ci : use local ggml

.github/workflows/build.yml

@@ -313,7 +313,7 @@ jobs:
             cmake --build build --config ${{ matrix.build }} -j $(nproc)
 
   windows:
-    runs-on: windows-2019
+    runs-on: windows-latest
 
     strategy:
       matrix:
@@ -373,7 +373,7 @@ jobs:
           path: build/bin/${{ matrix.build }}
 
   windows-blas:
-    runs-on: windows-2019
+    runs-on: windows-latest
 
     strategy:
       matrix:

.gitignore

@@ -1,6 +1,5 @@
 *.o
 *.a
-*.d
 .cache/
 .coreml/
 .test/
@@ -20,9 +19,6 @@ 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 1.7.2)
+project("whisper.cpp" VERSION 1.7.1)
 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 # RISCV
+endif
 
 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 -DGGML_BLAS_USE_ACCELERATE
+		MK_CPPFLAGS += -DGGML_USE_ACCELERATE -DGGML_USE_BLAS
 		MK_CPPFLAGS += -DACCELERATE_NEW_LAPACK
 		MK_CPPFLAGS += -DACCELERATE_LAPACK_ILP64
 		MK_LDFLAGS  += -framework Accelerate
-		OBJ_GGML    += ggml/src/ggml-blas/ggml-blas.o
+		OBJ_GGML    += ggml/src/ggml-blas.o
 	endif
 endif # GGML_NO_ACCELERATE
 
@@ -464,38 +464,29 @@ 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/ggml-blas.o
+	OBJ_GGML    += ggml/src/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/ggml-blas.o
+	OBJ_GGML    += ggml/src/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/ggml-blas.o
+	OBJ_GGML    += ggml/src/ggml-blas.o
 endif # GGML_BLIS
 
 ifdef GGML_RPC
 	MK_CPPFLAGS += -DGGML_USE_RPC
-	OBJ_GGML    += ggml/src/ggml-rpc/ggml-rpc.o
+	OBJ_GGML    += ggml/src/ggml-rpc.o
 endif # GGML_RPC
 
 OBJ_CUDA_TMPL      = $(patsubst %.cu,%.o,$(wildcard ggml/src/ggml-cuda/template-instances/fattn-wmma*.cu))
@@ -522,7 +513,7 @@ 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/ggml-cuda.o
+	OBJ_GGML += ggml/src/ggml-cuda.o
 	OBJ_GGML += $(patsubst %.cu,%.o,$(wildcard ggml/src/ggml-cuda/*.cu))
 	OBJ_GGML += $(OBJ_CUDA_TMPL)
 ifdef WHISPER_FATAL_WARNINGS
@@ -624,11 +615,11 @@ ggml/src/ggml-cuda/%.o: \
 	ggml/src/ggml-cuda/common.cuh
 	$(NVCC_COMPILE)
 
-ggml/src/ggml-cuda/ggml-cuda.o: \
+ggml/src/ggml-cuda.o: \
-	ggml/src/ggml-cuda/ggml-cuda.cu \
+	ggml/src/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)
@@ -751,43 +742,50 @@ 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/ggml-metal.o
+	OBJ_GGML	+= ggml/src/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/ggml-metal-embed.o
+	OBJ_GGML    += ggml/src/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/ggml-metal.o: \
+ggml/src/ggml-metal.o: \
-	ggml/src/ggml-metal/ggml-metal.m \
+	ggml/src/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/ggml-metal-embed.o: \
+ggml/src/ggml-metal-embed.o: \
-	ggml/src/ggml-metal/ggml-metal.metal \
+	ggml/src/ggml-metal.metal \
-	ggml/src/ggml-metal/ggml-metal-impl.h \
 	ggml/src/ggml-common.h
 	@echo "Embedding Metal library"
-	@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-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
-	@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))
-	$(eval TEMP_ASSEMBLY=$(shell mktemp -d))
+	@echo ".section __DATA, __ggml_metallib"            >  $(TEMP_ASSEMBLY)
-	@echo ".section __DATA, __ggml_metallib"                       >  $(TEMP_ASSEMBLY)/ggml-metal-embed.s
+	@echo ".globl _ggml_metallib_start"                 >> $(TEMP_ASSEMBLY)
-	@echo ".globl _ggml_metallib_start"                            >> $(TEMP_ASSEMBLY)/ggml-metal-embed.s
+	@echo "_ggml_metallib_start:"                       >> $(TEMP_ASSEMBLY)
-	@echo "_ggml_metallib_start:"                                  >> $(TEMP_ASSEMBLY)/ggml-metal-embed.s
+	@echo ".incbin \"ggml/src/ggml-metal-embed.metal\"" >> $(TEMP_ASSEMBLY)
-	@echo ".incbin \"ggml/src/ggml-metal/ggml-metal-embed.metal\"" >> $(TEMP_ASSEMBLY)/ggml-metal-embed.s
+	@echo ".globl _ggml_metallib_end"                   >> $(TEMP_ASSEMBLY)
-	@echo ".globl _ggml_metallib_end"                              >> $(TEMP_ASSEMBLY)/ggml-metal-embed.s
+	@echo "_ggml_metallib_end:"                         >> $(TEMP_ASSEMBLY)
-	@echo "_ggml_metallib_end:"                                    >> $(TEMP_ASSEMBLY)/ggml-metal-embed.s
+	@$(AS) $(TEMP_ASSEMBLY) -o $@
-	$(CC) $(CFLAGS) -c $(TEMP_ASSEMBLY)/ggml-metal-embed.s -o $@
+	@rm -f ${TEMP_ASSEMBLY}
-	@rm -f ${TEMP_ASSEMBLY}/ggml-metal-embed.s
-	@rmdir ${TEMP_ASSEMBLY}
 endif
 endif # GGML_METAL
 
@@ -803,17 +801,11 @@ endif
 
 OBJ_GGML += \
 	ggml/src/ggml.o \
-	ggml/src/ggml-aarch64.o \
+	ggml/src/ggml-cpu.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-threading.o \
+	ggml/src/ggml-aarch64.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
@@ -918,64 +910,114 @@ endif
 # Build libraries
 #
 
-LIB_GGML   = libggml.so
+# ggml
-LIB_GGML_S = libggml.a
 
-LIB_LLAMA   = libllama.so
+ggml/src/ggml.o: \
-LIB_LLAMA_S = libllama.a
+	ggml/src/ggml.c \
+	ggml/include/ggml.h
+	$(CC)  $(CFLAGS)   -c $< -o $@
 
-LIB_COMMON   = libcommon.so
+ggml/src/ggml-cpu.o: \
-LIB_COMMON_S = libcommon.a
+	ggml/src/ggml-cpu.c \
-
-LIB_COMMON_SDL   = libcommon-sdl.so
-LIB_COMMON_SDL_S = libcommon-sdl.a
-
-# Targets
-BUILD_TARGETS += $(LIB_GGML) $(LIB_GGML_S) $(LIB_LLAMA) $(LIB_LLAMA_S) $(LIB_COMMON) $(LIB_COMMON_S)
-
-# Dependency files
-DEP_FILES = $(OBJ_GGML:.o=.d) $(OBJ_LLAMA:.o=.d) $(OBJ_COMMON:.o=.d)
-
-# Default target
-all: $(BUILD_TARGETS)
-
-# 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/src/ggml-common.h
-	ggml/src/ggml-backend-impl.h \
+	$(CC)  $(CFLAGS)   -c $< -o $@
-	ggml/include/ggml-cpu.h \
-	ggml/src/ggml-impl.h
-	$(CXX) $(CXXFLAGS)   -c $< -o $@
 
-# Rules for building object files
+ggml/src/ggml-alloc.o: \
-ggml/%.o: ggml/%.c
+	ggml/src/ggml-alloc.c \
-	$(CC) $(CFLAGS) -MMD -c $< -o $@
+	ggml/include/ggml.h \
+	ggml/include/ggml-alloc.h
+	$(CC)  $(CFLAGS)   -c $< -o $@
 
-ggml/%.o: ggml/%.cpp
+ggml/src/ggml-backend.o: \
-	$(CXX) $(CXXFLAGS) -MMD -c $< -o $@
+	ggml/src/ggml-backend.cpp \
+	ggml/include/ggml.h \
+	ggml/include/ggml-backend.h
+	$(CXX) $(CXXFLAGS) -c $< -o $@
 
-src/%.o: src/%.cpp
+ggml/src/ggml-quants.o: \
-	$(CXX) $(CXXFLAGS) -MMD -c $< -o $@
+	ggml/src/ggml-quants.c \
+	ggml/include/ggml.h \
+	ggml/src/ggml-quants.h \
+	ggml/src/ggml-common.h
+	$(CC) $(CFLAGS)    -c $< -o $@
 
-examples/%.o: examples/%.cpp
+ggml/src/ggml-aarch64.o: \
-	$(CXX) $(CXXFLAGS) -MMD -c $< -o $@
+	ggml/src/ggml-aarch64.c \
+	ggml/include/ggml.h \
+	ggml/src/ggml-aarch64.h \
+	ggml/src/ggml-common.h
+	$(CC) $(CFLAGS)    -c $< -o $@
 
-# Rules for building libraries
+ggml/src/ggml-blas.o: \
-$(LIB_GGML): $(OBJ_GGML)
+	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): $(OBJ_GGML)
+$(LIB_GGML_S): \
+	$(OBJ_GGML)
 	ar rcs $(LIB_GGML_S) $^
 
-$(LIB_LLAMA): $(OBJ_LLAMA) $(LIB_GGML)
+# whisper
+
+src/whisper.o: \
+	src/whisper.cpp \
+	include/whisper.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 $@
+
+$(LIB_WHISPER): \
+	$(OBJ_WHISPER) \
+	$(LIB_GGML)
 	$(CXX) $(CXXFLAGS) -shared -fPIC -o $@ $^ $(LDFLAGS)
 
-$(LIB_LLAMA_S): $(OBJ_LLAMA)
+$(LIB_WHISPER_S): \
-	ar rcs $(LIB_LLAMA_S) $^
+	$(OBJ_WHISPER) \
+	$(OBJ_GGML)
+	ar rcs $(LIB_WHISPER_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)
+	$(CXX) $(CXXFLAGS) -shared -fPIC -o $@ $^ $(LDFLAGS)
+
+$(LIB_COMMON_S): \
+	$(OBJ_COMMON)
+	ar rcs $(LIB_COMMON_S) $^
 
 # common-sdl
 
@@ -987,21 +1029,34 @@ examples/common-sdl.o: \
 	examples/common-sdl.h
 	$(CXX) $(CXXFLAGS) $(CFLAGS_SDL) -c $< -o $@
 
-$(LIB_COMMON): $(OBJ_COMMON) $(LIB_LLAMA) $(LIB_GGML)
+$(LIB_COMMON_SDL): \
-	$(CXX) $(CXXFLAGS) -shared -fPIC -o $@ $^ $(LDFLAGS)
+	$(OBJ_SDL)
+	$(CXX) $(CXXFLAGS) -shared -fPIC -o $@ $^ $(LDFLAGS) $(LDFLAGS_SDL)
 
-$(LIB_COMMON_S): $(OBJ_COMMON)
+$(LIB_COMMON_SDL_S): \
-	ar rcs $(LIB_COMMON_S) $^
+	$(OBJ_SDL)
+	ar rcs $(LIB_COMMON_SDL_S) $^
 
-# Include dependency files
--include $(DEP_FILES)
-
-# Clean rule
 clean:
-	rm -vrf $(BUILD_TARGETS) $(TEST_TARGETS)
+	rm -vrf *.dot $(BUILD_TARGETS) $(TEST_TARGETS)
-	rm -rvf *.a *.dll *.so *.dot
+	rm -rvf src/*.o
-	find ggml src tests examples -type f -name "*.o" -delete
+	rm -rvf src/coreml/*.o
-	find ggml src tests examples -type f -name "*.d" -delete
+	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
 
 #
 # Examples

Package.swift

@@ -28,7 +28,7 @@ let package = Package(
                "tests",
                "CMakeLists.txt",
                "Makefile",
-               "ggml/src/ggml-metal/ggml-metal-embed.metal"
+               "ggml/src/ggml-metal-embed.metal"
             ],
             sources: [
                 "ggml/src/ggml.c",
@@ -36,22 +36,16 @@ let package = Package(
                 "ggml/src/ggml-aarch64.c",
                 "ggml/src/ggml-alloc.c",
                 "ggml/src/ggml-backend.cpp",
-                "ggml/src/ggml-backend-reg.cpp",
+                "ggml/src/ggml-cpu.c",
-                "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-threading.cpp",
+                "ggml/src/ggml-metal.m"
-                "ggml/src/ggml-metal/ggml-metal.m"
             ],
-            resources: [.process("ggml/src/ggml-metal/ggml-metal.metal")],
+            resources: [.process("ggml/src/ggml-metal.metal")],
             publicHeadersPath: "spm-headers",
             cSettings: [
                 .unsafeFlags(["-Wno-shorten-64-to-32", "-O3", "-DNDEBUG"]),
-                .unsafeFlags(["-fno-objc-arc"]),
-                .headerSearchPath("ggml/src"),
                 .define("GGML_USE_ACCELERATE"),
+                .unsafeFlags(["-fno-objc-arc"]),
                 .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: [v1.7.2](https://github.com/ggerganov/whisper.cpp/releases/tag/v1.7.2) / [Roadmap | F.A.Q.](https://github.com/ggerganov/whisper.cpp/discussions/126)
+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)
 
 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": "1.7.2",
+  "version": "1.7.1",
   "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(ls-sycl-device PROPERTIES FOLDER "examples")
+        set_target_properties(sycl 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 whisper ${CMAKE_THREAD_LIBS_INIT})
+target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
-target_compile_features(${TARGET} PRIVATE cxx_std_17)
+target_compile_features(${TARGET} PRIVATE cxx_std_17)

examples/sycl/build.sh

@@ -7,16 +7,13 @@ cd build
 source /opt/intel/oneapi/setvars.sh
 
 #for FP16
-#cmake .. -DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DWHISPER_SYCL_F16=ON # faster for long-prompt inference
+#cmake .. -DWHISPER_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DWHISPER_SYCL_F16=ON # faster for long-prompt inference
 
 #for FP32
-cmake .. -DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
+cmake .. -DWHISPER_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

@@ -179,7 +179,6 @@ 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,
@@ -233,7 +232,6 @@ 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"       },
@@ -1209,25 +1207,6 @@ 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,
         {
@@ -2928,15 +2907,9 @@ 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;
 
@@ -3481,13 +3454,21 @@ 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();
 
-        ggml_backend_buffer_type_t buft;
+        const llama_model::buft_list_t * buft_list;
         if (offload) {
-            auto * dev = model.dev_layer.at(i).dev;
+            buft_list = model.dev_layer.at(i).buft_list;
-            buft = ggml_backend_dev_buffer_type(dev);
         } else {
-            buft = ggml_backend_cpu_buffer_type();
+            buft_list = &model.cpu_buft_list;
         }
+        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) {
@@ -4294,8 +4275,8 @@ struct llama_model_loader {
     int n_tensors = 0;
     int n_created = 0;
 
-    uint64_t n_elements = 0;
+    int64_t n_elements = 0;
-    size_t  n_bytes     = 0;
+    size_t  n_bytes    = 0;
 
     bool use_mmap = false;
     bool check_tensors;
@@ -5363,11 +5344,6 @@ 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) {
@@ -5898,17 +5874,6 @@ 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);
@@ -7289,7 +7254,7 @@ 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_dev_get_extra_bufts");
+        ggml_backend_reg_get_proc_address(cpu_reg, "ggml_backend_cpu_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) {
@@ -7556,7 +7521,7 @@ 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_dev && buft == ggml_backend_dev_host_buffer_type(buft_dev)) {
+            if (ml.use_mmap && 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);
             }
@@ -8591,31 +8556,6 @@ 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);
@@ -9188,10 +9128,6 @@ 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;
@@ -9316,7 +9252,6 @@ 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 &&
@@ -14481,130 +14416,6 @@ 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
@@ -16797,10 +16608,6 @@ 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();
@@ -18213,7 +18020,7 @@ 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 (!llama_kv_cache_can_shift(&lctx)) {
+        if (lctx.model.arch == LLM_ARCH_DEEPSEEK2) { // not supported due to MLA
             GGML_ABORT("Deepseek2 does not support K-shift");
         }
 
@@ -18790,7 +18597,6 @@ 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);
 
@@ -20070,7 +19876,6 @@ 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:
@@ -20144,11 +19949,19 @@ 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) {
-    return model->n_bytes;
+    uint64_t size = 0;
+    for (const auto & it : model->tensors_by_name) {
+        size += ggml_nbytes(it.second);
+    }
+    return size;
 }
 
 uint64_t llama_model_n_params(const struct llama_model * model) {
-    return model->n_elements;
+    uint64_t nparams = 0;
+    for (const auto & it : model->tensors_by_name) {
+        nparams += ggml_nelements(it.second);
+    }
+    return nparams;
 }
 
 struct ggml_tensor * llama_get_model_tensor(struct llama_model * model, const char * name) {
@@ -20462,10 +20275,6 @@ 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);
@@ -22212,6 +22021,7 @@ 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())         + " | ";
@@ -22257,6 +22067,28 @@ 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

@@ -667,9 +667,6 @@ 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
     //
@@ -1247,6 +1244,8 @@ 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

@@ -19,16 +19,11 @@ 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

@@ -25,11 +25,6 @@
 		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 */; };
@@ -55,8 +50,8 @@
 		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/ggml-metal.m"; 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/ggml-metal.metal"; 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>"; };
 		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>"; };
@@ -82,17 +77,8 @@
 		18ABE1572AF556340044A204 /* ggml-backend.cpp */ = {isa = PBXFileReference; explicitFileType = sourcecode.cpp.cpp; fileEncoding = 4; name = "ggml-backend.cpp"; path = "../../../ggml/src/ggml-backend.cpp"; sourceTree = "<group>"; };
 		18ABE1582AF556340044A204 /* ggml-impl.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; name = "ggml-impl.h"; path = "../../../ggml/src/ggml-impl.h"; sourceTree = "<group>"; };
 		18ABE1592AF556340044A204 /* ggml-quants.c */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.c; name = "ggml-quants.c"; path = "../../../ggml/src/ggml-quants.c"; sourceTree = "<group>"; };
-		18E864A82CE73C1E0094B8B3 /* ggml-cpu.c */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.c.c; name = "ggml-cpu.c"; path = "../../../ggml/src/ggml-cpu/ggml-cpu.c"; sourceTree = "<group>"; };
+		18E864A82CE73C1E0094B8B3 /* ggml-cpu.c */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.c.c; name = "ggml-cpu.c"; path = "../../../ggml/src/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>"; };
@@ -132,15 +118,6 @@
 		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 */,
@@ -275,16 +252,11 @@
 				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 */,
 			);
@@ -363,7 +335,6 @@
 				GCC_WARN_UNINITIALIZED_AUTOS = YES_AGGRESSIVE;
 				GCC_WARN_UNUSED_FUNCTION = YES;
 				GCC_WARN_UNUSED_VARIABLE = YES;
-				HEADER_SEARCH_PATHS = "";
 				IPHONEOS_DEPLOYMENT_TARGET = 16.0;
 				MTL_ENABLE_DEBUG_INFO = INCLUDE_SOURCE;
 				MTL_FAST_MATH = YES;
@@ -417,7 +388,6 @@
 				GCC_WARN_UNINITIALIZED_AUTOS = YES_AGGRESSIVE;
 				GCC_WARN_UNUSED_FUNCTION = YES;
 				GCC_WARN_UNUSED_VARIABLE = YES;
-				HEADER_SEARCH_PATHS = "";
 				IPHONEOS_DEPLOYMENT_TARGET = 16.0;
 				MTL_ENABLE_DEBUG_INFO = NO;
 				MTL_FAST_MATH = YES;
@@ -440,7 +410,6 @@
 				DEVELOPMENT_TEAM = P8JZH34X63;
 				GCC_WARN_64_TO_32_BIT_CONVERSION = NO;
 				GENERATE_INFOPLIST_FILE = YES;
-				HEADER_SEARCH_PATHS = ../../../ggml/src/;
 				INFOPLIST_FILE = whisper.objc/Info.plist;
 				INFOPLIST_KEY_UIApplicationSupportsIndirectInputEvents = YES;
 				INFOPLIST_KEY_UILaunchStoryboardName = LaunchScreen;
@@ -470,7 +439,6 @@
 				DEVELOPMENT_TEAM = P8JZH34X63;
 				GCC_WARN_64_TO_32_BIT_CONVERSION = NO;
 				GENERATE_INFOPLIST_FILE = YES;
-				HEADER_SEARCH_PATHS = ../../../ggml/src/;
 				INFOPLIST_FILE = whisper.objc/Info.plist;
 				INFOPLIST_KEY_UIApplicationSupportsIndirectInputEvents = YES;
 				INFOPLIST_KEY_UILaunchStoryboardName = LaunchScreen;

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

@@ -67,6 +67,8 @@ 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

@@ -92,7 +92,6 @@ 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})
@@ -117,7 +116,6 @@ 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)
@@ -128,9 +126,14 @@ option(GGML_LLAMAFILE                       "ggml: use LLAMAFILE"
 
 option(GGML_CUDA                            "ggml: use CUDA"                                  OFF)
 option(GGML_MUSA                            "ggml: use MUSA"                                  OFF)
+option(GGML_CUDA_FORCE_DMMV                 "ggml: use dmmv instead of mmvq CUDA kernels"     OFF)
 option(GGML_CUDA_FORCE_MMQ                  "ggml: use mmq kernels instead of cuBLAS"         OFF)
 option(GGML_CUDA_FORCE_CUBLAS               "ggml: always use cuBLAS instead of mmq kernels"  OFF)
+set   (GGML_CUDA_DMMV_X   "32" CACHE STRING "ggml: x stride for dmmv CUDA kernels")
+set   (GGML_CUDA_MMV_Y     "1" CACHE STRING "ggml: y block size for mmv CUDA kernels")
 option(GGML_CUDA_F16                        "ggml: use 16 bit floats for some calculations"   OFF)
+set   (GGML_CUDA_KQUANTS_ITER "2" CACHE STRING
+                                            "ggml: iters./thread per block for Q2_K/Q6_K")
 set   (GGML_CUDA_PEER_MAX_BATCH_SIZE "128" CACHE STRING
                                             "ggml: max. batch size for using peer access")
 option(GGML_CUDA_NO_PEER_COPY               "ggml: do not use peer to peer copies"            OFF)
@@ -138,7 +141,7 @@ 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_HIP                             "ggml: use HIP"                                   OFF)
+option(GGML_HIPBLAS                         "ggml: use hipBLAS"                               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)
@@ -164,8 +167,6 @@ 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"    ${GGML_STANDALONE})
@@ -225,7 +226,6 @@ set(GGML_PUBLIC_HEADERS
     include/ggml-cann.h
     include/ggml-cuda.h
     include/ggml-kompute.h
-    include/ggml-opt.h
     include/ggml-metal.h
     include/ggml-rpc.h
     include/ggml-sycl.h
@@ -235,14 +235,15 @@ set_target_properties(ggml PROPERTIES PUBLIC_HEADER "${GGML_PUBLIC_HEADERS}")
 #if (GGML_METAL)
 #    set_target_properties(ggml PROPERTIES RESOURCE "${CMAKE_CURRENT_SOURCE_DIR}/src/ggml-metal.metal")
 #endif()
-install(TARGETS ggml LIBRARY PUBLIC_HEADER)
+install(TARGETS ggml PUBLIC_HEADER)
-install(TARGETS ggml-base LIBRARY)
+
+if (BUILD_SHARED_LIBS)
+    install(TARGETS ggml LIBRARY)
+endif()
 
-# 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/ggml-metal.metal
+        FILES src/ggml-metal.metal
         PERMISSIONS
             OWNER_READ
             OWNER_WRITE

ggml/ggml_vk_generate_shaders.py

@@ -0,0 +1,220 @@
+#!/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_BACKEND_API ggml_backend_buffer_type_t ggml_backend_amx_buffer_type(void);
+GGML_API ggml_backend_buffer_type_t ggml_backend_amx_buffer_type(void);
 
-GGML_BACKEND_API bool ggml_backend_is_amx(ggml_backend_t backend);
+GGML_API bool ggml_backend_is_amx(ggml_backend_t backend);
 
 // backend API
-GGML_BACKEND_API ggml_backend_t ggml_backend_amx_init(void);
+GGML_API ggml_backend_t ggml_backend_amx_init(void);
 
-GGML_BACKEND_API void ggml_backend_amx_set_n_threads(ggml_backend_t backend_amx, int n_threads);
+GGML_API void ggml_backend_amx_set_n_threads(ggml_backend_t backend_amx, int n_threads);
 
-GGML_BACKEND_API ggml_backend_reg_t ggml_backend_amx_reg(void);
+GGML_API ggml_backend_reg_t ggml_backend_amx_reg(void);
 
 #ifdef  __cplusplus
 }

ggml/include/ggml-backend.h

@@ -3,20 +3,6 @@
 #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
@@ -86,7 +72,7 @@ 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 in tensor->data for setting/getting data
+    // "offset" refers to the offset of the 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);
@@ -242,20 +228,14 @@ extern "C" {
         ggml_backend_sched_reserve(sched, reserve_graph);
 
         // compute
-        graph = build_graph(sched); // the graph and its tensors are single-use in terms of allocation, multi-use in terms of computation
+        graph = build_graph(sched);
-        for (int i = 0; i < 10; ++i) {
+        ggml_backend_sched_graph_compute(sched, graph);
-            ggml_backend_sched_graph_compute(sched, graph); // on the first iteration the graph is allocated automatically
-        }
 
         // if there are graph inputs:
-        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);
-        ggml_backend_sched_reset(sched); // clear the allocation of the previous graph
+        ggml_backend_sched_alloc_graph(sched, graph);
-        ggml_backend_sched_alloc_graph(sched, graph); // explicitly allocate the new graph but do not execute it
+        ggml_backend_tensor_set(input_tensor, ...);
-        ggml_backend_tensor_set(input_tensor, ...); // copy data to the newly allocated graph tensors
+        ggml_backend_sched_graph_compute(sched, graph);
-        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
     }
     */
 
@@ -270,7 +250,7 @@ extern "C" {
     //
     typedef bool (*ggml_backend_sched_eval_callback)(struct ggml_tensor * t, bool ask, void * user_data);
 
-    // Initialize a backend scheduler, backends with low index are given priority over backends with high index
+    // Initialize a backend scheduler
     GGML_API ggml_backend_sched_t ggml_backend_sched_new(ggml_backend_t * backends, ggml_backend_buffer_type_t * bufts, int n_backends, size_t graph_size, bool parallel);
     GGML_API void                 ggml_backend_sched_free(ggml_backend_sched_t sched);
 
@@ -295,9 +275,7 @@ 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 or allocating a new graph.
+    // Reset all assignments and allocators - must be called before changing the node backends
-    // 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_BACKEND_API ggml_backend_t ggml_backend_blas_init(void);
+GGML_API ggml_backend_t ggml_backend_blas_init(void);
 
-GGML_BACKEND_API bool ggml_backend_is_blas(ggml_backend_t backend);
+GGML_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_BACKEND_API void ggml_backend_blas_set_n_threads(ggml_backend_t backend_blas, int n_threads);
+GGML_API void ggml_backend_blas_set_n_threads(ggml_backend_t backend_blas, int n_threads);
 
-GGML_BACKEND_API ggml_backend_reg_t ggml_backend_blas_reg(void);
+GGML_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_BACKEND_API ggml_backend_reg_t ggml_backend_cann_reg(void);
+GGML_API ggml_backend_reg_t ggml_backend_cann_reg(void);
 
 /**
  * @brief Initializes the CANN backend for a specified device.
@@ -46,7 +46,7 @@ GGML_BACKEND_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_BACKEND_API ggml_backend_t ggml_backend_cann_init(int32_t device);
+GGML_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_BACKEND_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_BACKEND_API bool ggml_backend_is_cann(ggml_backend_t backend);
+GGML_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_BACKEND_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_BACKEND_API ggml_backend_buffer_type_t
+GGML_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_BACKEND_API int32_t ggml_backend_cann_get_device_count(void);
+GGML_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_BACKEND_API ggml_backend_buffer_type_t ggml_backend_cann_host_buffer_type(void);
+GGML_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_BACKEND_API ggml_backend_buffer_type_t ggml_backend_cann_host_buffer_type(v
  * @param description Pointer to a buffer where the description will be written.
  * @param description_size Size of the description buffer.
  */
-GGML_BACKEND_API void ggml_backend_cann_get_device_description(
+GGML_API void ggml_backend_cann_get_device_description(
     int32_t device, char* description, size_t description_size);
 
 /**
@@ -114,7 +114,7 @@ GGML_BACKEND_API void ggml_backend_cann_get_device_description(
  * @param total Pointer to a variable where the total memory size will be
  * stored.
  */
-GGML_BACKEND_API void ggml_backend_cann_get_device_memory(int32_t device,
+GGML_API void ggml_backend_cann_get_device_memory(int32_t device,
                                                   size_t* free,
                                                   size_t* total);
 

ggml/include/ggml-cpu.h

@@ -54,77 +54,54 @@ extern "C" {
         GGML_NUMA_STRATEGY_COUNT
     };
 
-    GGML_BACKEND_API void    ggml_numa_init(enum ggml_numa_strategy numa); // call once for better performance on NUMA systems
+    GGML_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 bool    ggml_is_numa(void); // true if init detected that system has >1 NUMA node
 
-    GGML_BACKEND_API struct ggml_tensor * ggml_new_i32(struct ggml_context * ctx, int32_t value);
+    GGML_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_new_f32(struct ggml_context * ctx, float value);
 
-    GGML_BACKEND_API struct ggml_tensor * ggml_set_i32 (struct ggml_tensor * tensor, int32_t value);
+    GGML_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 struct ggml_tensor * ggml_set_f32 (struct ggml_tensor * tensor, float value);
 
-    GGML_BACKEND_API int32_t ggml_get_i32_1d(const struct ggml_tensor * tensor, int i);
+    GGML_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 void    ggml_set_i32_1d(const struct ggml_tensor * tensor, int i, int32_t value);
 
-    GGML_BACKEND_API int32_t ggml_get_i32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3);
+    GGML_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 void    ggml_set_i32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3, int32_t value);
 
-    GGML_BACKEND_API float   ggml_get_f32_1d(const struct ggml_tensor * tensor, int i);
+    GGML_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 void    ggml_set_f32_1d(const struct ggml_tensor * tensor, int i, float value);
 
-    GGML_BACKEND_API float   ggml_get_f32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3);
+    GGML_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 void    ggml_set_f32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3, float value);
 
-    GGML_BACKEND_API struct ggml_threadpool_params ggml_threadpool_params_default(int n_threads);
+    GGML_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_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_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_API struct ggml_threadpool *      ggml_threadpool_new          (struct ggml_threadpool_params  * params);
-    GGML_BACKEND_API void                          ggml_threadpool_free         (struct ggml_threadpool * threadpool);
+    GGML_API void                          ggml_threadpool_free         (struct ggml_threadpool * threadpool);
-    GGML_BACKEND_API int                           ggml_threadpool_get_n_threads(struct ggml_threadpool * threadpool);
+    GGML_API int                           ggml_threadpool_get_n_threads(struct ggml_threadpool * threadpool);
-    GGML_BACKEND_API void                          ggml_threadpool_pause        (struct ggml_threadpool * threadpool);
+    GGML_API void                          ggml_threadpool_pause        (struct ggml_threadpool * threadpool);
-    GGML_BACKEND_API void                          ggml_threadpool_resume       (struct ggml_threadpool * threadpool);
+    GGML_API void                          ggml_threadpool_resume       (struct ggml_threadpool * threadpool);
 
     // ggml_graph_plan() has to be called before ggml_graph_compute()
     // when plan.work_size > 0, caller must allocate memory for plan.work_data
-    GGML_BACKEND_API struct ggml_cplan ggml_graph_plan(
+    GGML_API struct ggml_cplan ggml_graph_plan(
                   const struct ggml_cgraph * cgraph,
                                        int   n_threads, /* = GGML_DEFAULT_N_THREADS */
                     struct ggml_threadpool * threadpool /* = NULL */ );
-    GGML_BACKEND_API enum ggml_status  ggml_graph_compute(struct ggml_cgraph * cgraph, struct ggml_cplan * cplan);
+    GGML_API enum ggml_status  ggml_graph_compute(struct ggml_cgraph * cgraph, struct ggml_cplan * cplan);
 
     // same as ggml_graph_compute() but the work data is allocated as a part of the context
     // note: the drawback of this API is that you must have ensured that the context has enough memory for the work data
-    GGML_BACKEND_API enum ggml_status  ggml_graph_compute_with_ctx(struct ggml_context * ctx, struct ggml_cgraph * cgraph, int n_threads);
+    GGML_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
-    // system info
+    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);
-    // x86
+    // get the sve vector length in bytes
-    GGML_BACKEND_API int ggml_cpu_has_sse3       (void);
+    GGML_API int ggml_cpu_get_sve_cnt(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
 
@@ -138,7 +115,6 @@ 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;
@@ -148,30 +124,27 @@ extern "C" {
         ggml_gemm_t              gemm;
     };
 
-    GGML_BACKEND_API const struct ggml_type_traits_cpu * ggml_get_type_traits_cpu(enum ggml_type type);
+    GGML_API const struct ggml_type_traits_cpu * ggml_get_type_traits_cpu(enum ggml_type type);
 
-    GGML_BACKEND_API void ggml_cpu_init(void);
+    GGML_API void ggml_cpu_init(void);
 
     //
     // CPU backend
     //
 
-    GGML_BACKEND_API ggml_backend_t ggml_backend_cpu_init(void);
+    GGML_API ggml_backend_t ggml_backend_cpu_init(void);
 
-    GGML_BACKEND_API bool ggml_backend_is_cpu                (ggml_backend_t backend);
+    GGML_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_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_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 void ggml_backend_cpu_set_abort_callback(ggml_backend_t backend_cpu, ggml_abort_callback abort_callback, void * abort_callback_data);
 
-    GGML_BACKEND_API ggml_backend_reg_t ggml_backend_cpu_reg(void);
+    GGML_API ggml_backend_reg_t ggml_backend_cpu_reg(void);
 
 #ifdef GGML_USE_CPU_HBM
-    GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_cpu_hbm_buffer_type(void);
+    GGML_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_HIP
+#ifdef GGML_USE_HIPBLAS
 #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_BACKEND_API ggml_backend_t ggml_backend_cuda_init(int device);
+GGML_API ggml_backend_t ggml_backend_cuda_init(int device);
 
-GGML_BACKEND_API bool ggml_backend_is_cuda(ggml_backend_t backend);
+GGML_API bool ggml_backend_is_cuda(ggml_backend_t backend);
 
 // device buffer
-GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_cuda_buffer_type(int device);
+GGML_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_BACKEND_API ggml_backend_buffer_type_t ggml_backend_cuda_split_buffer_type(int main_device, const float * tensor_split);
+GGML_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_BACKEND_API ggml_backend_buffer_type_t ggml_backend_cuda_host_buffer_type(void);
+GGML_API ggml_backend_buffer_type_t ggml_backend_cuda_host_buffer_type(void);
 
-GGML_BACKEND_API int  ggml_backend_cuda_get_device_count(void);
+GGML_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_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 void ggml_backend_cuda_get_device_memory(int device, size_t * free, size_t * total);
 
-GGML_BACKEND_API bool ggml_backend_cuda_register_host_buffer(void * buffer, size_t size);
+GGML_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 void ggml_backend_cuda_unregister_host_buffer(void * buffer);
 
-GGML_BACKEND_API ggml_backend_reg_t ggml_backend_cuda_reg(void);
+GGML_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_BACKEND_API ggml_backend_t ggml_backend_kompute_init(int device);
+GGML_API ggml_backend_t ggml_backend_kompute_init(int device);
 
-GGML_BACKEND_API bool ggml_backend_is_kompute(ggml_backend_t backend);
+GGML_API bool ggml_backend_is_kompute(ggml_backend_t backend);
 
-GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_kompute_buffer_type(int device);
+GGML_API ggml_backend_buffer_type_t ggml_backend_kompute_buffer_type(int device);
 
-GGML_BACKEND_API ggml_backend_reg_t ggml_backend_kompute_reg(void);
+GGML_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_BACKEND_API ggml_backend_t ggml_backend_metal_init(void);
+GGML_API ggml_backend_t ggml_backend_metal_init(void);
 
-GGML_BACKEND_API bool ggml_backend_is_metal(ggml_backend_t backend);
+GGML_API bool ggml_backend_is_metal(ggml_backend_t backend);
 
 GGML_DEPRECATED(
-        GGML_BACKEND_API ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void * data, size_t size, size_t max_size),
+        GGML_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_BACKEND_API void ggml_backend_metal_set_abort_callback(ggml_backend_t backend, ggml_abort_callback abort_callback, void * user_data);
+GGML_API void ggml_backend_metal_set_abort_callback(ggml_backend_t backend, ggml_abort_callback abort_callback, void * user_data);
 
-GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_metal_buffer_type(void);
+GGML_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_BACKEND_API bool ggml_backend_metal_supports_family(ggml_backend_t backend, int family);
+GGML_API bool ggml_backend_metal_supports_family(ggml_backend_t backend, int family);
 
 // capture all command buffers committed the next time `ggml_backend_graph_compute` is called
-GGML_BACKEND_API void ggml_backend_metal_capture_next_compute(ggml_backend_t backend);
+GGML_API void ggml_backend_metal_capture_next_compute(ggml_backend_t backend);
 
-GGML_BACKEND_API ggml_backend_reg_t ggml_backend_metal_reg(void);
+GGML_API ggml_backend_reg_t ggml_backend_metal_reg(void);
 
 #ifdef __cplusplus
 }

ggml/include/ggml-opt.h

@@ -1,216 +0,0 @@
-// 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_BACKEND_API ggml_backend_t ggml_backend_rpc_init(const char * endpoint);
+GGML_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 bool ggml_backend_is_rpc(ggml_backend_t backend);
 
-GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_rpc_buffer_type(const char * endpoint);
+GGML_API ggml_backend_buffer_type_t ggml_backend_rpc_buffer_type(const char * endpoint);
 
-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_get_device_memory(const char * endpoint, size_t * free, size_t * total);
 
-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 void ggml_backend_rpc_start_server(ggml_backend_t backend, const char * endpoint, size_t free_mem, size_t total_mem);
 
-GGML_BACKEND_API ggml_backend_reg_t ggml_backend_rpc_reg(void);
+GGML_API ggml_backend_reg_t ggml_backend_rpc_reg(void);
 
-GGML_BACKEND_API ggml_backend_dev_t ggml_backend_rpc_add_device(const char * endpoint);
+GGML_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_BACKEND_API ggml_backend_t ggml_backend_sycl_init(int device);
+GGML_API ggml_backend_t ggml_backend_sycl_init(int device);
 
-GGML_BACKEND_API bool ggml_backend_is_sycl(ggml_backend_t backend);
+GGML_API bool ggml_backend_is_sycl(ggml_backend_t backend);
 
 // devide buffer
-GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_sycl_buffer_type(int device);
+GGML_API ggml_backend_buffer_type_t ggml_backend_sycl_buffer_type(int device);
 
 // split tensor buffer that splits matrices by rows across multiple devices
-GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_sycl_split_buffer_type(const float * tensor_split);
+GGML_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_BACKEND_API ggml_backend_buffer_type_t ggml_backend_sycl_host_buffer_type(void);
+GGML_API ggml_backend_buffer_type_t ggml_backend_sycl_host_buffer_type(void);
 
-GGML_BACKEND_API void ggml_backend_sycl_print_sycl_devices(void);
+GGML_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_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,
+GGML_API void ggml_backend_sycl_get_device_description(int device,
                                                        char *description,
                                                        size_t description_size);
-GGML_BACKEND_API int  ggml_backend_sycl_get_device_count();
+GGML_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);
+GGML_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_BACKEND_API bool ggml_backend_sycl_register_host_buffer(void * buffer, size_t size);
+// GGML_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 void ggml_backend_sycl_unregister_host_buffer(void * buffer);
 
-GGML_BACKEND_API ggml_backend_reg_t ggml_backend_sycl_reg(void);
+GGML_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_BACKEND_API void ggml_vk_instance_init(void);
+GGML_API void ggml_vk_instance_init(void);
 
 // backend API
-GGML_BACKEND_API ggml_backend_t ggml_backend_vk_init(size_t dev_num);
+GGML_API ggml_backend_t ggml_backend_vk_init(size_t dev_num);
 
