disable

* 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/build.yml

@@ -549,8 +549,9 @@ jobs:
           cp models/for-tests-ggml-base.en.bin models/ggml-base.en.bin
           mkdir models/ggml-base.en-encoder.mlmodelc
 
-      - name: Build objc example
-        run: xcodebuild -project examples/whisper.objc/whisper.objc.xcodeproj -scheme whisper.objc -configuration ${{ matrix.build }} -sdk iphonesimulator build
+# TODO: disabled because it fails for some reason with Github Actions
+#      - name: Build objc example
+#        run: xcodebuild -project examples/whisper.objc/whisper.objc.xcodeproj -scheme whisper.objc -configuration ${{ matrix.build }} -sdk iphonesimulator build
 
       - name: Build swiftui example
         run: xcodebuild -project examples/whisper.swiftui/whisper.swiftui.xcodeproj -scheme WhisperCppDemo -configuration ${{ matrix.build }} -sdk iphonesimulator build
@@ -564,12 +565,6 @@ 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:
@@ -588,7 +583,7 @@ 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:

.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

@@ -444,17 +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
 
@@ -464,29 +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))
@@ -513,7 +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
@@ -615,11 +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)
@@ -742,50 +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
 
@@ -801,11 +803,17 @@ endif
 
 OBJ_GGML += \
 	ggml/src/ggml.o \
-	ggml/src/ggml-cpu.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
@@ -910,114 +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-cpu.o: \
-	ggml/src/ggml-cpu.c \
-	ggml/include/ggml.h \
-	ggml/src/ggml-common.h
-	$(CC)  $(CFLAGS)   -c $< -o $@
+LIB_COMMON   = libcommon.so
+LIB_COMMON_S = libcommon.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_SDL   = libcommon-sdl.so
+LIB_COMMON_SDL_S = libcommon-sdl.a
 
-ggml/src/ggml-backend.o: \
-	ggml/src/ggml-backend.cpp \
-	ggml/include/ggml.h \
-	ggml/include/ggml-backend.h
-	$(CXX) $(CXXFLAGS) -c $< -o $@
+# Targets
+BUILD_TARGETS += $(LIB_GGML) $(LIB_GGML_S) $(LIB_LLAMA) $(LIB_LLAMA_S) $(LIB_COMMON) $(LIB_COMMON_S)
 
-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 $@
+# Dependency files
+DEP_FILES = $(OBJ_GGML:.o=.d) $(OBJ_LLAMA:.o=.d) $(OBJ_COMMON:.o=.d)
 
-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 $@
+# Default target
+all: $(BUILD_TARGETS)
 
-ggml/src/ggml-blas.o: \
-	ggml/src/ggml-blas.cpp \
-	ggml/include/ggml-blas.h
-	$(CXX) $(CXXFLAGS) -c $< -o $@
-
-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
 
@@ -1029,34 +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

@@ -28,7 +28,7 @@ let package = Package(
                "tests",
                "CMakeLists.txt",
                "Makefile",
-               "ggml/src/ggml-metal-embed.metal"
+               "ggml/src/ggml-metal/ggml-metal-embed.metal"
             ],
             sources: [
                 "ggml/src/ggml.c",
@@ -36,16 +36,22 @@ let package = Package(
                 "ggml/src/ggml-aarch64.c",
                 "ggml/src/ggml-alloc.c",
                 "ggml/src/ggml-backend.cpp",
-                "ggml/src/ggml-cpu.c",
+                "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/src/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

@@ -7,7 +7,7 @@
 [![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:
 

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": {

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/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.cpp

@@ -179,6 +179,7 @@ enum llm_arch {
     LLM_ARCH_COMMAND_R,
     LLM_ARCH_DBRX,
     LLM_ARCH_OLMO,
+    LLM_ARCH_OLMO_1124,
     LLM_ARCH_OLMOE,
     LLM_ARCH_OPENELM,
     LLM_ARCH_ARCTIC,
@@ -232,6 +233,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_COMMAND_R,       "command-r"    },
     { LLM_ARCH_DBRX,            "dbrx"         },
     { LLM_ARCH_OLMO,            "olmo"         },
+    { LLM_ARCH_OLMO_1124,       "olmo_1124"    },
     { LLM_ARCH_OLMOE,           "olmoe"        },
     { LLM_ARCH_OPENELM,         "openelm"      },
     { LLM_ARCH_ARCTIC,          "arctic"       },
@@ -1207,6 +1209,25 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
         },
     },
+    {
+        LLM_ARCH_OLMO_1124,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,     "output_norm" },
+            { LLM_TENSOR_OUTPUT,          "output" },
+            { LLM_TENSOR_ATTN_Q,          "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_K,          "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_V,          "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,        "blk.%d.attn_output" },
+            { LLM_TENSOR_ATTN_POST_NORM,  "blk.%d.post_attention_norm" },
+            { LLM_TENSOR_ATTN_Q_NORM,     "blk.%d.attn_q_norm" },
+            { LLM_TENSOR_ATTN_K_NORM,     "blk.%d.attn_k_norm" },
+            { LLM_TENSOR_FFN_POST_NORM,   "blk.%d.post_ffw_norm" },
+            { LLM_TENSOR_FFN_GATE,        "blk.%d.ffn_gate" },
+            { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
+        },
+    },
     {
         LLM_ARCH_OLMOE,
         {
@@ -2907,9 +2928,15 @@ struct llama_model {
     // for quantize-stats only
     std::vector<std::pair<std::string, struct ggml_tensor *>> tensors_by_name;
 
-    int64_t t_load_us = 0;
+    int64_t t_load_us  = 0;
     int64_t t_start_us = 0;
 
+    // total number of parameters in the model
+    uint64_t n_elements = 0;
+
+    // total size of all the tensors in the model in bytes
+    size_t  n_bytes     = 0;
+
     // keep track of loaded lora adapters
     std::set<struct llama_lora_adapter *> lora_adapters;
 
@@ -3454,21 +3481,13 @@ static bool llama_kv_cache_init(
         const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa(i) + hparams.n_embd_k_s();
         const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(i) + hparams.n_embd_v_s();
 
-        const llama_model::buft_list_t * buft_list;
+        ggml_backend_buffer_type_t buft;
         if (offload) {
-            buft_list = model.dev_layer.at(i).buft_list;
+            auto * dev = model.dev_layer.at(i).dev;
+            buft = ggml_backend_dev_buffer_type(dev);
         } else {
-            buft_list = &model.cpu_buft_list;
+            buft = ggml_backend_cpu_buffer_type();
         }
-        ggml_backend_buffer_type_t buft = select_buft(*buft_list,
-            [&](ggml_context * ctx) {
-                ggml_tensor * k = ggml_new_tensor_1d(ctx, type_k, n_embd_k_gqa*kv_size);
-                if (hparams.rope_type == LLAMA_ROPE_TYPE_NONE) {
-                    return k;
-                }
-                ggml_tensor * p = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, 1);
-                return ggml_rope(ctx, k, p, hparams.n_rot, hparams.rope_type);
-            });
         ggml_context * ctx = ctx_for_buft(buft);
 
         if (!ctx) {
@@ -4275,8 +4294,8 @@ struct llama_model_loader {
     int n_tensors = 0;
     int n_created = 0;
 
-    int64_t n_elements = 0;
-    size_t  n_bytes    = 0;
+    uint64_t n_elements = 0;
+    size_t  n_bytes     = 0;
 
     bool use_mmap = false;
     bool check_tensors;
@@ -5344,6 +5363,11 @@ static const char * llama_model_vocab_type_name(enum llama_vocab_type type){
     }
 }
 
+static void llm_load_stats(llama_model_loader & ml, llama_model & model) {
+    model.n_elements = ml.n_elements;
+    model.n_bytes = ml.n_bytes;
+}
+
 static void llm_load_arch(llama_model_loader & ml, llama_model & model) {
     model.arch = ml.get_arch();
     if (model.arch == LLM_ARCH_UNKNOWN) {
@@ -5874,6 +5898,17 @@ static void llm_load_hparams(
                     default: model.type = e_model::MODEL_UNKNOWN;
                 }
             } break;
+        case LLM_ARCH_OLMO_1124:
+            {
+                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+
+                switch (hparams.n_layer) {
+                    case 16: model.type = e_model::MODEL_1B; break;
+                    case 32: model.type = e_model::MODEL_7B; break;
+                    case 40: model.type = e_model::MODEL_13B; break;
+                    default: model.type = e_model::MODEL_UNKNOWN;
+                }
+            } break;
         case LLM_ARCH_OLMOE:
             {
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
@@ -7254,7 +7289,7 @@ static llama_model::buft_list_t make_cpu_buft_list(llama_model & model) {
     auto * cpu_dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU);
     auto * cpu_reg = ggml_backend_dev_backend_reg(cpu_dev);
     auto ggml_backend_dev_get_extra_bufts_fn = (ggml_backend_dev_get_extra_bufts_t)
-        ggml_backend_reg_get_proc_address(cpu_reg, "ggml_backend_cpu_get_extra_bufts");
+        ggml_backend_reg_get_proc_address(cpu_reg, "ggml_backend_dev_get_extra_bufts");
     if (ggml_backend_dev_get_extra_bufts_fn) {
         ggml_backend_buffer_type_t * extra_bufts = ggml_backend_dev_get_extra_bufts_fn(cpu_dev);
         while (extra_bufts && *extra_bufts) {
@@ -7521,7 +7556,7 @@ static bool llm_load_tensors(
 
             // avoid using a host buffer when using mmap
             auto * buft_dev = ggml_backend_buft_get_device(buft);
-            if (ml.use_mmap && buft == ggml_backend_dev_host_buffer_type(buft_dev)) {
+            if (ml.use_mmap && buft_dev && buft == ggml_backend_dev_host_buffer_type(buft_dev)) {
                 auto * cpu_dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU);
                 buft = ggml_backend_dev_buffer_type(cpu_dev);
             }
@@ -8556,6 +8591,31 @@ static bool llm_load_tensors(
                         layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
                     }
                 } break;
+            case LLM_ARCH_OLMO_1124:
+                {
+                    model.tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+
+                    // output
+                    model.output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+                    model.output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, 0);
+
+                    for (int i = 0; i < n_layer; ++i) {
+                        auto & layer = model.layers[i];
+
+                        layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q,   "weight", i), {n_embd, n_embd}, 0);
+                        layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K,   "weight", i), {n_embd, n_embd_gqa}, 0);
+                        layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V,   "weight", i), {n_embd, n_embd_gqa}, 0);
+                        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);
+                        layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd}, 0);
+                        layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd}, 0);
+                        layer.attn_post_norm = create_tensor(tn(LLM_TENSOR_ATTN_POST_NORM, "weight", i), {n_embd}, 0);
+
+                        layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);
+                        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
+                        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);
+                        layer.ffn_post_norm = create_tensor(tn(LLM_TENSOR_FFN_POST_NORM, "weight", i), {n_embd}, 0);
+                    }
+                } break;
             case LLM_ARCH_OLMOE:
                 {
                     model.tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
@@ -9128,6 +9188,10 @@ static bool llm_load_tensors(
 
         // check if it is possible to use buffer_from_host_ptr with this buffer type
         ggml_backend_dev_t dev = ggml_backend_buft_get_device(buft);
+        if (!dev) {
+            // FIXME: workaround for CPU backend buft having a NULL device
+            dev = ggml_backend_reg_dev_get(ggml_backend_cpu_reg(), 0);
+        }
         ggml_backend_dev_props props;
         ggml_backend_dev_get_props(dev, &props);
         bool buffer_from_host_ptr_supported = props.caps.buffer_from_host_ptr;
@@ -9252,6 +9316,7 @@ static int llama_model_load(const std::string & fname, llama_model & model, llam
             throw std::runtime_error("error loading model vocabulary: " + std::string(e.what()));
         }
 
+        llm_load_stats(ml, model);
         llm_load_print_meta(ml, model);
 
         if (model.vocab.type != LLAMA_VOCAB_TYPE_NONE &&
@@ -14416,6 +14481,130 @@ struct llm_build_context {
         return gf;
     }
 
+    struct ggml_cgraph * build_olmo_1124() {
+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);
+
+        // mutable variable, needed during the last layer of the computation to skip unused tokens
+        int32_t n_tokens = this->n_tokens;
+
+        const int64_t n_embd_head = hparams.n_embd_head_v;
+        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+        GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+        struct ggml_tensor * cur;
+        struct ggml_tensor * inpL;
+
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
+
+        // inp_pos - contains the positions
+        struct ggml_tensor * inp_pos = build_inp_pos();
+
+        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
+        struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
+
+        for (int il = 0; il < n_layer; ++il) {
+            struct ggml_tensor * inpSA = inpL;
+
+            cur = inpL;
+
+            // self_attention
+            {
+                // compute Q and K and RoPE them
+                struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur);
+                cb(Qcur, "Qcur", il);
+
+                struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur);
+                cb(Kcur, "Kcur", il);
+
+                struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur);
+                cb(Vcur, "Vcur", il);
+
+                Qcur = llm_build_norm(ctx0, Qcur, hparams, model.layers[il].attn_q_norm, NULL,
+                        LLM_NORM_RMS, cb, il);
+                cb(Qcur, "Qcur_normed", il);
+
+                Kcur = llm_build_norm(ctx0, Kcur, hparams, model.layers[il].attn_k_norm, NULL,
+                        LLM_NORM_RMS, cb, il);
+                cb(Kcur, "Kcur_normed", il);
+
+                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
+                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+
+                Qcur = ggml_rope_ext(
+                    ctx0, Qcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                );
+                cb(Qcur, "Qcur_rope", il);
+
+                Kcur = ggml_rope_ext(
+                    ctx0, Kcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                );
+                cb(Kcur, "Kcur_rope", il);
+
+                cur = llm_build_kv(ctx0, lctx, kv_self, gf,
+                        model.layers[il].wo, NULL,
+                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+            }
+
+            cur = llm_build_norm(ctx0, cur, hparams,
+                    model.layers[il].attn_post_norm, NULL,
+                    LLM_NORM_RMS, cb, il);
+            cb(cur, "attn_post_norm", il);
+
+            if (il == n_layer - 1) {
+                // skip computing output for unused tokens
+                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
+                n_tokens = n_outputs;
+                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+            }
+
+            struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+            cb(ffn_inp, "ffn_inp", il);
+
+            // feed-forward network
+            cur = llm_build_ffn(ctx0, lctx, ffn_inp,
+                    model.layers[il].ffn_up,   NULL, NULL,
+                    model.layers[il].ffn_gate, NULL, NULL,
+                    model.layers[il].ffn_down, NULL, NULL,
+                    NULL,
+                    LLM_FFN_SILU, LLM_FFN_PAR, cb, il);
+            cb(cur, "ffn_out", il);
+
+            cur = llm_build_norm(ctx0, cur, hparams,
+                model.layers[il].ffn_post_norm, NULL,
+                LLM_NORM_RMS, cb, -1);
+            cb(cur, "ffn_post_norm", -1);
+
+            cur = ggml_add(ctx0, cur, ffn_inp);
+            cb(cur, "ffn_out", il);
+
+            cur = lctx.cvec.apply_to(ctx0, cur, il);
+            cb(cur, "l_out", il);
+
+            // input for next layer
+            inpL = cur;
+        }
+
+        cur = inpL;
+
+        cur = llm_build_norm(ctx0, cur, hparams,
+                model.output_norm, NULL,
+                LLM_NORM_RMS, cb, -1);
+        cb(cur, "result_norm", -1);
+
+        // lm_head
+        cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
+        cb(cur, "result_output", -1);
+
+        ggml_build_forward_expand(gf, cur);
+
+        return gf;
+    }
+
     // based on the build_qwen2moe() function, changes:
     //   * removed shared experts
     //   * removed bias
@@ -16608,6 +16797,10 @@ static struct ggml_cgraph * llama_build_graph(
             {
                 result = llm.build_olmo();
             } break;
+        case LLM_ARCH_OLMO_1124:
+            {
+                result = llm.build_olmo_1124();
+            } break;
         case LLM_ARCH_OLMOE:
             {
                 result = llm.build_olmoe();
@@ -18020,7 +18213,7 @@ static void llama_kv_cache_update_internal(struct llama_context & lctx) {
 
     // apply K-shift if needed
     if (lctx.model.hparams.rope_type != LLAMA_ROPE_TYPE_NONE && lctx.kv_self.has_shift) {
-        if (lctx.model.arch == LLM_ARCH_DEEPSEEK2) { // not supported due to MLA
+        if (!llama_kv_cache_can_shift(&lctx)) {
             GGML_ABORT("Deepseek2 does not support K-shift");
         }
 
@@ -18597,6 +18790,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
     llama_model model;
     llm_load_arch(ml, model);
     llm_load_hparams(ml, model);
+    llm_load_stats(ml, model);
 
     struct quantize_state_internal qs(model, params);
 
@@ -19876,6 +20070,7 @@ enum llama_rope_type llama_rope_type(const struct llama_model * model) {
         case LLM_ARCH_QWEN:
         case LLM_ARCH_QWEN2:
         case LLM_ARCH_QWEN2MOE:
+        case LLM_ARCH_OLMO_1124:
         case LLM_ARCH_OLMOE:
         case LLM_ARCH_PHI2:
         case LLM_ARCH_PHI3:
@@ -19949,19 +20144,11 @@ int32_t llama_model_desc(const struct llama_model * model, char * buf, size_t bu
 }
 
 uint64_t llama_model_size(const struct llama_model * model) {
-    uint64_t size = 0;
-    for (const auto & it : model->tensors_by_name) {
-        size += ggml_nbytes(it.second);
-    }
-    return size;
+    return model->n_bytes;
 }
 
 uint64_t llama_model_n_params(const struct llama_model * model) {
-    uint64_t nparams = 0;
-    for (const auto & it : model->tensors_by_name) {
-        nparams += ggml_nelements(it.second);
-    }
-    return nparams;
+    return model->n_elements;
 }
 
 struct ggml_tensor * llama_get_model_tensor(struct llama_model * model, const char * name) {
@@ -20275,6 +20462,10 @@ void llama_kv_cache_update(struct llama_context * ctx) {
     llama_kv_cache_update_internal(*ctx);
 }
 
+bool llama_kv_cache_can_shift(struct llama_context * ctx) {
+    return ctx->model.arch != LLM_ARCH_DEEPSEEK2; // not supported due to MLA
+}
+
 // deprecated
 size_t llama_get_state_size(struct llama_context * ctx) {
     return llama_state_get_size(ctx);
@@ -22021,7 +22212,6 @@ const char * llama_print_system_info(void) {
     s += "FP16_VA = "     + std::to_string(ggml_cpu_has_fp16_va())     + " | ";
     s += "RISCV_VECT = "  + std::to_string(ggml_cpu_has_riscv_v())     + " | ";
     s += "WASM_SIMD = "   + std::to_string(ggml_cpu_has_wasm_simd())   + " | ";
-    s += "BLAS = "        + std::to_string(ggml_cpu_has_blas())        + " | ";
     s += "SSE3 = "        + std::to_string(ggml_cpu_has_sse3())        + " | ";
     s += "SSSE3 = "       + std::to_string(ggml_cpu_has_ssse3())       + " | ";
     s += "VSX = "         + std::to_string(ggml_cpu_has_vsx())         + " | ";
@@ -22067,28 +22257,6 @@ void llama_perf_context_reset(struct llama_context * ctx) {
     ctx->t_p_eval_us = ctx->n_p_eval = 0;
 }
 
-void llama_perf_dump_yaml(FILE * stream, const llama_context * ctx) {
-    fprintf(stream, "\n");
-    fprintf(stream, "###########\n");
-    fprintf(stream, "# Timings #\n");
-    fprintf(stream, "###########\n");
-    fprintf(stream, "\n");
-
-    fprintf(stream, "mst_eval: %.2f  # ms / token during generation\n",
-            1.0e-3 * ctx->t_eval_us / ctx->n_eval);
-    fprintf(stream, "mst_p_eval: %.2f  # ms / token during prompt processing\n",
-            1.0e-3 * ctx->t_p_eval_us / ctx->n_p_eval);
-    fprintf(stream, "n_eval: %d  # number of tokens generated (excluding the first one)\n", ctx->n_eval);
-    fprintf(stream, "n_p_eval: %d  # number of tokens processed in batches at the beginning\n", ctx->n_p_eval);
-    fprintf(stream, "t_eval_us: %" PRId64 "  # total microseconds spent generating tokens\n", ctx->t_eval_us);
-    fprintf(stream, "t_load_us: %" PRId64 "  # total microseconds spent loading the model\n", ctx->t_load_us);
-    fprintf(stream, "t_p_eval_us: %" PRId64 "  # total microseconds spent prompt processing\n", ctx->t_p_eval_us);
-    fprintf(stream, "ts_eval: %.2f  # tokens / second during generation\n",
-            1.0e6 * ctx->n_eval / ctx->t_eval_us);
-    fprintf(stream, "ts_p_eval: %.2f  # tokens / second during prompt processing\n",
-            1.0e6 * ctx->n_p_eval / ctx->t_p_eval_us);
-}
-
 // For internal test use
 const std::vector<std::pair<std::string, struct ggml_tensor *>> & llama_internal_get_tensor_map(
     struct llama_context * ctx

examples/talk-llama/llama.h

@@ -667,6 +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
     //
@@ -1244,8 +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/whisper.android/lib/src/main/jni/whisper/CMakeLists.txt

@@ -19,11 +19,16 @@ if (NOT GGML_HOME)
         SOURCE_FILES
         ${SOURCE_FILES}
         ${WHISPER_LIB_DIR}/ggml/src/ggml.c
-        ${WHISPER_LIB_DIR}/ggml/src/ggml-cpu.c
         ${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

@@ -25,6 +25,11 @@
 		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 */; };
@@ -50,8 +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>"; };
@@ -77,8 +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.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>"; };
@@ -118,6 +132,15 @@
 		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 */,
@@ -252,11 +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 */,
 			);
@@ -335,6 +363,8 @@
 				GCC_WARN_UNINITIALIZED_AUTOS = YES_AGGRESSIVE;
 				GCC_WARN_UNUSED_FUNCTION = YES;
 				GCC_WARN_UNUSED_VARIABLE = YES;
+				GENERATE_INFOPLIST_FILE = YES;
+				HEADER_SEARCH_PATHS = ../../../ggml/src/;
 				IPHONEOS_DEPLOYMENT_TARGET = 16.0;
 				MTL_ENABLE_DEBUG_INFO = INCLUDE_SOURCE;
 				MTL_FAST_MATH = YES;
@@ -388,6 +418,8 @@
 				GCC_WARN_UNINITIALIZED_AUTOS = YES_AGGRESSIVE;
 				GCC_WARN_UNUSED_FUNCTION = YES;
 				GCC_WARN_UNUSED_VARIABLE = YES;
+				GENERATE_INFOPLIST_FILE = YES;
+				HEADER_SEARCH_PATHS = ../../../ggml/src/;
 				IPHONEOS_DEPLOYMENT_TARGET = 16.0;
 				MTL_ENABLE_DEBUG_INFO = NO;
 				MTL_FAST_MATH = YES;
@@ -400,64 +432,6 @@
 			};
 			name = Release;
 		};
-		18627C9029052BE000BD2A04 /* Debug */ = {
-			isa = XCBuildConfiguration;
-			buildSettings = {
-				ASSETCATALOG_COMPILER_APPICON_NAME = AppIcon;
-				ASSETCATALOG_COMPILER_GLOBAL_ACCENT_COLOR_NAME = AccentColor;
-				CODE_SIGN_STYLE = Automatic;
-				CURRENT_PROJECT_VERSION = 1;
-				DEVELOPMENT_TEAM = P8JZH34X63;
-				GCC_WARN_64_TO_32_BIT_CONVERSION = NO;
-				GENERATE_INFOPLIST_FILE = YES;
-				INFOPLIST_FILE = whisper.objc/Info.plist;
-				INFOPLIST_KEY_UIApplicationSupportsIndirectInputEvents = YES;
-				INFOPLIST_KEY_UILaunchStoryboardName = LaunchScreen;
-				INFOPLIST_KEY_UIMainStoryboardFile = Main;
-				INFOPLIST_KEY_UISupportedInterfaceOrientations_iPad = "UIInterfaceOrientationPortrait UIInterfaceOrientationPortraitUpsideDown UIInterfaceOrientationLandscapeLeft UIInterfaceOrientationLandscapeRight";
-				INFOPLIST_KEY_UISupportedInterfaceOrientations_iPhone = "UIInterfaceOrientationPortrait UIInterfaceOrientationLandscapeLeft UIInterfaceOrientationLandscapeRight";
-				LD_RUNPATH_SEARCH_PATHS = (
-					"$(inherited)",
-					"@executable_path/Frameworks",
-				);
-				MARKETING_VERSION = 1.0;
-				MTL_HEADER_SEARCH_PATHS = "";
-				PRODUCT_BUNDLE_IDENTIFIER = "com.ggerganov.whisper-objc";
-				PRODUCT_NAME = "$(TARGET_NAME)";
-				SWIFT_EMIT_LOC_STRINGS = YES;
-				TARGETED_DEVICE_FAMILY = "1,2";
-			};
-			name = Debug;
-		};
-		18627C9129052BE000BD2A04 /* Release */ = {
-			isa = XCBuildConfiguration;
-			buildSettings = {
-				ASSETCATALOG_COMPILER_APPICON_NAME = AppIcon;
-				ASSETCATALOG_COMPILER_GLOBAL_ACCENT_COLOR_NAME = AccentColor;
-				CODE_SIGN_STYLE = Automatic;
-				CURRENT_PROJECT_VERSION = 1;
-				DEVELOPMENT_TEAM = P8JZH34X63;
-				GCC_WARN_64_TO_32_BIT_CONVERSION = NO;
-				GENERATE_INFOPLIST_FILE = YES;
-				INFOPLIST_FILE = whisper.objc/Info.plist;
-				INFOPLIST_KEY_UIApplicationSupportsIndirectInputEvents = YES;
-				INFOPLIST_KEY_UILaunchStoryboardName = LaunchScreen;
-				INFOPLIST_KEY_UIMainStoryboardFile = Main;
-				INFOPLIST_KEY_UISupportedInterfaceOrientations_iPad = "UIInterfaceOrientationPortrait UIInterfaceOrientationPortraitUpsideDown UIInterfaceOrientationLandscapeLeft UIInterfaceOrientationLandscapeRight";
-				INFOPLIST_KEY_UISupportedInterfaceOrientations_iPhone = "UIInterfaceOrientationPortrait UIInterfaceOrientationLandscapeLeft UIInterfaceOrientationLandscapeRight";
-				LD_RUNPATH_SEARCH_PATHS = (
-					"$(inherited)",
-					"@executable_path/Frameworks",
-				);
-				MARKETING_VERSION = 1.0;
-				MTL_HEADER_SEARCH_PATHS = "";
-				PRODUCT_BUNDLE_IDENTIFIER = "com.ggerganov.whisper-objc";
-				PRODUCT_NAME = "$(TARGET_NAME)";
-				SWIFT_EMIT_LOC_STRINGS = YES;
-				TARGETED_DEVICE_FAMILY = "1,2";
-			};
-			name = Release;
-		};
 /* End XCBuildConfiguration section */
 
