
+// backend reg interface
+
+static const char * ggml_backend_amx_reg_get_name(ggml_backend_reg_t reg) {
+    return "AMX";
+
+    GGML_UNUSED(reg);
+}
+
+static size_t ggml_backend_amx_reg_get_device_count(ggml_backend_reg_t reg) {
+    return 1;
+
+    GGML_UNUSED(reg);
+}
+
+static ggml_backend_dev_t ggml_backend_amx_reg_get_device(ggml_backend_reg_t reg, size_t index) {
+    GGML_ASSERT(index == 0);
+
+    static ggml_backend_device ggml_backend_amx_device = {
+        /* .iface   = */ ggml_backend_amx_device_i,
+        /* .reg     = */ reg,
+        /* .context = */ nullptr,
+    };
+
+    return &ggml_backend_amx_device;
+
+    GGML_UNUSED(reg);
+    GGML_UNUSED(index);
+}
+
+static void * ggml_backend_amx_get_proc_address(ggml_backend_reg_t reg, const char * name) {
+    if (std::strcmp(name, "ggml_backend_set_n_threads") == 0) {
+        return (void *)ggml_backend_amx_set_n_threads;
+    }
+    return NULL;
+
+    GGML_UNUSED(reg);
+    GGML_UNUSED(name);
+}
+
+static const struct ggml_backend_reg_i ggml_backend_amx_reg_i = {
+    /* .get_name         = */ ggml_backend_amx_reg_get_name,
+    /* .get_device_count = */ ggml_backend_amx_reg_get_device_count,
+    /* .get_device       = */ ggml_backend_amx_reg_get_device,
+    /* .get_proc_address = */ ggml_backend_amx_get_proc_address,
+};
+
+ggml_backend_reg_t ggml_backend_amx_reg(void) {
+    static struct ggml_backend_reg ggml_backend_amx_reg = {
+        /* .iface   = */ ggml_backend_amx_reg_i,
+        /* .context = */ NULL,
+    };
+
+    return &ggml_backend_amx_reg;
+}
+
+#else // if defined(__AMX_INT8__)
+
+ggml_backend_buffer_type_t ggml_backend_amx_buffer_type(void) {
+    return nullptr;
+}
+
+bool ggml_backend_is_amx(ggml_backend_t backend) {
+    GGML_UNUSED(backend);
+    return false;
+}
+
+ggml_backend_t ggml_backend_amx_init(void) {
+    fprintf(stderr, "GGML is not compiled with AMX support!\n");
+    return nullptr;
+}
+
+void ggml_backend_amx_set_n_threads(ggml_backend_t backend_amx, int n_threads) {
+    fprintf(stderr, "GGML is not compiled with AMX support!\n");
+
+    GGML_UNUSED(backend_amx);
+    GGML_UNUSED(n_threads);
+}
+
+ggml_backend_reg_t ggml_backend_amx_reg(void) {
+    return nullptr;
+}
+
+#endif

ggml/src/ggml-amx/mmq.h

@@ -0,0 +1,17 @@
+#pragma once
+#include "common.h"
+#include <stdint.h>
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+size_t ggml_backend_amx_get_alloc_size(const struct ggml_tensor * tensor);
+
+void ggml_backend_amx_convert_weight(struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
+
+void ggml_backend_amx_mul_mat(ggml_backend_amx_context * ctx, struct ggml_tensor * dst);
+
+#ifdef __cplusplus
+}
+#endif

ggml/src/ggml-backend-impl.h

@@ -22,7 +22,7 @@ extern "C" {
         size_t                (*get_max_size)  (ggml_backend_buffer_type_t buft);
         // (optional) data size needed to allocate the tensor, including padding (defaults to ggml_nbytes)
         size_t                (*get_alloc_size)(ggml_backend_buffer_type_t buft, const struct ggml_tensor * tensor);
-        // (optional) check if tensor data is in host memory (defaults to false)
+        // (optional) check if tensor data is in host memory and uses standard ggml tensor layout (defaults to false)
         bool                  (*is_host)       (ggml_backend_buffer_type_t buft);
     };
 
@@ -37,7 +37,6 @@ extern "C" {
     //
 
     struct ggml_backend_buffer_i {
-        const char * (*get_name)     (ggml_backend_buffer_t buffer);
         // (optional) free the buffer
         void         (*free_buffer)  (ggml_backend_buffer_t buffer);
         // base address of the buffer
@@ -88,18 +87,16 @@ extern "C" {
 
         void (*free)(ggml_backend_t backend);
 
-        // buffer allocation
-        ggml_backend_buffer_type_t (*get_default_buffer_type)(ggml_backend_t backend);
-
         // (optional) asynchronous tensor data access
         void (*set_tensor_async)(ggml_backend_t backend,       struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
         void (*get_tensor_async)(ggml_backend_t backend, const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
         bool (*cpy_tensor_async)(ggml_backend_t backend_src, ggml_backend_t backend_dst, const struct ggml_tensor * src, struct ggml_tensor * dst);
 
-        // (optional) complete all pending operations
+        // (optional) complete all pending operations (required if the backend supports async operations)
         void (*synchronize)(ggml_backend_t backend);
 
-        // (optional) compute graph with a plan (not used currently)
+        // (optional) graph plans (not used currently)
+        // compute graph with a plan
         ggml_backend_graph_plan_t (*graph_plan_create) (ggml_backend_t backend, const struct ggml_cgraph * cgraph);
         void                      (*graph_plan_free)   (ggml_backend_t backend, ggml_backend_graph_plan_t plan);
         // update the plan with a new graph - this should be faster than creating a new plan when the graph has the same topology
@@ -110,21 +107,6 @@ extern "C" {
         // compute graph (always async if supported by the backend)
         enum ggml_status          (*graph_compute)     (ggml_backend_t backend, struct ggml_cgraph * cgraph);
 
-        // IMPORTANT: these functions have been moved to the device interface and will be removed from the backend interface
-        //            new backends should implement the device interface instead
-
-        // These functions are being moved to the device interface
-        // check if the backend can compute an operation
-        bool (*supports_op)  (ggml_backend_t backend, const struct ggml_tensor * op);
-
-        // check if the backend can use tensors allocated in a buffer type
-        bool (*supports_buft)(ggml_backend_t backend, ggml_backend_buffer_type_t buft);
-
-        // check if the backend wants to run an operation, even if the weights are allocated in a CPU buffer
-        // these should be expensive operations with large batch sizes that may benefit from running on this backend
-        // even if the weight has to be copied from the CPU temporarily
-        bool (*offload_op)   (ggml_backend_t backend, const struct ggml_tensor * op);
-
         // (optional) event synchronization
         // record an event on this stream
         void (*event_record)(ggml_backend_t backend, ggml_backend_event_t event);
@@ -184,9 +166,8 @@ extern "C" {
         // check if the backend can use tensors allocated in a buffer type
         bool (*supports_buft)(ggml_backend_dev_t dev, ggml_backend_buffer_type_t buft);
 
-        // check if the backend wants to run an operation, even if the weights are allocated in a CPU buffer
-        // these should be expensive operations with large batch sizes that may benefit from running on this backend
-        // even if the weight has to be copied from the CPU temporarily
+        // (optional) check if the backend wants to run an operation, even if the weights are allocated in an incompatible buffer
+        // these should be expensive operations that may benefit from running on this backend instead of the CPU backend
         bool (*offload_op)(ggml_backend_dev_t dev, const struct ggml_tensor * op);
 
         // (optional) event synchronization

ggml/src/ggml-backend-reg.cpp

@@ -0,0 +1,195 @@
+#include "ggml-backend-impl.h"
+#include "ggml-backend.h"
+#include "ggml-cpu.h"
+#include "ggml-impl.h"
+#include <cstring>
+#include <vector>
+
+// Backend registry
+
+#ifdef GGML_USE_CUDA
+#include "ggml-cuda.h"
+#endif
+
+#ifdef GGML_USE_METAL
+#include "ggml-metal.h"
+#endif
+
+#ifdef GGML_USE_SYCL
+#include "ggml-sycl.h"
+#endif
+
+#ifdef GGML_USE_VULKAN
+#include "ggml-vulkan.h"
+#endif
+
+#ifdef GGML_USE_BLAS
+#include "ggml-blas.h"
+#endif
+
+#ifdef GGML_USE_RPC
+#include "ggml-rpc.h"
+#endif
+
+#ifdef GGML_USE_AMX
+#  include "ggml-amx.h"
+#endif
+
+#ifdef GGML_USE_CANN
+#include "ggml-cann.h"
+#endif
+
+#ifdef GGML_USE_KOMPUTE
+#include "ggml-kompute.h"
+#endif
+
+struct ggml_backend_registry {
+    std::vector<ggml_backend_reg_t> backends;
+    std::vector<ggml_backend_dev_t> devices;
+
+    ggml_backend_registry() {
+#ifdef GGML_USE_CUDA
+        register_backend(ggml_backend_cuda_reg());
+#endif
+#ifdef GGML_USE_METAL
+        register_backend(ggml_backend_metal_reg());
+#endif
+#ifdef GGML_USE_SYCL
+        register_backend(ggml_backend_sycl_reg());
+#endif
+#ifdef GGML_USE_VULKAN
+        register_backend(ggml_backend_vk_reg());
+#endif
+#ifdef GGML_USE_CANN
+        register_backend(ggml_backend_cann_reg());
+#endif
+#ifdef GGML_USE_BLAS
+        register_backend(ggml_backend_blas_reg());
+#endif
+#ifdef GGML_USE_RPC
+        register_backend(ggml_backend_rpc_reg());
+#endif
+#ifdef GGML_USE_AMX
+        register_backend(ggml_backend_amx_reg());
+#endif
+#ifdef GGML_USE_KOMPUTE
+        register_backend(ggml_backend_kompute_reg());
+#endif
+
+        register_backend(ggml_backend_cpu_reg());
+    }
+
+    void register_backend(ggml_backend_reg_t reg) {
+        if (!reg) {
+            return;
+        }
+
+#ifndef NDEBUG
+        GGML_LOG_DEBUG("%s: registered backend %s (%zu devices)\n",
+            __func__, ggml_backend_reg_name(reg), ggml_backend_reg_dev_count(reg));
+#endif
+        backends.push_back(reg);
+        for (size_t i = 0; i < ggml_backend_reg_dev_count(reg); i++) {
+            register_device(ggml_backend_reg_dev_get(reg, i));
+        }
+    }
+
+    void register_device(ggml_backend_dev_t device) {
+#ifndef NDEBUG
+        GGML_LOG_DEBUG("%s: registered device %s (%s)\n", __func__, ggml_backend_dev_name(device), ggml_backend_dev_description(device));
+#endif
+        devices.push_back(device);
+    }
+};
+
+static ggml_backend_registry & get_reg() {
+    static ggml_backend_registry reg;
+    return reg;
+}
+
+// Internal API
+void ggml_backend_register(ggml_backend_reg_t reg) {
+    get_reg().register_backend(reg);
+}
+
+void ggml_backend_device_register(ggml_backend_dev_t device) {
+    get_reg().register_device(device);
+}
+
+// Backend (reg) enumeration
+size_t ggml_backend_reg_count() {
+    return get_reg().backends.size();
+}
+
+ggml_backend_reg_t ggml_backend_reg_get(size_t index) {
+    GGML_ASSERT(index < ggml_backend_reg_count());
+    return get_reg().backends[index];
+}
+
+ggml_backend_reg_t ggml_backend_reg_by_name(const char * name) {
+    for (size_t i = 0; i < ggml_backend_reg_count(); i++) {
+        ggml_backend_reg_t reg = ggml_backend_reg_get(i);
+        if (std::strcmp(ggml_backend_reg_name(reg), name) == 0) {
+            return reg;
+        }
+    }
+    return NULL;
+}
+
+// Device enumeration
+size_t ggml_backend_dev_count() {
+    return get_reg().devices.size();
+}
+
+ggml_backend_dev_t ggml_backend_dev_get(size_t index) {
+    GGML_ASSERT(index < ggml_backend_dev_count());
+    return get_reg().devices[index];
+}
+
+ggml_backend_dev_t ggml_backend_dev_by_name(const char * name) {
+    for (size_t i = 0; i < ggml_backend_dev_count(); i++) {
+        ggml_backend_dev_t dev = ggml_backend_dev_get(i);
+        if (strcmp(ggml_backend_dev_name(dev), name) == 0) {
+            return dev;
+        }
+    }
+    return NULL;
+}
+
+ggml_backend_dev_t ggml_backend_dev_by_type(enum ggml_backend_dev_type type) {
+    for (size_t i = 0; i < ggml_backend_dev_count(); i++) {
+        ggml_backend_dev_t dev = ggml_backend_dev_get(i);
+        if (ggml_backend_dev_type(dev) == type) {
+            return dev;
+        }
+    }
+    return NULL;
+}
+
+// Convenience functions
+ggml_backend_t ggml_backend_init_by_name(const char * name, const char * params) {
+    ggml_backend_dev_t dev = ggml_backend_dev_by_name(name);
+    if (!dev) {
+        return NULL;
+    }
+    return ggml_backend_dev_init(dev, params);
+}
+
+ggml_backend_t ggml_backend_init_by_type(enum ggml_backend_dev_type type, const char * params) {
+    ggml_backend_dev_t dev = ggml_backend_dev_by_type(type);
+    if (!dev) {
+        return NULL;
+    }
+    return ggml_backend_dev_init(dev, params);
+}
+
+ggml_backend_t ggml_backend_init_best(void) {
+    ggml_backend_dev_t dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_GPU);
+    if (!dev) {
+        dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU);
+    }
+    if (!dev) {
+        return NULL;
+    }
+    return ggml_backend_dev_init(dev, NULL);
+}

