Sync with upstream

.flake8

@@ -0,0 +1,2 @@
+[flake8]
+max-line-length = 125

.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 }}

.gitignore

@@ -22,6 +22,7 @@ build-metal/
 build-no-accel/
 build-sanitize-addr/
 build-sanitize-thread/
+out/
 
 models/*
 *.bin
@@ -32,14 +33,18 @@ models/*
 /result
 /perplexity
 /embedding
+/train-text-from-scratch
+/simple
 /benchmark-matmult
 /vdot
+/server
 /Pipfile
 /libllama.so
 
 build-info.h
 arm_neon.h
 compile_commands.json
+CMakeSettings.json
 
 __pycache__
 

.pre-commit-config.yaml

@@ -0,0 +1,15 @@
+# See https://pre-commit.com for more information
+# See https://pre-commit.com/hooks.html for more hooks
+exclude: prompts/.*.txt
+repos:
+- repo: https://github.com/pre-commit/pre-commit-hooks
+  rev: v3.2.0
+  hooks:
+  - id: trailing-whitespace
+  - id: end-of-file-fixer
+  - id: check-yaml
+  - id: check-added-large-files
+- repo: https://github.com/PyCQA/flake8
+  rev: 6.0.0
+  hooks:
+  -   id: flake8

CMakeLists.txt

@@ -70,6 +70,8 @@ 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_CUDA_DMMV_F16                   "llama: use 16 bit floats for dmmv CUDA kernels"   OFF)
+set(LLAMA_CUDA_KQUANTS_ITER "2" CACHE STRING "llama: iters./thread per block for Q2_K/Q6_K")
 option(LLAMA_CLBLAST                         "llama: use CLBlast"                               OFF)
 option(LLAMA_METAL                           "llama: use Metal"                                 OFF)
 option(LLAMA_K_QUANTS                        "llama: use k-quants"                              ON)
@@ -158,17 +160,64 @@ if (LLAMA_BLAS)
     if ($(CMAKE_VERSION) VERSION_GREATER_EQUAL 3.22)
         set(BLA_SIZEOF_INTEGER 8)
     endif()
+
     set(BLA_VENDOR ${LLAMA_BLAS_VENDOR})
     find_package(BLAS)
+
     if (BLAS_FOUND)
         message(STATUS "BLAS found, Libraries: ${BLAS_LIBRARIES}")
 
+        if ("${BLAS_INCLUDE_DIRS}" STREQUAL "")
+            # BLAS_INCLUDE_DIRS is missing in FindBLAS.cmake.
+            # see https://gitlab.kitware.com/cmake/cmake/-/issues/20268
+            find_package(PkgConfig REQUIRED)
+            if (${LLAMA_BLAS_VENDOR} MATCHES "Generic")
+                pkg_check_modules(DepBLAS REQUIRED blas)
+            elseif (${LLAMA_BLAS_VENDOR} MATCHES "OpenBLAS")
+                pkg_check_modules(DepBLAS REQUIRED openblas)
+            elseif (${LLAMA_BLAS_VENDOR} MATCHES "FLAME")
+                pkg_check_modules(DepBLAS REQUIRED blis)
+            elseif (${LLAMA_BLAS_VENDOR} MATCHES "ATLAS")
+                pkg_check_modules(DepBLAS REQUIRED blas-atlas)
+            elseif (${LLAMA_BLAS_VENDOR} MATCHES "FlexiBLAS")
+                pkg_check_modules(DepBLAS REQUIRED flexiblas_api)
+            elseif (${LLAMA_BLAS_VENDOR} MATCHES "Intel")
+                # all Intel* libraries share the same include path
+                pkg_check_modules(DepBLAS REQUIRED mkl-sdl)
+            elseif (${LLAMA_BLAS_VENDOR} MATCHES "NVHPC")
+                # this doesn't provide pkg-config
+                # suggest to assign BLAS_INCLUDE_DIRS on your own
+                if ("${NVHPC_VERSION}" STREQUAL "")
+                    message(WARNING "Better to set NVHPC_VERSION")
+                else()
+                    set(DepBLAS_FOUND ON)
+                    set(DepBLAS_INCLUDE_DIRS "/opt/nvidia/hpc_sdk/${CMAKE_SYSTEM_NAME}_${CMAKE_SYSTEM_PROCESSOR}/${NVHPC_VERSION}/math_libs/include")
+                endif()
+            endif()
+            if (DepBLAS_FOUND)
+                set(BLAS_INCLUDE_DIRS ${DepBLAS_INCLUDE_DIRS})
+            else()
+                message(WARNING "BLAS_INCLUDE_DIRS neither been provided nor been automatically"
+                " detected by pkgconfig, trying to find cblas.h from possible paths...")
+                find_path(BLAS_INCLUDE_DIRS
+                    NAMES cblas.h
+                    HINTS
+                        /usr/include
+                        /usr/local/include
+                        /usr/include/openblas
+                        /opt/homebrew/opt/openblas/include
+                        /usr/local/opt/openblas/include
+                        /usr/include/x86_64-linux-gnu/openblas/include
+                )
+            endif()
+        endif()
+
+        message(STATUS "BLAS found, Includes: ${BLAS_INCLUDE_DIRS}")
         add_compile_options(${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"
@@ -190,6 +239,10 @@ if (LLAMA_CUBLAS)
         add_compile_definitions(GGML_USE_CUBLAS)
         add_compile_definitions(GGML_CUDA_DMMV_X=${LLAMA_CUDA_DMMV_X})
         add_compile_definitions(GGML_CUDA_DMMV_Y=${LLAMA_CUDA_DMMV_Y})
+        if (LLAMA_CUDA_DMMV_F16)
+            add_compile_definitions(GGML_CUDA_DMMV_F16)
+        endif()
+        add_compile_definitions(K_QUANTS_PER_ITERATION=${LLAMA_CUDA_KQUANTS_ITER})
 