 /* Begin XCConfigurationList section */
@@ -470,15 +444,6 @@
 			defaultConfigurationIsVisible = 0;
 			defaultConfigurationName = Release;
 		};
-		18627C8F29052BE000BD2A04 /* Build configuration list for PBXNativeTarget "whisper.objc" */ = {
-			isa = XCConfigurationList;
-			buildConfigurations = (
-				18627C9029052BE000BD2A04 /* Debug */,
-				18627C9129052BE000BD2A04 /* Release */,
-			);
-			defaultConfigurationIsVisible = 0;
-			defaultConfigurationName = Release;
-		};
 /* End XCConfigurationList section */
 	};
 	rootObject = 18627C6E29052BDF00BD2A04 /* Project object */;

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

@@ -67,8 +67,6 @@ actor WhisperContext {
     private func systemInfo() -> String {
         var info = ""
         if (ggml_cpu_has_neon() != 0) { info += "NEON " }
-        if (ggml_cpu_has_metal() != 0) { info += "METAL " }
-        if (ggml_cpu_has_blas() != 0) { info += "BLAS " }
         return String(info.dropLast())
     }
 

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})
@@ -116,6 +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)
@@ -126,14 +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)
@@ -141,7 +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)
@@ -167,6 +164,8 @@ option(GGML_SYCL                            "ggml: use SYCL"
 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"    ${GGML_STANDALONE})
@@ -226,6 +225,7 @@ set(GGML_PUBLIC_HEADERS
     include/ggml-cann.h
     include/ggml-cuda.h
     include/ggml-kompute.h
+    include/ggml-opt.h
     include/ggml-metal.h
     include/ggml-rpc.h
     include/ggml-sycl.h
@@ -235,15 +235,14 @@ set_target_properties(ggml PROPERTIES PUBLIC_HEADER "${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

@@ -9,16 +9,16 @@ extern "C" {
 #endif
 
 // buffer_type API
-GGML_API ggml_backend_buffer_type_t ggml_backend_amx_buffer_type(void);
+GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_amx_buffer_type(void);
 
-GGML_API bool ggml_backend_is_amx(ggml_backend_t backend);
+GGML_BACKEND_API bool ggml_backend_is_amx(ggml_backend_t backend);
 
 // backend API
-GGML_API ggml_backend_t ggml_backend_amx_init(void);
+GGML_BACKEND_API ggml_backend_t ggml_backend_amx_init(void);
 
-GGML_API void ggml_backend_amx_set_n_threads(ggml_backend_t backend_amx, int n_threads);
+GGML_BACKEND_API void ggml_backend_amx_set_n_threads(ggml_backend_t backend_amx, int n_threads);
 
-GGML_API ggml_backend_reg_t ggml_backend_amx_reg(void);
+GGML_BACKEND_API ggml_backend_reg_t ggml_backend_amx_reg(void);
 
 #ifdef  __cplusplus
 }

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);
@@ -228,14 +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
     }
     */
 
@@ -250,7 +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);
 
@@ -275,7 +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

ggml/include/ggml-blas.h

@@ -9,15 +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_API ggml_backend_reg_t ggml_backend_blas_reg(void);
+GGML_BACKEND_API ggml_backend_reg_t ggml_backend_blas_reg(void);
 
 
 #ifdef  __cplusplus

ggml/include/ggml-cann.h

@@ -34,7 +34,7 @@ extern "C" {
  */
 #define GGML_CANN_MAX_DEVICES 16
 
-GGML_API ggml_backend_reg_t ggml_backend_cann_reg(void);
+GGML_BACKEND_API ggml_backend_reg_t ggml_backend_cann_reg(void);
 
 /**
  * @brief Initializes the CANN backend for a specified device.
@@ -46,7 +46,7 @@ GGML_API ggml_backend_reg_t ggml_backend_cann_reg(void);
  * @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.
@@ -57,7 +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.
@@ -69,7 +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);
 
 /**
@@ -80,14 +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.
@@ -99,7 +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);
 
 /**
@@ -114,7 +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-cpu.h

@@ -54,54 +54,77 @@ extern "C" {
         GGML_NUMA_STRATEGY_COUNT
     };
 
-    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_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_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_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_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);
+    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_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_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_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_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_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_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_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_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_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_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_API struct ggml_cplan ggml_graph_plan(
+    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_API enum ggml_status  ggml_graph_compute(struct ggml_cgraph * cgraph, struct ggml_cplan * cplan);
+    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_API enum ggml_status  ggml_graph_compute_with_ctx(struct ggml_context * ctx, struct ggml_cgraph * cgraph, int n_threads);
+    GGML_BACKEND_API enum ggml_status  ggml_graph_compute_with_ctx(struct ggml_context * ctx, struct ggml_cgraph * cgraph, int n_threads);
 
-    // TODO: move to backend interface
-    GGML_API int ggml_cpu_has_neon       (void);
-    GGML_API int ggml_cpu_has_sve        (void);
-    GGML_API int ggml_cpu_has_matmul_int8(void);
-    // get the sve vector length in bytes
-    GGML_API int ggml_cpu_get_sve_cnt(void);
+    //
+    // 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
 
@@ -115,6 +138,7 @@ extern "C" {
                                        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;
@@ -124,27 +148,30 @@ extern "C" {
         ggml_gemm_t              gemm;
     };
 
-    GGML_API const struct ggml_type_traits_cpu * ggml_get_type_traits_cpu(enum ggml_type type);
+    GGML_BACKEND_API const struct ggml_type_traits_cpu * ggml_get_type_traits_cpu(enum ggml_type type);
 
-    GGML_API void ggml_cpu_init(void);
+    GGML_BACKEND_API void ggml_cpu_init(void);
 
     //
     // CPU backend
     //
 
-    GGML_API ggml_backend_t ggml_backend_cpu_init(void);
+    GGML_BACKEND_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);
+    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_API ggml_backend_reg_t ggml_backend_cpu_reg(void);
+    GGML_BACKEND_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);
+    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(int main_device, 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

@@ -37,13 +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_API ggml_backend_reg_t ggml_backend_kompute_reg(void);
+GGML_BACKEND_API ggml_backend_reg_t ggml_backend_kompute_reg(void);
 
 #ifdef __cplusplus
 }

ggml/include/ggml-metal.h

@@ -39,27 +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_DEPRECATED(
-        GGML_API ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void * data, size_t size, size_t max_size),
+        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_API ggml_backend_reg_t ggml_backend_metal_reg(void);
+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

@@ -10,18 +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 ggml_backend_rpc_start_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_API ggml_backend_reg_t ggml_backend_rpc_reg(void);
+GGML_BACKEND_API ggml_backend_reg_t ggml_backend_rpc_reg(void);
 
-GGML_API ggml_backend_dev_t ggml_backend_rpc_add_device(const char * endpoint);
+GGML_BACKEND_API ggml_backend_dev_t ggml_backend_rpc_add_device(const char * endpoint);
 
 #ifdef  __cplusplus
 }

ggml/include/ggml-sycl.h

@@ -17,32 +17,32 @@ 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_API bool ggml_backend_is_sycl(ggml_backend_t backend);
+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_backend_sycl_get_gpu_list(int *id_list, int max_len);
-GGML_API void ggml_backend_sycl_get_device_description(int device,
+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_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 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_API ggml_backend_reg_t ggml_backend_sycl_reg(void);
+GGML_BACKEND_API ggml_backend_reg_t ggml_backend_sycl_reg(void);
 
 #ifdef  __cplusplus
 }

ggml/include/ggml-vulkan.h

@@ -10,21 +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_API ggml_backend_reg_t ggml_backend_vk_reg(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
@@ -602,7 +602,6 @@ extern "C" {
 
         int32_t flags;
 
-        struct ggml_tensor * grad;
         struct ggml_tensor * src[GGML_MAX_SRC];
 
         // source tensor and offset for views
@@ -615,7 +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);
@@ -1490,7 +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
@@ -1985,28 +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_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
@@ -2026,7 +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_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);
@@ -2037,198 +2030,15 @@ 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_tensor * ggml_set_zero(struct ggml_tensor * tensor);
 
-    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);
-
     //
     // quantization
     //
@@ -2384,38 +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_amx_int8   (void);
-    GGML_API int ggml_cpu_has_fma        (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_cann       (void);
-    GGML_API int ggml_cpu_has_llamafile  (void);
-
 #ifdef  __cplusplus
 // restrict not standard in C++
 #define GGML_RESTRICT
@@ -2432,7 +2210,6 @@ extern "C" {
         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/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

@@ -466,18 +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)) {
@@ -816,7 +810,11 @@ 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;
 }
 

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

@@ -1,7 +1,8 @@
 #pragma once
 
 #include "ggml.h"
-#include "ggml-cpu-impl.h" // <immintrin.h>
+// hack until AMX is moved into the CPU backend
+#include "../ggml-cpu/ggml-cpu-impl.h" // <immintrin.h>
 
 #include <algorithm>
 #include <memory>

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

@@ -317,8 +317,6 @@ static bool ggml_backend_amx_device_supports_op(ggml_backend_dev_t dev, const st
             const enum ggml_type type = src0->type;
             const int64_t ne0 = op->ne[0];
 
-            bool is_training = src0->grad || src1->grad;
-
             // 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);
@@ -326,7 +324,6 @@ static bool ggml_backend_amx_device_supports_op(ggml_backend_dev_t dev, const st
             bool can_use_amx =
                 is_contiguous_2d(src0) &&       // src0 must be contiguous
                 is_contiguous_2d(src1) &&       // src1 must be contiguous
-                !is_training &&                 // inference only
                 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
@@ -421,9 +418,18 @@ ggml_backend_reg_t ggml_backend_amx_reg(void) {
 
 #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 ggml_backend_t{};
+    return nullptr;
 }
 
 void ggml_backend_amx_set_n_threads(ggml_backend_t backend_amx, int n_threads) {
@@ -433,4 +439,8 @@ void ggml_backend_amx_set_n_threads(ggml_backend_t backend_amx, int n_threads) {
     GGML_UNUSED(n_threads);
 }
 
+ggml_backend_reg_t ggml_backend_amx_reg(void) {
+    return nullptr;
+}
+
 #endif

ggml/src/ggml-amx/mmq.cpp

@@ -496,19 +496,20 @@ inline void from_float(const float * x, char * vy, int64_t k);
 
 template <>
 inline void from_float<block_q8_0>(const float * x, char * vy, int64_t k) {
-    quantize_row_q8_0(x, vy, k);
+    // FIXME: using unoptimized reference impl until moved to CPU backend
+    quantize_row_q8_0_ref(x, (block_q8_0 *)vy, k);
 }
 
 template <>
 inline void from_float<block_q8_1>(const float * x, char * vy, int64_t k) {
-    quantize_row_q8_1(x, vy, k);
+    quantize_row_q8_1_ref(x, (block_q8_1 *)vy, k);
 }
 
 template <>
 inline void from_float<block_q8_K>(const float * x, char * vy, int64_t k) {
 #if 1
     // TODO: this is reference impl!
-    quantize_row_q8_K(x, vy, k);
+    quantize_row_q8_K_ref(x, (block_q8_K *)vy, k);
 #else
     quantize_row_q8_K_vnni(x, vy, k);
 #endif

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-backend.cpp

@@ -279,7 +279,7 @@ void ggml_backend_tensor_get(const struct ggml_tensor * tensor, void * data, siz
     buf->iface.get_tensor(buf, tensor, data, offset, size);
 }
 
-GGML_API void ggml_backend_tensor_memset(struct ggml_tensor * tensor, uint8_t value, size_t offset, size_t size) {
+void ggml_backend_tensor_memset(struct ggml_tensor * tensor, uint8_t value, size_t offset, size_t size) {
     ggml_backend_buffer_t buf = tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
 
     if (size == 0) {
@@ -525,197 +525,6 @@ void * ggml_backend_reg_get_proc_address(ggml_backend_reg_t reg, const char * na
     return reg->iface.get_proc_address(reg, name);
 }
 
-// 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
-
-#ifndef __AMX_INT8__
-#undef GGML_USE_AMX
-#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
-
-#include "ggml-cpu.h"
-
-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) {
-#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 (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);
-}
-
 // multi-buffer buffer
 
 struct ggml_backend_multi_buffer_context {
@@ -880,7 +689,7 @@ static int ggml_backend_sched_backend_id(ggml_backend_sched_t sched, ggml_backen
 }
 
 static int ggml_backend_sched_backend_from_buffer(ggml_backend_sched_t sched, const struct ggml_tensor * tensor, const struct ggml_tensor * op) {
-    ggml_backend_buffer_t buffer = tensor->buffer;
+    ggml_backend_buffer_t buffer = tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
     if (buffer == NULL) {
         return -1;
     }
@@ -913,8 +722,6 @@ static char causes[GGML_DEFAULT_GRAPH_SIZE*16 + GGML_SCHED_MAX_SPLITS_DEBUG*GGML
 
 // returns the backend that should be used for the node based on the current locations
 static int ggml_backend_sched_backend_id_from_cur(ggml_backend_sched_t sched, struct ggml_tensor * tensor) {
-    // TODO: use supports_op to check if the backend supports the op
-
     // assign pre-allocated nodes to their backend
     int cur_backend_id = ggml_backend_sched_backend_from_buffer(sched, tensor, tensor);
     if (cur_backend_id != -1) {
@@ -933,7 +740,7 @@ static int ggml_backend_sched_backend_id_from_cur(ggml_backend_sched_t sched, st
 
     if (tensor->buffer || (tensor->view_src && tensor->view_src->buffer)) {
         // since the tensor is pre-allocated, it cannot be moved to another backend
-        GGML_ABORT("pre-allocated tensor in a backend that cannot run the operation");
+        GGML_ABORT("pre-allocated tensor (%s) in a backend that cannot run the operation", tensor->name);
     }
 
     // graph input
@@ -1640,7 +1447,7 @@ ggml_backend_sched_t ggml_backend_sched_new(
         bool parallel) {
     GGML_ASSERT(n_backends > 0);
     GGML_ASSERT(n_backends <= GGML_SCHED_MAX_BACKENDS);
-    GGML_ASSERT(ggml_backend_is_cpu(backends[n_backends - 1])); // last backend must be CPU
+    GGML_ASSERT(ggml_backend_dev_type(ggml_backend_get_device(backends[n_backends - 1])) == GGML_BACKEND_DEVICE_TYPE_CPU);
 
     struct ggml_backend_sched * sched = (ggml_backend_sched *) calloc(1, sizeof(struct ggml_backend_sched));
 
@@ -1729,12 +1536,13 @@ bool ggml_backend_sched_reserve(ggml_backend_sched_t sched, struct ggml_cgraph *
 
     ggml_backend_sched_split_graph(sched, measure_graph);
 
+    ggml_backend_sched_synchronize(sched);
+
     if (!ggml_gallocr_reserve_n(sched->galloc, &sched->graph, sched->node_backend_ids, sched->leaf_backend_ids)) {
         return false;
     }
 
     ggml_backend_sched_reset(sched);
-    ggml_backend_sched_synchronize(sched);
 
     return true;
 }
@@ -2036,17 +1844,6 @@ bool ggml_backend_compare_graph_backend(ggml_backend_t backend1, ggml_backend_t
     return true;
 }
 
-
-
-#include "ggml-backend.h"
-#include "ggml-backend-impl.h"
-#include "ggml-cpu.h"
-#include "ggml-impl.h"
-#include <cctype>
-#include <string>
-
-// ggml-backend interface
-
 // CPU backend - buffer
 
 static void * ggml_backend_cpu_buffer_get_base(ggml_backend_buffer_t buffer) {
@@ -2120,7 +1917,9 @@ static const struct ggml_backend_buffer_i ggml_backend_cpu_buffer_from_ptr_i = {
     /* .reset           = */ NULL,
 };
 
-// CPU backend - buffer type
+// CPU backend buffer type
+
+// this buffer type is defined here to make it available to all backends
 
 static const char * ggml_backend_cpu_buffer_type_get_name(ggml_backend_buffer_type_t buft) {
     return "CPU";
@@ -2161,7 +1960,7 @@ ggml_backend_buffer_type_t ggml_backend_cpu_buffer_type(void) {
             /* .get_alloc_size   = */ NULL, // defaults to ggml_nbytes
             /* .is_host          = */ ggml_backend_cpu_buffer_type_is_host,
         },
-        /* .device  = */ ggml_backend_reg_dev_get(ggml_backend_cpu_reg(), 0),
+        /* .device  = */ NULL, // FIXME ggml_backend_reg_dev_get(ggml_backend_cpu_reg(), 0),
         /* .context = */ NULL,
     };
 
@@ -2184,478 +1983,14 @@ static ggml_backend_buffer_type_t ggml_backend_cpu_buffer_from_ptr_type(void) {
             /* .get_alloc_size   = */ NULL, // defaults to ggml_nbytes
             /* .is_host          = */ ggml_backend_cpu_buffer_type_is_host,
         },
-        /* .device  = */ ggml_backend_reg_dev_get(ggml_backend_cpu_reg(), 0),
+        /* .device  = */ NULL, // FIXME ggml_backend_reg_dev_get(ggml_backend_cpu_reg(), 0),
         /* .context = */ NULL,
     };
 
     return &ggml_backend_cpu_buffer_type;
 }
 
-#ifdef GGML_USE_CPU_HBM
-
-// buffer type HBM
-
-#include <hbwmalloc.h>
-
-static const char * ggml_backend_cpu_hbm_buffer_type_get_name(ggml_backend_buffer_type_t buft) {
-    return "CPU_HBM";
-
-    GGML_UNUSED(buft);
-}
-
-static void ggml_backend_cpu_hbm_buffer_free_buffer(ggml_backend_buffer_t buffer) {
-    hbw_free(buffer->context);
-}
-
-static ggml_backend_buffer_t ggml_backend_cpu_hbm_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
-    void * ptr;
-    int result = hbw_posix_memalign(&ptr, ggml_backend_cpu_buffer_type_get_alignment(buft), size);
-    if (result != 0) {
-        GGML_LOG_ERROR("failed to allocate HBM buffer of size %zu\n", size);
-        return NULL;
-    }
-
-    ggml_backend_buffer_t buffer = ggml_backend_cpu_buffer_from_ptr(ptr, size);
-    buffer->buft = buft;
-    buffer->iface.free_buffer = ggml_backend_cpu_hbm_buffer_free_buffer;
-
-    return buffer;
-}
-
-ggml_backend_buffer_type_t ggml_backend_cpu_hbm_buffer_type(void) {
-    static struct ggml_backend_buffer_type ggml_backend_cpu_buffer_type_hbm = {
-        /* .iface    = */ {
-            /* .get_name         = */ ggml_backend_cpu_hbm_buffer_type_get_name,
-            /* .alloc_buffer     = */ ggml_backend_cpu_hbm_buffer_type_alloc_buffer,
-            /* .get_alignment    = */ ggml_backend_cpu_buffer_type_get_alignment,
-            /* .get_max_size     = */ NULL, // defaults to SIZE_MAX
-            /* .get_alloc_size   = */ NULL, // defaults to ggml_nbytes
-            /* .is_host          = */ ggml_backend_cpu_buffer_type_is_host,
-        },
-        /* .context  = */ NULL,
-    };
-
-    return &ggml_backend_cpu_buffer_type_hbm;
-}
-#endif
-
-static ggml_backend_buffer_type_t * ggml_backend_cpu_get_extra_bufts(ggml_backend_dev_t device) {
-    static ggml_backend_buffer_type_t bufts[] = {
-#ifdef GGML_USE_CPU_HBM
-        ggml_backend_cpu_hbm_buffer_type(),
-#endif
-        NULL
-    };
-
-    return bufts;
-
-    GGML_UNUSED(device);
-}
-
-// CPU backend - backend (stream)
-
-struct ggml_backend_cpu_context {
-    int                 n_threads;
-    ggml_threadpool_t   threadpool;
-
-    uint8_t *           work_data;
-    size_t              work_size;
-
-    ggml_abort_callback abort_callback;
-    void *              abort_callback_data;
-};
-
-static const char * ggml_backend_cpu_get_name(ggml_backend_t backend) {
-    return "CPU";
-
-    GGML_UNUSED(backend);
-}
-
-static void ggml_backend_cpu_free(ggml_backend_t backend) {
-    struct ggml_backend_cpu_context * cpu_ctx = (struct ggml_backend_cpu_context *)backend->context;
-    delete[] cpu_ctx->work_data;
-    delete cpu_ctx;
-    delete backend;
-}
-
-struct ggml_backend_plan_cpu {
-    struct ggml_cplan cplan;
-    struct ggml_cgraph cgraph;
-};
-
-static ggml_backend_graph_plan_t ggml_backend_cpu_graph_plan_create(ggml_backend_t backend, const struct ggml_cgraph * cgraph) {
-    struct ggml_backend_cpu_context * cpu_ctx = (struct ggml_backend_cpu_context *)backend->context;
-
-    struct ggml_backend_plan_cpu * cpu_plan = new ggml_backend_plan_cpu;
-
-    cpu_plan->cplan = ggml_graph_plan(cgraph, cpu_ctx->n_threads, cpu_ctx->threadpool);
-    cpu_plan->cgraph = *cgraph; // FIXME: deep copy
-
-    if (cpu_plan->cplan.work_size > 0) {
-        cpu_plan->cplan.work_data = new uint8_t[cpu_plan->cplan.work_size];
-        if (cpu_plan->cplan.work_data == NULL) {
-            delete cpu_plan;
-            return NULL;
-        }
-    }
-
-    cpu_plan->cplan.abort_callback      = cpu_ctx->abort_callback;
-    cpu_plan->cplan.abort_callback_data = cpu_ctx->abort_callback_data;
-
-    return cpu_plan;
-}
-
-static void ggml_backend_cpu_graph_plan_free(ggml_backend_t backend, ggml_backend_graph_plan_t plan) {
-    struct ggml_backend_plan_cpu * cpu_plan = (struct ggml_backend_plan_cpu *)plan;
-
-    delete[] cpu_plan->cplan.work_data;
-    delete cpu_plan;
-
-    GGML_UNUSED(backend);
-}
-
-static enum ggml_status ggml_backend_cpu_graph_plan_compute(ggml_backend_t backend, ggml_backend_graph_plan_t plan) {
-    struct ggml_backend_plan_cpu * cpu_plan = (struct ggml_backend_plan_cpu *)plan;
-
-    return ggml_graph_compute(&cpu_plan->cgraph, &cpu_plan->cplan);
-
-    GGML_UNUSED(backend);
-}
-
-static enum ggml_status ggml_backend_cpu_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
-    struct ggml_backend_cpu_context * cpu_ctx = (struct ggml_backend_cpu_context *)backend->context;
-
-    struct ggml_cplan cplan = ggml_graph_plan(cgraph, cpu_ctx->n_threads, cpu_ctx->threadpool);
-
-    if (cpu_ctx->work_size < cplan.work_size) {
-        delete[] cpu_ctx->work_data;
-        cpu_ctx->work_data = new uint8_t[cplan.work_size];
-        if (cpu_ctx->work_data == NULL) {
-            cpu_ctx->work_size = 0;
-            return GGML_STATUS_ALLOC_FAILED;
-        }
-        cpu_ctx->work_size = cplan.work_size;
-    }
-    cplan.work_data = (uint8_t *)cpu_ctx->work_data;
-
-    cplan.abort_callback      = cpu_ctx->abort_callback;
-    cplan.abort_callback_data = cpu_ctx->abort_callback_data;
-
-    return ggml_graph_compute(cgraph, &cplan);
-}
-
-static const struct ggml_backend_i ggml_backend_cpu_i = {
-    /* .get_name                = */ ggml_backend_cpu_get_name,
-    /* .free                    = */ ggml_backend_cpu_free,
-    /* .set_tensor_async        = */ NULL,
-    /* .get_tensor_async        = */ NULL,
-    /* .cpy_tensor_async        = */ NULL,
-    /* .synchronize             = */ NULL,
-    /* .graph_plan_create       = */ ggml_backend_cpu_graph_plan_create,
-    /* .graph_plan_free         = */ ggml_backend_cpu_graph_plan_free,
-    /* .graph_plan_update       = */ NULL,
-    /* .graph_plan_compute      = */ ggml_backend_cpu_graph_plan_compute,
-    /* .graph_compute           = */ ggml_backend_cpu_graph_compute,
-    /* .event_record            = */ NULL,
-    /* .event_wait              = */ NULL,
-};
-
-static ggml_guid_t ggml_backend_cpu_guid(void) {
-    static ggml_guid guid = { 0xaa, 0x67, 0xc7, 0x43, 0x96, 0xe6, 0xa3, 0x8a, 0xe3, 0xaf, 0xea, 0x92, 0x36, 0xbc, 0xfc, 0x89 };
-    return &guid;
-}
-
-ggml_backend_t ggml_backend_cpu_init(void) {
-    // initialize CPU backend now to avoid slowing the first graph computation
-    ggml_cpu_init();
-
-    struct ggml_backend_cpu_context * ctx = new ggml_backend_cpu_context;
-    if (ctx == NULL) {
-        return NULL;
-    }
-
-    ctx->n_threads           = GGML_DEFAULT_N_THREADS;
-    ctx->threadpool          = NULL;
-    ctx->work_data           = NULL;
-    ctx->work_size           = 0;
-    ctx->abort_callback      = NULL;
-    ctx->abort_callback_data = NULL;
-
-    ggml_backend_t cpu_backend = new ggml_backend {
-        /* .guid      = */ ggml_backend_cpu_guid(),
-        /* .interface = */ ggml_backend_cpu_i,
-        /* .device    = */ ggml_backend_reg_dev_get(ggml_backend_cpu_reg(), 0),
-        /* .context   = */ ctx,
-    };
-
-    if (cpu_backend == NULL) {
-        delete ctx;
-        return NULL;
-    }
-
-    return cpu_backend;
-}
-
-bool ggml_backend_is_cpu(ggml_backend_t backend) {
-    return backend != NULL && ggml_guid_matches(backend->guid, ggml_backend_cpu_guid());
-}
-
-void ggml_backend_cpu_set_n_threads(ggml_backend_t backend_cpu, int n_threads) {
-    GGML_ASSERT(ggml_backend_is_cpu(backend_cpu));
-
-    struct ggml_backend_cpu_context * ctx = (struct ggml_backend_cpu_context *)backend_cpu->context;
-    ctx->n_threads = n_threads;
-}
-
-void ggml_backend_cpu_set_threadpool(ggml_backend_t backend_cpu, ggml_threadpool_t threadpool) {
-    GGML_ASSERT(ggml_backend_is_cpu(backend_cpu));
-
-    struct ggml_backend_cpu_context * ctx = (struct ggml_backend_cpu_context *)backend_cpu->context;
-
-    if (ctx->threadpool && ctx->threadpool != threadpool) {
-        // already had a different threadpool, pause/suspend it before switching
-        ggml_threadpool_pause(ctx->threadpool);
-    }
-    ctx->threadpool = threadpool;
-}
-
-void ggml_backend_cpu_set_abort_callback(ggml_backend_t backend_cpu, ggml_abort_callback abort_callback, void * abort_callback_data) {
-    GGML_ASSERT(ggml_backend_is_cpu(backend_cpu));
-
-    struct ggml_backend_cpu_context * ctx = (struct ggml_backend_cpu_context *)backend_cpu->context;
-    ctx->abort_callback = abort_callback;
-    ctx->abort_callback_data = abort_callback_data;
-}
-
 ggml_backend_buffer_t ggml_backend_cpu_buffer_from_ptr(void * ptr, size_t size) {
     GGML_ASSERT((uintptr_t)ptr % TENSOR_ALIGNMENT == 0 && "buffer pointer must be aligned");
     return ggml_backend_buffer_init(ggml_backend_cpu_buffer_from_ptr_type(), ggml_backend_cpu_buffer_from_ptr_i, ptr, size);
 }
-
-// CPU backend - device
-
-struct ggml_backend_cpu_device_context {
-    std::string description = "CPU";
-
-    ggml_backend_cpu_device_context() {
-#ifdef __APPLE__
-        size_t len = 0;
-        if (!sysctlbyname("machdep.cpu.brand_string", NULL, &len, NULL, 0)) {
-            description.resize(len);
-            sysctlbyname("machdep.cpu.brand_string", &description[0], &len, NULL, 0); // NOLINT
-        }
-#elif defined(__linux__)
-        FILE * f = fopen("/proc/cpuinfo", "r");
-        if (f) {
-            char buf[1024];
-            while (fgets(buf, sizeof(buf), f)) {
-                if (strncmp(buf, "model name", 10) == 0) {
-                    char * p = strchr(buf, ':');
-                    if (p) {
-                        p++;
-                        while (std::isspace(*p)) {
-                            p++;
-                        }
-                        while (std::isspace(p[strlen(p) - 1])) {
-                            p[strlen(p) - 1] = '\0';
-                        }
-                        description = p;
-                        break;
-                    }
-                }
-            }
-            fclose(f);
-        }
-#elif defined(_WIN32)
-        HKEY hKey;
-        if (RegOpenKeyEx(HKEY_LOCAL_MACHINE,
-                        TEXT("HARDWARE\\DESCRIPTION\\System\\CentralProcessor\\0"),
-                        0,
-                        KEY_READ,
-                        &hKey) == ERROR_SUCCESS) {
-            DWORD cpu_brand_size = 0;
-            if (RegQueryValueExA(hKey,
-                                TEXT("ProcessorNameString"),
-                                NULL,
-                                NULL,
-                                NULL,
-                                &cpu_brand_size) == ERROR_SUCCESS) {
-                description.resize(cpu_brand_size);
-                if (RegQueryValueExA(hKey,
-                                    TEXT("ProcessorNameString"),
-                                    NULL,
-                                    NULL,
-                                    (LPBYTE)&description[0], // NOLINT
-                                    &cpu_brand_size) == ERROR_SUCCESS) {
-                    if (description.find('\0') != std::string::npos) {
-                        description.resize(description.find('\0'));
-                    }
-                }
-            }
-            RegCloseKey(hKey);
-        }
-#endif
-    }
-};
-
-static const char * ggml_backend_cpu_device_get_name(ggml_backend_dev_t dev) {
-    return "CPU";
-
-    GGML_UNUSED(dev);
-}
-
-static const char * ggml_backend_cpu_device_get_description(ggml_backend_dev_t dev) {
-    struct ggml_backend_cpu_device_context * ctx = (struct ggml_backend_cpu_device_context *)dev->context;
-
-    return ctx->description.c_str();
-}
-
-static void ggml_backend_cpu_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_cpu_device_get_type(ggml_backend_dev_t dev) {
-    return GGML_BACKEND_DEVICE_TYPE_CPU;
-
-    GGML_UNUSED(dev);
-}
-
-static void ggml_backend_cpu_device_get_props(ggml_backend_dev_t dev, struct ggml_backend_dev_props * props) {
-    props->name        = ggml_backend_cpu_device_get_name(dev);
-    props->description = ggml_backend_cpu_device_get_description(dev);
-    props->type        = ggml_backend_cpu_device_get_type(dev);
-    ggml_backend_cpu_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_cpu_device_init_backend(ggml_backend_dev_t dev, const char * params) {
-    return ggml_backend_cpu_init();
-
-    GGML_UNUSED(dev);
-    GGML_UNUSED(params);
-}
-
-static ggml_backend_buffer_type_t ggml_backend_cpu_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_cpu_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_cpu_device_supports_op(ggml_backend_dev_t dev, const struct ggml_tensor * op) {
-    switch (op->op) {
-        case GGML_OP_CPY:
-            return
-                op->type != GGML_TYPE_IQ2_XXS &&
-                op->type != GGML_TYPE_IQ2_XS  &&
-                op->type != GGML_TYPE_IQ1_S   &&
-                op->type != GGML_TYPE_IQ1_M; // missing type_traits.from_float
-        case GGML_OP_MUL_MAT:
-            return op->src[1]->type == GGML_TYPE_F32;// FIXME || op->src[1]->type == ggml_get_type_traits(op->src[0]->type)->vec_dot_type;
-        case GGML_OP_ROPE_BACK:
-            return op->src[2] == NULL && (op->op_params[2] & 4) == 0;
-        case GGML_OP_IM2COL_BACK:
-            return op->src[0]->type == GGML_TYPE_F32 && op->src[1]->type == GGML_TYPE_F32;
-        case GGML_OP_OUT_PROD:
-            return (op->src[0]->type == GGML_TYPE_F32 || ggml_is_quantized(op->src[0]->type)) && op->src[1]->type == GGML_TYPE_F32;
-        default:
-            return true;
-    }
-
-    GGML_UNUSED(dev);
-}
-
-static bool ggml_backend_cpu_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_cpu_device_i = {
-    /* .get_name             = */ ggml_backend_cpu_device_get_name,
-    /* .get_description      = */ ggml_backend_cpu_device_get_description,
-    /* .get_memory           = */ ggml_backend_cpu_device_get_memory,
-    /* .get_type             = */ ggml_backend_cpu_device_get_type,
-    /* .get_props            = */ ggml_backend_cpu_device_get_props,
-    /* .init_backend         = */ ggml_backend_cpu_device_init_backend,
-    /* .get_buffer_type      = */ ggml_backend_cpu_device_get_buffer_type,
-    /* .get_host_buffer_type = */ NULL,
-    /* .buffer_from_host_ptr = */ ggml_backend_cpu_device_buffer_from_host_ptr,
-    /* .supports_op          = */ ggml_backend_cpu_device_supports_op,
-    /* .supports_buft        = */ ggml_backend_cpu_device_supports_buft,
-    /* .offload_op           = */ NULL,
-    /* .event_new            = */ NULL,
-    /* .event_free           = */ NULL,
-    /* .event_synchronize    = */ NULL,
-};
-
-// CPU backend - backend (reg)
-
-static const char * ggml_backend_cpu_reg_get_name(ggml_backend_reg_t reg) {
-    return "CPU";
-
-    GGML_UNUSED(reg);
-}
-
-static size_t ggml_backend_cpu_reg_get_device_count(ggml_backend_reg_t reg) {
-    return 1;
-
-    GGML_UNUSED(reg);
-}
-
-static ggml_backend_dev_t ggml_backend_cpu_reg_get_device(ggml_backend_reg_t reg, size_t index) {
-    GGML_ASSERT(index == 0);
-
-    static ggml_backend_cpu_device_context ctx;
-    static ggml_backend_device ggml_backend_cpu_device = {
-        /* .iface   = */ ggml_backend_cpu_device_i,
-        /* .reg     = */ reg,
-        /* .context = */ &ctx,
-    };
-
-    return &ggml_backend_cpu_device;
-}
-
-static void * ggml_backend_cpu_get_proc_address(ggml_backend_reg_t reg, const char * name) {
-    if (strcmp(name, "ggml_backend_set_n_threads") == 0) {
-        return (void *)ggml_backend_cpu_set_n_threads;
-    }
-    if (strcmp(name, "ggml_backend_dev_get_extra_bufts") == 0) {
-        return (void *)ggml_backend_cpu_get_extra_bufts;
-    }
-
-    return NULL;
-
-    GGML_UNUSED(reg);
-}
-
-static const struct ggml_backend_reg_i ggml_backend_cpu_reg_i = {
-    /* .get_name         = */ ggml_backend_cpu_reg_get_name,
-    /* .get_device_count = */ ggml_backend_cpu_reg_get_device_count,
-    /* .get_device       = */ ggml_backend_cpu_reg_get_device,
-    /* .get_proc_address = */ ggml_backend_cpu_get_proc_address,
-};
-
-ggml_backend_reg_t ggml_backend_cpu_reg(void) {
-    static struct ggml_backend_reg ggml_backend_cpu_reg = {
-        /* .iface   = */ ggml_backend_cpu_reg_i,
-        /* .context = */ NULL,
-    };
-
-    return &ggml_backend_cpu_reg;
-}

