cmake : set include path for OpenBlas (#1830)

* Ignore metal file in spm

* Add ggml.h to spm public Headers

---------

Co-authored-by: Vogel Frederik <vogel.frederik@linecorp.com>

.devops/full.Dockerfile

@@ -16,4 +16,6 @@ COPY . .
 
 RUN make
 
+ENV LC_ALL=C.utf8
+
 ENTRYPOINT ["/app/.devops/tools.sh"]

.devops/main.Dockerfile

@@ -15,4 +15,6 @@ FROM ubuntu:$UBUNTU_VERSION as runtime
 
 COPY --from=build /app/main /main
 
+ENV LC_ALL=C.utf8
+
 ENTRYPOINT [ "/main" ]

.devops/tools.sh

@@ -11,7 +11,7 @@ shift
 arg2="$@"
 
 if [[ $arg1 == '--convert' || $arg1 == '-c' ]]; then
-    python3 ./convert-pth-to-ggml.py $arg2
+    python3 ./convert.py $arg2
 elif [[ $arg1 == '--quantize' || $arg1 == '-q' ]]; then
     ./quantize $arg2
 elif [[ $arg1 == '--run' || $arg1 == '-r' ]]; then
@@ -32,7 +32,7 @@ else
     echo "  --run (-r): Run a model previously converted into ggml"
     echo "              ex: -m /models/7B/ggml-model-q4_0.bin -p \"Building a website can be done in 10 simple steps:\" -n 512"
     echo "  --convert (-c): Convert a llama model into ggml"
-    echo "              ex: \"/models/7B/\" 1"
+    echo "              ex: --outtype f16 \"/models/7B/\" "
     echo "  --quantize (-q): Optimize with quantization process ggml"
     echo "              ex: \"/models/7B/ggml-model-f16.bin\" \"/models/7B/ggml-model-q4_0.bin\" 2"
     echo "  --all-in-one (-a): Execute --convert & --quantize"

.github/workflows/build.yml

@@ -10,10 +10,10 @@ on:
   push:
     branches:
       - master
-    paths: ['.github/workflows/**', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp']
+    paths: ['.github/workflows/**', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu']
   pull_request:
     types: [opened, synchronize, reopened]
-    paths: ['**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp']
+    paths: ['**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu']
 
 env:
  BRANCH_NAME: ${{ github.head_ref || github.ref_name }}

.github/workflows/tidy-post.yml

@@ -1,7 +1,7 @@
 name: clang-tidy review post comments
 
 on:
-  workflow_run:
+  workflow_dispatch:
     workflows: ["clang-tidy-review"]
     types:
       - completed

.gitignore

@@ -7,6 +7,7 @@
 .envrc
 .swiftpm
 .venv
+.clang-tidy
 .vs/
 .vscode/
 
@@ -17,6 +18,7 @@ build-release/
 build-static/
 build-cublas/
 build-opencl/
+build-metal/
 build-no-accel/
 build-sanitize-addr/
 build-sanitize-thread/
@@ -30,9 +32,11 @@ models/*
 /result
 /perplexity
 /embedding
+/train-text-from-scratch
 /benchmark-matmult
 /vdot
 /Pipfile
+/libllama.so
 
 build-info.h
 arm_neon.h

CMakeLists.txt

@@ -64,13 +64,15 @@ if (NOT MSVC)
 endif()
 
 # 3rd party libs
-option(LLAMA_ACCELERATE                 "llama: enable Accelerate framework"                    ON)
-option(LLAMA_BLAS                       "llama: use BLAS"                                       OFF)
+option(LLAMA_ACCELERATE                      "llama: enable Accelerate framework"               ON)
+option(LLAMA_BLAS                            "llama: use BLAS"                                  OFF)
 set(LLAMA_BLAS_VENDOR "Generic" CACHE STRING "llama: BLAS library vendor")
-option(LLAMA_CUBLAS                     "llama: use cuBLAS"                                     OFF)
-set(LLAMA_CUDA_DMMV_X "32" CACHE STRING "llama: x stride for dmmv CUDA kernels")
-set(LLAMA_CUDA_DMMV_Y "1" CACHE STRING  "llama: y block size for dmmv CUDA kernels")
-option(LLAMA_CLBLAST                    "llama: use CLBlast"                                    OFF)
+option(LLAMA_CUBLAS                          "llama: use cuBLAS"                                OFF)
+set(LLAMA_CUDA_DMMV_X      "32" CACHE STRING "llama: x stride for dmmv CUDA kernels")
+set(LLAMA_CUDA_DMMV_Y       "1" CACHE STRING "llama: y block size for dmmv CUDA kernels")
+option(LLAMA_CLBLAST                         "llama: use CLBlast"                               OFF)
+option(LLAMA_METAL                           "llama: use Metal"                                 OFF)
+option(LLAMA_K_QUANTS                        "llama: use k-quants"                              ON)
 
 option(LLAMA_BUILD_TESTS                "llama: build tests"    ${LLAMA_STANDALONE})
 option(LLAMA_BUILD_EXAMPLES             "llama: build examples" ${LLAMA_STANDALONE})
@@ -161,12 +163,24 @@ if (LLAMA_BLAS)
     if (BLAS_FOUND)
         message(STATUS "BLAS found, Libraries: ${BLAS_LIBRARIES}")
 
+        # BLAS_INCLUDE_DIRS is missing in FindBLAS.cmake.
+        # see https://gitlab.kitware.com/cmake/cmake/-/issues/20268
+        find_path(BLAS_INCLUDE_DIRS
+            NAMES cblas.h
+            HINTS
+                /usr/include
+                /usr/local/include
+                /usr/include/openblas
+        )
+
+        
+        message(STATUS "BLAS found, Includes: ${BLAS_INCLUDE_DIRS}")
+
         add_compile_options(${BLAS_LINKER_FLAGS})
         add_compile_definitions(GGML_USE_OPENBLAS)
         set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} ${BLAS_LIBRARIES})
+        set(LLAMA_EXTRA_INCLUDES ${LLAMA_EXTRA_INCLUDES} ${BLAS_INCLUDE_DIRS})
 
-        message("${BLAS_LIBRARIES} ${BLAS_INCLUDE_DIRS}")
-        include_directories(${BLAS_INCLUDE_DIRS})
     else()
         message(WARNING "BLAS not found, please refer to "
         "https://cmake.org/cmake/help/latest/module/FindBLAS.html#blas-lapack-vendors"
@@ -183,7 +197,7 @@ if (LLAMA_CUBLAS)
 
         enable_language(CUDA)
 
-        set(GGML_CUDA_SOURCES ggml-cuda.cu ggml-cuda.h)
+        set(GGML_SOURCES_CUDA ggml-cuda.cu ggml-cuda.h)
 
         add_compile_definitions(GGML_USE_CUBLAS)
         add_compile_definitions(GGML_CUDA_DMMV_X=${LLAMA_CUDA_DMMV_X})
@@ -200,12 +214,42 @@ if (LLAMA_CUBLAS)
     endif()
 endif()
 
+if (LLAMA_METAL)
+    find_library(FOUNDATION_LIBRARY         Foundation              REQUIRED)
+    find_library(METAL_FRAMEWORK            Metal                   REQUIRED)
+    find_library(METALKIT_FRAMEWORK         MetalKit                REQUIRED)
+    find_library(METALPERFORMANCE_FRAMEWORK MetalPerformanceShaders REQUIRED)
+
+    set(GGML_SOURCES_METAL ggml-metal.m ggml-metal.h)
+
+    add_compile_definitions(GGML_USE_METAL)
+    add_compile_definitions(GGML_METAL_NDEBUG)
+
+    # get full path to the file
+    #add_compile_definitions(GGML_METAL_DIR_KERNELS="${CMAKE_CURRENT_SOURCE_DIR}/")
+
+    # copy ggml-metal.metal to bin directory
+    configure_file(ggml-metal.metal bin/ggml-metal.metal COPYONLY)
+
+    set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS}
+        ${FOUNDATION_LIBRARY}
+        ${METAL_FRAMEWORK}
+        ${METALKIT_FRAMEWORK}
+        ${METALPERFORMANCE_FRAMEWORK}
+        )
+endif()
+
+if (LLAMA_K_QUANTS)
+    set(GGML_SOURCES_EXTRA ${GGML_SOURCES_EXTRA} k_quants.c k_quants.h)
+    add_compile_definitions(GGML_USE_K_QUANTS)
+endif()
+
 if (LLAMA_CLBLAST)
     find_package(CLBlast)
     if (CLBlast_FOUND)
         message(STATUS "CLBlast found")
 
-        set(GGML_OPENCL_SOURCES ggml-opencl.cpp ggml-opencl.h)
+        set(GGML_SOURCES_OPENCL ggml-opencl.cpp ggml-opencl.h)
 
         add_compile_definitions(GGML_USE_CLBLAST)
 
@@ -370,10 +414,13 @@ endif()
 add_library(ggml OBJECT
             ggml.c
             ggml.h
-            ${GGML_CUDA_SOURCES}
-            ${GGML_OPENCL_SOURCES})
+            ${GGML_SOURCES_CUDA}
+            ${GGML_SOURCES_OPENCL}
+            ${GGML_SOURCES_METAL}
+            ${GGML_SOURCES_EXTRA}
+            )
 
-target_include_directories(ggml PUBLIC .)
+target_include_directories(ggml PUBLIC . ${LLAMA_EXTRA_INCLUDES})
 target_compile_features(ggml PUBLIC c_std_11) # don't bump
 target_link_libraries(ggml PUBLIC Threads::Threads ${LLAMA_EXTRA_LIBS})
 
@@ -384,21 +431,28 @@ endif()
 add_library(llama
             llama.cpp
             llama.h
-            llama-util.h)
+            llama-util.h
+            )
 
 target_include_directories(llama PUBLIC .)
 target_compile_features(llama PUBLIC cxx_std_11) # don't bump
-target_link_libraries(llama PRIVATE ggml ${LLAMA_EXTRA_LIBS})
+target_link_libraries(llama PRIVATE
+    ggml
+    ${LLAMA_EXTRA_LIBS}
+    )
 
 if (BUILD_SHARED_LIBS)
     set_target_properties(llama PROPERTIES POSITION_INDEPENDENT_CODE ON)
     target_compile_definitions(llama PRIVATE LLAMA_SHARED LLAMA_BUILD)
+    if (LLAMA_METAL)
+        set_target_properties(llama PROPERTIES RESOURCE "${CMAKE_CURRENT_SOURCE_DIR}/ggml-metal.metal")
+    endif()
 endif()
 
-if (GGML_CUDA_SOURCES)
+if (GGML_SOURCES_CUDA)
     message(STATUS "GGML CUDA sources found, configuring CUDA architecture")
-    set_property(TARGET ggml PROPERTY CUDA_ARCHITECTURES OFF)
-    set_property(TARGET ggml PROPERTY CUDA_SELECT_NVCC_ARCH_FLAGS "Auto")
+    set_property(TARGET ggml  PROPERTY CUDA_ARCHITECTURES OFF)
+    set_property(TARGET ggml  PROPERTY CUDA_SELECT_NVCC_ARCH_FLAGS "Auto")
     set_property(TARGET llama PROPERTY CUDA_ARCHITECTURES OFF)
 endif()
 

Makefile

@@ -1,5 +1,5 @@
 # Define the default target now so that it is always the first target
-BUILD_TARGETS = main quantize quantize-stats perplexity embedding vdot
+BUILD_TARGETS = main quantize quantize-stats perplexity embedding vdot train-text-from-scratch
 
 ifdef LLAMA_BUILD_SERVER
 	BUILD_TARGETS += server
@@ -40,8 +40,11 @@ endif
 #
 
 # keep standard at C11 and C++11
-CFLAGS   = -I.              -O3 -std=c11   -fPIC
-CXXFLAGS = -I. -I./examples -O3 -std=c++11 -fPIC
+# -Ofast tends to produce faster code, but may not be available for some compilers.
+#OPT = -Ofast
+OPT = -O3
+CFLAGS   = -I.              $(OPT) -std=c11   -fPIC
+CXXFLAGS = -I. -I./examples $(OPT) -std=c++11 -fPIC
 LDFLAGS  =
 
 ifdef LLAMA_DEBUG
@@ -104,7 +107,12 @@ ifeq ($(UNAME_M),$(filter $(UNAME_M),x86_64 i686))
 	# Usage AVX-only
 	#CFLAGS   += -mfma -mf16c -mavx
 	#CXXFLAGS += -mfma -mf16c -mavx
+
+	# Usage SSSE3-only (Not is SSE3!)
+	#CFLAGS   += -mssse3
+	#CXXFLAGS += -mssse3
 endif
+
 ifneq ($(filter ppc64%,$(UNAME_M)),)
 	POWER9_M := $(shell grep "POWER9" /proc/cpuinfo)
 	ifneq (,$(findstring POWER9,$(POWER9_M)))
@@ -116,6 +124,13 @@ ifneq ($(filter ppc64%,$(UNAME_M)),)
 		CXXFLAGS += -std=c++23 -DGGML_BIG_ENDIAN
 	endif
 endif
+
+ifndef LLAMA_NO_K_QUANTS
+	CFLAGS   += -DGGML_USE_K_QUANTS
+	CXXFLAGS += -DGGML_USE_K_QUANTS
+	OBJS     += k_quants.o
+endif
+
 ifndef LLAMA_NO_ACCELERATE
 	# Mac M1 - include Accelerate framework.
 	# `-framework Accelerate` works on Mac Intel as well, with negliable performance boost (as of the predict time).
@@ -123,7 +138,8 @@ ifndef LLAMA_NO_ACCELERATE
 		CFLAGS  += -DGGML_USE_ACCELERATE
 		LDFLAGS += -framework Accelerate
 	endif
-endif
+endif # LLAMA_NO_ACCELERATE
+
 ifdef LLAMA_OPENBLAS
 	CFLAGS  += -DGGML_USE_OPENBLAS -I/usr/local/include/openblas -I/usr/include/openblas
 	ifneq ($(shell grep -e "Arch Linux" -e "ID_LIKE=arch" /etc/os-release 2>/dev/null),)
@@ -131,11 +147,13 @@ ifdef LLAMA_OPENBLAS
 	else
 		LDFLAGS += -lopenblas
 	endif
-endif
+endif # LLAMA_OPENBLAS
+
 ifdef LLAMA_BLIS
-	CFLAGS += -DGGML_USE_OPENBLAS -I/usr/local/include/blis -I/usr/include/blis
+	CFLAGS  += -DGGML_USE_OPENBLAS -I/usr/local/include/blis -I/usr/include/blis
 	LDFLAGS += -lblis -L/usr/local/lib
-endif
+endif # LLAMA_BLIS
+
 ifdef LLAMA_CUBLAS
 	CFLAGS    += -DGGML_USE_CUBLAS -I/usr/local/cuda/include -I/opt/cuda/include -I$(CUDA_PATH)/targets/x86_64-linux/include
 	CXXFLAGS  += -DGGML_USE_CUBLAS -I/usr/local/cuda/include -I/opt/cuda/include -I$(CUDA_PATH)/targets/x86_64-linux/include
@@ -156,9 +174,10 @@ endif # LLAMA_CUDA_DMMV_Y
 ggml-cuda.o: ggml-cuda.cu ggml-cuda.h
 	$(NVCC) $(NVCCFLAGS) $(CXXFLAGS) -Wno-pedantic -c $< -o $@
 endif # LLAMA_CUBLAS
+
 ifdef LLAMA_CLBLAST
-	CFLAGS  += -DGGML_USE_CLBLAST
-	CXXFLAGS  += -DGGML_USE_CLBLAST
+	CFLAGS   += -DGGML_USE_CLBLAST
+	CXXFLAGS += -DGGML_USE_CLBLAST
 	# Mac provides OpenCL as a framework
 	ifeq ($(UNAME_S),Darwin)
 		LDFLAGS += -lclblast -framework OpenCL
@@ -166,28 +185,48 @@ ifdef LLAMA_CLBLAST
 		LDFLAGS += -lclblast -lOpenCL
 	endif
 	OBJS    += ggml-opencl.o
+
 ggml-opencl.o: ggml-opencl.cpp ggml-opencl.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@
-endif
+endif # LLAMA_CLBLAST
+
+ifdef LLAMA_METAL
+	CFLAGS   += -DGGML_USE_METAL -DGGML_METAL_NDEBUG
+	CXXFLAGS += -DGGML_USE_METAL
+	LDFLAGS  += -framework Foundation -framework Metal -framework MetalKit -framework MetalPerformanceShaders
+	OBJS     += ggml-metal.o
+
+ggml-metal.o: ggml-metal.m ggml-metal.h
+	$(CC) $(CFLAGS) -c $< -o $@
+endif # LLAMA_METAL
+
 ifneq ($(filter aarch64%,$(UNAME_M)),)
 	# Apple M1, M2, etc.
 	# Raspberry Pi 3, 4, Zero 2 (64-bit)
 	CFLAGS   += -mcpu=native
 	CXXFLAGS += -mcpu=native
 endif
+
 ifneq ($(filter armv6%,$(UNAME_M)),)
 	# Raspberry Pi 1, Zero
 	CFLAGS += -mfpu=neon-fp-armv8 -mfp16-format=ieee -mno-unaligned-access
 endif
+
 ifneq ($(filter armv7%,$(UNAME_M)),)
 	# Raspberry Pi 2
 	CFLAGS += -mfpu=neon-fp-armv8 -mfp16-format=ieee -mno-unaligned-access -funsafe-math-optimizations
 endif
+
 ifneq ($(filter armv8%,$(UNAME_M)),)
 	# Raspberry Pi 3, 4, Zero 2 (32-bit)
 	CFLAGS += -mfp16-format=ieee -mno-unaligned-access
 endif
 
+ifdef LLAMA_NO_K_QUANTS
+k_quants.o: k_quants.c k_quants.h
+	$(CC) $(CFLAGS) -c $< -o $@
+endif # LLAMA_NO_K_QUANTS
+
 #
 # Print build information
 #
@@ -220,28 +259,28 @@ libllama.so: llama.o ggml.o $(OBJS)
 	$(CXX) $(CXXFLAGS) -shared -fPIC -o $@ $^ $(LDFLAGS)
 
 clean:
-	rm -vf *.o main quantize quantize-stats perplexity embedding benchmark-matmult save-load-state server vdot build-info.h
+	rm -vf *.o *.so main quantize quantize-stats perplexity embedding benchmark-matmult save-load-state server vdot train-text-from-scratch build-info.h
 
 #
 # Examples
 #
 
-main: examples/main/main.cpp build-info.h ggml.o llama.o common.o $(OBJS)
+main: examples/main/main.cpp                                  build-info.h ggml.o llama.o common.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 	@echo
 	@echo '====  Run ./main -h for help.  ===='
 	@echo
 
-quantize: examples/quantize/quantize.cpp build-info.h ggml.o llama.o $(OBJS)
+quantize: examples/quantize/quantize.cpp                      build-info.h ggml.o llama.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-quantize-stats: examples/quantize-stats/quantize-stats.cpp build-info.h ggml.o llama.o $(OBJS)
+quantize-stats: examples/quantize-stats/quantize-stats.cpp    build-info.h ggml.o llama.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-perplexity: examples/perplexity/perplexity.cpp build-info.h ggml.o llama.o common.o $(OBJS)
+perplexity: examples/perplexity/perplexity.cpp                build-info.h ggml.o llama.o common.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-embedding: examples/embedding/embedding.cpp build-info.h ggml.o llama.o common.o $(OBJS)
+embedding: examples/embedding/embedding.cpp                   build-info.h ggml.o llama.o common.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
 save-load-state: examples/save-load-state/save-load-state.cpp build-info.h ggml.o llama.o common.o $(OBJS)
@@ -250,6 +289,9 @@ save-load-state: examples/save-load-state/save-load-state.cpp build-info.h ggml.
 server: examples/server/server.cpp examples/server/httplib.h examples/server/json.hpp build-info.h ggml.o llama.o common.o $(OBJS)
 	$(CXX) $(CXXFLAGS) -Iexamples/server $(filter-out %.h,$(filter-out %.hpp,$^)) -o $@ $(LDFLAGS)
 
+train-text-from-scratch: examples/train-text-from-scratch/train-text-from-scratch.cpp    build-info.h ggml.o llama.o $(OBJS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+
 build-info.h: $(wildcard .git/index) scripts/build-info.sh
 	@sh scripts/build-info.sh > $@.tmp
 	@if ! cmp -s $@.tmp $@; then \

Package.swift

@@ -11,6 +11,7 @@ let package = Package(
         .target(
             name: "llama",
             path: ".",
+            exclude: ["ggml-metal.metal"],
             sources: ["ggml.c", "llama.cpp"],
             publicHeadersPath: "spm-headers",
             cSettings: [.unsafeFlags(["-Wno-shorten-64-to-32"]), .define("GGML_USE_ACCELERATE")],

README.md

@@ -9,9 +9,12 @@ Inference of [LLaMA](https://arxiv.org/abs/2302.13971) model in pure C/C++
 
 **Hot topics:**
 
-- Quantization formats `Q4` and `Q8` have changed again (19 May) - [(info)](https://github.com/ggerganov/llama.cpp/pull/1508)
-- Quantization formats `Q4` and `Q5` have changed - requantize any old models [(info)](https://github.com/ggerganov/llama.cpp/pull/1405)
-- [Roadmap May 2023](https://github.com/ggerganov/llama.cpp/discussions/1220)
+- Roadmap June 2023: https://github.com/ggerganov/llama.cpp/discussions/1729
+- GPU support with Metal (Apple Silicon): https://github.com/ggerganov/llama.cpp/pull/1642
+- High-quality 2,3,4,5,6-bit quantization: https://github.com/ggerganov/llama.cpp/pull/1684
+- Multi-GPU support: https://github.com/ggerganov/llama.cpp/pull/1607
+- Training LLaMA models from scratch: https://github.com/ggerganov/llama.cpp/pull/1652
+- CPU threading improvements: https://github.com/ggerganov/llama.cpp/pull/1632
 
 <details>
   <summary>Table of Contents</summary>
@@ -51,11 +54,10 @@ Inference of [LLaMA](https://arxiv.org/abs/2302.13971) model in pure C/C++
 The main goal of `llama.cpp` is to run the LLaMA model using 4-bit integer quantization on a MacBook
 
 - Plain C/C++ implementation without dependencies
-- Apple silicon first-class citizen - optimized via ARM NEON and Accelerate framework
+- Apple silicon first-class citizen - optimized via ARM NEON, Accelerate and Metal frameworks
 - AVX, AVX2 and AVX512 support for x86 architectures
 - Mixed F16 / F32 precision
 - 4-bit, 5-bit and 8-bit integer quantization support
-- Runs on the CPU
 - Supports OpenBLAS/Apple BLAS/ARM Performance Lib/ATLAS/BLIS/Intel MKL/NVHPC/ACML/SCSL/SGIMATH and [more](https://cmake.org/cmake/help/latest/module/FindBLAS.html#blas-lapack-vendors) in BLAS
 - cuBLAS and CLBlast support
 
@@ -236,15 +238,41 @@ In order to build llama.cpp you have three different options.
     zig build -Drelease-fast
     ```
 
+### Metal Build
+
+Using Metal allows the computation to be executed on the GPU for Apple devices:
+
+- Using `make`:
+
+  ```bash
+  LLAMA_METAL=1 make
+  ```
+
+- Using `CMake`:
+
+    ```bash
+    mkdir build-metal
+    cd build-metal
+    cmake -DLLAMA_METAL=ON ..
+    cmake --build . --config Release
+    ```
+
+When built with Metal support, you can enable GPU inference with the `--gpu-layers|-ngl` command-line argument.
+Any value larger than 0 will offload the computation to the GPU. For example:
+
+```bash
+./main -m ./models/7B/ggml-model-q4_0.bin -n 128 -ngl 1
+```
+
 ### BLAS Build
 
 Building the program with BLAS support may lead to some performance improvements in prompt processing using batch sizes higher than 32 (the default is 512). BLAS doesn't affect the normal generation performance. There are currently three different implementations of it:
 
-- **Accelerate Framework**:
+- #### Accelerate Framework:
 
   This is only available on Mac PCs and it's enabled by default. You can just build using the normal instructions.
 