ggml/src/ggml-blas/CMakeLists.txt

@@ -0,0 +1,90 @@
+if (GGML_STATIC)
+    set(BLA_STATIC ON)
+endif()
+#if (CMAKE_VERSION VERSION_GREATER_EQUAL 3.22)
+#    set(BLA_SIZEOF_INTEGER 8)
+#endif()
+
+set(BLA_VENDOR ${GGML_BLAS_VENDOR})
+find_package(BLAS)
+
+if (BLAS_FOUND)
+    message(STATUS "BLAS found, Libraries: ${BLAS_LIBRARIES}")
+
+    add_library(ggml-blas
+                ggml-blas.cpp
+                )
+
+    target_link_libraries(ggml-blas PRIVATE ggml-base)
+    target_include_directories(ggml-blas PRIVATE . ..)
+
+    if (${GGML_BLAS_VENDOR} MATCHES "Apple")
+        add_compile_definitions(ACCELERATE_NEW_LAPACK)
+        add_compile_definitions(ACCELERATE_LAPACK_ILP64)
+        add_compile_definitions(GGML_BLAS_USE_ACCELERATE)
+    elseif ("${BLAS_INCLUDE_DIRS}" STREQUAL "")
+        # BLAS_INCLUDE_DIRS is missing in FindBLAS.cmake.
+        # see https://gitlab.kitware.com/cmake/cmake/-/issues/20268
+        find_package(PkgConfig REQUIRED)
+        if (${GGML_BLAS_VENDOR} MATCHES "Generic")
+            pkg_check_modules(DepBLAS blas)
+        elseif (${GGML_BLAS_VENDOR} MATCHES "OpenBLAS")
+            # As of openblas v0.3.22, the 64-bit is named openblas64.pc
+            pkg_check_modules(DepBLAS openblas64)
+            if (NOT DepBLAS_FOUND)
+                pkg_check_modules(DepBLAS openblas)
+            endif()
+        elseif (${GGML_BLAS_VENDOR} MATCHES "FLAME")
+            add_compile_definitions(GGML_BLAS_USE_BLIS)
+            pkg_check_modules(DepBLAS blis)
+        elseif (${GGML_BLAS_VENDOR} MATCHES "ATLAS")
+            pkg_check_modules(DepBLAS blas-atlas)
+        elseif (${GGML_BLAS_VENDOR} MATCHES "FlexiBLAS")
+            pkg_check_modules(DepBLAS flexiblas_api)
+        elseif (${GGML_BLAS_VENDOR} MATCHES "Intel")
+            add_compile_definitions(GGML_BLAS_USE_MKL)
+            # all Intel* libraries share the same include path
+            pkg_check_modules(DepBLAS mkl-sdl)
+        elseif (${GGML_BLAS_VENDOR} MATCHES "NVHPC")
+            # this doesn't provide pkg-config
+            # suggest to assign BLAS_INCLUDE_DIRS on your own
+            if ("${NVHPC_VERSION}" STREQUAL "")
+                message(WARNING "Better to set NVHPC_VERSION")
+            else()
+                set(DepBLAS_FOUND ON)
+                set(DepBLAS_INCLUDE_DIRS "/opt/nvidia/hpc_sdk/${CMAKE_SYSTEM_NAME}_${CMAKE_SYSTEM_PROCESSOR}/${NVHPC_VERSION}/math_libs/include")
+            endif()
+        endif()
+        if (DepBLAS_FOUND)
+            set(BLAS_INCLUDE_DIRS ${DepBLAS_INCLUDE_DIRS})
+        else()
+            message(WARNING "BLAS_INCLUDE_DIRS neither been provided nor been automatically"
+            " detected by pkgconfig, trying to find cblas.h from possible paths...")
+            find_path(BLAS_INCLUDE_DIRS
+                NAMES cblas.h
+                HINTS
+                    /usr/include
+                    /usr/local/include
+                    /usr/include/openblas
+                    /opt/homebrew/opt/openblas/include
+                    /usr/local/opt/openblas/include
+                    /usr/include/x86_64-linux-gnu/openblas/include
+            )
+        endif()
+    endif()
+
+    message(STATUS "BLAS found, Includes: ${BLAS_INCLUDE_DIRS}")
+
+    target_compile_options(ggml-blas PRIVATE ${BLAS_LINKER_FLAGS})
+
+    if (${BLAS_INCLUDE_DIRS} MATCHES "mkl" AND (${GGML_BLAS_VENDOR} MATCHES "Generic" OR ${GGML_BLAS_VENDOR} MATCHES "Intel"))
+        add_compile_definitions(GGML_BLAS_USE_MKL)
+    endif()
+
+    target_link_libraries     (ggml-blas PRIVATE ${BLAS_LIBRARIES})
+    target_include_directories(ggml-blas PRIVATE ${BLAS_INCLUDE_DIRS})
+else()
+    message(ERROR "BLAS not found, please refer to "
+                  "https://cmake.org/cmake/help/latest/module/FindBLAS.html#blas-lapack-vendors"
+                  " to set correct GGML_BLAS_VENDOR")
+endif()

ggml/src/ggml-blas/ggml-blas.cpp

@@ -4,8 +4,9 @@
 
 #include <future>
 #include <vector>
+#include <cstring>
 
-#if defined(GGML_USE_ACCELERATE)
+#if defined(GGML_BLAS_USE_ACCELERATE)
 #   include <Accelerate/Accelerate.h>
 #elif defined(GGML_BLAS_USE_MKL)
 #   include <mkl.h>
@@ -26,30 +27,6 @@ struct ggml_backend_blas_context {
 #endif
 };
 
-// helper function to determine if it is better to use BLAS or not
-// for large matrices, BLAS is faster
-static bool ggml_backend_blas_use_blas(const struct ggml_tensor * dst) {
-    const struct ggml_tensor * src0 = dst->src[0];
-    const struct ggml_tensor * src1 = dst->src[1];
-
-    const int64_t ne10 = src1->ne[0];
-
-    const int64_t ne0 = dst->ne[0];
-    const int64_t ne1 = dst->ne[1];
-
-    // TODO: find the optimal values for these
-    if (ggml_is_contiguous(src0) &&
-        ggml_is_contiguous(src1) &&
-        src1->type == GGML_TYPE_F32 &&
-        (ne0 >= 32 && ne1 >= 32 && ne10 >= 32)) {
-
-        /*printf("BLAS: %d %d %d %d %d\n", ne0, ne1, ne10, ne00, ne01);*/
-        return true;
-    }
-
-    return false;
-}
-
 static void ggml_backend_blas_mul_mat(ggml_backend_blas_context * ctx, struct ggml_tensor * dst) {
     const struct ggml_tensor * src0 = dst->src[0];
     const struct ggml_tensor * src1 = dst->src[1];
@@ -88,8 +65,8 @@ static void ggml_backend_blas_mul_mat(ggml_backend_blas_context * ctx, struct gg
 
     // convert src0 to float
     if (type != GGML_TYPE_F32) {
-        ggml_type_traits_t type_traits = ggml_internal_get_type_traits(type);
-        ggml_to_float_t const to_float = type_traits.to_float;
+        const auto * type_traits = ggml_get_type_traits(type);
+        ggml_to_float_t const to_float = type_traits->to_float;
 
         for (int64_t i03 = 0; i03 < ne03; i03++) {
             for (int64_t i02 = 0; i02 < ne02; i02++) {
@@ -235,7 +212,7 @@ static void ggml_backend_blas_out_prod(ggml_backend_blas_context * ctx, struct g
 
 // backend interface
 
-static const char * ggml_backend_blas_name(ggml_backend_t backend) {
+static const char * ggml_backend_blas_get_name(ggml_backend_t backend) {
     return "BLAS";
 
     GGML_UNUSED(backend);
@@ -247,12 +224,6 @@ static void ggml_backend_blas_free(ggml_backend_t backend) {
     delete backend;
 }
 
-static ggml_backend_buffer_type_t ggml_backend_blas_get_default_buffer_type(ggml_backend_t backend) {
-    return ggml_backend_cpu_buffer_type();
-
-    GGML_UNUSED(backend);
-}
-
 static enum ggml_status ggml_backend_blas_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
     ggml_backend_blas_context * ctx = (ggml_backend_blas_context *)backend->context;
 
@@ -285,31 +256,9 @@ static enum ggml_status ggml_backend_blas_graph_compute(ggml_backend_t backend,
     GGML_UNUSED(backend);
 }
 
-static bool ggml_backend_blas_supports_op(ggml_backend_t backend, const struct ggml_tensor * op) {
-    const struct ggml_tensor * src0 = op->src[0];
-    const struct ggml_tensor * src1 = op->src[1];
-
-    return (op->op == GGML_OP_MUL_MAT  && ggml_backend_blas_use_blas(op)) ||
-           (op->op == GGML_OP_OUT_PROD && op->src[0]->type == GGML_TYPE_F32 &&
-                                          op->src[1]->type == GGML_TYPE_F32 &&
-                                          ggml_is_matrix(src0) &&
-                                          ggml_is_matrix(src1) &&
-                                          ggml_is_contiguous(src0) &&
-                                          (ggml_is_contiguous(src1) || ggml_is_transposed(src1)));
-
-    GGML_UNUSED(backend);
-}
-
-static bool ggml_backend_blas_supports_buft(ggml_backend_t backend, ggml_backend_buffer_type_t buft) {
-    return ggml_backend_buft_is_host(buft);
-
-    GGML_UNUSED(backend);
-}
-
 static struct ggml_backend_i blas_backend_i = {
-    /* .get_name                = */ ggml_backend_blas_name,
+    /* .get_name                = */ ggml_backend_blas_get_name,
     /* .free                    = */ ggml_backend_blas_free,
-    /* .get_default_buffer_type = */ ggml_backend_blas_get_default_buffer_type,
     /* .set_tensor_async        = */ NULL,
     /* .get_tensor_async        = */ NULL,
     /* .cpy_tensor_async        = */ NULL,
@@ -319,9 +268,6 @@ static struct ggml_backend_i blas_backend_i = {
     /* .graph_plan_update       = */ NULL,
     /* .graph_plan_compute      = */ NULL,
     /* .graph_compute           = */ ggml_backend_blas_graph_compute,
-    /* .supports_op             = */ ggml_backend_blas_supports_op,
-    /* .supports_buft           = */ ggml_backend_blas_supports_buft,
-    /* .offload_op              = */ NULL,
     /* .event_record            = */ NULL,
     /* .event_wait              = */ NULL,
 };
@@ -337,18 +283,18 @@ ggml_backend_t ggml_backend_blas_init(void) {
     ggml_backend_t backend = new ggml_backend {
         /* .guid      = */ ggml_backend_blas_guid(),
         /* .interface = */ blas_backend_i,
-        /* .device    = */ nullptr,
+        /* .device    = */ ggml_backend_reg_dev_get(ggml_backend_blas_reg(), 0),
         /* .context   = */ ctx,
     };
 
-#if !defined(NDEBUG) && defined(OPENBLAS_VERSION) && defined(GGML_USE_OPENMP)
+#if defined(OPENBLAS_VERSION) && defined(GGML_USE_OPENMP)
     if (openblas_get_parallel() != OPENBLAS_OPENMP) {
-        fprintf(stderr, "%s: warning: ggml is using OpenMP, but OpenBLAS was compiled without OpenMP support\n", __func__);
+        GGML_LOG_DEBUG("%s: warning: ggml is using OpenMP, but OpenBLAS was compiled without OpenMP support\n", __func__);
     }
 #endif
 