-GGML_BACKEND_API bool ggml_backend_is_vk(ggml_backend_t backend);
+GGML_API bool ggml_backend_is_vk(ggml_backend_t backend);
-GGML_BACKEND_API int  ggml_backend_vk_get_device_count(void);
+GGML_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_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 void ggml_backend_vk_get_device_memory(int device, size_t * free, size_t * total);
 
-GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_vk_buffer_type(size_t dev_num);
+GGML_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_BACKEND_API ggml_backend_buffer_type_t ggml_backend_vk_host_buffer_type(void);
+GGML_API ggml_backend_buffer_type_t ggml_backend_vk_host_buffer_type(void);
 
-GGML_BACKEND_API ggml_backend_reg_t ggml_backend_vk_reg(void);
+GGML_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) extern
+#            define GGML_API __declspec(dllexport)
 #        else
-#            define GGML_API __declspec(dllimport) extern
+#            define GGML_API __declspec(dllimport)
 #        endif
 #    else
-#        define GGML_API __attribute__ ((visibility ("default"))) extern
+#        define GGML_API __attribute__ ((visibility ("default")))
 #    endif
 #else
-#    define GGML_API extern
+#    define GGML_API
 #endif
 
 // TODO: support for clang
@@ -602,6 +602,7 @@ extern "C" {
 
         int32_t flags;
 
+        struct ggml_tensor * grad;
         struct ggml_tensor * src[GGML_MAX_SRC];
 
         // source tensor and offset for views
@@ -614,7 +615,7 @@ extern "C" {
 
         void * extra; // extra things e.g. for ggml-cuda.cu
 
-        char padding[8];
+        // char padding[4];
     };
 
     static const size_t GGML_TENSOR_SIZE = sizeof(struct ggml_tensor);
@@ -1489,7 +1490,7 @@ extern "C" {
         "use ggml_rope_ext_inplace instead");
 
     // compute correction dims for YaRN RoPE scaling
-    GGML_API void ggml_rope_yarn_corr_dims(
+    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
@@ -1984,20 +1985,28 @@ extern "C" {
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
             struct ggml_tensor  * grad,
-            struct ggml_tensor  * m,
+            float                 alpha,
-            struct ggml_tensor  * v,
+            float                 beta1,
-            struct ggml_tensor  * adamw_params); // parameters such a the learning rate
+            float                 beta2,
+            float                 eps,
+            float                 wd); // weight decay
 
     //
     // automatic differentiation
     //
 
-    GGML_API void ggml_build_forward_expand(struct ggml_cgraph * cgraph, struct ggml_tensor * tensor);
+    GGML_API void ggml_build_forward_expand (struct ggml_cgraph * cgraph, struct ggml_tensor * tensor);
-    GGML_API void ggml_build_backward_expand(
+    GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool accumulate);
-        struct ggml_context * ctx_static,  // context for static gradients (loss + gradient accumulation)
+
-        struct ggml_context * ctx_compute, // context for gradient computation
+    GGML_API void ggml_build_opt_adamw(
-        struct ggml_cgraph  * cgraph,
+            struct ggml_context * ctx,
-        bool                  accumulate); // whether or not gradients should be accumulated, requires static allocation of tensors in ctx_static
+            struct ggml_cgraph  * gf,
+            struct ggml_cgraph  * gb,
+            float                 alpha,
+            float                 beta1,
+            float                 beta2,
+            float                 eps,
+            float                 wd); // weight decay
 
     // graph allocation in a context
     GGML_API struct ggml_cgraph * ggml_new_graph       (struct ggml_context * ctx); // size = GGML_DEFAULT_GRAPH_SIZE, grads = false
@@ -2017,9 +2026,7 @@ 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  (const struct ggml_cgraph * cgraph, const char * name);
+    GGML_API struct ggml_tensor * ggml_graph_get_tensor(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);
@@ -2030,15 +2037,198 @@ 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);
 
-    // TODO these functions were sandwiched in the old optimization interface, is there a better place for them?
+    // 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);
     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
     //
@@ -2194,6 +2384,38 @@ 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
@@ -2210,6 +2432,7 @@ 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

@@ -1,5 +1,9 @@
+// SPDX-FileCopyrightText: Copyright 2024 Arm Ltd.
 #pragma once
 
+#define GGML_COMMON_DECL_C
+#include "ggml-common.h"
+
 #include "ggml.h"
 
 // GGML internal header
@@ -8,11 +12,27 @@
 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

@@ -466,12 +466,18 @@ 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)) {
@@ -810,11 +816,7 @@ 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 = 0;
+    size_t node_size = (node->data || node->view_src) ? 0 : ggml_backend_buft_get_alloc_size(galloc->bufts[talloc->buffer_id], node);
-    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.cpp

@@ -317,6 +317,8 @@ 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);
@@ -324,6 +326,7 @@ 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
@@ -418,18 +421,9 @@ 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 nullptr;
+    return ggml_backend_t{};
 }
 
 void ggml_backend_amx_set_n_threads(ggml_backend_t backend_amx, int n_threads) {
@@ -439,8 +433,4 @@ 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/CMakeLists.txt

@@ -1,107 +0,0 @@
-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

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

ggml/src/ggml-amx/mmq.cpp

@@ -496,20 +496,19 @@ 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) {
-    // FIXME: using unoptimized reference impl until moved to CPU backend
+    quantize_row_q8_0(x, vy, k);
-    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_ref(x, (block_q8_1 *)vy, k);
+    quantize_row_q8_1(x, 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_ref(x, (block_q8_K *)vy, k);
+    quantize_row_q8_K(x, vy, k);
 #else
     quantize_row_q8_K_vnni(x, vy, k);
 #endif

ggml/src/ggml-backend-reg.cpp

@@ -1,195 +0,0 @@
-#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);
 }
 
-void ggml_backend_tensor_memset(struct ggml_tensor * tensor, uint8_t value, 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) {
     ggml_backend_buffer_t buf = tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
 
     if (size == 0) {
@@ -525,6 +525,197 @@ 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 {
@@ -689,7 +880,7 @@ 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->view_src ? tensor->view_src->buffer : tensor->buffer;
+    ggml_backend_buffer_t buffer = tensor->buffer;
     if (buffer == NULL) {
         return -1;
     }
@@ -722,6 +913,8 @@ 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) {
@@ -740,7 +933,7 @@ 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 (%s) in a backend that cannot run the operation", tensor->name);
+        GGML_ABORT("pre-allocated tensor in a backend that cannot run the operation");
     }
 
     // graph input
@@ -1447,7 +1640,7 @@ 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_dev_type(ggml_backend_get_device(backends[n_backends - 1])) == GGML_BACKEND_DEVICE_TYPE_CPU);
+    GGML_ASSERT(ggml_backend_is_cpu(backends[n_backends - 1])); // last backend must be CPU
 
     struct ggml_backend_sched * sched = (ggml_backend_sched *) calloc(1, sizeof(struct ggml_backend_sched));
 
@@ -1536,13 +1729,12 @@ 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;
 }
@@ -1844,6 +2036,17 @@ 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) {
@@ -1917,9 +2120,7 @@ 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";
@@ -1960,7 +2161,7 @@ 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  = */ NULL, // FIXME ggml_backend_reg_dev_get(ggml_backend_cpu_reg(), 0),
+        /* .device  = */ ggml_backend_reg_dev_get(ggml_backend_cpu_reg(), 0),
         /* .context = */ NULL,
     };
 
@@ -1983,14 +2184,479 @@ 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  = */ NULL, // FIXME ggml_backend_reg_dev_get(ggml_backend_cpu_reg(), 0),
+        /* .device  = */ 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; // TMP: workaround until sync with latest ggml
+            return op->src[1]->type == GGML_TYPE_F32 || op->src[1]->type == ggml_get_type_traits_cpu(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.cpp

@@ -6,7 +6,7 @@
 #include <vector>
 #include <cstring>
 
-#if defined(GGML_BLAS_USE_ACCELERATE)
+#if defined(GGML_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_BLAS_USE_ACCELERATE)
+    #if defined(GGML_USE_ACCELERATE)
         return "Accelerate";
     #elif defined(GGML_BLAS_USE_MKL)
         return "MKL";

ggml/src/ggml-blas/CMakeLists.txt

@@ -1,90 +0,0 @@
-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-cann/CMakeLists.txt

@@ -1,46 +0,0 @@
-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-impl.h

@@ -27,6 +27,80 @@ 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__
@@ -314,6 +388,28 @@ 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
@@ -366,6 +462,153 @@ 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.c

@@ -1,15 +1,13 @@
 #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__)
@@ -44,7 +42,7 @@
 #endif
 
 #ifdef GGML_USE_LLAMAFILE
-#include "llamafile/sgemm.h"
+#include <llamafile/sgemm.h>
 #endif
 
 #if defined(_MSC_VER)
@@ -106,6 +104,9 @@ 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;
@@ -260,13 +261,11 @@ 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)
@@ -276,7 +275,6 @@ 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)
@@ -285,20 +283,27 @@ 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,
@@ -309,106 +314,85 @@ 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,
@@ -417,7 +401,6 @@ 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,
@@ -426,7 +409,6 @@ 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,
@@ -435,13 +417,11 @@ 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,
@@ -1469,12 +1449,8 @@ 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__) || defined(__AVX__)
+#elif defined(__AVX2__)
-#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();
@@ -2274,7 +2250,22 @@ 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);
@@ -2369,7 +2360,7 @@ 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__ > 33) || defined(__COSMOPOLITAN__)
+#if __GLIBC__ > 2 || (__GLIBC__ == 2 && __GLIBC_MINOR__ > 28) || 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
@@ -3006,8 +2997,8 @@ static void ggml_compute_forward_dup_f16(
                         id += ne00 * (ne01 - ir1);
                     }
                 }
-            } else if (ggml_get_type_traits_cpu(dst->type)->from_float) {
+            } else if (ggml_get_type_traits(dst->type)->from_float) {
-                ggml_from_float_t const quantize_row_q = ggml_get_type_traits_cpu(dst->type)->from_float;
+                ggml_from_float_t const quantize_row_q = ggml_get_type_traits(dst->type)->from_float;
                 float * src0_f32 = (float *) params->wdata + (ne00 + CACHE_LINE_SIZE_F32) * ith;
 
                 size_t id = 0;
@@ -3287,8 +3278,8 @@ static void ggml_compute_forward_dup_bf16(
                         id += ne00 * (ne01 - ir1);
                     }
                 }
-            } else if (ggml_get_type_traits_cpu(dst->type)->from_float) {
+            } else if (ggml_get_type_traits(dst->type)->from_float) {
-                ggml_from_float_t const quantize_row_q = ggml_get_type_traits_cpu(dst->type)->from_float;
+                ggml_from_float_t const quantize_row_q = ggml_get_type_traits(dst->type)->from_float;
                 float * src0_f32 = (float *) params->wdata + (ne00 + CACHE_LINE_SIZE_F32) * ith;
 
                 size_t id = 0;
@@ -3603,8 +3594,8 @@ static void ggml_compute_forward_dup_f32(
                         id += rs * (ne01 - ir1);
                     }
                 }
-            } else if (ggml_get_type_traits_cpu(dst->type)->from_float) {
+            } else if (ggml_get_type_traits(dst->type)->from_float) {
-                ggml_from_float_t const quantize_row_q = ggml_get_type_traits_cpu(dst->type)->from_float;
+                ggml_from_float_t const quantize_row_q = ggml_get_type_traits(dst->type)->from_float;
 
                 size_t id = 0;
                 size_t rs = nb0 * (ne00 / ggml_blck_size(dst->type));
@@ -4386,7 +4377,7 @@ 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_cpu(dtype)->from_float;
+    ggml_from_float_t const quantize_row_q = ggml_get_type_traits(dtype)->from_float;
 
     // we don't support permuted src0 or src1
     GGML_ASSERT(nb00 == ggml_type_size(type));
@@ -4688,7 +4679,7 @@ 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_cpu(type)->from_float;
+    ggml_from_float_t const quantize_row_q = ggml_get_type_traits(type)->from_float;
 
     // we don't support permuted src0
     GGML_ASSERT(nb00 == ggml_type_size(type));
@@ -7334,7 +7325,6 @@ 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,
@@ -7346,6 +7336,8 @@ 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;
@@ -7433,14 +7425,10 @@ static void ggml_compute_forward_mul_mat(
     const int ith = params->ith;
     const int nth = params->nth;
 
-    enum ggml_type type = src0->type;
+    const 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           = type_traits_cpu[vec_dot_type].from_float;
+    ggml_from_float_t        const from_float           = ggml_get_type_traits(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;
@@ -7476,15 +7464,15 @@ 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(type),
+                if (!llamafile_sgemm(ne01, ne11, ne00/ggml_blck_size(src0->type),
                                      (const char *)src0->data + i12/r2*nb02 + i13/r3*nb03,
-                                     nb01/ggml_type_size(type),
+                                     nb01/ggml_type_size(src0->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,
-                                     type,
+                                     src0->type,
                                      src1->type,
                                      dst->type))
                     goto UseGgmlGemm1;
@@ -7537,15 +7525,15 @@ 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(type),
+                if (!llamafile_sgemm(ne01, ne11, ne00/ggml_blck_size(src0->type),
                                      (const char *)src0->data + i12/r2*nb02 + i13/r3*nb03,
-                                     nb01/ggml_type_size(type),
+                                     nb01/ggml_type_size(src0->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,
-                                     type,
+                                     src0->type,
                                      vec_dot_type,
                                      dst->type))
                     goto UseGgmlGemm2;
@@ -7630,7 +7618,7 @@ 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, type, num_rows_per_vec_dot, ir0_start, ir0_end, ir1_start, ir1_end);
+        ggml_compute_forward_mul_mat_one_chunk(params, dst, num_rows_per_vec_dot, ir0_start, ir0_end, ir1_start, ir1_end);
 
         if (nth >= nchunk0 * nchunk1) {
             break;
@@ -7661,7 +7649,7 @@ 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      = type_traits_cpu[vec_dot_type].from_float;
+    ggml_from_float_t const from_float      = ggml_get_type_traits(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;
 
@@ -9171,6 +9159,12 @@ 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) {
@@ -9187,6 +9181,16 @@ 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,
@@ -10664,7 +10668,7 @@ 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   = type_traits_cpu[k_vec_dot_type].from_float;
+    ggml_from_float_t const q_to_vec_dot   = ggml_get_type_traits(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;
 
@@ -12216,16 +12220,11 @@ 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        = dst->src[0];
-    const struct ggml_tensor * src0_grad    = dst->src[1];
+    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_m = dst->src[2];
-    const struct ggml_tensor * src0_grad_v  = dst->src[3];
+    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;
@@ -12242,14 +12241,16 @@ static void ggml_compute_forward_opt_step_adamw_f32(
     const int ir0 = dr*ith;
     const int ir1 = MIN(ir0 + dr, nr);
 
-    const float * adamw_params_ptr = ggml_get_data_f32(adamw_params);
+    /* const float   gnorm = 1.0f; */
-    const float alpha  = adamw_params_ptr[0];
+    int64_t       iter;   memcpy(&iter, &dst->op_params[0], sizeof(int64_t));
-    const float beta1  = adamw_params_ptr[1];
+    const float   alpha = ggml_get_op_params_f32(dst, 2);
-    const float beta2  = adamw_params_ptr[2];
+    const float   beta1 = ggml_get_op_params_f32(dst, 3);
-    const float eps    = adamw_params_ptr[3];
+    const float   beta2 = ggml_get_op_params_f32(dst, 4);
-    const float wd     = adamw_params_ptr[4];
+    const float   eps   = ggml_get_op_params_f32(dst, 5);
-    const float beta1h = adamw_params_ptr[5];
+    const float   wd    = ggml_get_op_params_f32(dst, 6);
-    const float beta2h = adamw_params_ptr[6];
+
+    const float beta1h  = alpha/(1.0f - powf(beta1, iter));
+    const float beta2h  =  1.0f/(1.0f - powf(beta2, iter));
 
     for (int ir = ir0; ir < ir1; ++ir) {
         const int64_t i03 = ir/(ne02*ne01);
@@ -12273,9 +12274,17 @@ 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) - alpha*mh/vh;
+            w[i00] = w[i00]*(1.0f - alpha*wd) - 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(
@@ -13750,151 +13759,6 @@ 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/CMakeLists.txt

@@ -1,261 +0,0 @@
-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

@@ -1,30 +0,0 @@
-#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-quants.h

@@ -1,63 +0,0 @@
-#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.cpp

@@ -1,663 +0,0 @@
-#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.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_HIP) && defined(GGML_HIP_UMA)
+#if defined(GGML_USE_HIPBLAS) && 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_HIP)
+#if !defined(GGML_USE_HIPBLAS)
     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_HIP)
+#endif // !defined(GGML_USE_HIPBLAS)
 
 #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_HIP) && !defined(GGML_CUDA_NO_VMM)
+#if !defined(GGML_USE_HIPBLAS) && !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_HIP) && !defined(GGML_CUDA_NO_VMM)
+#endif // !defined(GGML_USE_HIPBLAS) && !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_HIP) && defined(__HIP_PLATFORM_AMD__)
+#if defined(GGML_USE_HIPBLAS) && 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_HIP) && defined(__HIP_PLATFORM_AMD__)
+#endif // defined(GGML_USE_HIPBLAS) && 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_HIP) && !defined(GGML_CUDA_NO_VMM)
+#if !defined(GGML_USE_HIPBLAS) && !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_HIP) && !defined(GGML_CUDA_NO_VMM)
+#endif // !defined(GGML_USE_HIPBLAS) && !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_HIP) && !defined(GGML_CUDA_NO_VMM)
+#if !defined(GGML_USE_HIPBLAS) && !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_HIP) && !defined(GGML_CUDA_NO_VMM)
+#endif // !defined(GGML_USE_HIPBLAS) && !defined(GGML_CUDA_NO_VMM)
     return std::unique_ptr<ggml_cuda_pool>(new ggml_cuda_pool_leg(device));
 }
 
@@ -1020,6 +1020,114 @@ 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) {
 
@@ -1217,7 +1325,7 @@ 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_HIP) && !defined(GGML_USE_MUSA)
+#if !defined(GGML_USE_HIPBLAS) && !defined(GGML_USE_MUSA)
     // cudaMemcpy2DAsync may fail with copies between vmm pools of different devices
     cudaMemcpy3DPeerParms p = {};
     p.dstDevice = dstDevice;
@@ -1231,7 +1339,7 @@ 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_HIP) && !defined(GGML_USE_MUSA)
+#endif // !defined(GGML_USE_HIPBLAS) && !defined(GGML_USE_MUSA)
 }
 
 static void ggml_cuda_op_mul_mat(
@@ -1546,6 +1654,58 @@ 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,
@@ -1719,17 +1879,21 @@ 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_mul_mat_vec   = src0->type == GGML_TYPE_F16
+    bool use_dequantize_mul_mat_vec = ggml_cuda_dmmv_type_supported(src0->type)
         && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32
-        && src0->ne[0] % 2 == 0 && src1->ne[1] == 1;
+        && src0->ne[0] % (GGML_CUDA_DMMV_X*2) == 0 && src1->ne[1] == 1;
-    bool use_mul_mat_vec_q = ggml_is_quantized(src0->type)
+    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;
 
-    bool any_gpus_with_slow_fp16   = false;
+    // if mmvq is available it's a better choice than dmmv:
-    bool any_gpus_without_fp16_mma = false;
+#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;
 
     if (split) {
         ggml_backend_cuda_split_buffer_type_context * buft_ctx = (ggml_backend_cuda_split_buffer_type_context *) src0->buffer->buft->context;
@@ -1740,16 +1904,14 @@ static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor
                 continue;
             }
 
-            const int cc              = ggml_cuda_info().devices[id].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]);
+            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_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;
+        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]);
+        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_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
@@ -1760,16 +1922,18 @@ 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 && use_mul_mat_vec && dst->ne[3] == 1 && (src0->ne[1] < MMV_MAX_ROWS || any_gpus_without_fp16_mma)) {
+    if (!split && any_gpus_with_slow_fp16 && src0->type == GGML_TYPE_F16 && ggml_is_permuted(src0) && ggml_is_permuted(src1) && src1->ne[1] == 1) {
-        // the custom F16 vector kernel can be used over batched cuBLAS GEMM
+        // FP32 precision KQ single-batch for batch size 1 without FlashAttention
-        // 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_p021(ctx, src0, src1, dst);
-        ggml_cuda_mul_mat_vec(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);
     } 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) {
-        // general KQ + KQV multi-batch without FlashAttention
+        // KQ + KQV multi-batch without FlashAttention
         ggml_cuda_mul_mat_batched_cublas(ctx, src0, src1, dst);
-    } else if (use_mul_mat_vec) {
+    } else if (use_dequantize_mul_mat_vec) {
-        ggml_cuda_op_mul_mat(ctx, src0, src1, dst, ggml_cuda_op_mul_mat_vec, nullptr);
+        ggml_cuda_op_mul_mat(ctx, src0, src1, dst, ggml_cuda_op_dequantize_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) {
@@ -2814,17 +2978,6 @@ 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/CMakeLists.txt

@@ -1,155 +0,0 @@
-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_HIP)
+#if defined(GGML_USE_HIPBLAS)
 #define GGML_COMMON_DECL_HIP
 #define GGML_COMMON_IMPL_HIP
 #else
@@ -26,13 +26,13 @@
 #include <string>
 #include <vector>
 
-#if defined(GGML_USE_HIP)
+#if defined(GGML_USE_HIPBLAS)
 #include "vendors/hip.h"
 #elif defined(GGML_USE_MUSA)
 #include "vendors/musa.h"
 #else
 #include "vendors/cuda.h"
-#endif // defined(GGML_USE_HIP)
+#endif // defined(GGML_USE_HIPBLAS)
 
 #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_HIP)
+#if !defined(GGML_USE_HIPBLAS)
 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_HIP) && defined(__HIP_PLATFORM_AMD__)) || __CUDA_ARCH__ >= CC_PASCAL
+#if (defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) || __CUDA_ARCH__ >= CC_PASCAL
 #define FP16_AVAILABLE
-#endif // (defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) || __CUDA_ARCH__ >= CC_PASCAL
+#endif // (defined(GGML_USE_HIPBLAS) && 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_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_VOLTA
+#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_VOLTA
 #define FP16_MMA_AVAILABLE
-#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_VOLTA
+#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_VOLTA
 
-#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_TURING
+#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_TURING
 #define INT8_MMA_AVAILABLE
-#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_TURING
+#endif // !(defined(GGML_USE_HIPBLAS) && 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_HIP) && defined(__HIP_PLATFORM_AMD__)
+#if defined(GGML_USE_HIPBLAS) && 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_HIP) && defined(__HIP_PLATFORM_AMD__)
+#endif // defined(GGML_USE_HIPBLAS) && 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_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_AMPERE
+#if !(defined(GGML_USE_HIPBLAS) && 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_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_AMPERE
+#endif // !(defined(GGML_USE_HIPBLAS) && 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_HIP) && defined(__HIP_PLATFORM_AMD__)
+#if defined(GGML_USE_HIPBLAS) && 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_HIP) && defined(__HIP_PLATFORM_AMD__)
+#endif // defined(GGML_USE_HIPBLAS) && 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_HIP) && defined(__HIP_PLATFORM_AMD__)) && CUDART_VERSION < CUDART_HMAX
+#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && CUDART_VERSION < CUDART_HMAX
     return __float2half(fmaxf(__half2float(a), __half2float(b)));
 #else
     return __hmax(a, b);
-#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && CUDART_VERSION < CUDART_HMAX
+#endif // !(defined(GGML_USE_HIPBLAS) && 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_HIP) && defined(__HIP_PLATFORM_AMD__))
+#if !(defined(GGML_USE_HIPBLAS) && 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_HIP) && defined(__HIP_PLATFORM_AMD__))
+#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
 }
 
 static __device__ __forceinline__ half2 warp_reduce_max(half2 x) {
-#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL
+#if !(defined(GGML_USE_HIPBLAS) && 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_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL
+#endif // !(defined(GGML_USE_HIPBLAS) && 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_HIP) && defined(__HIP_PLATFORM_AMD__)
+#if defined(GGML_USE_HIPBLAS) && 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_HIP) && defined(__HIP_PLATFORM_AMD__)
+#else // defined(GGML_USE_HIPBLAS) && 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_HIP) && defined(__HIP_PLATFORM_AMD__)
+#endif // defined(GGML_USE_HIPBLAS) && 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_HIP) && defined(__HIP_PLATFORM_AMD__))
+#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
 __launch_bounds__(D, 1)
-#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
+#endif // !(defined(GGML_USE_HIPBLAS) && 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_HIP) && defined(__HIP_PLATFORM_AMD__))
+#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
 __launch_bounds__(nwarps*WARP_SIZE, 1)
-#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
+#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
 static __global__ void flash_attn_tile_ext_f16(
         const char * __restrict__ Q,
         const char * __restrict__ K,

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_HIP) && defined(__HIP_PLATFORM_AMD__))
+#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
 __launch_bounds__(nwarps*WARP_SIZE, 1)
-#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
+#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
 static __global__ void flash_attn_tile_ext_f32(
         const char * __restrict__ Q,
         const char * __restrict__ K,

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_HIP) && defined(__HIP_PLATFORM_AMD__))
+#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
 __launch_bounds__(D, 1)
-#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
+#endif // !(defined(GGML_USE_HIPBLAS) && 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_HIP) && defined(__HIP_PLATFORM_AMD__))
+#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
 __launch_bounds__(D, 1)
-#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
+#endif // !(defined(GGML_USE_HIPBLAS) && 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_HIP) && defined(__HIP_PLATFORM_AMD__))
+#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
 __launch_bounds__(nwarps*WARP_SIZE, 1)
-#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
+#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
 static __global__ void flash_attn_ext_f16(
         const char * __restrict__ Q,
         const char * __restrict__ K,

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

@@ -1,155 +0,0 @@
-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_HIP) && defined(__HIP_PLATFORM_AMD__)
+#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
     return 128;
-#else // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
+#else // defined(GGML_USE_HIPBLAS) && 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_HIP) && defined(__HIP_PLATFORM_AMD__)
+#endif // defined(GGML_USE_HIPBLAS) && 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_HIP) && defined(__HIP_PLATFORM_AMD__)
+#if defined(GGML_USE_HIPBLAS) && 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_HIP) && defined(__HIP_PLATFORM_AMD__)
+#endif // defined(GGML_USE_HIPBLAS) && 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_HIP) && defined(__HIP_PLATFORM_AMD__)
+#if defined(GGML_USE_HIPBLAS) && 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_HIP) && defined(__HIP_PLATFORM_AMD__)
+#endif // defined(GGML_USE_HIPBLAS) && 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_HIP) && defined(__HIP_PLATFORM_AMD__)) || __CUDA_ARCH__ < CC_VOLTA
+#if (defined(GGML_USE_HIPBLAS) && 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_HIP) && defined(__HIP_PLATFORM_AMD__)) || __CUDA_ARCH__ < CC_VOLTA
+#endif // (defined(GGML_USE_HIPBLAS) && 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_HIP) && defined(__HIP_PLATFORM_AMD__))
+#if !(defined(GGML_USE_HIPBLAS) && 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_HIP) && defined(__HIP_PLATFORM_AMD__))
+#endif // !(defined(GGML_USE_HIPBLAS) && 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

@@ -1,223 +0,0 @@
-#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

@@ -1,12 +0,0 @@
-#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_HIP) && defined(__HIP_PLATFORM_AMD__))
+#if !(defined(GGML_USE_HIPBLAS) && 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_HIP) && defined(__HIP_PLATFORM_AMD__))
+#endif // !(defined(GGML_USE_HIPBLAS) && 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_HIP) && defined(__HIP_PLATFORM_AMD__) && (defined(RDNA2) || defined(RDNA3))
+#if defined(GGML_USE_HIPBLAS) && 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_HIP) && defined(__HIP_PLATFORM_AMD__) && !defined(RDNA2) && !defined(RDNA3)
+#endif // defined(GGML_USE_HIPBLAS) && 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,
+    float * __restrict__ x, const float * __restrict__ g, float * __restrict__ g_m, float * __restrict__ g_v, const int64_t k,
-    const float * __restrict__ pars, 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) {
 
     const int64_t i = (int64_t) blockIdx.x*blockDim.x + threadIdx.x;
 
@@ -13,14 +13,6 @@ 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);
@@ -31,48 +23,58 @@ 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) - alpha*mh/vh;
+    x[i] = x[i]*(1.0f - alpha*wd) - mh/vh;
 }
 
 static void opt_step_adamw_f32_cuda(
-    float * x, const float * g, float * g_m, float * g_v, const float * pars, const int64_t k, cudaStream_t stream) {
+    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) {
 
     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, pars, k);
+    opt_step_adamw_f32<<<block_nums, block_dims, 0, stream>>>(x, g, g_m, g_v, k, alpha, beta1, beta2, eps, wd, beta1h, beta2h);
 }
 
 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        = dst->src[0];
-    const ggml_tensor * src0_grad    = dst->src[1];
+    const ggml_tensor * src0_grad   = dst->src[1];
-    const ggml_tensor * src0_grad_m  = dst->src[2];
+    const ggml_tensor * src0_grad_m = dst->src[2];
-    const ggml_tensor * src0_grad_v  = dst->src[3];
+    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->type        == GGML_TYPE_F32);
-    GGML_ASSERT(src0_grad->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_m->type == GGML_TYPE_F32);
-    GGML_ASSERT(src0_grad_v->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;
+    float       * src0_d        = (float       *) src0->data;
-    const float * src0_grad_d    = (const float *) src0_grad->data;
+    const float * src0_grad_d   = (const float *) src0_grad->data;
-    float       * src0_grad_m_d  = (float       *) src0_grad_m->data;
+    float       * src0_grad_m_d = (float       *) src0_grad_m->data;
-    float       * src0_grad_v_d  = (float       *) src0_grad_v->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);
 
-    opt_step_adamw_f32_cuda(src0_d, src0_grad_d, src0_grad_m_d, src0_grad_v_d, adamw_params_d, ne, stream);
+    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));
 }

ggml/src/ggml-cuda/sum.cu

@@ -1,6 +1,6 @@
-#if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA) && CUDART_VERSION >= 11700
+#if !defined(GGML_USE_HIPBLAS) && !defined(GGML_USE_MUSA) && CUDART_VERSION >= 11700
 #define USE_CUB
-#endif // !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA) && CUDART_VERSION >= 11700
+#endif // !defined(GGML_USE_HIPBLAS) && !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

@@ -1,106 +0,0 @@
-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

@@ -3,29 +3,13 @@
 // 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
@@ -44,13 +28,13 @@ extern "C" {
 // if C99 - static_assert is noop
 // ref: https://stackoverflow.com/a/53923785/4039976
 #ifndef __cplusplus
-    #ifndef static_assert
+#ifndef static_assert
-        #if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201100L)
+#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201100L)
-            #define static_assert(cond, msg) _Static_assert(cond, msg)
+#define static_assert(cond, msg) _Static_assert(cond, msg)
-        #else
+#else
-            #define static_assert(cond, msg) struct global_scope_noop_trick
+#define static_assert(cond, msg) struct global_scope_noop_trick
-        #endif
+#endif
-    #endif
+#endif
 #endif
 
 static inline int ggml_up32(int n) {
@@ -136,12 +120,14 @@ 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;
@@ -196,7 +182,7 @@ 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, const struct ggml_tensor * key);
+static size_t ggml_hash_find(const struct ggml_hash_set * hash_set, 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);
@@ -210,7 +196,7 @@ 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, const struct ggml_tensor * key) {
+static size_t ggml_hash_find(const struct ggml_hash_set * hash_set, struct ggml_tensor * key) {
     size_t h = ggml_hash(key) % hash_set->size;
 
     // linear probing
@@ -281,14 +267,13 @@ enum ggml_cgraph_eval_order {
 };
 
 struct ggml_cgraph {
-    int size;    // maximum number of nodes/leafs/grads/grad_accs
+    int size;
-    int n_nodes; // number of nodes currently in use
+    int n_nodes;
-    int n_leafs; // number of leafs currently in use
+    int n_leafs;
 
-    struct ggml_tensor ** nodes;     // tensors with data that can change if the graph is evaluated
+    struct ggml_tensor ** nodes;
-    struct ggml_tensor ** grads;     // the outputs of these tensors are the gradients of the nodes
+    struct ggml_tensor ** grads;
-    struct ggml_tensor ** grad_accs; // accumulators for node gradients
+    struct ggml_tensor ** leafs;
-    struct ggml_tensor ** leafs;     // tensors with constant data
 
     struct ggml_hash_set visited_hash_set;
 
@@ -302,249 +287,9 @@ 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);
 
-// FP16 to FP32 conversion
+// TODO: move to threading file
-
+void ggml_critical_section_start(void);
-#if defined(__ARM_NEON)
+void ggml_critical_section_end(void);
-    #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

@@ -1,162 +0,0 @@
-
-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/op_mul_mat_q4_k.comp

@@ -1,133 +0,0 @@
-#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;
-        }
-    }
-}

ggml/src/ggml-metal.m

@@ -2,7 +2,6 @@
 
 #import "ggml-impl.h"
 #import "ggml-backend-impl.h"
-#import "ggml-metal-impl.h"
 
 #import <Foundation/Foundation.h>
 
@@ -126,7 +125,6 @@ enum ggml_metal_kernel_type {
     GGML_METAL_KERNEL_TYPE_GELU_QUICK_4,
     GGML_METAL_KERNEL_TYPE_SILU,
     GGML_METAL_KERNEL_TYPE_SILU_4,
-    GGML_METAL_KERNEL_TYPE_ELU,
     GGML_METAL_KERNEL_TYPE_SOFT_MAX_F16,
     GGML_METAL_KERNEL_TYPE_SOFT_MAX_F16_4,
     GGML_METAL_KERNEL_TYPE_SOFT_MAX_F32,
@@ -518,10 +516,6 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de
                     [prep setObject:@"1" forKey:@"GGML_METAL_USE_BF16"];
                 }
 
-#if GGML_METAL_EMBED_LIBRARY
-                [prep setObject:@"1" forKey:@"GGML_METAL_EMBED_LIBRARY"];
-#endif
-
                 MTLCompileOptions * options = [MTLCompileOptions new];
                 options.preprocessorMacros = prep;
 
@@ -650,7 +644,6 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GELU_QUICK_4,                  gelu_quick_4,                   true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SILU,                          silu,                           true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SILU_4,                        silu_4,                         true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ELU,                           elu,                            true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SOFT_MAX_F16,                  soft_max_f16,                   has_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SOFT_MAX_F16_4,                soft_max_f16_4,                 has_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SOFT_MAX_F32,                  soft_max_f32,                   has_simdgroup_reduction);
@@ -970,7 +963,6 @@ static bool ggml_metal_supports_op(const struct ggml_backend_metal_device_contex
                 case GGML_UNARY_OP_GELU:
                 case GGML_UNARY_OP_GELU_QUICK:
                 case GGML_UNARY_OP_SILU:
-                case GGML_UNARY_OP_ELU:
                     return ggml_is_contiguous(op->src[0]);
                 default:
                     return false;
@@ -1197,39 +1189,35 @@ static void ggml_metal_encode_node(
 
                 const int32_t dim = ((const int32_t *) dst->op_params)[0];
 
-                ggml_metal_kargs_concat args = {
-                    /*.ne00 =*/ ne00,
-                    /*.ne01 =*/ ne01,
-                    /*.ne02 =*/ ne02,
-                    /*.ne03 =*/ ne03,
-                    /*.nb00 =*/ nb00,
-                    /*.nb01 =*/ nb01,
-                    /*.nb02 =*/ nb02,
-                    /*.nb03 =*/ nb03,
-                    /*.ne10 =*/ ne10,
-                    /*.ne11 =*/ ne11,
-                    /*.ne12 =*/ ne12,
-                    /*.ne13 =*/ ne13,
-                    /*.nb10 =*/ nb10,
-                    /*.nb11 =*/ nb11,
-                    /*.nb12 =*/ nb12,
-                    /*.nb13 =*/ nb13,
-                    /*.ne0  =*/ ne0,
-                    /*.ne1  =*/ ne1,
-                    /*.ne2  =*/ ne2,
-                    /*.ne3  =*/ ne3,
-                    /*.nb0  =*/ nb0,
-                    /*.nb1  =*/ nb1,
-                    /*.nb2  =*/ nb2,
-                    /*.nb3  =*/ nb3,
-                    /*.dim  =*/ dim,
-                };
-
                 [encoder setComputePipelineState:pipeline];
-                [encoder setBytes:&args length:sizeof(args) atIndex:0];
+                [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
-                [encoder setBuffer:id_src0 offset:offs_src0 atIndex:1];
+                [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
-                [encoder setBuffer:id_src1 offset:offs_src1 atIndex:2];
+                [encoder setBuffer:id_dst  offset:offs_dst  atIndex:2];
-                [encoder setBuffer:id_dst  offset:offs_dst  atIndex:3];
+                [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:3];
+                [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:4];
+                [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:5];
+                [encoder setBytes:&ne03 length:sizeof(ne03) atIndex:6];
+                [encoder setBytes:&nb00 length:sizeof(nb00) atIndex:7];
+                [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:8];
+                [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:9];
+                [encoder setBytes:&nb03 length:sizeof(nb03) atIndex:10];
+                [encoder setBytes:&ne10 length:sizeof(ne10) atIndex:11];
+                [encoder setBytes:&ne11 length:sizeof(ne11) atIndex:12];
+                [encoder setBytes:&ne12 length:sizeof(ne12) atIndex:13];
+                [encoder setBytes:&ne13 length:sizeof(ne13) atIndex:14];
+                [encoder setBytes:&nb10 length:sizeof(nb10) atIndex:15];
+                [encoder setBytes:&nb11 length:sizeof(nb11) atIndex:16];
+                [encoder setBytes:&nb12 length:sizeof(nb12) atIndex:17];
+                [encoder setBytes:&nb13 length:sizeof(nb13) atIndex:18];
+                [encoder setBytes:&ne0  length:sizeof(ne0)  atIndex:19];
+                [encoder setBytes:&ne1  length:sizeof(ne1)  atIndex:20];
+                [encoder setBytes:&ne2  length:sizeof(ne2)  atIndex:21];
+                [encoder setBytes:&ne3  length:sizeof(ne3)  atIndex:22];
+                [encoder setBytes:&nb0  length:sizeof(nb0)  atIndex:23];
+                [encoder setBytes:&nb1  length:sizeof(nb1)  atIndex:24];
+                [encoder setBytes:&nb2  length:sizeof(nb2)  atIndex:25];
+                [encoder setBytes:&nb3  length:sizeof(nb3)  atIndex:26];
+                [encoder setBytes:&dim  length:sizeof(dim)  atIndex:27];
 
                 const int nth = MIN(1024, ne0);
 
@@ -1247,6 +1235,8 @@ static void ggml_metal_encode_node(
 
                 bool bcast_row = false;
 
+                int64_t nb = ne00; // used by the "row" kernels
+
                 id<MTLComputePipelineState> pipeline = nil;
 
                 if (ggml_nelements(src1) == ne10 && ggml_is_contiguous(src1) && ne00 % 4 == 0 && ne10 % 4 == 0) {
@@ -1255,6 +1245,7 @@ static void ggml_metal_encode_node(
                     // src1 is a row
                     GGML_ASSERT(ne11 == 1);
 
+                    nb = ne00 / 4;
                     switch (dst->op) {
                         case GGML_OP_ADD: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ADD_ROW].pipeline; break;
                         case GGML_OP_SUB: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SUB_ROW].pipeline; break;
@@ -1274,39 +1265,36 @@ static void ggml_metal_encode_node(
                     }
                 }
 