-- **OpenBLAS**:
+- #### OpenBLAS:
 
   This provides BLAS acceleration using only the CPU. Make sure to have OpenBLAS installed on your machine.
 
@@ -278,11 +306,11 @@ Building the program with BLAS support may lead to some performance improvements
       cmake --build . --config Release
       ```
 
-- **BLIS**
+- #### BLIS
 
-  Check [BLIS.md](BLIS.md) for more information.
+  Check [BLIS.md](docs/BLIS.md) for more information.
 
-- **Intel MKL**
+- #### Intel MKL
 
   By default, `LLAMA_BLAS_VENDOR` is set to `Generic`, so if you already sourced intel environment script and assign `-DLLAMA_BLAS=ON` in cmake, the mkl version of Blas will automatically been selected. You may also specify it by:
 
@@ -290,10 +318,10 @@ Building the program with BLAS support may lead to some performance improvements
   mkdir build
   cd build
   cmake .. -DLLAMA_BLAS=ON -DLLAMA_BLAS_VENDOR=Intel10_64lp -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
-  cmake --build . -config Release
+  cmake --build . --config Release
   ```
 
-- **cuBLAS**
+- #### cuBLAS
 
   This provides BLAS acceleration using the CUDA cores of your Nvidia GPU. Make sure to have the CUDA toolkit installed. You can download it from your Linux distro's package manager or from here: [CUDA Toolkit](https://developer.nvidia.com/cuda-downloads).
   - Using `make`:
@@ -310,7 +338,9 @@ Building the program with BLAS support may lead to some performance improvements
     ```
   Note: Because llama.cpp uses multiple CUDA streams for matrix multiplication results [are not guaranteed to be reproducible](https://docs.nvidia.com/cuda/cublas/index.html#results-reproducibility). If you need reproducibility, set `GGML_CUDA_MAX_STREAMS` in the file `ggml-cuda.cu` to 1.
 
-- **CLBlast**
+  The environment variable [`CUDA_VISIBLE_DEVICES`](https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#env-vars) can be used to specify which GPU(s) will be used.
+
+- #### CLBlast
 
   OpenCL acceleration is provided by the matrix multiplication kernels from the [CLBlast](https://github.com/CNugteren/CLBlast) project and custom kernels for ggml that can generate tokens on the GPU.
 
@@ -348,7 +378,7 @@ Building the program with BLAS support may lead to some performance improvements
       cmake --install . --prefix /some/path
       ```
 
-      Where `/some/path` is where the built library will be installed (default is `/usr/loca`l`).
+      Where `/some/path` is where the built library will be installed (default is `/usr/local`).
     </details>
 
   Building:
@@ -367,7 +397,7 @@ Building the program with BLAS support may lead to some performance improvements
 
   Running:
 
-  The CLBlast build supports `--gpu-layers|-ngl` like  the CUDA version does.
+  The CLBlast build supports `--gpu-layers|-ngl` like the CUDA version does.
 
   To select the correct platform (driver) and device (GPU), you can use the environment variables `GGML_OPENCL_PLATFORM` and `GGML_OPENCL_DEVICE`.
   The selection can be a number (starting from 0) or a text string to search:
@@ -655,3 +685,4 @@ docker run -v /path/to/models:/models ghcr.io/ggerganov/llama.cpp:light -m /mode
 ### Docs
 
 - [GGML tips & tricks](https://github.com/ggerganov/llama.cpp/wiki/GGML-Tips-&-Tricks)
+- [Performance troubleshooting](./docs/token_generation_performance_tips.md)

SHA256SUMS

@@ -1,6 +1,6 @@
 700df0d3013b703a806d2ae7f1bfb8e59814e3d06ae78be0c66368a50059f33d  models/7B/consolidated.00.pth
 666a4bb533b303bdaf89e1b6a3b6f93535d868de31d903afdc20983dc526c847  models/7B/ggml-model-f16.bin
-ffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff  models/7B/ggml-model-q4_0.bin
+ec2f2d1f0dfb73b72a4cbac7fa121abbe04c37ab327125a38248f930c0f09ddf  models/7B/ggml-model-q4_0.bin
 ffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff  models/7B/ggml-model-q4_1.bin
 ffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff  models/7B/ggml-model-q5_0.bin
 ffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff  models/7B/ggml-model-q5_1.bin
@@ -8,7 +8,7 @@ ffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff  models/7B/ggml
 745bf4e29a4dd6f411e72976d92b452da1b49168a4f41c951cfcc8051823cf08  models/13B/consolidated.00.pth
 d5ccbcc465c71c0de439a5aeffebe8344c68a519bce70bc7f9f92654ee567085  models/13B/consolidated.01.pth
 2b206e9b21fb1076f11cafc624e2af97c9e48ea09312a0962153acc20d45f808  models/13B/ggml-model-f16.bin
-ffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff  models/13B/ggml-model-q4_0.bin
+fad169e6f0f575402cf75945961cb4a8ecd824ba4da6be2af831f320c4348fa5  models/13B/ggml-model-q4_0.bin
 ffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff  models/13B/ggml-model-q4_1.bin
 ffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff  models/13B/ggml-model-q5_0.bin
 ffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff  models/13B/ggml-model-q5_1.bin
@@ -18,7 +18,7 @@ e23294a58552d8cdec5b7e8abb87993b97ea6eced4178ff2697c02472539d067  models/30B/con
 24a87f01028cbd3a12de551dcedb712346c0b5cbdeff1454e0ddf2df9b675378  models/30B/consolidated.02.pth
 1adfcef71420886119544949767f6a56cb6339b4d5fcde755d80fe68b49de93b  models/30B/consolidated.03.pth
 7e1b524061a9f4b27c22a12d6d2a5bf13b8ebbea73e99f218809351ed9cf7d37  models/30B/ggml-model-f16.bin
-ffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff  models/30B/ggml-model-q4_0.bin
+d2a441403944819492ec8c2002cc36fa38468149bfb4b7b4c52afc7bd9a7166d  models/30B/ggml-model-q4_0.bin
 ffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff  models/30B/ggml-model-q4_1.bin
 ffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff  models/30B/ggml-model-q5_0.bin
 ffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff  models/30B/ggml-model-q5_1.bin
@@ -32,7 +32,7 @@ a287c0dfe49081626567c7fe87f74cce5831f58e459b427b5e05567641f47b78  models/65B/con
 72b4eba67a1a3b18cb67a85b70f8f1640caae9b40033ea943fb166bd80a7b36b  models/65B/consolidated.06.pth
 d27f5b0677d7ff129ceacd73fd461c4d06910ad7787cf217b249948c3f3bc638  models/65B/consolidated.07.pth
 60758f2384d74e423dffddfd020ffed9d3bb186ebc54506f9c4a787d0f5367b0  models/65B/ggml-model-f16.bin
-ffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff  models/65B/ggml-model-q4_0.bin
+cde053439fa4910ae454407e2717cc46cc2c2b4995c00c93297a2b52e790fa92  models/65B/ggml-model-q4_0.bin
 ffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff  models/65B/ggml-model-q4_1.bin
 ffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff  models/65B/ggml-model-q5_0.bin
 ffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff  models/65B/ggml-model-q5_1.bin

convert-pth-to-ggml.py

@@ -4,7 +4,9 @@ import argparse
 
 import convert
 
-parser = argparse.ArgumentParser(description='Convert a LLaMA model checkpoint to a ggml compatible file')
+parser = argparse.ArgumentParser(
+    description="""[DEPRECATED - use `convert.py` instead]
+    Convert a LLaMA model checkpoint to a ggml compatible file""")
 parser.add_argument('dir_model',  help='directory containing the model checkpoint')
 parser.add_argument('ftype',      help='file type (0: float32, 1: float16)', type=int, choices=[0, 1], default=1)
 args = parser.parse_args()

docs/token_generation_performance_tips.md

@@ -0,0 +1,40 @@
+# Token generation performance troubleshooting
+
+## Verifying that the model is running on the GPU with cuBLAS
+Make sure you compiled llama with the correct env variables according to [this guide](../README.md#cublas), so that llama accepts the `-ngl N` (or `--n-gpu-layers N`) flag. When running llama, you may configure `N` to be very large, and llama will offload the maximum possible number of layers to the GPU, even if it's less than the number you configured. For example:
+```shell
+./main -m "path/to/model.bin" -ngl 200000 -p "Please sir, may I have some "
+```
+
+When running llama, before it starts the inference work, it will output diagnostic information that shows whether cuBLAS is offloading work to the GPU. Look for these lines:
+```shell
+llama_model_load_internal: [cublas] offloading 60 layers to GPU
+llama_model_load_internal: [cublas] offloading output layer to GPU
+llama_model_load_internal: [cublas] total VRAM used: 17223 MB
+... rest of inference
+```
+
+If you see these lines, then the GPU is being used.
+
+## Verifying that the CPU is not oversaturated
+llama accepts a `-t N` (or `--threads N`) parameter. It's extremely important that this parameter is not too large. If your token generation is extremely slow, try setting this number to 1. If this significantly improves your token generation speed, then your CPU is being oversaturated and you need to explicitly set this parameter to the number of the physicial CPU cores on your machine (even if you utilize a GPU). If in doubt, start with 1 and double the amount until you hit a performance bottleneck, then scale the number down.
+
+# Example of runtime flags effect on inference speed benchmark
+These runs were tested on the following machine:
+GPU: A6000 (48GB VRAM)
+CPU: 7 physical cores
+RAM: 32GB
+
+Model: `TheBloke_Wizard-Vicuna-30B-Uncensored-GGML/Wizard-Vicuna-30B-Uncensored.ggmlv3.q4_0.bin` (30B parameters, 4bit quantization, GGML)
+
+Run command: `./main -m "path/to/model.bin" -p "-p "An extremely detailed description of the 10 best ethnic dishes will follow, with recipes: " -n 1000 [additional benchmark flags]`
+
+Result:
+
+| command | tokens/second (higher is better) |
+| - | - |
+| -ngl 2000000 | N/A (less than 0.1) |
+| -t 7 | 1.7 |
+| -t 1 -ngl 2000000 | 5.5 |
+| -t 7 -ngl 2000000 | 8.7 |
+| -t 4 -ngl 2000000 | 9.1 |

examples/CMakeLists.txt

@@ -37,7 +37,11 @@ else()
     add_subdirectory(save-load-state)
     add_subdirectory(benchmark)
     add_subdirectory(baby-llama)
-    if(LLAMA_BUILD_SERVER)
+    add_subdirectory(train-text-from-scratch)
+    if (LLAMA_METAL)
+        add_subdirectory(metal)
+    endif()
+    if (LLAMA_BUILD_SERVER)
         add_subdirectory(server)
     endif()
 endif()

examples/baby-llama/baby-llama.cpp

@@ -79,34 +79,39 @@ struct ggml_tensor * randomize_tensor_normal(
         int ndims,
         const int64_t ne[],
         struct random_normal_distribution * rnd) {
+    float scale = 1.0; // xavier
     switch (ndims) {
         case 1:
+            scale /= sqrtf(ne[0]);
             for (int i0 = 0; i0 < ne[0]; i0++) {
-                ((float *)tensor->data)[i0] = frand_normal(rnd);
+                ((float *)tensor->data)[i0] = scale * frand_normal(rnd);
             }
             break;
         case 2:
+            scale /= sqrtf(ne[0]+ne[1]);
             for (int i1 = 0; i1 < ne[1]; i1++) {
                 for (int i0 = 0; i0 < ne[0]; i0++) {
-                    ((float *)tensor->data)[i1*ne[0] + i0] = frand_normal(rnd);
+                    ((float *)tensor->data)[i1*ne[0] + i0] = scale * frand_normal(rnd);
                 }
             }
             break;
         case 3:
+            scale /= sqrtf(ne[0]+ne[1]);
             for (int i2 = 0; i2 < ne[2]; i2++) {
                 for (int i1 = 0; i1 < ne[1]; i1++) {
                     for (int i0 = 0; i0 < ne[0]; i0++) {
-                        ((float *)tensor->data)[i2*ne[1]*ne[0] + i1*ne[0] + i0] = frand_normal(rnd);
+                        ((float *)tensor->data)[i2*ne[1]*ne[0] + i1*ne[0] + i0] = scale * frand_normal(rnd);
                     }
                 }
             }
             break;
         case 4:
+            scale /= sqrtf(ne[0]+ne[1]);
             for (int i3 = 0; i3 < ne[3]; i3++) {
                 for (int i2 = 0; i2 < ne[2]; i2++) {
                     for (int i1 = 0; i1 < ne[1]; i1++) {
                         for (int i0 = 0; i0 < ne[0]; i0++) {
-                            ((float *)tensor->data)[i3*ne[2]*ne[1]*ne[0] + i2*ne[1]*ne[0] + i1*ne[0] + i0] = frand_normal(rnd);
+                            ((float *)tensor->data)[i3*ne[2]*ne[1]*ne[0] + i2*ne[1]*ne[0] + i1*ne[0] + i0] = scale * frand_normal(rnd);
                         }
                     }
                 }
@@ -148,8 +153,8 @@ struct llama_hparams_lora {
     uint32_t n_rot   = 64;
     uint32_t n_lora  = 64;
 
-    bool operator!=(const llama_hparams & other) const {
-        return memcmp(this, &other, sizeof(llama_hparams));
+    bool operator!=(const llama_hparams_lora & other) const {
+        return memcmp(this, &other, sizeof(llama_hparams_lora)) != 0;
     }
 };
 

examples/common.cpp

@@ -9,6 +9,7 @@
 #include <algorithm>
 #include <sstream>
 #include <unordered_set>
+#include <regex>
 
 #if defined(__APPLE__) && defined(__MACH__)
 #include <sys/types.h>
@@ -131,6 +132,8 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
             params.path_prompt_cache = argv[i];
         } else if (arg == "--prompt-cache-all") {
             params.prompt_cache_all = true;
+        } else if (arg == "--prompt-cache-ro") {
+            params.prompt_cache_ro = true;
         } else if (arg == "-f" || arg == "--file") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -295,10 +298,52 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
             fprintf(stderr, "warning: not compiled with GPU offload support, --n-gpu-layers option will be ignored\n");
             fprintf(stderr, "warning: see main README.md for information on enabling GPU BLAS support\n");
 #endif
+        } else if (arg == "--main-gpu" || arg == "-mg") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+#ifdef GGML_USE_CUBLAS
+            params.main_gpu = std::stoi(argv[i]);
+#else
+      fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS. It is not possible to set a main GPU.\n");
+#endif
+        } else if (arg == "--tensor-split" || arg == "-ts") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+#ifdef GGML_USE_CUBLAS
+            std::string arg_next = argv[i];
+
+            // split string by , and /
+            const std::regex regex{R"([,/]+)"};
+            std::sregex_token_iterator it{arg_next.begin(), arg_next.end(), regex, -1};
+            std::vector<std::string> split_arg{it, {}};
+            GGML_ASSERT(split_arg.size() <= LLAMA_MAX_DEVICES);
+
+            for (size_t i = 0; i < LLAMA_MAX_DEVICES; ++i) {
+                if (i < split_arg.size()) {
+                    params.tensor_split[i] = std::stof(split_arg[i]);
+                } else {
+                    params.tensor_split[i] = 0.0f;
+                }
+            }
+#else
+      fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS. It is not possible to set a tensor split.\n");
+#endif // GGML_USE_CUBLAS
+        } else if (arg == "--low-vram" || arg == "-lv") {
+#ifdef GGML_USE_CUBLAS
+            params.low_vram = true;
+#else
+      fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS. It is not possible to set lower vram usage.\n");
+#endif // GGML_USE_CUBLAS
         } else if (arg == "--no-mmap") {
             params.use_mmap = false;
         } else if (arg == "--mtest") {
             params.mem_test = true;
+        } else if (arg == "--export") {
+            params.export_cgraph = true;
         } else if (arg == "--verbose-prompt") {
             params.verbose_prompt = true;
         } else if (arg == "-r" || arg == "--reverse-prompt") {
@@ -367,6 +412,14 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
         gpt_print_usage(argc, argv, default_params);
         exit(1);
     }
+
+#ifdef GGML_USE_CUBLAS
+    if (!params.lora_adapter.empty() && params.n_gpu_layers > 0) {
+        fprintf(stderr, "%s: error: the simultaneous use of LoRAs and GPU acceleration is not supported", __func__);
+        exit(1);
+    }
+#endif // GGML_USE_CUBLAS
+
     if (escape_prompt) {
         process_escapes(params.prompt);
     }
@@ -395,6 +448,7 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     fprintf(stderr, "  --prompt-cache FNAME  file to cache prompt state for faster startup (default: none)\n");
     fprintf(stderr, "  --prompt-cache-all    if specified, saves user input and generations to cache as well.\n");
     fprintf(stderr, "                        not supported with --interactive or other interactive options\n");
+    fprintf(stderr, "  --prompt-cache-ro     if specified, uses the prompt cache but does not update it.\n");
     fprintf(stderr, "  --random-prompt       start with a randomized prompt.\n");
     fprintf(stderr, "  --in-prefix STRING    string to prefix user inputs with (default: empty)\n");
     fprintf(stderr, "  --in-suffix STRING    string to suffix after user inputs with (default: empty)\n");
@@ -436,8 +490,13 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
 #ifdef LLAMA_SUPPORTS_GPU_OFFLOAD
     fprintf(stderr, "  -ngl N, --n-gpu-layers N\n");
     fprintf(stderr, "                        number of layers to store in VRAM\n");
+    fprintf(stderr, "  -ts SPLIT --tensor-split SPLIT\n");
+    fprintf(stderr, "                        how to split tensors across multiple GPUs, comma-separated list of proportions, e.g. 3,1\n");
+    fprintf(stderr, "  -mg i, --main-gpu i   the GPU to use for scratch and small tensors\n" );
+    fprintf(stderr, "  -lv, --low-vram       don't allocate VRAM scratch buffer\n" );
 #endif
     fprintf(stderr, "  --mtest               compute maximum memory usage\n");
+    fprintf(stderr, "  --export              export the computation graph to 'llama.ggml'\n");
     fprintf(stderr, "  --verbose-prompt      print prompt before generation\n");
     fprintf(stderr, "  --lora FNAME          apply LoRA adapter (implies --no-mmap)\n");
     fprintf(stderr, "  --lora-base FNAME     optional model to use as a base for the layers modified by the LoRA adapter\n");
@@ -480,7 +539,11 @@ struct llama_context * llama_init_from_gpt_params(const gpt_params & params) {
     auto lparams = llama_context_default_params();
 
     lparams.n_ctx        = params.n_ctx;
+    lparams.n_batch      = params.n_batch;
     lparams.n_gpu_layers = params.n_gpu_layers;
+    lparams.main_gpu     = params.main_gpu;
+    memcpy(lparams.tensor_split, params.tensor_split, LLAMA_MAX_DEVICES*sizeof(float));
+    lparams.low_vram     = params.low_vram;
     lparams.seed         = params.seed;
     lparams.f16_kv       = params.memory_f16;
     lparams.use_mmap     = params.use_mmap;
@@ -585,6 +648,9 @@ void console_set_color(console_state & con_st, console_color_t color) {
             case CONSOLE_COLOR_USER_INPUT:
                 fprintf(con_st.out, ANSI_BOLD ANSI_COLOR_GREEN);
                 break;
+            case CONSOLE_COLOR_ERROR:
+                fprintf(con_st.out, ANSI_BOLD ANSI_COLOR_RED);
+                break;
         }
         con_st.color = color;
         fflush(con_st.out);

examples/common.h

@@ -21,13 +21,16 @@
 int32_t get_num_physical_cores();
 
 struct gpt_params {
-    int32_t seed          = -1;  // RNG seed
-    int32_t n_threads     = get_num_physical_cores();
-    int32_t n_predict     = -1;  // new tokens to predict
-    int32_t n_ctx         = 512; // context size
-    int32_t n_batch       = 512; // batch size for prompt processing (must be >=32 to use BLAS)
-    int32_t n_keep        = 0;   // number of tokens to keep from initial prompt
-    int32_t n_gpu_layers  = 0;   // number of layers to store in VRAM
+    int32_t seed                            = -1;  // RNG seed
+    int32_t n_threads                       = get_num_physical_cores();
+    int32_t n_predict                       = -1;  // new tokens to predict
+    int32_t n_ctx                           = 512; // context size
+    int32_t n_batch                         = 512; // batch size for prompt processing (must be >=32 to use BLAS)
+    int32_t n_keep                          = 0;   // number of tokens to keep from initial prompt
+    int32_t n_gpu_layers                    = 0;   // number of layers to store in VRAM
+    int32_t main_gpu                        = 0;   // the GPU that is used for scratch and small tensors
+    float   tensor_split[LLAMA_MAX_DEVICES] = {0}; // how split tensors should be distributed across GPUs
+    bool    low_vram                        = 0;   // if true, reduce VRAM usage at the cost of performance
 
     // sampling parameters
     std::unordered_map<llama_token, float> logit_bias; // logit bias for specific tokens
@@ -60,6 +63,7 @@ struct gpt_params {
     bool use_color         = false; // use color to distinguish generations and inputs
     bool interactive       = false; // interactive mode
     bool prompt_cache_all  = false; // save user input and generations to prompt cache
+    bool prompt_cache_ro   = false; // open the prompt cache read-only and do not update it
 