-#if !defined(NDEBUG) && defined(BLIS_ENABLE_CBLAS) && defined(GGML_USE_OPENMP) && !defined(BLIS_ENABLE_OPENMP)
-    fprintf(stderr, "%s: warning: ggml is using OpenMP, but BLIS was compiled without OpenMP support\n", __func__);
+#if defined(BLIS_ENABLE_CBLAS) && defined(GGML_USE_OPENMP) && !defined(BLIS_ENABLE_OPENMP)
+    GGML_LOG_DEBUG("%s: warning: ggml is using OpenMP, but BLIS was compiled without OpenMP support\n", __func__);
 #endif
 
     return backend;
@@ -364,3 +310,205 @@ void ggml_backend_blas_set_n_threads(ggml_backend_t backend_blas, int n_threads)
     ggml_backend_blas_context * ctx = (ggml_backend_blas_context *)backend_blas->context;
     ctx->n_threads = n_threads;
 }
+
+// device interface
+
+static const char * ggml_backend_blas_device_get_name(ggml_backend_dev_t dev) {
+    return "BLAS";
+
+    GGML_UNUSED(dev);
+}
+
+static const char * ggml_backend_blas_device_get_description(ggml_backend_dev_t dev) {
+    #if defined(GGML_BLAS_USE_ACCELERATE)
+        return "Accelerate";
+    #elif defined(GGML_BLAS_USE_MKL)
+        return "MKL";
+    #elif defined(GGML_BLAS_USE_BLIS)
+        return "BLIS";
+    #elif defined(GGML_BLAS_USE_NVPL)
+        return "NVPL";
+    #elif defined(OPENBLAS_VERSION)
+        return "OpenBLAS";
+    #else
+        return "BLAS";
+    #endif
+
+    GGML_UNUSED(dev);
+}
+
+static void ggml_backend_blas_device_get_memory(ggml_backend_dev_t dev, size_t * free, size_t * total) {
+    // TODO
+    *free = 0;
+    *total = 0;
+
+    GGML_UNUSED(dev);
+}
+
+static enum ggml_backend_dev_type ggml_backend_blas_device_get_type(ggml_backend_dev_t dev) {
+    return GGML_BACKEND_DEVICE_TYPE_ACCEL;
+
+    GGML_UNUSED(dev);
+}
+
+static void ggml_backend_blas_device_get_props(ggml_backend_dev_t dev, struct ggml_backend_dev_props * props) {
+    props->name        = ggml_backend_blas_device_get_name(dev);
+    props->description = ggml_backend_blas_device_get_description(dev);
+    props->type        = ggml_backend_blas_device_get_type(dev);
+    ggml_backend_blas_device_get_memory(dev, &props->memory_free, &props->memory_total);
+    props->caps = {
+        /* .async                 = */ false,
+        /* .host_buffer           = */ false,
+        /* .buffer_from_host_ptr  = */ true,
+        /* .events                = */ false,
+    };
+}
+
+static ggml_backend_t ggml_backend_blas_device_init_backend(ggml_backend_dev_t dev, const char * params) {
+    return ggml_backend_blas_init();
+
+    GGML_UNUSED(dev);
+    GGML_UNUSED(params);
+}
+
+static ggml_backend_buffer_type_t ggml_backend_blas_device_get_buffer_type(ggml_backend_dev_t dev) {
+    return ggml_backend_cpu_buffer_type();
+
+    GGML_UNUSED(dev);
+}
+
+static ggml_backend_buffer_t ggml_backend_blas_device_buffer_from_host_ptr(ggml_backend_dev_t dev, void * ptr, size_t size, size_t max_tensor_size) {
+    return ggml_backend_cpu_buffer_from_ptr(ptr, size);
+
+    GGML_UNUSED(dev);
+    GGML_UNUSED(max_tensor_size);
+}
+
+static bool ggml_backend_blas_device_supports_op(ggml_backend_dev_t dev, const struct ggml_tensor * op) {
+    const struct ggml_tensor * src0 = op->src[0];
+    const struct ggml_tensor * src1 = op->src[1];
+
+    switch (op->op) {
+        case GGML_OP_NONE:
+        case GGML_OP_RESHAPE:
+        case GGML_OP_VIEW:
+        case GGML_OP_PERMUTE:
+        case GGML_OP_TRANSPOSE:
+            return true;
+
+        case GGML_OP_MUL_MAT:
+        {
+            // BLAS usually is only faster for large matrices
+            const struct ggml_tensor * src0 = op->src[0];
+            const struct ggml_tensor * src1 = op->src[1];
+
+            const int64_t ne10 = src1->ne[0];
+
+            const int64_t ne0 = op->ne[0];
+            const int64_t ne1 = op->ne[1];
+
+            // TODO: find the optimal value
+            const int64_t min_batch = 32;
+
+            return ggml_is_contiguous(src0) &&
+                   ggml_is_contiguous(src1) &&
+                   src1->type == GGML_TYPE_F32 &&
+                   (ne0 >= min_batch && ne1 >= min_batch && ne10 >= min_batch) &&
+                   (src0->type == GGML_TYPE_F32 || ggml_get_type_traits(src0->type)->to_float != NULL);
+        }
+
+        case GGML_OP_OUT_PROD:
+            return op->src[0]->type == GGML_TYPE_F32 &&
+                   op->src[1]->type == GGML_TYPE_F32 &&
+                   ggml_is_matrix(src0) &&
+                   ggml_is_matrix(src1) &&
+                   ggml_is_contiguous(src0) &&
+                   (ggml_is_contiguous(src1) || ggml_is_transposed(src1)) &&
+                   (src0->type == GGML_TYPE_F32 || ggml_get_type_traits(src0->type)->to_float != NULL);
+
+        default:
+            return false;
+
+    }
+
+    GGML_UNUSED(dev);
+}
+
+static bool ggml_backend_blas_device_supports_buft(ggml_backend_dev_t dev, ggml_backend_buffer_type_t buft) {
+    return ggml_backend_buft_is_host(buft);
+
+    GGML_UNUSED(dev);
+}
+
+static const struct ggml_backend_device_i ggml_backend_blas_device_i = {
+    /* .get_name             = */ ggml_backend_blas_device_get_name,
+    /* .get_description      = */ ggml_backend_blas_device_get_description,
+    /* .get_memory           = */ ggml_backend_blas_device_get_memory,
+    /* .get_type             = */ ggml_backend_blas_device_get_type,
+    /* .get_props            = */ ggml_backend_blas_device_get_props,
+    /* .init_backend         = */ ggml_backend_blas_device_init_backend,
+    /* .get_buffer_type      = */ ggml_backend_blas_device_get_buffer_type,
+    /* .get_host_buffer_type = */ NULL,
+    /* .buffer_from_host_ptr = */ ggml_backend_blas_device_buffer_from_host_ptr,
+    /* .supports_op          = */ ggml_backend_blas_device_supports_op,
+    /* .supports_buft        = */ ggml_backend_blas_device_supports_buft,
+    /* .offload_op           = */ NULL,
+    /* .event_new            = */ NULL,
+    /* .event_free           = */ NULL,
+    /* .event_synchronize    = */ NULL,
+};
+
+// backend reg interface
+
+static const char * ggml_backend_blas_reg_get_name(ggml_backend_reg_t reg) {
+    return "BLAS";
+
+    GGML_UNUSED(reg);
+}
+
+static size_t ggml_backend_blas_reg_get_device_count(ggml_backend_reg_t reg) {
+    return 1;
+
+    GGML_UNUSED(reg);
+}
+
+static ggml_backend_dev_t ggml_backend_blas_reg_get_device(ggml_backend_reg_t reg, size_t index) {
+    GGML_ASSERT(index == 0);
+
+    static ggml_backend_device ggml_backend_blas_device = {
+        /* .iface   = */ ggml_backend_blas_device_i,
+        /* .reg     = */ reg,
+        /* .context = */ nullptr,
+    };
+
+    return &ggml_backend_blas_device;
+
+    GGML_UNUSED(reg);
+    GGML_UNUSED(index);
+}
+
+static void * ggml_backend_blas_get_proc_address(ggml_backend_reg_t reg, const char * name) {
+    if (std::strcmp(name, "ggml_backend_set_n_threads") == 0) {
+        return (void *)ggml_backend_blas_set_n_threads;
+    }
+    return NULL;
+
+    GGML_UNUSED(reg);
+    GGML_UNUSED(name);
+}
+
+static const struct ggml_backend_reg_i ggml_backend_blas_reg_i = {
+    /* .get_name         = */ ggml_backend_blas_reg_get_name,
+    /* .get_device_count = */ ggml_backend_blas_reg_get_device_count,
+    /* .get_device       = */ ggml_backend_blas_reg_get_device,
+    /* .get_proc_address = */ ggml_backend_blas_get_proc_address,
+};
+
+ggml_backend_reg_t ggml_backend_blas_reg(void) {
+    static struct ggml_backend_reg ggml_backend_blas_reg = {
+        /* .iface   = */ ggml_backend_blas_reg_i,
+        /* .context = */ NULL,
+    };
+
+    return &ggml_backend_blas_reg;
+}

ggml/src/ggml-cann/CMakeLists.txt

@@ -0,0 +1,46 @@
+if ("cann${CANN_INSTALL_DIR}" STREQUAL "cann" AND DEFINED ENV{ASCEND_TOOLKIT_HOME})
+    set(CANN_INSTALL_DIR $ENV{ASCEND_TOOLKIT_HOME})
+    message(STATUS "CANN: updated CANN_INSTALL_DIR from ASCEND_TOOLKIT_HOME=$ENV{ASCEND_TOOLKIT_HOME}")
+endif()
+
+if (CANN_INSTALL_DIR)
+    # Only Support Linux.
+    if (NOT UNIX)
+        message(FATAL_ERROR "CANN: CANN toolkit supports unix but not ${CMAKE_SYSTEM_NAME}")
+    endif()
+
+    # Supported platforms: x86-64, arm64
+    if (CMAKE_SYSTEM_PROCESSOR STREQUAL "aarch64")
+    elseif (CMAKE_SYSTEM_PROCESSOR STREQUAL "x86_64" OR CMAKE_SYSTEM_PROCESSOR STREQUAL "amd64")
+    else()
+        message(FATAL_ERROR "CANN: CANN toolkit supports x86-64 and arm64 but not ${CMAKE_SYSTEM_PROCESSOR}")
+    endif()
+
+    # Set header and libs
+    set(CANN_INCLUDE_DIRS
+        ${CANN_INSTALL_DIR}/include
+        ${CANN_INSTALL_DIR}/include/aclnn
+        ${CANN_INSTALL_DIR}/acllib/include
+    )
+
+    add_subdirectory(kernels)
+    list(APPEND CANN_LIBRARIES
+        ascendcl
+        nnopbase
+        opapi
+        acl_op_compiler
+        ascendc_kernels
+    )
+
+    file(GLOB GGML_SOURCES_CANN "*.cpp")
+
+    add_library(ggml-cann ${GGML_SOURCES_CANN})
+    target_link_libraries(ggml-cann PRIVATE ggml-base ${CANN_LIBRARIES})
+    target_include_directories(ggml-cann PRIVATE . .. ${CANN_INCLUDE_DIRS})
+    target_link_directories(ggml-cann PRIVATE ${CANN_INSTALL_DIR}/lib64)
+
+    message(STATUS "CANN: CANN_INCLUDE_DIRS =  ${CANN_INCLUDE_DIRS}")
+    message(STATUS "CANN: CANN_LIBRARIES =  ${CANN_LIBRARIES}")
+else()
+    message(FATAL_ERROR "CANN: Can't find CANN_INSTALL_DIR, did you forget to source set_var.sh?")
+endif()

ggml/src/ggml-cann/ggml-cann.cpp

@@ -39,6 +39,8 @@
 
 #include "ggml-common.h"
 
+#define GGML_CANN_NAME "CANN"
+
 /**
  * @brief Handles CANN errors by printing an error message and aborting.
  *
@@ -487,23 +489,6 @@ struct ggml_backend_cann_buffer_context {
     ~ggml_backend_cann_buffer_context() { ACL_CHECK(aclrtFree(dev_ptr)); }
 };
 
-/**
- * @brief Retrieve the name associated with a CANN buffer.
- *
- * This function returns the name of a CANN buffer, which is stored in the
- * context of the buffer.
- *
- * @param buffer The CANN buffer whose name is to be retrieved.
- * @return A pointer to a C-string containing the name of the buffer.
- */
-
-static const char* ggml_backend_cann_buffer_get_name(
-    ggml_backend_buffer_t buffer) {
-    return "CANN";
-
-    GGML_UNUSED(buffer);
-}
-
 /**
  * @brief Check if a buffer is a CANN buffer.
  *
@@ -513,9 +498,10 @@ static const char* ggml_backend_cann_buffer_get_name(
  * @param buffer The buffer to check.
  * @return true if the buffer is a CANN buffer, false otherwise.
  */
+static bool ggml_backend_buft_is_cann(ggml_backend_buffer_type_t buft);
 static bool ggml_backend_buffer_is_cann(
     ggml_backend_buffer_t buffer) {
-    return buffer->iface.get_name == ggml_backend_cann_buffer_get_name;
+    return ggml_backend_buft_is_cann(buffer->buft);
 }
 
 /**
@@ -851,13 +837,6 @@ static void ggml_backend_cann_buffer_set_tensor(
         void *transform_buffer = malloc(size);
         ggml_backend_cann_transform(tensor, data, transform_buffer);
 
-#ifndef NDEBUG
-        void *check_buffer = malloc(size);
-        ggml_backend_cann_transform_back(tensor, transform_buffer,
-                                         check_buffer);
-        GGML_ASSERT(memcmp(data, check_buffer, size) == 0);
-        free(check_buffer);
-#endif
         ACL_CHECK(aclrtMemcpy((char *)tensor->data + offset, size,
                               transform_buffer, size,
                               ACL_MEMCPY_HOST_TO_DEVICE));
@@ -969,8 +948,7 @@ static void ggml_backend_cann_buffer_clear(
  * This structure defines function pointers to operations that can be performed
  * on a CANN buffer within the backend.
  */
-static ggml_backend_buffer_i ggml_backend_cann_buffer_interface = {
-    /* .get_name        = */ ggml_backend_cann_buffer_get_name,
+static const ggml_backend_buffer_i ggml_backend_cann_buffer_interface = {
     /* .free_buffer     = */ ggml_backend_cann_buffer_free_buffer,
     /* .get_base        = */ ggml_backend_cann_buffer_get_base,
     /* .init_tensor     = */ ggml_backend_cann_buffer_init_tensor,
@@ -1004,9 +982,10 @@ struct ggml_backend_cann_buffer_type_context {
  */
 static const char* ggml_backend_cann_buffer_type_name(
     ggml_backend_buffer_type_t buft) {
-    return "CANN";
+    ggml_backend_cann_buffer_type_context* buft_ctx =
+        (ggml_backend_cann_buffer_type_context*)buft->context;
 
-    GGML_UNUSED(buft);
+    return buft_ctx->name.c_str();
 }
 
 /**
@@ -1105,19 +1084,25 @@ static size_t ggml_backend_cann_buffer_type_get_alloc_size(
     GGML_UNUSED(buft);
 }
 