-                ggml_metal_kargs_bin args = {
-                    /*.ne00 =*/ ne00,
-                    /*.ne01 =*/ ne01,
-                    /*.ne02 =*/ ne02,
-                    /*.ne03 =*/ ne03,
-                    /*.nb00 =*/ nb00,
-                    /*.nb01 =*/ nb01,
-                    /*.nb02 =*/ nb02,
-                    /*.nb03 =*/ nb03,
-                    /*.ne10 =*/ ne10,
-                    /*.ne11 =*/ ne11,
-                    /*.ne12 =*/ ne12,
-                    /*.ne13 =*/ ne13,
-                    /*.nb10 =*/ nb10,
-                    /*.nb11 =*/ nb11,
-                    /*.nb12 =*/ nb12,
-                    /*.nb13 =*/ nb13,
-                    /*.ne0  =*/ ne0,
-                    /*.ne1  =*/ ne1,
-                    /*.ne2  =*/ ne2,
-                    /*.ne3  =*/ ne3,
-                    /*.nb0  =*/ nb0,
-                    /*.nb1  =*/ nb1,
-                    /*.nb2  =*/ nb2,
-                    /*.nb3  =*/ nb3,
-                    /*.offs =*/ offs,
-                };
-
                 [encoder setComputePipelineState:pipeline];
-                [encoder setBytes:&args length:sizeof(args) atIndex:0];
+                [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
-                [encoder setBuffer:id_src0 offset:offs_src0 atIndex:1];
+                [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
-                [encoder setBuffer:id_src1 offset:offs_src1 atIndex:2];
+                [encoder setBuffer:id_dst  offset:offs_dst  atIndex:2];
-                [encoder setBuffer:id_dst  offset:offs_dst  atIndex:3];
+                [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:3];
+                [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:4];
+                [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:5];
+                [encoder setBytes:&ne03 length:sizeof(ne03) atIndex:6];
+                [encoder setBytes:&nb00 length:sizeof(nb00) atIndex:7];
+                [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:8];
+                [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:9];
+                [encoder setBytes:&nb03 length:sizeof(nb03) atIndex:10];
+                [encoder setBytes:&ne10 length:sizeof(ne10) atIndex:11];
+                [encoder setBytes:&ne11 length:sizeof(ne11) atIndex:12];
+                [encoder setBytes:&ne12 length:sizeof(ne12) atIndex:13];
+                [encoder setBytes:&ne13 length:sizeof(ne13) atIndex:14];
+                [encoder setBytes:&nb10 length:sizeof(nb10) atIndex:15];
+                [encoder setBytes:&nb11 length:sizeof(nb11) atIndex:16];
+                [encoder setBytes:&nb12 length:sizeof(nb12) atIndex:17];
+                [encoder setBytes:&nb13 length:sizeof(nb13) atIndex:18];
+                [encoder setBytes:&ne0  length:sizeof(ne0)  atIndex:19];
+                [encoder setBytes:&ne1  length:sizeof(ne1)  atIndex:20];
+                [encoder setBytes:&ne2  length:sizeof(ne2)  atIndex:21];
+                [encoder setBytes:&ne3  length:sizeof(ne3)  atIndex:22];
+                [encoder setBytes:&nb0  length:sizeof(nb0)  atIndex:23];
+                [encoder setBytes:&nb1  length:sizeof(nb1)  atIndex:24];
+                [encoder setBytes:&nb2  length:sizeof(nb2)  atIndex:25];
+                [encoder setBytes:&nb3  length:sizeof(nb3)  atIndex:26];
+                [encoder setBytes:&offs length:sizeof(offs) atIndex:27];
+                [encoder setBytes:&nb   length:sizeof(nb)   atIndex:28];
 
                 if (bcast_row) {
                     const int64_t n = ggml_nelements(dst)/4;
@@ -1330,29 +1318,25 @@ static void ggml_metal_encode_node(
                     default: GGML_ABORT("fatal error");
                 }
 
-                ggml_metal_kargs_repeat args = {
-                    /*.ne00 =*/ ne00,
-                    /*.ne01 =*/ ne01,
-                    /*.ne02 =*/ ne02,
-                    /*.ne03 =*/ ne03,
-                    /*.nb00 =*/ nb00,
-                    /*.nb01 =*/ nb01,
-                    /*.nb02 =*/ nb02,
-                    /*.nb03 =*/ nb03,
-                    /*.ne0  =*/ ne0,
-                    /*.ne1  =*/ ne1,
-                    /*.ne2  =*/ ne2,
-                    /*.ne3  =*/ ne3,
-                    /*.nb0  =*/ nb0,
-                    /*.nb1  =*/ nb1,
-                    /*.nb2  =*/ nb2,
-                    /*.nb3  =*/ nb3,
-                };
-
                 [encoder setComputePipelineState:pipeline];
-                [encoder setBytes:&args length:sizeof(args) atIndex:0];
+                [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
-                [encoder setBuffer:id_src0 offset:offs_src0 atIndex:1];
+                [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
-                [encoder setBuffer:id_dst  offset:offs_dst  atIndex:2];
+                [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:2];
+                [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:3];
+                [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:4];
+                [encoder setBytes:&ne03 length:sizeof(ne03) atIndex:5];
+                [encoder setBytes:&nb00 length:sizeof(nb00) atIndex:6];
+                [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:7];
+                [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:8];
+                [encoder setBytes:&nb03 length:sizeof(nb03) atIndex:9];
+                [encoder setBytes:&ne0  length:sizeof(ne0)  atIndex:10];
+                [encoder setBytes:&ne1  length:sizeof(ne1)  atIndex:11];
+                [encoder setBytes:&ne2  length:sizeof(ne2)  atIndex:12];
+                [encoder setBytes:&ne3  length:sizeof(ne3)  atIndex:13];
+                [encoder setBytes:&nb0  length:sizeof(nb0)  atIndex:14];
+                [encoder setBytes:&nb1  length:sizeof(nb1)  atIndex:15];
+                [encoder setBytes:&nb2  length:sizeof(nb2)  atIndex:16];
+                [encoder setBytes:&nb3  length:sizeof(nb3)  atIndex:17];
 
                 const int nth = MIN((int) pipeline.maxTotalThreadsPerThreadgroup, ne0);
 
@@ -1381,29 +1365,25 @@ static void ggml_metal_encode_node(
 
                     const id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_F32_F32].pipeline;
 
-                    ggml_metal_kargs_cpy args = {
-                        /*.ne00 =*/ ne00,
-                        /*.ne01 =*/ ne01,
-                        /*.ne02 =*/ ne02,
-                        /*.ne03 =*/ ne03,
-                        /*.nb00 =*/ nb00,
-                        /*.nb01 =*/ nb01,
-                        /*.nb02 =*/ nb02,
-                        /*.nb03 =*/ nb03,
-                        /*.ne0  =*/ ne0,
-                        /*.ne1  =*/ ne1,
-                        /*.ne2  =*/ ne2,
-                        /*.ne3  =*/ ne3,
-                        /*.nb0  =*/ nb0,
-                        /*.nb1  =*/ nb1,
-                        /*.nb2  =*/ nb2,
-                        /*.nb3  =*/ nb3,
-                    };
-
                     [encoder setComputePipelineState:pipeline];
-                    [encoder setBytes:&args length:sizeof(args) atIndex:0];
+                    [encoder setBuffer:id_src0 offset:offs_src0        atIndex:0];
-                    [encoder setBuffer:id_src0 offset:offs_src0 atIndex:1];
+                    [encoder setBuffer:id_dst  offset:offs_dst         atIndex:1];
-                    [encoder setBuffer:id_dst  offset:offs_dst  atIndex:2];
+                    [encoder setBytes:&ne00    length:sizeof( int64_t) atIndex:2];
+                    [encoder setBytes:&ne01    length:sizeof( int64_t) atIndex:3];
+                    [encoder setBytes:&ne02    length:sizeof( int64_t) atIndex:4];
+                    [encoder setBytes:&ne03    length:sizeof( int64_t) atIndex:5];
+                    [encoder setBytes:&nb00    length:sizeof(uint64_t) atIndex:6];
+                    [encoder setBytes:&nb01    length:sizeof(uint64_t) atIndex:7];
+                    [encoder setBytes:&nb02    length:sizeof(uint64_t) atIndex:8];
+                    [encoder setBytes:&nb03    length:sizeof(uint64_t) atIndex:9];
+                    [encoder setBytes:&ne0     length:sizeof( int64_t) atIndex:10];
+                    [encoder setBytes:&ne1     length:sizeof( int64_t) atIndex:11];
+                    [encoder setBytes:&ne2     length:sizeof( int64_t) atIndex:12];
+                    [encoder setBytes:&ne3     length:sizeof( int64_t) atIndex:13];
+                    [encoder setBytes:&nb0     length:sizeof(uint64_t) atIndex:14];
+                    [encoder setBytes:&nb1     length:sizeof(uint64_t) atIndex:15];
+                    [encoder setBytes:&nb2     length:sizeof(uint64_t) atIndex:16];
+                    [encoder setBytes:&nb3     length:sizeof(uint64_t) atIndex:17];
 
                     const int nth = MIN((int) pipeline.maxTotalThreadsPerThreadgroup, ne00);
 
@@ -1412,39 +1392,35 @@ static void ggml_metal_encode_node(
 
                 const id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ADD].pipeline;
 
-                ggml_metal_kargs_bin args = {
-                    /*.ne00 =*/ ne00,
-                    /*.ne01 =*/ ne01,
-                    /*.ne02 =*/ ne02,
-                    /*.ne03 =*/ ne03,
-                    /*.nb00 =*/ nb00,
-                    /*.nb01 =*/ pnb1,
-                    /*.nb02 =*/ pnb2,
-                    /*.nb03 =*/ pnb3,
-                    /*.ne10 =*/ ne10,
-                    /*.ne11 =*/ ne11,
-                    /*.ne12 =*/ ne12,
-                    /*.ne13 =*/ ne13,
-                    /*.nb10 =*/ nb10,
-                    /*.nb11 =*/ nb11,
-                    /*.nb12 =*/ nb12,
-                    /*.nb13 =*/ nb13,
-                    /*.ne0  =*/ ne0,
-                    /*.ne1  =*/ ne1,
-                    /*.ne2  =*/ ne2,
-                    /*.ne3  =*/ ne3,
-                    /*.nb0  =*/ nb0,
-                    /*.nb1  =*/ pnb1,
-                    /*.nb2  =*/ pnb2,
-                    /*.nb3  =*/ pnb3,
-                    /*.offs =*/ offs,
-                };
-
                 [encoder setComputePipelineState:pipeline];
-                [encoder setBytes:&args length:sizeof(args) atIndex:0];
+                [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
-                [encoder setBuffer:id_src0 offset:offs_src0 atIndex:1];
+                [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
-                [encoder setBuffer:id_src1 offset:offs_src1 atIndex:2];
+                [encoder setBuffer:id_dst  offset:offs_dst  atIndex:2];
-                [encoder setBuffer:id_dst  offset:offs_dst  atIndex:3];
+                [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:3];
+                [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:4];
+                [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:5];
+                [encoder setBytes:&ne03 length:sizeof(ne03) atIndex:6];
+                [encoder setBytes:&nb00 length:sizeof(nb00) atIndex:7];
+                [encoder setBytes:&pnb1 length:sizeof(pnb1) atIndex:8];
+                [encoder setBytes:&pnb2 length:sizeof(pnb2) atIndex:9];
+                [encoder setBytes:&pnb3 length:sizeof(pnb3) atIndex:10];
+                [encoder setBytes:&ne10 length:sizeof(ne10) atIndex:11];
+                [encoder setBytes:&ne11 length:sizeof(ne11) atIndex:12];
+                [encoder setBytes:&ne12 length:sizeof(ne12) atIndex:13];
+                [encoder setBytes:&ne13 length:sizeof(ne13) atIndex:14];
+                [encoder setBytes:&nb10 length:sizeof(nb10) atIndex:15];
+                [encoder setBytes:&nb11 length:sizeof(nb11) atIndex:16];
+                [encoder setBytes:&nb12 length:sizeof(nb12) atIndex:17];
+                [encoder setBytes:&nb13 length:sizeof(nb13) atIndex:18];
+                [encoder setBytes:&ne0  length:sizeof(ne0)  atIndex:19];
+                [encoder setBytes:&ne1  length:sizeof(ne1)  atIndex:20];
+                [encoder setBytes:&ne2  length:sizeof(ne2)  atIndex:21];
+                [encoder setBytes:&ne3  length:sizeof(ne3)  atIndex:22];
+                [encoder setBytes:&nb0  length:sizeof(nb0)  atIndex:23];
+                [encoder setBytes:&pnb1 length:sizeof(pnb1) atIndex:24];
+                [encoder setBytes:&pnb2 length:sizeof(pnb2) atIndex:25];
+                [encoder setBytes:&pnb3 length:sizeof(pnb3) atIndex:26];
+                [encoder setBytes:&offs length:sizeof(offs) atIndex:27];
 
                 const int nth = MIN((int) pipeline.maxTotalThreadsPerThreadgroup, ne00);
 
@@ -1485,10 +1461,10 @@ static void ggml_metal_encode_node(
                 memcpy(&max, ((const int32_t *) dst->op_params) + 1, sizeof(float));
 
                 [encoder setComputePipelineState:pipeline];
-                [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                [encoder setBuffer:id_src0   offset:offs_src0 atIndex:0];
-                [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
+                [encoder setBuffer:id_dst    offset:offs_dst  atIndex:1];
-                [encoder setBytes:&min   length:sizeof(min) atIndex:2];
+                [encoder setBytes:&min length:sizeof(min) atIndex:2];
-                [encoder setBytes:&max   length:sizeof(max) atIndex:3];
+                [encoder setBytes:&max length:sizeof(max) atIndex:3];
 
                 const int64_t n = ggml_nelements(dst);
 
@@ -1592,18 +1568,6 @@ static void ggml_metal_encode_node(
 
                     [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
                 } break;
-                case GGML_UNARY_OP_ELU:
-                {
-                    id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ELU].pipeline;
-
-                    [encoder setComputePipelineState:pipeline];
-                    [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
-                    [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
-
-                    const int64_t n = ggml_nelements(dst);
-
-                    [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
-                } break;
                 default:
                 {
                     GGML_LOG_WARN("%s: node %3d, op = %8s not implemented\n", __func__, idx, ggml_op_name(dst->op));
@@ -1672,7 +1636,6 @@ static void ggml_metal_encode_node(
 
                 id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SUM_ROWS].pipeline;
 
-                // TODO: add ggml_metal_kargs struct
                 [encoder setComputePipelineState:pipeline];
                 [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
                 [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
@@ -1748,8 +1711,6 @@ static void ggml_metal_encode_node(
                 const float m0 = powf(2.0f, -(max_bias       ) / n_head_log2);
                 const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2);
 
-                // TODO: add ggml_metal_kargs struct
-                // TODO: optimize (see https://github.com/ggerganov/llama.cpp/pull/10238/commits/7941b6b9ec29a2866fec6fa6c51612515ca509f6)
                 [encoder setComputePipelineState:pipeline];
                 [encoder setBuffer:id_src0 offset:offs_src0   atIndex:0];
                 if (id_src1) {
@@ -1766,7 +1727,6 @@ static void ggml_metal_encode_node(
                 [encoder setBytes:&m0          length:sizeof(m0)          atIndex:8];
                 [encoder setBytes:&m1          length:sizeof(m1)          atIndex:9];
                 [encoder setBytes:&n_head_log2 length:sizeof(n_head_log2) atIndex:10];
-
                 [encoder setThreadgroupMemoryLength:32*sizeof(float) atIndex:0];
 
                 [encoder dispatchThreadgroups:MTLSizeMake(ne01*ne02*ne03, 1, 1) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
@@ -1783,7 +1743,6 @@ static void ggml_metal_encode_node(
                     pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_DIAG_MASK_INF].pipeline;
                 }
 
-                // TODO: add ggml_metal_kargs struct
                 [encoder setComputePipelineState:pipeline];
                 [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
                 [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
@@ -1808,7 +1767,6 @@ static void ggml_metal_encode_node(
 
                 id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SSM_CONV_F32].pipeline;
 
-                // TODO: add ggml_metal_kargs struct
                 [encoder setComputePipelineState:pipeline];
                 [encoder setBuffer:id_src0 offset:offs_src0    atIndex:0];
                 [encoder setBuffer:id_src1 offset:offs_src1    atIndex:1];
@@ -1879,7 +1837,6 @@ static void ggml_metal_encode_node(
 
                 id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SSM_SCAN_F32].pipeline;
 
-                // TODO: add ggml_metal_kargs struct
                 [encoder setComputePipelineState:pipeline];
                 [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
                 [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
@@ -1998,29 +1955,24 @@ static void ggml_metal_encode_node(
                                 default: GGML_ABORT("MUL MAT-MAT not implemented");
                             }
 
-                            ggml_metal_kargs_mul_mm args = {
-                                /*.ne00 =*/ ne00,
-                                /*.ne02 =*/ ne02,
-                                /*.nb01 =*/ nb01,
-                                /*.nb02 =*/ nb02,
-                                /*.nb03 =*/ nb03,
-                                /*.ne12 =*/ ne12,
-                                /*.nb10 =*/ nb10,
-                                /*.nb11 =*/ nb11,
-                                /*.nb12 =*/ nb12,
-                                /*.nb13 =*/ nb13,
-                                /*.ne0  =*/ ne0,
-                                /*.ne1  =*/ ne1,
-                                /*.r2   =*/ r2,
-                                /*.r3   =*/ r3,
-                            };
-
                             [encoder setComputePipelineState:pipeline];
-                            [encoder setBytes:&args    length:sizeof(args) atIndex:0];
+                            [encoder setBuffer:id_src0 offset:offs_src0    atIndex:0];
-                            [encoder setBuffer:id_src0 offset:offs_src0    atIndex:1];
+                            [encoder setBuffer:id_src1 offset:offs_src1    atIndex:1];
-                            [encoder setBuffer:id_src1 offset:offs_src1    atIndex:2];
+                            [encoder setBuffer:id_dst  offset:offs_dst     atIndex:2];
-                            [encoder setBuffer:id_dst  offset:offs_dst     atIndex:3];
+                            [encoder setBytes:&ne00    length:sizeof(ne00) atIndex:3];
-
+                            [encoder setBytes:&ne02    length:sizeof(ne02) atIndex:4];
+                            [encoder setBytes:&nb01    length:sizeof(nb01) atIndex:5];
+                            [encoder setBytes:&nb02    length:sizeof(nb02) atIndex:6];
+                            [encoder setBytes:&nb03    length:sizeof(nb03) atIndex:7];
+                            [encoder setBytes:&ne12    length:sizeof(ne12) atIndex:8];
+                            [encoder setBytes:&nb10    length:sizeof(nb10) atIndex:9];
+                            [encoder setBytes:&nb11    length:sizeof(nb11) atIndex:10];
+                            [encoder setBytes:&nb12    length:sizeof(nb12) atIndex:11];
+                            [encoder setBytes:&nb13    length:sizeof(nb13) atIndex:12];
+                            [encoder setBytes:&ne0     length:sizeof(ne0)  atIndex:13];
+                            [encoder setBytes:&ne1     length:sizeof(ne1)  atIndex:14];
+                            [encoder setBytes:&r2      length:sizeof(r2)   atIndex:15];
+                            [encoder setBytes:&r3      length:sizeof(r3)   atIndex:16];
                             [encoder setThreadgroupMemoryLength:8192 atIndex:0];
                             [encoder dispatchThreadgroups:MTLSizeMake( (ne11 + 31)/32, (ne01 + 63)/64, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(128, 1, 1)];
                         } else {
@@ -2198,32 +2150,28 @@ static void ggml_metal_encode_node(
                                     }
                             };
 
-                            ggml_metal_kargs_mul_mv args = {
-                                /*.ne00 =*/ ne00,
-                                /*.ne01 =*/ ne01,
-                                /*.ne02 =*/ ne02,
-                                /*.nb00 =*/ nb00,
-                                /*.nb01 =*/ nb01,
-                                /*.nb02 =*/ nb02,
-                                /*.nb03 =*/ nb03,
-                                /*.ne10 =*/ ne10,
-                                /*.ne11 =*/ ne11,
-                                /*.ne12 =*/ ne12,
-                                /*.nb10 =*/ nb10,
-                                /*.nb11 =*/ nb11,
-                                /*.nb12 =*/ nb12,
-                                /*.nb13 =*/ nb13,
-                                /*.ne0  =*/ ne0,
-                                /*.ne1  =*/ ne1,
-                                /*.r2   =*/ r2,
-                                /*.r3   =*/ r3,
-                            };
-
                             [encoder setComputePipelineState:pipeline];
-                            [encoder setBytes:&args length:sizeof(args) atIndex:0];
+                            [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
-                            [encoder setBuffer:id_src0 offset:offs_src0 atIndex:1];
+                            [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
-                            [encoder setBuffer:id_src1 offset:offs_src1 atIndex:2];
+                            [encoder setBuffer:id_dst  offset:offs_dst  atIndex:2];
-                            [encoder setBuffer:id_dst  offset:offs_dst  atIndex:3];
+                            [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:3];
+                            [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:4];
+                            [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:5];
+                            [encoder setBytes:&nb00 length:sizeof(nb00) atIndex:6];
+                            [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:7];
+                            [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:8];
+                            [encoder setBytes:&nb03 length:sizeof(nb03) atIndex:9];
+                            [encoder setBytes:&ne10 length:sizeof(ne10) atIndex:10];
+                            [encoder setBytes:&ne11 length:sizeof(ne11) atIndex:11];
+                            [encoder setBytes:&ne12 length:sizeof(ne12) atIndex:12];
+                            [encoder setBytes:&nb10 length:sizeof(nb10) atIndex:13];
+                            [encoder setBytes:&nb11 length:sizeof(nb11) atIndex:14];
+                            [encoder setBytes:&nb12 length:sizeof(nb12) atIndex:15];
+                            [encoder setBytes:&nb13 length:sizeof(nb13) atIndex:16];
+                            [encoder setBytes:&ne0  length:sizeof(ne0)  atIndex:17];
+                            [encoder setBytes:&ne1  length:sizeof(ne1)  atIndex:18];
+                            [encoder setBytes:&r2   length:sizeof(r2)   atIndex:19];
+                            [encoder setBytes:&r3   length:sizeof(r3)   atIndex:20];
 
                             if (src0t == GGML_TYPE_Q4_0  || src0t == GGML_TYPE_Q4_1  || src0t == GGML_TYPE_Q5_0 ||
                                 src0t == GGML_TYPE_Q5_1  || src0t == GGML_TYPE_Q8_0  || src0t == GGML_TYPE_Q2_K ||
@@ -2336,30 +2284,27 @@ static void ggml_metal_encode_node(
                         default: GGML_ABORT("MUL_MAT_ID not implemented");
                     }
 
-                    ggml_metal_kargs_mul_mm_id args = {
-                        /*.nei0 =*/ ne20,
-                        /*.nei1 =*/ ne21,
-                        /*.nbi1 =*/ nb21,
-                        /*.ne00 =*/ ne00,
-                        /*.ne02 =*/ ne02,
-                        /*.nb01 =*/ nb01,
-                        /*.nb02 =*/ nb02,
-                        /*.ne11 =*/ ne11,
-                        /*.ne12 =*/ ne12,
-                        /*.ne13 =*/ ne13,
-                        /*.nb10 =*/ nb10,
-                        /*.nb11 =*/ nb11,
-                        /*.nb12 =*/ nb12,
-                        /*.ne0  =*/ ne0,
-                        /*.ne1  =*/ ne1,
-                    };
-
                     [encoder setComputePipelineState:pipeline];
-                    [encoder setBytes:&args    length:sizeof(args) atIndex:0];
+                    [encoder setBuffer:id_src0 offset:offs_src0    atIndex:0];
-                    [encoder setBuffer:id_src0 offset:offs_src0    atIndex:1];
+                    [encoder setBuffer:id_src1 offset:offs_src1    atIndex:1];
-                    [encoder setBuffer:id_src1 offset:offs_src1    atIndex:2];
+                    [encoder setBuffer:id_dst  offset:offs_dst     atIndex:2];
-                    [encoder setBuffer:id_dst  offset:offs_dst     atIndex:3];
+                    [encoder setBuffer:id_src2 offset:offs_src2    atIndex:3];
-                    [encoder setBuffer:id_src2 offset:offs_src2    atIndex:4];
+                    [encoder setBytes:&ne20    length:sizeof(ne20) atIndex:4];
+                    [encoder setBytes:&ne21    length:sizeof(ne21) atIndex:5];
+                    [encoder setBytes:&nb21    length:sizeof(nb21) atIndex:6];
+                    [encoder setBytes:&ne00    length:sizeof(ne00) atIndex:7];
+                    [encoder setBytes:&ne02    length:sizeof(ne02) atIndex:8];
+                    [encoder setBytes:&nb01    length:sizeof(nb01) atIndex:9];
+                    [encoder setBytes:&nb02    length:sizeof(nb02) atIndex:10];
+                    [encoder setBytes:&ne11    length:sizeof(ne11) atIndex:11];
+                    [encoder setBytes:&ne12    length:sizeof(ne12) atIndex:12];
+                    [encoder setBytes:&ne13    length:sizeof(ne13) atIndex:13];
+                    [encoder setBytes:&nb10    length:sizeof(nb10) atIndex:14];
+                    [encoder setBytes:&nb11    length:sizeof(nb11) atIndex:15];
+                    [encoder setBytes:&nb12    length:sizeof(nb12) atIndex:16];
+                    [encoder setBytes:&ne0     length:sizeof(ne0)  atIndex:17];
+                    [encoder setBytes:&ne1     length:sizeof(ne1)  atIndex:18];
+                    [encoder setBytes:&nb1     length:sizeof(nb1)  atIndex:19];
 
                     [encoder setThreadgroupMemoryLength:GGML_PAD(8192 + dst_rows*4/*sizeof(ushort2)*/, 16) atIndex:0];
 
@@ -2518,34 +2463,30 @@ static void ggml_metal_encode_node(
                         GGML_ASSERT(ne00 >= nth0*nth1);
                     }
 
-                    ggml_metal_kargs_mul_mv_id args = {
-                        /*.nei0 =*/ ne20,
-                        /*.nei1 =*/ ne21,
-                        /*.nbi1 =*/ nb21,
-                        /*.ne00 =*/ ne00,
-                        /*.ne01 =*/ ne01,
-                        /*.ne02 =*/ ne02,
-                        /*.nb00 =*/ nb00,
-                        /*.nb01 =*/ nb01,
-                        /*.nb02 =*/ nb02,
-                        /*.ne10 =*/ ne10,
-                        /*.ne11 =*/ ne11,
-                        /*.ne12 =*/ ne12,
-                        /*.ne13 =*/ ne13,
-                        /*.nb10 =*/ nb10,
-                        /*.nb11 =*/ nb11,
-                        /*.nb12 =*/ nb12,
-                        /*.ne0  =*/ ne0,
-                        /*.ne1  =*/ ne1,
-                        /*.nb1  =*/ nb1,
-                    };
-
                     [encoder setComputePipelineState:pipeline];
-                    [encoder setBytes:&args length:sizeof(args) atIndex:0];
+                    [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
-                    [encoder setBuffer:id_src0 offset:offs_src0 atIndex:1];
+                    [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
-                    [encoder setBuffer:id_src1 offset:offs_src1 atIndex:2];
+                    [encoder setBuffer:id_dst  offset:offs_dst  atIndex:2];
-                    [encoder setBuffer:id_dst  offset:offs_dst  atIndex:3];
+                    [encoder setBuffer:id_src2 offset:offs_src2 atIndex:3];
-                    [encoder setBuffer:id_src2 offset:offs_src2 atIndex:4];
+                    [encoder setBytes:&ne20 length:sizeof(ne20) atIndex:4];
+                    [encoder setBytes:&ne21 length:sizeof(ne21) atIndex:5];
+                    [encoder setBytes:&nb21 length:sizeof(nb21) atIndex:6];
+                    [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:7];
+                    [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:8];
+                    [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:9];
+                    [encoder setBytes:&nb00 length:sizeof(nb00) atIndex:10];
+                    [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:11];
+                    [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:12];
+                    [encoder setBytes:&ne10 length:sizeof(ne10) atIndex:13];
+                    [encoder setBytes:&ne11 length:sizeof(ne11) atIndex:14];
+                    [encoder setBytes:&ne12 length:sizeof(ne12) atIndex:15];
+                    [encoder setBytes:&ne13 length:sizeof(ne13) atIndex:16];
+                    [encoder setBytes:&nb10 length:sizeof(nb10) atIndex:17];
+                    [encoder setBytes:&nb11 length:sizeof(nb11) atIndex:18];
+                    [encoder setBytes:&nb12 length:sizeof(nb12) atIndex:19];
+                    [encoder setBytes:&ne0  length:sizeof(ne0)  atIndex:20];
+                    [encoder setBytes:&ne1  length:sizeof(ne1)  atIndex:21];
+                    [encoder setBytes:&nb1  length:sizeof(nb1)  atIndex:22];
 
                     const int64_t _ne1 = 1;
                     const int tgz = dst_rows;
@@ -2618,7 +2559,6 @@ static void ggml_metal_encode_node(
                     default: GGML_ABORT("not implemented");
                 }
 
-                // TODO: add ggml_metal_kargs struct
                 [encoder setComputePipelineState:pipeline];
                 [encoder setBuffer:id_src0     offset:offs_src0 atIndex:0];
                 [encoder setBuffer:id_src1     offset:offs_src1 atIndex:1];
@@ -2642,28 +2582,20 @@ static void ggml_metal_encode_node(
                 float eps;
                 memcpy(&eps, dst->op_params, sizeof(float));
 
-                id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_RMS_NORM].pipeline;
-
                 int nth = 32; // SIMD width
 
-                while (nth < ne00/4 && nth < (int) pipeline.maxTotalThreadsPerThreadgroup) {
+                while (nth < ne00/4 && nth < 1024) {
                     nth *= 2;
                 }
 
-                nth = MIN(nth, ne00/4);
+                id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_RMS_NORM].pipeline;
-
-                ggml_metal_kargs_rms_norm args = {
-                    /*.ne00   =*/ ne00,
-                    /*.ne00_4 =*/ ne00/4,
-                    /*.nb01   =*/ nb01,
-                    /*.eps    =*/ eps,
-                };
 
                 [encoder setComputePipelineState:pipeline];
-                [encoder setBytes:&args length:sizeof(args) atIndex:0];
+                [encoder setBuffer:id_src0 offset:offs_src0        atIndex:0];
-                [encoder setBuffer:id_src0 offset:offs_src0 atIndex:1];
+                [encoder setBuffer:id_dst  offset:offs_dst         atIndex:1];
-                [encoder setBuffer:id_dst  offset:offs_dst  atIndex:2];
+                [encoder setBytes:&ne00    length:sizeof( int64_t) atIndex:2];
-
+                [encoder setBytes:&nb01    length:sizeof(uint64_t) atIndex:3];
+                [encoder setBytes:&eps     length:sizeof(   float) atIndex:4];
                 [encoder setThreadgroupMemoryLength:32*sizeof(float) atIndex:0];
 
                 const int64_t nrows = ggml_nrows(src0);
@@ -2688,7 +2620,6 @@ static void ggml_metal_encode_node(
 
                 id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GROUP_NORM].pipeline;
 
-                // TODO: add ggml_metal_kargs struct
                 [encoder setComputePipelineState:pipeline];
                 [encoder setBuffer:id_src0  offset:offs_src0        atIndex:0];
                 [encoder setBuffer:id_dst   offset:offs_dst         atIndex:1];
@@ -2706,35 +2637,22 @@ static void ggml_metal_encode_node(
             } break;
         case GGML_OP_NORM:
             {
-                GGML_ASSERT(ne00 % 4 == 0);
                 GGML_ASSERT(ggml_is_contiguous_1(src0));
 
                 float eps;
                 memcpy(&eps, dst->op_params, sizeof(float));
 
+                const int nth = MIN(256, ne00);
+
                 id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_NORM].pipeline;
 
-                int nth = 32; // SIMD width
-
-                while (nth < ne00/4 && nth < (int) pipeline.maxTotalThreadsPerThreadgroup) {
-                    nth *= 2;
-                }
-
-                nth = MIN(nth, ne00/4);
-
-                ggml_metal_kargs_norm args = {
-                    /*.ne00   =*/ ne00,
-                    /*.ne00_4 =*/ ne00/4,
-                    /*.nb01   =*/ nb01,
-                    /*.eps    =*/ eps,
-                };
-
                 [encoder setComputePipelineState:pipeline];
-                [encoder setBytes:&args length:sizeof(args) atIndex:0];
+                [encoder setBuffer:id_src0 offset:offs_src0        atIndex:0];
-                [encoder setBuffer:id_src0 offset:offs_src0 atIndex:1];
+                [encoder setBuffer:id_dst  offset:offs_dst         atIndex:1];
-                [encoder setBuffer:id_dst  offset:offs_dst  atIndex:2];
+                [encoder setBytes:&ne00    length:sizeof( int64_t) atIndex:2];
-
+                [encoder setBytes:&nb01    length:sizeof(uint64_t) atIndex:3];
-                [encoder setThreadgroupMemoryLength:32*sizeof(float) atIndex:0];
+                [encoder setBytes:&eps     length:sizeof(   float) atIndex:4];
+                [encoder setThreadgroupMemoryLength:GGML_PAD(nth*sizeof(float), 16) atIndex:0];
 
                 const int64_t nrows = ggml_nrows(src0);
 
@@ -2784,44 +2702,40 @@ static void ggml_metal_encode_node(
                     };
                 }
 
-                ggml_metal_kargs_rope args = {
-                    /*.ne00        =*/ ne00,
-                    /*.ne01        =*/ ne01,
-                    /*.ne02        =*/ ne02,
-                    /*.ne03        =*/ ne03,
-                    /*.nb00        =*/ nb00,
-                    /*.nb01        =*/ nb01,
-                    /*.nb02        =*/ nb02,
-                    /*.nb03        =*/ nb03,
-                    /*.ne0         =*/ ne0,
-                    /*.ne1         =*/ ne1,
-                    /*.ne2         =*/ ne2,
-                    /*.ne3         =*/ ne3,
-                    /*.nb0         =*/ nb0,
-                    /*.nb1         =*/ nb1,
-                    /*.nb2         =*/ nb2,
-                    /*.nb3         =*/ nb3,
-                    /*.n_past      =*/ n_past,
-                    /*.n_dims      =*/ n_dims,
-                    /*.n_ctx_orig  =*/ n_ctx_orig,
-                    /*.freq_base   =*/ freq_base,
-                    /*.freq_scale  =*/ freq_scale,
-                    /*.ext_factor  =*/ ext_factor,
-                    /*.attn_factor =*/ attn_factor,
-                    /*.beta_fast   =*/ beta_fast,
-                    /*.beta_slow   =*/ beta_slow,
-                };
-
                 [encoder setComputePipelineState:pipeline];
-                [encoder setBytes:&args length:sizeof(args)     atIndex:0];
+                [encoder setBuffer:id_src0     offset:offs_src0        atIndex:0];
-                [encoder setBuffer:id_src0 offset:offs_src0     atIndex:1];
+                [encoder setBuffer:id_src1     offset:offs_src1        atIndex:1];
-                [encoder setBuffer:id_src1 offset:offs_src1     atIndex:2];
                 if (id_src2 != nil) {
-                    [encoder setBuffer:id_src2 offset:offs_src2 atIndex:3];
+                    [encoder setBuffer:id_src2 offset:offs_src2        atIndex:2];
                 } else {
-                    [encoder setBuffer:id_src0 offset:offs_src0 atIndex:3];
+                    [encoder setBuffer:id_src0 offset:offs_src0        atIndex:2];
                 }
-                [encoder setBuffer:id_dst  offset:offs_dst      atIndex:4];
+                [encoder setBuffer:id_dst      offset:offs_dst         atIndex:3];
+                [encoder setBytes:&ne00        length:sizeof( int64_t) atIndex:4];
+                [encoder setBytes:&ne01        length:sizeof( int64_t) atIndex:5];
+                [encoder setBytes:&ne02        length:sizeof( int64_t) atIndex:6];
+                [encoder setBytes:&ne03        length:sizeof( int64_t) atIndex:7];
+                [encoder setBytes:&nb00        length:sizeof(uint64_t) atIndex:8];
+                [encoder setBytes:&nb01        length:sizeof(uint64_t) atIndex:9];
+                [encoder setBytes:&nb02        length:sizeof(uint64_t) atIndex:10];
+                [encoder setBytes:&nb03        length:sizeof(uint64_t) atIndex:11];
+                [encoder setBytes:&ne0         length:sizeof( int64_t) atIndex:12];
+                [encoder setBytes:&ne1         length:sizeof( int64_t) atIndex:13];
+                [encoder setBytes:&ne2         length:sizeof( int64_t) atIndex:14];
+                [encoder setBytes:&ne3         length:sizeof( int64_t) atIndex:15];
+                [encoder setBytes:&nb0         length:sizeof(uint64_t) atIndex:16];
+                [encoder setBytes:&nb1         length:sizeof(uint64_t) atIndex:17];
+                [encoder setBytes:&nb2         length:sizeof(uint64_t) atIndex:18];
+                [encoder setBytes:&nb3         length:sizeof(uint64_t) atIndex:19];
+                [encoder setBytes:&n_past      length:sizeof(     int) atIndex:20];
+                [encoder setBytes:&n_dims      length:sizeof(     int) atIndex:21];
+                [encoder setBytes:&n_ctx_orig  length:sizeof(     int) atIndex:22];
+                [encoder setBytes:&freq_base   length:sizeof(   float) atIndex:23];
+                [encoder setBytes:&freq_scale  length:sizeof(   float) atIndex:24];
+                [encoder setBytes:&ext_factor  length:sizeof(   float) atIndex:25];
+                [encoder setBytes:&attn_factor length:sizeof(   float) atIndex:26];
+                [encoder setBytes:&beta_fast   length:sizeof(   float) atIndex:27];
+                [encoder setBytes:&beta_slow   length:sizeof(   float) atIndex:28];
 
                 [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
             } break;
@@ -2878,7 +2792,6 @@ static void ggml_metal_encode_node(
                     default: GGML_ABORT("fatal error");
                 };
 
-                // TODO: add ggml_metal_kargs struct
                 [encoder setComputePipelineState:pipeline];
                 [encoder setBuffer:id_src1 offset:offs_src1       atIndex:0];
                 [encoder setBuffer:id_dst  offset:offs_dst        atIndex:1];
@@ -2919,7 +2832,6 @@ static void ggml_metal_encode_node(
 
                 const id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_UPSCALE_F32].pipeline;
 
-                // TODO: add ggml_metal_kargs struct
                 [encoder setComputePipelineState:pipeline];
                 [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
                 [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
@@ -2954,7 +2866,6 @@ static void ggml_metal_encode_node(
 
                 id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_PAD_F32].pipeline;
 
-                // TODO: add ggml_metal_kargs struct
                 [encoder setComputePipelineState:pipeline];
                 [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
                 [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
@@ -2991,7 +2902,6 @@ static void ggml_metal_encode_node(
 
                 id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ARANGE_F32].pipeline;
 
-                // TODO: add ggml_metal_kargs struct
                 [encoder setComputePipelineState:pipeline];
                 [encoder setBuffer:id_dst  offset:offs_dst    atIndex:0];
                 [encoder setBytes:&ne0   length:sizeof(ne0)   atIndex:1];
@@ -3013,7 +2923,6 @@ static void ggml_metal_encode_node(
 
                 id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_TIMESTEP_EMBEDDING_F32].pipeline;
 
-                // TODO: add ggml_metal_kargs struct
                 [encoder setComputePipelineState:pipeline];
                 [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
                 [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
@@ -3052,7 +2961,6 @@ static void ggml_metal_encode_node(
                     default: GGML_ABORT("fatal error");
                 };
 
-                // TODO: add ggml_metal_kargs struct
                 [encoder setComputePipelineState:pipeline];
                 [encoder setBuffer:id_src0     offset:offs_src0        atIndex:0];
                 [encoder setBuffer:id_dst      offset:offs_dst         atIndex:1];
@@ -3071,7 +2979,6 @@ static void ggml_metal_encode_node(
 
                 id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_LEAKY_RELU_F32].pipeline;
 
-                // TODO: add ggml_metal_kargs struct
                 [encoder setComputePipelineState:pipeline];
                 [encoder setBuffer:id_src0 offset:offs_src0   atIndex:0];
                 [encoder setBuffer:id_dst  offset:offs_dst    atIndex:1];
@@ -3313,41 +3220,37 @@ static void ggml_metal_encode_node(
                     }
                 }
 