     bool embedding         = false; // get only sentence embedding
     bool interactive_first = false; // wait for user input immediately
@@ -71,6 +75,7 @@ struct gpt_params {
     bool use_mmap          = true;  // use mmap for faster loads
     bool use_mlock         = false; // use mlock to keep model in memory
     bool mem_test          = false; // compute maximum memory usage
+    bool export_cgraph     = false; // export the computation graph
     bool verbose_prompt    = false; // print prompt tokens before generation
 };
 
@@ -108,7 +113,8 @@ struct llama_context * llama_init_from_gpt_params(const gpt_params & params);
 enum console_color_t {
     CONSOLE_COLOR_DEFAULT=0,
     CONSOLE_COLOR_PROMPT,
-    CONSOLE_COLOR_USER_INPUT
+    CONSOLE_COLOR_USER_INPUT,
+    CONSOLE_COLOR_ERROR
 };
 
 struct console_state {

examples/main/README.md

@@ -286,5 +286,8 @@ These options provide extra functionality and customization when running the LLa
 -   `--verbose-prompt`: Print the prompt before generating text.
 -   `--mtest`: Test the model's functionality by running a series of tests to ensure it's working properly.
 -   `-ngl N, --n-gpu-layers N`: When compiled with appropriate support (currently CLBlast or cuBLAS), this option allows offloading some layers to the GPU for computation. Generally results in increased performance.
+-   `-mg i, --main-gpu i`: When using multiple GPUs this option controls which GPU is used for small tensors for which the overhead of splitting the computation across all GPUs is not worthwhile. The GPU in question will use slightly more VRAM to store a scratch buffer for temporary results. By default GPU 0 is used. Requires cuBLAS.
+-   `-ts SPLIT, --tensor-split SPLIT`: When using multiple GPUs this option controls how large tensors should be split across all GPUs. `SPLIT` is a comma-separated list of non-negative values that assigns the proportion of data that each GPU should get in order. For example, "3,2" will assign 60% of the data to GPU 0 and 40% to GPU 1. By default the data is split in proportion to VRAM but this may not be optimal for performance. Requires cuBLAS.
+-   `-lv, --low-vram`: Do not allocate a VRAM scratch buffer for holding temporary results. Reduces VRAM usage at the cost of performance, particularly prompt processing speed. Requires cuBLAS.
 -   `--lora FNAME`: Apply a LoRA (Low-Rank Adaptation) adapter to the model (implies --no-mmap). This allows you to adapt the pretrained model to specific tasks or domains.
 -   `--lora-base FNAME`: Optional model to use as a base for the layers modified by the LoRA adapter. This flag is used in conjunction with the `--lora` flag, and specifies the base model for the adaptation.

examples/main/main.cpp

@@ -81,6 +81,9 @@ int main(int argc, char ** argv) {
     if (params.n_ctx > 2048) {
         fprintf(stderr, "%s: warning: model does not support context sizes greater than 2048 tokens (%d specified);"
                 "expect poor results\n", __func__, params.n_ctx);
+    } else if (params.n_ctx < 8) {
+        fprintf(stderr, "%s: warning: minimum context size is 8, using minimum size.\n", __func__);
+        params.n_ctx = 8;
     }
 
     fprintf(stderr, "%s: build = %d (%s)\n", __func__, BUILD_NUMBER, BUILD_COMMIT);
@@ -134,6 +137,13 @@ int main(int argc, char ** argv) {
         return 0;
     }
 
+    // export the cgraph and exit
+    if (params.export_cgraph) {
+        llama_eval_export(ctx, "llama.ggml");
+        llama_free(ctx);
+
+        return 0;
+    }
 
     std::string path_session = params.path_prompt_cache;
     std::vector<llama_token> session_tokens;
@@ -202,6 +212,13 @@ int main(int argc, char ** argv) {
         }
     }
 
+    // if we will use the cache for the full prompt without reaching the end of the cache, force
+    // reevaluation of the last token token to recalculate the cached logits
+    if (!embd_inp.empty() && n_matching_session_tokens == embd_inp.size() &&
+            session_tokens.size() > embd_inp.size()) {
+        session_tokens.resize(embd_inp.size() - 1);
+    }
+
     // number of tokens to keep when resetting context
     if (params.n_keep < 0 || params.n_keep > (int) embd_inp.size() || params.instruct) {
         params.n_keep = (int)embd_inp.size();
@@ -314,9 +331,29 @@ int main(int argc, char ** argv) {
 
     std::vector<llama_token> embd;
 
+    // do one empty run to warm up the model
+    {
+        const std::vector<llama_token> tmp = { llama_token_bos(), };
+        llama_eval(ctx, tmp.data(), tmp.size(), 0, params.n_threads);
+        llama_reset_timings(ctx);
+    }
+
     while ((n_remain != 0 && !is_antiprompt) || params.interactive) {
         // predict
         if (embd.size() > 0) {
+            // Note: n_ctx - 4 here is to match the logic for commandline prompt handling via
+            // --prompt or --file which uses the same value.
+            auto max_embd_size = n_ctx - 4;
+            // Ensure the input doesn't exceed the context size by truncating embd if necessary.
+            if ((int)embd.size() > max_embd_size) {
+                auto skipped_tokens = embd.size() - max_embd_size;
+                console_set_color(con_st, CONSOLE_COLOR_ERROR);
+                printf("<<input too long: skipped %ld token%s>>", skipped_tokens, skipped_tokens != 1 ? "s" : "");
+                console_set_color(con_st, CONSOLE_COLOR_DEFAULT);
+                fflush(stdout);
+                embd.resize(max_embd_size);
+            }
+
             // infinite text generation via context swapping
             // if we run out of context:
             // - take the n_keep first tokens from the original prompt (via n_past)
@@ -360,12 +397,6 @@ int main(int argc, char ** argv) {
                     }
                 }
                 if (i > 0) {
-                    // check if we've used up all the prompt but not all cached tokens
-                    if (embd.size() == i && n_session_consumed < (int) session_tokens.size()) {
-                        // force revaluation of the last token to recalculate logits
-                        i--;
-                        n_past--;
-                    }
                     embd.erase(embd.begin(), embd.begin() + i);
                 }
             }
@@ -409,7 +440,7 @@ int main(int argc, char ** argv) {
             const bool    penalize_nl     = params.penalize_nl;
 
             // optionally save the session on first sample (for faster prompt loading next time)
-            if (!path_session.empty() && need_to_save_session) {
+            if (!path_session.empty() && need_to_save_session && !params.prompt_cache_ro) {
                 need_to_save_session = false;
                 llama_save_session_file(ctx, path_session.c_str(), session_tokens.data(), session_tokens.size());
             }
@@ -622,7 +653,7 @@ int main(int argc, char ** argv) {
         }
     }
 
-    if (!path_session.empty() && params.prompt_cache_all) {
+    if (!path_session.empty() && params.prompt_cache_all && !params.prompt_cache_ro) {
         fprintf(stderr, "\n%s: saving final output to session file '%s'\n", __func__, path_session.c_str());
         llama_save_session_file(ctx, path_session.c_str(), session_tokens.data(), session_tokens.size());
     }

examples/metal/CMakeLists.txt

@@ -0,0 +1,3 @@
+set(TEST_TARGET metal)
+add_executable(${TEST_TARGET} metal.cpp)
+target_link_libraries(${TEST_TARGET} PRIVATE ggml)

examples/metal/metal.cpp

@@ -0,0 +1,102 @@
+// Evaluate a statically exported ggml computation graph with Metal
+//
+// - First, export a LLaMA graph:
+//
+//  $ ./bin/main -m ../models/7B/ggml-model-q4_0.bin --export
+//
+// - Run this tool to evaluate the exported graph:
+//
+//  $ ./bin/metal llama.ggml
+//
+// The purpose of this tool is mostly for debugging and demonstration purposes.
+// The main limitation of exporting computation graphs is that their sizes are static which often
+// can be a problem for real-world applications.
+//
+
+#include "ggml.h"
+#include "ggml-metal.h"
+
+#include <cstdio>
+#include <cstring>
+#include <cstdlib>
+
+int main(int argc, char ** argv) {
+    ggml_time_init();
+
+    if (argc != 2) {
+        fprintf(stderr, "Usage: %s llama.ggml\n", argv[0]);
+        return -1;
+    }
+
+    const char * fname_cgraph = argv[1];
+
+    // load the compute graph
+    struct ggml_context * ctx_data = NULL;
+    struct ggml_context * ctx_eval = NULL;
+
+    struct ggml_cgraph gf = ggml_graph_import(fname_cgraph, &ctx_data, &ctx_eval);
+    gf.n_threads = 1;
+
+    // this allocates all Metal resources and memory buffers
+    auto * ctx_metal = ggml_metal_init();
+
+    ggml_metal_add_buffer(ctx_metal, "data", ggml_get_mem_buffer(ctx_data), ggml_get_mem_size(ctx_data));
+    ggml_metal_add_buffer(ctx_metal, "eval", ggml_get_mem_buffer(ctx_eval), ggml_get_mem_size(ctx_eval));
+
+    // main
+    {
+        struct ggml_tensor * input = ggml_graph_get_tensor(&gf, "embd");
+        *(int32_t *) input->data = 1; // BOS
+
+        ggml_metal_set_tensor(ctx_metal, input);
+
+        // warmup
+        ggml_metal_graph_compute(ctx_metal, &gf);
+
+        const int n_iter = 16;
+
+        const int64_t t0 = ggml_time_us();
+
+        // the actual inference happens here
+        for (int i = 0; i < n_iter; ++i) {
+            ggml_metal_graph_compute(ctx_metal, &gf);
+        }
+
+        const int64_t t1 = ggml_time_us();
+
+        printf("time: %.2f ms, %.2f ms/tok\n", (t1 - t0) / 1000.0, (t1 - t0) / 1000.0 / n_iter);
+    }
+
+    // debug output
+    {
+        struct ggml_tensor * logits = gf.nodes[gf.n_nodes - 1];
+        ggml_metal_get_tensor(ctx_metal, logits);
+
+        float * ptr = (float *) ggml_get_data(logits);
+
+        printf("logits: ");
+        for (int i = 0; i < 10; i++) {
+            printf("%8.4f ", ptr[i]);
+        }
+        printf("\n");
+        int imax = 0;
+        double sum = 0.0;
+        double vmax = -1e9;
+        for (int i = 0; i < 32000; i++) {
+            sum += (double) ptr[i];
+            if (ptr[i] > vmax) {
+                vmax = ptr[i];
+                imax = i;
+            }
+        }
+        printf("sum: %f, imax = %d, vmax = %f\n", sum, imax, vmax);
+    }
+
+    ggml_metal_free(ctx_metal);
+
+    ggml_free(ctx_data);
+    ggml_free(ctx_eval);
+
+    return 0;
+}
+

examples/quantize-stats/quantize-stats.cpp

@@ -282,8 +282,9 @@ int main(int argc, char ** argv) {
                 break;
             }
             int j;
-            for (j = 0; j < GGML_TYPE_COUNT && strcmp(argv[i], ggml_type_name((ggml_type) j)) != 0; j++) {
-                // find match
+            for (j = 0; j < GGML_TYPE_COUNT; ++j) {
+               const auto * name = ggml_type_name((ggml_type) j);
+               if (name && strcmp(argv[i], name) == 0) break;
             }
             if (j < GGML_TYPE_COUNT) {
                 params.include_types.push_back((ggml_type) j);

examples/quantize/quantize.cpp

@@ -3,31 +3,136 @@
 #include "llama.h"
 
 #include <cstdio>
-#include <map>
+#include <cstring>
+#include <vector>
 #include <string>
 
-static const std::map<std::string, llama_ftype> LLAMA_FTYPE_MAP = {
-  {"q4_0", LLAMA_FTYPE_MOSTLY_Q4_0},
-  {"q4_1", LLAMA_FTYPE_MOSTLY_Q4_1},
-  {"q5_0", LLAMA_FTYPE_MOSTLY_Q5_0},
-  {"q5_1", LLAMA_FTYPE_MOSTLY_Q5_1},
-  {"q8_0", LLAMA_FTYPE_MOSTLY_Q8_0},
+struct quant_option {
+    std::string name;
+    llama_ftype ftype;
+    std::string desc;
 };
 
-bool try_parse_ftype(const std::string & ftype_str, llama_ftype & ftype, std::string & ftype_str_out) {
-    auto it = LLAMA_FTYPE_MAP.find(ftype_str);
-    if (it != LLAMA_FTYPE_MAP.end()) {
-        ftype = it->second;
-        ftype_str_out = it->first;
-        return true;
+static const std::vector<struct quant_option> QUANT_OPTIONS = {
+    {
+        "Q4_0",
+        LLAMA_FTYPE_MOSTLY_Q4_0,
+        " 3.50G, +0.2499 ppl @ 7B - small, very high quality loss - legacy, prefer using Q3_K_M",
+    },
+    {
+        "Q4_1",
+        LLAMA_FTYPE_MOSTLY_Q4_1,
+        " 3.90G, +0.1846 ppl @ 7B - small, substantial quality loss - legacy, prefer using Q3_K_L",
+    },
+    {
+        "Q5_0",
+        LLAMA_FTYPE_MOSTLY_Q5_0,
+        " 4.30G, +0.0796 ppl @ 7B - medium, balanced quality - legacy, prefer using Q4_K_M",
+    },
+    {
+        "Q5_1",
+        LLAMA_FTYPE_MOSTLY_Q5_1,
+        " 4.70G, +0.0415 ppl @ 7B - medium, low quality loss - legacy, prefer using Q5_K_M",
+    },
+#ifdef GGML_USE_K_QUANTS
+    {
+        "Q2_K",
+        LLAMA_FTYPE_MOSTLY_Q2_K,
+        " 2.67G, +0.8698 ppl @ 7B - smallest, extreme quality loss - not recommended",
+    },
+    {
+        "Q3_K",
+        LLAMA_FTYPE_MOSTLY_Q3_K_M,
+        "alias for Q3_K_M"
+    },
+    {
+        "Q3_K_S",
+        LLAMA_FTYPE_MOSTLY_Q3_K_S,
+        " 2.75G, +0.5505 ppl @ 7B - very small, very high quality loss",
+    },
+    {
+        "Q3_K_M",
+        LLAMA_FTYPE_MOSTLY_Q3_K_M,
+        " 3.06G, +0.2437 ppl @ 7B - very small, very high quality loss",
+    },
+    {
+        "Q3_K_L",
+        LLAMA_FTYPE_MOSTLY_Q3_K_L,
+        " 3.35G, +0.1803 ppl @ 7B - small, substantial quality loss",
+    },
+    {
+        "Q4_K",
+        LLAMA_FTYPE_MOSTLY_Q4_K_M,
+        "alias for Q4_K_M",
+    },
+    {
+        "Q4_K_S",
+        LLAMA_FTYPE_MOSTLY_Q4_K_S,
+        " 3.56G, +0.1149 ppl @ 7B - small, significant quality loss",
+    },
+    {
+        "Q4_K_M",
+        LLAMA_FTYPE_MOSTLY_Q4_K_M,
+        " 3.80G, +0.0535 ppl @ 7B - medium, balanced quality - *recommended*",
+    },
+    {
+        "Q5_K",
+        LLAMA_FTYPE_MOSTLY_Q5_K_M,
+        "alias for Q5_K_M",
+    },
+    {
+        "Q5_K_S",
+        LLAMA_FTYPE_MOSTLY_Q5_K_S,
+        " 4.33G, +0.0353 ppl @ 7B - large, low quality loss - *recommended*",
+    },
+    {
+        "Q5_K_M",
+        LLAMA_FTYPE_MOSTLY_Q5_K_M,
+        " 4.45G, +0.0142 ppl @ 7B - large, very low quality loss - *recommended*",
+    },
+    {
+        "Q6_K",
+        LLAMA_FTYPE_MOSTLY_Q6_K,
+        " 5.15G, +0.0044 ppl @ 7B - very large, extremely low quality loss",
+    },
+#endif
+    {
+        "Q8_0",
+        LLAMA_FTYPE_MOSTLY_Q8_0,
+        " 6.70G, +0.0004 ppl @ 7B - very large, extremely low quality loss - not recommended",
+    },
+    {
+        "F16",
+        LLAMA_FTYPE_MOSTLY_F16,
+        "13.00G              @ 7B - extremely large, virtually no quality loss - not recommended",
+    },
+    {
+        "F32",
+        LLAMA_FTYPE_ALL_F32,
+        "26.00G              @ 7B - absolutely huge, lossless - not recommended",
+    },
+};
+
+
+bool try_parse_ftype(const std::string & ftype_str_in, llama_ftype & ftype, std::string & ftype_str_out) {
+    std::string ftype_str;
+
+    for (auto ch : ftype_str_in) {
+        ftype_str.push_back(std::toupper(ch));
+    }
+    for (auto & it : QUANT_OPTIONS) {
+        if (it.name == ftype_str) {
+            ftype = it.ftype;
+            ftype_str_out = it.name;
+            return true;
+        }
     }
-    // try to parse as an integer
     try {
         int ftype_int = std::stoi(ftype_str);
-        for (auto it = LLAMA_FTYPE_MAP.begin(); it != LLAMA_FTYPE_MAP.end(); it++) {
-            if (it->second == ftype_int) {
-                ftype = it->second;
-                ftype_str_out = it->first;
+        for (auto & it : QUANT_OPTIONS) {
+            if (it.ftype == ftype_int) {
+                ftype = it.ftype;
+                ftype_str_out = it.name;
                 return true;
             }
         }
@@ -39,29 +144,51 @@ bool try_parse_ftype(const std::string & ftype_str, llama_ftype & ftype, std::st
 }
 
 // usage:
-//  ./quantize models/llama/ggml-model.bin [models/llama/ggml-model-quant.bin] type [nthreads]
+//  ./quantize [--allow-requantize] [--leave-output-tensor] models/llama/ggml-model.bin [models/llama/ggml-model-quant.bin] type [nthreads]
 //
+void usage(const char * executable) {
+    fprintf(stderr, "usage: %s [--help] [--allow-requantize] [--leave-output-tensor] model-f32.bin [model-quant.bin] type [nthreads]\n\n", executable);
+    fprintf(stderr, "  --allow-requantize: Allows requantizing tensors that have already been quantized. Warning: This can severely reduce quality compared to quantizing from 16bit or 32bit\n");
+    fprintf(stderr, "  --leave-output-tensor: Will leave output.weight un(re)quantized. Increases model size but may also increase quality, especially when requantizing\n");
+    fprintf(stderr, "\nAllowed quantization types:\n");
+    for (auto & it : QUANT_OPTIONS) {
+        printf("  %2d  or  %-6s : %s\n", it.ftype, it.name.c_str(), it.desc.c_str());
+    }
+    exit(1);
+}
+
 int main(int argc, char ** argv) {
     if (argc < 3) {
-        fprintf(stderr, "usage: %s model-f32.bin [model-quant.bin] type [nthreads]\n", argv[0]);
-        for (auto it = LLAMA_FTYPE_MAP.begin(); it != LLAMA_FTYPE_MAP.end(); it++) {
-            fprintf(stderr, "  type = \"%s\" or %d\n", it->first.c_str(), it->second);
+        usage(argv[0]);
+    }
+
+    llama_model_quantize_params params = llama_model_quantize_default_params();
+
+    int arg_idx = 1;
+
+    for (; arg_idx < argc && strncmp(argv[arg_idx], "--", 2) == 0; arg_idx++) {
+        if (strcmp(argv[arg_idx], "--leave-output-tensor") == 0) {
+            params.quantize_output_tensor = false;
+        } else if (strcmp(argv[arg_idx], "--allow-requantize") == 0) {
+            params.allow_requantize = true;
+        } else {
+            usage(argv[0]);
         }
-        return 1;
+    }
+
+    if (argc - arg_idx < 3) {
+        usage(argv[0]);
     }
 
     llama_init_backend();
 
     // parse command line arguments
-    const std::string fname_inp = argv[1];
+    const std::string fname_inp = argv[arg_idx];
+    arg_idx++;
     std::string fname_out;
-    int nthread;
-    llama_ftype ftype;
 
-    int arg_idx = 2;
     std::string ftype_str;
-    if (try_parse_ftype(argv[arg_idx], ftype, ftype_str)) {
-        // argv[2] is the ftype
+    if (try_parse_ftype(argv[arg_idx], params.ftype, ftype_str)) {
         std::string fpath;
         const size_t pos = fname_inp.find_last_of('/');
         if (pos != std::string::npos) {
@@ -72,7 +199,6 @@ int main(int argc, char ** argv) {
         arg_idx++;
     }
     else {
-        // argv[2] is the output path
         fname_out = argv[arg_idx];
         arg_idx++;
 
@@ -80,8 +206,7 @@ int main(int argc, char ** argv) {
             fprintf(stderr, "%s: missing ftype\n", __func__);
             return 1;
         }
-        // argv[3] is the ftype
-        if (!try_parse_ftype(argv[arg_idx], ftype, ftype_str)) {
+        if (!try_parse_ftype(argv[arg_idx], params.ftype, ftype_str)) {
             fprintf(stderr, "%s: invalid ftype '%s'\n", __func__, argv[3]);
             return 1;
         }
@@ -91,21 +216,19 @@ int main(int argc, char ** argv) {
     // parse nthreads
     if (argc > arg_idx) {
         try {
-            nthread = std::stoi(argv[arg_idx]);
+            params.nthread = std::stoi(argv[arg_idx]);
         }
         catch (const std::exception & e) {
             fprintf(stderr, "%s: invalid nthread '%s' (%s)\n", __func__, argv[arg_idx], e.what());
             return 1;
         }
-    } else {
-        nthread = 0;
     }
 
     fprintf(stderr, "%s: build = %d (%s)\n", __func__, BUILD_NUMBER, BUILD_COMMIT);
 
     fprintf(stderr, "%s: quantizing '%s' to '%s' as %s", __func__, fname_inp.c_str(), fname_out.c_str(), ftype_str.c_str());
-    if (nthread > 0) {
-        fprintf(stderr, " using %d threads", nthread);
+    if (params.nthread > 0) {
+        fprintf(stderr, " using %d threads", params.nthread);
     }
     fprintf(stderr, "\n");
 
@@ -117,7 +240,7 @@ int main(int argc, char ** argv) {
     {
         const int64_t t_start_us = llama_time_us();
 
-        if (llama_model_quantize(fname_inp.c_str(), fname_out.c_str(), ftype, nthread)) {
+        if (llama_model_quantize(fname_inp.c_str(), fname_out.c_str(), &params)) {
             fprintf(stderr, "%s: failed to quantize model from '%s'\n", __func__, fname_inp.c_str());
             return 1;
         }

examples/server/README.md

@@ -16,6 +16,10 @@ This example allow you to have a llama.cpp http server to interact from a web pa
 To get started right away, run the following command, making sure to use the correct path for the model you have:
 
 #### Unix-based systems (Linux, macOS, etc.):
+Make sure to build with the server option on
+```bash
+LLAMA_BUILD_SERVER=1 make
+```
 