+static bool ggml_backend_cann_buffer_type_is_host(ggml_backend_buffer_type_t buft) {
+    return false;
+
+    GGML_UNUSED(buft);
+}
+
 /**
  * @brief Interface for managing CANN buffer types in the GGML backend.
  *
  * Provides function pointers for allocating, querying properties, and managing
  * memory for CANN buffer types in the GGML backend.
  */
-static ggml_backend_buffer_type_i ggml_backend_cann_buffer_type_interface = {
+static const ggml_backend_buffer_type_i ggml_backend_cann_buffer_type_interface = {
     /* .get_name         = */ ggml_backend_cann_buffer_type_name,
     /* .alloc_buffer     = */ ggml_backend_cann_buffer_type_alloc_buffer,
     /* .get_alignment    = */ ggml_backend_cann_buffer_type_get_alignment,
     /* .get_max_size     = */ NULL,  // defaults to SIZE_MAX
     /* .get_alloc_size   = */ ggml_backend_cann_buffer_type_get_alloc_size,
-    /* .is_host          = */ NULL,
+    /* .is_host          = */ ggml_backend_cann_buffer_type_is_host,
 };
 
 /**
@@ -1148,6 +1133,7 @@ ggml_backend_cann_buffer_type(int32_t device) {
         for (int32_t i = 0; i < GGML_CANN_MAX_DEVICES; i++) {
             ggml_backend_cann_buffer_types[i] = {
                 /* .iface    = */ ggml_backend_cann_buffer_type_interface,
+                /* .device    = */ ggml_backend_reg_dev_get(ggml_backend_cann_reg(), device),
                 /* .context  = */
                  new ggml_backend_cann_buffer_type_context{
                     i, "CANN" + std::to_string(i)},
@@ -1241,7 +1227,6 @@ static ggml_backend_buffer_t ggml_backend_cann_host_buffer_type_alloc_buffer(ggm
 
     ggml_backend_buffer_t buffer = ggml_backend_cpu_buffer_from_ptr(hostPtr, size);
     buffer->buft = buft;
-    buffer->iface.get_name = ggml_backend_cann_host_buffer_name;
     buffer->iface.free_buffer = ggml_backend_cann_host_buffer_free;
 
     return buffer;
@@ -1263,7 +1248,7 @@ ggml_backend_buffer_type_t ggml_backend_cann_host_buffer_type() {
             /* .get_alloc_size   = */ ggml_backend_cpu_buffer_type()->iface.get_alloc_size,
             /* .is_host          = */ ggml_backend_cpu_buffer_type()->iface.is_host,
         },
-        /* .device   = */ nullptr,
+        /* .device   = */ ggml_backend_reg_dev_get(ggml_backend_cann_reg(), 0),
         /* .context  = */ nullptr,
     };
 
@@ -1463,24 +1448,6 @@ static void ggml_backend_cann_free(ggml_backend_t backend) {
     delete backend;
 }
 
-/**
- * @brief Retrieves the default buffer type associated with the CANN backend.
- *
- * This function returns the buffer type specific to the device associated
- * with the CANN backend. It is used to allocate buffers for computations
- * performed by the backend.
- *
- * @param backend Pointer to the CANN backend structure.
- * @return Pointer to the buffer type structure for the CANN backend.
- */
-static ggml_backend_buffer_type_t
-ggml_backend_cann_get_default_buffer_type(ggml_backend_t backend) {
-    ggml_backend_cann_context* cann_ctx =
-        (ggml_backend_cann_context*)backend->context;
-
-    return ggml_backend_cann_buffer_type(cann_ctx->device);
-}
-
 /**
  * @brief Sets tensor data asynchronously in the CANN backend.
  *
@@ -1510,13 +1477,6 @@ static void ggml_backend_cann_set_tensor_async(ggml_backend_t backend,
         void *transform_buffer = malloc(size);
         ggml_backend_cann_transform(tensor, data, transform_buffer);
 
-#ifndef NDEBUG
-        void *check_buffer = malloc(size);
-        ggml_backend_cann_transform_back(tensor, transform_buffer,
-                                         check_buffer);
-        GGML_ASSERT(memcmp(data, check_buffer, size));
-        free(check_buffer);
-#endif
         ACL_CHECK(aclrtMemcpyAsync(
             (char *)tensor->data + offset, size, transform_buffer, size,
             ACL_MEMCPY_HOST_TO_DEVICE, cann_ctx->stream()));
@@ -1691,7 +1651,7 @@ static enum ggml_status ggml_backend_cann_graph_compute(
  * @return bool Returns true if the operation is supported by the backend,
  *              otherwise false.
  */
-static bool ggml_backend_cann_supports_op(ggml_backend_t backend,
+static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
                                                     const ggml_tensor* op) {
     switch (op->op) {
         case GGML_OP_UNARY:
@@ -1782,7 +1742,7 @@ static bool ggml_backend_cann_supports_op(ggml_backend_t backend,
             return false;
     }
 
-    GGML_UNUSED(backend);
+    GGML_UNUSED(dev);
 }
 
 /**
@@ -1800,31 +1760,6 @@ static bool ggml_backend_buft_is_cann(ggml_backend_buffer_type_t buft) {
     return buft->iface.get_name == ggml_backend_cann_buffer_type_name;
 }
 
-/**
- * @brief Checks if the CANN backend supports a specific backend buffer type.
- *
- * This function determines whether the CANN backend supports the given backend
- * buffer type by comparing the device context of the backend and buffer type.
- * It returns true if the devices are same between the backend context and
- * buffer type context.
- *
- * @param backend Pointer to the CANN backend.
- * @param buft Pointer to the backend buffer type to check.
- * @return bool Returns true if the CANN backend supports the buffer type,
- *              otherwise false.
- */
-static bool ggml_backend_cann_supports_buft(
-    ggml_backend_t backend, ggml_backend_buffer_type_t buft) {
-    if (ggml_backend_buft_is_cann(buft)) {
-        ggml_backend_cann_context * cann_ctx =
-                        (ggml_backend_cann_context *)backend->context;
-        ggml_backend_cann_buffer_type_context * buft_ctx =
-                        (ggml_backend_cann_buffer_type_context *)buft->context;
-        return buft_ctx->device == cann_ctx->device;
-    }
-    return false;
-}
-
 /**
  * @brief Determines if a tensor operation should be offloaded to the CANN
  * backend.
@@ -1839,54 +1774,14 @@ static bool ggml_backend_cann_supports_buft(
  * @return bool Returns true if the operation should be offloaded, otherwise
  * false.
  */
-static bool ggml_backend_cann_offload_op(ggml_backend_t backend,
+static bool ggml_backend_cann_offload_op(ggml_backend_dev_t dev,
                                                    const ggml_tensor* op) {
     const int min_batch_size = 32;
-    GGML_UNUSED(backend);
+    GGML_UNUSED(dev);
 
     return op->ne[1] >= min_batch_size && op->op != GGML_OP_GET_ROWS;
 }
 
-/**
- * @brief Creates a new event for the CANN backend.
- *
- * This function initializes a new event for the CANN backend by setting the
- * device and creating an ACL runtime event. The created event is then wrapped
- * in a ggml_backend_event structure and returned.
- *
- * @param backend Pointer to the CANN backend.
- * @return ggml_backend_event_t Returns a pointer to the new event structure.
- */
-static ggml_backend_event_t ggml_backend_cann_event_new(
-    ggml_backend_t backend) {
-    ggml_backend_cann_context* cann_ctx =
-        (ggml_backend_cann_context*)backend->context;
-
-    ggml_cann_set_device(cann_ctx->device);
-
-    aclrtEvent event;
-    ACL_CHECK(aclrtCreateEvent(&event));
-
-    return new ggml_backend_event{
-        /* .backend = */ backend,
-        /* .context = */ event,
-    };
-}
-
-/**
- * @brief Frees a CANN backend event.
- *
- * This function destroys the ACL runtime event associated with the given CANN
- * backend event and then deletes the event structure itself.
- *
- * @param event Pointer to the event structure to be freed.
- */
-static void ggml_backend_cann_event_free(ggml_backend_event_t event) {
-    ACL_CHECK(aclrtDestroyEvent((aclrtEvent)event->context));
-
-    delete event;
-}
-
 /**
  * @brief Records an event on the CANN backend stream.
  *
@@ -1895,10 +1790,9 @@ static void ggml_backend_cann_event_free(ggml_backend_event_t event) {
  *
  * @param event Pointer to the event structure to be recorded.
  */
-static void ggml_backend_cann_event_record(ggml_backend_event_t event) {
+static void ggml_backend_cann_event_record(ggml_backend_t backend, ggml_backend_event_t event) {
     ggml_backend_cann_context* cann_ctx =
-        (ggml_backend_cann_context*)event->backend->context;
-
+        (ggml_backend_cann_context*)backend->context;
     ACL_CHECK(aclrtRecordEvent((aclrtEvent)event->context, cann_ctx->stream()));
 }
 
@@ -1916,8 +1810,7 @@ static void ggml_backend_cann_event_wait(ggml_backend_t backend,
                                          ggml_backend_event_t event) {
     ggml_backend_cann_context* cann_ctx =
         (ggml_backend_cann_context*)backend->context;
-
-    if (ggml_backend_is_cann(event->backend)) {
+    if (ggml_backend_is_cann(backend)) {
         ACL_CHECK(aclrtStreamWaitEvent(cann_ctx->stream(),
                                        (aclrtEvent)event->context));
     } else {
@@ -1925,17 +1818,6 @@ static void ggml_backend_cann_event_wait(ggml_backend_t backend,
     }
 }
 
-/**
- * @brief Synchronizes the given event on the CANN backend.
- *
- * This function waits for the specified event to complete on the ACL runtime.
- *
- * @param event Pointer to the event structure to be synchronized.
- */
-static void ggml_backend_cann_event_synchronize(ggml_backend_event_t event) {
-    ACL_CHECK(aclrtSynchronizeEvent((aclrtEvent)event->context));
-}
-
 /**
  * @brief Structure defining the interface for the CANN backend.
  *
@@ -1943,10 +1825,9 @@ static void ggml_backend_cann_event_synchronize(ggml_backend_event_t event) {
  * supported by the CANN backend, including name retrieval, memory
  * management, tensor operations, synchronization, and event handling.
  */
-static ggml_backend_i ggml_backend_cann_interface = {
+static const ggml_backend_i ggml_backend_cann_interface = {
     /* .get_name                = */ ggml_backend_cann_name,
     /* .free                    = */ ggml_backend_cann_free,
-    /* .get_default_buffer_type = */ ggml_backend_cann_get_default_buffer_type,
     /* .set_tensor_async        = */ ggml_backend_cann_set_tensor_async,
     /* .get_tensor_async        = */ ggml_backend_cann_get_tensor_async,
     /* .cpy_tensor_async        = */ ggml_backend_cann_cpy_tensor_async,
@@ -1956,9 +1837,6 @@ static ggml_backend_i ggml_backend_cann_interface = {
     /* .graph_plan_update       = */ NULL,
     /* .graph_plan_compute      = */ NULL,
     /* .graph_compute           = */ ggml_backend_cann_graph_compute,
-    /* .supports_op             = */ ggml_backend_cann_supports_op,
-    /* .supports_buft           = */ ggml_backend_cann_supports_buft,
-    /* .offload_op              = */ ggml_backend_cann_offload_op,
     /* .event_record            = */ ggml_backend_cann_event_record,
     /* .event_wait              = */ ggml_backend_cann_event_wait,
 };
@@ -1977,6 +1855,234 @@ static ggml_guid_t ggml_backend_cann_guid() {
     return &guid;
 }
 