-                ggml_metal_kargs_flash_attn_ext args = {
-                    /*.ne01          =*/ ne01,
-                    /*.ne02          =*/ ne02,
-                    /*.ne03          =*/ ne03,
-                    /*.nb01          =*/ nb01,
-                    /*.nb02          =*/ nb02,
-                    /*.nb03          =*/ nb03,
-                    /*.ne11          =*/ ne11,
-                    /*.ne_12_2       =*/ ne12,
-                    /*.ne_12_3       =*/ ne13,
-                    /*.nb_12_1       =*/ nb11,
-                    /*.nb_12_2       =*/ nb12,
-                    /*.nb_12_3       =*/ nb13,
-                    /*.nb31          =*/ nb31,
-                    /*.ne1           =*/ ne1,
-                    /*.ne2           =*/ ne2,
-                    /*.scale         =*/ scale,
-                    /*.max_bias      =*/ max_bias,
-                    /*.m0            =*/ m0,
-                    /*.m1            =*/ m1,
-                    /*.n_head_log2   =*/ n_head_log2,
-                    /*.logit_softcap =*/ logit_softcap,
-                };
-
                 [encoder setComputePipelineState:pipeline];
-                [encoder setBytes:&args length:sizeof(args)     atIndex:0];
+                [encoder setBuffer:id_src0     offset:offs_src0           atIndex:0];
-                [encoder setBuffer:id_src0 offset:offs_src0     atIndex:1];
+                [encoder setBuffer:id_src1     offset:offs_src1           atIndex:1];
-                [encoder setBuffer:id_src1 offset:offs_src1     atIndex:2];
+                [encoder setBuffer:id_src2     offset:offs_src2           atIndex:2];
-                [encoder setBuffer:id_src2 offset:offs_src2     atIndex:3];
                 if (id_src3) {
-                    [encoder setBuffer:id_src3 offset:offs_src3 atIndex:4];
+                    [encoder setBuffer:id_src3     offset:offs_src3           atIndex:3];
                 } else {
-                    [encoder setBuffer:id_src0 offset:offs_src0 atIndex:4];
+                    [encoder setBuffer:id_src0     offset:offs_src0           atIndex:3];
                 }
-                [encoder setBuffer:id_dst offset:offs_dst       atIndex:5];
+                [encoder setBuffer:id_dst        offset:offs_dst              atIndex:4];
+                [encoder setBytes:&ne01          length:sizeof( int64_t)      atIndex:5];
+                [encoder setBytes:&ne02          length:sizeof( int64_t)      atIndex:6];
+                [encoder setBytes:&ne03          length:sizeof( int64_t)      atIndex:7];
+                [encoder setBytes:&nb01          length:sizeof(uint64_t)      atIndex:8];
+                [encoder setBytes:&nb02          length:sizeof(uint64_t)      atIndex:9];
+                [encoder setBytes:&nb03          length:sizeof(uint64_t)      atIndex:10];
+                [encoder setBytes:&ne11          length:sizeof( int64_t)      atIndex:11];
+                [encoder setBytes:&ne12          length:sizeof( int64_t)      atIndex:12];
+                [encoder setBytes:&ne13          length:sizeof( int64_t)      atIndex:13];
+                [encoder setBytes:&nb11          length:sizeof(uint64_t)      atIndex:14];
+                [encoder setBytes:&nb12          length:sizeof(uint64_t)      atIndex:15];
+                [encoder setBytes:&nb13          length:sizeof(uint64_t)      atIndex:16];
+                [encoder setBytes:&nb31          length:sizeof(uint64_t)      atIndex:17];
+                [encoder setBytes:&ne1           length:sizeof( int64_t)      atIndex:18];
+                [encoder setBytes:&ne2           length:sizeof( int64_t)      atIndex:19];
+                [encoder setBytes:&scale         length:sizeof(   float)      atIndex:20];
+                [encoder setBytes:&max_bias      length:sizeof(   float)      atIndex:21];
+                [encoder setBytes:&m0            length:sizeof(m0)            atIndex:22];
+                [encoder setBytes:&m1            length:sizeof(m1)            atIndex:23];
+                [encoder setBytes:&n_head_log2   length:sizeof(n_head_log2)   atIndex:24];
+                [encoder setBytes:&logit_softcap length:sizeof(logit_softcap) atIndex:25];
 
                 if (!use_vec_kernel) {
                     // half8x8 kernel
@@ -3482,29 +3385,25 @@ static void ggml_metal_encode_node(
                     default: GGML_ABORT("not implemented");
                 }
 
-                ggml_metal_kargs_cpy args = {
-                    /*.ne00 =*/ ne00,
-                    /*.ne01 =*/ ne01,
-                    /*.ne02 =*/ ne02,
-                    /*.ne03 =*/ ne03,
-                    /*.nb00 =*/ nb00,
-                    /*.nb01 =*/ nb01,
-                    /*.nb02 =*/ nb02,
-                    /*.nb03 =*/ nb03,
-                    /*.ne0  =*/ ne0,
-                    /*.ne1  =*/ ne1,
-                    /*.ne2  =*/ ne2,
-                    /*.ne3  =*/ ne3,
-                    /*.nb0  =*/ nb0,
-                    /*.nb1  =*/ nb1,
-                    /*.nb2  =*/ nb2,
-                    /*.nb3  =*/ nb3,
-                };
-
                 [encoder setComputePipelineState:pipeline];
-                [encoder setBytes:&args length:sizeof(args) atIndex:0];
+                [encoder setBuffer:id_src0 offset:offs_src0        atIndex:0];
-                [encoder setBuffer:id_src0 offset:offs_src0 atIndex:1];
+                [encoder setBuffer:id_dst  offset:offs_dst         atIndex:1];
-                [encoder setBuffer:id_dst  offset:offs_dst  atIndex:2];
+                [encoder setBytes:&ne00    length:sizeof( int64_t) atIndex:2];
+                [encoder setBytes:&ne01    length:sizeof( int64_t) atIndex:3];
+                [encoder setBytes:&ne02    length:sizeof( int64_t) atIndex:4];
+                [encoder setBytes:&ne03    length:sizeof( int64_t) atIndex:5];
+                [encoder setBytes:&nb00    length:sizeof(uint64_t) atIndex:6];
+                [encoder setBytes:&nb01    length:sizeof(uint64_t) atIndex:7];
+                [encoder setBytes:&nb02    length:sizeof(uint64_t) atIndex:8];
+                [encoder setBytes:&nb03    length:sizeof(uint64_t) atIndex:9];
+                [encoder setBytes:&ne0     length:sizeof( int64_t) atIndex:10];
+                [encoder setBytes:&ne1     length:sizeof( int64_t) atIndex:11];
+                [encoder setBytes:&ne2     length:sizeof( int64_t) atIndex:12];
+                [encoder setBytes:&ne3     length:sizeof( int64_t) atIndex:13];
+                [encoder setBytes:&nb0     length:sizeof(uint64_t) atIndex:14];
+                [encoder setBytes:&nb1     length:sizeof(uint64_t) atIndex:15];
+                [encoder setBytes:&nb2     length:sizeof(uint64_t) atIndex:16];
+                [encoder setBytes:&nb3     length:sizeof(uint64_t) atIndex:17];
 
                 [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
             } break;
@@ -3549,7 +3448,6 @@ static void ggml_metal_encode_node(
                 const int64_t n_threads = MIN((int64_t)[pipeline maxTotalThreadsPerThreadgroup], parallel_elements);
                 const int64_t n_tg = (parallel_elements + n_threads - 1) / n_threads;
 
-                // TODO: add ggml_metal_kargs struct
                 [encoder setComputePipelineState:pipeline];
                 [encoder setBuffer:id_src0 offset:offs_src0       atIndex:0];
                 [encoder setBuffer:id_dst  offset:offs_dst        atIndex:1];
@@ -3737,12 +3635,6 @@ static void * ggml_backend_metal_buffer_get_base(ggml_backend_buffer_t buffer) {
     return ctx->all_data;
 }
 
-static void ggml_backend_metal_buffer_memset_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, uint8_t value, size_t offset, size_t size) {
-    memset((char *)tensor->data + offset, value, size);
-
-    UNUSED(buffer);
-}
-
 static void ggml_backend_metal_buffer_set_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
     memcpy((char *)tensor->data + offset, data, size);
 
@@ -3775,7 +3667,7 @@ static struct ggml_backend_buffer_i ggml_backend_metal_buffer_i = {
     /* .free_buffer     = */ ggml_backend_metal_buffer_free_buffer,
     /* .get_base        = */ ggml_backend_metal_buffer_get_base,
     /* .init_tensor     = */ NULL,
-    /* .memset_tensor   = */ ggml_backend_metal_buffer_memset_tensor,
+    /* .memset_tensor   = */ NULL,
     /* .set_tensor      = */ ggml_backend_metal_buffer_set_tensor,
     /* .get_tensor      = */ ggml_backend_metal_buffer_get_tensor,
     /* .cpy_tensor      = */ ggml_backend_metal_buffer_cpy_tensor,

ggml/src/ggml-metal/CMakeLists.txt

@@ -1,108 +0,0 @@
-find_library(FOUNDATION_LIBRARY Foundation REQUIRED)
-find_library(METAL_FRAMEWORK    Metal      REQUIRED)
-find_library(METALKIT_FRAMEWORK MetalKit   REQUIRED)
-
-message(STATUS "Metal framework found")
-
-add_library(ggml-metal
-            ggml-metal.m
-            )
-
-target_link_libraries(ggml-metal PRIVATE
-                      ggml-base
-                      ${FOUNDATION_LIBRARY}
-                      ${METAL_FRAMEWORK}
-                      ${METALKIT_FRAMEWORK}
-                      )
-
-target_include_directories(ggml-metal PRIVATE . ..)
-
-if (GGML_METAL_NDEBUG)
-    add_compile_definitions(GGML_METAL_NDEBUG)
-endif()
-
-if (GGML_METAL_USE_BF16)
-    add_compile_definitions(GGML_METAL_USE_BF16)
-endif()
-
-# copy metal files to bin directory
-configure_file(../ggml-common.h  ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-common.h     COPYONLY)
-configure_file(ggml-metal.metal  ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.metal  COPYONLY)
-configure_file(ggml-metal-impl.h ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal-impl.h COPYONLY)
-
-if (GGML_METAL_EMBED_LIBRARY)
-    enable_language(ASM)
-
-    add_compile_definitions(GGML_METAL_EMBED_LIBRARY)
-
-    set(METALLIB_COMMON "${CMAKE_CURRENT_SOURCE_DIR}/../ggml-common.h")
-    set(METALLIB_SOURCE "${CMAKE_CURRENT_SOURCE_DIR}/ggml-metal.metal")
-    set(METALLIB_IMPL   "${CMAKE_CURRENT_SOURCE_DIR}/ggml-metal-impl.h")
-
-    file(MAKE_DIRECTORY "${CMAKE_BINARY_DIR}/autogenerated")
-
-    # merge ggml-common.h and ggml-metal.metal into a single file
-    set(METALLIB_EMBED_ASM        "${CMAKE_BINARY_DIR}/autogenerated/ggml-metal-embed.s")
-    set(METALLIB_SOURCE_EMBED     "${CMAKE_BINARY_DIR}/autogenerated/ggml-metal-embed.metal")
-    set(METALLIB_SOURCE_EMBED_TMP "${CMAKE_BINARY_DIR}/autogenerated/ggml-metal-embed.metal.tmp")
-
-    add_custom_command(
-        OUTPUT ${METALLIB_EMBED_ASM}
-        COMMAND echo "Embedding Metal library"
-        COMMAND sed -e '/__embed_ggml-common.h__/r         ${METALLIB_COMMON}' -e '/__embed_ggml-common.h__/d'         < ${METALLIB_SOURCE}           > ${METALLIB_SOURCE_EMBED_TMP}
-        COMMAND sed -e '/\#include \"ggml-metal-impl.h\"/r ${METALLIB_IMPL}'   -e '/\#include \"ggml-metal-impl.h\"/d' < ${METALLIB_SOURCE_EMBED_TMP} > ${METALLIB_SOURCE_EMBED}
-        COMMAND echo ".section __DATA,__ggml_metallib"          >  ${METALLIB_EMBED_ASM}
-        COMMAND echo ".globl _ggml_metallib_start"              >> ${METALLIB_EMBED_ASM}
-        COMMAND echo "_ggml_metallib_start:"                    >> ${METALLIB_EMBED_ASM}
-        COMMAND echo ".incbin \\\"${METALLIB_SOURCE_EMBED}\\\"" >> ${METALLIB_EMBED_ASM}
-        COMMAND echo ".globl _ggml_metallib_end"                >> ${METALLIB_EMBED_ASM}
-        COMMAND echo "_ggml_metallib_end:"                      >> ${METALLIB_EMBED_ASM}
-        DEPENDS ../ggml-common.h ggml-metal.metal ggml-metal-impl.h
-        COMMENT "Generate assembly for embedded Metal library"
-    )
-
-    target_sources(ggml-metal PRIVATE ${METALLIB_EMBED_ASM})
-else()
-    if (GGML_METAL_SHADER_DEBUG)
-        # custom command to do the following:
-        #   xcrun -sdk macosx metal    -fno-fast-math -c ggml-metal.metal -o ggml-metal.air
-        #   xcrun -sdk macosx metallib                   ggml-metal.air   -o default.metallib
-        #
-        # note: this is the only way I found to disable fast-math in Metal. it's ugly, but at least it works
-        #       disabling fast math is needed in order to pass tests/test-backend-ops
-        # note: adding -fno-inline fixes the tests when using MTL_SHADER_VALIDATION=1
-        # note: unfortunately, we have to call it default.metallib instead of ggml.metallib
-        #       ref: https://github.com/ggerganov/whisper.cpp/issues/1720
-        set(XC_FLAGS -fno-fast-math -fno-inline -g)
-    else()
-        set(XC_FLAGS -O3)
-    endif()
-
-    # Append macOS metal versioning flags
-    if (GGML_METAL_MACOSX_VERSION_MIN)
-        message(STATUS "Adding  -mmacosx-version-min=${GGML_METAL_MACOSX_VERSION_MIN} flag to metal compilation")
-        list   (APPEND XC_FLAGS -mmacosx-version-min=${GGML_METAL_MACOSX_VERSION_MIN})
-    endif()
-
-    if (GGML_METAL_STD)
-        message(STATUS "Adding  -std=${GGML_METAL_STD} flag to metal compilation")
-        list   (APPEND XC_FLAGS -std=${GGML_METAL_STD})
-    endif()
-
-    add_custom_command(
-        OUTPUT ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
-        COMMAND xcrun -sdk macosx metal    ${XC_FLAGS} -c ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.metal -o ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.air
-        COMMAND xcrun -sdk macosx metallib                ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.air   -o ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
-        COMMAND rm -f ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.air
-        COMMAND rm -f ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-common.h
-        COMMAND rm -f ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.metal
-        DEPENDS ggml-metal.metal ggml-common.h
-        COMMENT "Compiling Metal kernels"
-        )
-
-    # FIXME: only add to the ggml-metal target?
-    add_custom_target(
-        ggml-metal-lib ALL
-        DEPENDS ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
-        )
-endif() # GGML_METAL_EMBED_LIBRARY

ggml/src/ggml-metal/ggml-metal-impl.h

@@ -1,249 +0,0 @@
-#ifndef GGML_METAL_IMPL
-#define GGML_METAL_IMPL
-
-// kernel argument structs
-//
-// - element counters (e.g. ne00) typically use int32_t to reduce register usage
-//   however, be careful from int overflows when using those in the kernel implementation
-//
-// - strides (e.g. nb00) use uint64_t
-
-typedef struct {
-    int32_t  ne00;
-    int32_t  ne01;
-    int32_t  ne02;
-    int32_t  ne03;
-    uint64_t nb00;
-    uint64_t nb01;
-    uint64_t nb02;
-    uint64_t nb03;
-    int32_t  ne10;
-    int32_t  ne11;
-    int32_t  ne12;
-    int32_t  ne13;
-    uint64_t nb10;
-    uint64_t nb11;
-    uint64_t nb12;
-    uint64_t nb13;
-    int32_t  ne0;
-    int32_t  ne1;
-    int32_t  ne2;
-    int32_t  ne3;
-    uint64_t nb0;
-    uint64_t nb1;
-    uint64_t nb2;
-    uint64_t nb3;
-    int32_t  dim;
-} ggml_metal_kargs_concat;
-
-typedef struct {
-    int32_t  ne00;
-    int32_t  ne01;
-    int32_t  ne02;
-    int32_t  ne03;
-    uint64_t nb00;
-    uint64_t nb01;
-    uint64_t nb02;
-    uint64_t nb03;
-    int32_t  ne10;
-    int32_t  ne11;
-    int32_t  ne12;
-    int32_t  ne13;
-    uint64_t nb10;
-    uint64_t nb11;
-    uint64_t nb12;
-    uint64_t nb13;
-    int32_t  ne0;
-    int32_t  ne1;
-    int32_t  ne2;
-    int32_t  ne3;
-    uint64_t nb0;
-    uint64_t nb1;
-    uint64_t nb2;
-    uint64_t nb3;
-    uint64_t offs;
-} ggml_metal_kargs_bin;
-
-typedef struct {
-    int32_t  ne00;
-    int32_t  ne01;
-    int32_t  ne02;
-    int32_t  ne03;
-    uint64_t nb00;
-    uint64_t nb01;
-    uint64_t nb02;
-    uint64_t nb03;
-    int32_t  ne0;
-    int32_t  ne1;
-    int32_t  ne2;
-    int32_t  ne3;
-    uint64_t nb0;
-    uint64_t nb1;
-    uint64_t nb2;
-    uint64_t nb3;
-} ggml_metal_kargs_repeat;
-
-typedef struct {
-    int64_t  ne00;
-    int64_t  ne01;
-    int64_t  ne02;
-    int64_t  ne03;
-    uint64_t nb00;
-    uint64_t nb01;
-    uint64_t nb02;
-    uint64_t nb03;
-    int64_t  ne0;
-    int64_t  ne1;
-    int64_t  ne2;
-    int64_t  ne3;
-    uint64_t nb0;
-    uint64_t nb1;
-    uint64_t nb2;
-    uint64_t nb3;
-} ggml_metal_kargs_cpy;
-
-typedef struct {
-    int32_t  ne00;
-    int32_t  ne01;
-    int32_t  ne02;
-    int32_t  ne03;
-    uint64_t nb00;
-    uint64_t nb01;
-    uint64_t nb02;
-    uint64_t nb03;
-    int32_t  ne0;
-    int32_t  ne1;
-    int32_t  ne2;
-    int32_t  ne3;
-    uint64_t nb0;
-    uint64_t nb1;
-    uint64_t nb2;
-    uint64_t nb3;
-    int32_t  n_past;
-    int32_t  n_dims;
-    int32_t  n_ctx_orig;
-    float    freq_base;
-    float    freq_scale;
-    float    ext_factor;
-    float    attn_factor;
-    float    beta_fast;
-    float    beta_slow;
-} ggml_metal_kargs_rope;
-
-typedef struct {
-    int32_t  ne01;
-    int32_t  ne02;
-    int32_t  ne03;
-    uint64_t nb01;
-    uint64_t nb02;
-    uint64_t nb03;
-    int32_t  ne11;
-    int32_t  ne_12_2; // assume K and V are same shape
-    int32_t  ne_12_3;
-    uint64_t nb_12_1;
-    uint64_t nb_12_2;
-    uint64_t nb_12_3;
-    uint64_t nb31;
-    int32_t  ne1;
-    int32_t  ne2;
-    float    scale;
-    float    max_bias;
-    float    m0;
-    float    m1;
-    uint16_t n_head_log2;
-    float    logit_softcap;
-} ggml_metal_kargs_flash_attn_ext;
-
-typedef struct {
-    int32_t  ne00;
-    int32_t  ne02;
-    uint64_t nb01;
-    uint64_t nb02;
-    uint64_t nb03;
-    int32_t  ne12;
-    uint64_t nb10;
-    uint64_t nb11;
-    uint64_t nb12;
-    uint64_t nb13;
-    int32_t  ne0;
-    int32_t  ne1;
-    int16_t  r2;
-    int16_t  r3;
-} ggml_metal_kargs_mul_mm;
-
-typedef struct {
-    int32_t  ne00;
-    int32_t  ne01;
-    int32_t  ne02;
-    uint64_t nb00;
-    uint64_t nb01;
-    uint64_t nb02;
-    uint64_t nb03;
-    int32_t  ne10;
-    int32_t  ne11;
-    int32_t  ne12;
-    uint64_t nb10;
-    uint64_t nb11;
-    uint64_t nb12;
-    uint64_t nb13;
-    int32_t  ne0;
-    int32_t  ne1;
-    int16_t  r2;
-    int16_t  r3;
-} ggml_metal_kargs_mul_mv;
-
-typedef struct {
-    int32_t  nei0;
-    int32_t  nei1;
-    uint64_t nbi1;
-    int32_t  ne00;
-    int32_t  ne02;
-    uint64_t nb01;
-    uint64_t nb02;
-    int32_t  ne11;
-    int32_t  ne12;
-    int32_t  ne13;
-    uint64_t nb10;
-    uint64_t nb11;
-    uint64_t nb12;
-    int32_t  ne0;
-    int32_t  ne1;
-} ggml_metal_kargs_mul_mm_id;
-
-typedef struct {
-    int32_t  nei0;
-    int32_t  nei1;
-    uint64_t nbi1;
-    int32_t  ne00;
-    int32_t  ne01;
-    int32_t  ne02;
-    uint64_t nb00;
-    uint64_t nb01;
-    uint64_t nb02;
-    int32_t  ne10;
-    int32_t  ne11;
-    int32_t  ne12;
-    int32_t  ne13;
-    uint64_t nb10;
-    uint64_t nb11;
-    uint64_t nb12;
-    int32_t  ne0;
-    int32_t  ne1;
-    uint64_t nb1;
-} ggml_metal_kargs_mul_mv_id;
-
-typedef struct {
-    int32_t  ne00;
-    int32_t  ne00_4;
-    uint64_t nb01;
-    float    eps;
-} ggml_metal_kargs_norm;
-
-typedef struct {
-    int32_t  ne00;
-    int32_t  ne00_4;
-    uint64_t nb01;
-    float    eps;
-} ggml_metal_kargs_rms_norm;
-
-#endif // GGML_METAL_IMPL

ggml/src/ggml-musa/CMakeLists.txt

@@ -1,100 +0,0 @@
-if (NOT EXISTS $ENV{MUSA_PATH})
-    if (NOT EXISTS /opt/musa)
-        set(MUSA_PATH /usr/local/musa)
-    else()
-        set(MUSA_PATH /opt/musa)
-    endif()
-else()
-    set(MUSA_PATH $ENV{MUSA_PATH})
-endif()
-
-set(CMAKE_C_COMPILER "${MUSA_PATH}/bin/clang")
-set(CMAKE_C_EXTENSIONS OFF)
-set(CMAKE_CXX_COMPILER "${MUSA_PATH}/bin/clang++")
-set(CMAKE_CXX_EXTENSIONS OFF)
-
-list(APPEND CMAKE_MODULE_PATH "${MUSA_PATH}/cmake")
-
-find_package(MUSAToolkit)
-
-if (MUSAToolkit_FOUND)
-    message(STATUS "MUSA Toolkit found")
-
-    file(GLOB   GGML_HEADERS_MUSA "../ggml-cuda/*.cuh")
-    list(APPEND GGML_HEADERS_MUSA "../../include/ggml-cuda.h")
-
-    file(GLOB   GGML_SOURCES_MUSA "../ggml-cuda/*.cu")
-    file(GLOB   SRCS "../ggml-cuda/template-instances/fattn-wmma*.cu")
-    list(APPEND GGML_SOURCES_MUSA ${SRCS})
-    file(GLOB   SRCS "../ggml-cuda/template-instances/mmq*.cu")
-    list(APPEND GGML_SOURCES_MUSA ${SRCS})
-
-    if (GGML_CUDA_FA_ALL_QUANTS)
-        file(GLOB   SRCS "../ggml-cuda/template-instances/fattn-vec*.cu")
-        list(APPEND GGML_SOURCES_MUSA ${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_MUSA ${SRCS})
-        file(GLOB   SRCS "../ggml-cuda/template-instances/fattn-vec*q8_0-q8_0.cu")
-        list(APPEND GGML_SOURCES_MUSA ${SRCS})
-        file(GLOB   SRCS "../ggml-cuda/template-instances/fattn-vec*f16-f16.cu")
-        list(APPEND GGML_SOURCES_MUSA ${SRCS})
-    endif()
-
-    set_source_files_properties(${GGML_SOURCES_MUSA} PROPERTIES LANGUAGE CXX)
-    foreach(SOURCE ${GGML_SOURCES_MUSA})
-        set_property(SOURCE ${SOURCE} PROPERTY COMPILE_FLAGS "-x musa -mtgpu --cuda-gpu-arch=mp_21 --cuda-gpu-arch=mp_22")
-    endforeach()
-
-    add_library(ggml-musa
-                ${GGML_HEADERS_MUSA}
-                ${GGML_SOURCES_MUSA})
-
-    target_link_libraries(ggml-musa PRIVATE ggml-base)
-    target_include_directories(ggml-musa PRIVATE . ..)
-
-    # TODO: do not use CUDA definitions for MUSA
-    target_compile_definitions(ggml PUBLIC GGML_USE_CUDA)
-
-    add_compile_definitions(GGML_USE_MUSA)
-    add_compile_definitions(GGML_CUDA_PEER_MAX_BATCH_SIZE=${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)
-        target_link_libraries(ggml-musa PRIVATE MUSA::musart_static MUSA::mublas_static)
-    else()
-        target_link_libraries(ggml-musa PRIVATE MUSA::musart MUSA::mublas)
-    endif()
-
-    if (GGML_CUDA_NO_VMM)
-        # No VMM requested, no need to link directly with the musa driver lib (libmusa.so)
-    else()
-        target_link_libraries(ggml-musa PRIVATE MUSA::musa_driver)
-    endif()
-else()
-    message(FATAL_ERROR "MUSA Toolkit not found")
-endif()

ggml/src/ggml-musa/ggml/CMakeLists.txt

@@ -1,100 +0,0 @@
-if (NOT EXISTS $ENV{MUSA_PATH})
-    if (NOT EXISTS /opt/musa)
-        set(MUSA_PATH /usr/local/musa)
-    else()
-        set(MUSA_PATH /opt/musa)
-    endif()
-else()
-    set(MUSA_PATH $ENV{MUSA_PATH})
-endif()
-
-set(CMAKE_C_COMPILER "${MUSA_PATH}/bin/clang")
-set(CMAKE_C_EXTENSIONS OFF)
-set(CMAKE_CXX_COMPILER "${MUSA_PATH}/bin/clang++")
-set(CMAKE_CXX_EXTENSIONS OFF)
-
-list(APPEND CMAKE_MODULE_PATH "${MUSA_PATH}/cmake")
-
-find_package(MUSAToolkit)
-
-if (MUSAToolkit_FOUND)
-    message(STATUS "MUSA Toolkit found")
-
-    file(GLOB   GGML_HEADERS_MUSA "../ggml-cuda/*.cuh")
-    list(APPEND GGML_HEADERS_MUSA "../../include/ggml-cuda.h")
-
-    file(GLOB   GGML_SOURCES_MUSA "../ggml-cuda/*.cu")
-    file(GLOB   SRCS "../ggml-cuda/template-instances/fattn-wmma*.cu")
-    list(APPEND GGML_SOURCES_MUSA ${SRCS})
-    file(GLOB   SRCS "../ggml-cuda/template-instances/mmq*.cu")
-    list(APPEND GGML_SOURCES_MUSA ${SRCS})
-
-    if (GGML_CUDA_FA_ALL_QUANTS)
-        file(GLOB   SRCS "../ggml-cuda/template-instances/fattn-vec*.cu")
-        list(APPEND GGML_SOURCES_MUSA ${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_MUSA ${SRCS})
-        file(GLOB   SRCS "../ggml-cuda/template-instances/fattn-vec*q8_0-q8_0.cu")
-        list(APPEND GGML_SOURCES_MUSA ${SRCS})
-        file(GLOB   SRCS "../ggml-cuda/template-instances/fattn-vec*f16-f16.cu")
-        list(APPEND GGML_SOURCES_MUSA ${SRCS})
-    endif()
-
-    set_source_files_properties(${GGML_SOURCES_MUSA} PROPERTIES LANGUAGE CXX)
-    foreach(SOURCE ${GGML_SOURCES_MUSA})
-        set_property(SOURCE ${SOURCE} PROPERTY COMPILE_FLAGS "-x musa -mtgpu --cuda-gpu-arch=mp_21 --cuda-gpu-arch=mp_22")
-    endforeach()
-
-    add_library(ggml-musa
-                ${GGML_HEADERS_MUSA}
-                ${GGML_SOURCES_MUSA})
-
-    target_link_libraries(ggml-musa PRIVATE ggml-base)
-    target_include_directories(ggml-musa PRIVATE . ..)
-
-    # TODO: do not use CUDA definitions for MUSA
-    target_compile_definitions(ggml PUBLIC GGML_USE_CUDA)
-
-    add_compile_definitions(GGML_USE_MUSA)
-    add_compile_definitions(GGML_CUDA_PEER_MAX_BATCH_SIZE=${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)
-        target_link_libraries(ggml-musa PRIVATE MUSA::musart_static MUSA::mublas_static)
-    else()
-        target_link_libraries(ggml-musa PRIVATE MUSA::musart MUSA::mublas)
-    endif()
-
-    if (GGML_CUDA_NO_VMM)
-        # No VMM requested, no need to link directly with the musa driver lib (libmusa.so)
-    else()
-        target_link_libraries(ggml-musa PRIVATE MUSA::musa_driver)
-    endif()
-else()
-    message(FATAL_ERROR "MUSA Toolkit not found")
-endif()