 ```bash
 ./server -m models/7B/ggml-model.bin --ctx_size 2048
@@ -287,6 +291,9 @@ Test();
 -   `-m FNAME, --model FNAME`: Specify the path to the LLaMA model file (e.g., `models/7B/ggml-model.bin`).
 -   `-c N, --ctx-size N`: Set the size of the prompt context. The default is 512, but LLaMA models were built with a context of 2048, which will provide better results for longer input/inference.
 -   `-ngl N, --n-gpu-layers N`: When compiled with appropriate support (currently CLBlast or cuBLAS), this option allows offloading some layers to the GPU for computation. Generally results in increased performance.
+-   `-mg i, --main-gpu i`: When using multiple GPUs this option controls which GPU is used for small tensors for which the overhead of splitting the computation across all GPUs is not worthwhile. The GPU in question will use slightly more VRAM to store a scratch buffer for temporary results. By default GPU 0 is used. Requires cuBLAS.
+-   `-ts SPLIT, --tensor-split SPLIT`: When using multiple GPUs this option controls how large tensors should be split across all GPUs. `SPLIT` is a comma-separated list of non-negative values that assigns the proportion of data that each GPU should get in order. For example, "3,2" will assign 60% of the data to GPU 0 and 40% to GPU 1. By default the data is split in proportion to VRAM but this may not be optimal for performance. Requires cuBLAS.
+-   `-lv, --low-vram`: Do not allocate a VRAM scratch buffer for holding temporary results. Reduces VRAM usage at the cost of performance, particularly prompt processing speed. Requires cuBLAS.
 -   `--embedding`: Enable the embedding mode. **Completion function doesn't work in this mode**.
 -   `--host`: Set the hostname or ip address to listen. Default `127.0.0.1`;
 -   `--port`: Set the port to listen. Default: `8080`.

examples/server/server.cpp

@@ -401,6 +401,11 @@ void server_print_usage(int /*argc*/, char **argv, const gpt_params &params)
 #ifdef LLAMA_SUPPORTS_GPU_OFFLOAD
   fprintf(stderr, "  -ngl N, --n-gpu-layers N\n");
   fprintf(stderr, "                        number of layers to store in VRAM\n");
+  fprintf(stderr, "  -ts SPLIT --tensor-split SPLIT\n");
+  fprintf(stderr, "                        how to split tensors across multiple GPUs, comma-separated list of proportions, e.g. 3,1\n");
+  fprintf(stderr, "                        how to split tensors across multiple GPUs, comma-separated list of proportions, e.g. 3,1\n");
+  fprintf(stderr, "  -mg i, --main-gpu i   the GPU to use for scratch and small tensors\n" );
+  fprintf(stderr, "  -lv, --low-vram       don't allocate VRAM scratch buffer\n" );
 #endif
   fprintf(stderr, "  -m FNAME, --model FNAME\n");
   fprintf(stderr, "                        model path (default: %s)\n", params.model.c_str());
@@ -502,6 +507,58 @@ bool server_params_parse(int argc, char **argv, server_params &sparams, gpt_para
 #else
       fprintf(stderr, "warning: not compiled with GPU offload support, --n-gpu-layers option will be ignored\n");
       fprintf(stderr, "warning: see main README.md for information on enabling GPU BLAS support\n");
+#endif
+    }
+    else if (arg == "--tensor-split" || arg == "-ts")
+    {
+      if (++i >= argc)
+      {
+        invalid_param = true;
+        break;
+      }
+#ifdef GGML_USE_CUBLAS
+      std::string arg_next = argv[i];
+
+      // split string by , and /
+      const std::regex regex{R"([,/]+)"};
+      std::sregex_token_iterator it{arg_next.begin(), arg_next.end(), regex, -1};
+      std::vector<std::string> split_arg{it, {}};
+      GGML_ASSERT(split_arg.size() <= LLAMA_MAX_DEVICES);
+
+      for (size_t i = 0; i < LLAMA_MAX_DEVICES; ++i)
+      {
+        if (i < split_arg.size())
+        {
+          params.tensor_split[i] = std::stof(split_arg[i]);
+        }
+        else
+        {
+          params.tensor_split[i] = 0.0f;
+        }
+      }
+#else
+      fprintf(stderr, "WARNING: llama.cpp was compiled without cuBLAS. It is not possible to set a tensor split.\n");
+#endif // GGML_USE_CUBLAS
+    }
+    else if (arg == "--low-vram" || arg == "-lv")
+    {
+#ifdef GGML_USE_CUBLAS
+      params.low_vram = true;
+#else
+      fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS. It is not possible to set lower vram usage.\n");
+#endif // GGML_USE_CUBLAS
+    }
+    else if (arg == "--main-gpu" || arg == "-mg")
+    {
+      if (++i >= argc)
+      {
+        invalid_param = true;
+        break;
+      }
+#ifdef GGML_USE_CUBLAS
+      params.main_gpu = std::stoi(argv[i]);
+#else
+      fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS. It is not possible to set a main GPU.\n");
 #endif
     }
     else

examples/train-text-from-scratch/CMakeLists.txt

@@ -0,0 +1,4 @@
+set(TARGET train-text-from-scratch)
+add_executable(${TARGET} train-text-from-scratch.cpp)
+target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_11)

examples/train-text-from-scratch/README.md

@@ -0,0 +1,22 @@
+# train-text-from-scratch
+
+Basic usage instructions:
+
+```bash
+# get training data
+wget https://github.com/brunoklein99/deep-learning-notes/blob/master/shakespeare.txt
+
+# train
+./bin/train-text-from-scratch \
+        --vocab-model ../models/ggml-vocab.bin \
+        --ctx 64 --embd 256 --head 8 --layer 16 \
+        --checkpoint-in  chk-shakespeare-256x16.bin \
+        --checkpoint-out chk-shakespeare-256x16.bin \
+        --model-out ggml-shakespeare-256x16-f32.bin \
+        --train-data "shakespeare.txt" \
+        -t 6 -b 16 -n 32 --seed 1 --adam-iter 16 \
+        --print-details-interval 0 --predict 16 --use-flash
+
+# predict
+./bin/main -m ggml-shakespeare-256x16-f32.bin
+```

flake.lock

@@ -1,12 +1,15 @@
 {
   "nodes": {
     "flake-utils": {
+      "inputs": {
+        "systems": "systems"
+      },
       "locked": {
-        "lastModified": 1676283394,
-        "narHash": "sha256-XX2f9c3iySLCw54rJ/CZs+ZK6IQy7GXNY4nSOyu2QG4=",
+        "lastModified": 1685518550,
+        "narHash": "sha256-o2d0KcvaXzTrPRIo0kOLV0/QXHhDQ5DTi+OxcjO8xqY=",
         "owner": "numtide",
         "repo": "flake-utils",
-        "rev": "3db36a8b464d0c4532ba1c7dda728f4576d6d073",
+        "rev": "a1720a10a6cfe8234c0e93907ffe81be440f4cef",
         "type": "github"
       },
       "original": {
@@ -17,11 +20,11 @@
     },
     "nixpkgs": {
       "locked": {
-        "lastModified": 1678470307,
-        "narHash": "sha256-OEeMUr3ueLIXyW/OaFUX5jUdimyQwMg/7e+/Q0gC/QE=",
+        "lastModified": 1685931219,
+        "narHash": "sha256-8EWeOZ6LKQfgAjB/USffUSELPRjw88A+xTcXnOUvO5M=",
         "owner": "NixOS",
         "repo": "nixpkgs",
-        "rev": "0c4800d579af4ed98ecc47d464a5e7b0870c4b1f",
+        "rev": "7409480d5c8584a1a83c422530419efe4afb0d19",
         "type": "github"
       },
       "original": {
@@ -36,6 +39,21 @@
         "flake-utils": "flake-utils",
         "nixpkgs": "nixpkgs"
       }
+    },
+    "systems": {
+      "locked": {
+        "lastModified": 1681028828,
+        "narHash": "sha256-Vy1rq5AaRuLzOxct8nz4T6wlgyUR7zLU309k9mBC768=",
+        "owner": "nix-systems",
+        "repo": "default",
+        "rev": "da67096a3b9bf56a91d16901293e51ba5b49a27e",
+        "type": "github"
+      },
+      "original": {
+        "owner": "nix-systems",
+        "repo": "default",
+        "type": "github"
+      }
     }
   },
   "root": "root",

flake.nix

@@ -6,6 +6,13 @@
   outputs = { self, nixpkgs, flake-utils }:
     flake-utils.lib.eachDefaultSystem (system:
       let
+        inherit (pkgs.stdenv) isAarch64 isDarwin;
+        inherit (pkgs.lib) optionals;
+        isM1 = isAarch64 && isDarwin;
+        osSpecific =
+          if isM1 then with pkgs.darwin.apple_sdk_11_0.frameworks; [ Accelerate MetalKit MetalPerformanceShaders MetalPerformanceShadersGraph ]
+          else if isDarwin then with pkgs.darwin.apple_sdk.frameworks; [ Accelerate CoreGraphics CoreVideo ]
+          else [ ];
         pkgs = import nixpkgs {
           inherit system;
         };
@@ -18,17 +25,22 @@
         packages.default = pkgs.stdenv.mkDerivation {
           name = "llama.cpp";
           src = ./.;
+          postPatch =
+            if isM1 then ''
+              substituteInPlace ./ggml-metal.m \
+                --replace '[bundle pathForResource:@"ggml-metal" ofType:@"metal"];' "@\"$out/ggml-metal.metal\";"
+            '' else "";
           nativeBuildInputs = with pkgs; [ cmake ];
-          buildInputs = with pkgs; lib.optionals stdenv.isDarwin [
-            darwin.apple_sdk.frameworks.Accelerate
-          ];
-          cmakeFlags = with pkgs; lib.optionals (system == "aarch64-darwin") [
+          buildInputs = osSpecific;
+          cmakeFlags = [ "-DLLAMA_BUILD_SERVER=ON" ] ++ (optionals isM1 [
             "-DCMAKE_C_FLAGS=-D__ARM_FEATURE_DOTPROD=1"
-          ];
+            "-DLLAMA_METAL=ON"
+          ]);
           installPhase = ''
             mkdir -p $out/bin
             mv bin/* $out/bin/
             mv $out/bin/main $out/bin/llama
+            mv $out/bin/server $out/bin/llama-server
 
             echo "#!${llama-python}/bin/python" > $out/bin/convert.py
             cat ${./convert.py} >> $out/bin/convert.py
@@ -40,9 +52,7 @@
           packages = with pkgs; [
             cmake
             llama-python
-          ] ++ lib.optionals stdenv.isDarwin [
-            darwin.apple_sdk.frameworks.Accelerate
-          ];
+          ] ++ osSpecific;
         };
       }
     );

ggml-cuda.h

@@ -1,10 +1,19 @@
+#pragma once
+
 #include "ggml.h"
 
 #ifdef  __cplusplus
 extern "C" {
 #endif
 
+#define GGML_CUDA_MAX_DEVICES       16
+
+struct ggml_tensor_extra_gpu {
+    void * data_device[GGML_CUDA_MAX_DEVICES]; // 1 pointer for each device for split tensors
+};
+
 void   ggml_init_cublas(void);
+void   ggml_cuda_set_tensor_split(const float * tensor_split);
 
 void   ggml_cuda_mul(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
 bool   ggml_cuda_can_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
@@ -15,8 +24,15 @@ void   ggml_cuda_mul_mat(const struct ggml_tensor * src0, const struct ggml_tens
 void * ggml_cuda_host_malloc(size_t size);
 void   ggml_cuda_host_free(void * ptr);
 
-void ggml_cuda_transform_tensor(struct ggml_tensor * tensor);
-void ggml_cuda_load_data(const char * fname, struct ggml_tensor * tensors, size_t offset);
+void   ggml_cuda_transform_tensor(void * data, struct ggml_tensor * tensor);
+
+void   ggml_cuda_free_data(struct ggml_tensor * tensor);
+void   ggml_cuda_assign_buffers(struct ggml_tensor * tensor);
+void   ggml_cuda_assign_buffers_no_scratch(struct ggml_tensor * tensor);
+void   ggml_cuda_set_main_device(int main_device);
+void   ggml_cuda_set_scratch_size(size_t scratch_size);
+void   ggml_cuda_free_scratch(void);
+bool   ggml_cuda_compute_forward(struct ggml_compute_params * params, struct ggml_tensor * tensor);
 
 #ifdef  __cplusplus
 }

ggml-metal.h

@@ -0,0 +1,64 @@
+// An interface allowing to compute ggml_cgraph with Metal
+//
+// This is a fully functional interface that extends ggml with GPU support for Apple devices.
+// A similar interface can be created for other GPU backends (e.g. Vulkan, CUDA, OpenCL, etc.)
+//
+// How it works?
+//
+// As long as your program can create and evaluate a ggml_cgraph on the CPU, you can use this
+// interface to evaluate the same graph on the GPU. Instead of using ggml_graph_compute(), you
+// use ggml_metal_graph_compute() (or ggml_vulkan_graph_compute(), etc.)
+//
+// You only need to make sure that all memory buffers that you used during the graph creation
+// are mapped to the device memory with the ggml_metal_add_buffer() function. This mapping is
+// used during the graph evaluation to determine the arguments of the compute kernels.
+//
+// Synchronization between device and host memory (for example for input and output tensors)
+// is done with the ggml_metal_set_tensor() and ggml_metal_get_tensor() functions.
+//
+
+#pragma once
+
+#include <stddef.h>
+#include <stdbool.h>
+
+// max memory buffers that can be mapped to the device
+#define GGML_METAL_MAX_BUFFERS 16
+
+struct ggml_tensor;
+struct ggml_cgraph;
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct ggml_metal_context;
+
+struct ggml_metal_context * ggml_metal_init(void);
+void ggml_metal_free(struct ggml_metal_context * ctx);
+
+// creates a mapping between a host memory buffer and a device memory buffer
+// - make sure to map all buffers used in the graph before calling ggml_metal_graph_compute
+// - the mapping is used during computation to determine the arguments of the compute kernels
+// - you don't need to keep the host memory buffer allocated as it is never accessed by Metal
+//
+bool ggml_metal_add_buffer(
+        struct ggml_metal_context * ctx,
+                       const char * name,
+                             void * data,
+                           size_t   size);
+
+// set data from host memory into the device
+void ggml_metal_set_tensor(struct ggml_metal_context * ctx, struct ggml_tensor * t);
+
+// get data from the device into host memory
+void ggml_metal_get_tensor(struct ggml_metal_context * ctx, struct ggml_tensor * t);
+
+// same as ggml_graph_compute but uses Metal
+// creates gf->n_threads command buffers in parallel
+void ggml_metal_graph_compute(struct ggml_metal_context * ctx, struct ggml_cgraph * gf);
+
+#ifdef __cplusplus
+}
+#endif
+