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

@@ -6,7 +6,7 @@
 #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>
@@ -320,7 +320,7 @@ static const char * ggml_backend_blas_device_get_name(ggml_backend_dev_t dev) {
 }
 
 static const char * ggml_backend_blas_device_get_description(ggml_backend_dev_t dev) {
-    #if defined(GGML_USE_ACCELERATE)
+    #if defined(GGML_BLAS_USE_ACCELERATE)
         return "Accelerate";
     #elif defined(GGML_BLAS_USE_MKL)
         return "MKL";

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-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

ggml/src/ggml-cpu/ggml-cpu.c

@@ -1,13 +1,15 @@
 #define _CRT_SECURE_NO_DEPRECATE // Disables "unsafe" warnings on Windows
 #define _USE_MATH_DEFINES // For M_PI on MSVC
 
-#include "ggml-aarch64.h"
 #include "ggml-backend-impl.h"
 #include "ggml-backend.h"
+#include "ggml-cpu-aarch64.h"
 #include "ggml-cpu-impl.h"
 #include "ggml-cpu.h"
 #include "ggml-impl.h"
 #include "ggml-quants.h"
+#include "ggml-cpu-quants.h"
+#include "ggml-threading.h"
 #include "ggml.h"
 
 #if defined(_MSC_VER) || defined(__MINGW32__)
@@ -42,7 +44,7 @@
 #endif
 
 #ifdef GGML_USE_LLAMAFILE
-#include <llamafile/sgemm.h>
+#include "llamafile/sgemm.h"
 #endif
 
 #if defined(_MSC_VER)
@@ -104,9 +106,6 @@ static ggml_fp16_t ggml_table_gelu_f16[1 << 16];
 // precomputed quick gelu table for f16 (128 KB)
 static ggml_fp16_t ggml_table_gelu_quick_f16[1 << 16];
 
-// precomputed f32 table for f16 (256 KB) (ggml-impl.h)
-float ggml_table_f32_f16[1 << 16];
-
 #if defined(__ARM_ARCH)
 struct ggml_arm_arch_features_type {
     int has_neon;
@@ -261,11 +260,13 @@ static const struct ggml_type_traits_cpu type_traits_cpu[GGML_TYPE_COUNT] = {
         .nrows                    = 1,
     },
     [GGML_TYPE_F16] = {
+        .from_float               = (ggml_from_float_t) ggml_fp32_to_fp16_row,
         .vec_dot                  = (ggml_vec_dot_t) ggml_vec_dot_f16,
         .vec_dot_type             = GGML_TYPE_F16,
         .nrows                    = 1,
     },
     [GGML_TYPE_Q4_0] = {
+        .from_float               = quantize_row_q4_0,
         .vec_dot                  = ggml_vec_dot_q4_0_q8_0,
         .vec_dot_type             = GGML_TYPE_Q8_0,
 #if defined (__ARM_FEATURE_MATMUL_INT8)
@@ -275,6 +276,7 @@ static const struct ggml_type_traits_cpu type_traits_cpu[GGML_TYPE_COUNT] = {
 #endif
     },
     [GGML_TYPE_Q4_1] = {
+        .from_float               = quantize_row_q4_1,
         .vec_dot                  = ggml_vec_dot_q4_1_q8_1,
         .vec_dot_type             = GGML_TYPE_Q8_1,
 #if defined (__ARM_FEATURE_MATMUL_INT8)
@@ -283,27 +285,20 @@ static const struct ggml_type_traits_cpu type_traits_cpu[GGML_TYPE_COUNT] = {
         .nrows                    = 1,
 #endif
     },
-    [4] = { // GGML_TYPE_Q4_2
-        .vec_dot                  = NULL,
-        .vec_dot_type             = GGML_TYPE_COUNT,
-        .nrows                    = 1,
-    },
-    [5] = { // GGML_TYPE_Q4_3
-        .vec_dot                  = NULL,
-        .vec_dot_type             = GGML_TYPE_COUNT,
-        .nrows                    = 1,
-    },
     [GGML_TYPE_Q5_0] = {
+        .from_float               = quantize_row_q5_0,
         .vec_dot                  = ggml_vec_dot_q5_0_q8_0,
         .vec_dot_type             = GGML_TYPE_Q8_0,
         .nrows                    = 1,
     },
     [GGML_TYPE_Q5_1] = {
+        .from_float               = quantize_row_q5_1,
         .vec_dot                  = ggml_vec_dot_q5_1_q8_1,
         .vec_dot_type             = GGML_TYPE_Q8_1,
         .nrows                    = 1,
     },
     [GGML_TYPE_Q8_0] = {
+        .from_float               = quantize_row_q8_0,
         .from_float_to_mat        = quantize_mat_q8_0,
         .vec_dot                  = ggml_vec_dot_q8_0_q8_0,
         .vec_dot_type             = GGML_TYPE_Q8_0,
@@ -314,85 +309,106 @@ static const struct ggml_type_traits_cpu type_traits_cpu[GGML_TYPE_COUNT] = {
 #endif
     },
     [GGML_TYPE_Q8_1] = {
+        .from_float               = quantize_row_q8_1,
         .vec_dot_type             = GGML_TYPE_Q8_1,
         .nrows                    = 1,
     },
     [GGML_TYPE_Q2_K] = {
+        .from_float               = quantize_row_q2_K,
         .vec_dot                  = ggml_vec_dot_q2_K_q8_K,
         .vec_dot_type             = GGML_TYPE_Q8_K,
         .nrows                    = 1,
     },
     [GGML_TYPE_Q3_K] = {
+        .from_float               = quantize_row_q3_K,
         .vec_dot                  = ggml_vec_dot_q3_K_q8_K,
         .vec_dot_type             = GGML_TYPE_Q8_K,
         .nrows                    = 1,
     },
     [GGML_TYPE_Q4_K] = {
+        .from_float               = quantize_row_q4_K,
         .vec_dot                  = ggml_vec_dot_q4_K_q8_K,
         .vec_dot_type             = GGML_TYPE_Q8_K,
         .nrows                    = 1,
     },
     [GGML_TYPE_Q5_K] = {
+        .from_float               = quantize_row_q5_K,
         .vec_dot                  = ggml_vec_dot_q5_K_q8_K,
         .vec_dot_type             = GGML_TYPE_Q8_K,
         .nrows                    = 1,
     },
     [GGML_TYPE_Q6_K] = {
+        .from_float               = quantize_row_q6_K,
         .vec_dot                  = ggml_vec_dot_q6_K_q8_K,
         .vec_dot_type             = GGML_TYPE_Q8_K,
         .nrows                    = 1,
     },
     [GGML_TYPE_IQ2_XXS] = {
+        .from_float               = NULL,
         .vec_dot                  = ggml_vec_dot_iq2_xxs_q8_K,
         .vec_dot_type             = GGML_TYPE_Q8_K,
         .nrows                    = 1,
     },
     [GGML_TYPE_IQ2_XS] = {
+        .from_float               = NULL,
         .vec_dot                  = ggml_vec_dot_iq2_xs_q8_K,
         .vec_dot_type             = GGML_TYPE_Q8_K,
         .nrows                    = 1,
     },
     [GGML_TYPE_IQ3_XXS] = {
+        // NOTE: from_float for iq3 and iq2_s was removed because these quants require initialization in ggml_quantize_init
+        //.from_float               = quantize_row_iq3_xxs,
         .vec_dot                  = ggml_vec_dot_iq3_xxs_q8_K,
         .vec_dot_type             = GGML_TYPE_Q8_K,
         .nrows                    = 1,
     },
     [GGML_TYPE_IQ3_S] = {
+        //.from_float               = quantize_row_iq3_s,
         .vec_dot                  = ggml_vec_dot_iq3_s_q8_K,
         .vec_dot_type             = GGML_TYPE_Q8_K,
         .nrows                    = 1,
     },
     [GGML_TYPE_IQ2_S] = {
+        //.from_float               = quantize_row_iq2_s,
         .vec_dot                  = ggml_vec_dot_iq2_s_q8_K,
         .vec_dot_type             = GGML_TYPE_Q8_K,
         .nrows                    = 1,
     },
     [GGML_TYPE_IQ1_S] = {
+        .from_float               = NULL,
         .vec_dot                  = ggml_vec_dot_iq1_s_q8_K,
         .vec_dot_type             = GGML_TYPE_Q8_K,
         .nrows                    = 1,
     },
     [GGML_TYPE_IQ1_M] = {
+        .from_float               = NULL,
         .vec_dot                  = ggml_vec_dot_iq1_m_q8_K,
         .vec_dot_type             = GGML_TYPE_Q8_K,
         .nrows                    = 1,
     },
     [GGML_TYPE_IQ4_NL] = {
+        .from_float               = quantize_row_iq4_nl,
         .vec_dot                  = ggml_vec_dot_iq4_nl_q8_0,
         .vec_dot_type             = GGML_TYPE_Q8_0,
         .nrows                    = 1,
     },
     [GGML_TYPE_IQ4_XS] = {
+        .from_float               = quantize_row_iq4_xs,
         .vec_dot                  = ggml_vec_dot_iq4_xs_q8_K,
         .vec_dot_type             = GGML_TYPE_Q8_K,
         .nrows                    = 1,
     },
+    [GGML_TYPE_Q8_K] = {
+        .from_float               = quantize_row_q8_K,
+    },
     [GGML_TYPE_BF16] = {
+        .from_float               = (ggml_from_float_t) ggml_fp32_to_bf16_row,
         .vec_dot                  = (ggml_vec_dot_t) ggml_vec_dot_bf16,
         .vec_dot_type             = GGML_TYPE_BF16,
         .nrows                    = 1,
     },
     [GGML_TYPE_Q4_0_4_4] = {
+        .from_float               = NULL,
         .vec_dot                  = NULL,
         .vec_dot_type             = GGML_TYPE_Q8_0,
         .nrows                    = 1,
@@ -401,6 +417,7 @@ static const struct ggml_type_traits_cpu type_traits_cpu[GGML_TYPE_COUNT] = {
         .gemm                     = ggml_gemm_q4_0_4x4_q8_0,
     },
     [GGML_TYPE_Q4_0_4_8] = {
+        .from_float               = NULL,
         .vec_dot                  = NULL,
         .vec_dot_type             = GGML_TYPE_Q8_0,
         .nrows                    = 1,
@@ -409,6 +426,7 @@ static const struct ggml_type_traits_cpu type_traits_cpu[GGML_TYPE_COUNT] = {
         .gemm                     = ggml_gemm_q4_0_4x8_q8_0,
     },
     [GGML_TYPE_Q4_0_8_8] = {
+        .from_float               = NULL,
         .vec_dot                  = NULL,
         .vec_dot_type             = GGML_TYPE_Q8_0,
         .nrows                    = 1,
@@ -417,11 +435,13 @@ static const struct ggml_type_traits_cpu type_traits_cpu[GGML_TYPE_COUNT] = {
         .gemm                     = ggml_gemm_q4_0_8x8_q8_0,
     },
     [GGML_TYPE_TQ1_0] = {
+        .from_float               = quantize_row_tq1_0,
         .vec_dot                  = ggml_vec_dot_tq1_0_q8_K,
         .vec_dot_type             = GGML_TYPE_Q8_K,
         .nrows                    = 1,
     },
     [GGML_TYPE_TQ2_0] = {
+        .from_float               = quantize_row_tq2_0,
         .vec_dot                  = ggml_vec_dot_tq2_0_q8_K,
         .vec_dot_type             = GGML_TYPE_Q8_K,
         .nrows                    = 1,
@@ -1449,8 +1469,12 @@ static void ggml_vec_dot_bf16(int n, float * restrict s, size_t bs, ggml_bf16_t
     sumf += (ggml_float)_mm512_reduce_add_ps(c2);
 
 #undef LOAD
-#elif defined(__AVX2__)
+#elif defined(__AVX2__) || defined(__AVX__)
+#if defined(__AVX2__)
 #define LOAD(p) _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_cvtepu16_epi32(_mm_loadu_si128((const __m128i *)(p))), 16))
+#else
+#define LOAD(p) _mm256_castsi256_ps(_mm256_insertf128_si256(_mm256_castsi128_si256(_mm_slli_epi32(_mm_cvtepu16_epi32(_mm_loadu_si128((const __m128i *)(p))), 16)), (_mm_slli_epi32(_mm_cvtepu16_epi32(_mm_bsrli_si128(_mm_loadu_si128((const __m128i *)(p)), 8)), 16)), 1))
+#endif
     __m256 c1 = _mm256_setzero_ps();
     __m256 c2 = _mm256_setzero_ps();
     __m256 c3 = _mm256_setzero_ps();
@@ -2250,22 +2274,7 @@ struct ggml_state {
     struct ggml_numa_nodes numa;
 };
 
-// global state
 static struct ggml_state g_state = {0};
-static atomic_flag g_state_critical = ATOMIC_FLAG_INIT;
-
-// TODO: move to threading file
-// critical section via spin lock
-void ggml_critical_section_start(void) {
-    while (atomic_flag_test_and_set(&g_state_critical)) {
-        // spin
-        sched_yield();
-    }
-}
-
-void ggml_critical_section_end(void) {
-    atomic_flag_clear(&g_state_critical);
-}
 
 static void ggml_barrier(struct ggml_threadpool * tp) {
     int n_threads = atomic_load_explicit(&tp->n_threads_cur, memory_order_relaxed);
@@ -2360,7 +2369,7 @@ void ggml_numa_init(enum ggml_numa_strategy numa_flag) {
     // figure out which node we're on
     uint current_cpu;
     int getcpu_ret = 0;
-#if __GLIBC__ > 2 || (__GLIBC__ == 2 && __GLIBC_MINOR__ > 28) || defined(__COSMOPOLITAN__)
+#if __GLIBC__ > 2 || (__GLIBC__ == 2 && __GLIBC_MINOR__ > 33) || defined(__COSMOPOLITAN__)
     getcpu_ret = getcpu(&current_cpu, &g_state.numa.current_node);
 #else
     // old glibc doesn't have a wrapper for this call. Fall back on direct syscall
@@ -2997,8 +3006,8 @@ static void ggml_compute_forward_dup_f16(
                         id += ne00 * (ne01 - ir1);
                     }
                 }
-            } else if (ggml_get_type_traits(dst->type)->from_float) {
-                ggml_from_float_t const quantize_row_q = ggml_get_type_traits(dst->type)->from_float;
+            } else if (ggml_get_type_traits_cpu(dst->type)->from_float) {
+                ggml_from_float_t const quantize_row_q = ggml_get_type_traits_cpu(dst->type)->from_float;
                 float * src0_f32 = (float *) params->wdata + (ne00 + CACHE_LINE_SIZE_F32) * ith;
 
                 size_t id = 0;
@@ -3278,8 +3287,8 @@ static void ggml_compute_forward_dup_bf16(
                         id += ne00 * (ne01 - ir1);
                     }
                 }
-            } else if (ggml_get_type_traits(dst->type)->from_float) {
-                ggml_from_float_t const quantize_row_q = ggml_get_type_traits(dst->type)->from_float;
+            } else if (ggml_get_type_traits_cpu(dst->type)->from_float) {
+                ggml_from_float_t const quantize_row_q = ggml_get_type_traits_cpu(dst->type)->from_float;
                 float * src0_f32 = (float *) params->wdata + (ne00 + CACHE_LINE_SIZE_F32) * ith;
 
                 size_t id = 0;
@@ -3594,8 +3603,8 @@ static void ggml_compute_forward_dup_f32(
                         id += rs * (ne01 - ir1);
                     }
                 }
-            } else if (ggml_get_type_traits(dst->type)->from_float) {
-                ggml_from_float_t const quantize_row_q = ggml_get_type_traits(dst->type)->from_float;
+            } else if (ggml_get_type_traits_cpu(dst->type)->from_float) {
+                ggml_from_float_t const quantize_row_q = ggml_get_type_traits_cpu(dst->type)->from_float;
 
                 size_t id = 0;
                 size_t rs = nb0 * (ne00 / ggml_blck_size(dst->type));
@@ -4377,7 +4386,7 @@ static void ggml_compute_forward_add_q_f32(
     const enum ggml_type type = src0->type;
     const enum ggml_type dtype = dst->type;
     ggml_to_float_t const dequantize_row_q = ggml_get_type_traits(type)->to_float;
-    ggml_from_float_t const quantize_row_q = ggml_get_type_traits(dtype)->from_float;
+    ggml_from_float_t const quantize_row_q = ggml_get_type_traits_cpu(dtype)->from_float;
 
     // we don't support permuted src0 or src1
     GGML_ASSERT(nb00 == ggml_type_size(type));
@@ -4679,7 +4688,7 @@ static void ggml_compute_forward_add1_q_f32(
 
     const enum ggml_type type = src0->type;
     ggml_to_float_t const dequantize_row_q = ggml_get_type_traits(type)->to_float;
-    ggml_from_float_t const quantize_row_q = ggml_get_type_traits(type)->from_float;
+    ggml_from_float_t const quantize_row_q = ggml_get_type_traits_cpu(type)->from_float;
 
     // we don't support permuted src0
     GGML_ASSERT(nb00 == ggml_type_size(type));
@@ -7325,6 +7334,7 @@ static void ggml_compute_forward_group_norm(
 static void ggml_compute_forward_mul_mat_one_chunk(
     const struct ggml_compute_params * params,
     struct ggml_tensor * dst,
+    const enum ggml_type type,
     const int64_t num_rows_per_vec_dot,
     const int64_t ir0_start,
     const int64_t ir0_end,
@@ -7336,8 +7346,6 @@ static void ggml_compute_forward_mul_mat_one_chunk(
 
     GGML_TENSOR_BINARY_OP_LOCALS
 
-    const enum ggml_type type = src0->type;
-
     const bool src1_cont = ggml_is_contiguous(src1);
 
     ggml_vec_dot_t const vec_dot      = type_traits_cpu[type].vec_dot;
@@ -7425,10 +7433,14 @@ static void ggml_compute_forward_mul_mat(
     const int ith = params->ith;
     const int nth = params->nth;
 