ggml/src/ggml-opt.cpp

@@ -1,867 +0,0 @@
-#include "ggml-opt.h"
-
-#include "ggml.h"
-#include "ggml-alloc.h"
-#include "ggml-backend.h"
-#include "ggml-impl.h"
-
-#include <algorithm>
-#include <cmath>
-#include <cstdint>
-#include <cinttypes>
-#include <map>
-#include <random>
-#include <vector>
-
-struct ggml_opt_dataset {
-    struct ggml_context   * ctx;
-    ggml_backend_buffer_t   buf;
-    struct ggml_tensor    * data;
-    struct ggml_tensor    * labels;
-
-    int64_t ndata;
-    int64_t ndata_shard;
-    size_t  nbs_data;
-    size_t  nbs_labels;
-
-    std::vector<int64_t> permutation;
-};
-
-struct ggml_opt_context {
-    ggml_backend_sched_t    backend_sched;
-    ggml_cgraph           * allocated_graph;
-    ggml_cgraph           * allocated_graph_copy;
-    struct ggml_context   * ctx_static;
-    struct ggml_context   * ctx_static_cpu;
-    struct ggml_context   * ctx_compute;
-    struct ggml_context   * ctx_copy;
-    ggml_backend_buffer_t   buf_static;
-    ggml_backend_buffer_t   buf_static_cpu;
-    std::mt19937            rng;
-
-    struct ggml_tensor * inputs;
-    struct ggml_tensor * outputs;
-    struct ggml_tensor * labels;
-
-    struct ggml_tensor * loss;
-    struct ggml_tensor * pred;
-    struct ggml_tensor * ncorrect;
-
-    struct ggml_cgraph * gf;
-    struct ggml_cgraph * gb_grad;
-    struct ggml_cgraph * gb_opt;
-
-    int64_t iter;
-    int32_t opt_period;
-    int32_t opt_i;
-    bool    loss_per_datapoint;
-
-    ggml_opt_get_optimizer_params get_opt_pars;
-    void * get_opt_pars_ud;
-    struct ggml_tensor * adamw_params;
-};
-
-struct ggml_opt_result {
-    int64_t              ndata    = 0;
-    std::vector<float>   loss;
-    std::vector<int32_t> pred;
-    int64_t              ncorrect = 0;
-
-    bool loss_per_datapoint = false;
-    int64_t opt_period = -1;
-};
-
-// ====== Dataset ======
-
-ggml_opt_dataset_t ggml_opt_dataset_init(int64_t ne_datapoint, int64_t ne_label, int64_t ndata, int64_t ndata_shard) {
-    GGML_ASSERT(ne_datapoint >  0);
-    GGML_ASSERT(ne_label     >= 0);
-    GGML_ASSERT(ndata        >  0);
-    GGML_ASSERT(ndata_shard  >  0);
-
-    ggml_opt_dataset_t result = new ggml_opt_dataset;
-    result->ndata       = ndata;
-    result->ndata_shard = ndata_shard;
-
-    {
-        struct ggml_init_params params = {
-            /*.mem_size   =*/ 2*ggml_tensor_overhead(),
-            /*.mem_buffer =*/ nullptr,
-            /*.no_alloc   =*/ true,
-        };
-        result->ctx = ggml_init(params);
-    }
-
-    result->data = ggml_new_tensor_2d(result->ctx, GGML_TYPE_F32, ne_datapoint, ndata);
-    result->nbs_data = ggml_nbytes(result->data) * ndata_shard/ndata;
-
-    if (ne_label > 0) {
-        result->labels = ggml_new_tensor_2d(result->ctx, GGML_TYPE_F32, ne_label, ndata);
-        result->nbs_labels = ggml_nbytes(result->labels) * ndata_shard/ndata;
-    } else {
-        result->labels = nullptr;
-        result->nbs_labels = 0;
-    }
-
-    result->buf = ggml_backend_alloc_ctx_tensors_from_buft(result->ctx, ggml_backend_cpu_buffer_type());
-
-    const int64_t nshards = ndata/ndata_shard;
-    result->permutation.resize(nshards);
-    for (int64_t i = 0; i < nshards; ++i) {
-        result->permutation[i] = i;
-    }
-    return result;
-}
-
-void ggml_opt_dataset_free(ggml_opt_dataset_t dataset) {
-    ggml_backend_buffer_free(dataset->buf);
-    ggml_free(dataset->ctx);
-    delete dataset;
-}
-
-struct ggml_tensor * ggml_opt_dataset_data(ggml_opt_dataset_t dataset) {
-    return dataset->data;
-}
-
-struct ggml_tensor * ggml_opt_dataset_labels(ggml_opt_dataset_t dataset) {
-    return dataset->labels;
-}
-
-void ggml_opt_dataset_shuffle(ggml_opt_context_t opt_ctx, ggml_opt_dataset_t dataset, int64_t idata) {
-    GGML_ASSERT(idata <= dataset->ndata);
-
-    if (idata < 0) {
-        std::shuffle(dataset->permutation.begin(), dataset->permutation.end(), opt_ctx->rng);
-        return;
-    }
-
-    GGML_ASSERT(idata % dataset->ndata_shard == 0);
-    const int64_t ishard_max = idata / dataset->ndata_shard;
-    std::shuffle(dataset->permutation.begin(), dataset->permutation.begin() + ishard_max, opt_ctx->rng);
-}
-
-void ggml_opt_dataset_get_batch(ggml_opt_dataset_t dataset, struct ggml_tensor * data_batch, struct ggml_tensor * labels_batch, int64_t ibatch) {
-    GGML_ASSERT(   data_batch && ggml_is_contiguous(data_batch));
-    GGML_ASSERT(!labels_batch || ggml_is_contiguous(labels_batch));
-    GGML_ASSERT((labels_batch == nullptr) == (dataset->labels == nullptr));
-
-    const size_t nb_data_batch = ggml_nbytes(data_batch);
-    GGML_ASSERT(nb_data_batch % dataset->nbs_data == 0);
-    const int64_t shards_per_batch = nb_data_batch / dataset->nbs_data;
-
-    if (labels_batch) {
-        const size_t nb_labels_batch = ggml_nbytes(labels_batch);
-        GGML_ASSERT(nb_labels_batch == shards_per_batch*dataset->nbs_labels);
-    }
-
-    GGML_ASSERT((ibatch + 1)*shards_per_batch <= int64_t(dataset->permutation.size()));
-
-    for (int64_t ishard_batch = 0; ishard_batch < shards_per_batch; ++ishard_batch) {
-        const int64_t ishard = dataset->permutation[ibatch*shards_per_batch + ishard_batch];
-
-        const char * ptr_data = (const char *) dataset->data->data + ishard*dataset->nbs_data;
-        ggml_backend_tensor_set(data_batch, ptr_data, ishard_batch*dataset->nbs_data, dataset->nbs_data);
-
-        if (!labels_batch) {
-            continue;
-        }
-
-        const char * ptr_labels = (const char *) dataset->labels->data + ishard*dataset->nbs_labels;
-        ggml_backend_tensor_set(labels_batch, ptr_labels, ishard_batch*dataset->nbs_labels, dataset->nbs_labels);
-    }
-}
-
-// ====== Model / Context ======
-
-struct ggml_opt_optimizer_params ggml_opt_get_default_optimizer_params(void * userdata) {
-    GGML_UNUSED(userdata);
-
-    ggml_opt_optimizer_params result;
-
-    result.adamw.alpha = 0.001f;
-    result.adamw.beta1 = 0.9f;
-    result.adamw.beta2 = 0.999f;
-    result.adamw.eps   = 1e-8f;
-    result.adamw.wd    = 0.0f;
-
-    return result;
-}
-
-struct 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) {
-    return {
-        /*backend_sched   =*/ backend_sched,
-        /*ctx_compute     =*/ ctx_compute,
-        /*inputs          =*/ inputs,
-        /*logits          =*/ outputs,
-        /*loss_type       =*/ loss_type,
-        /*build_type      =*/ GGML_OPT_BUILD_TYPE_OPT,
-        /*opt_period      =*/ 1,
-        /*get_opt_pars    =*/ ggml_opt_get_default_optimizer_params,
-        /*get_opt_pars_ud =*/ nullptr,
-    };
-}
-
-static ggml_tensor * map_tensor(std::map<ggml_tensor *, ggml_tensor *> & tensor_map, ggml_context * ctx, ggml_tensor * tensor) {
-    if (!tensor) {
-        return nullptr;
-    }
-
-    if (tensor_map.find(tensor) != tensor_map.end()) {
-        return tensor_map[tensor];
-    }
-
-    ggml_tensor * new_tensor = ggml_dup_tensor(ctx, tensor);
-    tensor_map[tensor] = new_tensor;
-
-    new_tensor->op = tensor->op;
-    for (int i = 0; i < GGML_MAX_DIMS; i++) {
-        new_tensor->nb[i] = tensor->nb[i];
-    }
-    new_tensor->flags = tensor->flags;
-    memcpy(new_tensor->op_params, tensor->op_params, sizeof(tensor->op_params));
-    strcpy(new_tensor->name, tensor->name);
-    new_tensor->data = tensor->data;
-    new_tensor->buffer = tensor->buffer;
-    new_tensor->extra = tensor->extra;
-    new_tensor->view_offs = tensor->view_offs;
-    new_tensor->view_src = map_tensor(tensor_map, ctx, tensor->view_src);
-    for (int i = 0; i < GGML_MAX_SRC; i++) {
-        new_tensor->src[i] = map_tensor(tensor_map, ctx, tensor->src[i]);
-    }
-
-    return new_tensor;
-}
-
-static ggml_cgraph * dup_graph(ggml_context * ctx, ggml_cgraph * graph) {
-    std::map<ggml_tensor *, ggml_tensor *> tensor_map;
-
-    ggml_cgraph * new_graph = ggml_new_graph_custom(ctx, GGML_DEFAULT_GRAPH_SIZE, /*grads =*/ true);
-
-    for (int i = 0; i < graph->n_leafs; i++) {
-        ggml_build_forward_expand(new_graph, map_tensor(tensor_map, ctx, graph->leafs[i]));
-    }
-    for (int i = 0; i < graph->n_nodes; i++) {
-        ggml_build_forward_expand(new_graph, map_tensor(tensor_map, ctx, graph->nodes[i]));
-    }
-    for (int i = 0; i < graph->n_nodes; ++i) {
-        const size_t igrad_src = ggml_hash_find(&graph->visited_hash_set, graph->nodes[i]);
-        const size_t igrad_dst = ggml_hash_find(&new_graph->visited_hash_set, new_graph->nodes[i]);
-        graph->grads[igrad_dst]     = new_graph->grads[igrad_src];
-        graph->grad_accs[igrad_dst] = new_graph->grad_accs[igrad_src];
-    }
-
-    return new_graph;
-}
-
-static void ggml_opt_alloc_graph(ggml_opt_context_t opt_ctx, ggml_cgraph * graph) {
-    GGML_ASSERT(graph);
-    if (opt_ctx->allocated_graph == graph) {
-        return;
-    }
-
-    ggml_backend_sched_reset(opt_ctx->backend_sched); // clear allocation of previous graph
-
-    {
-        ggml_init_params params = {
-            /*.mem_size   =*/ ggml_tensor_overhead() * GGML_DEFAULT_GRAPH_SIZE,
-            /*.mem_buffer =*/ nullptr,
-            /*.no_alloc   =*/ true,
-        };
-        ggml_free(opt_ctx->ctx_copy);
-        opt_ctx->ctx_copy = ggml_init(params);
-    }
-
-    opt_ctx->allocated_graph_copy = dup_graph(opt_ctx->ctx_copy, graph);
-
-    ggml_backend_sched_alloc_graph(opt_ctx->backend_sched, opt_ctx->allocated_graph_copy);
-    opt_ctx->allocated_graph = graph;
-}
-
-ggml_opt_context_t ggml_opt_init(struct ggml_opt_params params) {
-    ggml_opt_context_t result = new struct ggml_opt_context;
-    result->backend_sched        = params.backend_sched;
-    result->allocated_graph      = nullptr;
-    result->allocated_graph_copy = nullptr;
-    result->ctx_compute          = params.ctx_compute;
-    result->ctx_copy             = nullptr;
-    result->inputs               = params.inputs;
-    result->outputs              = params.outputs;
-    result->iter                 = 1;
-    result->opt_period           = params.opt_period;
-    result->opt_i                = 0;
-    result->get_opt_pars         = params.get_opt_pars;
-    result->get_opt_pars_ud      = params.get_opt_pars_ud;
-
-    GGML_ASSERT(result->inputs->data && "the inputs must be allocated statically");
-    GGML_ASSERT(result->opt_period >= 1);
-
-    const bool accumulate = params.build_type == GGML_OPT_BUILD_TYPE_GRAD ||
-        (params.build_type == GGML_OPT_BUILD_TYPE_OPT && result->opt_period > 1);
-
-    ggml_set_input(result->inputs);
-    ggml_set_output(result->outputs);
-
-    result->gf = ggml_new_graph_custom(result->ctx_compute, GGML_DEFAULT_GRAPH_SIZE, /*grads =*/ true); // Forward pass.
-    ggml_build_forward_expand(result->gf, result->outputs);
-
-    int n_param = 0;
-    for (int i = 0; i < result->gf->n_nodes; ++i) {
-        if (result->gf->nodes[i]->flags & GGML_TENSOR_FLAG_PARAM) {
-            n_param++;
-        }
-    }
-
-    {
-        // The static context is used for:
-        //   - gradients (1 tensor per param if using gradient accumulation)
-        //   - optimizer momenta (2 tensors per param)
-        //   - labels
-        //   - loss + its gradient (up to 5 tensors)
-        //   - pred
-        //   - ncorrect (2 tensors).
-        const size_t tensors_per_param = (accumulate ? 1 : 0) + (params.build_type == GGML_OPT_BUILD_TYPE_OPT ? 2 : 0);
-        const size_t size_meta = (tensors_per_param*n_param + 9) * ggml_tensor_overhead();
-        struct ggml_init_params params = {
-            /*.mem_size   =*/ size_meta,
-            /*.mem_buffer =*/ nullptr,
-            /*.no_alloc   =*/ true,
-        };
-        result->ctx_static = ggml_init(params);
-    }
-    {
-        // The static cpu context is used for:
-        //   - optimizer parameters (1 for the entire context)
-        const size_t size_meta = 1 * ggml_tensor_overhead();
-        struct ggml_init_params params = {
-            /*.mem_size   =*/ size_meta,
-            /*.mem_buffer =*/ nullptr,
-            /*.no_alloc   =*/ true,
-        };
-        result->ctx_static_cpu = ggml_init(params);
-    }
-
-
-    switch (params.loss_type) {
-        case GGML_OPT_LOSS_TYPE_MEAN: {
-            result->labels = nullptr;
-            result->loss = ggml_sum(result->ctx_static, result->outputs);
-            ggml_set_name(result->loss, "loss_sum");
-            const float scale = 1.0f / (result->opt_period * ggml_nelements(result->outputs));
-            result->loss = ggml_scale(result->ctx_static, result->loss, scale);
-            ggml_set_name(result->loss, "loss_mean");
-            result->loss_per_datapoint = true;
-            break;
-        }
-        case GGML_OPT_LOSS_TYPE_SUM: {
-            result->labels = nullptr;
-            result->loss = ggml_sum(result->ctx_static, result->outputs);
-            ggml_set_name(result->loss, "loss_sum");
-            result->loss_per_datapoint = false;
-            break;
-        }
-        case GGML_OPT_LOSS_TYPE_CROSS_ENTROPY: {
-            result->labels = ggml_dup_tensor(result->ctx_static, result->outputs);
-            ggml_set_input(result->labels);
-            ggml_set_name(result->labels, "labels");
-            result->loss = ggml_cross_entropy_loss(result->ctx_static, result->outputs, result->labels);
-            ggml_set_name(result->loss, "loss_cross_entropy");
-            if (result->opt_period > 1) {
-                result->loss = ggml_scale(result->ctx_static, result->loss, 1.0f / result->opt_period);
-                ggml_set_name(result->loss, "loss_cross_entropy_scaled");
-            }
-            result->loss_per_datapoint = true;
-            break;
-        }
-        case GGML_OPT_LOSS_TYPE_MEAN_SQUARED_ERROR: {
-            result->labels = ggml_dup_tensor(result->ctx_static, result->outputs);
-            ggml_set_input(result->labels);
-            ggml_set_name(result->labels, "labels");
-            result->loss = ggml_sub(result->ctx_static, result->outputs, result->labels);
-            ggml_set_name(result->loss, "loss_error");
-            result->loss = ggml_sqr(result->ctx_static, result->loss);
-            ggml_set_name(result->loss, "loss_squared_error");
-            result->loss = ggml_sum(result->ctx_static, result->loss);
-            ggml_set_name(result->loss, "loss_sum_squared_error");
-            const float scale = 1.0f / (result->opt_period * ggml_nelements(result->outputs));
-            result->loss = ggml_scale(result->ctx_static, result->loss, scale);
-            ggml_set_name(result->loss, "loss_mean_squared_error");
-            result->loss_per_datapoint = true;
-            break;
-        }
-    }
-    ggml_set_output(result->loss);
-    ggml_set_loss(result->loss);
-    ggml_build_forward_expand(result->gf, result->loss);
-
-    result->pred = ggml_argmax(result->ctx_static, result->outputs);
-    ggml_set_name(result->pred, "pred");
-    ggml_set_output(result->pred);
-    ggml_build_forward_expand(result->gf, result->pred);
-
-    if (result->labels) {
-        result->ncorrect = ggml_count_equal(result->ctx_static, result->pred, ggml_argmax(result->ctx_static, result->labels));
-        ggml_set_name(result->ncorrect, "ncorrect");
-        ggml_set_output(result->ncorrect);
-        ggml_build_forward_expand(result->gf, result->ncorrect);
-    } else {
-        result->ncorrect = nullptr;
-    }
-
-    if (params.build_type == GGML_OPT_BUILD_TYPE_FORWARD) {
-        result->gb_grad = nullptr;
-        result->gb_opt  = nullptr;
-
-        result->buf_static = ggml_backend_alloc_ctx_tensors(result->ctx_static, ggml_backend_sched_get_backend(result->backend_sched, 0));
-        result->buf_static_cpu = nullptr;
-
-        ggml_opt_alloc_graph(result, result->gf);
-
-        return result;
-    }
-
-    // gb_grad == graph backward gradients, forward pass, then backward pass to calculate gradients.
-    result->gb_grad = ggml_graph_dup(result->ctx_compute, result->gf);
-    ggml_build_backward_expand(result->ctx_static, result->ctx_compute, result->gb_grad, accumulate);
-
-    if (params.build_type == GGML_OPT_BUILD_TYPE_GRAD) {
-        result->gb_opt  = nullptr;
-
-        result->buf_static = ggml_backend_alloc_ctx_tensors(result->ctx_static, ggml_backend_sched_get_backend(result->backend_sched, 0));
-        result->buf_static_cpu = nullptr;
-
-        ggml_opt_alloc_graph(result, result->gb_grad);
-        ggml_graph_reset(result->gb_grad);
-
-        return result;
-    }
-
-    GGML_ASSERT(params.build_type == GGML_OPT_BUILD_TYPE_OPT);
-
-    // gb_opt == graph backward optimize, forward pass, then backward pass to calculate gradients, then optimizer step.
-    result->gb_opt = ggml_graph_dup(result->ctx_compute, result->gb_grad);
-
-    result->adamw_params = ggml_new_tensor_1d(result->ctx_static_cpu, GGML_TYPE_F32, 7);
-    ggml_set_input(result->adamw_params);
-    ggml_set_name(result->adamw_params, "adamw_params");
-
-    for (int i = result->gf->n_nodes-1; i >= 0; --i) {
-        struct ggml_tensor * node = result->gb_opt->nodes[i];
-        struct ggml_tensor * grad = ggml_graph_get_grad(result->gb_opt, node);
-
-        if (node->flags & GGML_TENSOR_FLAG_PARAM) {
-            struct ggml_tensor * m        = ggml_dup_tensor(result->ctx_static, node);
-            struct ggml_tensor * v        = ggml_dup_tensor(result->ctx_static, node);
-            struct ggml_tensor * opt_step = ggml_opt_step_adamw(result->ctx_compute, node, grad, m, v, result->adamw_params);
-            ggml_build_forward_expand(result->gb_opt, opt_step);
-        }
-    }
-
-    result->buf_static = ggml_backend_alloc_ctx_tensors(
-        result->ctx_static, ggml_backend_sched_get_backend(result->backend_sched, 0));
-
-    result->buf_static_cpu = ggml_backend_alloc_ctx_tensors_from_buft(result->ctx_static_cpu, ggml_backend_cpu_buffer_type());
-
-    ggml_opt_alloc_graph(result, result->gb_opt);
-    ggml_graph_reset(result->gb_opt);
-
-    return result;
-}
-
-void ggml_opt_free(ggml_opt_context_t opt_ctx) {
-    if (opt_ctx == nullptr) {
-        return;
-    }
-    ggml_backend_buffer_free(opt_ctx->buf_static);
-    ggml_backend_buffer_free(opt_ctx->buf_static_cpu);
-    ggml_free(opt_ctx->ctx_static);
-    ggml_free(opt_ctx->ctx_static_cpu);
-    delete opt_ctx;
-}
-
-void ggml_opt_reset(ggml_opt_context_t opt_ctx, bool optimizer) {
-    if (optimizer) {
-        ggml_graph_reset(opt_ctx->gb_opt);
-        opt_ctx->iter = 1;
-    } else {
-        ggml_graph_reset(opt_ctx->gb_grad);
-    }
-}
-
-struct ggml_tensor * ggml_opt_inputs(ggml_opt_context_t opt_ctx) {
-    return opt_ctx->inputs;
-}
-
-struct ggml_tensor * ggml_opt_outputs(ggml_opt_context_t opt_ctx) {
-    return opt_ctx->outputs;
-}
-
-struct ggml_tensor * ggml_opt_labels(ggml_opt_context_t opt_ctx) {
-    return opt_ctx->labels;
-}
-
-struct ggml_tensor * ggml_opt_loss(ggml_opt_context_t opt_ctx) {
-    return opt_ctx->loss;
-}
-
-struct ggml_tensor * ggml_opt_pred(ggml_opt_context_t opt_ctx) {
-    return opt_ctx->pred;
-}
-
-struct ggml_tensor * ggml_opt_ncorrect(ggml_opt_context_t opt_ctx) {
-    return opt_ctx->ncorrect;
-}
-
-struct ggml_tensor * ggml_opt_grad_acc(ggml_opt_context_t opt_ctx, struct ggml_tensor * node) {
-    return ggml_graph_get_grad_acc(opt_ctx->gb_opt, node);
-}
-
-// ====== Optimization Result ======
-
-ggml_opt_result_t ggml_opt_result_init() {
-    return new ggml_opt_result;
-}
-
-void ggml_opt_result_free(ggml_opt_result_t result) {
-    delete result;
-}
-
-void ggml_opt_result_reset(ggml_opt_result_t result) {
-    result->ndata = 0;
-    result->loss.clear();
-    result->pred.clear();
-    result->ncorrect = 0;
-}
-
-void ggml_opt_result_ndata(ggml_opt_result_t result, int64_t * ndata) {
-    *ndata = result->ndata;
-}
-
-void ggml_opt_result_loss(ggml_opt_result_t result, double * loss, double * unc) {
-    const int64_t nbatches = result->loss.size(); // Number of physical batches.
-
-    if (nbatches == 0) {
-        *loss = 0.0;
-        *unc  = NAN;
-        return;
-    }
-
-    double sum         = 0.0;
-    double sum_squared = 0.0;
-
-    for (const float & loss : result->loss) {
-        // If the loss is per datapoint it was scaled by 1.0f/opt_period for each physical batch.
-        const float loss_scaled = result->loss_per_datapoint ? loss*result->opt_period : loss;
-        sum         += loss_scaled;
-        sum_squared += loss_scaled*loss_scaled;
-    }
-
-    const double mean = sum/nbatches;
-    *loss = result->loss_per_datapoint ? mean : sum;
-
-    if (!unc) {
-        return;
-    }
-
-    if (nbatches < 2) {
-        *unc = NAN;
-        return;
-    }
-
-    const double var_sum = sum_squared/nbatches - mean*mean; // variance without Bessel's correction, i.e. nbatches/(nbatches-1)
-    *unc = result->loss_per_datapoint ? sqrt(var_sum / (nbatches - 1)) : sqrt(var_sum * nbatches/(nbatches - 1));
-}
-
-void ggml_opt_result_pred(ggml_opt_result_t result, int32_t * pred) {
-    for (size_t i = 0; i < result->pred.size(); ++i) {
-        pred[i] = result->pred[i];
-    }
-}
-
-void ggml_opt_result_accuracy(ggml_opt_result_t result, double * accuracy, double * unc) {
-    *accuracy = result->ncorrect >= 0 ? double(result->ncorrect) / double(result->ndata) : NAN;
-
-    if (!unc) {
-        return;
-    }
-
-    *unc = result->ncorrect >= 0 && result->ndata >= 2 ?
-        sqrt((*accuracy) * (1.0 - (*accuracy)) / double(result->ndata - 1)) : NAN;
-}
-
-// ====== Computation ======
-
-static void ggml_opt_eval_graph(ggml_opt_context_t opt_ctx, ggml_cgraph * graph, ggml_opt_result * result) {
-    if (graph != opt_ctx->gf) {
-        struct ggml_opt_optimizer_params opt_pars = opt_ctx->get_opt_pars(opt_ctx->get_opt_pars_ud);
-
-        GGML_ASSERT(opt_pars.adamw.alpha >  0.0f);
-        GGML_ASSERT(opt_pars.adamw.beta1 >= 0.0f);
-        GGML_ASSERT(opt_pars.adamw.beta1 <= 1.0f);
-        GGML_ASSERT(opt_pars.adamw.beta2 >= 0.0f);
-        GGML_ASSERT(opt_pars.adamw.beta2 <= 1.0f);
-        GGML_ASSERT(opt_pars.adamw.eps   >= 0.0f);
-        GGML_ASSERT(opt_pars.adamw.wd    >= 0.0f);
-        GGML_ASSERT(opt_pars.adamw.wd    <= 1.0f);
-
-        // beta1, beta2 after applying warmup
-        const float beta1h = 1.0f/(1.0f - powf(opt_pars.adamw.beta1, opt_ctx->iter));
-        const float beta2h = 1.0f/(1.0f - powf(opt_pars.adamw.beta2, opt_ctx->iter));
-
-        float * adamw_par_data = ggml_get_data_f32(opt_ctx->adamw_params);
-        adamw_par_data[0] = opt_pars.adamw.alpha;
-        adamw_par_data[1] = opt_pars.adamw.beta1;
-        adamw_par_data[2] = opt_pars.adamw.beta2;
-        adamw_par_data[3] = opt_pars.adamw.eps;
-        adamw_par_data[4] = opt_pars.adamw.wd;
-        adamw_par_data[5] = beta1h;
-        adamw_par_data[6] = beta2h;
-    }
-
-    ggml_opt_alloc_graph(opt_ctx, graph);
-    ggml_backend_sched_graph_compute(opt_ctx->backend_sched, opt_ctx->allocated_graph_copy);
-    opt_ctx->iter += opt_ctx->allocated_graph == opt_ctx->gb_opt;
-
-    if (!result) {
-        return;
-    }
-
-    if (result->ndata == 0) {
-        result->loss_per_datapoint = opt_ctx->loss_per_datapoint;
-        result->opt_period         = opt_ctx->opt_period;
-    } else {
-        GGML_ASSERT(result->loss_per_datapoint == opt_ctx->loss_per_datapoint);
-        GGML_ASSERT(result->opt_period         == opt_ctx->opt_period);
-    }
-
-    const int64_t ndata = opt_ctx->outputs->ne[1];
-    GGML_ASSERT(result->ndata == ndata*int64_t(result->loss.size()) && "varying batch size not supported");
-    result->ndata += ndata;
-
-    GGML_ASSERT(ggml_is_scalar(opt_ctx->loss));
-    GGML_ASSERT(opt_ctx->loss->type == GGML_TYPE_F32);
-    float loss;
-    ggml_backend_tensor_get(opt_ctx->loss, &loss, 0, ggml_nbytes(opt_ctx->loss));
-    result->loss.push_back(loss);
-
-    GGML_ASSERT(opt_ctx->pred->type == GGML_TYPE_I32);
-    std::vector<int32_t> pred(ndata);
-    ggml_backend_tensor_get(opt_ctx->pred, pred.data(), 0, ggml_nbytes(opt_ctx->pred));
-    result->pred.insert(result->pred.end(), pred.begin(), pred.end());
-
-    if (!opt_ctx->labels || result->ncorrect < 0) {
-        result->ncorrect = -1;
-        return;
-    }
-
-    GGML_ASSERT(ggml_is_scalar(opt_ctx->ncorrect));
-    GGML_ASSERT(opt_ctx->ncorrect->type == GGML_TYPE_I64);
-    int64_t ncorrect;
-    ggml_backend_tensor_get(opt_ctx->ncorrect, &ncorrect, 0, ggml_nbytes(opt_ctx->ncorrect));
-    result->ncorrect += ncorrect;
-}
-
-void ggml_opt_forward(ggml_opt_context_t opt_ctx, ggml_opt_result * result) {
-    ggml_opt_eval_graph(opt_ctx, opt_ctx->gf, result);
-}
-
-void ggml_opt_forward_backward(ggml_opt_context_t opt_ctx, ggml_opt_result * result) {
-    if (opt_ctx->opt_period == 1) {
-        ggml_opt_eval_graph(opt_ctx, opt_ctx->gb_opt, result);
-        return;
-    }
-
-    const int32_t opt_i_next = (opt_ctx->opt_i + 1) % opt_ctx->opt_period;
-    if (opt_i_next == 0) {
-        ggml_opt_eval_graph(opt_ctx, opt_ctx->gb_opt, result);
-        ggml_opt_reset(opt_ctx, /*optimizer =*/ false);
-    } else {
-        ggml_opt_eval_graph(opt_ctx, opt_ctx->gb_grad, result);
-    }
-    opt_ctx->opt_i = opt_i_next;
-}
-
-// ====== High-Level Functions ======
-
-void ggml_opt_epoch(
-        ggml_opt_context_t      opt_ctx,
-        ggml_opt_dataset_t      dataset,
-        ggml_opt_result_t       result_train,
-        ggml_opt_result_t       result_eval,
-        int64_t                 idata_split,
-        ggml_opt_epoch_callback callback_train,
-        ggml_opt_epoch_callback callback_eval) {
-    struct ggml_tensor * inputs = ggml_opt_inputs(opt_ctx);
-    struct ggml_tensor * labels = ggml_opt_labels(opt_ctx);
-    struct ggml_tensor * data   = ggml_opt_dataset_data(dataset);
-    GGML_ASSERT(data->ne[0] == inputs->ne[0]);
-
-    const int64_t ndata       =   data->ne[1];
-    const int64_t ndata_batch = inputs->ne[1];
-
-    GGML_ASSERT(data->ne[1] % inputs->ne[1] == 0);
-    const int64_t nbatches = ndata/ndata_batch;
-
-    idata_split = idata_split < 0 ? ndata : idata_split;
-    GGML_ASSERT(idata_split % ndata_batch == 0);
-    const int64_t ibatch_split = idata_split / ndata_batch;
-
-    int64_t ibatch = 0;
-    int64_t t_loop_start = ggml_time_us();
-    for (; ibatch < ibatch_split; ++ibatch) {
-        ggml_opt_dataset_get_batch(dataset, inputs, labels, ibatch);
-        ggml_opt_forward_backward(opt_ctx, result_train);
-        if (callback_train) {
-            callback_train(true, opt_ctx, dataset, result_train, ibatch+1, ibatch_split, t_loop_start);
-        }
-    }
-    t_loop_start = ggml_time_us();
-    for (; ibatch < nbatches; ++ibatch) {
-        ggml_opt_dataset_get_batch(dataset, inputs, labels, ibatch);
-        ggml_opt_forward(opt_ctx, result_eval);
-        if (callback_eval) {
-            callback_eval(false, opt_ctx, dataset, result_eval, ibatch+1-ibatch_split, nbatches-ibatch_split, t_loop_start);
-        }
-    }
-}
-
-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) {
-    fprintf(stderr, "%s[", train ? "train: " : "val:   ");
-
-    constexpr int64_t bar_length = 25;
-    for (int64_t j = 0; j < bar_length; ++j) {
-        const int64_t ibatch_j = ibatch_max * j/bar_length;
-        if (ibatch_j < ibatch) {
-            fprintf(stderr, "=");
-        } else if (ibatch_max * (j - 1)/bar_length < ibatch) {
-            fprintf(stderr, ">");
-        } else {
-            fprintf(stderr, " ");
-        }
-    }
-
-    const int64_t batch_size = ggml_opt_inputs(opt_ctx)->ne[1];
-    const int64_t idata      = ibatch*batch_size;
-    const int64_t idata_max  = ibatch_max*batch_size;
-
-    double loss;
-    double loss_unc;
-    ggml_opt_result_loss(result, &loss, &loss_unc);
-
-    double accuracy;
-    double accuracy_unc;
-    ggml_opt_result_accuracy(result, &accuracy, &accuracy_unc);
-
-    const int64_t t_ibatch_us = ggml_time_us() - t_start_us;
-    int64_t t_ibatch_s = t_ibatch_us / 1000000;
-    const int64_t t_ibatch_h = t_ibatch_s / 3600;
-    t_ibatch_s -= t_ibatch_h * 3600;
-    const int64_t t_ibatch_m = t_ibatch_s / 60;
-    t_ibatch_s -= t_ibatch_m * 60;
-
-    const int64_t t_eta_us = t_ibatch_us * (ibatch_max - ibatch)/ibatch;
-    int64_t t_eta_s = t_eta_us / 1000000;
-    const int64_t t_eta_h = t_eta_s / 3600;
-    t_eta_s -= t_eta_h * 3600;
-    const int64_t t_eta_m = t_eta_s / 60;
-    t_eta_s -= t_eta_m * 60;
-
-    fprintf(stderr, "| data=%06" PRId64 "/%06" PRId64 ", loss=%.6lf+-%.6lf, accuracy=%.2lf+-%.2lf%%, "
-            "t=%02" PRId64 ":%02" PRId64 ":%02" PRId64 ", ETA=%02" PRId64 ":%02" PRId64 ":%02" PRId64 "]\r",
-            idata, idata_max, loss, loss_unc, 100.0*accuracy, 100.0*accuracy_unc,
-            t_ibatch_h, t_ibatch_m, t_ibatch_s, t_eta_h, t_eta_m, t_eta_s);
-    if (ibatch == ibatch_max) {
-        fprintf(stderr, "\n");
-    }
-    fflush(stderr);
-
-    GGML_UNUSED(dataset);
-}
-
-void ggml_opt_fit(
-        ggml_backend_sched_t            backend_sched,
-        ggml_context                  * ctx_compute,
-        ggml_tensor                   * inputs,
-        ggml_tensor                   * outputs,
-        ggml_opt_dataset_t              dataset,
-        enum ggml_opt_loss_type         loss_type,
-        ggml_opt_get_optimizer_params   get_opt_pars,
-        int64_t                         nepoch,
-        int64_t                         nbatch_logical,
-        float                           val_split,
-        bool                            silent) {
-    ggml_time_init();
-    const int64_t t_start_us = ggml_time_us();
-
-    const int64_t ndata           = ggml_opt_dataset_data(dataset)->ne[1];
-    const int64_t nbatch_physical = inputs->ne[1];
-    GGML_ASSERT(ndata          % nbatch_logical  == 0);
-    GGML_ASSERT(nbatch_logical % nbatch_physical == 0);
-
-    const int64_t opt_period       = nbatch_logical / nbatch_physical;
-    const int64_t nbatches_logical = ndata / nbatch_logical;
-
-    GGML_ASSERT(val_split >= 0.0f);
-    GGML_ASSERT(val_split <  1.0f);
-    const int64_t ibatch_split = int64_t(((1.0f - val_split) * nbatches_logical)) * opt_period; // train <-> val split index (physical)
-    const int64_t idata_split  = ibatch_split * nbatch_physical;
-
-    int64_t epoch = 1;
-
-    ggml_opt_params params = ggml_opt_default_params(backend_sched, ctx_compute, inputs, outputs, loss_type);
-    params.opt_period      = opt_period;
-    params.get_opt_pars    = get_opt_pars;
-    params.get_opt_pars_ud = &epoch;
-    ggml_opt_context_t opt_ctx = ggml_opt_init(params);
-
-    // Shuffling the data is generally useful but there is only a point if not all data is used in a single batch.
-    if (nbatch_logical < ndata) {
-        ggml_opt_dataset_shuffle(opt_ctx, dataset, -1); // Shuffle all data (train + validation).
-    }
-
-    ggml_opt_result_t result_train = ggml_opt_result_init();
-    ggml_opt_result_t result_val   = ggml_opt_result_init();
-
-    ggml_opt_epoch_callback epoch_callback = silent ? nullptr : ggml_opt_epoch_callback_progress_bar;
-
-    for (; epoch <= nepoch; ++epoch) {
-        if (nbatch_logical < idata_split) {
-            ggml_opt_dataset_shuffle(opt_ctx, dataset, idata_split);
-        }
-
-        ggml_opt_result_reset(result_train);
-        ggml_opt_result_reset(result_val);
-
-        if (!silent) {
-            fprintf(stderr, "%s: epoch %04" PRId64 "/%04" PRId64 ":\n", __func__, epoch, nepoch);
-        }
-        ggml_opt_epoch(opt_ctx, dataset, result_train, result_val, idata_split, epoch_callback, epoch_callback);
-        if (!silent) {
-            fprintf(stderr, "\n");
-        }
-    }
-
-    if (!silent) {
-        int64_t t_total_s = (ggml_time_us() - t_start_us) / 1000000;
-        const int64_t t_total_h = t_total_s / 3600;
-        t_total_s -= t_total_h * 3600;
-        const int64_t t_total_m = t_total_s / 60;
-        t_total_s -= t_total_m * 60;
-        fprintf(stderr, "%s: training took %02" PRId64 ":%02" PRId64 ":%02" PRId64 "\n", __func__, t_total_h, t_total_m, t_total_s);
-    }
-
-    ggml_opt_free(opt_ctx);
-    ggml_opt_result_free(result_train);
-    ggml_opt_result_free(result_val);
-}

ggml/src/ggml-quants.h

@@ -11,89 +11,136 @@
 extern "C" {
 #endif
 
-// NOTE: these functions are defined as GGML_API because they used by the CPU backend
-
 // Quantization
-GGML_API void quantize_row_q4_0_ref(const float * GGML_RESTRICT x, block_q4_0 * GGML_RESTRICT y, int64_t k);
+void quantize_row_q4_0_ref(const float * GGML_RESTRICT x, block_q4_0 * GGML_RESTRICT y, int64_t k);
-GGML_API void quantize_row_q4_1_ref(const float * GGML_RESTRICT x, block_q4_1 * GGML_RESTRICT y, int64_t k);
+void quantize_row_q4_1_ref(const float * GGML_RESTRICT x, block_q4_1 * GGML_RESTRICT y, int64_t k);
-GGML_API void quantize_row_q5_0_ref(const float * GGML_RESTRICT x, block_q5_0 * GGML_RESTRICT y, int64_t k);
+void quantize_row_q5_0_ref(const float * GGML_RESTRICT x, block_q5_0 * GGML_RESTRICT y, int64_t k);
-GGML_API void quantize_row_q5_1_ref(const float * GGML_RESTRICT x, block_q5_1 * GGML_RESTRICT y, int64_t k);
+void quantize_row_q5_1_ref(const float * GGML_RESTRICT x, block_q5_1 * GGML_RESTRICT y, int64_t k);
-GGML_API void quantize_row_q8_0_ref(const float * GGML_RESTRICT x, block_q8_0 * GGML_RESTRICT y, int64_t k);
+void quantize_row_q8_0_ref(const float * GGML_RESTRICT x, block_q8_0 * GGML_RESTRICT y, int64_t k);
-GGML_API void quantize_row_q8_1_ref(const float * GGML_RESTRICT x, block_q8_1 * GGML_RESTRICT y, int64_t k);
+void quantize_row_q8_1_ref(const float * GGML_RESTRICT x, block_q8_1 * GGML_RESTRICT y, int64_t k);
 
-GGML_API void quantize_row_q2_K_ref(const float * GGML_RESTRICT x, block_q2_K * GGML_RESTRICT y, int64_t k);
+void quantize_row_q2_K_ref(const float * GGML_RESTRICT x, block_q2_K * GGML_RESTRICT y, int64_t k);
-GGML_API void quantize_row_q3_K_ref(const float * GGML_RESTRICT x, block_q3_K * GGML_RESTRICT y, int64_t k);
+void quantize_row_q3_K_ref(const float * GGML_RESTRICT x, block_q3_K * GGML_RESTRICT y, int64_t k);
-GGML_API void quantize_row_q4_K_ref(const float * GGML_RESTRICT x, block_q4_K * GGML_RESTRICT y, int64_t k);
+void quantize_row_q4_K_ref(const float * GGML_RESTRICT x, block_q4_K * GGML_RESTRICT y, int64_t k);
-GGML_API void quantize_row_q5_K_ref(const float * GGML_RESTRICT x, block_q5_K * GGML_RESTRICT y, int64_t k);
+void quantize_row_q5_K_ref(const float * GGML_RESTRICT x, block_q5_K * GGML_RESTRICT y, int64_t k);
-GGML_API void quantize_row_q6_K_ref(const float * GGML_RESTRICT x, block_q6_K * GGML_RESTRICT y, int64_t k);
+void quantize_row_q6_K_ref(const float * GGML_RESTRICT x, block_q6_K * GGML_RESTRICT y, int64_t k);
-GGML_API void quantize_row_q8_K_ref(const float * GGML_RESTRICT x, block_q8_K * GGML_RESTRICT y, int64_t k);
+void quantize_row_q8_K_ref(const float * GGML_RESTRICT x, block_q8_K * GGML_RESTRICT y, int64_t k);
 
-GGML_API void quantize_row_tq1_0_ref(const float * GGML_RESTRICT x, block_tq1_0 * GGML_RESTRICT y, int64_t k);
+void quantize_row_tq1_0_ref(const float * GGML_RESTRICT x, block_tq1_0 * GGML_RESTRICT y, int64_t k);
-GGML_API void quantize_row_tq2_0_ref(const float * GGML_RESTRICT x, block_tq2_0 * GGML_RESTRICT y, int64_t k);
+void quantize_row_tq2_0_ref(const float * GGML_RESTRICT x, block_tq2_0 * GGML_RESTRICT y, int64_t k);
 
-GGML_API void quantize_row_iq3_xxs_ref(const float * GGML_RESTRICT x, block_iq3_xxs * GGML_RESTRICT y, int64_t k);
+void quantize_row_iq3_xxs_ref(const float * GGML_RESTRICT x, block_iq3_xxs * GGML_RESTRICT y, int64_t k);
-GGML_API void quantize_row_iq4_nl_ref (const float * GGML_RESTRICT x, block_iq4_nl  * GGML_RESTRICT y, int64_t k);
+void quantize_row_iq4_nl_ref (const float * GGML_RESTRICT x, block_iq4_nl  * GGML_RESTRICT y, int64_t k);
-GGML_API void quantize_row_iq4_xs_ref (const float * GGML_RESTRICT x, block_iq4_xs  * GGML_RESTRICT y, int64_t k);
+void quantize_row_iq4_xs_ref (const float * GGML_RESTRICT x, block_iq4_xs  * GGML_RESTRICT y, int64_t k);
-GGML_API void quantize_row_iq3_s_ref  (const float * GGML_RESTRICT x, block_iq3_s   * GGML_RESTRICT y, int64_t k);
+void quantize_row_iq3_s_ref  (const float * GGML_RESTRICT x, block_iq3_s   * GGML_RESTRICT y, int64_t k);
-GGML_API void quantize_row_iq2_s_ref  (const float * GGML_RESTRICT x, block_iq2_s   * GGML_RESTRICT y, int64_t k);
+void quantize_row_iq2_s_ref  (const float * GGML_RESTRICT x, block_iq2_s   * GGML_RESTRICT y, int64_t k);
+
+void quantize_row_q4_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q4_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q5_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q5_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+
+void quantize_row_q2_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q3_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q4_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q5_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q6_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q8_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+
+void quantize_row_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_iq3_xxs(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_iq4_nl (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_iq4_xs (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_iq3_s  (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_iq2_s  (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
 
 // Dequantization
-GGML_API void dequantize_row_q4_0(const block_q4_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+void dequantize_row_q4_0(const block_q4_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-GGML_API void dequantize_row_q4_1(const block_q4_1 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+void dequantize_row_q4_1(const block_q4_1 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-GGML_API void dequantize_row_q5_0(const block_q5_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+void dequantize_row_q5_0(const block_q5_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-GGML_API void dequantize_row_q5_1(const block_q5_1 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+void dequantize_row_q5_1(const block_q5_1 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-GGML_API void dequantize_row_q8_0(const block_q8_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+void dequantize_row_q8_0(const block_q8_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-//GGML_API void dequantize_row_q8_1(const block_q8_1 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+//void dequantize_row_q8_1(const block_q8_1 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
 
-GGML_API void dequantize_row_q2_K(const block_q2_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+void dequantize_row_q2_K(const block_q2_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-GGML_API void dequantize_row_q3_K(const block_q3_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+void dequantize_row_q3_K(const block_q3_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-GGML_API void dequantize_row_q4_K(const block_q4_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+void dequantize_row_q4_K(const block_q4_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-GGML_API void dequantize_row_q5_K(const block_q5_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+void dequantize_row_q5_K(const block_q5_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-GGML_API void dequantize_row_q6_K(const block_q6_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+void dequantize_row_q6_K(const block_q6_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-GGML_API void dequantize_row_q8_K(const block_q8_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+void dequantize_row_q8_K(const block_q8_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
 
-GGML_API void dequantize_row_tq1_0(const block_tq1_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+void dequantize_row_tq1_0(const block_tq1_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-GGML_API void dequantize_row_tq2_0(const block_tq2_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+void dequantize_row_tq2_0(const block_tq2_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
 
-GGML_API void dequantize_row_iq2_xxs(const block_iq2_xxs * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+void dequantize_row_iq2_xxs(const block_iq2_xxs * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-GGML_API void dequantize_row_iq2_xs (const block_iq2_xs  * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+void dequantize_row_iq2_xs (const block_iq2_xs  * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-GGML_API void dequantize_row_iq2_s  (const block_iq2_s   * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+void dequantize_row_iq2_s  (const block_iq2_s   * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-GGML_API void dequantize_row_iq3_xxs(const block_iq3_xxs * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+void dequantize_row_iq3_xxs(const block_iq3_xxs * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-GGML_API void dequantize_row_iq1_s  (const block_iq1_s   * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+void dequantize_row_iq1_s  (const block_iq1_s   * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-GGML_API void dequantize_row_iq1_m  (const block_iq1_m   * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+void dequantize_row_iq1_m  (const block_iq1_m   * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-GGML_API void dequantize_row_iq4_nl (const block_iq4_nl  * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+void dequantize_row_iq4_nl (const block_iq4_nl  * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-GGML_API void dequantize_row_iq4_xs (const block_iq4_xs  * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+void dequantize_row_iq4_xs (const block_iq4_xs  * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-GGML_API void dequantize_row_iq3_s  (const block_iq3_s   * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+void dequantize_row_iq3_s  (const block_iq3_s   * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+
+// Dot product
+void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+
+void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+
+void ggml_vec_dot_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);
 
 // Quantization utilizing an importance matrix (a.k.a. "Activation aWare Quantization")
-GGML_API size_t quantize_iq2_xxs(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+size_t quantize_iq2_xxs(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-GGML_API size_t quantize_iq2_xs (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+size_t quantize_iq2_xs (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-GGML_API size_t quantize_iq2_s  (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+size_t quantize_iq2_s  (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-GGML_API size_t quantize_iq3_xxs(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+size_t quantize_iq3_xxs(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-GGML_API size_t quantize_iq1_s  (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+size_t quantize_iq1_s  (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-GGML_API size_t quantize_iq1_m  (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+size_t quantize_iq1_m  (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-GGML_API size_t quantize_iq4_nl (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+size_t quantize_iq4_nl (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-GGML_API size_t quantize_iq4_xs (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+size_t quantize_iq4_xs (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-GGML_API size_t quantize_iq3_s  (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+size_t quantize_iq3_s  (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
 
-GGML_API size_t quantize_tq1_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+size_t quantize_tq1_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-GGML_API size_t quantize_tq2_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+size_t quantize_tq2_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
 
-GGML_API size_t quantize_q2_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+size_t quantize_q2_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-GGML_API size_t quantize_q3_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+size_t quantize_q3_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-GGML_API size_t quantize_q4_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+size_t quantize_q4_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-GGML_API size_t quantize_q5_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+size_t quantize_q5_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-GGML_API size_t quantize_q6_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+size_t quantize_q6_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-GGML_API size_t quantize_q4_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+size_t quantize_q4_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-GGML_API size_t quantize_q4_1(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+size_t quantize_q4_1(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-GGML_API size_t quantize_q5_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+size_t quantize_q5_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-GGML_API size_t quantize_q5_1(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+size_t quantize_q5_1(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-GGML_API size_t quantize_q8_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+size_t quantize_q8_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
 
-GGML_API void iq2xs_init_impl(enum ggml_type type);
+void iq2xs_init_impl(enum ggml_type type);
-GGML_API void iq2xs_free_impl(enum ggml_type type);
+void iq2xs_free_impl(enum ggml_type type);
-GGML_API void iq3xs_init_impl(int grid_size);
+void iq3xs_init_impl(int grid_size);
-GGML_API void iq3xs_free_impl(int grid_size);
+void iq3xs_free_impl(int grid_size);
 
 #ifdef __cplusplus
 }

ggml/src/ggml-rpc.cpp

@@ -671,7 +671,7 @@ static ggml_backend_i ggml_backend_rpc_interface = {
     /* .event_wait              = */ NULL,
 };
 
-ggml_backend_buffer_type_t ggml_backend_rpc_buffer_type(const char * endpoint) {
+GGML_API ggml_backend_buffer_type_t ggml_backend_rpc_buffer_type(const char * endpoint) {
     static std::mutex mutex;
     std::lock_guard<std::mutex> lock(mutex);
     // NOTE: buffer types are allocated and never freed; this is by design
@@ -718,7 +718,7 @@ ggml_backend_t ggml_backend_rpc_init(const char * endpoint) {
     return backend;
 }
 
-bool ggml_backend_is_rpc(ggml_backend_t backend) {
+GGML_API bool ggml_backend_is_rpc(ggml_backend_t backend) {
     return backend != NULL && ggml_guid_matches(backend->guid, ggml_backend_rpc_guid());
 }
 
@@ -730,7 +730,7 @@ static void get_device_memory(const std::shared_ptr<socket_t> & sock, size_t * f
     *total = response.total_mem;
 }
 