+// backend device
+struct ggml_backend_cann_device_context {
+    int device;
+    std::string name;
+    std::string description;
+};
+
+static const char * ggml_backend_cann_device_get_name(ggml_backend_dev_t dev) {
+    ggml_backend_cann_device_context * ctx = (ggml_backend_cann_device_context *)dev->context;
+    return ctx->name.c_str();
+}
+
+static const char* ggml_backend_cann_device_get_description(ggml_backend_dev_t dev) {
+    ggml_backend_cann_device_context * ctx = (ggml_backend_cann_device_context *)dev->context;
+    return ctx->description.c_str();
+}
+
+static void ggml_backend_cann_device_get_memory(ggml_backend_dev_t dev, size_t * free, size_t * total) {
+    ggml_backend_cann_device_context * ctx = (ggml_backend_cann_device_context *)dev->context;
+    ggml_backend_cann_get_device_memory(ctx->device, free, total);
+}
+
+static enum ggml_backend_dev_type ggml_backend_cann_device_get_type(ggml_backend_dev_t dev) {
+    GGML_UNUSED(dev);
+    return GGML_BACKEND_DEVICE_TYPE_GPU;
+}
+
+static void ggml_backend_cann_device_get_props(ggml_backend_dev_t dev, ggml_backend_dev_props * props) {
+    props->name        = ggml_backend_cann_device_get_name(dev);
+    props->description = ggml_backend_cann_device_get_description(dev);
+    props->type        = ggml_backend_cann_device_get_type(dev);
+    ggml_backend_cann_device_get_memory(dev, &props->memory_free, &props->memory_total);
+
+    bool host_buffer = getenv("GGML_CANN_NO_PINNED") == nullptr;
+
+    props->caps = {
+        /* .async                 = */ false,
+        /* .host_buffer           = */ host_buffer,
+        /* .buffer_from_host_ptr  = */ false,
+        /* .events                = */ true,
+    };
+}
+
+static ggml_backend_t ggml_backend_cann_device_init(ggml_backend_dev_t dev, const char * params) {
+    GGML_UNUSED(params);
+    ggml_backend_cann_device_context * ctx = (ggml_backend_cann_device_context *)dev->context;
+    return ggml_backend_cann_init(ctx->device);
+}
+
+/**
+ * @brief Checks if the CANN backend supports a specific backend buffer type.
+ *
+ * This function determines whether the CANN backend supports the given backend
+ * buffer type by comparing the device context of the backend and buffer type.
+ * It returns true if the devices are same between the backend context and
+ * buffer type context.
+ *
+ * @param backend Pointer to the CANN backend.
+ * @param buft Pointer to the backend buffer type to check.
+ * @return bool Returns true if the CANN backend supports the buffer type,
+ *              otherwise false.
+ */
+static bool ggml_backend_cann_supports_buft(
+    ggml_backend_dev_t dev, ggml_backend_buffer_type_t buft) {
+    if (ggml_backend_buft_is_cann(buft)) {
+        ggml_backend_cann_device_context * dev_ctx = (ggml_backend_cann_device_context *)dev->context;
+        ggml_backend_cann_buffer_type_context * buft_ctx =
+                        (ggml_backend_cann_buffer_type_context *)buft->context;
+        return buft_ctx->device == dev_ctx->device;
+    }
+    return false;
+}
+
+static ggml_backend_buffer_type_t ggml_backend_cann_device_get_buffer_type(ggml_backend_dev_t dev) {
+    ggml_backend_cann_device_context * ctx = (ggml_backend_cann_device_context *)dev->context;
+    return ggml_backend_cann_buffer_type(ctx->device);
+}
+
+static ggml_backend_buffer_type_t ggml_backend_cann_device_get_host_buffer_type(ggml_backend_dev_t dev) {
+    GGML_UNUSED(dev);
+    return ggml_backend_cann_host_buffer_type();
+}
+
+/**
+ * @brief Creates a new event for the CANN backend device.
+ *
+ * This function initializes a new event for the CANN backend by setting the
+ * device and creating an ACL runtime event. The created event is then wrapped
+ * in a ggml_backend_event structure and returned.
+ *
+ * @param backend Pointer to the CANN backend.
+ * @return ggml_backend_event_t Returns a pointer to the new event structure.
+ */
+static ggml_backend_event_t ggml_backend_cann_device_event_new(
+    ggml_backend_dev_t dev) {
+    ggml_backend_cann_device_context * dev_ctx = (ggml_backend_cann_device_context *)dev->context;
+
+    ggml_cann_set_device(dev_ctx->device);
+
+    aclrtEvent event;
+    ACL_CHECK(aclrtCreateEvent(&event));
+
+    return new ggml_backend_event{
+        /* .device = */ ggml_backend_reg_dev_get(ggml_backend_cann_reg(), dev_ctx->device),
+        /* .context = */ event,
+    };
+}
+
+/**
+ * @brief Frees a CANN backend event.
+ *
+ * This function destroys the ACL runtime event associated with the given CANN
+ * backend event and then deletes the event structure itself.
+ *
+ * @param event Pointer to the event structure to be freed.
+ */
+static void ggml_backend_cann_device_event_free(ggml_backend_dev_t dev, ggml_backend_event_t event) {
+    ACL_CHECK(aclrtDestroyEvent((aclrtEvent)event->context));
+
+    delete event;
+    GGML_UNUSED(dev);
+}
+
+/**
+ * @brief Synchronizes the given event on the CANN backend.
+ *
+ * This function waits for the specified event to complete on the ACL runtime.
+ *
+ * @param event Pointer to the event structure to be synchronized.
+ */
+static void ggml_backend_cann_device_event_synchronize(ggml_backend_dev_t dev, ggml_backend_event_t event) {
+    ACL_CHECK(aclrtSynchronizeEvent((aclrtEvent)event->context));
+
+    GGML_UNUSED(dev);
+}
+
+static const ggml_backend_device_i ggml_backend_cann_device_interface = {
+    /* .get_name                = */ ggml_backend_cann_device_get_name,
+    /* .get_description         = */ ggml_backend_cann_device_get_description,
+    /* .get_memory              = */ ggml_backend_cann_device_get_memory,
+    /* .get_type                = */ ggml_backend_cann_device_get_type,
+    /* .get_props               = */ ggml_backend_cann_device_get_props,
+    /* .init_backend            = */ ggml_backend_cann_device_init,    // called for every card
+    /* .get_buffer_type         = */ ggml_backend_cann_device_get_buffer_type,
+    /* .get_host_buffer_type    = */ ggml_backend_cann_device_get_host_buffer_type,
+    /* .buffer_from_host_ptr    = */ NULL, // not supported for CANN
+    /* .supports_op             = */ ggml_backend_cann_supports_op,
+    /* .supports_buft           = */ ggml_backend_cann_supports_buft,
+    /* .offload_op              = */ ggml_backend_cann_offload_op,
+    /* .event_new               = */ ggml_backend_cann_device_event_new,
+    /* .event_free              = */ ggml_backend_cann_device_event_free,
+    /* .event_synchronize       = */ ggml_backend_cann_device_event_synchronize,
+};
+
+
+// backend reg
+struct ggml_backend_cann_reg_context {
+    std::vector<ggml_backend_dev_t> devices;
+};
+
+static const char * ggml_backend_cann_reg_get_name(ggml_backend_reg_t reg) {
+    GGML_UNUSED(reg);
+    return GGML_CANN_NAME;
+}
+
+static size_t ggml_backend_cann_reg_get_device_count(ggml_backend_reg_t reg) {
+    ggml_backend_cann_reg_context * ctx = (ggml_backend_cann_reg_context *)reg->context;
+    return ctx->devices.size();
+}
+
+static ggml_backend_dev_t ggml_backend_cann_reg_get_device(ggml_backend_reg_t reg, size_t index) {
+    ggml_backend_cann_reg_context * ctx = (ggml_backend_cann_reg_context *)reg->context;
+    GGML_ASSERT(index < ctx->devices.size());
+    return ctx->devices[index];
+}
+
+static void * ggml_backend_cann_reg_get_proc_address(ggml_backend_reg_t reg, const char * name) {
+    GGML_UNUSED(reg);
+    GGML_UNUSED(name);
+    // reserved for future use
+    return nullptr;
+}
+
+static const ggml_backend_reg_i ggml_backend_cann_reg_interface = {
+    /* .get_name          = */ ggml_backend_cann_reg_get_name,
+    /* .get_device_count  = */ ggml_backend_cann_reg_get_device_count,
+    /* .get_device_get    = */ ggml_backend_cann_reg_get_device,
+    /* .get_proc_address  = */ ggml_backend_cann_reg_get_proc_address,
+};
+
+// backend registry, called only once for cann backend
+ggml_backend_reg_t ggml_backend_cann_reg() {
+    static ggml_backend_reg reg;
+    static bool initialized = false;
+
+    {
+        static std::mutex mutex;
+        std::lock_guard<std::mutex> lock(mutex);
+        if (!initialized) {
+            aclInit(nullptr);
+            ggml_backend_cann_reg_context * ctx = new ggml_backend_cann_reg_context;
+
+            for (int i = 0; i < ggml_cann_info().device_count; i++) {
+                ggml_backend_cann_device_context* dev_ctx = new ggml_backend_cann_device_context();
+                dev_ctx->description = aclrtGetSocName();
+                dev_ctx->device = i;
+                dev_ctx->name = GGML_CANN_NAME + std::to_string(i);
+                ggml_cann_set_device(i);
+                ggml_backend_dev_t dev = new ggml_backend_device {
+                    /* .interface = */ ggml_backend_cann_device_interface,
+                    /* .reg       = */ &reg,
+                    /* .context   = */ dev_ctx
+                };
+                ctx->devices.push_back(dev);
+            }
+
+            reg = ggml_backend_reg {
+                /* .interface = */ ggml_backend_cann_reg_interface,
+                /* .context   = */ ctx
+            };
+        }
+
+        initialized = true;
+    }
+
+    return &reg;
+}
+
 ggml_backend_t ggml_backend_cann_init(int32_t device) {
     aclInit(nullptr);
     if (device < 0 || device >= ggml_backend_cann_get_device_count()) {
@@ -1993,7 +2099,7 @@ ggml_backend_t ggml_backend_cann_init(int32_t device) {
     ggml_backend_t cann_backend =
         new ggml_backend{/* .guid      = */ ggml_backend_cann_guid(),
                          /* .interface = */ ggml_backend_cann_interface,
-                         /* .device    = */ nullptr,
+                         /* .device    = */ ggml_backend_reg_dev_get(ggml_backend_cann_reg(), device),
                          /* .context   = */ ctx};
 
     return cann_backend;