ggml-metal.m

@@ -0,0 +1,834 @@
+#import "ggml-metal.h"
+
+#import "ggml.h"
+
+#import <Foundation/Foundation.h>
+
+#import <Metal/Metal.h>
+#import <MetalPerformanceShaders/MetalPerformanceShaders.h>
+
+#ifdef GGML_METAL_NDEBUG
+#define metal_printf(...)
+#else
+#define metal_printf(...) fprintf(stderr, __VA_ARGS__)
+#endif
+
+#define UNUSED(x) (void)(x)
+
+struct ggml_metal_buffer {
+    const char * name;
+
+    void   * data;
+    size_t   size;
+
+    id<MTLBuffer> metal;
+};
+
+struct ggml_metal_context {
+    float * logits;
+
+    id<MTLDevice>       device;
+    id<MTLCommandQueue> queue;
+    id<MTLLibrary>      library;
+
+    int n_buffers;
+    struct ggml_metal_buffer buffers[GGML_METAL_MAX_BUFFERS];
+
+    // custom kernels
+#define GGML_METAL_DECL_KERNEL(name) \
+    id<MTLFunction>             function_##name; \
+    id<MTLComputePipelineState> pipeline_##name
+
+    GGML_METAL_DECL_KERNEL(add);
+    GGML_METAL_DECL_KERNEL(mul);
+    GGML_METAL_DECL_KERNEL(mul_row); // TODO: avoid this extra kernel, instead extend the "mul" kernel to support broadcast
+    GGML_METAL_DECL_KERNEL(scale);
+    GGML_METAL_DECL_KERNEL(silu);
+    GGML_METAL_DECL_KERNEL(relu);
+    GGML_METAL_DECL_KERNEL(gelu);
+    GGML_METAL_DECL_KERNEL(soft_max);
+    GGML_METAL_DECL_KERNEL(diag_mask_inf);
+    GGML_METAL_DECL_KERNEL(get_rows_f16);
+    GGML_METAL_DECL_KERNEL(get_rows_q4_0);
+    GGML_METAL_DECL_KERNEL(get_rows_q4_1);
+    GGML_METAL_DECL_KERNEL(get_rows_q2_k);
+    GGML_METAL_DECL_KERNEL(get_rows_q3_k);
+    GGML_METAL_DECL_KERNEL(get_rows_q4_k);
+    GGML_METAL_DECL_KERNEL(get_rows_q5_k);
+    GGML_METAL_DECL_KERNEL(get_rows_q6_k);
+    GGML_METAL_DECL_KERNEL(rms_norm);
+    GGML_METAL_DECL_KERNEL(mul_mat_f16_f32);
+    GGML_METAL_DECL_KERNEL(mul_mat_q4_0_f32);
+    GGML_METAL_DECL_KERNEL(mul_mat_q4_1_f32);
+    GGML_METAL_DECL_KERNEL(mul_mat_q2_k_f32);
+    GGML_METAL_DECL_KERNEL(mul_mat_q3_k_f32);
+    GGML_METAL_DECL_KERNEL(mul_mat_q4_k_f32);
+    GGML_METAL_DECL_KERNEL(mul_mat_q5_k_f32);
+    GGML_METAL_DECL_KERNEL(mul_mat_q6_k_f32);
+    GGML_METAL_DECL_KERNEL(rope);
+    GGML_METAL_DECL_KERNEL(cpy_f32_f16);
+    GGML_METAL_DECL_KERNEL(cpy_f32_f32);
+
+#undef GGML_METAL_DECL_KERNEL
+};
+
+// MSL code
+// TODO: move the contents here when ready
+//       for now it is easier to work in a separate file
+static NSString * const msl_library_source = @"see metal.metal";
+
+// Here to assist with NSBundle Path Hack
+@interface GGMLMetalClass : NSObject
+@end
+@implementation GGMLMetalClass
+@end
+
+struct ggml_metal_context * ggml_metal_init(void) {
+    fprintf(stderr, "%s: allocating\n", __func__);
+
+    struct ggml_metal_context * ctx = malloc(sizeof(struct ggml_metal_context));
+
+    ctx->device = MTLCreateSystemDefaultDevice();
+    ctx->queue  = [ctx->device newCommandQueue];
+    ctx->n_buffers = 0;
+
+    // determine if we can use MPS
+    if (MPSSupportsMTLDevice(ctx->device)) {
+        fprintf(stderr, "%s: using MPS\n", __func__);
+    } else {
+        fprintf(stderr, "%s: not using MPS\n", __func__);
+        GGML_ASSERT(false && "MPS not supported");
+    }
+
+#if 0
+    // compile from source string and show compile log
+    {
+        NSError * error = nil;
+
+        ctx->library = [ctx->device newLibraryWithSource:msl_library_source options:nil error:&error];
+        if (error) {
+            fprintf(stderr, "%s: error: %s\n", __func__, [[error description] UTF8String]);
+            exit(1);
+        }
+    }
+#else
+    UNUSED(msl_library_source);
+
+    // read the source from "ggml-metal.metal" into a string and use newLibraryWithSource
+    {
+        NSError * error = nil;
+
+        //NSString * path = [[NSBundle mainBundle] pathForResource:@"../../examples/metal/metal" ofType:@"metal"];
+        NSBundle * bundle = [NSBundle bundleForClass:[GGMLMetalClass class]];
+        NSString * path = [bundle pathForResource:@"ggml-metal" ofType:@"metal"];
+        fprintf(stderr, "%s: loading '%s'\n", __func__, [path UTF8String]);
+
+        NSString * src  = [NSString stringWithContentsOfFile:path encoding:NSUTF8StringEncoding error:&error];
+        if (error) {
+            fprintf(stderr, "%s: error: %s\n", __func__, [[error description] UTF8String]);
+            exit(1);
+        }
+
+        ctx->library = [ctx->device newLibraryWithSource:src options:nil error:&error];
+        if (error) {
+            fprintf(stderr, "%s: error: %s\n", __func__, [[error description] UTF8String]);
+            exit(1);
+        }
+    }
+#endif
+
+    // load kernels
+    {
+#define GGML_METAL_ADD_KERNEL(name) \
+        ctx->function_##name = [ctx->library newFunctionWithName:@"kernel_"#name]; \
+        ctx->pipeline_##name = [ctx->device newComputePipelineStateWithFunction:ctx->function_##name error:nil]; \
+        fprintf(stderr, "%s: loaded %-32s %16p\n", __func__, "kernel_"#name, (void *) ctx->pipeline_##name);
+
+        GGML_METAL_ADD_KERNEL(add);
+        GGML_METAL_ADD_KERNEL(mul);
+        GGML_METAL_ADD_KERNEL(mul_row);
+        GGML_METAL_ADD_KERNEL(scale);
+        GGML_METAL_ADD_KERNEL(silu);
+        GGML_METAL_ADD_KERNEL(relu);
+        GGML_METAL_ADD_KERNEL(gelu);
+        GGML_METAL_ADD_KERNEL(soft_max);
+        GGML_METAL_ADD_KERNEL(diag_mask_inf);
+        GGML_METAL_ADD_KERNEL(get_rows_f16);
+        GGML_METAL_ADD_KERNEL(get_rows_q4_0);
+        GGML_METAL_ADD_KERNEL(get_rows_q4_1);
+        GGML_METAL_ADD_KERNEL(get_rows_q2_k);
+        GGML_METAL_ADD_KERNEL(get_rows_q3_k);
+        GGML_METAL_ADD_KERNEL(get_rows_q4_k);
+        GGML_METAL_ADD_KERNEL(get_rows_q5_k);
+        GGML_METAL_ADD_KERNEL(get_rows_q6_k);
+        GGML_METAL_ADD_KERNEL(rms_norm);
+        GGML_METAL_ADD_KERNEL(mul_mat_f16_f32);
+        GGML_METAL_ADD_KERNEL(mul_mat_q4_0_f32);
+        GGML_METAL_ADD_KERNEL(mul_mat_q4_1_f32);
+        GGML_METAL_ADD_KERNEL(mul_mat_q2_k_f32);
+        GGML_METAL_ADD_KERNEL(mul_mat_q3_k_f32);
+        GGML_METAL_ADD_KERNEL(mul_mat_q4_k_f32);
+        GGML_METAL_ADD_KERNEL(mul_mat_q5_k_f32);
+        GGML_METAL_ADD_KERNEL(mul_mat_q6_k_f32);
+        GGML_METAL_ADD_KERNEL(rope);
+        GGML_METAL_ADD_KERNEL(cpy_f32_f16);
+        GGML_METAL_ADD_KERNEL(cpy_f32_f32);
+
+#undef GGML_METAL_ADD_KERNEL
+    }
+
+    return ctx;
+}
+
+void ggml_metal_free(struct ggml_metal_context * ctx) {
+    fprintf(stderr, "%s: deallocating\n", __func__);
+
+    free(ctx);
+}
+
+// finds the Metal buffer that contains the tensor data on the GPU device
+// the assumption is that there is 1-to-1 mapping between the host and device memory buffers, so we can find the
+// Metal buffer based on the host memory pointer
+//
+static id<MTLBuffer> ggml_metal_get_buffer(struct ggml_metal_context * ctx, struct ggml_tensor * t, size_t * offs) {
+    //fprintf(stderr, "%s: data tensor '%16s', offs_data = %8ld, offs_eval = %8ld, offs_cach = %8ld\n", __func__, t->name, offs_data, offs_eval, offs_cach);
+
+    for (int i = 0; i < ctx->n_buffers; ++i) {
+        const int64_t ioffs = (int64_t) t->data - (int64_t) ctx->buffers[i].data;
+
+        if (ioffs >= 0 && ioffs < (int64_t) ctx->buffers[i].size) {
+            *offs = (size_t) ioffs;
+
+            //fprintf(stderr, "%s: '%s' tensor '%16s', offs = %8ld\n", __func__, ctx->buffers[i].name, t->name, *offs);
+
+            return ctx->buffers[i].metal;
+        }
+    }
+
+    fprintf(stderr, "%s: error: buffer is nil\n", __func__);
+
+    return nil;
+}
+
+bool ggml_metal_add_buffer(
+        struct ggml_metal_context * ctx,
+                     const char * name,
+                           void * data,
+                         size_t   size) {
+    if (ctx->n_buffers >= GGML_METAL_MAX_BUFFERS) {
+        fprintf(stderr, "%s: too many buffers\n", __func__);
+        return false;
+    }
+
+    if (data) {
+        // verify that the buffer does not overlap with any of the existing buffers
+        for (int i = 0; i < ctx->n_buffers; ++i) {
+            const int64_t ioffs = (int64_t) data - (int64_t) ctx->buffers[i].data;
+
+            if (ioffs >= 0 && ioffs < (int64_t) ctx->buffers[i].size) {
+                fprintf(stderr, "%s: error: buffer '%s' overlaps with '%s'\n", __func__, name, ctx->buffers[i].name);
+                return false;
+            }
+        }
+
+        size_t page_size = getpagesize();
+        size_t aligned_size = size;
+        if ((aligned_size % page_size) != 0) {
+            aligned_size += (page_size - (aligned_size % page_size));
+        }
+
+        ctx->buffers[ctx->n_buffers].name = name;
+        ctx->buffers[ctx->n_buffers].data = data;
+        ctx->buffers[ctx->n_buffers].size = size;
+
+        if (ctx->device.maxBufferLength < aligned_size) {
+            fprintf(stderr, "%s: buffer '%s' size %zu is larger than buffer maximum of %zu\n", __func__, name, aligned_size, ctx->device.maxBufferLength);
+            return false;
+        }
+        ctx->buffers[ctx->n_buffers].metal = [ctx->device newBufferWithBytesNoCopy:data length:aligned_size options:MTLResourceStorageModeShared deallocator:nil];
+
+        if (ctx->buffers[ctx->n_buffers].metal == nil) {
+            fprintf(stderr, "%s: failed to allocate '%-16s' buffer, size = %8.2f MB\n", __func__, name, aligned_size / 1024.0 / 1024.0);
+            return false;
+        } else {
+            fprintf(stderr, "%s: allocated '%-16s' buffer, size = %8.2f MB\n", __func__, name, aligned_size / 1024.0 / 1024.0);
+        }
+
+        ++ctx->n_buffers;
+    }
+
+    return true;
+}
+
+void ggml_metal_set_tensor(
+        struct ggml_metal_context * ctx,
+        struct ggml_tensor * t) {
+    metal_printf("%s: set input for tensor '%s'\n", __func__, t->name);
+
+    size_t offs;
+    id<MTLBuffer> id_dst = ggml_metal_get_buffer(ctx, t, &offs);
+
+    memcpy((void *) ((uint8_t *) id_dst.contents + offs), t->data, ggml_nbytes(t));
+}
+
+void ggml_metal_get_tensor(
+        struct ggml_metal_context * ctx,
+        struct ggml_tensor * t) {
+    metal_printf("%s: extract results for tensor '%s'\n", __func__, t->name);
+
+    size_t offs;
+    id<MTLBuffer> id_src = ggml_metal_get_buffer(ctx, t, &offs);
+
+    memcpy(t->data, (void *) ((uint8_t *) id_src.contents + offs), ggml_nbytes(t));
+}
+
+void ggml_metal_graph_compute(
+        struct ggml_metal_context * ctx,
+               struct ggml_cgraph * gf) {
+    metal_printf("%s: evaluating graph\n", __func__);
+
+    // create multiple command buffers and enqueue them
+    // then, we encode the graph into the command buffers in parallel
+
+    const int n_cb = gf->n_threads;
+
+    NSMutableArray * command_buffers = [NSMutableArray arrayWithCapacity:n_cb];
+
+    for (int i = 0; i < n_cb; ++i) {
+        command_buffers[i] = [ctx->queue commandBuffer];
+
+        // enqueue the command buffers in order to specify their execution order
+        [command_buffers[i] enqueue];
+    }
+
+    // TODO: is this the best way to start threads?
+    dispatch_queue_t queue = dispatch_queue_create("llama.cpp", DISPATCH_QUEUE_CONCURRENT);
+
+    for (int cb_idx = 0; cb_idx < n_cb; ++cb_idx) {
+        const int n_nodes_per_cb = (gf->n_nodes + n_cb - 1) / n_cb;
+
+        dispatch_async(queue, ^{
+            size_t offs_src0 = 0;
+            size_t offs_src1 = 0;
+            size_t offs_dst  = 0;
+
+            id<MTLCommandBuffer> command_buffer = command_buffers[cb_idx];
+
+            id<MTLComputeCommandEncoder> encoder = nil;
+
+            const int node_start =                                      (cb_idx + 0) * n_nodes_per_cb;
+            const int node_end   = (cb_idx == n_cb - 1) ? gf->n_nodes : (cb_idx + 1) * n_nodes_per_cb;
+
+            for (int i = node_start; i < node_end; ++i) {
+                metal_printf("%s: encoding node %3d, op = %8s\n", __func__, i, ggml_op_name(gf->nodes[i]->op));
+
+                struct ggml_tensor * src0 = gf->nodes[i]->src0;
+                struct ggml_tensor * src1 = gf->nodes[i]->src1;
+                struct ggml_tensor * dst  = gf->nodes[i];
+
+                const int64_t  ne00 = src0 ? src0->ne[0] : 0;
+                const int64_t  ne01 = src0 ? src0->ne[1] : 0;
+                const int64_t  ne02 = src0 ? src0->ne[2] : 0;
+                const int64_t  ne03 = src0 ? src0->ne[3] : 0;
+
+                const uint64_t nb00 = src0 ? src0->nb[0] : 0;
+                const uint64_t nb01 = src0 ? src0->nb[1] : 0;
+                const uint64_t nb02 = src0 ? src0->nb[2] : 0;
+                const uint64_t nb03 = src0 ? src0->nb[3] : 0;
+
+                const int64_t  ne10 = src1 ? src1->ne[0] : 0;
+                const int64_t  ne11 = src1 ? src1->ne[1] : 0;
+                const int64_t  ne12 = src1 ? src1->ne[2] : 0;
+                const int64_t  ne13 = src1 ? src1->ne[3] : 0; UNUSED(ne13);
+
+                const uint64_t nb10 = src1 ? src1->nb[0] : 0;
+                const uint64_t nb11 = src1 ? src1->nb[1] : 0;
+                const uint64_t nb12 = src1 ? src1->nb[2] : 0;
+                const uint64_t nb13 = src1 ? src1->nb[3] : 0; UNUSED(nb13);
+
+                const int64_t  ne0  = dst ? dst->ne[0] : 0;
+                const int64_t  ne1  = dst ? dst->ne[1] : 0;
+                const int64_t  ne2  = dst ? dst->ne[2] : 0;
+                const int64_t  ne3  = dst ? dst->ne[3] : 0;
+
+                const uint64_t nb0  = dst ? dst->nb[0] : 0;
+                const uint64_t nb1  = dst ? dst->nb[1] : 0;
+                const uint64_t nb2  = dst ? dst->nb[2] : 0;
+                const uint64_t nb3  = dst ? dst->nb[3] : 0;
+
+                const enum ggml_type src0t = src0 ? src0->type : GGML_TYPE_COUNT;
+                const enum ggml_type src1t = src1 ? src1->type : GGML_TYPE_COUNT;
+                const enum ggml_type dstt  = dst  ? dst->type  : GGML_TYPE_COUNT;
+
+                id<MTLBuffer> id_src0 = src0 ? ggml_metal_get_buffer(ctx, src0, &offs_src0) : nil;
+                id<MTLBuffer> id_src1 = src1 ? ggml_metal_get_buffer(ctx, src1, &offs_src1) : nil;
+                id<MTLBuffer> id_dst  = dst  ? ggml_metal_get_buffer(ctx, dst,  &offs_dst)  : nil;
+
+                //metal_printf("%s: op - %s\n", __func__, ggml_op_name(dst->op));
+                //if (src0) {
+                //    metal_printf("%s: src0 - %4s [%5lld, %5lld, %5lld], %d, %s\n", __func__, ggml_type_name(src0t), ne00, ne01, ne02,
+                //            ggml_is_contiguous(src0), src0->name);
+                //}
+                //if (src1) {
+                //    metal_printf("%s: src1 - %4s [%5lld, %5lld, %5lld], %d, %s\n", __func__, ggml_type_name(src1t), ne10, ne11, ne12,
+                //            ggml_is_contiguous(src1), src1->name);
+                //}
+                //if (dst) {
+                //    metal_printf("%s: dst  - %4s [%5lld, %5lld, %5lld], 1, %s\n",  __func__, ggml_type_name(dstt),  ne0,  ne1,  ne2,
+                //            dst->name);
+                //}
+
+                switch (dst->op) {
+                    case GGML_OP_RESHAPE:
+                    case GGML_OP_VIEW:
+                    case GGML_OP_TRANSPOSE:
+                    case GGML_OP_PERMUTE:
+                        {
+                            // noop
+                        } break;
+                    case GGML_OP_ADD:
+                        {
+                            if (encoder == nil) {
+                                encoder = [command_buffer computeCommandEncoder];
+                            }
+
+                            [encoder setComputePipelineState:ctx->pipeline_add];
+                            [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                            [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
+                            [encoder setBuffer:id_dst  offset:offs_dst  atIndex:2];
+
+                            const int64_t n = ggml_nelements(dst);
+
+                            [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                        } break;
+                    case GGML_OP_MUL:
+                        {
+                            if (encoder == nil) {
+                                encoder = [command_buffer computeCommandEncoder];
+                            }
+
+                            if (ggml_nelements(src1) == ne10) {
+                                // src1 is a row
+                                [encoder setComputePipelineState:ctx->pipeline_mul_row];
+                            } else {
+                                [encoder setComputePipelineState:ctx->pipeline_mul];
+                            }
+                            [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                            [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
+                            [encoder setBuffer:id_dst  offset:offs_dst  atIndex:2];
+                            [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:3];
+
+                            const int64_t n = ggml_nelements(dst);
+
+                            [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                        } break;
+                    case GGML_OP_SCALE:
+                        {
+                            if (encoder == nil) {
+                                encoder = [command_buffer computeCommandEncoder];
+                            }
+
+                            const float scale = *(const float *) src1->data;
+
+                            [encoder setComputePipelineState:ctx->pipeline_scale];
+                            [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                            [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
+                            [encoder setBytes:&scale length:sizeof(scale) atIndex:2];
+
+                            const int64_t n = ggml_nelements(dst);
+
+                            [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                        } break;
+                    case GGML_OP_SILU:
+                        {
+                            if (encoder == nil) {
+                                encoder = [command_buffer computeCommandEncoder];
+                            }
+
+                            [encoder setComputePipelineState:ctx->pipeline_silu];
+                            [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;
+                    case GGML_OP_RELU:
+                        {
+                            if (encoder == nil) {
+                                encoder = [command_buffer computeCommandEncoder];
+                            }
+
+                            [encoder setComputePipelineState:ctx->pipeline_relu];
+                            [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;
+                    case GGML_OP_GELU:
+                    {
+                            if (encoder == nil) {
+                                encoder = [command_buffer computeCommandEncoder];
+                            }
+
+                            [encoder setComputePipelineState:ctx->pipeline_gelu];
+                            [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;
+                    case GGML_OP_SOFT_MAX:
+                        {
+                            if (encoder == nil) {
+                                encoder = [command_buffer computeCommandEncoder];
+                            }
+
+                            const int nth = 32;
+
+                            [encoder setComputePipelineState:ctx->pipeline_soft_max];
+                            [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                            [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
+                            [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 setThreadgroupMemoryLength:nth*sizeof(float) atIndex:0];
+
+                            [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
+                        } break;
+                    case GGML_OP_DIAG_MASK_INF:
+                        {
+                            if (encoder == nil) {
+                                encoder = [command_buffer computeCommandEncoder];
+                            }
+
+                            const int n_past = ((int32_t *)(src1->data))[0];
+
+                            [encoder setComputePipelineState:ctx->pipeline_diag_mask_inf];
+                            [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                            [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
+                            [encoder setBytes:&ne00   length:sizeof(ne00) atIndex:2];
+                            [encoder setBytes:&ne01   length:sizeof(ne01) atIndex:3];
+                            [encoder setBytes:&n_past length:sizeof(int)  atIndex:4];
+
+                            [encoder dispatchThreadgroups:MTLSizeMake(ne00, ne01, ne02) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                        } break;
+                    case GGML_OP_MUL_MAT:
+                        {
+                            // TODO: needs to be updated after PR: https://github.com/ggerganov/ggml/pull/224
+
+                            GGML_ASSERT(ne00 == ne10);
+                            GGML_ASSERT(ne02 == ne12);
+
+                            if (ggml_is_contiguous(src0) &&
+                                ggml_is_contiguous(src1) &&
+                                (src0t == GGML_TYPE_F32 || src0t == GGML_TYPE_F16) && ne11 > 1) {
+
+                                if (encoder != nil) {
+                                    [encoder endEncoding];
+                                    encoder = nil;
+                                }
+
+                                MPSDataType src0dt = src0t == GGML_TYPE_F32 ? MPSDataTypeFloat32 : MPSDataTypeFloat16;
+                                MPSDataType src1dt = src1t == GGML_TYPE_F32 ? MPSDataTypeFloat32 : MPSDataTypeFloat16;
+
+                                // for F32 x F32 we use MPS
+                                MPSMatrixDescriptor * desc0 = [MPSMatrixDescriptor
+                                    matrixDescriptorWithRows:ne01 columns:ne00 rowBytes:src0->nb[1] dataType:src0dt];
+
+                                MPSMatrixDescriptor * desc1 = [MPSMatrixDescriptor
+                                    matrixDescriptorWithRows:ne11 columns:ne10 rowBytes:src1->nb[1] dataType:src1dt];
+
+                                MPSMatrixDescriptor * desc  = [MPSMatrixDescriptor
+                                    matrixDescriptorWithRows:ne1 columns:ne0 rowBytes:dst->nb[1] dataType:MPSDataTypeFloat32];
+
+                                MPSMatrixMultiplication * mul = [[MPSMatrixMultiplication alloc]
+                                    initWithDevice:ctx->device transposeLeft:false transposeRight:true
+                                        resultRows:ne11 resultColumns:ne01 interiorColumns:ne00 alpha:1.0 beta:0.0];
+
+                                // we need to do ne02 multiplications
+                                // TODO: is there a way to do this in parallel - currently very slow ..
+                                // TODO: might be possible to offload part of the computation to ANE using Accelerate's CBLAS
+                                for (int64_t i02 = 0; i02 < ne02; ++i02) {
+                                    size_t offs_src0_cur = offs_src0 + i02*nb02;
+                                    size_t offs_src1_cur = offs_src1 + i02*nb12;
+                                    size_t offs_dst_cur  = offs_dst  + i02*nb2;
+
+                                    MPSMatrix * mat_src0 = [[MPSMatrix alloc] initWithBuffer:id_src0 offset:offs_src0_cur descriptor:desc0];
+                                    MPSMatrix * mat_src1 = [[MPSMatrix alloc] initWithBuffer:id_src1 offset:offs_src1_cur descriptor:desc1];
+                                    MPSMatrix * mat_dst  = [[MPSMatrix alloc] initWithBuffer:id_dst  offset:offs_dst_cur  descriptor:desc ];
+
+                                    [mul encodeToCommandBuffer:command_buffer leftMatrix:mat_src1 rightMatrix:mat_src0 resultMatrix:mat_dst];
+                                }
+                            } else {
+                                if (encoder == nil) {
+                                    encoder = [command_buffer computeCommandEncoder];
+                                }
+
+                                int nth0 = 32;
+                                int nth1 = 1;
+
+                                // use custom matrix x vector kernel
+                                switch (src0t) {
+                                    case GGML_TYPE_F16:
+                                        {
+                                            GGML_ASSERT(ne02 == ne12);
+
+                                            nth0 = 64;
+                                            nth1 = 1;
+                                            [encoder setComputePipelineState:ctx->pipeline_mul_mat_f16_f32];
+                                        } break;
+                                    case GGML_TYPE_Q4_0:
+                                        {
+                                            GGML_ASSERT(ne02 == 1);
+                                            GGML_ASSERT(ne12 == 1);
+
+                                            nth0 = 8;
+                                            nth1 = 8;
+                                            [encoder setComputePipelineState:ctx->pipeline_mul_mat_q4_0_f32];
+                                        } break;
+                                    case GGML_TYPE_Q4_1:
+                                        {
+                                            GGML_ASSERT(ne02 == 1);
+                                            GGML_ASSERT(ne12 == 1);
+
+                                            nth0 = 8;
+                                            nth1 = 8;
+                                            [encoder setComputePipelineState:ctx->pipeline_mul_mat_q4_1_f32];
+                                        } break;
+                                    case GGML_TYPE_Q2_K:
+                                        {
+                                            GGML_ASSERT(ne02 == 1);
+                                            GGML_ASSERT(ne12 == 1);
+
+                                            nth0 = 4;
+                                            nth1 = 16;
+                                            [encoder setComputePipelineState:ctx->pipeline_mul_mat_q2_k_f32];
+                                        } break;
+                                    case GGML_TYPE_Q3_K:
+                                        {
+                                            GGML_ASSERT(ne02 == 1);
+                                            GGML_ASSERT(ne12 == 1);
+
+                                            nth0 = 4;
+                                            nth1 = 16;
+                                            [encoder setComputePipelineState:ctx->pipeline_mul_mat_q3_k_f32];
+                                        } break;
+                                    case GGML_TYPE_Q4_K:
+                                        {
+                                            GGML_ASSERT(ne02 == 1);
+                                            GGML_ASSERT(ne12 == 1);
+
+                                            nth0 = 4;
+                                            nth1 = 16;
+                                            [encoder setComputePipelineState:ctx->pipeline_mul_mat_q4_k_f32];
+                                        } break;
+                                    case GGML_TYPE_Q5_K:
+                                        {
+                                            GGML_ASSERT(ne02 == 1);
+                                            GGML_ASSERT(ne12 == 1);
+
+                                            nth0 = 4;
+                                            nth1 = 16;
+                                            [encoder setComputePipelineState:ctx->pipeline_mul_mat_q5_k_f32];
+                                        } break;
+                                    case GGML_TYPE_Q6_K:
+                                        {
+                                            GGML_ASSERT(ne02 == 1);
+                                            GGML_ASSERT(ne12 == 1);
+
+                                            nth0 = 4;
+                                            nth1 = 16;
+                                            [encoder setComputePipelineState:ctx->pipeline_mul_mat_q6_k_f32];
+                                        } break;
+                                    default:
+                                        {
+                                            fprintf(stderr, "Asserting on type %d\n",(int)src0t);
+                                            GGML_ASSERT(false && "not implemented");
+                                        }
+                                };
+
+                                [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                                [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
+                                [encoder setBuffer:id_dst  offset:offs_dst  atIndex:2];
+                                [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:3];
+                                [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:4];
+                                [encoder setBytes:&nb00 length:sizeof(nb00) atIndex:5];
+                                [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:6];
+                                [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:7];
+                                [encoder setBytes:&ne10 length:sizeof(ne10) atIndex:8];
+                                [encoder setBytes:&ne11 length:sizeof(ne11) atIndex:9];
+                                [encoder setBytes:&nb10 length:sizeof(nb10) atIndex:10];
+                                [encoder setBytes:&nb11 length:sizeof(nb11) atIndex:11];
+                                [encoder setBytes:&nb12 length:sizeof(nb12) atIndex:12];
+                                [encoder setBytes:&ne0  length:sizeof(ne0)  atIndex:13];
+                                [encoder setBytes:&ne1  length:sizeof(ne1)  atIndex:14];
+
+                                if (src0t == GGML_TYPE_Q4_0 || src0t == GGML_TYPE_Q4_1) {
+                                    [encoder setThreadgroupMemoryLength:nth0*nth1*sizeof(float) atIndex:0];
+                                    [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne11, 1) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                                }
+                                else if (src0t == GGML_TYPE_Q2_K ||
+                                         src0t == GGML_TYPE_Q3_K ||
+                                         src0t == GGML_TYPE_Q4_K ||
+                                         src0t == GGML_TYPE_Q5_K ||
+                                         src0t == GGML_TYPE_Q6_K) {
+                                    [encoder setThreadgroupMemoryLength:nth0*nth1*sizeof(float) atIndex:0];
+                                    [encoder dispatchThreadgroups:MTLSizeMake(ne01, 1, 1) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                                } else {
+                                    [encoder setThreadgroupMemoryLength:nth0*sizeof(float) atIndex:0];
+                                    [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                                }
+                            }
+                        } break;
+                    case GGML_OP_GET_ROWS:
+                        {
+                            if (encoder == nil) {
+                                encoder = [command_buffer computeCommandEncoder];
+                            }
+
+                            switch (src0->type) {
+                                case GGML_TYPE_F16:  [encoder setComputePipelineState:ctx->pipeline_get_rows_f16]; break;
+                                case GGML_TYPE_Q4_0: [encoder setComputePipelineState:ctx->pipeline_get_rows_q4_0]; break;
+                                case GGML_TYPE_Q4_1: [encoder setComputePipelineState:ctx->pipeline_get_rows_q4_1]; break;
+                                case GGML_TYPE_Q2_K: [encoder setComputePipelineState:ctx->pipeline_get_rows_q2_k]; break;
+                                case GGML_TYPE_Q3_K: [encoder setComputePipelineState:ctx->pipeline_get_rows_q3_k]; break;
+                                case GGML_TYPE_Q4_K: [encoder setComputePipelineState:ctx->pipeline_get_rows_q4_k]; break;
+                                case GGML_TYPE_Q5_K: [encoder setComputePipelineState:ctx->pipeline_get_rows_q5_k]; break;
+                                case GGML_TYPE_Q6_K: [encoder setComputePipelineState:ctx->pipeline_get_rows_q6_k]; break;
+                                default: GGML_ASSERT(false && "not implemented");
+                            }
+
+                            [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                            [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
+                            [encoder setBuffer:id_dst  offset:offs_dst  atIndex:2];
+                            [encoder setBytes:&(src0->ne[0]) length:sizeof( int64_t) atIndex:3];
+                            [encoder setBytes:&(src0->nb[1]) length:sizeof(uint64_t) atIndex:4];
+                            [encoder setBytes:&(dst->nb[1])  length:sizeof(uint64_t) atIndex:5];
+
+                            const int64_t n = ggml_nelements(src1);
+
+                            [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                        } break;
+                    case GGML_OP_RMS_NORM:
+                        {
+                            if (encoder == nil) {
+                                encoder = [command_buffer computeCommandEncoder];
+                            }
+
+                            const float eps = 1e-6f;
+
+                            const int nth = 256;
+
+                            [encoder setComputePipelineState:ctx->pipeline_rms_norm];
+                            [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                            [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
+                            [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:nth*sizeof(float) atIndex:0];
+
+                            const int64_t nrows = ggml_nrows(src0);
+
+                            [encoder dispatchThreadgroups:MTLSizeMake(nrows, 1, 1) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
+                        } break;
+                    case GGML_OP_ROPE:
+                        {
+                            if (encoder == nil) {
+                                encoder = [command_buffer computeCommandEncoder];
+                            }
+
+                            const int n_dims = ((int32_t *) src1->data)[1];
+                            const int mode   = ((int32_t *) src1->data)[2];
+
+                            const int n_past = ((int32_t *)(src1->data))[0];
+
+                            [encoder setComputePipelineState:ctx->pipeline_rope];
+                            [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                            [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
+                            [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 setBytes:&n_past length:sizeof(     int) atIndex:18];
+                            [encoder setBytes:&n_dims length:sizeof(     int) atIndex:19];
+                            [encoder setBytes:&mode   length:sizeof(     int) atIndex:20];
+
+                            [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                        } break;
+                    case GGML_OP_CPY:
+                        {
+                            if (encoder == nil) {
+                                encoder = [command_buffer computeCommandEncoder];
+                            }
+
+                            const int nth = 32;
+
+                            switch (src0t) {
+                                case GGML_TYPE_F32:
+                                    {
+                                        switch (dstt) {
+                                            case GGML_TYPE_F16: [encoder setComputePipelineState:ctx->pipeline_cpy_f32_f16]; break;
+                                            case GGML_TYPE_F32: [encoder setComputePipelineState:ctx->pipeline_cpy_f32_f32]; break;
+                                            default: GGML_ASSERT(false && "not implemented");
+                                        };
+                                    } break;
+                                default: GGML_ASSERT(false && "not implemented");
+                            }
+
+                            [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                            [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
+                            [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;
+                    default:
+                        fprintf(stderr, "%s: node %3d, op = %8s not implemented\n", __func__, i, ggml_op_name(dst->op));
+                        GGML_ASSERT(false);
+                }
+            }
+
+            if (encoder != nil) {
+                [encoder endEncoding];
+                encoder = nil;
+            }
+
+            [command_buffer commit];
+        });
+    }
+
+    // wait for all threads to finish
+    dispatch_barrier_sync(queue, ^{});
+
+    [command_buffers[n_cb - 1] waitUntilCompleted];
+}