-void ggml_backend_rpc_get_device_memory(const char * endpoint, size_t * free, size_t * total) {
+GGML_API void ggml_backend_rpc_get_device_memory(const char * endpoint, size_t * free, size_t * total) {
     auto sock = get_socket(endpoint);
     if (sock == nullptr) {
         *free = 0;

ggml/src/ggml-rpc/CMakeLists.txt

@@ -1,11 +0,0 @@
-message(STATUS "Using RPC backend")
-
-add_library(ggml-rpc
-            ggml-rpc.cpp)
-
-target_link_libraries(ggml-rpc PRIVATE ggml-base)
-target_include_directories(ggml-rpc PRIVATE . ..)
-
-if (WIN32)
-    target_link_libraries(ggml-rpc PRIVATE ws2_32)
-endif()

ggml/src/ggml-sycl.cpp

@@ -4350,6 +4350,10 @@ static bool ggml_backend_sycl_device_supports_op(ggml_backend_dev_t dev, const g
                 if (op->op == GGML_OP_MUL_MAT) {
                     a = op->src[0];
                     b = op->src[1];
+                    if (ggml_is_permuted(a) || ggml_is_permuted(b)) {
+                        // TODO: fix like https://github.com/ggerganov/llama.cpp/pull/10021
+                        return false;
+                    }
                 } else {
                     a = op->src[2];
                     b = op->src[1];

ggml/src/ggml-sycl/CMakeLists.txt

@@ -1,85 +0,0 @@
-if (NOT GGML_SYCL_TARGET MATCHES "^(INTEL|NVIDIA|AMD)$")
-    message(FATAL_ERROR "Invalid backend chosen, supported options are INTEL, NVIDIA, or AMD")
-endif()
-
-check_cxx_compiler_flag("-fsycl" SUPPORTS_SYCL)
-
-if (DEFINED ENV{ONEAPI_ROOT})
-    message(STATUS "Using oneAPI Release SYCL compiler (icpx).")
-elseif(SUPPORTS_SYCL)
-    message(WARNING "Using open-source SYCL compiler (clang++). Didn't detect ENV {ONEAPI_ROOT}.
-        If you expected the oneAPI Release compiler, please install oneAPI & source it, like:
-        source /opt/intel/oneapi/setvars.sh")
-else()
-    message(FATAL_ERROR, "C++ compiler lacks SYCL support.")
-endif()
-message(STATUS "SYCL found")
-#todo: AOT
-
-add_library(ggml-sycl
-            ggml-sycl.cpp
-            ../../include/ggml-sycl.h)
-
-target_link_libraries(ggml-sycl PRIVATE ggml-base)
-target_include_directories(ggml-sycl PRIVATE . ..)
-
-if (GGML_SYCL_F16)
-    if (GGML_SYCL_TARGET STREQUAL "AMD")
-        message(WARNING "AMD target does not entirely support FP16 in the SYCL backend.")
-    endif()
-    add_compile_definitions(GGML_SYCL_F16)
-endif()
-
-set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wno-narrowing -fsycl")
-
-if (GGML_SYCL_TARGET STREQUAL "NVIDIA")
-    add_compile_definitions(GGML_SYCL_WARP_SIZE=32)
-elseif (GGML_SYCL_TARGET STREQUAL "AMD")
-    # INFO: Allowed Sub_group_sizes are not consistent through all
-    # hip targets. For example, 64 is used for certain models, but the backend
-    # does not support it.
-    # Target archs tested working: gfx1030, gfx1031, (Only tested sub_group_size = 32)
-    add_compile_definitions(GGML_SYCL_WARP_SIZE=32)
-else()
-    add_compile_definitions(GGML_SYCL_WARP_SIZE=16)
-endif()
-
-file(GLOB   GGML_HEADERS_SYCL "*.hpp")
-file(GLOB   GGML_SOURCES_SYCL "*.cpp")
-target_sources(ggml-sycl PRIVATE ${GGML_HEADERS_SYCL} ${GGML_SOURCES_SYCL})
-
-find_package(DNNL)
-message("-- DNNL found:" ${DNNL_FOUND})
-
-if (GGML_SYCL_TARGET STREQUAL "INTEL")
-    add_compile_definitions(GGML_SYCL_DNNL=${DNNL_FOUND})
-else()
-    add_compile_definitions(GGML_SYCL_DNNL=0)
-endif()
-
-if (${DNNL_FOUND} AND GGML_SYCL_TARGET STREQUAL "INTEL")
-    target_link_libraries(ggml-sycl PRIVATE DNNL::dnnl)
-endif()
-
-if (WIN32)
-    find_package(IntelSYCL REQUIRED)
-    find_package(MKL REQUIRED)
-    target_link_libraries(ggml-sycl PRIVATE IntelSYCL::SYCL_CXX MKL::MKL MKL::MKL_SYCL)
-else()
-    if (GGML_SYCL_TARGET STREQUAL "INTEL")
-        target_link_libraries(ggml-sycl PRIVATE sycl OpenCL mkl_core pthread m dl mkl_sycl_blas mkl_intel_ilp64 mkl_tbb_thread)
-    elseif (GGML_SYCL_TARGET STREQUAL "NVIDIA")
-        set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fsycl-targets=nvptx64-nvidia-cuda")
-        target_link_libraries(ggml-sycl PRIVATE sycl pthread m dl onemkl)
-    elseif (GGML_SYCL_TARGET STREQUAL "AMD")
-        if (NOT GGML_SYCL_DEVICE_ARCH)
-            message(ERROR "Can't enable SYCL hip backend, GGML_SYCL_DEVICE_ARCH has not been set.")
-        endif()
-        set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fsycl-targets=amdgcn-amd-amdhsa")
-        target_link_libraries(ggml-sycl PRIVATE sycl pthread m dl onemkl)
-    endif()
-
-    if (GGML_SYCL_DEVICE_ARCH)
-      set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Xsycl-target-backend --offload-arch=${GGML_SYCL_DEVICE_ARCH}")
-  endif()
-endif()

ggml/src/ggml-sycl/dpct/helper.hpp

@@ -15,7 +15,6 @@
 
 #include <sycl/sycl.hpp>
 #include <sycl/half_type.hpp>
-#include <syclcompat/math.hpp>
 #include <oneapi/mkl.hpp>
 #include <map>
 
@@ -1831,10 +1830,31 @@ namespace dpct
                                            : id);
     }
 
+    template <typename T>
+    sycl::vec<T, 4> extract_and_sign_or_zero_extend4(T val)
+    {
+        return sycl::vec<T, 1>(val)
+            .template as<sycl::vec<
+                std::conditional_t<std::is_signed_v<T>, int8_t, uint8_t>, 4>>()
+            .template convert<T>();
+    }
+
+    template <typename T1, typename T2>
+    using dot_product_acc_t =
+        std::conditional_t<std::is_unsigned_v<T1> && std::is_unsigned_v<T2>,
+                           uint32_t, int32_t>;
+
     template <typename T1, typename T2, typename T3>
     inline auto dp4a(T1 a, T2 b, T3 c)
     {
-        return syclcompat::dp4a(a, b, c);
+        dot_product_acc_t<T1, T2> res = c;
+        auto va = extract_and_sign_or_zero_extend4(a);
+        auto vb = extract_and_sign_or_zero_extend4(b);
+        res += va[0] * vb[0];
+        res += va[1] * vb[1];
+        res += va[2] * vb[2];
+        res += va[3] * vb[3];
+        return res;
     }
 
     struct sub_sat

ggml/src/ggml-sycl/vecdotq.hpp

@@ -968,8 +968,8 @@ vec_dot_iq3_xxs_q8_1(const void *__restrict__ vbq,
             grid1[0] ^ signs[0], signs[0], std::minus<>());
         const int grid_h = dpct::vectorized_binary<sycl::uchar4>(
             grid2[0] ^ signs[1], signs[1], std::minus<>());
-        sumi = dpct::dp4a(grid_l, *((const int *)q8 + 0), sumi);
+        sumi = dpct::dp4a(grid_l, *((int *)q8 + 0), sumi);
-        sumi = dpct::dp4a(grid_h, *((const int *)q8 + 1), sumi);
+        sumi = dpct::dp4a(grid_h, *((int *)q8 + 1), sumi);
         q8 += 8;
         aux32 >>= 7;
     }
@@ -1009,8 +1009,8 @@ vec_dot_iq3_s_q8_1(const void *__restrict__ vbq,
             grid1[0] ^ signs0, signs0, std::minus<>());
         const int grid_h = dpct::vectorized_binary<sycl::uchar4>(
             grid2[0] ^ signs1, signs1, std::minus<>());
-        sumi = dpct::dp4a(grid_l, *((const int *)q8 + 0), sumi);
+        sumi = dpct::dp4a(grid_l, *((int *)q8 + 0), sumi);
-        sumi = dpct::dp4a(grid_h, *((const int *)q8 + 1), sumi);
+        sumi = dpct::dp4a(grid_h, *((int *)q8 + 1), sumi);
         q8 += 8;
     }
     const float d =

ggml/src/ggml-threading.cpp

@@ -1,12 +0,0 @@
-#include "ggml-threading.h"
-#include <mutex>
-
-std::mutex ggml_critical_section_mutex;
-
-void ggml_critical_section_start() {
-    ggml_critical_section_mutex.lock();
-}
-
-void ggml_critical_section_end(void) {
-    ggml_critical_section_mutex.unlock();
-}

ggml/src/ggml-threading.h

@@ -1,12 +0,0 @@
-#pragma once
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-void ggml_critical_section_start(void);
-void ggml_critical_section_end(void);
-
-#ifdef __cplusplus
-}
-#endif

ggml/src/ggml-vulkan.cpp

@@ -106,15 +106,6 @@ struct vk_matmul_pipeline_struct {
 
 typedef std::shared_ptr<vk_matmul_pipeline_struct> vk_matmul_pipeline;
 
-struct vk_matmul_pipeline2 {
-    vk_matmul_pipeline2() {
-        f16acc = std::make_shared<vk_matmul_pipeline_struct>();
-        f32acc = std::make_shared<vk_matmul_pipeline_struct>();
-    }
-    vk_matmul_pipeline f32acc;
-    vk_matmul_pipeline f16acc;
-};
-
 struct vk_device_struct;
 typedef std::shared_ptr<vk_device_struct> vk_device;
 typedef std::weak_ptr<vk_device_struct> vk_device_ref;
@@ -170,11 +161,11 @@ struct vk_device_struct {
 
     vk_matmul_pipeline pipeline_matmul_f32;
     vk_matmul_pipeline pipeline_matmul_f32_f16;
-    vk_matmul_pipeline2 pipeline_matmul_f16;
+    vk_matmul_pipeline pipeline_matmul_f16;
-    vk_matmul_pipeline2 pipeline_matmul_f16_f32;
+    vk_matmul_pipeline pipeline_matmul_f16_f32;
     vk_pipeline pipeline_matmul_split_k_reduce;
 
-    vk_matmul_pipeline2 pipeline_dequant_mul_mat_mat[GGML_TYPE_COUNT];
+    vk_matmul_pipeline pipeline_dequant_mul_mat_mat[GGML_TYPE_COUNT];
 
     vk_matmul_pipeline pipeline_matmul_id_f32;
     vk_matmul_pipeline pipeline_matmul_id_f16;
@@ -192,10 +183,9 @@ struct vk_device_struct {
     vk_pipeline pipeline_get_rows[GGML_TYPE_COUNT];
     vk_pipeline pipeline_get_rows_f32[GGML_TYPE_COUNT];
     vk_pipeline pipeline_acc_f32;
-    vk_pipeline pipeline_add_f32, pipeline_add_f32_norepeat;
+    vk_pipeline pipeline_add_f32, pipeline_add_f16_f32_f16;
-    vk_pipeline pipeline_add_f16_f32_f16, pipeline_add_f16_f32_f16_norepeat;
+    vk_pipeline pipeline_mul_f32;
-    vk_pipeline pipeline_mul_f32, pipeline_mul_f32_norepeat;
+    vk_pipeline pipeline_div_f32;
-    vk_pipeline pipeline_div_f32, pipeline_div_f32_norepeat;
     vk_pipeline pipeline_concat_f32, pipeline_concat_f16, pipeline_concat_i32;
     vk_pipeline pipeline_upscale_f32;
     vk_pipeline pipeline_scale_f32;
@@ -218,7 +208,6 @@ struct vk_device_struct {
     vk_pipeline pipeline_tanh_f32;
     vk_pipeline pipeline_diag_mask_inf_f32;
     vk_pipeline pipeline_soft_max_f32, pipeline_soft_max_f32_f16;
-    vk_pipeline pipeline_soft_max_f32_wg512, pipeline_soft_max_f32_f16_wg512;
     vk_pipeline pipeline_rope_norm_f32, pipeline_rope_norm_f16;
     vk_pipeline pipeline_rope_neox_f32, pipeline_rope_neox_f16;
     vk_pipeline pipeline_argsort_f32;
@@ -389,7 +378,6 @@ struct vk_op_soft_max_push_constants {
     float m0;
     float m1;
     uint32_t n_head_log2;
-    uint32_t nrows_x;
 };
 
 struct vk_op_argsort_push_constants {
@@ -1222,28 +1210,37 @@ static void ggml_vk_load_shaders(vk_device& device) {
     std::cerr << "ggml_vulkan: Compiling shaders";
 
     // mulmat
-    std::vector<uint32_t> l_warptile, m_warptile, s_warptile, l_warptile_mmq, m_warptile_mmq, s_warptile_mmq;
+    std::initializer_list<uint32_t> warptile_l = { 128, 128, 128, 16, device->subgroup_size * 2, 64, 2, 4, 4, device->subgroup_size };
-    std::array<uint32_t, 3> l_wg_denoms, m_wg_denoms, s_wg_denoms;
+    std::initializer_list<uint32_t> warptile_m = { 128,  64,  64, 16, device->subgroup_size, 32, 2, 4, 2, device->subgroup_size };
-    uint32_t l_align, m_align, s_align;
+    std::initializer_list<uint32_t> warptile_s = { std::max(device->subgroup_size, 16u),  32,  32, 16, 32, 32, 2, 2, 2, device->subgroup_size };
 
-    l_warptile = { 128, 128, 128, 16, device->subgroup_size * 2, 64, 2, 4, 4, device->subgroup_size };
+    std::initializer_list<uint32_t> warptile_mmq_l = { 128, 128, 128, 32, device->subgroup_size * 2, 64, 2, 4, 4, device->subgroup_size };
-    m_warptile = { 128,  64,  64, 16, device->subgroup_size, 32, 2, 4, 2, device->subgroup_size };
+    std::initializer_list<uint32_t> warptile_mmq_m = { 128,  64,  64, 32, device->subgroup_size, 32, 2, 4, 2, device->subgroup_size };
-    s_warptile = { std::max(device->subgroup_size, 16u),  32,  32, 16, 32, 32, 2, 2, 2, device->subgroup_size };
+    std::initializer_list<uint32_t> warptile_mmq_s = { std::max(device->subgroup_size, 16u),  32,  32, 32, 32, 32, 2, 2, 2, device->subgroup_size };
 
-    l_warptile_mmq = { 128, 128, 128, 32, device->subgroup_size * 2, 64, 2, 4, 4, device->subgroup_size };
+    std::array<uint32_t, 3> l_wg_denoms = {128, 128, 1 };
-    m_warptile_mmq = { 128,  64,  64, 32, device->subgroup_size, 32, 2, 4, 2, device->subgroup_size };
+    std::array<uint32_t, 3> m_wg_denoms = { 64,  64, 1 };
-    s_warptile_mmq = { std::max(device->subgroup_size, 16u),  32,  32, 32, 32, 32, 2, 2, 2, device->subgroup_size };
+    std::array<uint32_t, 3> s_wg_denoms = { 32,  32, 1 };
 
-    l_wg_denoms = {128, 128, 1 };
+    uint32_t l_align = 128;
-    m_wg_denoms = { 64,  64, 1 };
+    uint32_t m_align =  64;
-    s_wg_denoms = { 32,  32, 1 };
+    uint32_t s_align =  32;
-
-    l_align = 128;
-    m_align =  64;
-    s_align =  32;
 
     device->pipeline_matmul_f32 = std::make_shared<vk_matmul_pipeline_struct>();
     device->pipeline_matmul_f32_f16 = std::make_shared<vk_matmul_pipeline_struct>();
+    device->pipeline_matmul_f16_f32 = std::make_shared<vk_matmul_pipeline_struct>();
+    device->pipeline_matmul_f16 = std::make_shared<vk_matmul_pipeline_struct>();
+    device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0] = std::make_shared<vk_matmul_pipeline_struct>();
+    device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1] = std::make_shared<vk_matmul_pipeline_struct>();
+    device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0] = std::make_shared<vk_matmul_pipeline_struct>();
+    device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1] = std::make_shared<vk_matmul_pipeline_struct>();
+    device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0] = std::make_shared<vk_matmul_pipeline_struct>();
+    device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K] = std::make_shared<vk_matmul_pipeline_struct>();
+    device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K] = std::make_shared<vk_matmul_pipeline_struct>();
+    device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K] = std::make_shared<vk_matmul_pipeline_struct>();
+    device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K] = std::make_shared<vk_matmul_pipeline_struct>();
+    device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K] = std::make_shared<vk_matmul_pipeline_struct>();
+    device->pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL] = std::make_shared<vk_matmul_pipeline_struct>();
 
     device->pipeline_matmul_id_f32 = std::make_shared<vk_matmul_pipeline_struct>();
     device->pipeline_matmul_id_f16_f32 = std::make_shared<vk_matmul_pipeline_struct>();
@@ -1261,7 +1258,7 @@ static void ggml_vk_load_shaders(vk_device& device) {
     device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL] = std::make_shared<vk_matmul_pipeline_struct>();
 
     std::vector<std::future<void>> compiles;
-    auto const &ggml_vk_create_pipeline = [&](vk_device& device, vk_pipeline& pipeline, const std::string &name, size_t spv_size, const void* spv_data, const std::string &entrypoint, uint32_t parameter_count, uint32_t push_constant_size, std::array<uint32_t, 3> wg_denoms, const std::vector<uint32_t>& specialization_constants, uint32_t align) {
+    auto const &ggml_vk_create_pipeline = [&](vk_device& device, vk_pipeline& pipeline, const std::string &name, size_t spv_size, const void* spv_data, const std::string &entrypoint, uint32_t parameter_count, uint32_t push_constant_size, std::array<uint32_t, 3> wg_denoms, std::vector<uint32_t>&& specialization_constants, uint32_t align) {
         {
             // wait until fewer than N compiles are in progress
             uint32_t N = std::max(1u, std::thread::hardware_concurrency());
@@ -1275,140 +1272,455 @@ static void ggml_vk_load_shaders(vk_device& device) {
     };
 
     if (device->fp16) {
-        // Create 6 variants, {s,m,l}x{unaligned,aligned}
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_f32->l, "matmul_f32_l", matmul_f32_f32_len, matmul_f32_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, 1);
-#define CREATE_MM(PIPELINE_NAME, NAMELC, F16ACC, WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT) \
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_f32->m, "matmul_f32_m", matmul_f32_f32_len, matmul_f32_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, 1);
-        ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->l, #NAMELC #F16ACC "_l", NAMELC ## F16ACC ## _len, NAMELC ## F16ACC ## _data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, 1);   \
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_f32->s, "matmul_f32_s", matmul_f32_f32_len, matmul_f32_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, 1);
-        ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->m, #NAMELC #F16ACC "_m", NAMELC ## F16ACC ## _len, NAMELC ## F16ACC ## _data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, 1);   \
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_f32->a_l, "matmul_f32_aligned_l", matmul_f32_f32_aligned_len, matmul_f32_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, l_align);
-        ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->s, #NAMELC #F16ACC "_s", NAMELC ## F16ACC ## _len, NAMELC ## F16ACC ## _data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, 1);   \
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_f32->a_m, "matmul_f32_aligned_m", matmul_f32_f32_aligned_len, matmul_f32_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, m_align);
-        ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_l, #NAMELC #F16ACC "_aligned_l", NAMELC ## _aligned ## F16ACC ## _len, NAMELC ## _aligned ## F16ACC ## _data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, l_align);   \
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_f32->a_s, "matmul_f32_aligned_s", matmul_f32_f32_aligned_len, matmul_f32_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, s_align);
-        ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_m, #NAMELC #F16ACC "_aligned_m", NAMELC ## _aligned ## F16ACC ## _len, NAMELC ## _aligned ## F16ACC ## _data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, m_align);   \
-        ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_s, #NAMELC #F16ACC "_aligned_s", NAMELC ## _aligned ## F16ACC ## _len, NAMELC ## _aligned ## F16ACC ## _data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, s_align);   \
 
-        CREATE_MM(pipeline_matmul_f32, matmul_f32_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 3);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_f32_f16->l, "matmul_f32_f16_l", matmul_f32_f16_len, matmul_f32_f16_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, 1);
-        CREATE_MM(pipeline_matmul_f32_f16, matmul_f32_f16, , wg_denoms, warptile, vk_mat_mat_push_constants, 3);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_f32_f16->m, "matmul_f32_f16_m", matmul_f32_f16_len, matmul_f32_f16_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, 1);
-        CREATE_MM(pipeline_matmul_f16.f32acc, matmul_f16, , wg_denoms, warptile, vk_mat_mat_push_constants, 3);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_f32_f16->s, "matmul_f32_f16_s", matmul_f32_f16_len, matmul_f32_f16_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, 1);
-        CREATE_MM(pipeline_matmul_f16_f32.f32acc, matmul_f16_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 3);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_f32_f16->a_l, "matmul_f32_f16_aligned_l", matmul_f32_f16_aligned_len, matmul_f32_f16_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_f32_f16->a_m, "matmul_f32_f16_aligned_m", matmul_f32_f16_aligned_len, matmul_f32_f16_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_f32_f16->a_s, "matmul_f32_f16_aligned_s", matmul_f32_f16_aligned_len, matmul_f32_f16_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, s_align);
 
-        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0].f32acc, matmul_q4_0_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_f16->l, "matmul_f16_l", matmul_f16_len, matmul_f16_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, 1);
-        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1].f32acc, matmul_q4_1_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_f16->m, "matmul_f16_m", matmul_f16_len, matmul_f16_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, 1);
-        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0].f32acc, matmul_q5_0_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_f16->s, "matmul_f16_s", matmul_f16_len, matmul_f16_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, 1);
-        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1].f32acc, matmul_q5_1_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_f16->a_l, "matmul_f16_aligned_l", matmul_f16_aligned_len, matmul_f16_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, l_align);
-        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0].f32acc, matmul_q8_0_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_f16->a_m, "matmul_f16_aligned_m", matmul_f16_aligned_len, matmul_f16_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_f16->a_s, "matmul_f16_aligned_s", matmul_f16_aligned_len, matmul_f16_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, s_align);
 
-        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K].f32acc, matmul_q2_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_f16_f32->l, "matmul_f16_f32_l", matmul_f16_f32_len, matmul_f16_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, 1);
-        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K].f32acc, matmul_q3_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_f16_f32->m, "matmul_f16_f32_m", matmul_f16_f32_len, matmul_f16_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, 1);
-        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K].f32acc, matmul_q4_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_f16_f32->s, "matmul_f16_f32_s", matmul_f16_f32_len, matmul_f16_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, 1);
-        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K].f32acc, matmul_q5_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_f16_f32->a_l, "matmul_f16_f32_aligned_l", matmul_f16_f32_aligned_len, matmul_f16_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, l_align);
-        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K].f32acc, matmul_q6_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_f16_f32->a_m, "matmul_f16_f32_aligned_m", matmul_f16_f32_aligned_len, matmul_f16_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, m_align);
-        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL].f32acc, matmul_iq4_nl_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_f16_f32->a_s, "matmul_f16_f32_aligned_s", matmul_f16_f32_aligned_len, matmul_f16_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, s_align);
 
-        CREATE_MM(pipeline_matmul_id_f32, matmul_id_f32_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 4);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0]->l, "matmul_q4_0_f32_l", matmul_q4_0_f32_len, matmul_q4_0_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        CREATE_MM(pipeline_matmul_id_f16, matmul_id_f16, , wg_denoms, warptile, vk_mat_mat_push_constants, 4);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0]->m, "matmul_q4_0_f32_m", matmul_q4_0_f32_len, matmul_q4_0_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        CREATE_MM(pipeline_matmul_id_f16_f32, matmul_id_f16_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 4);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0]->s, "matmul_q4_0_f32_s", matmul_q4_0_f32_len, matmul_q4_0_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0]->a_l, "matmul_q4_0_f32_aligned_l", matmul_q4_0_f32_aligned_len, matmul_q4_0_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0]->a_m, "matmul_q4_0_f32_aligned_m", matmul_q4_0_f32_aligned_len, matmul_q4_0_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0]->a_s, "matmul_q4_0_f32_aligned_s", matmul_q4_0_f32_aligned_len, matmul_q4_0_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
 
-        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0], matmul_id_q4_0_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1]->l, "matmul_q4_1_f32_l", matmul_q4_1_f32_len, matmul_q4_1_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1], matmul_id_q4_1_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1]->m, "matmul_q4_1_f32_m", matmul_q4_1_f32_len, matmul_q4_1_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0], matmul_id_q5_0_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1]->s, "matmul_q4_1_f32_s", matmul_q4_1_f32_len, matmul_q4_1_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1], matmul_id_q5_1_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1]->a_l, "matmul_q4_1_f32_aligned_l", matmul_q4_1_f32_aligned_len, matmul_q4_1_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0], matmul_id_q8_0_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1]->a_m, "matmul_q4_1_f32_aligned_m", matmul_q4_1_f32_aligned_len, matmul_q4_1_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1]->a_s, "matmul_q4_1_f32_aligned_s", matmul_q4_1_f32_aligned_len, matmul_q4_1_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
 
-        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K], matmul_id_q2_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0]->l, "matmul_q5_0_f32_l", matmul_q5_0_f32_len, matmul_q5_0_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K], matmul_id_q3_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0]->m, "matmul_q5_0_f32_m", matmul_q5_0_f32_len, matmul_q5_0_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K], matmul_id_q4_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0]->s, "matmul_q5_0_f32_s", matmul_q5_0_f32_len, matmul_q5_0_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K], matmul_id_q5_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0]->a_l, "matmul_q5_0_f32_aligned_l", matmul_q5_0_f32_aligned_len, matmul_q5_0_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K], matmul_id_q6_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0]->a_m, "matmul_q5_0_f32_aligned_m", matmul_q5_0_f32_aligned_len, matmul_q5_0_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL], matmul_id_iq4_nl_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0]->a_s, "matmul_q5_0_f32_aligned_s", matmul_q5_0_f32_aligned_len, matmul_q5_0_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-#undef CREATE_MM
+
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1]->l, "matmul_q5_1_f32_l", matmul_q5_1_f32_len, matmul_q5_1_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1]->m, "matmul_q5_1_f32_m", matmul_q5_1_f32_len, matmul_q5_1_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1]->s, "matmul_q5_1_f32_s", matmul_q5_1_f32_len, matmul_q5_1_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1]->a_l, "matmul_q5_1_f32_aligned_l", matmul_q5_1_f32_aligned_len, matmul_q5_1_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1]->a_m, "matmul_q5_1_f32_aligned_m", matmul_q5_1_f32_aligned_len, matmul_q5_1_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1]->a_s, "matmul_q5_1_f32_aligned_s", matmul_q5_1_f32_aligned_len, matmul_q5_1_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0]->l, "matmul_q8_0_f32_l", matmul_q8_0_f32_len, matmul_q8_0_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0]->m, "matmul_q8_0_f32_m", matmul_q8_0_f32_len, matmul_q8_0_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0]->s, "matmul_q8_0_f32_s", matmul_q8_0_f32_len, matmul_q8_0_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0]->a_l, "matmul_q8_0_f32_aligned_l", matmul_q8_0_f32_aligned_len, matmul_q8_0_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0]->a_m, "matmul_q8_0_f32_aligned_m", matmul_q8_0_f32_aligned_len, matmul_q8_0_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0]->a_s, "matmul_q8_0_f32_aligned_s", matmul_q8_0_f32_aligned_len, matmul_q8_0_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K]->l, "matmul_q2_k_f32_l", matmul_q2_k_f32_len, matmul_q2_k_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K]->m, "matmul_q2_k_f32_m", matmul_q2_k_f32_len, matmul_q2_k_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K]->s, "matmul_q2_k_f32_s", matmul_q2_k_f32_len, matmul_q2_k_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K]->a_l, "matmul_q2_k_f32_aligned_l", matmul_q2_k_f32_aligned_len, matmul_q2_k_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K]->a_m, "matmul_q2_k_f32_aligned_m", matmul_q2_k_f32_aligned_len, matmul_q2_k_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K]->a_s, "matmul_q2_k_f32_aligned_s", matmul_q2_k_f32_aligned_len, matmul_q2_k_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K]->l, "matmul_q3_k_f32_l", matmul_q3_k_f32_len, matmul_q3_k_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K]->m, "matmul_q3_k_f32_m", matmul_q3_k_f32_len, matmul_q3_k_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K]->s, "matmul_q3_k_f32_s", matmul_q3_k_f32_len, matmul_q3_k_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K]->a_l, "matmul_q3_k_f32_aligned_l", matmul_q3_k_f32_aligned_len, matmul_q3_k_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K]->a_m, "matmul_q3_k_f32_aligned_m", matmul_q3_k_f32_aligned_len, matmul_q3_k_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K]->a_s, "matmul_q3_k_f32_aligned_s", matmul_q3_k_f32_aligned_len, matmul_q3_k_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K]->l, "matmul_q4_k_f32_l", matmul_q4_k_f32_len, matmul_q4_k_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K]->m, "matmul_q4_k_f32_m", matmul_q4_k_f32_len, matmul_q4_k_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K]->s, "matmul_q4_k_f32_s", matmul_q4_k_f32_len, matmul_q4_k_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K]->a_l, "matmul_q4_k_f32_aligned_l", matmul_q4_k_f32_aligned_len, matmul_q4_k_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K]->a_m, "matmul_q4_k_f32_aligned_m", matmul_q4_k_f32_aligned_len, matmul_q4_k_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K]->a_s, "matmul_q4_k_f32_aligned_s", matmul_q4_k_f32_aligned_len, matmul_q4_k_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K]->l, "matmul_q5_k_f32_l", matmul_q5_k_f32_len, matmul_q5_k_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K]->m, "matmul_q5_k_f32_m", matmul_q5_k_f32_len, matmul_q5_k_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K]->s, "matmul_q5_k_f32_s", matmul_q5_k_f32_len, matmul_q5_k_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K]->a_l, "matmul_q5_k_f32_aligned_l", matmul_q5_k_f32_aligned_len, matmul_q5_k_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K]->a_m, "matmul_q5_k_f32_aligned_m", matmul_q5_k_f32_aligned_len, matmul_q5_k_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K]->a_s, "matmul_q5_k_f32_aligned_s", matmul_q5_k_f32_aligned_len, matmul_q5_k_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K]->l, "matmul_q6_k_f32_l", matmul_q6_k_f32_len, matmul_q6_k_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K]->m, "matmul_q6_k_f32_m", matmul_q6_k_f32_len, matmul_q6_k_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K]->s, "matmul_q6_k_f32_s", matmul_q6_k_f32_len, matmul_q6_k_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K]->a_l, "matmul_q6_k_f32_aligned_l", matmul_q6_k_f32_aligned_len, matmul_q6_k_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K]->a_m, "matmul_q6_k_f32_aligned_m", matmul_q6_k_f32_aligned_len, matmul_q6_k_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K]->a_s, "matmul_q6_k_f32_aligned_s", matmul_q6_k_f32_aligned_len, matmul_q6_k_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL]->l, "matmul_iq4_nl_f32_l", matmul_iq4_nl_f32_len, matmul_iq4_nl_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL]->m, "matmul_iq4_nl_f32_m", matmul_iq4_nl_f32_len, matmul_iq4_nl_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL]->s, "matmul_iq4_nl_f32_s", matmul_iq4_nl_f32_len, matmul_iq4_nl_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL]->a_l, "matmul_iq4_nl_f32_aligned_l", matmul_iq4_nl_f32_aligned_len, matmul_iq4_nl_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL]->a_m, "matmul_iq4_nl_f32_aligned_m", matmul_iq4_nl_f32_aligned_len, matmul_iq4_nl_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL]->a_s, "matmul_iq4_nl_f32_aligned_s", matmul_iq4_nl_f32_aligned_len, matmul_iq4_nl_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f32->l, "matmul_id_f32_l", matmul_id_f32_f32_len, matmul_id_f32_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_l, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f32->m, "matmul_id_f32_m", matmul_id_f32_f32_len, matmul_id_f32_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_m, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f32->s, "matmul_id_f32_s", matmul_id_f32_f32_len, matmul_id_f32_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_s, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f32->a_l, "matmul_id_f32_aligned_l", matmul_id_f32_f32_aligned_len, matmul_id_f32_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f32->a_m, "matmul_id_f32_aligned_m", matmul_id_f32_f32_aligned_len, matmul_id_f32_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f32->a_s, "matmul_id_f32_aligned_s", matmul_id_f32_f32_aligned_len, matmul_id_f32_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_s, s_align);
+
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16->l, "matmul_id_f16_l", matmul_id_f16_len, matmul_id_f16_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_l, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16->m, "matmul_id_f16_m", matmul_id_f16_len, matmul_id_f16_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_m, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16->s, "matmul_id_f16_s", matmul_id_f16_len, matmul_id_f16_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_s, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16->a_l, "matmul_id_f16_aligned_l", matmul_id_f16_aligned_len, matmul_id_f16_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16->a_m, "matmul_id_f16_aligned_m", matmul_id_f16_aligned_len, matmul_id_f16_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16->a_s, "matmul_id_f16_aligned_s", matmul_id_f16_aligned_len, matmul_id_f16_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_s, s_align);
+
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16_f32->l, "matmul_id_f16_f32_l", matmul_id_f16_f32_len, matmul_id_f16_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_l, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16_f32->m, "matmul_id_f16_f32_m", matmul_id_f16_f32_len, matmul_id_f16_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_m, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16_f32->s, "matmul_id_f16_f32_s", matmul_id_f16_f32_len, matmul_id_f16_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_s, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16_f32->a_l, "matmul_id_f16_f32_aligned_l", matmul_id_f16_f32_aligned_len, matmul_id_f16_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16_f32->a_m, "matmul_id_f16_f32_aligned_m", matmul_id_f16_f32_aligned_len, matmul_id_f16_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16_f32->a_s, "matmul_id_f16_f32_aligned_s", matmul_id_f16_f32_aligned_len, matmul_id_f16_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_s, s_align);
+
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0]->l, "matmul_id_q4_0_f32_l", matmul_id_q4_0_f32_len, matmul_id_q4_0_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0]->m, "matmul_id_q4_0_f32_m", matmul_id_q4_0_f32_len, matmul_id_q4_0_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0]->s, "matmul_id_q4_0_f32_s", matmul_id_q4_0_f32_len, matmul_id_q4_0_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0]->a_l, "matmul_id_q4_0_f32_aligned_l", matmul_id_q4_0_f32_aligned_len, matmul_id_q4_0_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0]->a_m, "matmul_id_q4_0_f32_aligned_m", matmul_id_q4_0_f32_aligned_len, matmul_id_q4_0_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0]->a_s, "matmul_id_q4_0_f32_aligned_s", matmul_id_q4_0_f32_aligned_len, matmul_id_q4_0_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1]->l, "matmul_id_q4_1_f32_l", matmul_id_q4_1_f32_len, matmul_id_q4_1_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1]->m, "matmul_id_q4_1_f32_m", matmul_id_q4_1_f32_len, matmul_id_q4_1_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1]->s, "matmul_id_q4_1_f32_s", matmul_id_q4_1_f32_len, matmul_id_q4_1_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1]->a_l, "matmul_id_q4_1_f32_aligned_l", matmul_id_q4_1_f32_aligned_len, matmul_id_q4_1_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1]->a_m, "matmul_id_q4_1_f32_aligned_m", matmul_id_q4_1_f32_aligned_len, matmul_id_q4_1_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1]->a_s, "matmul_id_q4_1_f32_aligned_s", matmul_id_q4_1_f32_aligned_len, matmul_id_q4_1_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0]->l, "matmul_id_q5_0_f32_l", matmul_id_q5_0_f32_len, matmul_id_q5_0_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0]->m, "matmul_id_q5_0_f32_m", matmul_id_q5_0_f32_len, matmul_id_q5_0_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0]->s, "matmul_id_q5_0_f32_s", matmul_id_q5_0_f32_len, matmul_id_q5_0_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0]->a_l, "matmul_id_q5_0_f32_aligned_l", matmul_id_q5_0_f32_aligned_len, matmul_id_q5_0_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0]->a_m, "matmul_id_q5_0_f32_aligned_m", matmul_id_q5_0_f32_aligned_len, matmul_id_q5_0_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0]->a_s, "matmul_id_q5_0_f32_aligned_s", matmul_id_q5_0_f32_aligned_len, matmul_id_q5_0_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1]->l, "matmul_id_q5_1_f32_l", matmul_id_q5_1_f32_len, matmul_id_q5_1_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1]->m, "matmul_id_q5_1_f32_m", matmul_id_q5_1_f32_len, matmul_id_q5_1_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1]->s, "matmul_id_q5_1_f32_s", matmul_id_q5_1_f32_len, matmul_id_q5_1_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1]->a_l, "matmul_id_q5_1_f32_aligned_l", matmul_id_q5_1_f32_aligned_len, matmul_id_q5_1_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1]->a_m, "matmul_id_q5_1_f32_aligned_m", matmul_id_q5_1_f32_aligned_len, matmul_id_q5_1_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1]->a_s, "matmul_id_q5_1_f32_aligned_s", matmul_id_q5_1_f32_aligned_len, matmul_id_q5_1_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0]->l, "matmul_id_q8_0_f32_l", matmul_id_q8_0_f32_len, matmul_id_q8_0_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0]->m, "matmul_id_q8_0_f32_m", matmul_id_q8_0_f32_len, matmul_id_q8_0_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0]->s, "matmul_id_q8_0_f32_s", matmul_id_q8_0_f32_len, matmul_id_q8_0_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0]->a_l, "matmul_id_q8_0_f32_aligned_l", matmul_id_q8_0_f32_aligned_len, matmul_id_q8_0_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0]->a_m, "matmul_id_q8_0_f32_aligned_m", matmul_id_q8_0_f32_aligned_len, matmul_id_q8_0_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0]->a_s, "matmul_id_q8_0_f32_aligned_s", matmul_id_q8_0_f32_aligned_len, matmul_id_q8_0_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K]->l, "matmul_id_q2_k_f32_l", matmul_id_q2_k_f32_len, matmul_id_q2_k_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K]->m, "matmul_id_q2_k_f32_m", matmul_id_q2_k_f32_len, matmul_id_q2_k_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K]->s, "matmul_id_q2_k_f32_s", matmul_id_q2_k_f32_len, matmul_id_q2_k_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K]->a_l, "matmul_id_q2_k_f32_aligned_l", matmul_id_q2_k_f32_aligned_len, matmul_id_q2_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K]->a_m, "matmul_id_q2_k_f32_aligned_m", matmul_id_q2_k_f32_aligned_len, matmul_id_q2_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K]->a_s, "matmul_id_q2_k_f32_aligned_s", matmul_id_q2_k_f32_aligned_len, matmul_id_q2_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K]->l, "matmul_id_q3_k_f32_l", matmul_id_q3_k_f32_len, matmul_id_q3_k_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K]->m, "matmul_id_q3_k_f32_m", matmul_id_q3_k_f32_len, matmul_id_q3_k_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K]->s, "matmul_id_q3_k_f32_s", matmul_id_q3_k_f32_len, matmul_id_q3_k_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K]->a_l, "matmul_id_q3_k_f32_aligned_l", matmul_id_q3_k_f32_aligned_len, matmul_id_q3_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K]->a_m, "matmul_id_q3_k_f32_aligned_m", matmul_id_q3_k_f32_aligned_len, matmul_id_q3_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K]->a_s, "matmul_id_q3_k_f32_aligned_s", matmul_id_q3_k_f32_aligned_len, matmul_id_q3_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K]->l, "matmul_id_q4_k_f32_l", matmul_id_q4_k_f32_len, matmul_id_q4_k_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K]->m, "matmul_id_q4_k_f32_m", matmul_id_q4_k_f32_len, matmul_id_q4_k_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K]->s, "matmul_id_q4_k_f32_s", matmul_id_q4_k_f32_len, matmul_id_q4_k_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K]->a_l, "matmul_id_q4_k_f32_aligned_l", matmul_id_q4_k_f32_aligned_len, matmul_id_q4_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K]->a_m, "matmul_id_q4_k_f32_aligned_m", matmul_id_q4_k_f32_aligned_len, matmul_id_q4_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K]->a_s, "matmul_id_q4_k_f32_aligned_s", matmul_id_q4_k_f32_aligned_len, matmul_id_q4_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K]->l, "matmul_id_q5_k_f32_l", matmul_id_q5_k_f32_len, matmul_id_q5_k_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K]->m, "matmul_id_q5_k_f32_m", matmul_id_q5_k_f32_len, matmul_id_q5_k_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K]->s, "matmul_id_q5_k_f32_s", matmul_id_q5_k_f32_len, matmul_id_q5_k_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K]->a_l, "matmul_id_q5_k_f32_aligned_l", matmul_id_q5_k_f32_aligned_len, matmul_id_q5_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K]->a_m, "matmul_id_q5_k_f32_aligned_m", matmul_id_q5_k_f32_aligned_len, matmul_id_q5_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K]->a_s, "matmul_id_q5_k_f32_aligned_s", matmul_id_q5_k_f32_aligned_len, matmul_id_q5_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K]->l, "matmul_id_q6_k_f32_l", matmul_id_q6_k_f32_len, matmul_id_q6_k_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K]->m, "matmul_id_q6_k_f32_m", matmul_id_q6_k_f32_len, matmul_id_q6_k_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K]->s, "matmul_id_q6_k_f32_s", matmul_id_q6_k_f32_len, matmul_id_q6_k_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K]->a_l, "matmul_id_q6_k_f32_aligned_l", matmul_id_q6_k_f32_aligned_len, matmul_id_q6_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K]->a_m, "matmul_id_q6_k_f32_aligned_m", matmul_id_q6_k_f32_aligned_len, matmul_id_q6_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K]->a_s, "matmul_id_q6_k_f32_aligned_s", matmul_id_q6_k_f32_aligned_len, matmul_id_q6_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL]->l, "matmul_id_iq4_nl_f32_l", matmul_id_iq4_nl_f32_len, matmul_id_iq4_nl_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL]->m, "matmul_id_iq4_nl_f32_m", matmul_id_iq4_nl_f32_len, matmul_id_iq4_nl_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL]->s, "matmul_id_iq4_nl_f32_s", matmul_id_iq4_nl_f32_len, matmul_id_iq4_nl_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL]->a_l, "matmul_id_iq4_nl_f32_aligned_l", matmul_id_iq4_nl_f32_aligned_len, matmul_id_iq4_nl_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL]->a_m, "matmul_id_iq4_nl_f32_aligned_m", matmul_id_iq4_nl_f32_aligned_len, matmul_id_iq4_nl_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL]->a_s, "matmul_id_iq4_nl_f32_aligned_s", matmul_id_iq4_nl_f32_aligned_len, matmul_id_iq4_nl_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
     } else {
-        // Create 6 variants, {s,m,l}x{unaligned,aligned}
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_f32->l, "matmul_f32_l", matmul_f32_f32_fp32_len, matmul_f32_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, 1);
-#define CREATE_MM(PIPELINE_NAME, NAMELC, F16ACC, WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT) \
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_f32->m, "matmul_f32_m", matmul_f32_f32_fp32_len, matmul_f32_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, 1);
-        ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->l, #NAMELC #F16ACC "_l", NAMELC ## F16ACC ## _fp32_len, NAMELC ## F16ACC ## _fp32_data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, 1);   \
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_f32->s, "matmul_f32_s", matmul_f32_f32_fp32_len, matmul_f32_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, 1);
-        ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->m, #NAMELC #F16ACC "_m", NAMELC ## F16ACC ## _fp32_len, NAMELC ## F16ACC ## _fp32_data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, 1);   \
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_f32->a_l, "matmul_f32_aligned_l", matmul_f32_f32_aligned_fp32_len, matmul_f32_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, l_align);
-        ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->s, #NAMELC #F16ACC "_s", NAMELC ## F16ACC ## _fp32_len, NAMELC ## F16ACC ## _fp32_data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, 1);   \
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_f32->a_m, "matmul_f32_aligned_m", matmul_f32_f32_aligned_fp32_len, matmul_f32_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, m_align);
-        ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_l, #NAMELC #F16ACC "_aligned_l", NAMELC ## _aligned ## F16ACC ## _fp32_len, NAMELC ## _aligned ## F16ACC ## _fp32_data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, l_align);   \
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_f32->a_s, "matmul_f32_aligned_s", matmul_f32_f32_aligned_fp32_len, matmul_f32_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, s_align);
-        ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_m, #NAMELC #F16ACC "_aligned_m", NAMELC ## _aligned ## F16ACC ## _fp32_len, NAMELC ## _aligned ## F16ACC ## _fp32_data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, m_align);   \
-        ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_s, #NAMELC #F16ACC "_aligned_s", NAMELC ## _aligned ## F16ACC ## _fp32_len, NAMELC ## _aligned ## F16ACC ## _fp32_data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, s_align);   \
 