-    const enum ggml_type type = src0->type;
+    enum ggml_type type = src0->type;
+
+    if (src0->buffer && ggml_backend_cpu_buft_is_aarch64(src0->buffer->buft)) {
+        type = (enum ggml_type)(intptr_t)src0->extra;
+    }
 
     enum ggml_type           const vec_dot_type         = type_traits_cpu[type].vec_dot_type;
-    ggml_from_float_t        const from_float           = ggml_get_type_traits(vec_dot_type)->from_float;
+    ggml_from_float_t        const from_float           = type_traits_cpu[vec_dot_type].from_float;
     ggml_from_float_to_mat_t const from_float_to_mat    = type_traits_cpu[vec_dot_type].from_float_to_mat;
     int64_t                  const vec_dot_num_rows     = type_traits_cpu[type].nrows;
     int64_t                  const matmul_num_cols      = type_traits_cpu[type].ncols;
@@ -7464,15 +7476,15 @@ static void ggml_compute_forward_mul_mat(
     if (src1_cont) {
         for (int64_t i13 = 0; i13 < ne13; i13++)
             for (int64_t i12 = 0; i12 < ne12; i12++)
-                if (!llamafile_sgemm(ne01, ne11, ne00/ggml_blck_size(src0->type),
+                if (!llamafile_sgemm(ne01, ne11, ne00/ggml_blck_size(type),
                                      (const char *)src0->data + i12/r2*nb02 + i13/r3*nb03,
-                                     nb01/ggml_type_size(src0->type),
+                                     nb01/ggml_type_size(type),
                                      (const char *)src1->data + i12*nb12 + i13*nb13,
                                      nb11/ggml_type_size(src1->type),
                                      (char *)dst->data + i12*nb2 + i13*nb3,
                                      nb1/ggml_type_size(dst->type),
                                      ith, nth,
-                                     src0->type,
+                                     type,
                                      src1->type,
                                      dst->type))
                     goto UseGgmlGemm1;
@@ -7525,15 +7537,15 @@ UseGgmlGemm1:;
 
         for (int64_t i13 = 0; i13 < ne13; i13++)
             for (int64_t i12 = 0; i12 < ne12; i12++)
-                if (!llamafile_sgemm(ne01, ne11, ne00/ggml_blck_size(src0->type),
+                if (!llamafile_sgemm(ne01, ne11, ne00/ggml_blck_size(type),
                                      (const char *)src0->data + i12/r2*nb02 + i13/r3*nb03,
-                                     nb01/ggml_type_size(src0->type),
+                                     nb01/ggml_type_size(type),
                                      (const char *)wdata + (i12*ne11 + i13*ne12*ne11)*row_size,
                                      row_size/ggml_type_size(vec_dot_type),
                                      (char *)dst->data + i12*nb2 + i13*nb3,
                                      nb1/ggml_type_size(dst->type),
                                      ith, nth,
-                                     src0->type,
+                                     type,
                                      vec_dot_type,
                                      dst->type))
                     goto UseGgmlGemm2;
@@ -7618,7 +7630,7 @@ UseGgmlGemm2:;
         const int64_t ir1_start = dr1 * ith1;
         const int64_t ir1_end = MIN(ir1_start + dr1, nr1);
 
-        ggml_compute_forward_mul_mat_one_chunk(params, dst, num_rows_per_vec_dot, ir0_start, ir0_end, ir1_start, ir1_end);
+        ggml_compute_forward_mul_mat_one_chunk(params, dst, type, num_rows_per_vec_dot, ir0_start, ir0_end, ir1_start, ir1_end);
 
         if (nth >= nchunk0 * nchunk1) {
             break;
@@ -7649,7 +7661,7 @@ static void ggml_compute_forward_mul_mat_id(
 
     ggml_vec_dot_t    const vec_dot         = type_traits_cpu[type].vec_dot;
     enum ggml_type    const vec_dot_type    = type_traits_cpu[type].vec_dot_type;
-    ggml_from_float_t const from_float      = ggml_get_type_traits(vec_dot_type)->from_float;
+    ggml_from_float_t const from_float      = type_traits_cpu[vec_dot_type].from_float;
     int64_t           const matmul_num_cols = type_traits_cpu[type].ncols;
     ggml_gemv_t       const gemv            = type_traits_cpu[type].gemv;
 
@@ -9159,12 +9171,6 @@ static void rope_yarn(
     *sin_theta = sinf(theta) * mscale;
 }
 
-// Apparently solving `n_rot = 2pi * x * base^((2 * max_pos_emb) / n_dims)` for x, we get
-// `corr_dim(n_rot) = n_dims * log(max_pos_emb / (n_rot * 2pi)) / (2 * log(base))`
-static float ggml_rope_yarn_corr_dim(int n_dims, int n_ctx_orig, float n_rot, float base) {
-    return n_dims * logf(n_ctx_orig / (n_rot * 2 * (float)M_PI)) / (2 * logf(base));
-}
-
 static void ggml_rope_cache_init(
      float theta_base, float freq_scale, const float * freq_factors, float corr_dims[2], int64_t ne0, float ext_factor, float mscale,
      float * cache, float sin_sign, float theta_scale) {
@@ -9181,16 +9187,6 @@ static void ggml_rope_cache_init(
     }
 }
 
-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]
-) {
-    // start and end correction dims
-    float start = floorf(ggml_rope_yarn_corr_dim(n_dims, n_ctx_orig, beta_fast, freq_base));
-    float end   =  ceilf(ggml_rope_yarn_corr_dim(n_dims, n_ctx_orig, beta_slow, freq_base));
-    dims[0] = MAX(0, start);
-    dims[1] = MIN(n_dims - 1, end);
-}
-
 static void ggml_compute_forward_rope_f32(
         const struct ggml_compute_params * params,
         struct ggml_tensor * dst,
@@ -10668,7 +10664,7 @@ static void ggml_compute_forward_flash_attn_ext_f16(
     const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2);
 
     enum ggml_type    const k_vec_dot_type = type_traits_cpu[k->type].vec_dot_type;
-    ggml_from_float_t const q_to_vec_dot   = ggml_get_type_traits(k_vec_dot_type)->from_float;
+    ggml_from_float_t const q_to_vec_dot   = type_traits_cpu[k_vec_dot_type].from_float;
     ggml_vec_dot_t    const kq_vec_dot     = type_traits_cpu[k->type].vec_dot;
     ggml_to_float_t   const v_to_float     = ggml_get_type_traits(v->type)->to_float;
 
@@ -12220,11 +12216,16 @@ static void ggml_compute_forward_opt_step_adamw_f32(
         const struct ggml_compute_params * params,
         struct ggml_tensor * dst) {
 
-    const struct ggml_tensor * src0        = dst->src[0];
-    const struct ggml_tensor * src0_grad   = dst->src[1];
-    const struct ggml_tensor * src0_grad_m = dst->src[2];
-    const struct ggml_tensor * src0_grad_v = dst->src[3];
+    const struct ggml_tensor * src0         = dst->src[0];
+    const struct ggml_tensor * src0_grad    = dst->src[1];
+    const struct ggml_tensor * src0_grad_m  = dst->src[2];
+    const struct ggml_tensor * src0_grad_v  = dst->src[3];
+    const struct ggml_tensor * adamw_params = dst->src[4];
+
     GGML_ASSERT(ggml_are_same_shape(src0, src0_grad));
+    GGML_ASSERT(ggml_are_same_shape(src0, src0_grad_m));
+    GGML_ASSERT(ggml_are_same_shape(src0, src0_grad_v));
+    GGML_ASSERT(ggml_nelements(adamw_params) == 7);
 
     const int ith = params->ith;
     const int nth = params->nth;
@@ -12241,16 +12242,14 @@ static void ggml_compute_forward_opt_step_adamw_f32(
     const int ir0 = dr*ith;
     const int ir1 = MIN(ir0 + dr, nr);
 
-    /* const float   gnorm = 1.0f; */
-    int64_t       iter;   memcpy(&iter, &dst->op_params[0], sizeof(int64_t));
-    const float   alpha = ggml_get_op_params_f32(dst, 2);
-    const float   beta1 = ggml_get_op_params_f32(dst, 3);
-    const float   beta2 = ggml_get_op_params_f32(dst, 4);
-    const float   eps   = ggml_get_op_params_f32(dst, 5);
-    const float   wd    = ggml_get_op_params_f32(dst, 6);
-
-    const float beta1h  = alpha/(1.0f - powf(beta1, iter));
-    const float beta2h  =  1.0f/(1.0f - powf(beta2, iter));
+    const float * adamw_params_ptr = ggml_get_data_f32(adamw_params);
+    const float alpha  = adamw_params_ptr[0];
+    const float beta1  = adamw_params_ptr[1];
+    const float beta2  = adamw_params_ptr[2];
+    const float eps    = adamw_params_ptr[3];
+    const float wd     = adamw_params_ptr[4];
+    const float beta1h = adamw_params_ptr[5];
+    const float beta2h = adamw_params_ptr[6];
 
     for (int ir = ir0; ir < ir1; ++ir) {
         const int64_t i03 = ir/(ne02*ne01);
@@ -12274,17 +12273,9 @@ static void ggml_compute_forward_opt_step_adamw_f32(
             // The weight decay is applied independently of the Adam momenta m and v.
             // This is NOT equivalent to l2 regularization that adds w[i00]*w[i00] to the loss.
             // See: https://arxiv.org/pdf/1711.05101v3.pdf
-            w[i00] = w[i00]*(1.0f - alpha*wd) - mh/vh;
+            w[i00] = w[i00]*(1.0f - alpha*wd) - alpha*mh/vh;
         }
     }
-
-    ggml_barrier(params->threadpool);
-    if (ith != 0) {
-        return;
-    }
-
-    iter++;
-    memcpy(&dst->op_params[0], &iter, sizeof(int64_t));
 }
 
 static void ggml_compute_forward_opt_step_adamw(
@@ -13759,6 +13750,151 @@ enum ggml_status ggml_graph_compute_with_ctx(struct ggml_context * ctx, struct g
     return ggml_graph_compute(cgraph, &cplan);
 }
 
+
+int ggml_cpu_has_avx(void) {
+#if defined(__AVX__)
+    return 1;
+#else
+    return 0;
+#endif
+}
+
+int ggml_cpu_has_avx_vnni(void) {
+#if defined(__AVXVNNI__)
+    return 1;
+#else
+    return 0;
+#endif
+}
+
+int ggml_cpu_has_avx2(void) {
+#if defined(__AVX2__)
+    return 1;
+#else
+    return 0;
+#endif
+}
+
+int ggml_cpu_has_avx512(void) {
+#if defined(__AVX512F__)
+    return 1;
+#else
+    return 0;
+#endif
+}
+
+int ggml_cpu_has_avx512_vbmi(void) {
+#if defined(__AVX512VBMI__)
+    return 1;
+#else
+    return 0;
+#endif
+}
+
+int ggml_cpu_has_avx512_vnni(void) {
+#if defined(__AVX512VNNI__)
+    return 1;
+#else
+    return 0;
+#endif
+}
+
+int ggml_cpu_has_avx512_bf16(void) {
+#if defined(__AVX512BF16__)
+    return 1;
+#else
+    return 0;
+#endif
+}
+
+int ggml_cpu_has_amx_int8(void) {
+#if defined(__AMX_INT8__)
+    return 1;
+#else
+    return 0;
+#endif
+}
+
+int ggml_cpu_has_fma(void) {
+#if defined(__FMA__)
+    return 1;
+#else
+    return 0;
+#endif
+}
+
+int ggml_cpu_has_arm_fma(void) {
+#if defined(__ARM_FEATURE_FMA)
+    return 1;
+#else
+    return 0;
+#endif
+}
+
+int ggml_cpu_has_riscv_v(void) {
+#if defined(__riscv_v_intrinsic)
+    return 1;
+#else
+    return 0;
+#endif
+}
+
+int ggml_cpu_has_f16c(void) {
+#if defined(__F16C__)
+    return 1;
+#else
+    return 0;
+#endif
+}
+
+int ggml_cpu_has_fp16_va(void) {
+#if defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC)
+    return 1;
+#else
+    return 0;
+#endif
+}
+
+int ggml_cpu_has_wasm_simd(void) {
+#if defined(__wasm_simd128__)
+    return 1;
+#else
+    return 0;
+#endif
+}
+
+int ggml_cpu_has_llamafile(void) {
+#if defined(GGML_USE_LLAMAFILE)
+    return 1;
+#else
+    return 0;
+#endif
+}
+
+int ggml_cpu_has_sse3(void) {
+#if defined(__SSE3__)
+    return 1;
+#else
+    return 0;
+#endif
+}
+
+int ggml_cpu_has_ssse3(void) {
+#if defined(__SSSE3__)
+    return 1;
+#else
+    return 0;
+#endif
+}
+
+int ggml_cpu_has_vsx(void) {
+#if defined(__POWER9_VECTOR__)
+    return 1;
+#else
+    return 0;
+#endif
+}
+
 int ggml_cpu_has_neon(void) {
 #if defined(__ARM_ARCH)
     return ggml_arm_arch_features.has_neon;

ggml/src/ggml-cpu/ggml-cpu.cpp

@@ -0,0 +1,663 @@
+#include "ggml-backend.h"
+#include "ggml-backend-impl.h"
+#include "ggml-cpu.h"
+#include "ggml-cpu-aarch64.h"
+#include "ggml-impl.h"
+#include <cctype>
+#include <string>
+#include <vector>
+
+#if defined(__APPLE__)
+#include <sys/types.h>
+#include <sys/sysctl.h>
+#endif
+
+#if defined(_WIN32)
+#define WIN32_LEAN_AND_MEAN
+#ifndef NOMINMAX
+    #define NOMINMAX
+#endif
+#include <windows.h>
+#endif
+
+// ggml-backend interface
+
+#ifdef GGML_USE_CPU_HBM
+
+// buffer type HBM
+
+#include <hbwmalloc.h>
+
+static const char * ggml_backend_cpu_hbm_buffer_type_get_name(ggml_backend_buffer_type_t buft) {
+    return "CPU_HBM";
+
+    GGML_UNUSED(buft);
+}
+
+static void ggml_backend_cpu_hbm_buffer_free_buffer(ggml_backend_buffer_t buffer) {
+    hbw_free(buffer->context);
+}
+
+static ggml_backend_buffer_t ggml_backend_cpu_hbm_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
+    void * ptr;
+    int result = hbw_posix_memalign(&ptr, ggml_backend_cpu_buffer_type_get_alignment(buft), size);
+    if (result != 0) {
+        GGML_LOG_ERROR("failed to allocate HBM buffer of size %zu\n", size);
+        return NULL;
+    }
+
+    ggml_backend_buffer_t buffer = ggml_backend_cpu_buffer_from_ptr(ptr, size);
+    buffer->buft = buft;
+    buffer->iface.free_buffer = ggml_backend_cpu_hbm_buffer_free_buffer;
+
+    return buffer;
+}
+
+ggml_backend_buffer_type_t ggml_backend_cpu_hbm_buffer_type(void) {
+    static struct ggml_backend_buffer_type ggml_backend_cpu_buffer_type_hbm = {
+        /* .iface    = */ {
+            /* .get_name         = */ ggml_backend_cpu_hbm_buffer_type_get_name,
+            /* .alloc_buffer     = */ ggml_backend_cpu_hbm_buffer_type_alloc_buffer,
+            /* .get_alignment    = */ ggml_backend_cpu_buffer_type_get_alignment,
+            /* .get_max_size     = */ NULL, // defaults to SIZE_MAX
+            /* .get_alloc_size   = */ NULL, // defaults to ggml_nbytes
+            /* .is_host          = */ ggml_backend_cpu_buffer_type_is_host,
+        },
+        /* .context  = */ NULL,
+    };
+
+    return &ggml_backend_cpu_buffer_type_hbm;
+}
+#endif
+
+// buffer type AARCH64
+
+static void ggml_backend_cpu_aarch64_buffer_init_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) {
+    tensor->extra = (void *)ggml_aarch64_get_optimal_repack_type(tensor); // NOLINT
+
+    GGML_UNUSED(buffer);
+}
+
+static void ggml_backend_cpu_aarch64_buffer_set_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
+    GGML_ASSERT(offset == 0);
+    GGML_ASSERT(size == ggml_nbytes(tensor));
+
+    enum ggml_type repack_type = (enum ggml_type)(intptr_t)tensor->extra;
+
+    ggml_aarch64_repack_tensor(tensor, repack_type, data, size);
+
+    GGML_UNUSED(buffer);
+}
+
+static const char * ggml_backend_cpu_aarch64_buffer_type_get_name(ggml_backend_buffer_type_t buft) {
+    return "CPU_AARCH64";
+
+    GGML_UNUSED(buft);
+}
+
+static ggml_backend_buffer_t ggml_backend_cpu_aarch64_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
+    auto * buffer = ggml_backend_buft_alloc_buffer(ggml_backend_cpu_buffer_type(), size);
+
+    if (buffer == NULL) {
+        return NULL;
+    }
+
+    buffer->buft = buft;
+    buffer->iface.init_tensor = ggml_backend_cpu_aarch64_buffer_init_tensor;
+    buffer->iface.set_tensor = ggml_backend_cpu_aarch64_buffer_set_tensor;
+
+    return buffer;
+}
+
+ggml_backend_buffer_type_t ggml_backend_cpu_aarch64_buffer_type(void) {
+    static struct ggml_backend_buffer_type ggml_backend_cpu_buffer_type_aarch64 = {
+        /* .iface    = */ {
+            /* .get_name         = */ ggml_backend_cpu_aarch64_buffer_type_get_name,
+            /* .alloc_buffer     = */ ggml_backend_cpu_aarch64_buffer_type_alloc_buffer,
+            /* .get_alignment    = */ ggml_backend_cpu_buffer_type()->iface.get_alignment,
+            /* .get_max_size     = */ NULL, // defaults to SIZE_MAX
+            /* .get_alloc_size   = */ NULL, // defaults to ggml_nbytes
+            /* .is_host          = */ NULL,
+        },
+        /* .device  = */ ggml_backend_reg_dev_get(ggml_backend_cpu_reg(), 0),
+        /* .context = */ NULL,
+    };
+
+    return &ggml_backend_cpu_buffer_type_aarch64;
+}
+
+bool ggml_backend_cpu_buft_is_aarch64(ggml_backend_buffer_type_t buft) {
+    return buft == ggml_backend_cpu_aarch64_buffer_type();
+}
+
+static ggml_backend_buffer_type_t * ggml_backend_cpu_get_extra_bufts(ggml_backend_dev_t device) {
+    static std::vector<ggml_backend_buffer_type_t> bufts = []() {
+        std::vector<ggml_backend_buffer_type_t> bufts;
+
+#ifdef GGML_USE_CPU_HBM
+        bufts.push_back(ggml_backend_cpu_hbm_buffer_type());
+#endif
+
+#ifdef GGML_USE_CPU_AARCH64
+        bufts.push_back(ggml_backend_cpu_aarch64_buffer_type());
+#endif
+
+        bufts.push_back(NULL);
+
+        return bufts;
+    }();
+
+    return bufts.data();
+
+    GGML_UNUSED(device);
+}
+
+// CPU backend - backend (stream)
+
+struct ggml_backend_cpu_context {
+    int                 n_threads;
+    ggml_threadpool_t   threadpool;
+
+    uint8_t *           work_data;
+    size_t              work_size;
+
+    ggml_abort_callback abort_callback;
+    void *              abort_callback_data;
+};
+
+static const char * ggml_backend_cpu_get_name(ggml_backend_t backend) {
+    return "CPU";
+
+    GGML_UNUSED(backend);
+}
+
+static void ggml_backend_cpu_free(ggml_backend_t backend) {
+    struct ggml_backend_cpu_context * cpu_ctx = (struct ggml_backend_cpu_context *)backend->context;
+    delete[] cpu_ctx->work_data;
+    delete cpu_ctx;
+    delete backend;
+}
+
+struct ggml_backend_plan_cpu {
+    struct ggml_cplan cplan;
+    struct ggml_cgraph cgraph;
+};
+
+static ggml_backend_graph_plan_t ggml_backend_cpu_graph_plan_create(ggml_backend_t backend, const struct ggml_cgraph * cgraph) {
+    struct ggml_backend_cpu_context * cpu_ctx = (struct ggml_backend_cpu_context *)backend->context;
+
+    struct ggml_backend_plan_cpu * cpu_plan = new ggml_backend_plan_cpu;
+
+    cpu_plan->cplan = ggml_graph_plan(cgraph, cpu_ctx->n_threads, cpu_ctx->threadpool);
+    cpu_plan->cgraph = *cgraph; // FIXME: deep copy
+
+    if (cpu_plan->cplan.work_size > 0) {
+        cpu_plan->cplan.work_data = new uint8_t[cpu_plan->cplan.work_size];
+        if (cpu_plan->cplan.work_data == NULL) {
+            delete cpu_plan;
+            return NULL;
+        }
+    }
+
+    cpu_plan->cplan.abort_callback      = cpu_ctx->abort_callback;
+    cpu_plan->cplan.abort_callback_data = cpu_ctx->abort_callback_data;
+
+    return cpu_plan;
+}
+
+static void ggml_backend_cpu_graph_plan_free(ggml_backend_t backend, ggml_backend_graph_plan_t plan) {
+    struct ggml_backend_plan_cpu * cpu_plan = (struct ggml_backend_plan_cpu *)plan;
+
+    delete[] cpu_plan->cplan.work_data;
+    delete cpu_plan;
+
+    GGML_UNUSED(backend);
+}
+
+static enum ggml_status ggml_backend_cpu_graph_plan_compute(ggml_backend_t backend, ggml_backend_graph_plan_t plan) {
+    struct ggml_backend_plan_cpu * cpu_plan = (struct ggml_backend_plan_cpu *)plan;
+
+    return ggml_graph_compute(&cpu_plan->cgraph, &cpu_plan->cplan);
+
+    GGML_UNUSED(backend);
+}
+
+static enum ggml_status ggml_backend_cpu_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
+    struct ggml_backend_cpu_context * cpu_ctx = (struct ggml_backend_cpu_context *)backend->context;
+
+    struct ggml_cplan cplan = ggml_graph_plan(cgraph, cpu_ctx->n_threads, cpu_ctx->threadpool);
+
+    if (cpu_ctx->work_size < cplan.work_size) {
+        delete[] cpu_ctx->work_data;
+        cpu_ctx->work_data = new uint8_t[cplan.work_size];
+        if (cpu_ctx->work_data == NULL) {
+            cpu_ctx->work_size = 0;
+            return GGML_STATUS_ALLOC_FAILED;
+        }
+        cpu_ctx->work_size = cplan.work_size;
+    }
+    cplan.work_data = (uint8_t *)cpu_ctx->work_data;
+
+    cplan.abort_callback      = cpu_ctx->abort_callback;
+    cplan.abort_callback_data = cpu_ctx->abort_callback_data;
+
+    return ggml_graph_compute(cgraph, &cplan);
+}
+
+static const struct ggml_backend_i ggml_backend_cpu_i = {
+    /* .get_name                = */ ggml_backend_cpu_get_name,
+    /* .free                    = */ ggml_backend_cpu_free,
+    /* .set_tensor_async        = */ NULL,
+    /* .get_tensor_async        = */ NULL,
+    /* .cpy_tensor_async        = */ NULL,
+    /* .synchronize             = */ NULL,
+    /* .graph_plan_create       = */ ggml_backend_cpu_graph_plan_create,
+    /* .graph_plan_free         = */ ggml_backend_cpu_graph_plan_free,
+    /* .graph_plan_update       = */ NULL,
+    /* .graph_plan_compute      = */ ggml_backend_cpu_graph_plan_compute,
+    /* .graph_compute           = */ ggml_backend_cpu_graph_compute,
+    /* .event_record            = */ NULL,
+    /* .event_wait              = */ NULL,
+};
+
+static ggml_guid_t ggml_backend_cpu_guid(void) {
+    static ggml_guid guid = { 0xaa, 0x67, 0xc7, 0x43, 0x96, 0xe6, 0xa3, 0x8a, 0xe3, 0xaf, 0xea, 0x92, 0x36, 0xbc, 0xfc, 0x89 };
+    return &guid;
+}
+
+ggml_backend_t ggml_backend_cpu_init(void) {
+    // initialize CPU backend now to avoid slowing the first graph computation
+    ggml_cpu_init();
+
+    struct ggml_backend_cpu_context * ctx = new ggml_backend_cpu_context;
+    if (ctx == NULL) {
+        return NULL;
+    }
+
+    ctx->n_threads           = GGML_DEFAULT_N_THREADS;
+    ctx->threadpool          = NULL;
+    ctx->work_data           = NULL;
+    ctx->work_size           = 0;
+    ctx->abort_callback      = NULL;
+    ctx->abort_callback_data = NULL;
+
+    ggml_backend_t cpu_backend = new ggml_backend {
+        /* .guid      = */ ggml_backend_cpu_guid(),
+        /* .interface = */ ggml_backend_cpu_i,
+        /* .device    = */ ggml_backend_reg_dev_get(ggml_backend_cpu_reg(), 0),
+        /* .context   = */ ctx,
+    };
+
+    if (cpu_backend == NULL) {
+        delete ctx;
+        return NULL;
+    }
+
+    return cpu_backend;
+}
+
+bool ggml_backend_is_cpu(ggml_backend_t backend) {
+    return backend != NULL && ggml_guid_matches(backend->guid, ggml_backend_cpu_guid());
+}
+
+void ggml_backend_cpu_set_n_threads(ggml_backend_t backend_cpu, int n_threads) {
+    GGML_ASSERT(ggml_backend_is_cpu(backend_cpu));
+
+    struct ggml_backend_cpu_context * ctx = (struct ggml_backend_cpu_context *)backend_cpu->context;
+    ctx->n_threads = n_threads;
+}
+
+void ggml_backend_cpu_set_threadpool(ggml_backend_t backend_cpu, ggml_threadpool_t threadpool) {
+    GGML_ASSERT(ggml_backend_is_cpu(backend_cpu));
+
+    struct ggml_backend_cpu_context * ctx = (struct ggml_backend_cpu_context *)backend_cpu->context;
+
+    if (ctx->threadpool && ctx->threadpool != threadpool) {
+        // already had a different threadpool, pause/suspend it before switching
+        ggml_threadpool_pause(ctx->threadpool);
+    }
+    ctx->threadpool = threadpool;
+}
+
+void ggml_backend_cpu_set_abort_callback(ggml_backend_t backend_cpu, ggml_abort_callback abort_callback, void * abort_callback_data) {
+    GGML_ASSERT(ggml_backend_is_cpu(backend_cpu));
+
+    struct ggml_backend_cpu_context * ctx = (struct ggml_backend_cpu_context *)backend_cpu->context;
+    ctx->abort_callback = abort_callback;
+    ctx->abort_callback_data = abort_callback_data;
+}
+
+// CPU backend - device
+
+struct ggml_backend_cpu_device_context {
+    std::string description = "CPU";
+
+    ggml_backend_cpu_device_context() {
+#ifdef __APPLE__
+        size_t len = 0;
+        if (!sysctlbyname("machdep.cpu.brand_string", NULL, &len, NULL, 0)) {
+            description.resize(len);
+            sysctlbyname("machdep.cpu.brand_string", &description[0], &len, NULL, 0); // NOLINT
+        }
+#elif defined(__linux__)
+        FILE * f = fopen("/proc/cpuinfo", "r");
+        if (f) {
+            char buf[1024];
+            while (fgets(buf, sizeof(buf), f)) {
+                if (strncmp(buf, "model name", 10) == 0) {
+                    char * p = strchr(buf, ':');
+                    if (p) {
+                        p++;
+                        while (std::isspace(*p)) {
+                            p++;
+                        }
+                        while (std::isspace(p[strlen(p) - 1])) {
+                            p[strlen(p) - 1] = '\0';
+                        }
+                        description = p;
+                        break;
+                    }
+                }
+            }
+            fclose(f);
+        }
+#elif defined(_WIN32)
+        HKEY hKey;
+        if (RegOpenKeyEx(HKEY_LOCAL_MACHINE,
+                        TEXT("HARDWARE\\DESCRIPTION\\System\\CentralProcessor\\0"),
+                        0,
+                        KEY_READ,
+                        &hKey) == ERROR_SUCCESS) {
+            DWORD cpu_brand_size = 0;
+            if (RegQueryValueExA(hKey,
+                                TEXT("ProcessorNameString"),
+                                NULL,
+                                NULL,
+                                NULL,
+                                &cpu_brand_size) == ERROR_SUCCESS) {
+                description.resize(cpu_brand_size);
+                if (RegQueryValueExA(hKey,
+                                    TEXT("ProcessorNameString"),
+                                    NULL,
+                                    NULL,
+                                    (LPBYTE)&description[0], // NOLINT
+                                    &cpu_brand_size) == ERROR_SUCCESS) {
+                    if (description.find('\0') != std::string::npos) {
+                        description.resize(description.find('\0'));
+                    }
+                }
+            }
+            RegCloseKey(hKey);
+        }
+#endif
+    }
+};
+
+static const char * ggml_backend_cpu_device_get_name(ggml_backend_dev_t dev) {
+    return "CPU";
+
+    GGML_UNUSED(dev);
+}
+
+static const char * ggml_backend_cpu_device_get_description(ggml_backend_dev_t dev) {
+    struct ggml_backend_cpu_device_context * ctx = (struct ggml_backend_cpu_device_context *)dev->context;
+
+    return ctx->description.c_str();
+}
+
+static void ggml_backend_cpu_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_cpu_device_get_type(ggml_backend_dev_t dev) {
+    return GGML_BACKEND_DEVICE_TYPE_CPU;
+
+    GGML_UNUSED(dev);
+}
+
+static void ggml_backend_cpu_device_get_props(ggml_backend_dev_t dev, struct ggml_backend_dev_props * props) {
+    props->name        = ggml_backend_cpu_device_get_name(dev);
+    props->description = ggml_backend_cpu_device_get_description(dev);
+    props->type        = ggml_backend_cpu_device_get_type(dev);
+    ggml_backend_cpu_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_cpu_device_init_backend(ggml_backend_dev_t dev, const char * params) {
+    return ggml_backend_cpu_init();
+
+    GGML_UNUSED(dev);
+    GGML_UNUSED(params);
+}
+
+static ggml_backend_buffer_type_t ggml_backend_cpu_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_cpu_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_cpu_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];
+
+    if (src0 && src0->buffer && ggml_backend_cpu_buft_is_aarch64(src0->buffer->buft)) {
+        if (op->op != GGML_OP_MUL_MAT || src0->type != GGML_TYPE_Q4_0 || ggml_aarch64_get_optimal_repack_type(src0) == GGML_TYPE_Q4_0) {
+            return false;
+        }
+    }
+
+    for (int i = 1; i < GGML_MAX_SRC; i++) {
+        if (op->src[i] && op->src[i]->buffer && ggml_backend_cpu_buft_is_aarch64(op->src[i]->buffer->buft)) {
+            return false;
+        }
+    }
+
+    switch (op->op) {
+        case GGML_OP_CPY:
+            return
+                op->type != GGML_TYPE_IQ2_XXS &&
+                op->type != GGML_TYPE_IQ2_XS  &&
+                op->type != GGML_TYPE_IQ1_S   &&
+                op->type != GGML_TYPE_IQ1_M; // missing type_traits.from_float
+        case GGML_OP_MUL_MAT:
+            return src1->type == GGML_TYPE_F32 || src1->type == ggml_get_type_traits_cpu(src0->type)->vec_dot_type;
+        case GGML_OP_ROPE_BACK:
+            return op->src[2] == NULL && (op->op_params[2] & 4) == 0;
+        case GGML_OP_IM2COL_BACK:
+            return src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32;
+        case GGML_OP_OUT_PROD:
+            return (src0->type == GGML_TYPE_F32 || ggml_is_quantized(src0->type)) && src1->type == GGML_TYPE_F32;
+        default:
+            return true;
+    }
+
+    GGML_UNUSED(dev);
+}
+
+static bool ggml_backend_cpu_device_supports_buft(ggml_backend_dev_t dev, ggml_backend_buffer_type_t buft) {
+    return ggml_backend_buft_is_host(buft) || ggml_backend_cpu_buft_is_aarch64(buft);
+
+    GGML_UNUSED(dev);
+}
+
+static const struct ggml_backend_device_i ggml_backend_cpu_device_i = {
+    /* .get_name             = */ ggml_backend_cpu_device_get_name,
+    /* .get_description      = */ ggml_backend_cpu_device_get_description,
+    /* .get_memory           = */ ggml_backend_cpu_device_get_memory,
+    /* .get_type             = */ ggml_backend_cpu_device_get_type,
+    /* .get_props            = */ ggml_backend_cpu_device_get_props,
+    /* .init_backend         = */ ggml_backend_cpu_device_init_backend,
+    /* .get_buffer_type      = */ ggml_backend_cpu_device_get_buffer_type,
+    /* .get_host_buffer_type = */ NULL,
+    /* .buffer_from_host_ptr = */ ggml_backend_cpu_device_buffer_from_host_ptr,
+    /* .supports_op          = */ ggml_backend_cpu_device_supports_op,
+    /* .supports_buft        = */ ggml_backend_cpu_device_supports_buft,
+    /* .offload_op           = */ NULL,
+    /* .event_new            = */ NULL,
+    /* .event_free           = */ NULL,
+    /* .event_synchronize    = */ NULL,
+};
+
+// CPU backend - backend (reg)
+
+static const char * ggml_backend_cpu_reg_get_name(ggml_backend_reg_t reg) {
+    return "CPU";
+
+    GGML_UNUSED(reg);
+}
+
+static size_t ggml_backend_cpu_reg_get_device_count(ggml_backend_reg_t reg) {
+    return 1;
+
+    GGML_UNUSED(reg);
+}
+
+static ggml_backend_dev_t ggml_backend_cpu_reg_get_device(ggml_backend_reg_t reg, size_t index) {
+    GGML_ASSERT(index == 0);
+
+    static ggml_backend_cpu_device_context ctx;
+    static ggml_backend_device ggml_backend_cpu_device = {
+        /* .iface   = */ ggml_backend_cpu_device_i,
+        /* .reg     = */ reg,
+        /* .context = */ &ctx,
+    };
+
+    return &ggml_backend_cpu_device;
+}
+
+struct ggml_backend_feature {
+    const char * name;
+    const char * value;
+};
+
+// Not used yet
+// This is intended to replace the the ggml_cpu_has_* functions when loading the CPU backend dynamically,
+// and additionally to allow other backends to expose their own list of features that applications can query using the same API.
+static ggml_backend_feature * ggml_backend_cpu_get_features(ggml_backend_reg_t reg) {
+    static std::vector<ggml_backend_feature> features = []() {
+        std::vector<ggml_backend_feature> features;
+        if (ggml_cpu_has_sse3()) {
+            features.push_back({ "SSE3", "1" });
+        }
+        if (ggml_cpu_has_ssse3()) {
+            features.push_back({ "SSSE3", "1" });
+        }
+        if (ggml_cpu_has_avx()) {
+            features.push_back({ "AVX", "1" });
+        }
+        if (ggml_cpu_has_avx2()) {
+            features.push_back({ "AVX2", "1" });
+        }
+        if (ggml_cpu_has_f16c()) {
+            features.push_back({ "F16C", "1" });
+        }
+        if (ggml_cpu_has_fma()) {
+            features.push_back({ "FMA", "1" });
+        }
+        if (ggml_cpu_has_avx_vnni()) {
+            features.push_back({ "AVX_VNNI", "1" });
+        }
+        if (ggml_cpu_has_avx512()) {
+            features.push_back({ "AVX512", "1" });
+        }
+        if (ggml_cpu_has_avx512_vbmi()) {
+            features.push_back({ "AVX512_VBMI", "1" });
+        }
+        if (ggml_cpu_has_avx512_vnni()) {
+            features.push_back({ "AVX512_VNNI", "1" });
+        }
+        if (ggml_cpu_has_avx512_bf16()) {
+            features.push_back({ "AVX512_BF16", "1" });
+        }
+        if (ggml_cpu_has_amx_int8()) {
+            features.push_back({ "AMX_INT8", "1" });
+        }
+        if (ggml_cpu_has_neon()) {
+            features.push_back({ "NEON", "1" });
+        }
+        if (ggml_cpu_has_arm_fma()) {
+            features.push_back({ "ARM_FMA", "1" });
+        }
+        if (ggml_cpu_has_fp16_va()) {
+            features.push_back({ "FP16_VA", "1" });
+        }
+        if (ggml_cpu_has_matmul_int8()) {
+            features.push_back({ "MATMUL_INT8", "1" });
+        }
+        if (ggml_cpu_has_sve()) {
+            features.push_back({ "SVE", "1" });
+        }
+        if (ggml_cpu_get_sve_cnt() > 0) {
+            static std::string sve_cnt = std::to_string(ggml_cpu_get_sve_cnt());
+            features.push_back({ "SVE_CNT", sve_cnt.c_str() });
+        }
+        if (ggml_cpu_has_riscv_v()) {
+            features.push_back({ "RISCV_V", "1" });
+        }
+        if (ggml_cpu_has_vsx()) {
+            features.push_back({ "VSX", "1" });
+        }
+        if (ggml_cpu_has_wasm_simd()) {
+            features.push_back({ "WASM_SIMD", "1" });
+        }
+        if (ggml_cpu_has_llamafile()) {
+            features.push_back({ "LLAMAFILE", "1" });
+        }
+
+        features.push_back({ nullptr, nullptr });
+
+        return features;
+    }();
+
+    return features.data();
+
+    GGML_UNUSED(reg);
+}
+
+static void * ggml_backend_cpu_get_proc_address(ggml_backend_reg_t reg, const char * name) {
+    if (strcmp(name, "ggml_backend_set_n_threads") == 0) {
+        return (void *)ggml_backend_cpu_set_n_threads;
+    }
+    if (strcmp(name, "ggml_backend_dev_get_extra_bufts") == 0) {
+        return (void *)ggml_backend_cpu_get_extra_bufts;
+    }
+
+    return NULL;
+
+    GGML_UNUSED(reg);
+}
+
+static const struct ggml_backend_reg_i ggml_backend_cpu_reg_i = {
+    /* .get_name         = */ ggml_backend_cpu_reg_get_name,
+    /* .get_device_count = */ ggml_backend_cpu_reg_get_device_count,
+    /* .get_device       = */ ggml_backend_cpu_reg_get_device,
+    /* .get_proc_address = */ ggml_backend_cpu_get_proc_address,
+};
+
+ggml_backend_reg_t ggml_backend_cpu_reg(void) {
+    // init CPU feature detection
+    ggml_cpu_init();
+
+    static struct ggml_backend_reg ggml_backend_cpu_reg = {
+        /* .iface   = */ ggml_backend_cpu_reg_i,
+        /* .context = */ NULL,
+    };
+
+    return &ggml_backend_cpu_reg;
+}