ggml-opencl.cpp

@@ -3,6 +3,8 @@
 #include <array>
 #include <atomic>
 #include <sstream>
+#include <vector>
+#include <limits>
 
 #define CL_TARGET_OPENCL_VERSION 110
 #include <clblast.h>
@@ -197,6 +199,18 @@ __kernel void KERNEL_NAME(__global X_TYPE* x, __local float* tmp, __global float
 }
 );
 
+std::string mul_template = MULTILINE_QUOTE(
+__kernel void KERNEL_NAME(__global TYPE* x, const int x_offset, __global TYPE* y, const int y_offset, __global TYPE* dst, const int dst_offset, const int ky) {
+    const int i = get_group_id(0)*get_local_size(0) + get_local_id(0);
+
+    if (i >= get_global_size(0)) {
+        return;
+    }
+
+    dst[dst_offset + i] = x[x_offset + i] * y[y_offset + i%ky];
+}
+);
+
 #define CL_CHECK(err)                                               \
     do {                                                            \
         cl_int err_ = (err);                                        \
@@ -239,6 +253,13 @@ std::array<std::string, 30> dequant_mul_mat_vec_str_values = {
     "convert_mul_mat_vec_f16", "half", "1", "1", "convert_f16"
 };
 
+std::array<std::string, 2> mul_str_keys = {
+    "KERNEL_NAME", "TYPE"
+};
+std::array<std::string, 2> mul_str_values = {
+    "mul_f32", "float"
+};
+
 std::string& replace(std::string& s, const std::string& from, const std::string& to) {
     size_t pos = 0;
     while ((pos = s.find(from, pos)) != std::string::npos) {
@@ -261,6 +282,13 @@ std::string generate_kernels() {
         src << dequant_kernel << '\n';
         src << dmmv_kernel << '\n';
     }
+    for (size_t i = 0; i < mul_str_values.size(); i += mul_str_keys.size()) {
+        std::string mul_kernel = mul_template;
+        for (size_t j = 0; j < mul_str_keys.size(); j++) {
+            replace(mul_kernel, mul_str_keys[j], mul_str_values[i + j]);
+        }
+        src << mul_kernel << '\n';
+    }
     return src.str();
 }
 
@@ -272,6 +300,7 @@ static cl_program program;
 static cl_kernel convert_row_f16_cl;
 static cl_kernel dequantize_row_q4_0_cl, dequantize_row_q4_1_cl, dequantize_row_q5_0_cl, dequantize_row_q5_1_cl, dequantize_row_q8_0_cl;
 static cl_kernel dequantize_mul_mat_vec_q4_0_cl, dequantize_mul_mat_vec_q4_1_cl, dequantize_mul_mat_vec_q5_0_cl, dequantize_mul_mat_vec_q5_1_cl, dequantize_mul_mat_vec_q8_0_cl, convert_mul_mat_vec_f16_cl;
+static cl_kernel mul_f32_cl;
 static bool fp16_support;
 
 static cl_program build_program_from_source(cl_context ctx, cl_device_id dev, const char* program_buffer) {
@@ -508,6 +537,9 @@ void ggml_cl_init(void) {
     CL_CHECK((dequantize_mul_mat_vec_q5_1_cl = clCreateKernel(program, "dequantize_mul_mat_vec_q5_1", &err), err));
     CL_CHECK((dequantize_mul_mat_vec_q8_0_cl = clCreateKernel(program, "dequantize_mul_mat_vec_q8_0", &err), err));
     CL_CHECK((convert_mul_mat_vec_f16_cl = clCreateKernel(program, "convert_mul_mat_vec_f16", &err), err));
+
+    // mul kernel
+    CL_CHECK((mul_f32_cl = clCreateKernel(program, "mul_f32", &err), err));
 }
 
 static cl_kernel* ggml_get_to_fp32_cl(ggml_type type) {
@@ -573,21 +605,44 @@ struct cl_buffer {
 static cl_buffer g_cl_buffer_pool[MAX_CL_BUFFERS];
 static std::atomic_flag g_cl_pool_lock = ATOMIC_FLAG_INIT;
 
-static cl_mem ggml_cl_pool_malloc(size_t size, size_t * actual_size, cl_mem_flags flags) {
+static cl_mem ggml_cl_pool_malloc(size_t size, size_t * actual_size) {
     scoped_spin_lock lock(g_cl_pool_lock);
     cl_int err;
 
+    int best_i = -1;
+    size_t best_size = std::numeric_limits<size_t>::max(); //smallest unused buffer that fits our needs
+    int worst_i = -1;
+    size_t worst_size = 0; //largest unused buffer seen so far
     for (int i = 0; i < MAX_CL_BUFFERS; ++i) {
-        cl_buffer& b = g_cl_buffer_pool[i];
-        if (b.size > 0 && b.size >= size) {
-            cl_mem mem = b.mem;
-            *actual_size = b.size;
-            b.size = 0;
-            return mem;
+        cl_buffer &b = g_cl_buffer_pool[i];
+        if (b.size > 0 && b.size >= size && b.size < best_size)
+        {
+            best_i = i;
+            best_size = b.size;
+        }
+        if (b.size > 0 && b.size > worst_size)
+        {
+            worst_i = i;
+            worst_size = b.size;
         }
     }
+    if(best_i!=-1) //found the smallest buffer that fits our needs
+    {
+        cl_buffer& b = g_cl_buffer_pool[best_i];
+        cl_mem mem = b.mem;
+        *actual_size = b.size;
+        b.size = 0;
+        return mem;
+    }
+    if(worst_i!=-1) //no buffer that fits our needs, resize largest one to save memory
+    {
+         cl_buffer& b = g_cl_buffer_pool[worst_i];
+         cl_mem mem = b.mem;
+         b.size = 0;
+         clReleaseMemObject(mem);
+    }
     cl_mem mem;
-    CL_CHECK((mem = clCreateBuffer(context, flags, size, NULL, &err), err));
+    CL_CHECK((mem = clCreateBuffer(context, CL_MEM_READ_WRITE, size, NULL, &err), err));
     *actual_size = size;
     return mem;
 }
@@ -607,6 +662,15 @@ static void ggml_cl_pool_free(cl_mem mem, size_t size) {
     clReleaseMemObject(mem);
 }
 
+void ggml_cl_free_data(const struct ggml_tensor* tensor) {
+    if (tensor->backend != GGML_BACKEND_GPU) {
+        return;
+    }
+
+    cl_mem mem = (cl_mem)tensor->data;
+    clReleaseMemObject(mem);
+}
+
 static cl_int ggml_cl_h2d_tensor_2d(cl_command_queue queue, cl_mem dst, size_t offset, const struct ggml_tensor * src, uint64_t i3, uint64_t i2, cl_event* ev) {
     cl_int err;
     const uint64_t ne0 = src->ne[0];
@@ -644,6 +708,99 @@ static cl_int ggml_cl_h2d_tensor_2d(cl_command_queue queue, cl_mem dst, size_t o
     return err;
 }
 
+static void ggml_cl_mul_f32(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_ASSERT(src1->backend == GGML_BACKEND_GPU);
+    const int64_t ne00 = src0->ne[0];
+    const int64_t ne01 = src0->ne[1];
+    const int64_t ne02 = src0->ne[2];
+    const int64_t ne03 = src0->ne[2];
+    const int64_t ne0 = ne00 * ne01 * ne02 * ne03;
+    const int64_t ne10 = src1->ne[0];
+    const int64_t ne11 = src1->ne[1];
+    const int64_t ne12 = src1->ne[2];
+    const int64_t ne13 = src1->ne[3];
+    const int64_t nb10 = src1->nb[0];
+    const int nb2  = dst->nb[2];
+    const int nb3  = dst->nb[3];
+    size_t x_size;
+    size_t d_size;
+
+    cl_mem d_X = ggml_cl_pool_malloc(ne0 * sizeof(float), &x_size); // src0
+    cl_mem d_Y = (cl_mem) src1->data; // src1 is already on device, broadcasted.
+    cl_mem d_D = ggml_cl_pool_malloc(ne0 * sizeof(float), &d_size); // dst
+
+
+    for (int64_t i03 = 0; i03 < ne03; i03++) {
+        for (int64_t i02 = 0; i02 < ne02; i02++) {
+            const int i0 = i03*ne02 + i02;
+
+            cl_event ev;
+
+            // copy src0 to device
+            CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_X, i0, src0, i03, i02, &ev));
+
+            if (nb10 == sizeof(float)) {
+                // Contiguous, avoid overhead from queueing many kernel runs
+                const int64_t i13 = i03%ne13;
+                const int64_t i12 = i02%ne12;
+                const int i1 = i13*ne12*ne11 + i12*ne11;
+
+                cl_int x_offset = 0;
+                cl_int y_offset = i1*ne10;
+                cl_int d_offset = 0;
+
+                size_t global = ne00 * ne01;
+                cl_int ky = ne10;
+                CL_CHECK(clSetKernelArg(mul_f32_cl, 0, sizeof(cl_mem), &d_X));
+                CL_CHECK(clSetKernelArg(mul_f32_cl, 1, sizeof(cl_int), &x_offset));
+                CL_CHECK(clSetKernelArg(mul_f32_cl, 2, sizeof(cl_mem), &d_Y));
+                CL_CHECK(clSetKernelArg(mul_f32_cl, 3, sizeof(cl_int), &y_offset));
+                CL_CHECK(clSetKernelArg(mul_f32_cl, 4, sizeof(cl_mem), &d_D));
+                CL_CHECK(clSetKernelArg(mul_f32_cl, 5, sizeof(cl_int), &d_offset));
+                CL_CHECK(clSetKernelArg(mul_f32_cl, 6, sizeof(cl_int), &ky));
+                CL_CHECK(clEnqueueNDRangeKernel(queue, mul_f32_cl, 1, NULL, &global, NULL, 1, &ev, NULL));
+            } else {
+                for (int64_t i01 = 0; i01 < ne01; i01++) {
+                    const int64_t i13 = i03%ne13;
+                    const int64_t i12 = i02%ne12;
+                    const int64_t i11 = i01%ne11;
+                    const int i1 = i13*ne12*ne11 + i12*ne11 + i11;
+
+                    cl_int x_offset = i01*ne00;
+                    cl_int y_offset = i1*ne10;
+                    cl_int d_offset = i01*ne00;
+
+                    // compute
+                    size_t global = ne00;
+                    cl_int ky = ne10;
+                    CL_CHECK(clSetKernelArg(mul_f32_cl, 0, sizeof(cl_mem), &d_X));
+                    CL_CHECK(clSetKernelArg(mul_f32_cl, 1, sizeof(cl_int), &x_offset));
+                    CL_CHECK(clSetKernelArg(mul_f32_cl, 2, sizeof(cl_mem), &d_Y));
+                    CL_CHECK(clSetKernelArg(mul_f32_cl, 3, sizeof(cl_int), &y_offset));
+                    CL_CHECK(clSetKernelArg(mul_f32_cl, 4, sizeof(cl_mem), &d_D));
+                    CL_CHECK(clSetKernelArg(mul_f32_cl, 5, sizeof(cl_int), &d_offset));
+                    CL_CHECK(clSetKernelArg(mul_f32_cl, 6, sizeof(cl_int), &ky));
+                    CL_CHECK(clEnqueueNDRangeKernel(queue, mul_f32_cl, 1, NULL, &global, NULL, 1, &ev, NULL));
+                }
+            }
+
+            CL_CHECK(clReleaseEvent(ev));
+            CL_CHECK(clFinish(queue));
+
+            // copy dst to host
+            float * d = (float *) ((char *) dst->data + i02*nb2 + i03*nb3);
+            CL_CHECK(clEnqueueReadBuffer(queue, d_D, true, 0, sizeof(float) * ne00*ne01, d, 0, NULL, NULL));
+        }
+    }
+    ggml_cl_pool_free(d_X, x_size);
+    ggml_cl_pool_free(d_D, d_size);
+}
+
+void ggml_cl_mul(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst) {
+    GGML_ASSERT(src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32);
+    ggml_cl_mul_f32(src0, src1, dst);
+}
+
 static void ggml_cl_mul_mat_f32(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     const int64_t ne00 = src0->ne[0];
     const int64_t ne01 = src0->ne[1];
@@ -666,18 +823,18 @@ static void ggml_cl_mul_mat_f32(const ggml_tensor * src0, const ggml_tensor * sr
     size_t y_size;
     size_t d_size;
     cl_mem d_X;
-    if (src0->backend == GGML_BACKEND_CL) {
+    if (src0->backend == GGML_BACKEND_GPU) { // NOLINT
         d_X = (cl_mem) src0->data;
     } else {
-        d_X = ggml_cl_pool_malloc(sizeof(ggml_fp16_t) * x_ne, &x_size, CL_MEM_READ_ONLY);
+        d_X = ggml_cl_pool_malloc(sizeof(ggml_fp16_t) * x_ne, &x_size);
     }
-    cl_mem d_Y = ggml_cl_pool_malloc(sizeof(float) * y_ne, &y_size, CL_MEM_READ_ONLY);
-    cl_mem d_D = ggml_cl_pool_malloc(sizeof(float) * d_ne, &d_size, CL_MEM_WRITE_ONLY);
+    cl_mem d_Y = ggml_cl_pool_malloc(sizeof(float) * y_ne, &y_size);
+    cl_mem d_D = ggml_cl_pool_malloc(sizeof(float) * d_ne, &d_size);
 