         if (LLAMA_STATIC)
             set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} CUDA::cudart_static CUDA::cublas_static CUDA::cublasLt_static)
@@ -408,12 +461,14 @@ add_library(ggml OBJECT
             ${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})
 
+add_library(ggml_static STATIC $<TARGET_OBJECTS:ggml>)
 if (BUILD_SHARED_LIBS)
     set_target_properties(ggml PROPERTIES POSITION_INDEPENDENT_CODE ON)
+    add_library(ggml_shared SHARED $<TARGET_OBJECTS:ggml>)
 endif()
 
 add_library(llama
@@ -439,9 +494,13 @@ endif()
 
 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_ARCHITECTURES "native")
     set_property(TARGET ggml  PROPERTY CUDA_SELECT_NVCC_ARCH_FLAGS "Auto")
-    set_property(TARGET llama PROPERTY CUDA_ARCHITECTURES OFF)
+
+    set_property(TARGET ggml_static PROPERTY CUDA_ARCHITECTURES "native")
+    set_property(TARGET ggml_static PROPERTY CUDA_SELECT_NVCC_ARCH_FLAGS "Auto")
+
+    set_property(TARGET llama PROPERTY CUDA_ARCHITECTURES "native")
 endif()
 
 

Makefile

@@ -1,8 +1,10 @@
 # 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 simple
 
 ifdef LLAMA_BUILD_SERVER
 	BUILD_TARGETS += server
+	LLAMA_SERVER_VERBOSE ?= 1
+server: private CXXFLAGS += -DSERVER_VERBOSE=$(LLAMA_SERVER_VERBOSE)
 endif
 
 default: $(BUILD_TARGETS)
@@ -127,6 +129,7 @@ endif
 
 ifndef LLAMA_NO_K_QUANTS
 	CFLAGS   += -DGGML_USE_K_QUANTS
+	CXXFLAGS += -DGGML_USE_K_QUANTS
 	OBJS     += k_quants.o
 endif
 
@@ -141,11 +144,7 @@ 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),)
-		LDFLAGS += -lopenblas -lcblas
-	else
 	LDFLAGS += -lopenblas
-	endif
 endif # LLAMA_OPENBLAS
 
 ifdef LLAMA_BLIS
@@ -170,6 +169,14 @@ ifdef LLAMA_CUDA_DMMV_Y
 else
 	NVCCFLAGS += -DGGML_CUDA_DMMV_Y=1
 endif # LLAMA_CUDA_DMMV_Y
+ifdef LLAMA_CUDA_DMMV_F16
+	NVCCFLAGS += -DGGML_CUDA_DMMV_F16
+endif # LLAMA_CUDA_DMMV_F16
+ifdef LLAMA_CUDA_KQUANTS_ITER
+	NVCCFLAGS += -DK_QUANTS_PER_ITERATION=$(LLAMA_CUDA_KQUANTS_ITER)
+else
+	NVCCFLAGS += -DK_QUANTS_PER_ITERATION=2
+endif
 ggml-cuda.o: ggml-cuda.cu ggml-cuda.h
 	$(NVCC) $(NVCCFLAGS) $(CXXFLAGS) -Wno-pedantic -c $< -o $@
 endif # LLAMA_CUBLAS
@@ -248,7 +255,7 @@ $(info )
 ggml.o: ggml.c ggml.h ggml-cuda.h
 	$(CC)  $(CFLAGS)   -c $< -o $@
 
-llama.o: llama.cpp ggml.h ggml-cuda.h llama.h llama-util.h
+llama.o: llama.cpp ggml.h ggml-cuda.h ggml-metal.h llama.h llama-util.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@
 
 common.o: examples/common.cpp examples/common.h
@@ -258,7 +265,7 @@ 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
@@ -270,6 +277,9 @@ main: examples/main/main.cpp                                  build-info.h ggml.
 	@echo '====  Run ./main -h for help.  ===='
 	@echo
 
+simple: examples/simple/simple.cpp                            build-info.h ggml.o llama.o common.o $(OBJS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+
 quantize: examples/quantize/quantize.cpp                      build-info.h ggml.o llama.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
@@ -288,6 +298,9 @@ 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

@@ -336,9 +336,15 @@ Building the program with BLAS support may lead to some performance improvements
     cmake .. -DLLAMA_CUBLAS=ON
     cmake --build . --config Release
     ```
-  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.
 