ggml/src/ggml-cuda/CMakeLists.txt

@@ -0,0 +1,155 @@
+cmake_minimum_required(VERSION 3.18)  # for CMAKE_CUDA_ARCHITECTURES
+
+find_package(CUDAToolkit)
+
+if (CUDAToolkit_FOUND)
+    message(STATUS "CUDA Toolkit found")
+
+    if (NOT DEFINED CMAKE_CUDA_ARCHITECTURES)
+        # native == GPUs available at build time
+        # 52     == Maxwell, lowest CUDA 12 standard
+        # 60     == P100, FP16 CUDA intrinsics
+        # 61     == Pascal, __dp4a instruction (per-byte integer dot product)
+        # 70     == V100, FP16 tensor cores
+        # 75     == Turing, int8 tensor cores
+        if (GGML_NATIVE AND CUDAToolkit_VERSION VERSION_GREATER_EQUAL "11.6" AND CMAKE_VERSION VERSION_GREATER_EQUAL "3.24")
+            set(CMAKE_CUDA_ARCHITECTURES "native")
+        elseif(GGML_CUDA_F16 OR GGML_CUDA_DMMV_F16)
+            set(CMAKE_CUDA_ARCHITECTURES "60;61;70;75")
+        else()
+            set(CMAKE_CUDA_ARCHITECTURES "52;61;70;75")
+        endif()
+    endif()
+    message(STATUS "Using CUDA architectures: ${CMAKE_CUDA_ARCHITECTURES}")
+
+    enable_language(CUDA)
+
+    file(GLOB   GGML_HEADERS_CUDA "*.cuh")
+    list(APPEND GGML_HEADERS_CUDA "../../include/ggml-cuda.h")
+
+    file(GLOB   GGML_SOURCES_CUDA "*.cu")
+    file(GLOB   SRCS "template-instances/fattn-wmma*.cu")
+    list(APPEND GGML_SOURCES_CUDA ${SRCS})
+    file(GLOB   SRCS "template-instances/mmq*.cu")
+    list(APPEND GGML_SOURCES_CUDA ${SRCS})
+
+    if (GGML_CUDA_FA_ALL_QUANTS)
+        file(GLOB   SRCS "template-instances/fattn-vec*.cu")
+        list(APPEND GGML_SOURCES_CUDA ${SRCS})
+        add_compile_definitions(GGML_CUDA_FA_ALL_QUANTS)
+    else()
+        file(GLOB   SRCS "template-instances/fattn-vec*q4_0-q4_0.cu")
+        list(APPEND GGML_SOURCES_CUDA ${SRCS})
+        file(GLOB   SRCS "template-instances/fattn-vec*q8_0-q8_0.cu")
+        list(APPEND GGML_SOURCES_CUDA ${SRCS})
+        file(GLOB   SRCS "template-instances/fattn-vec*f16-f16.cu")
+        list(APPEND GGML_SOURCES_CUDA ${SRCS})
+    endif()
+
+    add_library(ggml-cuda
+                ${GGML_HEADERS_CUDA}
+                ${GGML_SOURCES_CUDA}
+                )
+
+    target_link_libraries(ggml-cuda PRIVATE ggml-base)
+    target_include_directories(ggml-cuda PRIVATE . ..)
+
+    add_compile_definitions(GGML_CUDA_PEER_MAX_BATCH_SIZE=${GGML_CUDA_PEER_MAX_BATCH_SIZE})
+
+    if (GGML_CUDA_GRAPHS)
+        add_compile_definitions(GGML_CUDA_USE_GRAPHS)
+    endif()
+
+    if (GGML_CUDA_FORCE_MMQ)
+        add_compile_definitions(GGML_CUDA_FORCE_MMQ)
+    endif()
+
+    if (GGML_CUDA_FORCE_CUBLAS)
+        add_compile_definitions(GGML_CUDA_FORCE_CUBLAS)
+    endif()
+
+    if (GGML_CUDA_NO_VMM)
+        add_compile_definitions(GGML_CUDA_NO_VMM)
+    endif()
+
+    if (GGML_CUDA_F16 OR GGML_CUDA_DMMV_F16)
+        add_compile_definitions(GGML_CUDA_F16)
+    endif()
+
+    if (GGML_CUDA_NO_PEER_COPY)
+        add_compile_definitions(GGML_CUDA_NO_PEER_COPY)
+    endif()
+
+    if (GGML_STATIC)
+        if (WIN32)
+            # As of 12.3.1 CUDA Toolkit for Windows does not offer a static cublas library
+            target_link_libraries(ggml-cuda PRIVATE CUDA::cudart_static CUDA::cublas CUDA::cublasLt)
+        else ()
+            target_link_libraries(ggml-cuda PRIVATE  CUDA::cudart_static CUDA::cublas_static CUDA::cublasLt_static)
+        endif()
+    else()
+        target_link_libraries(ggml-cuda PRIVATE CUDA::cudart CUDA::cublas CUDA::cublasLt)
+    endif()
+
+    if (GGML_CUDA_NO_VMM)
+        # No VMM requested, no need to link directly with the cuda driver lib (libcuda.so)
+    else()
+        target_link_libraries(ggml-cuda PRIVATE CUDA::cuda_driver)
+    endif()
+
+    set(CUDA_CXX_FLAGS "")
+
+    set(CUDA_FLAGS -use_fast_math)
+
+    if (GGML_FATAL_WARNINGS)
+        list(APPEND CUDA_FLAGS -Werror all-warnings)
+    endif()
+
+    if (GGML_ALL_WARNINGS AND NOT MSVC)
+        set(NVCC_CMD ${CMAKE_CUDA_COMPILER} .c)
+        if (NOT CMAKE_CUDA_HOST_COMPILER STREQUAL "")
+            list(APPEND NVCC_CMD -ccbin ${CMAKE_CUDA_HOST_COMPILER})
+        endif()
+
+        execute_process(
+            COMMAND ${NVCC_CMD} -Xcompiler --version
+            OUTPUT_VARIABLE CUDA_CCFULLVER
+            ERROR_QUIET
+        )
+
+        if (NOT CUDA_CCFULLVER MATCHES clang)
+            set(CUDA_CCID "GNU")
+            execute_process(
+                COMMAND ${NVCC_CMD} -Xcompiler "-dumpfullversion -dumpversion"
+                OUTPUT_VARIABLE CUDA_CCVER
+                ERROR_QUIET
+            )
+        else()
+            if (CUDA_CCFULLVER MATCHES Apple)
+                set(CUDA_CCID "AppleClang")
+            else()
+                set(CUDA_CCID "Clang")
+            endif()
+            string(REGEX REPLACE "^.* version ([0-9.]*).*$" "\\1" CUDA_CCVER ${CUDA_CCFULLVER})
+        endif()
+
+        message("-- CUDA host compiler is ${CUDA_CCID} ${CUDA_CCVER}")
+
+        get_flags(${CUDA_CCID} ${CUDA_CCVER})
+        list(APPEND CUDA_CXX_FLAGS ${CXX_FLAGS} ${GF_CXX_FLAGS})  # This is passed to -Xcompiler later
+    endif()
+
+    if (NOT MSVC)
+        list(APPEND CUDA_CXX_FLAGS -Wno-pedantic)
+    endif()
+
+    list(JOIN   CUDA_CXX_FLAGS " " CUDA_CXX_FLAGS_JOINED)  # pass host compiler flags as a single argument
+
+    if (NOT CUDA_CXX_FLAGS_JOINED STREQUAL "")
+        list(APPEND CUDA_FLAGS -Xcompiler ${CUDA_CXX_FLAGS_JOINED})
+    endif()
+
+    target_compile_options(ggml-cuda PRIVATE "$<$<COMPILE_LANGUAGE:CUDA>:${CUDA_FLAGS}>")
+else()
+    message(FATAL_ERROR "CUDA Toolkit not found")
+endif()

ggml/src/ggml-cuda/common.cuh

@@ -6,7 +6,7 @@
 #include <cstdint>
 #include <memory>
 
-#if defined(GGML_USE_HIPBLAS)
+#if defined(GGML_USE_HIP)
 #define GGML_COMMON_DECL_HIP
 #define GGML_COMMON_IMPL_HIP
 #else
@@ -26,13 +26,13 @@
 #include <string>
 #include <vector>
 
-#if defined(GGML_USE_HIPBLAS)
+#if defined(GGML_USE_HIP)
 #include "vendors/hip.h"
 #elif defined(GGML_USE_MUSA)
 #include "vendors/musa.h"
 #else
 #include "vendors/cuda.h"
-#endif // defined(GGML_USE_HIPBLAS)
+#endif // defined(GGML_USE_HIP)
 
 #define STRINGIZE_IMPL(...) #__VA_ARGS__
 #define STRINGIZE(...) STRINGIZE_IMPL(__VA_ARGS__)
@@ -97,7 +97,7 @@ void ggml_cuda_error(const char * stmt, const char * func, const char * file, in
 
 #define CUBLAS_CHECK(err) CUDA_CHECK_GEN(err, CUBLAS_STATUS_SUCCESS, cublas_get_error_str)
 
-#if !defined(GGML_USE_HIPBLAS)
+#if !defined(GGML_USE_HIP)
 static const char * cu_get_error_str(CUresult err) {
     const char * err_str;
     cuGetErrorString(err, &err_str);
@@ -120,21 +120,21 @@ typedef float dfloat; // dequantize float
 typedef float2 dfloat2;
 #endif // GGML_CUDA_F16
 
-#if (defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) || __CUDA_ARCH__ >= CC_PASCAL
+#if (defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) || __CUDA_ARCH__ >= CC_PASCAL
 #define FP16_AVAILABLE
-#endif // (defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) || __CUDA_ARCH__ >= CC_PASCAL
+#endif // (defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) || __CUDA_ARCH__ >= CC_PASCAL
 
 #if defined(FP16_AVAILABLE) && __CUDA_ARCH__ != 610
 #define FAST_FP16_AVAILABLE
 #endif // defined(FP16_AVAILABLE) && __CUDA_ARCH__ != 610
 
-#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_VOLTA
+#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_VOLTA
 #define FP16_MMA_AVAILABLE
-#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_VOLTA
+#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_VOLTA
 
-#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_TURING
+#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_TURING
 #define INT8_MMA_AVAILABLE
-#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_TURING
+#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_TURING
 
 #if !(defined(GGML_USE_MUSA) && __MUSA_ARCH__ <= CC_QY1)
 #define FLASH_ATTN_AVAILABLE
@@ -156,14 +156,14 @@ static constexpr bool int8_mma_available(const int cc) {
 static __device__ void no_device_code(
     const char * file_name, const int line, const char * function_name, const int arch, const char * arch_list) {
 
-#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+#if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
     printf("%s:%d: ERROR: HIP kernel %s has no device code compatible with HIP arch %d.\n",
            file_name, line, function_name, arch);
     GGML_UNUSED(arch_list);
 #else
     printf("%s:%d: ERROR: CUDA kernel %s has no device code compatible with CUDA arch %d. ggml-cuda.cu was compiled for: %s\n",
            file_name, line, function_name, arch, arch_list);
-#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+#endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
     __trap();
 
     GGML_UNUSED(no_device_code); // suppress unused function warning
@@ -176,7 +176,7 @@ static __device__ void no_device_code(
 #endif // __CUDA_ARCH__
 
 static __device__ __forceinline__ int warp_reduce_sum(int x) {
-#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_AMPERE
+#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_AMPERE
     return __reduce_add_sync(0xffffffff, x);
 #else
 #pragma unroll
@@ -184,7 +184,7 @@ static __device__ __forceinline__ int warp_reduce_sum(int x) {
         x += __shfl_xor_sync(0xffffffff, x, mask, 32);
     }
     return x;
-#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_AMPERE
+#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_AMPERE
 }
 
 static __device__ __forceinline__ float warp_reduce_sum(float x) {
@@ -207,7 +207,7 @@ static __device__ __forceinline__ float2 warp_reduce_sum(float2 a) {
 static __device__ __forceinline__ half2 warp_reduce_sum(half2 a) {
 #ifdef FP16_AVAILABLE
 
-#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+#if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
 #pragma unroll
     for (int mask = 16; mask > 0; mask >>= 1) {
         const half2 a_other = __shfl_xor_sync(0xffffffff, a, mask, 32);
@@ -221,7 +221,7 @@ static __device__ __forceinline__ half2 warp_reduce_sum(half2 a) {
         a = __hadd2(a, __shfl_xor_sync(0xffffffff, a, mask, 32));
     }
     return a;
-#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+#endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
 
 #else
     NO_DEVICE_CODE;
@@ -240,11 +240,11 @@ static __device__ __forceinline__ float warp_reduce_max(float x) {
 static __device__ __forceinline__ half ggml_cuda_hmax(const half a, const half b) {
 #ifdef FP16_AVAILABLE
 
-#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && CUDART_VERSION < CUDART_HMAX
+#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && CUDART_VERSION < CUDART_HMAX
     return __float2half(fmaxf(__half2float(a), __half2float(b)));
 #else
     return __hmax(a, b);
-#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && CUDART_VERSION < CUDART_HMAX
+#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && CUDART_VERSION < CUDART_HMAX
 
 #else
    NO_DEVICE_CODE;
@@ -254,7 +254,7 @@ static __device__ __forceinline__ half ggml_cuda_hmax(const half a, const half b
 }
 
 static __device__ __forceinline__ half2 ggml_cuda_hmax2(const half2 a, const half2 b) {
-#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
 
 #if CUDART_VERSION >= CUDART_HMAX
     return __hmax2(a, b);
@@ -269,11 +269,11 @@ static __device__ __forceinline__ half2 ggml_cuda_hmax2(const half2 a, const hal
     GGML_UNUSED(a);
     GGML_UNUSED(b);
     NO_DEVICE_CODE;
-#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
 }
 
 static __device__ __forceinline__ half2 warp_reduce_max(half2 x) {
-#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL
+#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL
 #pragma unroll
    for (int mask = 16; mask > 0; mask >>= 1) {
        x = ggml_cuda_hmax2(x, __shfl_xor_sync(0xffffffff, x, mask, 32));
@@ -282,7 +282,7 @@ static __device__ __forceinline__ half2 warp_reduce_max(half2 x) {
 #else
    GGML_UNUSED(x);
    NO_DEVICE_CODE;
-#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL
+#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL
 }
 
 #if CUDART_VERSION < CUDART_HMASK
@@ -294,7 +294,7 @@ static __device__ __forceinline__ uint32_t __hgt2_mask(const half2 a, const half
 #endif // CUDART_VERSION < CUDART_HMASK
 
 static __device__ __forceinline__ int ggml_cuda_dp4a(const int a, const int b, int c) {
-#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+#if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
 #if defined(__gfx906__) || defined(__gfx908__) || defined(__gfx90a__) || defined(RDNA2)
     c = __builtin_amdgcn_sdot4(a, b, c, false);
 #elif defined(RDNA3)
@@ -320,7 +320,7 @@ static __device__ __forceinline__ int ggml_cuda_dp4a(const int a, const int b, i
 #endif
     return c;
 
-#else // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+#else // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
 
 #if __CUDA_ARCH__ >= MIN_CC_DP4A
     return __dp4a(a, b, c);
@@ -330,7 +330,7 @@ static __device__ __forceinline__ int ggml_cuda_dp4a(const int a, const int b, i
     return c + a8[0]*b8[0] + a8[1]*b8[1] + a8[2]*b8[2] + a8[3]*b8[3];
 #endif // __CUDA_ARCH__ >= MIN_CC_DP4A
 
-#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+#endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
 }
 