     for (int64_t i03 = 0; i03 < ne03; i03++) {
         for (int64_t i02 = 0; i02 < ne02; i02++) {
             // copy data to device
-            if (src0->backend != GGML_BACKEND_CL) {
+            if (src0->backend != GGML_BACKEND_GPU) {
                 CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_X, 0, src0, i03, i02, NULL));
             }
             CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_Y, 0, src1, i03, i02, NULL));
@@ -706,7 +863,7 @@ static void ggml_cl_mul_mat_f32(const ggml_tensor * src0, const ggml_tensor * sr
         }
     }
 
-    if (src0->backend != GGML_BACKEND_CL) {
+    if (src0->backend != GGML_BACKEND_GPU) {
         ggml_cl_pool_free(d_X, x_size);
     }
     ggml_cl_pool_free(d_Y, y_size);
@@ -742,13 +899,13 @@ static void ggml_cl_mul_mat_f16(const ggml_tensor * src0, const ggml_tensor * sr
     size_t y_size;
     size_t d_size;
     cl_mem d_X;
-    if (src0->backend == GGML_BACKEND_CL) {
+    if (src0->backend == GGML_BACKEND_GPU) { // NOLINT
         d_X = (cl_mem) src0->data;
     } else {
-        d_X = ggml_cl_pool_malloc(sizeof(ggml_fp16_t) * x_ne, &x_size, CL_MEM_READ_ONLY);
+        d_X = ggml_cl_pool_malloc(sizeof(ggml_fp16_t) * x_ne, &x_size);
     }
-    cl_mem d_Y = ggml_cl_pool_malloc(sizeof(ggml_fp16_t) * y_ne, &y_size, CL_MEM_READ_ONLY);
-    cl_mem d_D = ggml_cl_pool_malloc(sizeof(ggml_fp16_t) * d_ne, &d_size, CL_MEM_WRITE_ONLY);
+    cl_mem d_Y = ggml_cl_pool_malloc(sizeof(ggml_fp16_t) * y_ne, &y_size);
+    cl_mem d_D = ggml_cl_pool_malloc(sizeof(ggml_fp16_t) * d_ne, &d_size);
 
     bool src1_cont_rows = nb10 == sizeof(float);
     bool src1_cont_cols = (size_t)nb11 == ne11*sizeof(float);
@@ -756,7 +913,7 @@ static void ggml_cl_mul_mat_f16(const ggml_tensor * src0, const ggml_tensor * sr
     for (int64_t i03 = 0; i03 < ne03; i03++) {
         for (int64_t i02 = 0; i02 < ne02; i02++) {
             // copy src0 to device
-            if (src0->backend != GGML_BACKEND_CL) {
+            if (src0->backend != GGML_BACKEND_GPU) {
                 CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_X, 0, src0, i03, i02, NULL));
             }
 
@@ -813,7 +970,7 @@ static void ggml_cl_mul_mat_f16(const ggml_tensor * src0, const ggml_tensor * sr
         }
     }
 
-    if (src0->backend != GGML_BACKEND_CL) {
+    if (src0->backend != GGML_BACKEND_GPU) {
         ggml_cl_pool_free(d_X, x_size);
     }
     ggml_cl_pool_free(d_Y, y_size);
@@ -847,57 +1004,61 @@ static void ggml_cl_mul_mat_q_f32(const ggml_tensor * src0, const ggml_tensor *
     size_t q_size;
     cl_mem d_X;
     if (!mul_mat_vec) {
-        d_X = ggml_cl_pool_malloc(sizeof(float) * x_ne, &x_size, CL_MEM_READ_WRITE);
+        d_X = ggml_cl_pool_malloc(sizeof(float) * x_ne, &x_size);
     }
-    cl_mem d_Y = ggml_cl_pool_malloc(sizeof(float) * y_ne, &y_size, CL_MEM_READ_ONLY);
-    cl_mem d_D = ggml_cl_pool_malloc(sizeof(float) * d_ne, &d_size, CL_MEM_WRITE_ONLY);
+    cl_mem d_Y = ggml_cl_pool_malloc(sizeof(float) * y_ne, &y_size);
+    cl_mem d_D = ggml_cl_pool_malloc(sizeof(float) * d_ne, &d_size);
     cl_mem d_Q;
     if (src0->backend == GGML_BACKEND_CPU) {
-        d_Q = ggml_cl_pool_malloc(q_sz, &q_size, CL_MEM_READ_ONLY);
+        d_Q = ggml_cl_pool_malloc(q_sz, &q_size);
     }
 
     cl_kernel* to_fp32_cl = ggml_get_to_fp32_cl(type);
     cl_kernel* dmmv = ggml_get_dequantize_mul_mat_vec_cl(type);
     GGML_ASSERT(to_fp32_cl != nullptr);
 
+    size_t ev_idx = 0;
+    std::vector<cl_event> events;
+
     for (int64_t i03 = 0; i03 < ne03; i03++) {
         for (int64_t i02 = 0; i02 < ne02; i02++) {
-            cl_event ev_sgemm;
-
             // copy src0 to device if necessary
             if (src0->backend == GGML_BACKEND_CPU) {
-                CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_Q, 0, src0, i03, i02, NULL));
-            } else if (src0->backend == GGML_BACKEND_CL) {
+                events.emplace_back();
+                CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_Q, 0, src0, i03, i02, events.data() + ev_idx++));
+            } else if (src0->backend == GGML_BACKEND_GPU) {
                 d_Q = (cl_mem) src0->data;
             } else {
                 GGML_ASSERT(false);
             }
             if (mul_mat_vec) { // specialized dequantize_mul_mat_vec kernel
                 // copy src1 to device
-                CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_Y, 0, src1, i03, i02, NULL));
+                events.emplace_back();
+                CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_Y, 0, src1, i03, i02, events.data() + ev_idx++));
 
                 // compute
                 const size_t global = ne01 * CL_DMMV_BLOCK_SIZE;
                 const size_t local = CL_DMMV_BLOCK_SIZE;
                 const cl_int ncols = ne00;
+                events.emplace_back();
                 CL_CHECK(clSetKernelArg(*dmmv, 0, sizeof(cl_mem), &d_Q));
                 CL_CHECK(clSetKernelArg(*dmmv, 1, sizeof(float) * local, NULL));
                 CL_CHECK(clSetKernelArg(*dmmv, 2, sizeof(cl_mem), &d_Y));
                 CL_CHECK(clSetKernelArg(*dmmv, 3, sizeof(cl_mem), &d_D));
                 CL_CHECK(clSetKernelArg(*dmmv, 4, sizeof(cl_int), &ncols));
-                CL_CHECK(clFinish(queue));
-                CL_CHECK(clEnqueueNDRangeKernel(queue, *dmmv, 1, NULL, &global, &local, 0, NULL, &ev_sgemm));
+                CL_CHECK(clEnqueueNDRangeKernel(queue, *dmmv, 1, NULL, &global, &local, events.size() - 1, events.data(), events.data() + ev_idx++));
             } else { // general dequantization kernel + CLBlast matrix matrix multiplication
                 // convert src0 to fp32 on device
                 const size_t global = x_ne;
                 CL_CHECK(clSetKernelArg(*to_fp32_cl, 0, sizeof(cl_mem), &d_Q));
                 CL_CHECK(clSetKernelArg(*to_fp32_cl, 1, sizeof(cl_mem), &d_X));
-                CL_CHECK(clFinish(queue));
-                CL_CHECK(clEnqueueNDRangeKernel(queue, *to_fp32_cl, 1, NULL, &global, NULL, 0, NULL, NULL));
+                CL_CHECK(clEnqueueNDRangeKernel(queue, *to_fp32_cl, 1, NULL, &global, NULL, events.size(), !events.empty() ? events.data() : NULL, NULL));
 
                 // copy src1 to device
                 CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_Y, 0, src1, i03, i02, NULL));
 
+                events.emplace_back();
+
                 // wait for conversion
                 CL_CHECK(clFinish(queue));
 
@@ -910,7 +1071,7 @@ static void ggml_cl_mul_mat_q_f32(const ggml_tensor * src0, const ggml_tensor *
                                                            d_Y, 0, ne10,
                                                            beta,
                                                            d_D, 0, ne01,
-                                                           &queue, &ev_sgemm);
+                                                           &queue, events.data() + ev_idx++);
 
                 if (status != clblast::StatusCode::kSuccess) {
                     GGML_ASSERT(false);
@@ -919,8 +1080,13 @@ static void ggml_cl_mul_mat_q_f32(const ggml_tensor * src0, const ggml_tensor *
 
             // copy dst to host
             float * d = (float *) ((char *) dst->data + i02*nb2 + i03*nb3);
-            CL_CHECK(clEnqueueReadBuffer(queue, d_D, true, 0, sizeof(float) * d_ne, d, 1, &ev_sgemm, NULL));
-            clReleaseEvent(ev_sgemm);
+            CL_CHECK(clEnqueueReadBuffer(queue, d_D, true, 0, sizeof(float) * d_ne, d, 1, &events[events.size() - 1], NULL));
+            for (auto *event : events) {
+                clReleaseEvent(event);
+            }
+
+            ev_idx = 0;
+            events.clear();
         }
     }
 
@@ -945,7 +1111,7 @@ bool ggml_cl_can_mul_mat(const struct ggml_tensor * src0, const struct ggml_tens
     if ((src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 || ggml_is_quantized(src0->type)) &&
         src1->type == GGML_TYPE_F32 &&
         dst->type == GGML_TYPE_F32 &&
-        ((ne0 >= 32 && ne1 >= 32 && ne10 >= 32) || src0->backend == GGML_BACKEND_CL)) {
+        ((ne0 >= 32 && ne1 >= 32 && ne10 >= 32) || src0->backend == GGML_BACKEND_GPU)) {
         return true;
     }
 
@@ -1001,7 +1167,7 @@ size_t ggml_cl_mul_mat_get_wsize(const struct ggml_tensor * src0, const struct g
     return 0;
 }
 
-void ggml_cl_transform_tensor(ggml_tensor * tensor) {
+void ggml_cl_transform_tensor(void * data, ggml_tensor * tensor) {
     const int64_t ne0 = tensor->ne[0];
     const int64_t ne1 = tensor->ne[1];
     const int64_t ne2 = tensor->ne[2];
@@ -1011,8 +1177,9 @@ void ggml_cl_transform_tensor(ggml_tensor * tensor) {
     const size_t q_sz = ggml_type_size(type) * ne0 * ne1 * ne2 * ne3 / ggml_blck_size(type);
 
     size_t q_size;
-    cl_mem dst = ggml_cl_pool_malloc(q_sz, &q_size, CL_MEM_READ_ONLY);
+    cl_mem dst = ggml_cl_pool_malloc(q_sz, &q_size);
 
+    tensor->data = data;
     // copy tensor to device
     for (int64_t i3 = 0; i3 < ne3; i3++) {
         for (int64_t i2 = 0; i2 < ne2; i2++) {
@@ -1024,5 +1191,5 @@ void ggml_cl_transform_tensor(ggml_tensor * tensor) {
     CL_CHECK(clFinish(queue));
 
     tensor->data = dst;
-    tensor->backend = GGML_BACKEND_CL;
+    GGML_ASSERT(tensor->backend == GGML_BACKEND_GPU);
 }

ggml-opencl.h

@@ -8,6 +8,7 @@ extern "C" {
 
 void ggml_cl_init(void);
 
+void   ggml_cl_mul(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
 bool   ggml_cl_can_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
 size_t ggml_cl_mul_mat_get_wsize(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
 void   ggml_cl_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst, void * wdata, size_t wsize);
@@ -15,7 +16,9 @@ void   ggml_cl_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor
 void * ggml_cl_host_malloc(size_t size);
 void   ggml_cl_host_free(void * ptr);
 
-void ggml_cl_transform_tensor(struct ggml_tensor * tensor);
+void ggml_cl_free_data(const struct ggml_tensor* tensor);
+
+void ggml_cl_transform_tensor(void * data, struct ggml_tensor * tensor);
 
 #ifdef  __cplusplus
 }

ggml.h

@@ -241,6 +241,13 @@ extern "C" {
         GGML_TYPE_Q5_1 = 7,
         GGML_TYPE_Q8_0 = 8,
         GGML_TYPE_Q8_1 = 9,
+        // k-quantizations
+        GGML_TYPE_Q2_K = 10,
+        GGML_TYPE_Q3_K = 11,
+        GGML_TYPE_Q4_K = 12,
+        GGML_TYPE_Q5_K = 13,
+        GGML_TYPE_Q6_K = 14,
+        GGML_TYPE_Q8_K = 15,
         GGML_TYPE_I8,
         GGML_TYPE_I16,
         GGML_TYPE_I32,
@@ -249,8 +256,8 @@ extern "C" {
 
     enum ggml_backend {
         GGML_BACKEND_CPU = 0,
-        GGML_BACKEND_CUDA = 1,
-        GGML_BACKEND_CL = 2,
+        GGML_BACKEND_GPU = 10,
+        GGML_BACKEND_GPU_SPLIT = 20,
     };
 
     // model file types
@@ -264,6 +271,11 @@ extern "C" {
         GGML_FTYPE_MOSTLY_Q8_0 = 7,  // except 1d tensors
         GGML_FTYPE_MOSTLY_Q5_0 = 8,  // except 1d tensors
         GGML_FTYPE_MOSTLY_Q5_1 = 9,  // except 1d tensors
+        GGML_FTYPE_MOSTLY_Q2_K = 10, // except 1d tensors
+        GGML_FTYPE_MOSTLY_Q3_K = 11, // except 1d tensors
+        GGML_FTYPE_MOSTLY_Q4_K = 12, // except 1d tensors
+        GGML_FTYPE_MOSTLY_Q5_K = 13, // except 1d tensors
+        GGML_FTYPE_MOSTLY_Q6_K = 14, // except 1d tensors
     };
 
     // available tensor operations:
@@ -284,6 +296,7 @@ extern "C" {
         GGML_OP_SUM_ROWS,
         GGML_OP_MEAN,
         GGML_OP_REPEAT,
+        GGML_OP_REPEAT_BACK,
         GGML_OP_ABS,
         GGML_OP_SGN,
         GGML_OP_NEG,
@@ -297,6 +310,7 @@ extern "C" {
         GGML_OP_RMS_NORM_BACK,
 
         GGML_OP_MUL_MAT,
+        GGML_OP_OUT_PROD,
 
         GGML_OP_SCALE,
         GGML_OP_SET,
@@ -312,6 +326,7 @@ extern "C" {
         GGML_OP_DIAG_MASK_INF,
         GGML_OP_DIAG_MASK_ZERO,
         GGML_OP_SOFT_MAX,
+        GGML_OP_SOFT_MAX_BACK,
         GGML_OP_ROPE,
         GGML_OP_ROPE_BACK,
         GGML_OP_ALIBI,
@@ -321,10 +336,14 @@ extern "C" {
 
         GGML_OP_FLASH_ATTN,
         GGML_OP_FLASH_FF,
+        GGML_OP_FLASH_ATTN_BACK,
 
         GGML_OP_MAP_UNARY,
         GGML_OP_MAP_BINARY,
 
+        GGML_OP_CROSS_ENTROPY_LOSS,
+        GGML_OP_CROSS_ENTROPY_LOSS_BACK,
+
         GGML_OP_COUNT,
     };
 
@@ -375,7 +394,9 @@ extern "C" {
 
         char name[GGML_MAX_NAME];
 
-        char padding[16];
+        void * extra; // extra things e.g. for ggml-cuda.cu
+
+        char padding[4];
     };
 
     static const size_t GGML_TENSOR_SIZE = sizeof(struct ggml_tensor);
@@ -413,6 +434,25 @@ extern "C" {
         bool   no_alloc;   // don't allocate memory for the tensor data
     };
 
+
+    // compute types
+    enum ggml_task_type {
+        GGML_TASK_INIT = 0,
+        GGML_TASK_COMPUTE,
+        GGML_TASK_FINALIZE,
+    };
+
+    struct ggml_compute_params {
+        enum ggml_task_type type;
+
+        // ith = thread index, nth = number of threads
+        int ith, nth;
+
+        // work buffer for all threads
+        size_t wsize;
+        void * wdata;
+    };
+
     // misc
 
     GGML_API void    ggml_time_init(void); // call this once at the beginning of the program
@@ -424,8 +464,10 @@ extern "C" {
     GGML_API void    ggml_print_object (const struct ggml_object * obj);
     GGML_API void    ggml_print_objects(const struct ggml_context * ctx);
 
-    GGML_API int64_t ggml_nelements(const struct ggml_tensor * tensor);
-    GGML_API size_t  ggml_nbytes   (const struct ggml_tensor * tensor);
+    GGML_API int64_t ggml_nelements   (const struct ggml_tensor * tensor);
+    GGML_API int64_t ggml_nrows       (const struct ggml_tensor * tensor);
+    GGML_API size_t  ggml_nbytes      (const struct ggml_tensor * tensor);
+    GGML_API size_t  ggml_nbytes_split(const struct ggml_tensor * tensor, int nrows_split);
 
     GGML_API int     ggml_blck_size (enum ggml_type type);
     GGML_API size_t  ggml_type_size (enum ggml_type type); // size in bytes for all elements in a block
@@ -441,13 +483,17 @@ extern "C" {
     // TODO: temporary until model loading of ggml examples is refactored
     GGML_API enum ggml_type ggml_ftype_to_ggml_type(enum ggml_ftype ftype);
 
+    GGML_API bool ggml_is_transposed(const struct ggml_tensor * tensor);
+    GGML_API bool ggml_is_contiguous(const struct ggml_tensor * tensor);
+    GGML_API bool ggml_is_permuted  (const struct ggml_tensor * tensor);
+
     // use this to compute the memory overhead of a tensor
     GGML_API size_t ggml_tensor_overhead(void);
 
     // main
 
     GGML_API struct ggml_context * ggml_init(struct ggml_init_params params);
-    GGML_API void    ggml_free(struct ggml_context * ctx);
+    GGML_API void                  ggml_free(struct ggml_context * ctx);
 
     GGML_API size_t  ggml_used_mem(const struct ggml_context * ctx);
 
@@ -536,6 +582,11 @@ extern "C" {
             struct ggml_tensor  * a,
             struct ggml_tensor  * b);
 