-        CREATE_MM(pipeline_matmul_f32, matmul_f32_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 3);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_f32_f16->l, "matmul_f32_f16_l", matmul_f32_f16_fp32_len, matmul_f32_f16_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, 1);
-        CREATE_MM(pipeline_matmul_f32_f16, matmul_f32_f16, , wg_denoms, warptile, vk_mat_mat_push_constants, 3);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_f32_f16->m, "matmul_f32_f16_m", matmul_f32_f16_fp32_len, matmul_f32_f16_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, 1);
-        CREATE_MM(pipeline_matmul_f16.f32acc, matmul_f16, , wg_denoms, warptile, vk_mat_mat_push_constants, 3);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_f32_f16->s, "matmul_f32_f16_s", matmul_f32_f16_fp32_len, matmul_f32_f16_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, 1);
-        CREATE_MM(pipeline_matmul_f16_f32.f32acc, matmul_f16_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 3);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_f32_f16->a_l, "matmul_f32_f16_aligned_l", matmul_f32_f16_aligned_fp32_len, matmul_f32_f16_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_f32_f16->a_m, "matmul_f32_f16_aligned_m", matmul_f32_f16_aligned_fp32_len, matmul_f32_f16_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_f32_f16->a_s, "matmul_f32_f16_aligned_s", matmul_f32_f16_aligned_fp32_len, matmul_f32_f16_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, s_align);
 
-        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0].f32acc, matmul_q4_0_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_f16->l, "matmul_f16_l", matmul_f16_fp32_len, matmul_f16_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, 1);
-        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1].f32acc, matmul_q4_1_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_f16->m, "matmul_f16_m", matmul_f16_fp32_len, matmul_f16_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, 1);
-        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0].f32acc, matmul_q5_0_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_f16->s, "matmul_f16_s", matmul_f16_fp32_len, matmul_f16_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, 1);
-        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1].f32acc, matmul_q5_1_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_f16->a_l, "matmul_f16_aligned_l", matmul_f16_aligned_fp32_len, matmul_f16_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, l_align);
-        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0].f32acc, matmul_q8_0_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_f16->a_m, "matmul_f16_aligned_m", matmul_f16_aligned_fp32_len, matmul_f16_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_f16->a_s, "matmul_f16_aligned_s", matmul_f16_aligned_fp32_len, matmul_f16_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, s_align);
 
-        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K].f32acc, matmul_q2_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_f16_f32->l, "matmul_f16_f32_l", matmul_f16_f32_fp32_len, matmul_f16_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, 1);
-        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K].f32acc, matmul_q3_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_f16_f32->m, "matmul_f16_f32_m", matmul_f16_f32_fp32_len, matmul_f16_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, 1);
-        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K].f32acc, matmul_q4_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_f16_f32->s, "matmul_f16_f32_s", matmul_f16_f32_fp32_len, matmul_f16_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, 1);
-        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K].f32acc, matmul_q5_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_f16_f32->a_l, "matmul_f16_f32_aligned_l", matmul_f16_f32_aligned_fp32_len, matmul_f16_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, l_align);
-        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K].f32acc, matmul_q6_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_f16_f32->a_m, "matmul_f16_f32_aligned_m", matmul_f16_f32_aligned_fp32_len, matmul_f16_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, m_align);
-        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL].f32acc, matmul_iq4_nl_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_f16_f32->a_s, "matmul_f16_f32_aligned_s", matmul_f16_f32_aligned_fp32_len, matmul_f16_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, s_align);
 
-        CREATE_MM(pipeline_matmul_id_f32, matmul_id_f32_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 4);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0]->l, "matmul_q4_0_f32_l", matmul_q4_0_f32_fp32_len, matmul_q4_0_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        CREATE_MM(pipeline_matmul_id_f16, matmul_id_f16, , wg_denoms, warptile, vk_mat_mat_push_constants, 4);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0]->m, "matmul_q4_0_f32_m", matmul_q4_0_f32_fp32_len, matmul_q4_0_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        CREATE_MM(pipeline_matmul_id_f16_f32, matmul_id_f16_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 4);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0]->s, "matmul_q4_0_f32_s", matmul_q4_0_f32_fp32_len, matmul_q4_0_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0]->a_l, "matmul_q4_0_f32_aligned_l", matmul_q4_0_f32_aligned_fp32_len, matmul_q4_0_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0]->a_m, "matmul_q4_0_f32_aligned_m", matmul_q4_0_f32_aligned_fp32_len, matmul_q4_0_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0]->a_s, "matmul_q4_0_f32_aligned_s", matmul_q4_0_f32_aligned_fp32_len, matmul_q4_0_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
 
-        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0], matmul_id_q4_0_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1]->l, "matmul_q4_1_f32_l", matmul_q4_1_f32_fp32_len, matmul_q4_1_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1], matmul_id_q4_1_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1]->m, "matmul_q4_1_f32_m", matmul_q4_1_f32_fp32_len, matmul_q4_1_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0], matmul_id_q5_0_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1]->s, "matmul_q4_1_f32_s", matmul_q4_1_f32_fp32_len, matmul_q4_1_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1], matmul_id_q5_1_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1]->a_l, "matmul_q4_1_f32_aligned_l", matmul_q4_1_f32_aligned_fp32_len, matmul_q4_1_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0], matmul_id_q8_0_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1]->a_m, "matmul_q4_1_f32_aligned_m", matmul_q4_1_f32_aligned_fp32_len, matmul_q4_1_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1]->a_s, "matmul_q4_1_f32_aligned_s", matmul_q4_1_f32_aligned_fp32_len, matmul_q4_1_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
 
-        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K], matmul_id_q2_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0]->l, "matmul_q5_0_f32_l", matmul_q5_0_f32_fp32_len, matmul_q5_0_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K], matmul_id_q3_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0]->m, "matmul_q5_0_f32_m", matmul_q5_0_f32_fp32_len, matmul_q5_0_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K], matmul_id_q4_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0]->s, "matmul_q5_0_f32_s", matmul_q5_0_f32_fp32_len, matmul_q5_0_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K], matmul_id_q5_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0]->a_l, "matmul_q5_0_f32_aligned_l", matmul_q5_0_f32_aligned_fp32_len, matmul_q5_0_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K], matmul_id_q6_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0]->a_m, "matmul_q5_0_f32_aligned_m", matmul_q5_0_f32_aligned_fp32_len, matmul_q5_0_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL], matmul_id_iq4_nl_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0]->a_s, "matmul_q5_0_f32_aligned_s", matmul_q5_0_f32_aligned_fp32_len, matmul_q5_0_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-#undef CREATE_MM
+
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1]->l, "matmul_q5_1_f32_l", matmul_q5_1_f32_fp32_len, matmul_q5_1_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1]->m, "matmul_q5_1_f32_m", matmul_q5_1_f32_fp32_len, matmul_q5_1_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1]->s, "matmul_q5_1_f32_s", matmul_q5_1_f32_fp32_len, matmul_q5_1_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1]->a_l, "matmul_q5_1_f32_aligned_l", matmul_q5_1_f32_aligned_fp32_len, matmul_q5_1_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1]->a_m, "matmul_q5_1_f32_aligned_m", matmul_q5_1_f32_aligned_fp32_len, matmul_q5_1_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1]->a_s, "matmul_q5_1_f32_aligned_s", matmul_q5_1_f32_aligned_fp32_len, matmul_q5_1_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0]->l, "matmul_q8_0_f32_l", matmul_q8_0_f32_fp32_len, matmul_q8_0_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0]->m, "matmul_q8_0_f32_m", matmul_q8_0_f32_fp32_len, matmul_q8_0_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0]->s, "matmul_q8_0_f32_s", matmul_q8_0_f32_fp32_len, matmul_q8_0_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0]->a_l, "matmul_q8_0_f32_aligned_l", matmul_q8_0_f32_aligned_fp32_len, matmul_q8_0_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0]->a_m, "matmul_q8_0_f32_aligned_m", matmul_q8_0_f32_aligned_fp32_len, matmul_q8_0_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0]->a_s, "matmul_q8_0_f32_aligned_s", matmul_q8_0_f32_aligned_fp32_len, matmul_q8_0_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K]->l, "matmul_q2_k_f32_l", matmul_q2_k_f32_fp32_len, matmul_q2_k_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K]->m, "matmul_q2_k_f32_m", matmul_q2_k_f32_fp32_len, matmul_q2_k_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K]->s, "matmul_q2_k_f32_s", matmul_q2_k_f32_fp32_len, matmul_q2_k_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K]->a_l, "matmul_q2_k_f32_aligned_l", matmul_q2_k_f32_aligned_fp32_len, matmul_q2_k_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K]->a_m, "matmul_q2_k_f32_aligned_m", matmul_q2_k_f32_aligned_fp32_len, matmul_q2_k_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K]->a_s, "matmul_q2_k_f32_aligned_s", matmul_q2_k_f32_aligned_fp32_len, matmul_q2_k_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K]->l, "matmul_q3_k_f32_l", matmul_q3_k_f32_fp32_len, matmul_q3_k_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K]->m, "matmul_q3_k_f32_m", matmul_q3_k_f32_fp32_len, matmul_q3_k_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K]->s, "matmul_q3_k_f32_s", matmul_q3_k_f32_fp32_len, matmul_q3_k_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K]->a_l, "matmul_q3_k_f32_aligned_l", matmul_q3_k_f32_aligned_fp32_len, matmul_q3_k_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K]->a_m, "matmul_q3_k_f32_aligned_m", matmul_q3_k_f32_aligned_fp32_len, matmul_q3_k_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K]->a_s, "matmul_q3_k_f32_aligned_s", matmul_q3_k_f32_aligned_fp32_len, matmul_q3_k_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K]->l, "matmul_q4_k_f32_l", matmul_q4_k_f32_fp32_len, matmul_q4_k_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K]->m, "matmul_q4_k_f32_m", matmul_q4_k_f32_fp32_len, matmul_q4_k_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K]->s, "matmul_q4_k_f32_s", matmul_q4_k_f32_fp32_len, matmul_q4_k_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K]->a_l, "matmul_q4_k_f32_aligned_l", matmul_q4_k_f32_aligned_fp32_len, matmul_q4_k_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K]->a_m, "matmul_q4_k_f32_aligned_m", matmul_q4_k_f32_aligned_fp32_len, matmul_q4_k_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K]->a_s, "matmul_q4_k_f32_aligned_s", matmul_q4_k_f32_aligned_fp32_len, matmul_q4_k_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K]->l, "matmul_q5_k_f32_l", matmul_q5_k_f32_fp32_len, matmul_q5_k_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K]->m, "matmul_q5_k_f32_m", matmul_q5_k_f32_fp32_len, matmul_q5_k_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K]->s, "matmul_q5_k_f32_s", matmul_q5_k_f32_fp32_len, matmul_q5_k_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K]->a_l, "matmul_q5_k_f32_aligned_l", matmul_q5_k_f32_aligned_fp32_len, matmul_q5_k_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K]->a_m, "matmul_q5_k_f32_aligned_m", matmul_q5_k_f32_aligned_fp32_len, matmul_q5_k_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K]->a_s, "matmul_q5_k_f32_aligned_s", matmul_q5_k_f32_aligned_fp32_len, matmul_q5_k_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K]->l, "matmul_q6_k_f32_l", matmul_q6_k_f32_fp32_len, matmul_q6_k_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K]->m, "matmul_q6_k_f32_m", matmul_q6_k_f32_fp32_len, matmul_q6_k_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K]->s, "matmul_q6_k_f32_s", matmul_q6_k_f32_fp32_len, matmul_q6_k_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K]->a_l, "matmul_q6_k_f32_aligned_l", matmul_q6_k_f32_aligned_fp32_len, matmul_q6_k_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K]->a_m, "matmul_q6_k_f32_aligned_m", matmul_q6_k_f32_aligned_fp32_len, matmul_q6_k_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K]->a_s, "matmul_q6_k_f32_aligned_s", matmul_q6_k_f32_aligned_fp32_len, matmul_q6_k_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL]->l, "matmul_iq4_nl_f32_l", matmul_iq4_nl_f32_fp32_len, matmul_iq4_nl_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL]->m, "matmul_iq4_nl_f32_m", matmul_iq4_nl_f32_fp32_len, matmul_iq4_nl_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL]->s, "matmul_iq4_nl_f32_s", matmul_iq4_nl_f32_fp32_len, matmul_iq4_nl_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL]->a_l, "matmul_iq4_nl_f32_aligned_l", matmul_iq4_nl_f32_aligned_fp32_len, matmul_iq4_nl_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL]->a_m, "matmul_iq4_nl_f32_aligned_m", matmul_iq4_nl_f32_aligned_fp32_len, matmul_iq4_nl_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL]->a_s, "matmul_iq4_nl_f32_aligned_s", matmul_iq4_nl_f32_aligned_fp32_len, matmul_iq4_nl_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f32->l, "matmul_id_f32_l", matmul_id_f32_f32_fp32_len, matmul_id_f32_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_l, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f32->m, "matmul_id_f32_m", matmul_id_f32_f32_fp32_len, matmul_id_f32_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_m, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f32->s, "matmul_id_f32_s", matmul_id_f32_f32_fp32_len, matmul_id_f32_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_s, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f32->a_l, "matmul_id_f32_aligned_l", matmul_id_f32_f32_aligned_fp32_len, matmul_id_f32_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f32->a_m, "matmul_id_f32_aligned_m", matmul_id_f32_f32_aligned_fp32_len, matmul_id_f32_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f32->a_s, "matmul_id_f32_aligned_s", matmul_id_f32_f32_aligned_fp32_len, matmul_id_f32_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_s, s_align);
+
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16->l, "matmul_id_f16_l", matmul_id_f16_fp32_len, matmul_id_f16_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_l, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16->m, "matmul_id_f16_m", matmul_id_f16_fp32_len, matmul_id_f16_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_m, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16->s, "matmul_id_f16_s", matmul_id_f16_fp32_len, matmul_id_f16_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_s, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16->a_l, "matmul_id_f16_aligned_l", matmul_id_f16_aligned_fp32_len, matmul_id_f16_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16->a_m, "matmul_id_f16_aligned_m", matmul_id_f16_aligned_fp32_len, matmul_id_f16_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16->a_s, "matmul_id_f16_aligned_s", matmul_id_f16_aligned_fp32_len, matmul_id_f16_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_s, s_align);
+
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16_f32->l, "matmul_id_f16_f32_l", matmul_id_f16_f32_fp32_len, matmul_id_f16_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_l, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16_f32->m, "matmul_id_f16_f32_m", matmul_id_f16_f32_fp32_len, matmul_id_f16_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_m, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16_f32->s, "matmul_id_f16_f32_s", matmul_id_f16_f32_fp32_len, matmul_id_f16_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_s, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16_f32->a_l, "matmul_id_f16_f32_aligned_l", matmul_id_f16_f32_aligned_fp32_len, matmul_id_f16_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16_f32->a_m, "matmul_id_f16_f32_aligned_m", matmul_id_f16_f32_aligned_fp32_len, matmul_id_f16_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16_f32->a_s, "matmul_id_f16_f32_aligned_s", matmul_id_f16_f32_aligned_fp32_len, matmul_id_f16_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_s, s_align);
+
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0]->l, "matmul_id_q4_0_f32_l", matmul_id_q4_0_f32_fp32_len, matmul_id_q4_0_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0]->m, "matmul_id_q4_0_f32_m", matmul_id_q4_0_f32_fp32_len, matmul_id_q4_0_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0]->s, "matmul_id_q4_0_f32_s", matmul_id_q4_0_f32_fp32_len, matmul_id_q4_0_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0]->a_l, "matmul_id_q4_0_f32_aligned_l", matmul_id_q4_0_f32_aligned_fp32_len, matmul_id_q4_0_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0]->a_m, "matmul_id_q4_0_f32_aligned_m", matmul_id_q4_0_f32_aligned_fp32_len, matmul_id_q4_0_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0]->a_s, "matmul_id_q4_0_f32_aligned_s", matmul_id_q4_0_f32_aligned_fp32_len, matmul_id_q4_0_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1]->l, "matmul_id_q4_1_f32_l", matmul_id_q4_1_f32_fp32_len, matmul_id_q4_1_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1]->m, "matmul_id_q4_1_f32_m", matmul_id_q4_1_f32_fp32_len, matmul_id_q4_1_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1]->s, "matmul_id_q4_1_f32_s", matmul_id_q4_1_f32_fp32_len, matmul_id_q4_1_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1]->a_l, "matmul_id_q4_1_f32_aligned_l", matmul_id_q4_1_f32_aligned_fp32_len, matmul_id_q4_1_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1]->a_m, "matmul_id_q4_1_f32_aligned_m", matmul_id_q4_1_f32_aligned_fp32_len, matmul_id_q4_1_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1]->a_s, "matmul_id_q4_1_f32_aligned_s", matmul_id_q4_1_f32_aligned_fp32_len, matmul_id_q4_1_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0]->l, "matmul_id_q5_0_f32_l", matmul_id_q5_0_f32_fp32_len, matmul_id_q5_0_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0]->m, "matmul_id_q5_0_f32_m", matmul_id_q5_0_f32_fp32_len, matmul_id_q5_0_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0]->s, "matmul_id_q5_0_f32_s", matmul_id_q5_0_f32_fp32_len, matmul_id_q5_0_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0]->a_l, "matmul_id_q5_0_f32_aligned_l", matmul_id_q5_0_f32_aligned_fp32_len, matmul_id_q5_0_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0]->a_m, "matmul_id_q5_0_f32_aligned_m", matmul_id_q5_0_f32_aligned_fp32_len, matmul_id_q5_0_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0]->a_s, "matmul_id_q5_0_f32_aligned_s", matmul_id_q5_0_f32_aligned_fp32_len, matmul_id_q5_0_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1]->l, "matmul_id_q5_1_f32_l", matmul_id_q5_1_f32_fp32_len, matmul_id_q5_1_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1]->m, "matmul_id_q5_1_f32_m", matmul_id_q5_1_f32_fp32_len, matmul_id_q5_1_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1]->s, "matmul_id_q5_1_f32_s", matmul_id_q5_1_f32_fp32_len, matmul_id_q5_1_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1]->a_l, "matmul_id_q5_1_f32_aligned_l", matmul_id_q5_1_f32_aligned_fp32_len, matmul_id_q5_1_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1]->a_m, "matmul_id_q5_1_f32_aligned_m", matmul_id_q5_1_f32_aligned_fp32_len, matmul_id_q5_1_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1]->a_s, "matmul_id_q5_1_f32_aligned_s", matmul_id_q5_1_f32_aligned_fp32_len, matmul_id_q5_1_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0]->l, "matmul_id_q8_0_f32_l", matmul_id_q8_0_f32_fp32_len, matmul_id_q8_0_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0]->m, "matmul_id_q8_0_f32_m", matmul_id_q8_0_f32_fp32_len, matmul_id_q8_0_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0]->s, "matmul_id_q8_0_f32_s", matmul_id_q8_0_f32_fp32_len, matmul_id_q8_0_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0]->a_l, "matmul_id_q8_0_f32_aligned_l", matmul_id_q8_0_f32_aligned_fp32_len, matmul_id_q8_0_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0]->a_m, "matmul_id_q8_0_f32_aligned_m", matmul_id_q8_0_f32_aligned_fp32_len, matmul_id_q8_0_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0]->a_s, "matmul_id_q8_0_f32_aligned_s", matmul_id_q8_0_f32_aligned_fp32_len, matmul_id_q8_0_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K]->l, "matmul_id_q2_k_f32_l", matmul_id_q2_k_f32_fp32_len, matmul_id_q2_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K]->m, "matmul_id_q2_k_f32_m", matmul_id_q2_k_f32_fp32_len, matmul_id_q2_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K]->s, "matmul_id_q2_k_f32_s", matmul_id_q2_k_f32_fp32_len, matmul_id_q2_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K]->a_l, "matmul_id_q2_k_f32_aligned_l", matmul_id_q2_k_f32_aligned_fp32_len, matmul_id_q2_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K]->a_m, "matmul_id_q2_k_f32_aligned_m", matmul_id_q2_k_f32_aligned_fp32_len, matmul_id_q2_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K]->a_s, "matmul_id_q2_k_f32_aligned_s", matmul_id_q2_k_f32_aligned_fp32_len, matmul_id_q2_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K]->l, "matmul_id_q3_k_f32_l", matmul_id_q3_k_f32_fp32_len, matmul_id_q3_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K]->m, "matmul_id_q3_k_f32_m", matmul_id_q3_k_f32_fp32_len, matmul_id_q3_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K]->s, "matmul_id_q3_k_f32_s", matmul_id_q3_k_f32_fp32_len, matmul_id_q3_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K]->a_l, "matmul_id_q3_k_f32_aligned_l", matmul_id_q3_k_f32_aligned_fp32_len, matmul_id_q3_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K]->a_m, "matmul_id_q3_k_f32_aligned_m", matmul_id_q3_k_f32_aligned_fp32_len, matmul_id_q3_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K]->a_s, "matmul_id_q3_k_f32_aligned_s", matmul_id_q3_k_f32_aligned_fp32_len, matmul_id_q3_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K]->l, "matmul_id_q4_k_f32_l", matmul_id_q4_k_f32_fp32_len, matmul_id_q4_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K]->m, "matmul_id_q4_k_f32_m", matmul_id_q4_k_f32_fp32_len, matmul_id_q4_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K]->s, "matmul_id_q4_k_f32_s", matmul_id_q4_k_f32_fp32_len, matmul_id_q4_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K]->a_l, "matmul_id_q4_k_f32_aligned_l", matmul_id_q4_k_f32_aligned_fp32_len, matmul_id_q4_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K]->a_m, "matmul_id_q4_k_f32_aligned_m", matmul_id_q4_k_f32_aligned_fp32_len, matmul_id_q4_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K]->a_s, "matmul_id_q4_k_f32_aligned_s", matmul_id_q4_k_f32_aligned_fp32_len, matmul_id_q4_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K]->l, "matmul_id_q5_k_f32_l", matmul_id_q5_k_f32_fp32_len, matmul_id_q5_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K]->m, "matmul_id_q5_k_f32_m", matmul_id_q5_k_f32_fp32_len, matmul_id_q5_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K]->s, "matmul_id_q5_k_f32_s", matmul_id_q5_k_f32_fp32_len, matmul_id_q5_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K]->a_l, "matmul_id_q5_k_f32_aligned_l", matmul_id_q5_k_f32_aligned_fp32_len, matmul_id_q5_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K]->a_m, "matmul_id_q5_k_f32_aligned_m", matmul_id_q5_k_f32_aligned_fp32_len, matmul_id_q5_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K]->a_s, "matmul_id_q5_k_f32_aligned_s", matmul_id_q5_k_f32_aligned_fp32_len, matmul_id_q5_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K]->l, "matmul_id_q6_k_f32_l", matmul_id_q6_k_f32_fp32_len, matmul_id_q6_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K]->m, "matmul_id_q6_k_f32_m", matmul_id_q6_k_f32_fp32_len, matmul_id_q6_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K]->s, "matmul_id_q6_k_f32_s", matmul_id_q6_k_f32_fp32_len, matmul_id_q6_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K]->a_l, "matmul_id_q6_k_f32_aligned_l", matmul_id_q6_k_f32_aligned_fp32_len, matmul_id_q6_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K]->a_m, "matmul_id_q6_k_f32_aligned_m", matmul_id_q6_k_f32_aligned_fp32_len, matmul_id_q6_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K]->a_s, "matmul_id_q6_k_f32_aligned_s", matmul_id_q6_k_f32_aligned_fp32_len, matmul_id_q6_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL]->l, "matmul_id_iq4_nl_f32_l", matmul_id_iq4_nl_f32_fp32_len, matmul_id_iq4_nl_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL]->m, "matmul_id_iq4_nl_f32_m", matmul_id_iq4_nl_f32_fp32_len, matmul_id_iq4_nl_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL]->s, "matmul_id_iq4_nl_f32_s", matmul_id_iq4_nl_f32_fp32_len, matmul_id_iq4_nl_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL]->a_l, "matmul_id_iq4_nl_f32_aligned_l", matmul_id_iq4_nl_f32_aligned_fp32_len, matmul_id_iq4_nl_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL]->a_m, "matmul_id_iq4_nl_f32_aligned_m", matmul_id_iq4_nl_f32_aligned_fp32_len, matmul_id_iq4_nl_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL]->a_s, "matmul_id_iq4_nl_f32_aligned_s", matmul_id_iq4_nl_f32_aligned_fp32_len, matmul_id_iq4_nl_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
     }
 
     // mul mat vec
-    // computing two rows per workgroup is a benefit for Q4_0 -> Q5_1, but not for Q8_0.
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_F32 ], "mul_mat_vec_f32_f32_f32",  mul_mat_vec_f32_f32_f32_len,  mul_mat_vec_f32_f32_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_F32 ], "mul_mat_vec_f32_f32_f32",  mul_mat_vec_f32_f32_f32_len,  mul_mat_vec_f32_f32_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_F16 ], "mul_mat_vec_f16_f32_f32",  mul_mat_vec_f16_f32_f32_len,  mul_mat_vec_f16_f32_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_F16 ], "mul_mat_vec_f16_f32_f32",  mul_mat_vec_f16_f32_f32_len,  mul_mat_vec_f16_f32_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q4_0], "mul_mat_vec_q4_0_f32_f32", mul_mat_vec_q4_0_f32_f32_len, mul_mat_vec_q4_0_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q4_0], "mul_mat_vec_q4_0_f32_f32", mul_mat_vec_q4_0_f32_f32_len, mul_mat_vec_q4_0_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q4_1], "mul_mat_vec_q4_1_f32_f32", mul_mat_vec_q4_1_f32_f32_len, mul_mat_vec_q4_1_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q4_1], "mul_mat_vec_q4_1_f32_f32", mul_mat_vec_q4_1_f32_f32_len, mul_mat_vec_q4_1_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q5_0], "mul_mat_vec_q5_0_f32_f32", mul_mat_vec_q5_0_f32_f32_len, mul_mat_vec_q5_0_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q5_0], "mul_mat_vec_q5_0_f32_f32", mul_mat_vec_q5_0_f32_f32_len, mul_mat_vec_q5_0_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q5_1], "mul_mat_vec_q5_1_f32_f32", mul_mat_vec_q5_1_f32_f32_len, mul_mat_vec_q5_1_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q5_1], "mul_mat_vec_q5_1_f32_f32", mul_mat_vec_q5_1_f32_f32_len, mul_mat_vec_q5_1_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q8_0], "mul_mat_vec_q8_0_f32_f32", mul_mat_vec_q8_0_f32_f32_len, mul_mat_vec_q8_0_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q8_0], "mul_mat_vec_q8_0_f32_f32", mul_mat_vec_q8_0_f32_f32_len, mul_mat_vec_q8_0_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size, 1}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q2_K], "mul_mat_vec_q2_k_f32_f32", mul_mat_vec_q2_k_f32_f32_len, mul_mat_vec_q2_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q2_K], "mul_mat_vec_q2_k_f32_f32", mul_mat_vec_q2_k_f32_f32_len, mul_mat_vec_q2_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q3_K], "mul_mat_vec_q3_k_f32_f32", mul_mat_vec_q3_k_f32_f32_len, mul_mat_vec_q3_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q3_K], "mul_mat_vec_q3_k_f32_f32", mul_mat_vec_q3_k_f32_f32_len, mul_mat_vec_q3_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q4_K], "mul_mat_vec_q4_k_f32_f32", mul_mat_vec_q4_k_f32_f32_len, mul_mat_vec_q4_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q4_K], "mul_mat_vec_q4_k_f32_f32", mul_mat_vec_q4_k_f32_f32_len, mul_mat_vec_q4_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q5_K], "mul_mat_vec_q5_k_f32_f32", mul_mat_vec_q5_k_f32_f32_len, mul_mat_vec_q5_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q5_K], "mul_mat_vec_q5_k_f32_f32", mul_mat_vec_q5_k_f32_f32_len, mul_mat_vec_q5_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q6_K], "mul_mat_vec_q6_k_f32_f32", mul_mat_vec_q6_k_f32_f32_len, mul_mat_vec_q6_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q6_K], "mul_mat_vec_q6_k_f32_f32", mul_mat_vec_q6_k_f32_f32_len, mul_mat_vec_q6_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_IQ4_NL], "mul_mat_vec_iq4_nl_f32_f32", mul_mat_vec_iq4_nl_f32_f32_len, mul_mat_vec_iq4_nl_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_IQ4_NL], "mul_mat_vec_iq4_nl_f32_f32", mul_mat_vec_iq4_nl_f32_f32_len, mul_mat_vec_iq4_nl_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
 
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_F32 ], "mul_mat_vec_f32_f16_f32",  mul_mat_vec_f32_f16_f32_len,  mul_mat_vec_f32_f16_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_F32 ], "mul_mat_vec_f32_f16_f32",  mul_mat_vec_f32_f16_f32_len,  mul_mat_vec_f32_f16_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_F16 ], "mul_mat_vec_f16_f16_f32",  mul_mat_vec_f16_f16_f32_len,  mul_mat_vec_f16_f16_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_F16 ], "mul_mat_vec_f16_f16_f32",  mul_mat_vec_f16_f16_f32_len,  mul_mat_vec_f16_f16_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q4_0], "mul_mat_vec_q4_0_f16_f32", mul_mat_vec_q4_0_f16_f32_len, mul_mat_vec_q4_0_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q4_0], "mul_mat_vec_q4_0_f16_f32", mul_mat_vec_q4_0_f16_f32_len, mul_mat_vec_q4_0_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q4_1], "mul_mat_vec_q4_1_f16_f32", mul_mat_vec_q4_1_f16_f32_len, mul_mat_vec_q4_1_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q4_1], "mul_mat_vec_q4_1_f16_f32", mul_mat_vec_q4_1_f16_f32_len, mul_mat_vec_q4_1_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q5_0], "mul_mat_vec_q5_0_f16_f32", mul_mat_vec_q5_0_f16_f32_len, mul_mat_vec_q5_0_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q5_0], "mul_mat_vec_q5_0_f16_f32", mul_mat_vec_q5_0_f16_f32_len, mul_mat_vec_q5_0_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q5_1], "mul_mat_vec_q5_1_f16_f32", mul_mat_vec_q5_1_f16_f32_len, mul_mat_vec_q5_1_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q5_1], "mul_mat_vec_q5_1_f16_f32", mul_mat_vec_q5_1_f16_f32_len, mul_mat_vec_q5_1_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q8_0], "mul_mat_vec_q8_0_f16_f32", mul_mat_vec_q8_0_f16_f32_len, mul_mat_vec_q8_0_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size, 1}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q8_0], "mul_mat_vec_q8_0_f16_f32", mul_mat_vec_q8_0_f16_f32_len, mul_mat_vec_q8_0_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q2_K], "mul_mat_vec_q2_k_f16_f32", mul_mat_vec_q2_k_f16_f32_len, mul_mat_vec_q2_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q2_K], "mul_mat_vec_q2_k_f16_f32", mul_mat_vec_q2_k_f16_f32_len, mul_mat_vec_q2_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q3_K], "mul_mat_vec_q3_k_f16_f32", mul_mat_vec_q3_k_f16_f32_len, mul_mat_vec_q3_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q3_K], "mul_mat_vec_q3_k_f16_f32", mul_mat_vec_q3_k_f16_f32_len, mul_mat_vec_q3_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q4_K], "mul_mat_vec_q4_k_f16_f32", mul_mat_vec_q4_k_f16_f32_len, mul_mat_vec_q4_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q4_K], "mul_mat_vec_q4_k_f16_f32", mul_mat_vec_q4_k_f16_f32_len, mul_mat_vec_q4_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q5_K], "mul_mat_vec_q5_k_f16_f32", mul_mat_vec_q5_k_f16_f32_len, mul_mat_vec_q5_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q5_K], "mul_mat_vec_q5_k_f16_f32", mul_mat_vec_q5_k_f16_f32_len, mul_mat_vec_q5_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q6_K], "mul_mat_vec_q6_k_f16_f32", mul_mat_vec_q6_k_f16_f32_len, mul_mat_vec_q6_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q6_K], "mul_mat_vec_q6_k_f16_f32", mul_mat_vec_q6_k_f16_f32_len, mul_mat_vec_q6_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_IQ4_NL], "mul_mat_vec_iq4_nl_f16_f32", mul_mat_vec_iq4_nl_f16_f32_len, mul_mat_vec_iq4_nl_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_IQ4_NL], "mul_mat_vec_iq4_nl_f16_f32", mul_mat_vec_iq4_nl_f16_f32_len, mul_mat_vec_iq4_nl_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
 
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_F32 ], "mul_mat_vec_id_f32_f32",  mul_mat_vec_id_f32_f32_len,  mul_mat_vec_id_f32_f32_data,  "main", 4, sizeof(vk_mat_vec_id_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_F32 ], "mul_mat_vec_id_f32_f32",  mul_mat_vec_id_f32_f32_len,  mul_mat_vec_id_f32_f32_data,  "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_F16 ], "mul_mat_vec_id_f16_f32",  mul_mat_vec_id_f16_f32_len,  mul_mat_vec_id_f16_f32_data,  "main", 4, sizeof(vk_mat_vec_id_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_F16 ], "mul_mat_vec_id_f16_f32",  mul_mat_vec_id_f16_f32_len,  mul_mat_vec_id_f16_f32_data,  "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q4_0], "mul_mat_vec_id_q4_0_f32", mul_mat_vec_id_q4_0_f32_len, mul_mat_vec_id_q4_0_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q4_0], "mul_mat_vec_id_q4_0_f32", mul_mat_vec_id_q4_0_f32_len, mul_mat_vec_id_q4_0_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q4_1], "mul_mat_vec_id_q4_1_f32", mul_mat_vec_id_q4_1_f32_len, mul_mat_vec_id_q4_1_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q4_1], "mul_mat_vec_id_q4_1_f32", mul_mat_vec_id_q4_1_f32_len, mul_mat_vec_id_q4_1_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q5_0], "mul_mat_vec_id_q5_0_f32", mul_mat_vec_id_q5_0_f32_len, mul_mat_vec_id_q5_0_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q5_0], "mul_mat_vec_id_q5_0_f32", mul_mat_vec_id_q5_0_f32_len, mul_mat_vec_id_q5_0_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q5_1], "mul_mat_vec_id_q5_1_f32", mul_mat_vec_id_q5_1_f32_len, mul_mat_vec_id_q5_1_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q5_1], "mul_mat_vec_id_q5_1_f32", mul_mat_vec_id_q5_1_f32_len, mul_mat_vec_id_q5_1_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q8_0], "mul_mat_vec_id_q8_0_f32", mul_mat_vec_id_q8_0_f32_len, mul_mat_vec_id_q8_0_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, {device->subgroup_size, 1}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q8_0], "mul_mat_vec_id_q8_0_f32", mul_mat_vec_id_q8_0_f32_len, mul_mat_vec_id_q8_0_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q2_K], "mul_mat_vec_id_q2_k_f32", mul_mat_vec_id_q2_k_f32_len, mul_mat_vec_id_q2_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, {device->subgroup_size}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q2_K], "mul_mat_vec_id_q2_k_f32", mul_mat_vec_id_q2_k_f32_len, mul_mat_vec_id_q2_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q3_K], "mul_mat_vec_id_q3_k_f32", mul_mat_vec_id_q3_k_f32_len, mul_mat_vec_id_q3_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, {device->subgroup_size}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q3_K], "mul_mat_vec_id_q3_k_f32", mul_mat_vec_id_q3_k_f32_len, mul_mat_vec_id_q3_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q4_K], "mul_mat_vec_id_q4_k_f32", mul_mat_vec_id_q4_k_f32_len, mul_mat_vec_id_q4_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, {device->subgroup_size}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q4_K], "mul_mat_vec_id_q4_k_f32", mul_mat_vec_id_q4_k_f32_len, mul_mat_vec_id_q4_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q5_K], "mul_mat_vec_id_q5_k_f32", mul_mat_vec_id_q5_k_f32_len, mul_mat_vec_id_q5_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, {device->subgroup_size}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q5_K], "mul_mat_vec_id_q5_k_f32", mul_mat_vec_id_q5_k_f32_len, mul_mat_vec_id_q5_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q6_K], "mul_mat_vec_id_q6_k_f32", mul_mat_vec_id_q6_k_f32_len, mul_mat_vec_id_q6_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, {device->subgroup_size}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q6_K], "mul_mat_vec_id_q6_k_f32", mul_mat_vec_id_q6_k_f32_len, mul_mat_vec_id_q6_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_IQ4_NL], "mul_mat_vec_id_iq4_nl_f32", mul_mat_vec_id_iq4_nl_f32_len, mul_mat_vec_id_iq4_nl_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_IQ4_NL], "mul_mat_vec_id_iq4_nl_f32", mul_mat_vec_id_iq4_nl_f32_len, mul_mat_vec_id_iq4_nl_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
 