 // TODO: move to ggml-common.h

ggml/src/ggml-cuda/fattn-common.cuh

@@ -517,9 +517,9 @@ constexpr __device__ dequantize_1_f32_t get_dequantize_1_f32(ggml_type type_V) {
 }
 
 template<int D, int parallel_blocks> // D == head size
-#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
 __launch_bounds__(D, 1)
-#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
 static __global__ void flash_attn_combine_results(
         const float  * __restrict__ VKQ_parts,
         const float2 * __restrict__ VKQ_meta,

ggml/src/ggml-cuda/fattn-tile-f16.cu

@@ -5,9 +5,9 @@
 #define FATTN_KQ_STRIDE_TILE_F16 64
 
 template<int D, int ncols, int nwarps, int parallel_blocks, bool use_logit_softcap> // D == head size
-#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
 __launch_bounds__(nwarps*WARP_SIZE, 1)
-#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
 static __global__ void flash_attn_tile_ext_f16(
         const char * __restrict__ Q,
         const char * __restrict__ K,

ggml/src/ggml-cuda/fattn-tile-f32.cu

@@ -5,9 +5,9 @@
 #define FATTN_KQ_STRIDE_TILE_F32 32
 
 template<int D, int ncols, int nwarps, int parallel_blocks, bool use_logit_softcap> // D == head size
-#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
 __launch_bounds__(nwarps*WARP_SIZE, 1)
-#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
 static __global__ void flash_attn_tile_ext_f32(
         const char * __restrict__ Q,
         const char * __restrict__ K,

ggml/src/ggml-cuda/fattn-vec-f16.cuh

@@ -2,9 +2,9 @@
 #include "fattn-common.cuh"
 
 template<int D, int ncols, int parallel_blocks, ggml_type type_K, ggml_type type_V, bool use_logit_softcap> // D == head size
-#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
 __launch_bounds__(D, 1)
-#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
 static __global__ void flash_attn_vec_ext_f16(
         const char * __restrict__ Q,
         const char * __restrict__ K,

ggml/src/ggml-cuda/fattn-vec-f32.cuh

@@ -2,9 +2,9 @@
 #include "fattn-common.cuh"
 
 template<int D, int ncols, int parallel_blocks, ggml_type type_K, ggml_type type_V, bool use_logit_softcap> // D == head size
-#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
 __launch_bounds__(D, 1)
-#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
 static __global__ void flash_attn_vec_ext_f32(
         const char * __restrict__ Q,
         const char * __restrict__ K,

ggml/src/ggml-cuda/fattn-wmma-f16.cuh

@@ -7,9 +7,9 @@
 
 // D == head size, VKQ_stride == num VKQ rows calculated in parallel:
 template<int D, int ncols, int nwarps, int VKQ_stride, int parallel_blocks, typename KQ_acc_t, bool use_logit_softcap>
-#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
 __launch_bounds__(nwarps*WARP_SIZE, 1)
-#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
 static __global__ void flash_attn_ext_f16(
         const char * __restrict__ Q,
         const char * __restrict__ K,

ggml/src/ggml-cuda/ggml-cuda.cu

@@ -16,11 +16,11 @@
 #include "ggml-cuda/cpy.cuh"
 #include "ggml-cuda/cross-entropy-loss.cuh"
 #include "ggml-cuda/diagmask.cuh"
-#include "ggml-cuda/dmmv.cuh"
 #include "ggml-cuda/fattn.cuh"
 #include "ggml-cuda/getrows.cuh"
 #include "ggml-cuda/im2col.cuh"
 #include "ggml-cuda/mmq.cuh"
+#include "ggml-cuda/mmv.cuh"
 #include "ggml-cuda/mmvq.cuh"
 #include "ggml-cuda/norm.cuh"
 #include "ggml-cuda/opt-step-adamw.cuh"
@@ -91,7 +91,7 @@ int ggml_cuda_get_device() {
 
 static cudaError_t ggml_cuda_device_malloc(void ** ptr, size_t size, int device) {
     ggml_cuda_set_device(device);
-#if defined(GGML_USE_HIPBLAS) && defined(GGML_HIP_UMA)
+#if defined(GGML_USE_HIP) && defined(GGML_HIP_UMA)
     auto res = hipMallocManaged(ptr, size);
     if (res == hipSuccess) {
         // if error we "need" to know why...
@@ -100,7 +100,7 @@ static cudaError_t ggml_cuda_device_malloc(void ** ptr, size_t size, int device)
     return res;
 #else
 
-#if !defined(GGML_USE_HIPBLAS)
+#if !defined(GGML_USE_HIP)
     cudaError_t err;
     if (getenv("GGML_CUDA_ENABLE_UNIFIED_MEMORY") != nullptr)
     {
@@ -113,7 +113,7 @@ static cudaError_t ggml_cuda_device_malloc(void ** ptr, size_t size, int device)
     return err;
 #else
     return cudaMalloc(ptr, size);
-#endif // !defined(GGML_USE_HIPBLAS)
+#endif // !defined(GGML_USE_HIP)
 
 #endif
 }
@@ -151,7 +151,7 @@ static ggml_cuda_device_info ggml_cuda_init() {
     for (int id = 0; id < info.device_count; ++id) {
         int device_vmm = 0;
 
-#if !defined(GGML_USE_HIPBLAS) && !defined(GGML_CUDA_NO_VMM)
+#if !defined(GGML_USE_HIP) && !defined(GGML_CUDA_NO_VMM)
         CUdevice device;
         CU_CHECK(cuDeviceGet(&device, id));
         CU_CHECK(cuDeviceGetAttribute(&device_vmm, CU_DEVICE_ATTRIBUTE_VIRTUAL_MEMORY_MANAGEMENT_SUPPORTED, device));
@@ -163,7 +163,7 @@ static ggml_cuda_device_info ggml_cuda_init() {
             alloc_prop.location.id = id;
             CU_CHECK(cuMemGetAllocationGranularity(&info.devices[id].vmm_granularity, &alloc_prop, CU_MEM_ALLOC_GRANULARITY_RECOMMENDED));
         }
-#endif // !defined(GGML_USE_HIPBLAS) && !defined(GGML_CUDA_NO_VMM)
+#endif // !defined(GGML_USE_HIP) && !defined(GGML_CUDA_NO_VMM)
         info.devices[id].vmm = !!device_vmm;
 
         cudaDeviceProp prop;
@@ -175,13 +175,13 @@ static ggml_cuda_device_info ggml_cuda_init() {
 
         info.devices[id].nsm   = prop.multiProcessorCount;
         info.devices[id].smpb  = prop.sharedMemPerBlock;
-#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+#if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
         info.devices[id].smpbo = prop.sharedMemPerBlock;
         info.devices[id].cc = 100*prop.major + 10*prop.minor + CC_OFFSET_AMD;
 #else
         info.devices[id].smpbo = prop.sharedMemPerBlockOptin;
         info.devices[id].cc = 100*prop.major + 10*prop.minor;
-#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+#endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
     }
 
     for (int id = 0; id < info.device_count; ++id) {
@@ -299,7 +299,7 @@ struct ggml_cuda_pool_leg : public ggml_cuda_pool {
 };
 
 // pool with virtual memory
-#if !defined(GGML_USE_HIPBLAS) && !defined(GGML_CUDA_NO_VMM)
+#if !defined(GGML_USE_HIP) && !defined(GGML_CUDA_NO_VMM)
 struct ggml_cuda_pool_vmm : public ggml_cuda_pool {
     static const size_t CUDA_POOL_VMM_MAX_SIZE = 1ull << 35; // 32 GB
 
@@ -393,14 +393,14 @@ struct ggml_cuda_pool_vmm : public ggml_cuda_pool {
         GGML_ASSERT(ptr == (void *) (pool_addr + pool_used));
     }
 };
-#endif // !defined(GGML_USE_HIPBLAS) && !defined(GGML_CUDA_NO_VMM)
+#endif // !defined(GGML_USE_HIP) && !defined(GGML_CUDA_NO_VMM)
 
 std::unique_ptr<ggml_cuda_pool> ggml_backend_cuda_context::new_pool_for_device(int device) {
-#if !defined(GGML_USE_HIPBLAS) && !defined(GGML_CUDA_NO_VMM)
+#if !defined(GGML_USE_HIP) && !defined(GGML_CUDA_NO_VMM)
     if (ggml_cuda_info().devices[device].vmm) {
         return std::unique_ptr<ggml_cuda_pool>(new ggml_cuda_pool_vmm(device));
     }
-#endif // !defined(GGML_USE_HIPBLAS) && !defined(GGML_CUDA_NO_VMM)
+#endif // !defined(GGML_USE_HIP) && !defined(GGML_CUDA_NO_VMM)
     return std::unique_ptr<ggml_cuda_pool>(new ggml_cuda_pool_leg(device));
 }
 
@@ -1020,114 +1020,6 @@ typedef void (*ggml_cuda_op_mul_mat_t)(
 
 #define MUL_MAT_SRC1_COL_STRIDE 128
 
-static __global__ void mul_mat_p021_f16_f32(
-    const void * __restrict__ vx, const float * __restrict__ y, float * __restrict__ dst,
-    const int ncols_x, const int nrows_x, const int nchannels_x, const int nchannels_y) {
-
-    const half * x = (const half *) vx;
-
-    const int row_x = blockDim.y*blockIdx.y + threadIdx.y;
-    const int channel = blockDim.z*blockIdx.z + threadIdx.z;
-    const int channel_x = channel / (nchannels_y / nchannels_x);
-
-    const int nrows_y = ncols_x;
-    const int nrows_dst = nrows_x;
-    const int row_dst = row_x;
-
-    float tmp = 0.0f;
-
-    for (int col_x0 = 0; col_x0 < ncols_x; col_x0 += blockDim.x) {
-        const int col_x = col_x0 + threadIdx.x;
-
-        if (col_x >= ncols_x) {
-            break;
-        }
-
-        // x is transposed and permuted
-        const int ix = row_x*nchannels_x*ncols_x + channel_x*ncols_x + col_x;
-        const float xi = __half2float(x[ix]);
-
-        const int row_y = col_x;
-
-        // y is not transposed but permuted
-        const int iy = channel*nrows_y + row_y;
-
-        tmp += xi * y[iy];
-    }
-
-    // dst is not transposed and not permuted
-    const int idst = channel*nrows_dst + row_dst;
-
-    // sum up partial sums and write back result
-    tmp = warp_reduce_sum(tmp);
-
-    if (threadIdx.x == 0) {
-        dst[idst] = tmp;
-    }
-}
-
-static __global__ void mul_mat_vec_nc_f16_f32( // nc == non-contiguous
-    const void * __restrict__ vx, const float * __restrict__ y, float * __restrict__ dst, const int ncols_x, const int nrows_x,
-    const int row_stride_x, const int channel_stride_x, const int channel_x_divisor) {
-
-    const half * x = (const half *) vx;
-
-    const int row_x     = blockDim.y*blockIdx.y + threadIdx.y;
-    const int channel   = blockDim.z*blockIdx.z + threadIdx.z;
-    const int channel_x = channel / channel_x_divisor;
-
-    const int nrows_y   = ncols_x;
-    const int nrows_dst = nrows_x;
-    const int row_dst   = row_x;
-
-    const int idst = channel*nrows_dst + row_dst;
-
-    float tmp = 0.0f;
-
-    for (int col_x0 = 0; col_x0 < ncols_x; col_x0 += blockDim.x) {
-        const int col_x = col_x0 + threadIdx.x;
-
-        if (col_x >= ncols_x) {
-            break;
-        }
-
-        const int row_y = col_x;
-
-        const int ix = channel_x*channel_stride_x + row_x*row_stride_x + col_x;
-        const int iy = channel*nrows_y + row_y;
-
-        const float xi = __half2float(x[ix]);
-
-        tmp += xi * y[iy];
-    }
-
-    // sum up partial sums and write back result
-    tmp = warp_reduce_sum(tmp);
-
-    if (threadIdx.x == 0) {
-        dst[idst] = tmp;
-    }
-}
-
-static void ggml_mul_mat_p021_f16_f32_cuda(
-    const void * vx, const float * y, float * dst, const int ncols_x, const int nrows_x,
-    const int nchannels_x, const int nchannels_y, cudaStream_t stream) {
-
-    const dim3 block_nums(1, nrows_x, nchannels_y);
-    const dim3 block_dims(WARP_SIZE, 1, 1);
-    mul_mat_p021_f16_f32<<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols_x, nrows_x, nchannels_x, nchannels_y);
-}
-
-static void ggml_mul_mat_vec_nc_f16_f32_cuda(
-    const void * vx, const float * y, float * dst, const int ncols_x, const int nrows_x, const int row_stride_x,
-    const int nchannels_x, const int nchannels_y, const int channel_stride_x, cudaStream_t stream) {
-
-    const dim3 block_nums(1, nrows_x, nchannels_y);
-    const dim3 block_dims(WARP_SIZE, 1, 1);
-    mul_mat_vec_nc_f16_f32<<<block_nums, block_dims, 0, stream>>>
-        (vx, y, dst, ncols_x, nrows_x, row_stride_x, channel_stride_x, nchannels_y/nchannels_x);
-}
-
 static cudaError_t ggml_cuda_cpy_tensor_2d(
     void * dst, const struct ggml_tensor * src, int64_t i3, int64_t i2, int64_t i1_low, int64_t i1_high, cudaStream_t stream) {
 
@@ -1325,7 +1217,7 @@ static void ggml_cuda_set_peer_access(const int n_tokens, int main_device) {
 static cudaError_t ggml_cuda_Memcpy2DPeerAsync(
     void * dst, int dstDevice, size_t dpitch, void * src, int srcDevice, size_t spitch, size_t width, size_t height, cudaStream_t stream) {
 
-#if !defined(GGML_USE_HIPBLAS) && !defined(GGML_USE_MUSA)
+#if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
     // cudaMemcpy2DAsync may fail with copies between vmm pools of different devices
     cudaMemcpy3DPeerParms p = {};
     p.dstDevice = dstDevice;
@@ -1339,7 +1231,7 @@ static cudaError_t ggml_cuda_Memcpy2DPeerAsync(
     GGML_UNUSED(dstDevice);
     GGML_UNUSED(srcDevice);
     return cudaMemcpy2DAsync(dst, dpitch, src, spitch, width, height, cudaMemcpyDeviceToDevice, stream);
-#endif // !defined(GGML_USE_HIPBLAS) && !defined(GGML_USE_MUSA)
+#endif // !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
 }
 
 static void ggml_cuda_op_mul_mat(
@@ -1654,58 +1546,6 @@ static void ggml_cuda_op_mul_mat(
     }
 }
 
-static void ggml_cuda_mul_mat_vec_p021(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
-    GGML_ASSERT(ggml_is_permuted(src0) && ggml_is_permuted(src1));
-    GGML_ASSERT(ggml_backend_buffer_is_cuda(src0->buffer));
-    GGML_ASSERT(src0->nb[0] <= src0->nb[1] && src0->nb[2] <= src0->nb[3]); // 0213 permutation
-    GGML_ASSERT(src1->nb[0] <= src1->nb[1] && src1->nb[2] <= src1->nb[3]); // 0213 permutation
-    GGML_ASSERT(src0->type == GGML_TYPE_F16);
-    GGML_ASSERT(src1->type == GGML_TYPE_F32);
-
-    const int64_t ne00 = src0->ne[0];
-    const int64_t ne01 = src0->ne[1];
-    const int64_t ne02 = src0->ne[2];
-
-    const int64_t ne12 = src1->ne[2];
-
-    cudaStream_t main_stream = ctx.stream();
-
-    void  * src0_ddq = src0->data;
-    float * src1_ddf = (float *) src1->data;
-    float * dst_ddf  = (float *) dst->data;
-
-    ggml_mul_mat_p021_f16_f32_cuda(src0_ddq, src1_ddf, dst_ddf, ne00, ne01, ne02, ne12, main_stream);
-}
-
-static void ggml_cuda_mul_mat_vec_nc(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
-    GGML_ASSERT(!ggml_is_transposed(src0));
-    GGML_ASSERT(!ggml_is_transposed(src1));
-    GGML_ASSERT(!ggml_is_permuted(src0));
-    GGML_ASSERT(ggml_backend_buffer_is_cuda(src0->buffer));
-    GGML_ASSERT(src0->type == GGML_TYPE_F16);
-    GGML_ASSERT(src1->type == GGML_TYPE_F32);
-
-    const int64_t ne00 = src0->ne[0];
-    const int64_t ne01 = src0->ne[1];
-    const int64_t ne02 = src0->ne[2];
-
-    const int64_t nb01 = src0->nb[1];
-    const int64_t nb02 = src0->nb[2];
-
-    const int64_t ne12 = src1->ne[2];
-
-    cudaStream_t main_stream = ctx.stream();
-
-    void  * src0_ddq = src0->data;
-    float * src1_ddf = (float *) src1->data;
-    float * dst_ddf  = (float *) dst->data;
-
-    const int64_t row_stride_x = nb01 / sizeof(half);
-    const int64_t channel_stride_x = nb02 / sizeof(half);
-
-    ggml_mul_mat_vec_nc_f16_f32_cuda(src0_ddq, src1_ddf, dst_ddf, ne00, ne01, row_stride_x, ne02, ne12, channel_stride_x, main_stream);
-}
-
 static __global__ void k_compute_batched_ptrs(
         const half * src0_as_f16, const half * src1_as_f16, char * dst,
         const void ** ptrs_src, void ** ptrs_dst,
@@ -1879,21 +1719,17 @@ static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, co
 static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     const bool split = ggml_backend_buft_is_cuda_split(src0->buffer->buft);
 
-    bool use_dequantize_mul_mat_vec = ggml_cuda_dmmv_type_supported(src0->type)
+    bool use_mul_mat_vec   = src0->type == GGML_TYPE_F16
         && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32
-        && src0->ne[0] % (GGML_CUDA_DMMV_X*2) == 0 && src1->ne[1] == 1;
-    bool          use_mul_mat_vec_q =  ggml_is_quantized(src0->type)
+        && src0->ne[0] % 2 == 0 && src1->ne[1] == 1;
+    bool use_mul_mat_vec_q = ggml_is_quantized(src0->type)
         && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32
         && src1->ne[1] <= MMVQ_MAX_BATCH_SIZE;
-    bool              use_mul_mat_q =  ggml_is_quantized(src0->type)
+    bool use_mul_mat_q     = ggml_is_quantized(src0->type)
         && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32;
 
-    // if mmvq is available it's a better choice than dmmv:
-#ifndef GGML_CUDA_FORCE_DMMV
-    use_dequantize_mul_mat_vec = use_dequantize_mul_mat_vec && !use_mul_mat_vec_q;
-#endif // GGML_CUDA_FORCE_DMMV
-
-    bool any_gpus_with_slow_fp16 = false;
+    bool any_gpus_with_slow_fp16   = false;
+    bool any_gpus_without_fp16_mma = false;
 
     if (split) {
         ggml_backend_cuda_split_buffer_type_context * buft_ctx = (ggml_backend_cuda_split_buffer_type_context *) src0->buffer->buft->context;
@@ -1904,14 +1740,16 @@ static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor
                 continue;
             }
 
-            const int cc            = ggml_cuda_info().devices[id].cc;
-            use_mul_mat_q           = use_mul_mat_q           && ggml_cuda_should_use_mmq(src0->type, cc, src1->ne[1]);
-            any_gpus_with_slow_fp16 = any_gpus_with_slow_fp16 || !fast_fp16_available(cc);
+            const int cc              = ggml_cuda_info().devices[id].cc;
+            use_mul_mat_q             = use_mul_mat_q             && ggml_cuda_should_use_mmq(src0->type, cc, src1->ne[1]);
+            any_gpus_with_slow_fp16   = any_gpus_with_slow_fp16   || !fast_fp16_available(cc);
+            any_gpus_without_fp16_mma = any_gpus_without_fp16_mma || !fp16_mma_available(cc);
         }
     } else {
-        const int cc            = ggml_cuda_info().devices[ctx.device].cc;
-        use_mul_mat_q           = use_mul_mat_q           && ggml_cuda_should_use_mmq(src0->type, cc, src1->ne[1]);
-        any_gpus_with_slow_fp16 = any_gpus_with_slow_fp16 || !fast_fp16_available(cc);
+        const int cc              = ggml_cuda_info().devices[ctx.device].cc;
+        use_mul_mat_q             = use_mul_mat_q             && ggml_cuda_should_use_mmq(src0->type, cc, src1->ne[1]);
+        any_gpus_with_slow_fp16   = any_gpus_with_slow_fp16   || !fast_fp16_available(cc);
+        any_gpus_without_fp16_mma = any_gpus_without_fp16_mma || !fp16_mma_available(cc);
     }
 
     // debug helpers
@@ -1922,18 +1760,16 @@ static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor
     //printf("src0 is contiguous %d, transposed %d, type = %s, name = %s\n", ggml_is_contiguous(src0), ggml_is_transposed(src0), ggml_type_name(src0->type), src0->name);
     //printf("src1 is contiguous %d, transposed %d, type = %s, name = %s\n", ggml_is_contiguous(src1), ggml_is_transposed(src1), ggml_type_name(src1->type), src1->name);
 
-    if (!split && any_gpus_with_slow_fp16 && src0->type == GGML_TYPE_F16 && ggml_is_permuted(src0) && ggml_is_permuted(src1) && src1->ne[1] == 1) {
-        // FP32 precision KQ single-batch for batch size 1 without FlashAttention
-        ggml_cuda_mul_mat_vec_p021(ctx, src0, src1, dst);
-    } else if (!split && any_gpus_with_slow_fp16 && src0->type == GGML_TYPE_F16 && !ggml_is_contiguous(src0) && !ggml_is_transposed(src1) && src1->ne[1] == 1) {
-        // FP32 precision KQV single-batch for batch size 1 without FlashAttention
-        ggml_cuda_mul_mat_vec_nc(ctx, src0, src1, dst);
+    if (!split && use_mul_mat_vec && dst->ne[3] == 1 && (src0->ne[1] < MMV_MAX_ROWS || any_gpus_without_fp16_mma)) {
+        // the custom F16 vector kernel can be used over batched cuBLAS GEMM
+        // but this is only faster for GPUs without tensor cores or with a thin src0 matrix (particularly KQV in attention)
+        ggml_cuda_mul_mat_vec(ctx, src0, src1, dst);
     } else if (!split && src0->type == GGML_TYPE_F16 && (src1->type == GGML_TYPE_F16 || !any_gpus_with_slow_fp16)
                && !ggml_is_transposed(src0) && !ggml_is_transposed(src1) && src1->ne[2]*src1->ne[3] > 1) {
-        // KQ + KQV multi-batch without FlashAttention
+        // general KQ + KQV multi-batch without FlashAttention
         ggml_cuda_mul_mat_batched_cublas(ctx, src0, src1, dst);
-    } else if (use_dequantize_mul_mat_vec) {
-        ggml_cuda_op_mul_mat(ctx, src0, src1, dst, ggml_cuda_op_dequantize_mul_mat_vec, nullptr);
+    } else if (use_mul_mat_vec) {
+        ggml_cuda_op_mul_mat(ctx, src0, src1, dst, ggml_cuda_op_mul_mat_vec, nullptr);
     } else if (use_mul_mat_vec_q) {
         ggml_cuda_op_mul_mat(ctx, src0, src1, dst, ggml_cuda_op_mul_mat_vec_q, quantize_row_q8_1_cuda);
     } else if (use_mul_mat_q) {
@@ -2978,6 +2814,17 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
             {
                 struct ggml_tensor * a = op->src[0];
                 struct ggml_tensor * b = op->src[1];
+                // for small weight matrices the active device can end up without any rows, don't use row split in those cases
+                // this avoids some edge cases (and the performance would not be good anyways)
+                if (a->buffer && ggml_backend_buft_is_cuda_split(a->buffer->buft)) {
+                    ggml_backend_cuda_split_buffer_type_context * buft_ctx = (ggml_backend_cuda_split_buffer_type_context *) a->buffer->buft->context;
+                    int64_t row_low;
+                    int64_t row_high;
+                    get_row_split(&row_low, &row_high, a, buft_ctx->tensor_split, dev_ctx->device);
+                    if (row_low == row_high) {
+                        return false;
+                    }
+                }
                 if (b->type == GGML_TYPE_F16 && a->type != GGML_TYPE_F16) {
                     return false;
                 }