+    GGML_API struct ggml_tensor * ggml_add1_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b);
+
     GGML_API struct ggml_tensor * ggml_acc(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
@@ -607,6 +658,11 @@ extern "C" {
             struct ggml_tensor  * a,
             struct ggml_tensor  * b);
 
+    GGML_API struct ggml_tensor * ggml_repeat_back(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b);
+
     GGML_API struct ggml_tensor * ggml_abs(
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
@@ -660,14 +716,22 @@ extern "C" {
             struct ggml_tensor  * a,
             struct ggml_tensor  * b);
 
-    // A: m rows, n columns
-    // B: p rows, n columns (i.e. we transpose it internally)
+    // A: n columns, m rows
+    // B: n columns, p rows  (i.e. we transpose it internally)
     // result is m columns, p rows
     GGML_API struct ggml_tensor * ggml_mul_mat(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
             struct ggml_tensor  * b);
 
+    // A: m columns, n rows,
+    // B: p columns, n rows,
+    // result is m columns, p rows
+    GGML_API struct ggml_tensor * ggml_out_prod(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b);
+
     //
     // operations on tensors without backpropagation
     //
@@ -878,6 +942,17 @@ extern "C" {
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
 
+    GGML_API struct ggml_tensor * ggml_soft_max_back(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b);
+
+    // in-place, returns view(a)
+    GGML_API struct ggml_tensor * ggml_soft_max_back_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b);
+
     // rotary position embedding
     // if mode & 1 == 1, skip n_past elements
     // if mode & 2 == 1, GPT-NeoX style
@@ -944,6 +1019,14 @@ extern "C" {
             struct ggml_tensor  * v,
             bool                  masked);
 
+    GGML_API struct ggml_tensor * ggml_flash_attn_back(
+           struct ggml_context * ctx,
+           struct ggml_tensor  * q,
+           struct ggml_tensor  * k,
+           struct ggml_tensor  * v,
+           struct ggml_tensor  * d,
+           bool                  masked);
+
     GGML_API struct ggml_tensor * ggml_flash_ff(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
@@ -967,6 +1050,19 @@ extern "C" {
             struct ggml_tensor          * b,
                    ggml_binary_op_f32_t   fun);
 
+    // loss function
+
+    GGML_API struct ggml_tensor * ggml_cross_entropy_loss(
+            struct ggml_context         * ctx,
+            struct ggml_tensor          * a,
+            struct ggml_tensor          * b);
+
+    GGML_API struct ggml_tensor * ggml_cross_entropy_loss_back(
+            struct ggml_context         * ctx,
+            struct ggml_tensor          * a,
+            struct ggml_tensor          * b,
+            struct ggml_tensor          * c);
+
     //
     // automatic differentiation
     //
@@ -1061,6 +1157,8 @@ extern "C" {
         struct {
             int n_iter;
 
+            float sched; // schedule multiplier (fixed, decay or warmup)
+            float decay; // weight decay for AdamW, use 0.0f to disable
             float alpha; // learning rate
             float beta1;
             float beta2;
@@ -1085,6 +1183,49 @@ extern "C" {
         } 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;
+
+        struct {
+            struct ggml_tensor * x;  // view of the parameters
+            struct ggml_tensor * g1; // gradient
+            struct ggml_tensor * g2; // gradient squared
+            struct ggml_tensor * m;  // first moment
+            struct ggml_tensor * v;  // second moment
+            struct ggml_tensor * mh; // first moment hat
+            struct ggml_tensor * vh; // second moment hat
+            struct ggml_tensor * pf; // past function values
+            float fx_best;
+            float fx_prev;
+            int n_no_improvement;
+        } adam;
+
+        struct {
+            struct ggml_tensor * x;    // current parameters
+            struct ggml_tensor * xp;   // previous parameters
+            struct ggml_tensor * g;    // current gradient
+            struct ggml_tensor * gp;   // previous gradient
+            struct ggml_tensor * d;    // search direction
+            struct ggml_tensor * pf;   // past function values
+            struct ggml_tensor * lmal; // the L-BFGS memory alpha
+            struct ggml_tensor * lmys; // the L-BFGS memory ys
+            struct ggml_tensor * lms;  // the L-BFGS memory s
+            struct ggml_tensor * lmy;  // the L-BFGS memory y
+            float fx_best;
+            float step;
+            int j;
+            int k;
+            int end;
+            int n_no_improvement;
+        } lbfgs;
+    };
+
     GGML_API struct ggml_opt_params ggml_opt_default_params(enum ggml_opt_type type);
 
     // optimize the function defined by the tensor f
@@ -1093,6 +1234,27 @@ extern "C" {
             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);
+
     //
     // quantization
     //

k_quants.h

@@ -0,0 +1,122 @@
+#pragma once
+
+#include "ggml.h"
+
+#include <stdint.h>
+#include <assert.h>
+#include <stddef.h>
+
+// Super-block size
+#define QK_K 256
+
+//
+// Super-block quantization structures
+//
+
+// 2-bit quantization
+// weight is represented as x = a * q + b
+// 16 blocks of 16 elemenets each
+// Effectively 2.5625 bits per weight
+typedef struct {
+    uint8_t scales[QK_K/16]; // scales and mins, quantized with 4 bits
+    uint8_t qs[QK_K/4];      // quants
+    ggml_fp16_t d;           // super-block scale for quantized scales
+    ggml_fp16_t dmin;        // super-block scale for quantized mins
+} block_q2_K;
+static_assert(sizeof(block_q2_K) == 2*sizeof(ggml_fp16_t) + QK_K/16 + QK_K/4, "wrong q2_K block size/padding");
+
+// 3-bit quantization
+// weight is represented as x = a * q
+// 16 blocks of 16 elemenets each
+// Effectively 3.4375 bits per weight
+typedef struct {
+    uint8_t hmask[QK_K/8];     // quants - high bit
+    uint8_t qs[QK_K/4];        // quants - low 2 bits
+    uint8_t scales[3*QK_K/64]; // scales, quantized with 6 bits
+    ggml_fp16_t d;             // super-block scale
+} block_q3_K;
+static_assert(sizeof(block_q3_K) == sizeof(ggml_fp16_t) + QK_K / 4 + 11 * QK_K / 64, "wrong q3_K block size/padding");
+
+// 4-bit quantization
+// 16 blocks of 32 elements each
+// weight is represented as x = a * q + b
+// Effectively 4.5 bits per weight
+typedef struct {
+    ggml_fp16_t d;             // super-block scale for quantized scales
+    ggml_fp16_t dmin;          // super-block scale for quantized mins
+    uint8_t scales[3*QK_K/64]; // scales and mins, quantized with 6 bits
+    uint8_t qs[QK_K/2];        // 4--bit quants
+} block_q4_K;
+static_assert(sizeof(block_q4_K) == 2*sizeof(ggml_fp16_t) + 3*QK_K/64 + QK_K/2, "wrong q4_K block size/padding");
+
+// 5-bit quantization
+// 16 blocks of 32 elements each
+// weight is represented as x = a * q + b
+// Effectively 5.5 bits per weight
+typedef struct {
+    ggml_fp16_t d;               // super-block scale for quantized scales
+    ggml_fp16_t dmin;            // super-block scale for quantized mins
+    uint8_t scales[3*QK_K/64];   // scales and mins, quantized with 6 bits
+    uint8_t qh[QK_K/8];          // quants, high bit
+    uint8_t qs[QK_K/2];          // quants, low 4 bits
+} block_q5_K;
+static_assert(sizeof(block_q5_K) == 2*sizeof(ggml_fp16_t) + 3*QK_K/64 + QK_K/2 + QK_K/8, "wrong q5_K block size/padding");
+
+// 6-bit quantization
+// weight is represented as x = a * q
+// 16 blocks of 16 elemenets each
+// Effectively 6.5625 bits per weight
+typedef struct {
+    uint8_t ql[QK_K/2];      // quants, lower 4 bits
+    uint8_t qh[QK_K/4];      // quants, upper 2 bits
+    int8_t  scales[QK_K/16]; // scales, quantized with 8 bits
+    ggml_fp16_t d;           // super-block scale
+} block_q6_K;
+static_assert(sizeof(block_q6_K) == sizeof(ggml_fp16_t) + QK_K / 16 + 3*QK_K/4, "wrong q6_K block size/padding");
+
+// This is only used for intermediate quantization and dot products
+typedef struct {
+    float   d;              // delta
+    int8_t  qs[QK_K];       // quants
+    int16_t bsums[QK_K/16]; // sum of quants in groups of 16
+} block_q8_K;
+static_assert(sizeof(block_q8_K) == sizeof(float) + QK_K + QK_K/16*sizeof(int16_t), "wrong q8_K block size/padding");
+
+
+// Quantization
+void quantize_row_q2_K_reference(const float * restrict x, block_q2_K * restrict y, int k);
+void quantize_row_q3_K_reference(const float * restrict x, block_q3_K * restrict y, int k);
+void quantize_row_q4_K_reference(const float * restrict x, block_q4_K * restrict y, int k);
+void quantize_row_q5_K_reference(const float * restrict x, block_q5_K * restrict y, int k);
+void quantize_row_q6_K_reference(const float * restrict x, block_q6_K * restrict y, int k);
+void quantize_row_q8_K_reference(const float * restrict x, block_q8_K * restrict y, int k);
+
+void quantize_row_q2_K(const float * restrict x, void * restrict y, int k);
+void quantize_row_q3_K(const float * restrict x, void * restrict y, int k);
+void quantize_row_q4_K(const float * restrict x, void * restrict y, int k);
+void quantize_row_q5_K(const float * restrict x, void * restrict y, int k);
+void quantize_row_q6_K(const float * restrict x, void * restrict y, int k);
+void quantize_row_q8_K(const float * restrict x, void * restrict y, int k);
+
+// Dequantization
+void dequantize_row_q2_K(const block_q2_K * restrict x, float * restrict y, int k);
+void dequantize_row_q3_K(const block_q3_K * restrict x, float * restrict y, int k);
+void dequantize_row_q4_K(const block_q4_K * restrict x, float * restrict y, int k);
+void dequantize_row_q5_K(const block_q5_K * restrict x, float * restrict y, int k);
+void dequantize_row_q6_K(const block_q6_K * restrict x, float * restrict y, int k);
+void dequantize_row_q8_K(const block_q8_K * restrict x, float * restrict y, int k);
+
+// Dot product
+void ggml_vec_dot_q2_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
+void ggml_vec_dot_q3_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
+void ggml_vec_dot_q4_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
+void ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
+void ggml_vec_dot_q6_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
+
+// Quantization with histogram collection
+size_t ggml_quantize_q2_K(const float * src, void * dst, int n, int k, int64_t * hist);
+size_t ggml_quantize_q3_K(const float * src, void * dst, int n, int k, int64_t * hist);
+size_t ggml_quantize_q4_K(const float * src, void * dst, int n, int k, int64_t * hist);
+size_t ggml_quantize_q5_K(const float * src, void * dst, int n, int k, int64_t * hist);
+size_t ggml_quantize_q6_K(const float * src, void * dst, int n, int k, int64_t * hist);
+

llama-util.h

@@ -405,13 +405,29 @@ struct llama_buffer {
     llama_buffer() = default;
 
     void resize(size_t len) {
+#ifdef GGML_USE_METAL
+        free(addr);
+        int result = posix_memalign((void **) &addr, getpagesize(), len);
+        if (result == 0) {
+            memset(addr, 0, len);
+        }
+        else {
+            addr = NULL;
+        }
+#else
         delete[] addr;
         addr = new uint8_t[len];
+#endif
         size = len;
     }
 
     ~llama_buffer() {
+#ifdef GGML_USE_METAL
+        free(addr);
+#else
         delete[] addr;
+#endif
+        addr = NULL;
     }
 
     // disable copy and move

llama.h

@@ -1,6 +1,13 @@
 #ifndef LLAMA_H
 #define LLAMA_H
 
+#include "ggml.h"
+#ifdef GGML_USE_CUBLAS
+#include "ggml-cuda.h"
+#define LLAMA_MAX_DEVICES GGML_CUDA_MAX_DEVICES
+#else
+#define LLAMA_MAX_DEVICES 1
+#endif // GGML_USE_CUBLAS
 #include <stddef.h>
 #include <stdint.h>
 #include <stdbool.h>
@@ -31,7 +38,7 @@
 #define LLAMA_SESSION_MAGIC          LLAMA_FILE_MAGIC_GGSN
 #define LLAMA_SESSION_VERSION        1
 
-#if defined(GGML_USE_CUBLAS) || defined(GGML_USE_CLBLAST)
+#if defined(GGML_USE_CUBLAS) || defined(GGML_USE_CLBLAST) || defined(GGML_USE_METAL)
 // Defined when llama.cpp is compiled with support for offloading model layers to GPU.
 #define LLAMA_SUPPORTS_GPU_OFFLOAD
 #endif
@@ -65,9 +72,13 @@ extern "C" {
     typedef void (*llama_progress_callback)(float progress, void *ctx);
 
     struct llama_context_params {
-        int n_ctx;        // text context
-        int n_gpu_layers; // number of layers to store in VRAM
-        int seed;         // RNG seed, -1 for random
+        int n_ctx;                             // text context
+        int n_batch;                           // prompt processing batch size
+        int n_gpu_layers;                      // number of layers to store in VRAM
+        int main_gpu;                          // the GPU that is used for scratch and small tensors
+        float tensor_split[LLAMA_MAX_DEVICES]; // how to split layers across multiple GPUs
+        bool low_vram;                         // if true, reduce VRAM usage at the cost of performance
+        int seed;                              // RNG seed, -1 for random
 
         bool f16_kv;     // use fp16 for KV cache
         bool logits_all; // the llama_eval() call computes all logits, not just the last one
@@ -94,9 +105,27 @@ extern "C" {
         LLAMA_FTYPE_MOSTLY_Q8_0          = 7, // except 1d tensors
         LLAMA_FTYPE_MOSTLY_Q5_0          = 8, // except 1d tensors
         LLAMA_FTYPE_MOSTLY_Q5_1          = 9, // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q2_K          = 10,// except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q3_K_S        = 11,// except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q3_K_M        = 12,// except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q3_K_L        = 13,// except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q4_K_S        = 14,// except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q4_K_M        = 15,// except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q5_K_S        = 16,// except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q5_K_M        = 17,// except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q6_K          = 18,// except 1d tensors
     };
 
+    // model quantization parameters
+    typedef struct llama_model_quantize_params {
+        int nthread;                 // number of threads to use for quantizing, if <=0 will use std::thread::hardware_concurrency()
+        enum llama_ftype   ftype;    // quantize to this llama_ftype
+        bool allow_requantize;       // allow quantizing non-f32/f16 tensors
+        bool quantize_output_tensor; // quantize output.weight
+    } llama_model_quantize_params;
+
     LLAMA_API struct llama_context_params llama_context_default_params();
+    LLAMA_API struct llama_model_quantize_params llama_model_quantize_default_params();
 
     LLAMA_API bool llama_mmap_supported();
     LLAMA_API bool llama_mlock_supported();
@@ -118,14 +147,11 @@ extern "C" {
     // Frees all allocated memory
     LLAMA_API void llama_free(struct llama_context * ctx);
 
-    // TODO: not great API - very likely to change
     // Returns 0 on success
-    // nthread - how many threads to use. If <=0, will use std::thread::hardware_concurrency(), else the number given
     LLAMA_API int llama_model_quantize(
             const char * fname_inp,
             const char * fname_out,
-      enum llama_ftype   ftype,
-            int          nthread);
+            const llama_model_quantize_params * params);
 
     // Apply a LoRA adapter to a loaded model
     // path_base_model is the path to a higher quality model to use as a base for
@@ -173,6 +199,12 @@ extern "C" {
                              int   n_past,
                              int   n_threads);
 
+    // Export a static computation graph for context of 511 and batch size of 1
+    // NOTE: since this functionality is mostly for debugging and demonstration purposes, we hardcode these
+    //       parameters here to keep things simple
+    // IMPORTANT: do not use for anything else other than debugging and testing!
+    LLAMA_API int llama_eval_export(struct llama_context * ctx, const char * fname);
+
     // Convert the provided text into tokens.
     // The tokens pointer must be large enough to hold the resulting tokens.
     // Returns the number of tokens on success, no more than n_max_tokens
@@ -189,6 +221,14 @@ extern "C" {
     LLAMA_API int llama_n_ctx  (const struct llama_context * ctx);
     LLAMA_API int llama_n_embd (const struct llama_context * ctx);
 
+    // Get the vocabulary as output parameters.
+    // Returns number of results.
+    LLAMA_API int llama_get_vocab(
+            const struct llama_context * ctx,
+                          const char * * strings,
+                                 float * scores,
+                                   int   capacity);
+
     // Token logits obtained from the last call to llama_eval()
     // The logits for the last token are stored in the last row
     // Can be mutated in order to change the probabilities of the next token

spm-headers/ggml.h

@@ -0,0 +1 @@
+../ggml.h

tests/test-grad0.c

@@ -5,7 +5,7 @@
 #include <stdlib.h>
 #include <assert.h>
 
-#define MAX_NARGS 2
+#define MAX_NARGS 3
 
 #undef MIN
 #undef MAX
@@ -1090,6 +1090,25 @@ int main(int argc, const char ** argv) {
             }
         }
 
+        // cross_entropy_loss
+        {
+            const int nargs = 1;
+
+            int64_t ne2[4];
+            get_random_dims(ne2, 4);
+
+            for (int ndims = 1; ndims <= 3; ++ndims) {
+                x[0] = get_random_tensor(ctx0, ndims, ne2, -1.0f, 1.0f);
+                x[1] = get_random_tensor(ctx0, ndims, ne2, 0.0f, 1.0f);
+                ggml_set_param(ctx0, x[0]);
+
+                struct ggml_tensor * f = ggml_sum(ctx0, ggml_cross_entropy_loss(ctx0, x[0], x[1]));
+
+                check_gradient("cross_entropy_loss", ctx0, x, f, ndims, nargs, 1e-1f, 1e-2f, INFINITY);
+                // finite differences regularly fails!
+            }
+        }
+
         // rope
         {
             const int nargs = 1;
@@ -1124,6 +1143,45 @@ int main(int argc, const char ** argv) {
             }
         }
 
+        // flash_attn
+        {
+            const int nargs = 3;
+
+            int64_t ne2[4];
+
+            get_random_dims(ne2, 4);
+            int64_t D = ne2[0];
+            int64_t N = ne2[1];
+            int64_t M = ne2[2] + N;
+            int64_t B = ne2[3];
+
+            for (int masked = 0; masked <= 1; ++masked) {
+                for (int ndims = 2; ndims <= 4; ++ndims) {
+                    int64_t neq[4] = { D, N, B, ne[3] };
+                    int64_t nek[4] = { D, M, B, ne[3] };
+                    int64_t nev[4] = { M, D, B, ne[3] };
+                    if (ndims == 2) {
+                        neq[2] = 1; neq[3] = 1;
+                        nek[2] = 1; nek[3] = 1;
+                        nev[2] = 1; nev[3] = 1;
+                    } else if (ndims == 3) {
+                        neq[3] = 1;
+                        nek[3] = 1;
+                        nev[3] = 1;
+                    }
+                    x[0] = get_random_tensor(ctx0, ndims, neq, -0.1250f, 0.1250f);
+                    x[1] = get_random_tensor(ctx0, ndims, nek, -0.1250f, 0.1250f);
+                    x[2] = get_random_tensor(ctx0, ndims, nev, -0.1250f, 0.1250f);
+                    ggml_set_param(ctx0, x[0]);
+                    ggml_set_param(ctx0, x[1]);
+                    ggml_set_param(ctx0, x[2]);
+
+                    struct ggml_tensor * f = ggml_sum(ctx0, ggml_flash_attn(ctx0, x[0], x[1], x[2], (masked == 0)));
+
+                    check_gradient("flash_attn", ctx0, x, f, ndims, nargs, 1.5e-4f, INFINITY, 3.5f);
+                }
+            }
+        }
         ggml_free(ctx0);
     }
 

tests/test-quantize-fns.cpp

@@ -12,6 +12,8 @@
 
 const float MAX_QUANTIZATION_REFERENCE_ERROR = 0.0001;
 const float MAX_QUANTIZATION_TOTAL_ERROR = 0.002;
+const float MAX_QUANTIZATION_TOTAL_ERROR_2BITS = 0.0075;
+const float MAX_QUANTIZATION_TOTAL_ERROR_3BITS = 0.0040;
 const float MAX_DOT_PRODUCT_ERROR = 0.02;
 
 const char* RESULT_STR[] = {"ok", "FAILED"};
@@ -122,7 +124,10 @@ int main(int argc, char * argv[]) {
 
         if (qfns.quantize_row_q && qfns.dequantize_row_q) {
             const float total_error = total_quantization_error(qfns, test_size, test_data.data());
-            failed = !(total_error < MAX_QUANTIZATION_TOTAL_ERROR);
+            const float max_quantization_error =
+                type == GGML_TYPE_Q2_K ? MAX_QUANTIZATION_TOTAL_ERROR_2BITS :
+                type == GGML_TYPE_Q3_K ? MAX_QUANTIZATION_TOTAL_ERROR_3BITS : MAX_QUANTIZATION_TOTAL_ERROR;
+            failed = !(total_error < max_quantization_error);
             num_failed += failed;
             if (failed || verbose) {
                 printf("%5s absolute quantization error:    %s (%f)\n", ggml_type_name(type), RESULT_STR[failed], total_error);