ggml/src/ggml-cpu-impl.h

@@ -1,614 +0,0 @@
-#pragma once
-
-// GGML CPU internal header
-
-#include "ggml.h"
-#include "ggml-impl.h"
-#include <stdlib.h> // load `stdlib.h` before other headers to work around MinGW bug: https://sourceforge.net/p/mingw-w64/bugs/192/
-//#include <stddef.h>
-#include <stdbool.h>
-#include <string.h> // memcpy
-#include <math.h>   // fabsf
-
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-#if defined(_MSC_VER)
-
-#define m512bh(p) p
-#define m512i(p) p
-
-#else
-
-#define m512bh(p) (__m512bh)(p)
-#define m512i(p) (__m512i)(p)
-
-#endif
-
-/**
- * Converts brain16 to float32.
- *
- * The bfloat16 floating point format has the following structure:
- *
- *       ┌sign
- *       │
- *       │   ┌exponent
- *       │   │
- *       │   │      ┌mantissa
- *       │   │      │
- *       │┌──┴───┐┌─┴───┐
- *     0b0000000000000000 brain16
- *
- * Since bf16 has the same number of exponent bits as a 32bit float,
- * encoding and decoding numbers becomes relatively straightforward.
- *
- *       ┌sign
- *       │
- *       │   ┌exponent
- *       │   │
- *       │   │      ┌mantissa
- *       │   │      │
- *       │┌──┴───┐┌─┴───────────────────┐
- *     0b00000000000000000000000000000000 IEEE binary32
- *
- * For comparison, the standard fp16 format has fewer exponent bits.
- *
- *       ┌sign
- *       │
- *       │  ┌exponent
- *       │  │
- *       │  │    ┌mantissa
- *       │  │    │
- *       │┌─┴─┐┌─┴──────┐
- *     0b0000000000000000 IEEE binary16
- *
- * @see IEEE 754-2008
- */
-static inline float ggml_compute_bf16_to_fp32(ggml_bf16_t h) {
-    union {
-        float f;
-        uint32_t i;
-    } u;
-    u.i = (uint32_t)h.bits << 16;
-    return u.f;
-}
-
-/**
- * Converts float32 to brain16.
- *
- * This is binary identical with Google Brain float conversion.
- * Floats shall round to nearest even, and NANs shall be quiet.
- * Subnormals aren't flushed to zero, except perhaps when used.
- * This code should vectorize nicely if using modern compilers.
- */
-static inline ggml_bf16_t ggml_compute_fp32_to_bf16(float s) {
-    ggml_bf16_t h;
-    union {
-        float f;
-        uint32_t i;
-    } u;
-    u.f = s;
-    if ((u.i & 0x7fffffff) > 0x7f800000) { /* nan */
-        h.bits = (u.i >> 16) | 64; /* force to quiet */
-        return h;
-    }
-    h.bits = (u.i + (0x7fff + ((u.i >> 16) & 1))) >> 16;
-    return h;
-}
-
-#define GGML_FP32_TO_BF16(x) ggml_compute_fp32_to_bf16(x)
-#define GGML_BF16_TO_FP32(x) ggml_compute_bf16_to_fp32(x)
-
-// __FMA__ and __F16C__ are not defined in MSVC, however they are implied with AVX2/AVX512
-#if defined(_MSC_VER) && (defined(__AVX2__) || defined(__AVX512F__))
-#ifndef __FMA__
-#define __FMA__
-#endif
-#ifndef __F16C__
-#define __F16C__
-#endif
-#endif
-
-// __SSE3__ and __SSSE3__ are not defined in MSVC, but SSE3/SSSE3 are present when AVX/AVX2/AVX512 are available
-#if defined(_MSC_VER) && (defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__))
-#ifndef __SSE3__
-#define __SSE3__
-#endif
-#ifndef __SSSE3__
-#define __SSSE3__
-#endif
-#endif
-
-#if defined(__ARM_FEATURE_SVE)
-#include <arm_sve.h>
-#include <sys/prctl.h>
-#endif
-
-// 16-bit float
-// on Arm, we use __fp16
-// on x86, we use uint16_t
-#if defined(__ARM_NEON)
-
-// if YCM cannot find <arm_neon.h>, make a symbolic link to it, for example:
-//
-//   $ ln -sfn /Library/Developer/CommandLineTools/usr/lib/clang/13.1.6/include/arm_neon.h ./src/
-//
-#include <arm_neon.h>
-
-#ifdef _MSC_VER
-
-typedef uint16_t ggml_fp16_internal_t;
-
-#define ggml_vld1q_u32(w,x,y,z) { ((w) + ((uint64_t)(x) << 32)), ((y) + ((uint64_t)(z) << 32)) }
-
-#else
-
-typedef __fp16 ggml_fp16_internal_t;
-
-#define ggml_vld1q_u32(w,x,y,z) { (w), (x), (y), (z) }
-
-#endif // _MSC_VER
-
-#if !defined(__aarch64__)
-
-// 32-bit ARM compatibility
-
-// vaddlvq_s16
-// vpaddq_s16
-// vpaddq_s32
-// vaddvq_s32
-// vaddvq_f32
-// vmaxvq_f32
-// vcvtnq_s32_f32
-// vzip1_u8
-// vzip2_u8
-
-inline static int32_t vaddlvq_s16(int16x8_t v) {
-    int32x4_t v0 = vreinterpretq_s32_s64(vpaddlq_s32(vpaddlq_s16(v)));
-    return vgetq_lane_s32(v0, 0) + vgetq_lane_s32(v0, 2);
-}
-
-inline static int16x8_t vpaddq_s16(int16x8_t a, int16x8_t b) {
-    int16x4_t a0 = vpadd_s16(vget_low_s16(a), vget_high_s16(a));
-    int16x4_t b0 = vpadd_s16(vget_low_s16(b), vget_high_s16(b));
-    return vcombine_s16(a0, b0);
-}
-
-inline static int32x4_t vpaddq_s32(int32x4_t a, int32x4_t b) {
-    int32x2_t a0 = vpadd_s32(vget_low_s32(a), vget_high_s32(a));
-    int32x2_t b0 = vpadd_s32(vget_low_s32(b), vget_high_s32(b));
-    return vcombine_s32(a0, b0);
-}
-
-inline static int32_t vaddvq_s32(int32x4_t v) {
-    return vgetq_lane_s32(v, 0) + vgetq_lane_s32(v, 1) + vgetq_lane_s32(v, 2) + vgetq_lane_s32(v, 3);
-}
-
-inline static float vaddvq_f32(float32x4_t v) {
-    return vgetq_lane_f32(v, 0) + vgetq_lane_f32(v, 1) + vgetq_lane_f32(v, 2) + vgetq_lane_f32(v, 3);
-}
-
-inline static float vmaxvq_f32(float32x4_t v) {
-    return
-        MAX(MAX(vgetq_lane_f32(v, 0), vgetq_lane_f32(v, 1)),
-            MAX(vgetq_lane_f32(v, 2), vgetq_lane_f32(v, 3)));
-}
-
-inline static int32x4_t vcvtnq_s32_f32(float32x4_t v) {
-    int32x4_t res;
-
-    res[0] = roundf(vgetq_lane_f32(v, 0));
-    res[1] = roundf(vgetq_lane_f32(v, 1));
-    res[2] = roundf(vgetq_lane_f32(v, 2));
-    res[3] = roundf(vgetq_lane_f32(v, 3));
-
-    return res;
-}
-
-inline static uint8x8_t vzip1_u8(uint8x8_t a, uint8x8_t b) {
-    uint8x8_t res;
-
-    res[0] = a[0]; res[1] = b[0];
-    res[2] = a[1]; res[3] = b[1];
-    res[4] = a[2]; res[5] = b[2];
-    res[6] = a[3]; res[7] = b[3];
-
-    return res;
-}
-
-inline static uint8x8_t vzip2_u8(uint8x8_t a, uint8x8_t b) {
-    uint8x8_t res;
-
-    res[0] = a[4]; res[1] = b[4];
-    res[2] = a[5]; res[3] = b[5];
-    res[4] = a[6]; res[5] = b[6];
-    res[6] = a[7]; res[7] = b[7];
-
-    return res;
-}
-
-// vld1q_s16_x2
-// vld1q_u8_x2
-// vld1q_u8_x4
-// vld1q_s8_x2
-// vld1q_s8_x4
-// TODO: double-check these work correctly
-
-typedef struct ggml_int16x8x2_t {
-    int16x8_t val[2];
-} ggml_int16x8x2_t;
-
-inline static ggml_int16x8x2_t ggml_vld1q_s16_x2(const int16_t * ptr) {
-    ggml_int16x8x2_t res;
-
-    res.val[0] = vld1q_s16(ptr + 0);
-    res.val[1] = vld1q_s16(ptr + 8);
-
-    return res;
-}
-
-typedef struct ggml_uint8x16x2_t {
-    uint8x16_t val[2];
-} ggml_uint8x16x2_t;
-
-inline static ggml_uint8x16x2_t ggml_vld1q_u8_x2(const uint8_t * ptr) {
-    ggml_uint8x16x2_t res;
-
-    res.val[0] = vld1q_u8(ptr + 0);
-    res.val[1] = vld1q_u8(ptr + 16);
-
-    return res;
-}
-
-typedef struct ggml_uint8x16x4_t {
-    uint8x16_t val[4];
-} ggml_uint8x16x4_t;
-
-inline static ggml_uint8x16x4_t ggml_vld1q_u8_x4(const uint8_t * ptr) {
-    ggml_uint8x16x4_t res;
-
-    res.val[0] = vld1q_u8(ptr + 0);
-    res.val[1] = vld1q_u8(ptr + 16);
-    res.val[2] = vld1q_u8(ptr + 32);
-    res.val[3] = vld1q_u8(ptr + 48);
-
-    return res;
-}
-
-typedef struct ggml_int8x16x2_t {
-    int8x16_t val[2];
-} ggml_int8x16x2_t;
-
-inline static ggml_int8x16x2_t ggml_vld1q_s8_x2(const int8_t * ptr) {
-    ggml_int8x16x2_t res;
-
-    res.val[0] = vld1q_s8(ptr + 0);
-    res.val[1] = vld1q_s8(ptr + 16);
-
-    return res;
-}
-
-typedef struct ggml_int8x16x4_t {
-    int8x16_t val[4];
-} ggml_int8x16x4_t;
-
-inline static ggml_int8x16x4_t ggml_vld1q_s8_x4(const int8_t * ptr) {
-    ggml_int8x16x4_t res;
-
-    res.val[0] = vld1q_s8(ptr + 0);
-    res.val[1] = vld1q_s8(ptr + 16);
-    res.val[2] = vld1q_s8(ptr + 32);
-    res.val[3] = vld1q_s8(ptr + 48);
-
-    return res;
-}
-
-// NOTE: not tested
-inline static int8x16_t ggml_vqtbl1q_s8(int8x16_t a, uint8x16_t b) {
-    int8x16_t res;
-
-    res[ 0] = a[b[ 0]];
-    res[ 1] = a[b[ 1]];
-    res[ 2] = a[b[ 2]];
-    res[ 3] = a[b[ 3]];
-    res[ 4] = a[b[ 4]];
-    res[ 5] = a[b[ 5]];
-    res[ 6] = a[b[ 6]];
-    res[ 7] = a[b[ 7]];
-    res[ 8] = a[b[ 8]];
-    res[ 9] = a[b[ 9]];
-    res[10] = a[b[10]];
-    res[11] = a[b[11]];
-    res[12] = a[b[12]];
-    res[13] = a[b[13]];
-    res[14] = a[b[14]];
-    res[15] = a[b[15]];
-
-    return res;
-}
-
-// NOTE: not tested
-inline static uint8x16_t ggml_vqtbl1q_u8(uint8x16_t a, uint8x16_t b) {
-    uint8x16_t res;
-
-    res[ 0] = a[b[ 0]];
-    res[ 1] = a[b[ 1]];
-    res[ 2] = a[b[ 2]];
-    res[ 3] = a[b[ 3]];
-    res[ 4] = a[b[ 4]];
-    res[ 5] = a[b[ 5]];
-    res[ 6] = a[b[ 6]];
-    res[ 7] = a[b[ 7]];
-    res[ 8] = a[b[ 8]];
-    res[ 9] = a[b[ 9]];
-    res[10] = a[b[10]];
-    res[11] = a[b[11]];
-    res[12] = a[b[12]];
-    res[13] = a[b[13]];
-    res[14] = a[b[14]];
-    res[15] = a[b[15]];
-
-    return res;
-}
-
-#else
-
-#define ggml_int16x8x2_t  int16x8x2_t
-#define ggml_uint8x16x2_t uint8x16x2_t
-#define ggml_uint8x16x4_t uint8x16x4_t
-#define ggml_int8x16x2_t  int8x16x2_t
-#define ggml_int8x16x4_t  int8x16x4_t
-
-#define ggml_vld1q_s16_x2 vld1q_s16_x2
-#define ggml_vld1q_u8_x2  vld1q_u8_x2
-#define ggml_vld1q_u8_x4  vld1q_u8_x4
-#define ggml_vld1q_s8_x2  vld1q_s8_x2
-#define ggml_vld1q_s8_x4  vld1q_s8_x4
-#define ggml_vqtbl1q_s8   vqtbl1q_s8
-#define ggml_vqtbl1q_u8   vqtbl1q_u8
-
-#endif // !defined(__aarch64__)
-
-#if !defined(__ARM_FEATURE_DOTPROD)
-
-inline static int32x4_t ggml_vdotq_s32(int32x4_t acc, int8x16_t a, int8x16_t b) {
-    const int16x8_t p0 = vmull_s8(vget_low_s8 (a), vget_low_s8 (b));
-    const int16x8_t p1 = vmull_s8(vget_high_s8(a), vget_high_s8(b));
-
-    return vaddq_s32(acc, vaddq_s32(vpaddlq_s16(p0), vpaddlq_s16(p1)));
-}
-
-#else
-
-#define ggml_vdotq_s32(a, b, c) vdotq_s32(a, b, c)
-
-#endif // !defined(__ARM_FEATURE_DOTPROD)
-
-#endif // defined(__ARM_NEON)
-
-#if defined(__ARM_NEON) && !defined(_MSC_VER)
-
-#define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
-#define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x)
-
-#define GGML_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
-
-static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
-    ggml_fp16_internal_t tmp;
-    memcpy(&tmp, &h, sizeof(ggml_fp16_t));
-    return (float)tmp;
-}
-
-static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
-    ggml_fp16_t res;
-    ggml_fp16_internal_t tmp = f;
-    memcpy(&res, &tmp, sizeof(ggml_fp16_t));
-    return res;
-}
-
-#else
-
-#ifdef __wasm_simd128__
-#include <wasm_simd128.h>
-#else
-#ifdef __POWER9_VECTOR__
-#include <altivec.h>
-#undef bool
-#define bool _Bool
-#else
-#if defined(_MSC_VER) || defined(__MINGW32__)
-#include <intrin.h>
-#else
-#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__) || defined(__SSSE3__) || defined(__SSE3__) || defined(__SSE__)
-#if !defined(__riscv)
-#include <immintrin.h>
-#endif
-#endif
-#endif
-#endif
-#endif
-
-#ifdef __riscv_v_intrinsic
-#include <riscv_vector.h>
-#endif
-
-#if defined(__loongarch64)
-#if defined(__loongarch_asx)
-#include <lasxintrin.h>
-#endif
-#if defined(__loongarch_sx)
-#include <lsxintrin.h>
-#endif
-#endif
-
-#if defined(__loongarch_asx)
-
-typedef union {
-    int32_t i;
-    float f;
-} ft_union;
-
-/* float type data load instructions */
-static __m128 __lsx_vreplfr2vr_s(float val) {
-    ft_union fi_tmpval = {.f = val};
-    return (__m128)__lsx_vreplgr2vr_w(fi_tmpval.i);
-}
-
-static __m256 __lasx_xvreplfr2vr_s(float val) {
-    ft_union fi_tmpval = {.f = val};
-    return (__m256)__lasx_xvreplgr2vr_w(fi_tmpval.i);
-}
-#endif
-
-#ifdef __F16C__
-
-#ifdef _MSC_VER
-#define GGML_COMPUTE_FP16_TO_FP32(x) _mm_cvtss_f32(_mm_cvtph_ps(_mm_cvtsi32_si128(x)))
-#define GGML_COMPUTE_FP32_TO_FP16(x) _mm_extract_epi16(_mm_cvtps_ph(_mm_set_ss(x), 0), 0)
-#else
-#define GGML_COMPUTE_FP16_TO_FP32(x) _cvtsh_ss(x)
-#define GGML_COMPUTE_FP32_TO_FP16(x) _cvtss_sh(x, 0)
-#endif
-
-#elif defined(__POWER9_VECTOR__)
-
-#define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
-#define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x)
-/* the inline asm below is about 12% faster than the lookup method */
-#define GGML_FP16_TO_FP32(x) GGML_COMPUTE_FP16_TO_FP32(x)
-#define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x)
-
-static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
-    register float f;
-    register double d;
-    __asm__(
-        "mtfprd %0,%2\n"
-        "xscvhpdp %0,%0\n"
-        "frsp %1,%0\n" :
-        /* temp */ "=d"(d),
-        /* out */  "=f"(f):
-        /* in */   "r"(h));
-    return f;
-}
-
-static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
-    register double d;
-    register ggml_fp16_t r;
-    __asm__( /* xscvdphp can work on double or single precision */
-        "xscvdphp %0,%2\n"
-        "mffprd %1,%0\n" :
-        /* temp */ "=d"(d),
-        /* out */  "=r"(r):
-        /* in */   "f"(f));
-    return r;
-}
-
-#else
-
-// FP16 <-> FP32
-// ref: https://github.com/Maratyszcza/FP16
-
-static inline float fp32_from_bits(uint32_t w) {
-    union {
-        uint32_t as_bits;
-        float as_value;
-    } fp32;
-    fp32.as_bits = w;
-    return fp32.as_value;
-}
-
-static inline uint32_t fp32_to_bits(float f) {
-    union {
-        float as_value;
-        uint32_t as_bits;
-    } fp32;
-    fp32.as_value = f;
-    return fp32.as_bits;
-}
-
-static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
-    const uint32_t w = (uint32_t) h << 16;
-    const uint32_t sign = w & UINT32_C(0x80000000);
-    const uint32_t two_w = w + w;
-
-    const uint32_t exp_offset = UINT32_C(0xE0) << 23;
-#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) || defined(__GNUC__) && !defined(__STRICT_ANSI__)
-    const float exp_scale = 0x1.0p-112f;
-#else
-    const float exp_scale = fp32_from_bits(UINT32_C(0x7800000));
-#endif
-    const float normalized_value = fp32_from_bits((two_w >> 4) + exp_offset) * exp_scale;
-
-    const uint32_t magic_mask = UINT32_C(126) << 23;
-    const float magic_bias = 0.5f;
-    const float denormalized_value = fp32_from_bits((two_w >> 17) | magic_mask) - magic_bias;
-
-    const uint32_t denormalized_cutoff = UINT32_C(1) << 27;
-    const uint32_t result = sign |
-        (two_w < denormalized_cutoff ? fp32_to_bits(denormalized_value) : fp32_to_bits(normalized_value));
-    return fp32_from_bits(result);
-}
-
-static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
-#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) || defined(__GNUC__) && !defined(__STRICT_ANSI__)
-    const float scale_to_inf = 0x1.0p+112f;
-    const float scale_to_zero = 0x1.0p-110f;
-#else
-    const float scale_to_inf = fp32_from_bits(UINT32_C(0x77800000));
-    const float scale_to_zero = fp32_from_bits(UINT32_C(0x08800000));
-#endif
-    float base = (fabsf(f) * scale_to_inf) * scale_to_zero;
-
-    const uint32_t w = fp32_to_bits(f);
-    const uint32_t shl1_w = w + w;
-    const uint32_t sign = w & UINT32_C(0x80000000);
-    uint32_t bias = shl1_w & UINT32_C(0xFF000000);
-    if (bias < UINT32_C(0x71000000)) {
-        bias = UINT32_C(0x71000000);
-    }
-
-    base = fp32_from_bits((bias >> 1) + UINT32_C(0x07800000)) + base;
-    const uint32_t bits = fp32_to_bits(base);
-    const uint32_t exp_bits = (bits >> 13) & UINT32_C(0x00007C00);
-    const uint32_t mantissa_bits = bits & UINT32_C(0x00000FFF);
-    const uint32_t nonsign = exp_bits + mantissa_bits;
-    return (sign >> 16) | (shl1_w > UINT32_C(0xFF000000) ? UINT16_C(0x7E00) : nonsign);
-}
-
-#define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
-#define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x)
-
-#endif // __F16C__
-
-#endif // defined(__ARM_NEON) && (!defined(__MSC_VER)
-
-#ifdef __ARM_FEATURE_SVE
-#include <arm_sve.h>
-#endif // __ARM_FEATURE_SVE
-
-// precomputed f32 table for f16 (256 KB)
-// defined in ggml.c, initialized in ggml_init()
-extern float ggml_table_f32_f16[1 << 16];
-
-// On ARM NEON, it's quicker to directly convert x -> x instead of calling into ggml_lookup_fp16_to_fp32,
-// so we define GGML_FP16_TO_FP32 and GGML_FP32_TO_FP16 elsewhere for NEON.
-// This is also true for POWER9.
-#if !defined(GGML_FP16_TO_FP32)
-inline static float ggml_lookup_fp16_to_fp32(ggml_fp16_t f) {
-    uint16_t s;
-    memcpy(&s, &f, sizeof(uint16_t));
-    return ggml_table_f32_f16[s];
-}
-
-#define GGML_FP16_TO_FP32(x) ggml_lookup_fp16_to_fp32(x)
-#endif
-
-#if !defined(GGML_FP32_TO_FP16)
-#define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x)
-#endif
-
-#ifdef __cplusplus
-}
-#endif