ggml/src/ggml-cuda/ggml/CMakeLists.txt

@@ -0,0 +1,155 @@
+cmake_minimum_required(VERSION 3.18)  # for CMAKE_CUDA_ARCHITECTURES
+
+find_package(CUDAToolkit)
+
+if (CUDAToolkit_FOUND)
+    message(STATUS "CUDA Toolkit found")
+
+    if (NOT DEFINED CMAKE_CUDA_ARCHITECTURES)
+        # native == GPUs available at build time
+        # 52     == Maxwell, lowest CUDA 12 standard
+        # 60     == P100, FP16 CUDA intrinsics
+        # 61     == Pascal, __dp4a instruction (per-byte integer dot product)
+        # 70     == V100, FP16 tensor cores
+        # 75     == Turing, int8 tensor cores
+        if (GGML_NATIVE AND CUDAToolkit_VERSION VERSION_GREATER_EQUAL "11.6" AND CMAKE_VERSION VERSION_GREATER_EQUAL "3.24")
+            set(CMAKE_CUDA_ARCHITECTURES "native")
+        elseif(GGML_CUDA_F16 OR GGML_CUDA_DMMV_F16)
+            set(CMAKE_CUDA_ARCHITECTURES "60;61;70;75")
+        else()
+            set(CMAKE_CUDA_ARCHITECTURES "52;61;70;75")
+        endif()
+    endif()
+    message(STATUS "Using CUDA architectures: ${CMAKE_CUDA_ARCHITECTURES}")
+
+    enable_language(CUDA)
+
+    file(GLOB   GGML_HEADERS_CUDA "*.cuh")
+    list(APPEND GGML_HEADERS_CUDA "../../include/ggml-cuda.h")
+
+    file(GLOB   GGML_SOURCES_CUDA "*.cu")
+    file(GLOB   SRCS "template-instances/fattn-wmma*.cu")
+    list(APPEND GGML_SOURCES_CUDA ${SRCS})
+    file(GLOB   SRCS "template-instances/mmq*.cu")
+    list(APPEND GGML_SOURCES_CUDA ${SRCS})
+
+    if (GGML_CUDA_FA_ALL_QUANTS)
+        file(GLOB   SRCS "template-instances/fattn-vec*.cu")
+        list(APPEND GGML_SOURCES_CUDA ${SRCS})
+        add_compile_definitions(GGML_CUDA_FA_ALL_QUANTS)
+    else()
+        file(GLOB   SRCS "template-instances/fattn-vec*q4_0-q4_0.cu")
+        list(APPEND GGML_SOURCES_CUDA ${SRCS})
+        file(GLOB   SRCS "template-instances/fattn-vec*q8_0-q8_0.cu")
+        list(APPEND GGML_SOURCES_CUDA ${SRCS})
+        file(GLOB   SRCS "template-instances/fattn-vec*f16-f16.cu")
+        list(APPEND GGML_SOURCES_CUDA ${SRCS})
+    endif()
+
+    add_library(ggml-cuda
+                ${GGML_HEADERS_CUDA}
+                ${GGML_SOURCES_CUDA}
+                )
+
+    target_link_libraries(ggml-cuda PRIVATE ggml-base)
+    target_include_directories(ggml-cuda PRIVATE . ..)
+
+    add_compile_definitions(GGML_CUDA_PEER_MAX_BATCH_SIZE=${GGML_CUDA_PEER_MAX_BATCH_SIZE})
+
+    if (GGML_CUDA_GRAPHS)
+        add_compile_definitions(GGML_CUDA_USE_GRAPHS)
+    endif()
+
+    if (GGML_CUDA_FORCE_MMQ)
+        add_compile_definitions(GGML_CUDA_FORCE_MMQ)
+    endif()
+
+    if (GGML_CUDA_FORCE_CUBLAS)
+        add_compile_definitions(GGML_CUDA_FORCE_CUBLAS)
+    endif()
+
+    if (GGML_CUDA_NO_VMM)
+        add_compile_definitions(GGML_CUDA_NO_VMM)
+    endif()
+
+    if (GGML_CUDA_F16 OR GGML_CUDA_DMMV_F16)
+        add_compile_definitions(GGML_CUDA_F16)
+    endif()
+
+    if (GGML_CUDA_NO_PEER_COPY)
+        add_compile_definitions(GGML_CUDA_NO_PEER_COPY)
+    endif()
+
+    if (GGML_STATIC)
+        if (WIN32)
+            # As of 12.3.1 CUDA Toolkit for Windows does not offer a static cublas library
+            target_link_libraries(ggml-cuda PRIVATE CUDA::cudart_static CUDA::cublas CUDA::cublasLt)
+        else ()
+            target_link_libraries(ggml-cuda PRIVATE  CUDA::cudart_static CUDA::cublas_static CUDA::cublasLt_static)
+        endif()
+    else()
+        target_link_libraries(ggml-cuda PRIVATE CUDA::cudart CUDA::cublas CUDA::cublasLt)
+    endif()
+
+    if (GGML_CUDA_NO_VMM)
+        # No VMM requested, no need to link directly with the cuda driver lib (libcuda.so)
+    else()
+        target_link_libraries(ggml-cuda PRIVATE CUDA::cuda_driver)
+    endif()
+
+    set(CUDA_CXX_FLAGS "")
+
+    set(CUDA_FLAGS -use_fast_math)
+
+    if (GGML_FATAL_WARNINGS)
+        list(APPEND CUDA_FLAGS -Werror all-warnings)
+    endif()
+
+    if (GGML_ALL_WARNINGS AND NOT MSVC)
+        set(NVCC_CMD ${CMAKE_CUDA_COMPILER} .c)
+        if (NOT CMAKE_CUDA_HOST_COMPILER STREQUAL "")
+            list(APPEND NVCC_CMD -ccbin ${CMAKE_CUDA_HOST_COMPILER})
+        endif()
+
+        execute_process(
+            COMMAND ${NVCC_CMD} -Xcompiler --version
+            OUTPUT_VARIABLE CUDA_CCFULLVER
+            ERROR_QUIET
+        )
+
+        if (NOT CUDA_CCFULLVER MATCHES clang)
+            set(CUDA_CCID "GNU")
+            execute_process(
+                COMMAND ${NVCC_CMD} -Xcompiler "-dumpfullversion -dumpversion"
+                OUTPUT_VARIABLE CUDA_CCVER
+                ERROR_QUIET
+            )
+        else()
+            if (CUDA_CCFULLVER MATCHES Apple)
+                set(CUDA_CCID "AppleClang")
+            else()
+                set(CUDA_CCID "Clang")
+            endif()
+            string(REGEX REPLACE "^.* version ([0-9.]*).*$" "\\1" CUDA_CCVER ${CUDA_CCFULLVER})
+        endif()
+
+        message("-- CUDA host compiler is ${CUDA_CCID} ${CUDA_CCVER}")
+
+        get_flags(${CUDA_CCID} ${CUDA_CCVER})
+        list(APPEND CUDA_CXX_FLAGS ${CXX_FLAGS} ${GF_CXX_FLAGS})  # This is passed to -Xcompiler later
+    endif()
+
+    if (NOT MSVC)
+        list(APPEND CUDA_CXX_FLAGS -Wno-pedantic)
+    endif()
+
+    list(JOIN   CUDA_CXX_FLAGS " " CUDA_CXX_FLAGS_JOINED)  # pass host compiler flags as a single argument
+
+    if (NOT CUDA_CXX_FLAGS_JOINED STREQUAL "")
+        list(APPEND CUDA_FLAGS -Xcompiler ${CUDA_CXX_FLAGS_JOINED})
+    endif()
+
+    target_compile_options(ggml-cuda PRIVATE "$<$<COMPILE_LANGUAGE:CUDA>:${CUDA_FLAGS}>")
+else()
+    message(FATAL_ERROR "CUDA Toolkit not found")
+endif()

ggml/src/ggml-cuda/mmq.cuh

@@ -100,9 +100,9 @@ static constexpr __device__ int get_mmq_x_max_device() {
     return 128;
 #else // INT8_MMA_AVAILABLE
 
-#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+#if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
     return 128;
-#else // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+#else // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
 
 #if __CUDA_ARCH__ >= CC_VOLTA
 #ifdef GGML_CUDA_FORCE_MMQ
@@ -115,7 +115,7 @@ static constexpr __device__ int get_mmq_x_max_device() {
     return 64;
 #endif // __CUDA_ARCH__ >= CC_VOLTA
 
-#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+#endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
 #endif // INT8_MMA_AVAILABLE
 }
 
@@ -124,7 +124,7 @@ static constexpr int get_mmq_y_host(const int cc) {
 }
 
 static constexpr __device__ int get_mmq_y_device() {
-#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+#if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
 #if defined(RDNA1)
     return 64;
 #else
@@ -136,7 +136,7 @@ static constexpr __device__ int get_mmq_y_device() {
 #else
     return 64;
 #endif // __CUDA_ARCH__ >= CC_VOLTA
-#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+#endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
 }
 
 #define MMQ_DP4A_TXS_Q4_0    tile_x_sizes{mmq_y*WARP_SIZE   + mmq_y, mmq_y*WARP_SIZE/QI4_0   + mmq_y/QI4_0,     0}
@@ -2569,7 +2569,7 @@ static __device__ void mul_mat_q_process_tile(
 // The mul_mat_q kernel implements "stream-k" work partitioning as described in https://arxiv.org/abs/2301.03598
 
 template <ggml_type type, int mmq_x, int nwarps, bool need_check>
-#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+#if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
 #if defined(RDNA3) || defined(RDNA2)
     __launch_bounds__(WARP_SIZE*nwarps, 2)
 #endif // defined(RDNA3) || defined(RDNA2)
@@ -2579,7 +2579,7 @@ template <ggml_type type, int mmq_x, int nwarps, bool need_check>
 #else
     __launch_bounds__(WARP_SIZE*nwarps, 2)
 #endif // __CUDA_ARCH__ >= CC_VOLTA
-#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+#endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
 static __global__ void mul_mat_q(
     const char * __restrict__ x, const char * __restrict__ yc, float * __restrict__ dst, float * __restrict__ tmp_fixup,
     const int ne00, const int ne01, const int stride01, const int ne10, const int ne11, const int stride11, const int ne0) {
@@ -2594,7 +2594,7 @@ static __global__ void mul_mat_q(
     constexpr int mmq_y = get_mmq_y_device();
 
     // On AMD or old CUDA the performance with stream-k was worse, use conventional tiling instead:
-#if (defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) || __CUDA_ARCH__ < CC_VOLTA
+#if (defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) || __CUDA_ARCH__ < CC_VOLTA
     {
         constexpr bool fixup = false;
         mul_mat_q_process_tile<type, mmq_x, nwarps, need_check, fixup>
@@ -2602,7 +2602,7 @@ static __global__ void mul_mat_q(
                 blockIdx.x, blockIdx.y, 0, ne00/qk);
         return;
     }
-#endif // (defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) || __CUDA_ARCH__ < CC_VOLTA
+#endif // (defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) || __CUDA_ARCH__ < CC_VOLTA
 
     const     int64_t blocks_per_ne00 = ne00 / qk;
     constexpr int     blocks_per_iter = MMQ_ITER_K / qk;
@@ -2765,14 +2765,14 @@ static void launch_mul_mat_q(ggml_backend_cuda_context & ctx, const mmq_args & a
 
     const int shmem = mmq_get_shmem<type>(mmq_x, mmq_y, cc);
 
-#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
     static bool shmem_limit_raised[GGML_CUDA_MAX_DEVICES] = {false};
     if (!shmem_limit_raised[id]) {
         CUDA_CHECK(cudaFuncSetAttribute(mul_mat_q<type, mmq_x, MMQ_NWARPS, false>, cudaFuncAttributeMaxDynamicSharedMemorySize, shmem));
         CUDA_CHECK(cudaFuncSetAttribute(mul_mat_q<type, mmq_x, MMQ_NWARPS, true>,  cudaFuncAttributeMaxDynamicSharedMemorySize, shmem));
         shmem_limit_raised[id] = true;
     }
-#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
 
     const int nty = (args.ne01 + mmq_y - 1) / mmq_y;
     const int ntx = (args.ne11 + mmq_x - 1) / mmq_x;

ggml/src/ggml-cuda/mmv.cu

@@ -0,0 +1,223 @@
+#include "common.cuh"
+#include "mmv.cuh"
+
+template <typename type_acc, int block_size>
+static __global__ void mul_mat_vec(
+        const half * __restrict__ x, const float * __restrict__ y, float * __restrict__ dst, const int64_t ncols2, const int64_t stride_row,
+        const int64_t channel_ratio, const int64_t stride_channel_x, const int64_t stride_channel_y, const int64_t stride_channel_dst) {
+    const int64_t row     = blockIdx.x;
+    const int64_t channel = blockIdx.z;
+    const int     tid     = threadIdx.x;
+
+    x   += (channel/channel_ratio)*stride_channel_x + row*stride_row;
+    y   +=  channel               *stride_channel_y;
+    dst +=  channel               *stride_channel_dst;
+
+    const half2  * x2 = (const half2  *) x;
+    const float2 * y2 = (const float2 *) y;
+
+    extern __shared__ char data_mmv[];
+    float * buf_iw = (float *) data_mmv;
+
+    if (block_size > WARP_SIZE) {
+        if (tid < WARP_SIZE) {
+            buf_iw[tid] = 0.0f;
+        }
+        __syncthreads();
+    }
+
+    float sumf;
+
+    if (std::is_same<type_acc, float>::value) {
+        sumf = 0.0f;
+
+        for (int64_t col2 = tid; col2 < ncols2; col2 += block_size) {
+            const float2 tmpx = __half22float2(x2[col2]);
+            const float2 tmpy = y2[col2];
+            sumf += tmpx.x * tmpy.x;
+            sumf += tmpx.y * tmpy.y;
+        }
+    } else {
+#ifdef FP16_AVAILABLE
+        half2 sumh2 = make_half2(0.0f, 0.0f);
+
+        for (int64_t col2 = tid; col2 < ncols2; col2 += block_size) {
+            const float2 tmp = y2[col2];
+            sumh2 += x2[col2] * make_half2(tmp.x, tmp.y);
+        }
+
+        sumf = __low2float(sumh2) + __high2float(sumh2);
+#else
+        NO_DEVICE_CODE;
+#endif // FP16_AVAILABLE
+    }
+
+    sumf = warp_reduce_sum(sumf);
+
+    if (block_size > WARP_SIZE) {
+        buf_iw[tid/WARP_SIZE] = sumf;
+        __syncthreads();
+        if (tid > WARP_SIZE) {
+            return;
+        }
+        sumf = buf_iw[tid];
+        sumf = warp_reduce_sum(sumf);
+    }
+
+    if (tid != 0) {
+        return;
+    }
+
+    dst[row] = sumf;
+}
+
+template <typename type_acc>
+static void launch_mul_mat_vec_cuda(
+        const half * x, const float * y, float * dst,
+        const int64_t ncols, const int64_t nrows, const int64_t stride_row, const int64_t nchannels_x, const int64_t nchannels_y,
+        const int64_t stride_channel_x, const int64_t stride_channel_y, const int64_t stride_channel_dst,
+        cudaStream_t stream) {
+    GGML_ASSERT(ncols      % 2 == 0);
+    GGML_ASSERT(stride_row % 2 == 0);
+    GGML_ASSERT(nchannels_y % nchannels_x == 0);
+    const int64_t channel_ratio = nchannels_y / nchannels_x;
+
+    int64_t block_size_best = WARP_SIZE;
+    int64_t niter_best      = (ncols + 2*WARP_SIZE - 1) / (2*WARP_SIZE);
+    for (int64_t block_size = 2*WARP_SIZE; block_size <= 256; block_size += WARP_SIZE) {
+        const int64_t niter = (ncols + 2*block_size - 1) / (2*block_size);
+        if (niter < niter_best) {
+            niter_best      = niter;
+            block_size_best = block_size;
+        }
+    }
+
+    const int smem = WARP_SIZE*sizeof(float);
+    const dim3 block_nums(nrows, 1, nchannels_y);
+    const dim3 block_dims(block_size_best, 1, 1);
+    switch (block_size_best) {
+        case   32: {
+            mul_mat_vec<type_acc,  32><<<block_nums, block_dims, smem, stream>>>
+                (x, y, dst, ncols/2, stride_row, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst);
+        } break;
+        case   64: {
+            mul_mat_vec<type_acc,  64><<<block_nums, block_dims, smem, stream>>>
+                (x, y, dst, ncols/2, stride_row, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst);
+        } break;
+        case   96: {
+            mul_mat_vec<type_acc,  96><<<block_nums, block_dims, smem, stream>>>
+                (x, y, dst, ncols/2, stride_row, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst);
+        } break;
+        case  128: {
+            mul_mat_vec<type_acc, 128><<<block_nums, block_dims, smem, stream>>>
+                (x, y, dst, ncols/2, stride_row, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst);
+        } break;
+        case  160: {
+            mul_mat_vec<type_acc, 160><<<block_nums, block_dims, smem, stream>>>
+                (x, y, dst, ncols/2, stride_row, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst);
+        } break;
+        case  192: {
+            mul_mat_vec<type_acc, 192><<<block_nums, block_dims, smem, stream>>>
+                (x, y, dst, ncols/2, stride_row, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst);
+        } break;
+        case  224: {
+            mul_mat_vec<type_acc, 224><<<block_nums, block_dims, smem, stream>>>
+                (x, y, dst, ncols/2, stride_row, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst);
+        } break;
+        case  256: {
+            mul_mat_vec<type_acc, 256><<<block_nums, block_dims, smem, stream>>>
+                (x, y, dst, ncols/2, stride_row, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst);
+        } break;
+        default: {
+            GGML_ABORT("fatal error");
+        } break;
+    }
+}
+
+static void mul_mat_vec_cuda(
+        const half * x, const float * y, float * dst,
+        const int64_t ncols, const int64_t nrows, const int64_t stride_row, const int64_t nchannels_x, const int64_t nchannels_y,
+        const int64_t stride_channel_x, const int64_t stride_channel_y, const int64_t stride_channel_dst,
+        enum ggml_prec prec, cudaStream_t stream) {
+    switch (prec) {
+        case GGML_PREC_DEFAULT: {
+            launch_mul_mat_vec_cuda<half>(x, y, dst, ncols, nrows, stride_row, nchannels_x, nchannels_y,
+                stride_channel_x, stride_channel_y, stride_channel_dst, stream);
+        } break;
+        case GGML_PREC_F32: {
+            launch_mul_mat_vec_cuda<float>(x, y, dst, ncols, nrows, stride_row, nchannels_x, nchannels_y,
+                stride_channel_x, stride_channel_y, stride_channel_dst, stream);
+        } break;
+    }
+}
+
+void ggml_cuda_mul_mat_vec(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_ASSERT(src0->type == GGML_TYPE_F16);
+    GGML_ASSERT(src1->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type  == GGML_TYPE_F32);
+
+    const int64_t ne00 = src0->ne[0];
+    const int64_t ne01 = src0->ne[1];
+
+    GGML_ASSERT(src1->ne[1] == 1);
+
+    const int cc = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
+    const enum ggml_prec prec = fast_fp16_available(cc) ? ggml_prec(dst->op_params[0]) : GGML_PREC_F32;
+
+    const half  * src0_d = (const half  *) src0->data;
+    const float * src1_d = (const float *) src1->data;
+    float       *  dst_d = (float       *)  dst->data;
+
+    const int64_t ne02 = src0->ne[2];
+    const int64_t ne12 = src1->ne[2];
+    GGML_ASSERT(dst->ne[2] == ne12);
+
+    GGML_ASSERT(src0->ne[3] == 1);
+    GGML_ASSERT(src1->ne[3] == 1);
+    GGML_ASSERT( dst->ne[3] == 1);
+
+    const int64_t stride_row         = src0->nb[1] / ggml_type_size(src0->type);
+    const int64_t channel_stride_x   = src0->nb[2] / ggml_type_size(src0->type);
+    const int64_t channel_stride_y   = src1->nb[2] / ggml_type_size(src1->type);
+    const int64_t channel_stride_dst =  dst->nb[2] / ggml_type_size( dst->type);
+
+    mul_mat_vec_cuda(src0_d, src1_d, dst_d, ne00, ne01, stride_row, ne02, ne12, channel_stride_x, channel_stride_y, channel_stride_dst, prec, ctx.stream());
+}
+
+void ggml_cuda_op_mul_mat_vec(
+    ggml_backend_cuda_context & ctx,
+    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const char * src0_dd_i, const float * src1_ddf_i,
+    const char * src1_ddq_i, float * dst_dd_i, const int64_t row_low, const int64_t row_high, const int64_t src1_ncols,
+    const int64_t src1_padded_row_size, cudaStream_t stream) {
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F16);
+    GGML_ASSERT(src1->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type  == GGML_TYPE_F32);
+
+    const int64_t ne00 = src0->ne[0];
+    const int64_t row_diff = row_high - row_low;
+
+    GGML_ASSERT(src1_ncols == 1);
+
+    const int cc = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
+    const enum ggml_prec prec = fast_fp16_available(cc) ? ggml_prec(dst->op_params[0]) : GGML_PREC_F32;
+
+
+    // ggml_cuda_op provides single, contiguous matrices
+    const int64_t stride_row         = ne00;
+    const int64_t nchannels_x        = 1;
+    const int64_t nchannels_y        = 1;
+    const int64_t channel_stride_x   = 0;
+    const int64_t channel_stride_y   = 0;
+    const int64_t channel_stride_dst = 0;
+
+    mul_mat_vec_cuda((const half *) src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stride_row,
+        nchannels_x, nchannels_y, channel_stride_x, channel_stride_y, channel_stride_dst, prec, stream);
+
+    GGML_UNUSED(ctx);
+    GGML_UNUSED(src1);
+    GGML_UNUSED(dst);
+    GGML_UNUSED(src1_ddq_i);
+    GGML_UNUSED(src1_ncols);
+    GGML_UNUSED(src1_padded_row_size);
+}

ggml/src/ggml-cuda/mmv.cuh

@@ -0,0 +1,12 @@
+#include "common.cuh"
+
+// maximum number of src0 rows with which to use mul_mat_vec over cuBLAS if FP16 tensor cores are available
+#define MMV_MAX_ROWS 512
+
+void ggml_cuda_mul_mat_vec(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst);
+
+void ggml_cuda_op_mul_mat_vec(
+    ggml_backend_cuda_context & ctx,
+    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const char * src0_dd_i, const float * src1_ddf_i,
+    const char * src1_ddq_i, float * dst_dd_i, const int64_t row_low, const int64_t row_high, const int64_t src1_ncols,
+    const int64_t src1_padded_row_size, cudaStream_t stream);

ggml/src/ggml-cuda/mmvq.cu

@@ -48,10 +48,10 @@ static constexpr __device__ int get_vdr_mmvq(ggml_type type) {
 }
 
 template <ggml_type type, int ncols_y>
-#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
 // tell the compiler to use as many registers as it wants, see nwarps definition below
 __launch_bounds__((ncols_y <= 4 ? 4 : 2)*WARP_SIZE, 1)
-#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
 static __global__ void mul_mat_vec_q(
     const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst,
     const int ncols_x, const int nrows_x, const int nrows_y, const int nrows_dst) {
@@ -62,13 +62,13 @@ static __global__ void mul_mat_vec_q(
 
     constexpr vec_dot_q_cuda_t vec_dot_q_cuda = get_vec_dot_q_cuda(type);
 
-#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) && (defined(RDNA2) || defined(RDNA3))
+#if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__) && (defined(RDNA2) || defined(RDNA3))
     constexpr int nwarps              = 1;
     constexpr int rows_per_cuda_block = 1;
 #else
     constexpr int nwarps              = ncols_y <= 4 ? 4 : 2;
     constexpr int rows_per_cuda_block = ncols_y == 1 ? 1 : 2;
-#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) && !defined(RDNA2) && !defined(RDNA3)
+#endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__) && !defined(RDNA2) && !defined(RDNA3)
 
     const     int tid = WARP_SIZE*threadIdx.y + threadIdx.x;
     const     int row0 = rows_per_cuda_block*blockIdx.x;

ggml/src/ggml-cuda/opt-step-adamw.cu

@@ -1,11 +1,11 @@
+#include "ggml-impl.h"
 #include "opt-step-adamw.cuh"
 
 #include <cstdint>
 
 static __global__ void opt_step_adamw_f32(
-    float * __restrict__ x, const float * __restrict__ g, float * __restrict__ g_m, float * __restrict__ g_v, const int64_t k,
-    const float alpha, const float beta1, const float beta2, const float eps, const float wd,
-    const float beta1h, const float beta2h) {
+    float * __restrict__ x, const float * __restrict__ g, float * __restrict__ g_m, float * __restrict__ g_v,
+    const float * __restrict__ pars, const int64_t k) {
 
     const int64_t i = (int64_t) blockIdx.x*blockDim.x + threadIdx.x;
 
@@ -13,6 +13,14 @@ static __global__ void opt_step_adamw_f32(
         return;
     }
 
+    const float alpha  = pars[0];
+    const float beta1  = pars[1];
+    const float beta2  = pars[2];
+    const float eps    = pars[3];
+    const float wd     = pars[4];
+    const float beta1h = pars[5];
+    const float beta2h = pars[6];
+
     const float gi = g[i];
     const float gmi = g_m[i]*beta1 +    gi*(1.0f - beta1);
     const float gvi = g_v[i]*beta2 + gi*gi*(1.0f - beta2);
@@ -23,58 +31,48 @@ static __global__ void opt_step_adamw_f32(
     const float mh =       gmi*beta1h;
     const float vh = sqrtf(gvi*beta2h) + eps;
 
-    x[i] = x[i]*(1.0f - alpha*wd) - mh/vh;
+    x[i] = x[i]*(1.0f - alpha*wd) - alpha*mh/vh;
 }
 
 static void opt_step_adamw_f32_cuda(
-    float * x, const float * g, float * g_m, float * g_v, const int64_t k,
-    const float alpha, const float beta1, const float beta2, const float eps, const float wd,
-    const float beta1h, const float beta2h, cudaStream_t stream) {
+    float * x, const float * g, float * g_m, float * g_v, const float * pars, const int64_t k, cudaStream_t stream) {
 
     const dim3 block_dims(CUDA_OPT_STEP_ADAMW_BLOCK_SIZE, 1, 1);
     const dim3 block_nums((k + CUDA_OPT_STEP_ADAMW_BLOCK_SIZE - 1) / CUDA_OPT_STEP_ADAMW_BLOCK_SIZE, 1, 1);
-    opt_step_adamw_f32<<<block_nums, block_dims, 0, stream>>>(x, g, g_m, g_v, k, alpha, beta1, beta2, eps, wd, beta1h, beta2h);
+    opt_step_adamw_f32<<<block_nums, block_dims, 0, stream>>>(x, g, g_m, g_v, pars, k);
 }
 
 void ggml_cuda_opt_step_adamw(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
-    const ggml_tensor * src0        = dst->src[0];
-    const ggml_tensor * src0_grad   = dst->src[1];
-    const ggml_tensor * src0_grad_m = dst->src[2];
-    const ggml_tensor * src0_grad_v = dst->src[3];
+    const ggml_tensor * src0         = dst->src[0];
+    const ggml_tensor * src0_grad    = dst->src[1];
+    const ggml_tensor * src0_grad_m  = dst->src[2];
+    const ggml_tensor * src0_grad_v  = dst->src[3];
+    const ggml_tensor * adamw_params = dst->src[4];
 
-    GGML_ASSERT(src0->type        == GGML_TYPE_F32);
-    GGML_ASSERT(src0_grad->type   == GGML_TYPE_F32);
-    GGML_ASSERT(src0_grad_m->type == GGML_TYPE_F32);
-    GGML_ASSERT(src0_grad_v->type == GGML_TYPE_F32);
+    GGML_ASSERT(src0->type         == GGML_TYPE_F32);
+    GGML_ASSERT(src0_grad->type    == GGML_TYPE_F32);
+    GGML_ASSERT(src0_grad_m->type  == GGML_TYPE_F32);
+    GGML_ASSERT(src0_grad_v->type  == GGML_TYPE_F32);
+    GGML_ASSERT(adamw_params->type == GGML_TYPE_F32);
     GGML_ASSERT(ggml_is_contiguous(src0));
     GGML_ASSERT(ggml_is_contiguous(src0_grad));
     GGML_ASSERT(ggml_is_contiguous(src0_grad_m));
     GGML_ASSERT(ggml_is_contiguous(src0_grad_v));
+    GGML_ASSERT(ggml_is_contiguous(adamw_params));
     GGML_ASSERT(ggml_are_same_shape(src0, src0_grad));
     GGML_ASSERT(ggml_are_same_shape(src0, src0_grad_m));
     GGML_ASSERT(ggml_are_same_shape(src0, src0_grad_v));
+    GGML_ASSERT(ggml_nelements(adamw_params) == 7);
 
-    float       * src0_d        = (float       *) src0->data;
-    const float * src0_grad_d   = (const float *) src0_grad->data;
-    float       * src0_grad_m_d = (float       *) src0_grad_m->data;
-    float       * src0_grad_v_d = (float       *) src0_grad_v->data;
+    float       * src0_d         = (float       *) src0->data;
+    const float * src0_grad_d    = (const float *) src0_grad->data;
+    float       * src0_grad_m_d  = (float       *) src0_grad_m->data;
+    float       * src0_grad_v_d  = (float       *) src0_grad_v->data;
+    const float * adamw_params_d = (const float *) adamw_params->data;
 
     cudaStream_t stream = ctx.stream();
 
     const int64_t ne = ggml_nelements(src0);
 
-    int64_t iter;  memcpy(&iter,  &dst->op_params[0], sizeof(int64_t));
-    float   alpha; memcpy(&alpha, &dst->op_params[2], sizeof(float));
-    float   beta1; memcpy(&beta1, &dst->op_params[3], sizeof(float));
-    float   beta2; memcpy(&beta2, &dst->op_params[4], sizeof(float));
-    float   eps;   memcpy(&eps,   &dst->op_params[5], sizeof(float));
-    float   wd;    memcpy(&wd,    &dst->op_params[6], sizeof(float));
-
-    const float beta1h  = alpha/(1.0f - powf(beta1, iter));
-    const float beta2h  =  1.0f/(1.0f - powf(beta2, iter));
-
-    opt_step_adamw_f32_cuda(src0_d, src0_grad_d, src0_grad_m_d, src0_grad_v_d, ne, alpha, beta1, beta2, eps, wd, beta1h, beta2h, stream);
-
-    iter++;
-    memcpy(&dst->op_params[0], &iter, sizeof(int64_t));
+    opt_step_adamw_f32_cuda(src0_d, src0_grad_d, src0_grad_m_d, src0_grad_v_d, adamw_params_d, ne, stream);
 }

ggml/src/ggml-cuda/sum.cu

@@ -1,6 +1,6 @@
-#if !defined(GGML_USE_HIPBLAS) && !defined(GGML_USE_MUSA) && CUDART_VERSION >= 11700
+#if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA) && CUDART_VERSION >= 11700
 #define USE_CUB
-#endif // !defined(GGML_USE_HIPBLAS) && !defined(GGML_USE_MUSA) && CUDART_VERSION >= 11700
+#endif // !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA) && CUDART_VERSION >= 11700
 