     // dequant shaders
     ggml_vk_create_pipeline(device, device->pipeline_dequant[GGML_TYPE_F32 ], "f32_to_f16",   dequant_f32_len,  dequant_f32_data,  "main", 2, 5 * sizeof(uint32_t), {256 * 16, 1, 1}, {}, 1);
@@ -1460,17 +1772,13 @@ static void ggml_vk_load_shaders(vk_device& device) {
     ggml_vk_create_pipeline(device, device->pipeline_contig_cpy_f32_f16, "contig_cpy_f32_f16", contig_cpy_f32_f16_len, contig_cpy_f32_f16_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_contig_cpy_f16_f16, "contig_cpy_f16_f16", contig_cpy_f16_f16_len, contig_cpy_f16_f16_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
 
-    ggml_vk_create_pipeline(device, device->pipeline_add_f32, "add_f32", add_f32_len, add_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {0}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_add_f32, "add_f32", add_f32_len, add_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_add_f32_norepeat, "add_f32_norepeat", add_f32_len, add_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {1}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_add_f16_f32_f16, "add_f16_f32_f16", add_f16_f32_f16_len, add_f16_f32_f16_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_add_f16_f32_f16, "add_f16_f32_f16", add_f16_f32_f16_len, add_f16_f32_f16_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {0}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_add_f16_f32_f16_norepeat, "add_f16_f32_f16_norepeat", add_f16_f32_f16_len, add_f16_f32_f16_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {1}, 1);
 
     ggml_vk_create_pipeline(device, device->pipeline_acc_f32, "acc_f32", acc_f32_len, acc_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {}, 1);
 
-    ggml_vk_create_pipeline(device, device->pipeline_mul_f32, "mul_f32", mul_f32_len, mul_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {0}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_mul_f32, "mul_f32", mul_f32_len, mul_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_mul_f32_norepeat, "mul_f32_norepeat", mul_f32_len, mul_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {1}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_div_f32, "div_f32", div_f32_len, div_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_div_f32, "div_f32", div_f32_len, div_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {0}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_div_f32_norepeat, "div_f32_norepeat", div_f32_len, div_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {1}, 1);
 
     ggml_vk_create_pipeline(device, device->pipeline_concat_f32, "concat_f32", concat_f32_len, concat_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_concat_f16, "concat_f16", concat_f16_len, concat_f16_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {}, 1);
@@ -1499,10 +1807,8 @@ static void ggml_vk_load_shaders(vk_device& device) {
 
     ggml_vk_create_pipeline(device, device->pipeline_diag_mask_inf_f32, "diag_mask_inf_f32", diag_mask_inf_f32_len, diag_mask_inf_f32_data, "main", 2, sizeof(vk_op_diag_mask_push_constants), {512, 1, 1}, {}, 1);
 
-    ggml_vk_create_pipeline(device, device->pipeline_soft_max_f32, "soft_max_f32", soft_max_f32_len, soft_max_f32_data, "main", 3, sizeof(vk_op_soft_max_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_soft_max_f32, "soft_max_f32", soft_max_f32_len, soft_max_f32_data, "main", 3, sizeof(vk_op_soft_max_push_constants), {1, 1, 1}, {}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_soft_max_f32_wg512, "soft_max_f32_wg512", soft_max_f32_len, soft_max_f32_data, "main", 3, sizeof(vk_op_soft_max_push_constants), {1, 1, 1}, { 512 }, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_soft_max_f32_f16, "soft_max_f32_f16", soft_max_f32_f16_len, soft_max_f32_f16_data, "main", 3, sizeof(vk_op_soft_max_push_constants), {1, 1, 1}, {}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_soft_max_f32_f16, "soft_max_f32_f16", soft_max_f32_f16_len, soft_max_f32_f16_data, "main", 3, sizeof(vk_op_soft_max_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_soft_max_f32_f16_wg512, "soft_max_f32_f16_wg512", soft_max_f32_f16_len, soft_max_f32_f16_data, "main", 3, sizeof(vk_op_soft_max_push_constants), {1, 1, 1}, { 512 }, 1);
 
     ggml_vk_create_pipeline(device, device->pipeline_rope_norm_f32, "rope_norm_f32", rope_norm_f32_len, rope_norm_f32_data, "main", 4, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_rope_norm_f16, "rope_norm_f16", rope_norm_f16_len, rope_norm_f16_data, "main", 4, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1);
@@ -1768,11 +2074,11 @@ static void ggml_vk_print_gpu_info(size_t idx) {
     fp16 = fp16 && vk12_features.shaderFloat16;
 
     std::string device_name = props2.properties.deviceName.data();
-    GGML_LOG_DEBUG("ggml_vulkan: %zu = %s (%s) | uma: %d | fp16: %d | warp size: %zu\n",
+    GGML_LOG_DEBUG("ggml_vulkan: %d = %s (%s) | uma: %d | fp16: %d | warp size: %d\n",
-              idx, device_name.c_str(), driver_props.driverName.data(), uma, fp16, subgroup_size);
+              idx, device_name.c_str(), driver_props.driverName, uma, fp16, subgroup_size);
 
     if (props2.properties.deviceType == vk::PhysicalDeviceType::eCpu) {
-        GGML_LOG_DEBUG("ggml_vulkan: Warning: Device type is CPU. This is probably not the device you want.\n");
+        std::cerr << "ggml_vulkan: Warning: Device type is CPU. This is probably not the device you want." << std::endl;
     }
 }
 
@@ -1941,7 +2247,8 @@ void ggml_vk_instance_init() {
             vk_instance.device_indices.push_back(0);
         }
     }
-    GGML_LOG_DEBUG("ggml_vulkan: Found %zu Vulkan devices:\n", vk_instance.device_indices.size());
+    GGML_LOG_DEBUG("ggml_vulkan: Found %d Vulkan devices:\n", vk_instance.device_indices.size());
+
 
     for (size_t i = 0; i < vk_instance.device_indices.size(); i++) {
         ggml_vk_print_gpu_info(i);
@@ -2006,10 +2313,10 @@ static vk_matmul_pipeline ggml_vk_get_mul_mat_mat_pipeline(ggml_backend_vk_conte
         return ctx->device->pipeline_matmul_f32_f16;
     }
     if (src0_type == GGML_TYPE_F16 && src1_type == GGML_TYPE_F32) {
-        return ctx->device->pipeline_matmul_f16_f32.f32acc;
+        return ctx->device->pipeline_matmul_f16_f32;
     }
     if (src0_type == GGML_TYPE_F16 && src1_type == GGML_TYPE_F16) {
-        return ctx->device->pipeline_matmul_f16.f32acc;
+        return ctx->device->pipeline_matmul_f16;
     }
 
     if (src1_type != GGML_TYPE_F32) {
@@ -2033,7 +2340,7 @@ static vk_matmul_pipeline ggml_vk_get_mul_mat_mat_pipeline(ggml_backend_vk_conte
             return nullptr;
     }
 
-    return ctx->device->pipeline_dequant_mul_mat_mat[src0_type].f32acc;
+    return ctx->device->pipeline_dequant_mul_mat_mat[src0_type];
 }
 
 static vk_pipeline ggml_vk_get_dequantize_mul_mat_vec(ggml_backend_vk_context * ctx, ggml_type a_type, ggml_type b_type) {
@@ -3810,20 +4117,20 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
         return nullptr;
     case GGML_OP_ADD:
         if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
-            return ggml_are_same_shape(src0, src1) ? ctx->device->pipeline_add_f32_norepeat : ctx->device->pipeline_add_f32;
+            return ctx->device->pipeline_add_f32;
         }
         if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F16) {
-            return ggml_are_same_shape(src0, src1) ? ctx->device->pipeline_add_f16_f32_f16_norepeat : ctx->device->pipeline_add_f16_f32_f16;
+            return ctx->device->pipeline_add_f16_f32_f16;
         }
         return nullptr;
     case GGML_OP_MUL:
         if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
-            return ggml_are_same_shape(src0, src1) ? ctx->device->pipeline_mul_f32_norepeat : ctx->device->pipeline_mul_f32;
+            return ctx->device->pipeline_mul_f32;
         }
         return nullptr;
     case GGML_OP_DIV:
         if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
-            return ggml_are_same_shape(src0, src1) ? ctx->device->pipeline_div_f32_norepeat : ctx->device->pipeline_div_f32;
+            return ctx->device->pipeline_div_f32;
         }
         return nullptr;
     case GGML_OP_CONCAT:
@@ -3936,10 +4243,10 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
         GGML_ASSERT(!src1 || src1->type == GGML_TYPE_F32 || src1->type == GGML_TYPE_F16);
 
         if (src0->type == GGML_TYPE_F32 && (src1 == nullptr || src1->type == GGML_TYPE_F32) && dst->type == GGML_TYPE_F32) {
-            return src0->ne[0] > 1024 ? ctx->device->pipeline_soft_max_f32_wg512 : ctx->device->pipeline_soft_max_f32;
+            return ctx->device->pipeline_soft_max_f32;
         }
         if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F32) {
-            return src0->ne[0] > 1024 ? ctx->device->pipeline_soft_max_f32_f16_wg512 : ctx->device->pipeline_soft_max_f32_f16;
+            return ctx->device->pipeline_soft_max_f32_f16;
         }
         return nullptr;
     case GGML_OP_ROPE:
@@ -4585,7 +4892,6 @@ static void ggml_vk_soft_max(ggml_backend_vk_context * ctx, vk_context& subctx,
         scale, max_bias,
         m0, m1,
         n_head_log2,
-        nrows_x,
     }, dryrun);
 }
 
@@ -4764,10 +5070,10 @@ static void ggml_vk_test_matmul(ggml_backend_vk_context * ctx, size_t m, size_t
             p = ctx->device->pipeline_matmul_f32_f16->a_s;
             shname = "F32_F16_ALIGNED_S";
         } else if (std::is_same<ggml_fp16_t, X_TYPE>() && std::is_same<float, Y_TYPE>()) {
-            p = ctx->device->pipeline_matmul_f16_f32.f32acc->a_s;
+            p = ctx->device->pipeline_matmul_f16_f32->a_s;
             shname = "F16_F32_ALIGNED_S";
         } else if (std::is_same<ggml_fp16_t, X_TYPE>() && std::is_same<ggml_fp16_t, Y_TYPE>()) {
-            p = ctx->device->pipeline_matmul_f16.f32acc->a_s;
+            p = ctx->device->pipeline_matmul_f16->a_s;
             shname = "F16_ALIGNED_S";
         } else {
             GGML_ABORT("fatal error");
@@ -4780,10 +5086,10 @@ static void ggml_vk_test_matmul(ggml_backend_vk_context * ctx, size_t m, size_t
             p = ctx->device->pipeline_matmul_f32_f16->a_m;
             shname = "F32_F16_ALIGNED_M";
         } else if (std::is_same<ggml_fp16_t, X_TYPE>() && std::is_same<float, Y_TYPE>()) {
-            p = ctx->device->pipeline_matmul_f16_f32.f32acc->a_m;
+            p = ctx->device->pipeline_matmul_f16_f32->a_m;
             shname = "F16_F32_ALIGNED_M";
         } else if (std::is_same<ggml_fp16_t, X_TYPE>() && std::is_same<ggml_fp16_t, Y_TYPE>()) {
-            p = ctx->device->pipeline_matmul_f16.f32acc->a_m;
+            p = ctx->device->pipeline_matmul_f16->a_m;
             shname = "F16_ALIGNED_M";
         } else {
             GGML_ABORT("fatal error");
@@ -4796,10 +5102,10 @@ static void ggml_vk_test_matmul(ggml_backend_vk_context * ctx, size_t m, size_t
             p = ctx->device->pipeline_matmul_f32_f16->a_l;
             shname = "F32_F16_ALIGNED_L";
         } else if (std::is_same<ggml_fp16_t, X_TYPE>() && std::is_same<float, Y_TYPE>()) {
-            p = ctx->device->pipeline_matmul_f16_f32.f32acc->a_l;
+            p = ctx->device->pipeline_matmul_f16_f32->a_l;
             shname = "F16_F32_ALIGNED_L";
         } else if (std::is_same<ggml_fp16_t, X_TYPE>() && std::is_same<ggml_fp16_t, Y_TYPE>()) {
-            p = ctx->device->pipeline_matmul_f16.f32acc->a_l;
+            p = ctx->device->pipeline_matmul_f16->a_l;
             shname = "F16_ALIGNED_L";
         } else {
             GGML_ABORT("fatal error");
@@ -4819,10 +5125,10 @@ static void ggml_vk_test_matmul(ggml_backend_vk_context * ctx, size_t m, size_t
                 p = ctx->device->pipeline_matmul_f32_f16->s;
                 shname = "F32_F16_S";
             } else if (std::is_same<ggml_fp16_t, X_TYPE>() && std::is_same<float, Y_TYPE>()) {
-                p = ctx->device->pipeline_matmul_f16_f32.f32acc->s;
+                p = ctx->device->pipeline_matmul_f16_f32->s;
                 shname = "F16_F32_S";
             } else if (std::is_same<ggml_fp16_t, X_TYPE>() && std::is_same<ggml_fp16_t, Y_TYPE>()) {
-                p = ctx->device->pipeline_matmul_f16.f32acc->s;
+                p = ctx->device->pipeline_matmul_f16->s;
                 shname = "F16_S";
             }
         } else if (shader_size == 1) {
@@ -4833,10 +5139,10 @@ static void ggml_vk_test_matmul(ggml_backend_vk_context * ctx, size_t m, size_t
                 p = ctx->device->pipeline_matmul_f32_f16->m;
                 shname = "F32_F16_M";
             } else if (std::is_same<ggml_fp16_t, X_TYPE>() && std::is_same<float, Y_TYPE>()) {
-                p = ctx->device->pipeline_matmul_f16_f32.f32acc->m;
+                p = ctx->device->pipeline_matmul_f16_f32->m;
                 shname = "F16_F32_M";
             } else if (std::is_same<ggml_fp16_t, X_TYPE>() && std::is_same<ggml_fp16_t, Y_TYPE>()) {
-                p = ctx->device->pipeline_matmul_f16.f32acc->m;
+                p = ctx->device->pipeline_matmul_f16->m;
                 shname = "F16_M";
             }
         } else if (shader_size == 2) {
@@ -4847,10 +5153,10 @@ static void ggml_vk_test_matmul(ggml_backend_vk_context * ctx, size_t m, size_t
                 p = ctx->device->pipeline_matmul_f32_f16->l;
                 shname = "F32_F16_L";
             } else if (std::is_same<ggml_fp16_t, X_TYPE>() && std::is_same<float, Y_TYPE>()) {
-                p = ctx->device->pipeline_matmul_f16_f32.f32acc->l;
+                p = ctx->device->pipeline_matmul_f16_f32->l;
                 shname = "F16_F32_L";
             } else if (std::is_same<ggml_fp16_t, X_TYPE>() && std::is_same<ggml_fp16_t, Y_TYPE>()) {
-                p = ctx->device->pipeline_matmul_f16.f32acc->l;
+                p = ctx->device->pipeline_matmul_f16->l;
                 shname = "F16_L";
             }
         }
@@ -5179,13 +5485,13 @@ static void ggml_vk_test_dequant_matmul(ggml_backend_vk_context * ctx, size_t m,
     vk_pipeline p;
     std::string shname;
     if (shader_size == 0) {
-        p = ctx->device->pipeline_dequant_mul_mat_mat[quant].f32acc->a_s;
+        p = ctx->device->pipeline_dequant_mul_mat_mat[quant]->a_s;
         shname = std::string(ggml_type_name(quant)) + "_ALIGNED_S";
     } else if (shader_size == 1) {
-        p = ctx->device->pipeline_dequant_mul_mat_mat[quant].f32acc->a_m;
+        p = ctx->device->pipeline_dequant_mul_mat_mat[quant]->a_m;
         shname = std::string(ggml_type_name(quant)) + "_ALIGNED_M";
     } else if (shader_size == 2) {
-        p = ctx->device->pipeline_dequant_mul_mat_mat[quant].f32acc->a_l;
+        p = ctx->device->pipeline_dequant_mul_mat_mat[quant]->a_l;
         shname = std::string(ggml_type_name(quant)) + "_ALIGNED_L";
     } else {
         GGML_ASSERT(0);
@@ -5195,13 +5501,13 @@ static void ggml_vk_test_dequant_matmul(ggml_backend_vk_context * ctx, size_t m,
 
     if (k != kpad) {
         if (shader_size == 0) {
-            p = ctx->device->pipeline_dequant_mul_mat_mat[quant].f32acc->s;
+            p = ctx->device->pipeline_dequant_mul_mat_mat[quant]->s;
             shname = std::string(ggml_type_name(quant)) + "_S";
         } else if (shader_size == 1) {
-            p = ctx->device->pipeline_dequant_mul_mat_mat[quant].f32acc->m;
+            p = ctx->device->pipeline_dequant_mul_mat_mat[quant]->m;
             shname = std::string(ggml_type_name(quant)) + "_M";
         } else if (shader_size == 2) {
-            p = ctx->device->pipeline_dequant_mul_mat_mat[quant].f32acc->l;
+            p = ctx->device->pipeline_dequant_mul_mat_mat[quant]->l;
             shname = std::string(ggml_type_name(quant)) + "_L";
         } else {
             GGML_ASSERT(0);

ggml/src/ggml-vulkan/CMakeLists.txt

@@ -1,78 +0,0 @@
-find_package(Vulkan COMPONENTS glslc REQUIRED)
-
-if (Vulkan_FOUND)
-    message(STATUS "Vulkan found")
-
-    add_library(ggml-vulkan
-                ggml-vulkan.cpp
-                ../../include/ggml-vulkan.h
-                )
-
-    target_link_libraries(ggml-vulkan PRIVATE ggml-base Vulkan::Vulkan)
-    target_include_directories(ggml-vulkan PRIVATE . .. ${CMAKE_CURRENT_BINARY_DIR})
-
-    # Workaround to the "can't dereference invalidated vector iterator" bug in clang-cl debug build
-    # Posssibly relevant: https://stackoverflow.com/questions/74748276/visual-studio-no-displays-the-correct-length-of-stdvector
-    if (MSVC AND CMAKE_CXX_COMPILER_ID STREQUAL "Clang")
-        add_compile_definitions(_ITERATOR_DEBUG_LEVEL=0)
-    endif()
-
-    if (GGML_VULKAN_CHECK_RESULTS)
-        add_compile_definitions(GGML_VULKAN_CHECK_RESULTS)
-    endif()
-
-    if (GGML_VULKAN_DEBUG)
-        add_compile_definitions(GGML_VULKAN_DEBUG)
-    endif()
-
-    if (GGML_VULKAN_MEMORY_DEBUG)
-        add_compile_definitions(GGML_VULKAN_MEMORY_DEBUG)
-    endif()
-
-    if (GGML_VULKAN_SHADER_DEBUG_INFO)
-        add_compile_definitions(GGML_VULKAN_SHADER_DEBUG_INFO)
-    endif()
-
-    if (GGML_VULKAN_PERF)
-        add_compile_definitions(GGML_VULKAN_PERF)
-    endif()
-
-    if (GGML_VULKAN_VALIDATE)
-        add_compile_definitions(GGML_VULKAN_VALIDATE)
-    endif()
-
-    if (GGML_VULKAN_RUN_TESTS)
-        add_compile_definitions(GGML_VULKAN_RUN_TESTS)
-    endif()
-
-    add_subdirectory(vulkan-shaders)
-
-    set (_ggml_vk_genshaders_cmd vulkan-shaders-gen)
-    set (_ggml_vk_header     ${CMAKE_CURRENT_BINARY_DIR}/ggml-vulkan-shaders.hpp)
-    set (_ggml_vk_source     ${CMAKE_CURRENT_BINARY_DIR}/ggml-vulkan-shaders.cpp)
-    set (_ggml_vk_input_dir  ${CMAKE_CURRENT_SOURCE_DIR}/vulkan-shaders)
-    set (_ggml_vk_output_dir ${CMAKE_CURRENT_BINARY_DIR}/vulkan-shaders.spv)
-
-    file(GLOB _ggml_vk_shader_deps "${_ggml_vk_input_dir}/*.comp")
-
-    add_custom_command(
-        OUTPUT ${_ggml_vk_header}
-                ${_ggml_vk_source}
-
-        COMMAND ${_ggml_vk_genshaders_cmd}
-            --glslc      ${Vulkan_GLSLC_EXECUTABLE}
-            --input-dir  ${_ggml_vk_input_dir}
-            --output-dir ${_ggml_vk_output_dir}
-            --target-hpp ${_ggml_vk_header}
-            --target-cpp ${_ggml_vk_source}
-            --no-clean
-
-        DEPENDS ${_ggml_vk_shader_deps}
-        COMMENT "Generate vulkan shaders"
-    )
-
-    target_sources(ggml-vulkan PRIVATE ${_ggml_vk_source} ${_ggml_vk_header})
-
-else()
-    message(WARNING "Vulkan not found")
-endif()

ggml/src/ggml-vulkan/vulkan-shaders/add.comp

@@ -1,29 +0,0 @@
-#version 450
-
-#extension GL_EXT_shader_16bit_storage : require
-
-#include "types.comp"
-#include "generic_binary_head.comp"
-
-const uint num_threads = 256;
-
-layout(local_size_x = num_threads, local_size_y = 1, local_size_z = 1) in;
-
-void main() {
-    uint idx = get_idx();
-
-    // num_threads * num_iter must equal 512, to match the wg_denoms and get_idx calculation
-    const uint num_iter = 2;
-
-    [[unroll]] for (uint i = 0; i < num_iter; ++i) {
-        if (idx >= p.ne) {
-            continue;
-        }
-        uint i00, i01, i02, i03;
-        get_indices(idx, i00, i01, i02, i03);
-
-        data_d[p.d_offset + dst_idx(i00, i01, i02, i03)] = D_TYPE(FLOAT_TYPE(data_a[src0_idx(i00, i01, i02, i03)]) + FLOAT_TYPE(data_b[src1_idx(i00, i01, i02, i03)]));
-
-        idx += num_threads;
-    }
-}

ggml/src/ggml-vulkan/vulkan-shaders/div.comp

@@ -1,27 +0,0 @@
-#version 450
-
-#include "types.comp"
-#include "generic_binary_head.comp"
-
-const uint num_threads = 256;
-
-layout(local_size_x = num_threads, local_size_y = 1, local_size_z = 1) in;
-
-void main() {
-    uint idx = get_idx();
-
-    // num_threads * num_iter must equal 512, to match the wg_denoms and get_idx calculation
-    const uint num_iter = 2;
-
-    [[unroll]] for (uint i = 0; i < num_iter; ++i) {
-        if (idx >= p.ne) {
-            continue;
-        }
-        uint i00, i01, i02, i03;
-        get_indices(idx, i00, i01, i02, i03);
-
-        data_d[p.d_offset + dst_idx(i00, i01, i02, i03)] = D_TYPE(FLOAT_TYPE(data_a[src0_idx(i00, i01, i02, i03)]) / FLOAT_TYPE(data_b[src1_idx(i00, i01, i02, i03)]));
-
-        idx += num_threads;
-    }
-}

ggml/src/ggml-vulkan/vulkan-shaders/generic_binary_head.comp

@@ -1,60 +0,0 @@
-#extension GL_EXT_shader_16bit_storage : require
-#extension GL_EXT_control_flow_attributes : require
-
-layout (push_constant) uniform parameter
-{
-    uint ne;
-    uint ne00; uint ne01; uint ne02; uint ne03; uint nb00; uint nb01; uint nb02; uint nb03;
-    uint ne10; uint ne11; uint ne12; uint ne13; uint nb10; uint nb11; uint nb12; uint nb13;
-    uint ne20; uint ne21; uint ne22; uint ne23; uint nb20; uint nb21; uint nb22; uint nb23;
-    uint d_offset;
-    float param1; float param2; int param3;
-} p;
-
-layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
-layout (binding = 1) readonly buffer B {B_TYPE data_b[];};
-layout (binding = 2) writeonly buffer D {D_TYPE data_d[];};
-
-// true if src0/src1 are the same shape and the indices can be reused without additional modulus
-layout(constant_id = 0) const bool norepeat = false;
-
-uint get_idx() {
-    return gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
-}
-
-// mod and div are expensive and coordinates/dimensions are often power of 2 or equal to 1
-uint fastmod(uint a, uint b) {
-    if ((b & (b-1)) == 0) {
-        return a & (b-1);
-    }
-    return a % b;
-}
-
-uint fastdiv(uint a, uint b) {
-    return (a < b) ? 0 : (a / b);
-}
-
-void get_indices(uint idx, out uint i00, out uint i01, out uint i02, out uint i03) {
-    i03 = fastdiv(idx, (p.ne02*p.ne01*p.ne00));
-    const uint i03_offset = i03 * p.ne02*p.ne01*p.ne00;
-    i02 = fastdiv((idx - i03_offset), (p.ne01*p.ne00));
-    const uint i02_offset = i02*p.ne01*p.ne00;
-    i01 = (idx - i03_offset - i02_offset) / p.ne00;
-    i00 = idx - i03_offset - i02_offset - i01*p.ne00;
-}
-
-uint src0_idx(uint i00, uint i01, uint i02, uint i03) {
-    return i03*p.nb03 + i02*p.nb02 + i01*p.nb01 + i00*p.nb00;
-}
-
-uint src1_idx(uint i00, uint i01, uint i02, uint i03) {
-    if (norepeat) {
-        return i03*p.nb13 + i02*p.nb12 + i01*p.nb11 + i00*p.nb10;
-    } else {
-        return fastmod(i03, p.ne13)*p.nb13 + fastmod(i02, p.ne12)*p.nb12 + fastmod(i01, p.ne11)*p.nb11 + fastmod(i00, p.ne10)*p.nb10;
-    }
-}
-
-uint dst_idx(uint i00, uint i01, uint i02, uint i03) {
-    return i03*p.nb23 + i02*p.nb22 + i01*p.nb21 + i00*p.nb20;
-}

ggml/src/ggml-vulkan/vulkan-shaders/mul.comp

@@ -1,27 +0,0 @@
-#version 450
-
-#include "types.comp"
-#include "generic_binary_head.comp"
-
-const uint num_threads = 256;
-
-layout(local_size_x = num_threads, local_size_y = 1, local_size_z = 1) in;
-
-void main() {
-    uint idx = get_idx();
-
-    // num_threads * num_iter must equal 512, to match the wg_denoms and get_idx calculation
-    const uint num_iter = 2;
-
-    [[unroll]] for (uint i = 0; i < num_iter; ++i) {
-        if (idx >= p.ne) {
-            continue;
-        }
-        uint i00, i01, i02, i03;
-        get_indices(idx, i00, i01, i02, i03);
-
-        data_d[p.d_offset + dst_idx(i00, i01, i02, i03)] = D_TYPE(FLOAT_TYPE(data_a[src0_idx(i00, i01, i02, i03)]) * FLOAT_TYPE(data_b[src1_idx(i00, i01, i02, i03)]));
-
-        idx += num_threads;
-    }
-}

ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec.comp

@@ -1,111 +0,0 @@
-#version 450
-
-#ifdef FLOAT16
-#extension GL_EXT_shader_explicit_arithmetic_types_float16 : require
-#endif
-#extension GL_EXT_shader_explicit_arithmetic_types_int32 : require
-
-#include "mul_mat_vec_base.comp"
-
-layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
-
-layout (constant_id = 0) const uint BLOCK_SIZE = 32;
-layout (constant_id = 1) const uint NUM_ROWS = 1;
-
-uint a_offset, b_offset, d_offset, y_offset;
-
-shared FLOAT_TYPE tmpsh[NUM_ROWS][BLOCK_SIZE];
-
-void iter(inout FLOAT_TYPE temp[NUM_ROWS], const uint first_row, const uint num_rows, const uint tid, const uint i, bool lastiter)
-{
-    const uint col = i*BLOCK_SIZE + 2*tid;
-    const uint iqs = (col%QUANT_K)/QUANT_R; // quant index
-    const uint iybs = col - col%QUANT_K; // y block start index
-
-    // Check if the second of the pair of elements is OOB, and don't fetch B or
-    // accumulate it. We still fetch a pair of elements for A, which is fine for
-    // quantized formats since they'll be within the same block. We should
-    // probably skip fetching the second element for F16/F32, but as of now we
-    // still do.
-    const bool OOB = lastiter && (iybs + iqs + y_offset >= p.ncols);
-
-    FLOAT_TYPE b0 = 0, b1 = 0;
-    b0 = FLOAT_TYPE(data_b[b_offset + iybs + iqs]);
-    if (!OOB) {
-        b1 = FLOAT_TYPE(data_b[b_offset + iybs + iqs + y_offset]);
-    }
-    [[unroll]] for (uint n = 0; n < num_rows; ++n) {
-        const uint ib = ((first_row + n)*p.ncols + col)/QUANT_K; // block index
-
-        const vec2 v = dequantize(ib, iqs, a_offset);
-
-        // matrix multiplication
-        temp[n] = fma(FLOAT_TYPE(v.x), b0, temp[n]);
-        if (!OOB) {
-            temp[n] = fma(FLOAT_TYPE(v.y), b1, temp[n]);
-        }
-    }
-}
-
-void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
-    const uint tid = gl_LocalInvocationID.x;
-
-    get_offsets(a_offset, b_offset, d_offset);
-    a_offset /= QUANT_K;
-
-    y_offset = QUANT_R == 1 ? 1 : QUANT_K/2;
-
-    FLOAT_TYPE temp[NUM_ROWS];
-
-    for (uint i = 0; i < NUM_ROWS; ++i) {
-        temp[i] = FLOAT_TYPE(0);
-    }
-
-    const int unroll_count = 8;
-
-    const uint num_iters = (p.ncols >= 2*tid) ? ((p.ncols - 2*tid + BLOCK_SIZE - 1) / BLOCK_SIZE) : 0;
-    const uint unrolled_iters = num_iters & ~(2*unroll_count - 1);
-
-    uint i = 0;
-    while (i < unrolled_iters) {
-        // Manually partially unroll the loop
-        [[unroll]] for (uint k = 0; k < unroll_count; ++k) {
-            iter(temp, first_row, num_rows, tid, i, false);
-            i += 2;
-        }
-    }
-    while (i < num_iters) {
-        iter(temp, first_row, num_rows, tid, i, true);
-        i += 2;
-    }
-
-    // sum up partial sums and write back result
-    [[unroll]] for (uint n = 0; n < num_rows; ++n) {
-        tmpsh[n][tid] = temp[n];
-    }
-    barrier();
-    [[unroll]] for (uint s = BLOCK_SIZE/2; s > 0; s >>= 1) {
-        if (tid < s) {
-            [[unroll]] for (uint n = 0; n < num_rows; ++n) {
-                tmpsh[n][tid] += tmpsh[n][tid + s];
-            }
-        }
-        barrier();
-    }
-    if (tid == 0) {
-        [[unroll]] for (uint n = 0; n < num_rows; ++n) {
-            data_d[d_offset + first_row + n] = D_TYPE(tmpsh[n][0]);
-        }
-    }
-}
-
-void main() {
-    const uint first_row = NUM_ROWS * (gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z);
-
-    // do NUM_ROWS at a time, unless there aren't enough remaining rows
-    if (first_row + NUM_ROWS <= p.stride_d) {
-        compute_outputs(first_row, NUM_ROWS);
-    } else {
-        compute_outputs(first_row, p.stride_d - first_row);
-    }
-}

ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q4_k.comp

@@ -1,143 +0,0 @@
-#version 450
-
-#extension GL_EXT_shader_explicit_arithmetic_types : require
-
-#include "mul_mat_vec_base.comp"
-
-layout(local_size_x = 32, local_size_y = 1, local_size_z = 1) in;
-
-shared FLOAT_TYPE tmp[32];
-
-// Declare aliased versions of A and B bindings that can use 16b/32b loads for
-// the quantized values, and vec4 loads for B.
-struct block_q4_K_u32
-{
-    f16vec2 d;
-    uint32_t scales[3*QUANT_K/64/4];
-    uint32_t qs[QUANT_K/2/4];
-};
-
-struct block_q4_K_u16
-{
-    f16vec2 d;
-    uint16_t scales[3*QUANT_K/64/2];
-    uint16_t qs[QUANT_K/2/2];
-};
-
-layout (binding = 0) readonly buffer A_u32 {block_q4_K_u32 data_a_u32[];};
-layout (binding = 0) readonly buffer A_u16 {block_q4_K_u16 data_a_u16[];};
-layout (binding = 1) readonly buffer BV4 {B_TYPE_VEC4 data_b_v4[];};
-
-// This shader assumes K_QUANTS_PER_ITERATION == 2 for alignment of loads
-void main() {
-    const uint row = gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z;
-
-    uint a_offset, b_offset, d_offset;
-    get_offsets(a_offset, b_offset, d_offset);
-
-    const uint num_blocks_per_row = p.ncols / QUANT_K;
-    const uint ib0 = a_offset / QUANT_K + row*num_blocks_per_row;
-
-    const uint tid = gl_LocalInvocationID.x/K_QUANTS_PER_ITERATION;  // 0...31 or 0...16
-    const uint ix  = gl_LocalInvocationID.x%K_QUANTS_PER_ITERATION;  // 0 or 0, 1
-
-    const uint step = 8/K_QUANTS_PER_ITERATION;             // 8 or 4
-
-    const uint il = tid/step;                               // 0...3
-    const uint ir = tid - step*il;                          // 0...7 or 0...3
-    const uint n =  2 * K_QUANTS_PER_ITERATION;             // 2 or 4
-
-    const uint v_im = il / 2;  // 0 or 1. 0 computes 0,32 + 128,160, 1 computes 64,96 + 192,224
-    const uint v_in = il % 2;
-
-    const uint l0 = n * (2 * ir + v_in);            // 0...15
-    const uint q_offset = 32*v_im + l0;
-    const uint y_offset = 64*v_im + l0;
-
-    FLOAT_TYPE temp = FLOAT_TYPE(0.0); // partial sum for thread in warp
-
-    [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) {
-        const uint y1_idx = i * QUANT_K + y_offset;
-        const uint y2_idx = y1_idx + 128;
-
-        f16vec2 d = data_a[ib0 + i].d;
-        const FLOAT_TYPE dall = FLOAT_TYPE(d.x);
-        const FLOAT_TYPE dmin = FLOAT_TYPE(d.y);
-
-        uint32_t scale0_u32 = data_a_u16[ib0 + i].scales[v_im    ];
-        uint32_t scale4_u32 = data_a_u16[ib0 + i].scales[v_im + 2];
-        uint32_t scale8_u32 = data_a_u16[ib0 + i].scales[v_im + 4];
-        uvec4 scale0 = uvec4(unpack8(scale0_u32));
-        uvec4 scale4 = uvec4(unpack8(scale4_u32));
-        uvec4 scale8 = uvec4(unpack8(scale8_u32));
-
-        const uint32_t sc0 = (  scale0.x       & 0x3f);
-        const uint32_t sc1 = (  scale0.y       & 0x3f);
-        const uint32_t sc2 = (  scale4.x       & 0x3f);
-        const uint32_t sc3 = (  scale4.y       & 0x3f);
-        const uint32_t sc4 = (( scale8.x       & 0x0f) | ((scale0.x & 0xc0) >> 2));
-        const uint32_t sc5 = (( scale8.y       & 0x0f) | ((scale0.y & 0xc0) >> 2));
-        const uint32_t sc6 = (((scale8.x >> 4) & 0x0f) | ((scale4.x & 0xc0) >> 2));
-        const uint32_t sc7 = (((scale8.y >> 4) & 0x0f) | ((scale4.y & 0xc0) >> 2));
-
-        uint32_t qs0_u32 = data_a_u32[ib0 + i].qs[q_offset / 4];
-        uint32_t qs64_u32 = data_a_u32[ib0 + i].qs[q_offset / 4 + 16];
-
-        uint32_t qs0_u32_lo4 = qs0_u32 & 0x0F0F0F0F;
-        uint32_t qs0_u32_hi4 = (qs0_u32 >> 4) & 0x0F0F0F0F;
-        uint32_t qs64_u32_lo4 = qs64_u32 & 0x0F0F0F0F;
-        uint32_t qs64_u32_hi4 = (qs64_u32 >> 4) & 0x0F0F0F0F;
-
-        uvec4 qs0_lo4 = uvec4(unpack8(qs0_u32_lo4));
-        uvec4 qs64_lo4 = uvec4(unpack8(qs64_u32_lo4));
-        uvec4 qs0_hi4 = uvec4(unpack8(qs0_u32_hi4));
-        uvec4 qs64_hi4 = uvec4(unpack8(qs64_u32_hi4));
-
-        const uint32_t q4_0  = qs0_lo4.x;
-        const uint32_t q4_1  = qs0_lo4.y;
-        const uint32_t q4_2  = qs0_lo4.z;
-        const uint32_t q4_3  = qs0_lo4.w;
-        const uint32_t q4_4  = qs0_hi4.x;
-        const uint32_t q4_5  = qs0_hi4.y;
-        const uint32_t q4_6  = qs0_hi4.z;
-        const uint32_t q4_7  = qs0_hi4.w;
-        const uint32_t q4_8  = qs64_lo4.x;
-        const uint32_t q4_9  = qs64_lo4.y;
-        const uint32_t q4_10 = qs64_lo4.z;
-        const uint32_t q4_11 = qs64_lo4.w;
-        const uint32_t q4_12 = qs64_hi4.x;
-        const uint32_t q4_13 = qs64_hi4.y;
-        const uint32_t q4_14 = qs64_hi4.z;
-        const uint32_t q4_15 = qs64_hi4.w;
-
-        B_TYPE_VEC4 by10 =  data_b_v4[(b_offset + y1_idx) / 4];
-        B_TYPE_VEC4 by132 = data_b_v4[(b_offset + y1_idx) / 4 + 8];
-        B_TYPE_VEC4 by20 =  data_b_v4[(b_offset + y2_idx) / 4];
-        B_TYPE_VEC4 by232 = data_b_v4[(b_offset + y2_idx) / 4 + 8];
-
-        const FLOAT_TYPE sx = fma(FLOAT_TYPE(by10.x),      q4_0,  fma(FLOAT_TYPE(by10.y),  q4_1,  fma(FLOAT_TYPE(by10.z),  q4_2,  FLOAT_TYPE(by10.w) *  q4_3)));
-        const FLOAT_TYPE sy = fma(FLOAT_TYPE(by132.x),     q4_4,  fma(FLOAT_TYPE(by132.y), q4_5,  fma(FLOAT_TYPE(by132.z), q4_6,  FLOAT_TYPE(by132.w) * q4_7)));
-        const FLOAT_TYPE sz = fma(FLOAT_TYPE(by20.x),      q4_8,  fma(FLOAT_TYPE(by20.y),  q4_9,  fma(FLOAT_TYPE(by20.z),  q4_10, FLOAT_TYPE(by20.w) *  q4_11)));
-        const FLOAT_TYPE sw = fma(FLOAT_TYPE(by232.x),     q4_12, fma(FLOAT_TYPE(by232.y), q4_13, fma(FLOAT_TYPE(by232.z), q4_14, FLOAT_TYPE(by232.w) * q4_15)));
-        const FLOAT_TYPE smin =
-            fma(FLOAT_TYPE(by10.x), sc2, fma(FLOAT_TYPE(by132.x), sc3, fma(FLOAT_TYPE(by20.x), sc6, fma(FLOAT_TYPE(by232.x), sc7,
-            fma(FLOAT_TYPE(by10.y), sc2, fma(FLOAT_TYPE(by132.y), sc3, fma(FLOAT_TYPE(by20.y), sc6, fma(FLOAT_TYPE(by232.y), sc7,
-            fma(FLOAT_TYPE(by10.z), sc2, fma(FLOAT_TYPE(by132.z), sc3, fma(FLOAT_TYPE(by20.z), sc6, fma(FLOAT_TYPE(by232.z), sc7,
-            fma(FLOAT_TYPE(by10.w), sc2, fma(FLOAT_TYPE(by132.w), sc3, fma(FLOAT_TYPE(by20.w), sc6,     FLOAT_TYPE(by232.w) * sc7)))))))))))))));
-        temp = fma(dall, fma(sx, sc0, fma(sy, sc1, fma(sz, sc4, sw * sc5))), fma(-dmin, smin, temp));
-    }
-
-    tmp[gl_LocalInvocationID.x] = temp;
-
-    // sum up partial sums and write back result
-    barrier();
-    [[unroll]] for (uint s = 16; s > 0; s >>= 1) {
-        if (tid < s) {
-            tmp[tid] += tmp[tid + s];
-        }
-        barrier();
-    }
-    if (tid == 0) {
-        data_d[d_offset + row] = D_TYPE(tmp[0]);
-    }
-}