ggml/src/ggml-cpu/CMakeLists.txt

@@ -0,0 +1,261 @@
+add_library(ggml-cpu
+            ggml-cpu.c
+            ggml-cpu.cpp
+            ggml-cpu-aarch64.c
+            ggml-cpu-aarch64.h
+            ggml-cpu-quants.c
+            ggml-cpu-quants.h
+            )
+
+target_link_libraries(ggml-cpu PRIVATE ggml-base)
+target_include_directories(ggml-cpu PRIVATE . ..)
+
+if (APPLE AND GGML_ACCELERATE)
+    find_library(ACCELERATE_FRAMEWORK Accelerate)
+    if (ACCELERATE_FRAMEWORK)
+        message(STATUS "Accelerate framework found")
+
+        add_compile_definitions(GGML_USE_ACCELERATE)
+        add_compile_definitions(ACCELERATE_NEW_LAPACK)
+        add_compile_definitions(ACCELERATE_LAPACK_ILP64)
+
+        target_link_libraries(ggml-cpu PRIVATE ${ACCELERATE_FRAMEWORK})
+    else()
+        message(WARNING "Accelerate framework not found")
+    endif()
+endif()
+
+if (GGML_OPENMP)
+    find_package(OpenMP)
+    if (OpenMP_FOUND)
+        message(STATUS "OpenMP found")
+
+        add_compile_definitions(GGML_USE_OPENMP)
+
+        target_link_libraries(ggml-cpu PRIVATE OpenMP::OpenMP_C OpenMP::OpenMP_CXX)
+
+        # FIXME: should be replaced with a compiler id check
+        #if (GGML_MUSA)
+        #    list(APPEND GGML_CPU_EXTRA_INCLUDES     "/usr/lib/llvm-14/lib/clang/14.0.0/include")
+        #    list(APPEND GGML_CPU_EXTRA_LIBS_PRIVATE "/usr/lib/llvm-14/lib/libomp.so")
+        #endif()
+    else()
+        message(WARNING "OpenMP not found")
+    endif()
+endif()
+
+if (GGML_LLAMAFILE)
+    message(STATUS "Using llamafile")
+
+    add_compile_definitions(GGML_USE_LLAMAFILE)
+
+    target_sources(ggml-cpu PRIVATE
+                    llamafile/sgemm.cpp
+                    llamafile/sgemm.h)
+endif()
+
+if (GGML_CPU_HBM)
+    find_library(memkind memkind REQUIRED)
+
+    message(STATUS "Using memkind for CPU HBM")
+
+    add_compile_definitions(GGML_USE_CPU_HBM)
+
+    target_link_libraries(ggml-cpu PUBLIC memkind)
+endif()
+
+if (CMAKE_OSX_ARCHITECTURES      STREQUAL "arm64" OR
+    CMAKE_GENERATOR_PLATFORM_LWR STREQUAL "arm64" OR
+    (NOT CMAKE_OSX_ARCHITECTURES      AND
+     NOT CMAKE_GENERATOR_PLATFORM_LWR AND
+         CMAKE_SYSTEM_PROCESSOR MATCHES "^(aarch64|arm.*|ARM64)$"))
+
+    message(STATUS "ARM detected")
+
+    if (MSVC)
+        add_compile_definitions(__aarch64__) # MSVC defines _M_ARM64 instead
+        add_compile_definitions(__ARM_NEON)
+        add_compile_definitions(__ARM_FEATURE_FMA)
+
+        set(CMAKE_REQUIRED_FLAGS_PREV ${CMAKE_REQUIRED_FLAGS})
+        string(JOIN " " CMAKE_REQUIRED_FLAGS ${CMAKE_REQUIRED_FLAGS} "/arch:armv8.2")
+
+        check_cxx_source_compiles("#include <arm_neon.h>\nint main() { int8x16_t _a, _b; int32x4_t _s = vdotq_s32(_s, _a, _b); return 0; }" GGML_COMPILER_SUPPORT_DOTPROD)
+        if (GGML_COMPILER_SUPPORT_DOTPROD)
+            add_compile_definitions(__ARM_FEATURE_DOTPROD)
+        endif ()
+
+        check_cxx_source_compiles("#include <arm_neon.h>\nint main() { int8x16_t _a, _b; int32x4_t _s = vmlaq_f32(_s, _a, _b); return 0; }" GGML_COMPILER_SUPPORT_MATMUL_INT8)
+
+        if (GGML_COMPILER_SUPPORT_MATMUL_INT8)
+            add_compile_definitions(__ARM_FEATURE_MATMUL_INT8)
+        endif ()
+
+        check_cxx_source_compiles("#include <arm_neon.h>\nint main() { float16_t _a; float16x8_t _s = vdupq_n_f16(_a); return 0; }" GGML_COMPILER_SUPPORT_FP16_VECTOR_ARITHMETIC)
+        if (GGML_COMPILER_SUPPORT_FP16_VECTOR_ARITHMETIC)
+            add_compile_definitions(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC)
+        endif ()
+
+        set(CMAKE_REQUIRED_FLAGS ${CMAKE_REQUIRED_FLAGS_PREV})
+    else()
+        check_cxx_compiler_flag(-mfp16-format=ieee COMPILER_SUPPORTS_FP16_FORMAT_I3E)
+        if (NOT "${COMPILER_SUPPORTS_FP16_FORMAT_I3E}" STREQUAL "")
+            list(APPEND ARCH_FLAGS -mfp16-format=ieee)
+        endif()
+        if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "armv6")
+            # Raspberry Pi 1, Zero
+            list(APPEND ARCH_FLAGS -mfpu=neon-fp-armv8 -mno-unaligned-access)
+        endif()
+        if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "armv7")
+            if ("${CMAKE_SYSTEM_NAME}" STREQUAL "Android")
+                # Android armeabi-v7a
+                list(APPEND ARCH_FLAGS -mfpu=neon-vfpv4 -mno-unaligned-access -funsafe-math-optimizations)
+            else()
+                # Raspberry Pi 2
+                list(APPEND ARCH_FLAGS -mfpu=neon-fp-armv8 -mno-unaligned-access -funsafe-math-optimizations)
+            endif()
+        endif()
+        if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "armv8")
+            # Android arm64-v8a
+            # Raspberry Pi 3, 4, Zero 2 (32-bit)
+            list(APPEND ARCH_FLAGS -mno-unaligned-access)
+        endif()
+        if (GGML_SVE)
+            list(APPEND ARCH_FLAGS -march=armv8.6-a+sve)
+        endif()
+    endif()
+elseif (CMAKE_OSX_ARCHITECTURES STREQUAL "x86_64" OR CMAKE_GENERATOR_PLATFORM_LWR MATCHES "^(x86_64|i686|amd64|x64|win32)$" OR
+        (NOT CMAKE_OSX_ARCHITECTURES AND NOT CMAKE_GENERATOR_PLATFORM_LWR AND
+         CMAKE_SYSTEM_PROCESSOR MATCHES "^(x86_64|i686|AMD64)$"))
+    message(STATUS "x86 detected")
+    if (MSVC)
+        # instruction set detection for MSVC only
+        if (GGML_NATIVE)
+            # TODO: improve, should not reference files from the parent folder
+            include(cmake/FindSIMD.cmake)
+        endif ()
+        if (GGML_AVX512)
+            list(APPEND ARCH_FLAGS /arch:AVX512)
+            # MSVC has no compile-time flags enabling specific
+            # AVX512 extensions, neither it defines the
+            # macros corresponding to the extensions.
+            # Do it manually.
+            if (GGML_AVX512_VBMI)
+                add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AVX512VBMI__>)
+                add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AVX512VBMI__>)
+                if (CMAKE_C_COMPILER_ID STREQUAL "Clang")
+                    list(APPEND ARCH_FLAGS -mavx512vbmi)
+                endif()
+            endif()
+            if (GGML_AVX512_VNNI)
+                add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AVX512VNNI__>)
+                add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AVX512VNNI__>)
+                if (CMAKE_C_COMPILER_ID STREQUAL "Clang")
+                    list(APPEND ARCH_FLAGS -mavx512vnni)
+                endif()
+            endif()
+            if (GGML_AVX512_BF16)
+                add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AVX512BF16__>)
+                add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AVX512BF16__>)
+                if (CMAKE_C_COMPILER_ID STREQUAL "Clang")
+                    list(APPEND ARCH_FLAGS -mavx512bf16)
+                endif()
+            endif()
+            if (GGML_AMX_TILE)
+                add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AMX_TILE__>)
+                add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AMX_TILE__>)
+            endif()
+            if (GGML_AMX_INT8)
+                add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AMX_INT8__>)
+                add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AMX_INT8__>)
+            endif()
+            if (GGML_AMX_BF16)
+                add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AMX_BF16__>)
+                add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AMX_BF16__>)
+            endif()
+        elseif (GGML_AVX2)
+            list(APPEND ARCH_FLAGS /arch:AVX2)
+        elseif (GGML_AVX)
+            list(APPEND ARCH_FLAGS /arch:AVX)
+        endif()
+    else()
+        if (GGML_NATIVE)
+            list(APPEND ARCH_FLAGS -march=native)
+        endif()
+        if (GGML_F16C)
+            list(APPEND ARCH_FLAGS -mf16c)
+        endif()
+        if (GGML_FMA)
+            list(APPEND ARCH_FLAGS -mfma)
+        endif()
+        if (GGML_AVX)
+            list(APPEND ARCH_FLAGS -mavx)
+        endif()
+        if (GGML_AVX2)
+            list(APPEND ARCH_FLAGS -mavx2)
+        endif()
+        if (GGML_AVX512)
+            list(APPEND ARCH_FLAGS -mavx512f)
+            list(APPEND ARCH_FLAGS -mavx512dq)
+            list(APPEND ARCH_FLAGS -mavx512bw)
+        endif()
+        if (GGML_AVX512_VBMI)
+            list(APPEND ARCH_FLAGS -mavx512vbmi)
+        endif()
+        if (GGML_AVX512_VNNI)
+            list(APPEND ARCH_FLAGS -mavx512vnni)
+        endif()
+        if (GGML_AVX512_BF16)
+            list(APPEND ARCH_FLAGS -mavx512bf16)
+        endif()
+        if (GGML_AMX_TILE)
+            list(APPEND ARCH_FLAGS -mamx-tile)
+        endif()
+        if (GGML_AMX_INT8)
+            list(APPEND ARCH_FLAGS -mamx-int8)
+        endif()
+        if (GGML_AMX_BF16)
+            list(APPEND ARCH_FLAGS -mamx-bf16)
+        endif()
+    endif()
+elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc64")
+    message(STATUS "PowerPC detected")
+    execute_process(COMMAND bash -c "grep POWER10 /proc/cpuinfo | head -n 1" OUTPUT_VARIABLE POWER10_M)
+    string(FIND "${POWER10_M}" "POWER10" substring_index)
+    if (NOT DEFINED substring_index OR "${substring_index}" STREQUAL "")
+        set(substring_index -1)
+    endif()
+
+    if (${substring_index} GREATER_EQUAL 0)
+       list(APPEND ARCH_FLAGS -mcpu=power10)
+    elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc64le")
+       list(APPEND ARCH_FLAGS -mcpu=powerpc64le)
+    else()
+        list(APPEND ARCH_FLAGS -mcpu=native -mtune=native)
+        #TODO: Add  targets for Power8/Power9 (Altivec/VSX) and Power10(MMA) and query for big endian systems (ppc64/le/be)
+    endif()
+elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "loongarch64")
+    message(STATUS "loongarch64 detected")
+
+    list(APPEND ARCH_FLAGS -march=loongarch64)
+    if (GGML_LASX)
+        list(APPEND ARCH_FLAGS -mlasx)
+    endif()
+    if (GGML_LSX)
+        list(APPEND ARCH_FLAGS -mlsx)
+    endif()
+else()
+    message(STATUS "Unknown architecture")
+endif()
+
+if (GGML_CPU_AARCH64)
+    message(STATUS "Using runtime weight conversion of Q4_0 to Q4_0_x_x to enable optimized GEMM/GEMV kernels")
+    add_compile_definitions(GGML_USE_CPU_AARCH64)
+endif()
+
+target_compile_options(ggml-cpu PRIVATE "$<$<COMPILE_LANGUAGE:CXX>:${ARCH_FLAGS}>")
+target_compile_options(ggml-cpu PRIVATE "$<$<COMPILE_LANGUAGE:C>:${ARCH_FLAGS}>")
+
+if (EMSCRIPTEN)
+    set_target_properties(ggml-cpu PROPERTIES COMPILE_FLAGS "-msimd128")
+endif()