 #ifdef USE_CUB
 // On Windows CUB uses libraries with variables called CC_PASCAL which conflict with the define in common.cuh.

ggml/src/ggml-hip/CMakeLists.txt

@@ -0,0 +1,106 @@
+if (NOT EXISTS $ENV{ROCM_PATH})
+    if (NOT EXISTS /opt/rocm)
+        set(ROCM_PATH /usr)
+    else()
+        set(ROCM_PATH /opt/rocm)
+    endif()
+else()
+    set(ROCM_PATH $ENV{ROCM_PATH})
+endif()
+
+list(APPEND CMAKE_PREFIX_PATH  ${ROCM_PATH})
+list(APPEND CMAKE_PREFIX_PATH "${ROCM_PATH}/lib64/cmake")
+
+# CMake on Windows doesn't support the HIP language yet
+if (WIN32)
+    set(CXX_IS_HIPCC TRUE)
+else()
+    string(REGEX MATCH "hipcc(\.bat)?$" CXX_IS_HIPCC "${CMAKE_CXX_COMPILER}")
+endif()
+
+if (CXX_IS_HIPCC)
+    if (LINUX)
+        if (NOT ${CMAKE_CXX_COMPILER_ID} MATCHES "Clang")
+            message(WARNING "Only LLVM is supported for HIP, hint: CXX=/opt/rocm/llvm/bin/clang++")
+        endif()
+
+        message(WARNING "Setting hipcc as the C++ compiler is legacy behavior."
+                " Prefer setting the HIP compiler directly. See README for details.")
+    endif()
+else()
+    # Forward AMDGPU_TARGETS to CMAKE_HIP_ARCHITECTURES.
+    if (AMDGPU_TARGETS AND NOT CMAKE_HIP_ARCHITECTURES)
+        set(CMAKE_HIP_ARCHITECTURES ${AMDGPU_TARGETS})
+    endif()
+    cmake_minimum_required(VERSION 3.21)
+    enable_language(HIP)
+endif()
+
+find_package(hip     REQUIRED)
+find_package(hipblas REQUIRED)
+find_package(rocblas REQUIRED)
+
+message(STATUS "HIP and hipBLAS found")
+
+file(GLOB   GGML_HEADERS_ROCM "../ggml-cuda/*.cuh")
+list(APPEND GGML_HEADERS_ROCM "../../include/ggml-cuda.h")
+
+file(GLOB   GGML_SOURCES_ROCM "../ggml-cuda/*.cu")
+file(GLOB   SRCS "../ggml-cuda/template-instances/fattn-wmma*.cu")
+list(APPEND GGML_SOURCES_ROCM ${SRCS})
+file(GLOB   SRCS "../ggml-cuda/template-instances/mmq*.cu")
+list(APPEND GGML_SOURCES_ROCM ${SRCS})
+
+if (GGML_CUDA_FA_ALL_QUANTS)
+    file(GLOB   SRCS "../ggml-cuda/template-instances/fattn-vec*.cu")
+    list(APPEND GGML_SOURCES_ROCM ${SRCS})
+    add_compile_definitions(GGML_CUDA_FA_ALL_QUANTS)
+else()
+    file(GLOB   SRCS "../ggml-cuda/template-instances/fattn-vec*q4_0-q4_0.cu")
+    list(APPEND GGML_SOURCES_ROCM ${SRCS})
+    file(GLOB   SRCS "../ggml-cuda/template-instances/fattn-vec*q8_0-q8_0.cu")
+    list(APPEND GGML_SOURCES_ROCM ${SRCS})
+    file(GLOB   SRCS "../ggml-cuda/template-instances/fattn-vec*f16-f16.cu")
+    list(APPEND GGML_SOURCES_ROCM ${SRCS})
+endif()
+
+add_library(ggml-hip
+            ${GGML_HEADERS_ROCM}
+            ${GGML_SOURCES_ROCM})
+
+target_link_libraries(ggml-hip PRIVATE ggml-base)
+target_include_directories(ggml-hip PRIVATE . ..)
+
+# TODO: do not use CUDA definitions for HIP
+target_compile_definitions(ggml PUBLIC GGML_USE_CUDA)
+
+add_compile_definitions(GGML_USE_HIP)
+
+if (GGML_HIP_UMA)
+    add_compile_definitions(GGML_HIP_UMA)
+endif()
+
+if (GGML_CUDA_FORCE_MMQ)
+    add_compile_definitions(GGML_CUDA_FORCE_MMQ)
+endif()
+
+if (GGML_CUDA_FORCE_CUBLAS)
+    add_compile_definitions(GGML_CUDA_FORCE_CUBLAS)
+endif()
+
+if (GGML_CUDA_NO_PEER_COPY)
+    add_compile_definitions(GGML_CUDA_NO_PEER_COPY)
+endif()
+
+if (CXX_IS_HIPCC)
+    set_source_files_properties(${GGML_SOURCES_ROCM} PROPERTIES LANGUAGE CXX)
+    target_link_libraries(ggml-hip PRIVATE hip::device)
+else()
+    set_source_files_properties(${GGML_SOURCES_ROCM} PROPERTIES LANGUAGE HIP)
+endif()
+
+if (GGML_STATIC)
+    message(FATAL_ERROR "Static linking not supported for HIP/ROCm")
+endif()
+
+target_link_libraries(ggml-hip PRIVATE ggml-base hip::host roc::rocblas roc::hipblas)

ggml/src/ggml-impl.h

@@ -3,13 +3,29 @@
 // GGML internal header
 
 #include "ggml.h"
-
 #include <assert.h>
+#include <math.h>
 #include <stdlib.h> // load `stdlib.h` before other headers to work around MinGW bug: https://sourceforge.net/p/mingw-w64/bugs/192/
 #include <stdbool.h>
 #include <stdint.h>
 #include <string.h>
 
+#ifdef __ARM_FEATURE_SVE
+#include <arm_sve.h>
+#endif // __ARM_FEATURE_SVE
+
+#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>
+#endif
+
+#if defined(__F16C__)
+#include <immintrin.h>
+#endif
+
 #ifdef __cplusplus
 extern "C" {
 #endif
@@ -28,13 +44,13 @@ extern "C" {
 // 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
+    #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
 
 static inline int ggml_up32(int n) {
@@ -120,14 +136,12 @@ struct ggml_map_custom1_op_params {
     void             * userdata;
 };
 
-
 struct ggml_map_custom2_op_params {
     ggml_custom2_op_t   fun;
     int                 n_tasks;
     void              * userdata;
 };
 
-
 struct ggml_map_custom3_op_params {
     ggml_custom3_op_t fun;
     int n_tasks;
@@ -182,7 +196,7 @@ void ggml_hash_set_reset(struct ggml_hash_set * hash_set);
 static bool ggml_hash_contains(const struct ggml_hash_set * hash_set, struct ggml_tensor * key);
 
 // returns GGML_HASHSET_FULL if table is full, otherwise the current index of the key or where it should be inserted
-static size_t ggml_hash_find(const struct ggml_hash_set * hash_set, struct ggml_tensor * key);
+static size_t ggml_hash_find(const struct ggml_hash_set * hash_set, const struct ggml_tensor * key);
 
 // returns GGML_HASHSET_ALREADY_EXISTS if key already exists, index otherwise, asserts if table is full
 static size_t ggml_hash_insert(struct ggml_hash_set * hash_set, struct ggml_tensor * key);
@@ -196,7 +210,7 @@ static inline size_t ggml_hash(const struct ggml_tensor * p) {
     return (size_t)(uintptr_t)p >> 4;
 }
 
-static size_t ggml_hash_find(const struct ggml_hash_set * hash_set, struct ggml_tensor * key) {
+static size_t ggml_hash_find(const struct ggml_hash_set * hash_set, const struct ggml_tensor * key) {
     size_t h = ggml_hash(key) % hash_set->size;
 
     // linear probing
@@ -267,13 +281,14 @@ enum ggml_cgraph_eval_order {
 };
 
 struct ggml_cgraph {
-    int size;
-    int n_nodes;
-    int n_leafs;
+    int size;    // maximum number of nodes/leafs/grads/grad_accs
+    int n_nodes; // number of nodes currently in use
+    int n_leafs; // number of leafs currently in use
 
-    struct ggml_tensor ** nodes;
-    struct ggml_tensor ** grads;
-    struct ggml_tensor ** leafs;
+    struct ggml_tensor ** nodes;     // tensors with data that can change if the graph is evaluated
+    struct ggml_tensor ** grads;     // the outputs of these tensors are the gradients of the nodes
+    struct ggml_tensor ** grad_accs; // accumulators for node gradients
+    struct ggml_tensor ** leafs;     // tensors with constant data
 
     struct ggml_hash_set visited_hash_set;
 
@@ -287,9 +302,249 @@ struct ggml_cgraph ggml_graph_view(struct ggml_cgraph * cgraph, int i0, int i1);
 void * ggml_aligned_malloc(size_t size);
 void ggml_aligned_free(void * ptr, size_t size);
 
-// TODO: move to threading file
-void ggml_critical_section_start(void);
-void ggml_critical_section_end(void);
+// FP16 to FP32 conversion
+
+#if defined(__ARM_NEON)
+    #ifdef _MSC_VER
+        typedef uint16_t ggml_fp16_internal_t;
+    #else
+        typedef __fp16 ggml_fp16_internal_t;
+    #endif
+#endif
+
+#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;
+    }
+
+#elif defined(__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__)) && (!defined(__cplusplus) || __cplusplus >= 201703L)
+        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__)) && (!defined(__cplusplus) || __cplusplus >= 201703L)
+        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 // defined(__ARM_NEON) && (!defined(__MSC_VER)
+
+// precomputed f32 table for f16 (256 KB)
+// defined in ggml.c, initialized in ggml_init()
+GGML_API 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
+
+/**
+ * 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)
 
 #ifdef __cplusplus
 }

ggml/src/ggml-kompute/CMakeLists.txt

@@ -0,0 +1,162 @@
+
+find_package(Vulkan COMPONENTS glslc REQUIRED)
+find_program(glslc_executable NAMES glslc HINTS Vulkan::glslc)
+
+if (NOT glslc_executable)
+    message(FATAL_ERROR "glslc not found")
+endif()
+
+add_library(ggml-kompute
+            ggml-kompute.cpp
+            ../../include/ggml-kompute.h
+            )
+
+target_link_libraries(ggml-kompute PRIVATE ggml-base kompute)
+target_include_directories(ggml-kompute PRIVATE . .. ${CMAKE_CURRENT_BINARY_DIR})
+
+add_compile_definitions(VULKAN_HPP_DISPATCH_LOADER_DYNAMIC=1)
+
+function(compile_shader)
+    set(options)
+    set(oneValueArgs)
+    set(multiValueArgs SOURCES)
+    cmake_parse_arguments(compile_shader "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
+    foreach(source ${compile_shader_SOURCES})
+        get_filename_component(filename ${source} NAME)
+        set(spv_file ${filename}.spv)
+        add_custom_command(
+            OUTPUT ${spv_file}
+            DEPENDS ${CMAKE_CURRENT_SOURCE_DIR}/${source}
+            ${CMAKE_CURRENT_SOURCE_DIR}/kompute-shaders/common.comp
+            ${CMAKE_CURRENT_SOURCE_DIR}/kompute-shaders/op_getrows.comp
+            ${CMAKE_CURRENT_SOURCE_DIR}/kompute-shaders/op_mul_mv_q_n_pre.comp
+            ${CMAKE_CURRENT_SOURCE_DIR}/kompute-shaders/op_mul_mv_q_n.comp
+            COMMAND ${glslc_executable} --target-env=vulkan1.2 -o ${spv_file} ${CMAKE_CURRENT_SOURCE_DIR}/${source}
+            COMMENT "Compiling ${source} to ${spv_file}"
+            )
+
+        get_filename_component(RAW_FILE_NAME ${spv_file} NAME)
+        set(FILE_NAME "shader${RAW_FILE_NAME}")
+        string(REPLACE ".comp.spv" ".h" HEADER_FILE ${FILE_NAME})
+        string(TOUPPER ${HEADER_FILE} HEADER_FILE_DEFINE)
+        string(REPLACE "." "_" HEADER_FILE_DEFINE "${HEADER_FILE_DEFINE}")
+        set(OUTPUT_HEADER_FILE "${HEADER_FILE}")
+        message(STATUS "${HEADER_FILE} generating ${HEADER_FILE_DEFINE}")
+        if(CMAKE_GENERATOR MATCHES "Visual Studio")
+            add_custom_command(
+                OUTPUT ${OUTPUT_HEADER_FILE}
+                COMMAND ${CMAKE_COMMAND} -E echo "/*THIS FILE HAS BEEN AUTOMATICALLY GENERATED - DO NOT EDIT*/" > ${OUTPUT_HEADER_FILE}
+                COMMAND ${CMAKE_COMMAND} -E echo \"\#ifndef ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
+                COMMAND ${CMAKE_COMMAND} -E echo \"\#define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
+                COMMAND ${CMAKE_COMMAND} -E echo "namespace kp {" >> ${OUTPUT_HEADER_FILE}
+                COMMAND ${CMAKE_COMMAND} -E echo "namespace shader_data {" >> ${OUTPUT_HEADER_FILE}
+                COMMAND ${CMAKE_BINARY_DIR}/bin/$<CONFIG>/xxd -i ${RAW_FILE_NAME} >> ${OUTPUT_HEADER_FILE}
+                COMMAND ${CMAKE_COMMAND} -E echo "}}" >> ${OUTPUT_HEADER_FILE}
+                COMMAND ${CMAKE_COMMAND} -E echo \"\#endif // define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
+                DEPENDS ${spv_file} xxd
+                COMMENT "Converting to hpp: ${FILE_NAME} ${CMAKE_BINARY_DIR}/bin/$<CONFIG>/xxd"
+                )
+        else()
+            add_custom_command(
+                OUTPUT ${OUTPUT_HEADER_FILE}
+                COMMAND ${CMAKE_COMMAND} -E echo "/*THIS FILE HAS BEEN AUTOMATICALLY GENERATED - DO NOT EDIT*/" > ${OUTPUT_HEADER_FILE}
+                COMMAND ${CMAKE_COMMAND} -E echo \"\#ifndef ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
+                COMMAND ${CMAKE_COMMAND} -E echo \"\#define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
+                COMMAND ${CMAKE_COMMAND} -E echo "namespace kp {" >> ${OUTPUT_HEADER_FILE}
+                COMMAND ${CMAKE_COMMAND} -E echo "namespace shader_data {" >> ${OUTPUT_HEADER_FILE}
+                COMMAND ${CMAKE_BINARY_DIR}/bin/xxd -i ${RAW_FILE_NAME} >> ${OUTPUT_HEADER_FILE}
+                COMMAND ${CMAKE_COMMAND} -E echo "}}" >> ${OUTPUT_HEADER_FILE}
+                COMMAND ${CMAKE_COMMAND} -E echo \"\#endif // define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
+                DEPENDS ${spv_file} xxd
+                COMMENT "Converting to hpp: ${FILE_NAME} ${CMAKE_BINARY_DIR}/bin/xxd"
+                )
+        endif()
+    endforeach()
+endfunction()
+
+if (EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/kompute/CMakeLists.txt")
+    message(STATUS "Kompute found")
+    set(KOMPUTE_OPT_LOG_LEVEL Error CACHE STRING "Kompute log level")
+    add_subdirectory(kompute)
+
+    # Compile our shaders
+    compile_shader(SOURCES
+        kompute-shaders/op_scale.comp
+        kompute-shaders/op_scale_8.comp
+        kompute-shaders/op_add.comp
+        kompute-shaders/op_addrow.comp
+        kompute-shaders/op_mul.comp
+        kompute-shaders/op_silu.comp
+        kompute-shaders/op_relu.comp
+        kompute-shaders/op_gelu.comp
+        kompute-shaders/op_softmax.comp
+        kompute-shaders/op_norm.comp
+        kompute-shaders/op_rmsnorm.comp
+        kompute-shaders/op_diagmask.comp
+        kompute-shaders/op_mul_mat_mat_f32.comp
+        kompute-shaders/op_mul_mat_f16.comp
+        kompute-shaders/op_mul_mat_q8_0.comp
+        kompute-shaders/op_mul_mat_q4_0.comp
+        kompute-shaders/op_mul_mat_q4_1.comp
+        kompute-shaders/op_mul_mat_q4_k.comp
+        kompute-shaders/op_mul_mat_q6_k.comp
+        kompute-shaders/op_getrows_f32.comp
+        kompute-shaders/op_getrows_f16.comp
+        kompute-shaders/op_getrows_q4_0.comp
+        kompute-shaders/op_getrows_q4_1.comp
+        kompute-shaders/op_getrows_q6_k.comp
+        kompute-shaders/op_rope_f16.comp
+        kompute-shaders/op_rope_f32.comp
+        kompute-shaders/op_cpy_f16_f16.comp
+        kompute-shaders/op_cpy_f16_f32.comp
+        kompute-shaders/op_cpy_f32_f16.comp
+        kompute-shaders/op_cpy_f32_f32.comp
+    )
+
+    # Create a custom target for our generated shaders
+    add_custom_target(generated_shaders DEPENDS
+        shaderop_scale.h
+        shaderop_scale_8.h
+        shaderop_add.h
+        shaderop_addrow.h
+        shaderop_mul.h
+        shaderop_silu.h
+        shaderop_relu.h
+        shaderop_gelu.h
+        shaderop_softmax.h
+        shaderop_norm.h
+        shaderop_rmsnorm.h
+        shaderop_diagmask.h
+        shaderop_mul_mat_mat_f32.h
+        shaderop_mul_mat_f16.h
+        shaderop_mul_mat_q8_0.h
+        shaderop_mul_mat_q4_0.h
+        shaderop_mul_mat_q4_1.h
+        shaderop_mul_mat_q4_k.h
+        shaderop_mul_mat_q6_k.h
+        shaderop_getrows_f32.h
+        shaderop_getrows_f16.h
+        shaderop_getrows_q4_0.h
+        shaderop_getrows_q4_1.h
+        shaderop_getrows_q6_k.h
+        shaderop_rope_f16.h
+        shaderop_rope_f32.h
+        shaderop_cpy_f16_f16.h
+        shaderop_cpy_f16_f32.h
+        shaderop_cpy_f32_f16.h
+        shaderop_cpy_f32_f32.h
+    )
+
+    # Create a custom command that depends on the generated_shaders
+    add_custom_command(
+        OUTPUT ${CMAKE_CURRENT_BINARY_DIR}/ggml-kompute.stamp
+        COMMAND ${CMAKE_COMMAND} -E touch ${CMAKE_CURRENT_BINARY_DIR}/ggml-kompute.stamp
+        DEPENDS generated_shaders
+        COMMENT "Ensuring shaders are generated before compiling ggml-kompute.cpp"
+    )
+
+    # Add the stamp to the main sources to ensure dependency tracking
+    target_sources(ggml-kompute PRIVATE ${CMAKE_CURRENT_BINARY_DIR}/ggml-kompute.stamp)
+else()
+    message(WARNING "Kompute not found")
+endif()

ggml/src/ggml-kompute/kompute-shaders/common.comp

@@ -15,6 +15,7 @@
 #define TWOPI_F 6.283185307179586f
 
 #define QK_K 256
+#define K_SCALE_SIZE 12
 
 #define u8BufToU16(buf, idx) (((uint16_t(buf[idx + 1]) << 8)) | buf[idx])
 #define u8BufToFloat16(buf, idx) uint16BitsToHalf u8BufToU16(buf, idx)
@@ -64,6 +65,14 @@ mat4 dequantize_q4_1(const block_q4_1 xb, uint il) {
     return reg;
 }
 
+#define sizeof_block_q4_k 144
+struct block_q4_k {
+    float16_t d;
+    float16_t dmin;
+    uint8_t scales[K_SCALE_SIZE];
+    uint8_t qs[QK_K/2];
+};
+
 #define sizeof_block_q6_k 210
 struct block_q6_k {
     uint8_t ql[QK_K/2];      // quants, lower 4 bits

ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q4_k.comp

@@ -0,0 +1,133 @@
+#version 450
+
+#include "common.comp"
+
+#define N_DST 4
+#define SIZE_OF_BLOCK sizeof_block_q4_k
+
+layout(local_size_x = 4) in;
+layout(local_size_y = 8) in;
+layout(local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer tensorInA { block_q4_k inA[]; };
+layout (binding = 1) readonly buffer tensorInB { float inB[]; };
+layout (binding = 2) writeonly buffer tensorOut { float out_[]; };
+
+layout (push_constant) uniform parameter {
+    uint inAOff;
+    uint inBOff;
+    uint outOff;
+    int ne00;
+    int ne10;
+    int ne0;
+    int ne1;
+    int ne01;
+    int ne02;
+    int ne12;
+    int r2;
+    int r3;
+} pcs;
+
+void main() {
+    const uint16_t kmask1 = uint16_t(0x3f3f);
+    const uint16_t kmask2 = uint16_t(0x0f0f);
+    const uint16_t kmask3 = uint16_t(0xc0c0);
+
+    const uint ix = gl_SubgroupInvocationID/8;  // 0...3
+    const uint it = gl_SubgroupInvocationID%8;  // 0...7
+    const uint iq = it/4;     // 0 or 1
+    const uint ir = it%4;     // 0...3
+
+    const uint nb = pcs.ne00/QK_K;
+
+    const uint r0 = gl_WorkGroupID.x;
+    const uint r1 = gl_WorkGroupID.y;
+    const uint im = gl_WorkGroupID.z;
+
+    const uint first_row = r0 * N_DST;
+    const uint ib_row = first_row * nb;
+
+    const uint i12 = im%pcs.ne12;
+    const uint i13 = im/pcs.ne12;
+
+    const uint offset0 = (i12/pcs.r2)*(nb*pcs.ne01) + (i13/pcs.r3)*(nb*pcs.ne01*pcs.ne02);
+
+    const uint xblk = ib_row + offset0 + pcs.inAOff;
+    const uint y = r1*pcs.ne10 + im*pcs.ne00*pcs.ne1 + pcs.inBOff;
+
+    float yl[16];
+    float yh[16];
+    float sumf[N_DST] = {0.f, 0.f, 0.f, 0.f};
+    float all_sum = 0.f;
+
+    uint y4 = y + ix * QK_K + 64 * iq + 8 * ir;
+
+    for (uint ib = ix; ib < nb; ib += 4) {
+        const uint blk_idx = ib + xblk;
+
+        float sumy[4] = {0.f, 0.f, 0.f, 0.f};
+        for (int i = 0; i < 8; ++i) {
+            yl[i+0] = inB[y4+i+  0]; sumy[0] += yl[i+0];
+            yl[i+8] = inB[y4+i+ 32]; sumy[1] += yl[i+8];
+            yh[i+0] = inB[y4+i+128]; sumy[2] += yh[i+0];
+            yh[i+8] = inB[y4+i+160]; sumy[3] += yh[i+8];
+        }
+
+        for (int row = 0; row < N_DST; row++) {
+            uint row_idx = row * nb;
+
+            uint16_t sc_0 = u8BufToU16(inA[blk_idx + row_idx].scales, iq * 2 + 0);
+            uint16_t sc_1 = u8BufToU16(inA[blk_idx + row_idx].scales, iq * 2 + 2);
+            uint16_t sc_2 = u8BufToU16(inA[blk_idx + row_idx].scales, iq * 2 + 4);
+            uint16_t sc_3 = u8BufToU16(inA[blk_idx + row_idx].scales, iq * 2 + 6);
+            uint16_t sc_4 = u8BufToU16(inA[blk_idx + row_idx].scales, iq * 2 + 8);
+
+            uint16_t sc16[4];
+            sc16[0] = sc_0 & kmask1;
+            sc16[1] = sc_2 & kmask1;
+            sc16[2] = ((sc_4 >> 0) & kmask2) | ((sc_0 & kmask3) >> 2);
+            sc16[3] = ((sc_4 >> 4) & kmask2) | ((sc_2 & kmask3) >> 2);
+
+            float acc1[4] = {0.f, 0.f, 0.f, 0.f};
+            float acc2[4] = {0.f, 0.f, 0.f, 0.f};
+            for (int i = 0; i < 8; i += 2) {
+                uint16_t q1 = u8BufToU16(inA[blk_idx + row_idx].qs, 32 * iq + 8 * ir + i);
+                uint16_t q2 = u8BufToU16(inA[blk_idx + row_idx].qs, 64 + 32 * iq + 8 * ir + i);
+                acc1[0] += yl[i+0] * (q1 & 0x000F);
+                acc1[1] += yl[i+1] * (q1 & 0x0F00);
+                acc1[2] += yl[i+8] * (q1 & 0x00F0);
+                acc1[3] += yl[i+9] * (q1 & 0xF000);
+                acc2[0] += yh[i+0] * (q2 & 0x000F);
+                acc2[1] += yh[i+1] * (q2 & 0x0F00);
+                acc2[2] += yh[i+8] * (q2 & 0x00F0);
+                acc2[3] += yh[i+9] * (q2 & 0xF000);
+            }
+
+            uint8_t sc8_0 = uint8_t(sc16[0] & 0xFF);
+            uint8_t sc8_1 = uint8_t(sc16[0] >> 8 );
+            uint8_t sc8_2 = uint8_t(sc16[1] & 0xFF);
+            uint8_t sc8_3 = uint8_t(sc16[1] >> 8 );
+            uint8_t sc8_4 = uint8_t(sc16[2] & 0xFF);
+            uint8_t sc8_5 = uint8_t(sc16[2] >> 8 );
+            uint8_t sc8_6 = uint8_t(sc16[3] & 0xFF);
+            uint8_t sc8_7 = uint8_t(sc16[3] >> 8 );
+
+            float dall = float(inA[blk_idx + row_idx].d);
+            float dmin = float(inA[blk_idx + row_idx].dmin);
+            sumf[row] += dall * ((acc1[0] + 1.f/256.f * acc1[1]) * sc8_0 +
+                               (acc1[2] + 1.f/256.f * acc1[3]) * sc8_1 * 1.f/16.f +
+                               (acc2[0] + 1.f/256.f * acc2[1]) * sc8_4 +
+                               (acc2[2] + 1.f/256.f * acc2[3]) * sc8_5 * 1.f/16.f) -
+                dmin * (sumy[0] * sc8_2 + sumy[1] * sc8_3 + sumy[2] * sc8_6 + sumy[3] * sc8_7);
+        }
+
+        y4 += 4 * QK_K;
+    }
+
+    for (int row = 0; row < N_DST; ++row) {
+        all_sum = subgroupAdd(sumf[row]);
+        if (subgroupElect()) {
+            out_[r1*pcs.ne0 + im*pcs.ne0*pcs.ne1 + first_row + row + pcs.outOff] = all_sum;
+        }
+    }
+}