ggml/src/ggml-cpu/ggml-cpu-aarch64.h

@@ -0,0 +1,30 @@
+#pragma once
+
+#include "ggml.h"
+
+// GGML internal header
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+// Quantization
+void quantize_mat_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t nrows, int64_t n_per_row, int64_t blck_size_interleave);
+
+// GEMV
+void ggml_gemv_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+
+// GEMM
+void ggml_gemm_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+
+void           ggml_aarch64_repack_tensor(struct ggml_tensor * cur, enum ggml_type repack_type, const void * data, size_t data_size);
+enum ggml_type ggml_aarch64_get_optimal_repack_type(const struct ggml_tensor * cur);
+
+#ifdef __cplusplus
+}
+#endif
+

ggml/src/ggml-cpu/ggml-cpu-impl.h

@@ -27,80 +27,6 @@ extern "C" {
 
 #endif
 
-/**
- * Converts brain16 to float32.
- *
- * The bfloat16 floating point format has the following structure:
- *
- *       ┌sign
- *       │
- *       │   ┌exponent
- *       │   │
- *       │   │      ┌mantissa
- *       │   │      │
- *       │┌──┴───┐┌─┴───┐
- *     0b0000000000000000 brain16
- *
- * Since bf16 has the same number of exponent bits as a 32bit float,
- * encoding and decoding numbers becomes relatively straightforward.
- *
- *       ┌sign
- *       │
- *       │   ┌exponent
- *       │   │
- *       │   │      ┌mantissa
- *       │   │      │
- *       │┌──┴───┐┌─┴───────────────────┐
- *     0b00000000000000000000000000000000 IEEE binary32
- *
- * For comparison, the standard fp16 format has fewer exponent bits.
- *
- *       ┌sign
- *       │
- *       │  ┌exponent
- *       │  │
- *       │  │    ┌mantissa
- *       │  │    │
- *       │┌─┴─┐┌─┴──────┐
- *     0b0000000000000000 IEEE binary16
- *
- * @see IEEE 754-2008
- */
-static inline float ggml_compute_bf16_to_fp32(ggml_bf16_t h) {
-    union {
-        float f;
-        uint32_t i;
-    } u;
-    u.i = (uint32_t)h.bits << 16;
-    return u.f;
-}
-
-/**
- * Converts float32 to brain16.
- *
- * This is binary identical with Google Brain float conversion.
- * Floats shall round to nearest even, and NANs shall be quiet.
- * Subnormals aren't flushed to zero, except perhaps when used.
- * This code should vectorize nicely if using modern compilers.
- */
-static inline ggml_bf16_t ggml_compute_fp32_to_bf16(float s) {
-    ggml_bf16_t h;
-    union {
-        float f;
-        uint32_t i;
-    } u;
-    u.f = s;
-    if ((u.i & 0x7fffffff) > 0x7f800000) { /* nan */
-        h.bits = (u.i >> 16) | 64; /* force to quiet */
-        return h;
-    }
-    h.bits = (u.i + (0x7fff + ((u.i >> 16) & 1))) >> 16;
-    return h;
-}
-
-#define GGML_FP32_TO_BF16(x) ggml_compute_fp32_to_bf16(x)
-#define GGML_BF16_TO_FP32(x) ggml_compute_bf16_to_fp32(x)
-
 // __FMA__ and __F16C__ are not defined in MSVC, however they are implied with AVX2/AVX512
 #if defined(_MSC_VER) && (defined(__AVX2__) || defined(__AVX512F__))
 #ifndef __FMA__
@@ -388,28 +314,6 @@ inline static int32x4_t ggml_vdotq_s32(int32x4_t acc, int8x16_t a, int8x16_t b)
 
 #endif // defined(__ARM_NEON)
 
-#if defined(__ARM_NEON) && !defined(_MSC_VER)
-
-#define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
-#define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x)
-
-#define GGML_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
-
-static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
-    ggml_fp16_internal_t tmp;
-    memcpy(&tmp, &h, sizeof(ggml_fp16_t));
-    return (float)tmp;
-}
-
-static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
-    ggml_fp16_t res;
-    ggml_fp16_internal_t tmp = f;
-    memcpy(&res, &tmp, sizeof(ggml_fp16_t));
-    return res;
-}
-
-#else
-
 #ifdef __wasm_simd128__
 #include <wasm_simd128.h>
 #else
@@ -462,153 +366,6 @@ static __m256 __lasx_xvreplfr2vr_s(float val) {
 }
 #endif
 
-#ifdef __F16C__
-
-#ifdef _MSC_VER
-#define GGML_COMPUTE_FP16_TO_FP32(x) _mm_cvtss_f32(_mm_cvtph_ps(_mm_cvtsi32_si128(x)))
-#define GGML_COMPUTE_FP32_TO_FP16(x) _mm_extract_epi16(_mm_cvtps_ph(_mm_set_ss(x), 0), 0)
-#else
-#define GGML_COMPUTE_FP16_TO_FP32(x) _cvtsh_ss(x)
-#define GGML_COMPUTE_FP32_TO_FP16(x) _cvtss_sh(x, 0)
-#endif
-
-#elif defined(__POWER9_VECTOR__)
-
-#define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
-#define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x)
-/* the inline asm below is about 12% faster than the lookup method */
-#define GGML_FP16_TO_FP32(x) GGML_COMPUTE_FP16_TO_FP32(x)
-#define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x)
-
-static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
-    register float f;
-    register double d;
-    __asm__(
-        "mtfprd %0,%2\n"
-        "xscvhpdp %0,%0\n"
-        "frsp %1,%0\n" :
-        /* temp */ "=d"(d),
-        /* out */  "=f"(f):
-        /* in */   "r"(h));
-    return f;
-}
-
-static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
-    register double d;
-    register ggml_fp16_t r;
-    __asm__( /* xscvdphp can work on double or single precision */
-        "xscvdphp %0,%2\n"
-        "mffprd %1,%0\n" :
-        /* temp */ "=d"(d),
-        /* out */  "=r"(r):
-        /* in */   "f"(f));
-    return r;
-}
-
-#else
-
-// FP16 <-> FP32
-// ref: https://github.com/Maratyszcza/FP16
-
-static inline float fp32_from_bits(uint32_t w) {
-    union {
-        uint32_t as_bits;
-        float as_value;
-    } fp32;
-    fp32.as_bits = w;
-    return fp32.as_value;
-}
-
-static inline uint32_t fp32_to_bits(float f) {
-    union {
-        float as_value;
-        uint32_t as_bits;
-    } fp32;
-    fp32.as_value = f;
-    return fp32.as_bits;
-}
-
-static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
-    const uint32_t w = (uint32_t) h << 16;
-    const uint32_t sign = w & UINT32_C(0x80000000);
-    const uint32_t two_w = w + w;
-
-    const uint32_t exp_offset = UINT32_C(0xE0) << 23;
-#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) || defined(__GNUC__) && !defined(__STRICT_ANSI__)
-    const float exp_scale = 0x1.0p-112f;
-#else
-    const float exp_scale = fp32_from_bits(UINT32_C(0x7800000));
-#endif
-    const float normalized_value = fp32_from_bits((two_w >> 4) + exp_offset) * exp_scale;
-
-    const uint32_t magic_mask = UINT32_C(126) << 23;
-    const float magic_bias = 0.5f;
-    const float denormalized_value = fp32_from_bits((two_w >> 17) | magic_mask) - magic_bias;
-
-    const uint32_t denormalized_cutoff = UINT32_C(1) << 27;
-    const uint32_t result = sign |
-        (two_w < denormalized_cutoff ? fp32_to_bits(denormalized_value) : fp32_to_bits(normalized_value));
-    return fp32_from_bits(result);
-}
-
-static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
-#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) || defined(__GNUC__) && !defined(__STRICT_ANSI__)
-    const float scale_to_inf = 0x1.0p+112f;
-    const float scale_to_zero = 0x1.0p-110f;
-#else
-    const float scale_to_inf = fp32_from_bits(UINT32_C(0x77800000));
-    const float scale_to_zero = fp32_from_bits(UINT32_C(0x08800000));
-#endif
-    float base = (fabsf(f) * scale_to_inf) * scale_to_zero;
-
-    const uint32_t w = fp32_to_bits(f);
-    const uint32_t shl1_w = w + w;
-    const uint32_t sign = w & UINT32_C(0x80000000);
-    uint32_t bias = shl1_w & UINT32_C(0xFF000000);
-    if (bias < UINT32_C(0x71000000)) {
-        bias = UINT32_C(0x71000000);
-    }
-
-    base = fp32_from_bits((bias >> 1) + UINT32_C(0x07800000)) + base;
-    const uint32_t bits = fp32_to_bits(base);
-    const uint32_t exp_bits = (bits >> 13) & UINT32_C(0x00007C00);
-    const uint32_t mantissa_bits = bits & UINT32_C(0x00000FFF);
-    const uint32_t nonsign = exp_bits + mantissa_bits;
-    return (sign >> 16) | (shl1_w > UINT32_C(0xFF000000) ? UINT16_C(0x7E00) : nonsign);
-}
-
-#define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
-#define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x)
-
-#endif // __F16C__
-
-#endif // defined(__ARM_NEON) && (!defined(__MSC_VER)
-
-#ifdef __ARM_FEATURE_SVE
-#include <arm_sve.h>
-#endif // __ARM_FEATURE_SVE
-
-// precomputed f32 table for f16 (256 KB)
-// defined in ggml.c, initialized in ggml_init()
-extern float ggml_table_f32_f16[1 << 16];
-
-// On ARM NEON, it's quicker to directly convert x -> x instead of calling into ggml_lookup_fp16_to_fp32,
-// so we define GGML_FP16_TO_FP32 and GGML_FP32_TO_FP16 elsewhere for NEON.
-// This is also true for POWER9.
-#if !defined(GGML_FP16_TO_FP32)
-inline static float ggml_lookup_fp16_to_fp32(ggml_fp16_t f) {
-    uint16_t s;
-    memcpy(&s, &f, sizeof(uint16_t));
-    return ggml_table_f32_f16[s];
-}
-
-#define GGML_FP16_TO_FP32(x) ggml_lookup_fp16_to_fp32(x)
-#endif
-
-#if !defined(GGML_FP32_TO_FP16)
-#define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x)
-#endif
-
 #ifdef __cplusplus
 }
 #endif

ggml/src/ggml-cpu/ggml-cpu-quants.h

@@ -0,0 +1,63 @@
+#pragma once
+
+#define GGML_COMMON_DECL_C
+#include "ggml-common.h"
+
+#include "ggml.h"
+
+// GGML CPU internal header
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+// Quantization
+void quantize_row_q4_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q4_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q5_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q5_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+
+void quantize_row_q2_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q3_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q4_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q5_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q6_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q8_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+
+void quantize_row_tq1_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_tq2_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+
+void quantize_row_iq4_nl (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_iq4_xs (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+
+// Dot product
+void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+
+void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+
+void ggml_vec_dot_tq1_0_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_tq2_0_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+
+void ggml_vec_dot_iq2_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_iq2_xs_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_iq2_s_q8_K  (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_iq1_s_q8_K  (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_iq1_m_q8_K  (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_iq4_nl_q8_0 (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_iq4_xs_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_iq3_s_q8_K  (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+
+#ifdef __cplusplus
+}
+#endif