-  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.
+  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. The following compilation options are also available to tweak performance:
+
+  | Option                  | Legal values           | Default | Description |
+  |-------------------------|------------------------|---------|-------------|
+  | LLAMA_CUDA_DMMV_X       | Positive integer >= 32 |      32 | Number of values in x direction processed by the CUDA dequantization + matrix vector multiplication kernel per iteration. Increasing this value can improve performance on fast GPUs. Power of 2 heavily recommended. Does not affect k-quants. |
+  | LLAMA_CUDA_DMMV_Y       | Positive integer       |       1 | Block size in y direction for the CUDA dequantization + mul mat vec kernels. Increasing this value can improve performance on fast GPUs. Power of 2 recommended. Does not affect k-quants. |
+  | LLAMA_CUDA_DMMV_F16     | Boolean                |   false | If enabled, use half-precision floating point arithmetic for the CUDA dequantization + mul mat vec kernels. Can improve performance on relatively recent GPUs. |
+  | LLAMA_CUDA_KQUANTS_ITER | 1 or 2                 |       2 | Number of values processed per iteration and per CUDA thread for Q2_K and Q6_K quantization formats. Setting this value 2 1 can improve performance for slow GPUs. |
 
 - #### CLBlast
 
@@ -616,8 +622,14 @@ And after 4.45 hours, you will have the final perplexity.
 
 ### Android
 
+#### Building the Project using Android NDK
 You can easily run `llama.cpp` on Android device with [termux](https://termux.dev/).
-First, obtain the [Android NDK](https://developer.android.com/ndk) and then build with CMake:
+
+First, install the essential packages for termux:
+```
+pkg install clang wget git cmake
+```
+Second, obtain the [Android NDK](https://developer.android.com/ndk) and then build with CMake:
 ```
 $ mkdir build-android
 $ cd build-android
@@ -630,6 +642,46 @@ Finally, copy the `llama` binary and the model files to your device storage. Her
 
 https://user-images.githubusercontent.com/271616/225014776-1d567049-ad71-4ef2-b050-55b0b3b9274c.mp4
 
+#### Building the Project using Termux (F-Droid)
+Termux from F-Droid offers an alternative route to execute the project on an Android device. This method empowers you to construct the project right from within the terminal, negating the requirement for a rooted device or SD Card.
+
+Outlined below are the directives for installing the project using OpenBLAS and CLBlast. This combination is specifically designed to deliver peak performance on recent devices that feature a GPU.
+
+If you opt to utilize OpenBLAS, you'll need to install the corresponding package.
+```
+apt install libopenblas
+```
+
+Subsequently, if you decide to incorporate CLBlast, you'll first need to install the requisite OpenCL packages:
+```
+apt install ocl-icd opencl-headers opencl-clhpp clinfo
+```
+
+In order to compile CLBlast, you'll need to first clone the respective Git repository, which can be found at this URL: https://github.com/CNugteren/CLBlast. Alongside this, clone this repository into your home directory. Once this is done, navigate to the CLBlast folder and execute the commands detailed below:
+```
+cmake .
+make
+cp libclblast.so* $PREFIX/lib
+cp ./include/clblast.h ../llama.cpp
+```
+
+Following the previous steps, navigate to the LlamaCpp directory. To compile it with OpenBLAS and CLBlast, execute the command provided below:
+```
+cp /data/data/com.termux/files/usr/include/openblas/cblas.h .
+cp /data/data/com.termux/files/usr/include/openblas/openblas_config.h .
+make LLAMA_CLBLAST=1 //(sometimes you need to run this command twice)
+```
+
+Upon completion of the aforementioned steps, you will have successfully compiled the project. To run it using CLBlast, a slight adjustment is required: a command must be issued to direct the operations towards your device's physical GPU, rather than the virtual one. The necessary command is detailed below:
+```
+GGML_OPENCL_PLATFORM=0
+GGML_OPENCL_DEVICE=0
+export LD_LIBRARY_PATH=/system/vendor/lib64:$LD_LIBRARY_PATH
+./main (...)
+```
+
+For easy and swift re-execution, consider documenting this final part in a .sh script file. This will enable you to rerun the process with minimal hassle.
+
 ### Docker
 
 #### Prerequisites

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.py

@@ -512,7 +512,11 @@ class LazyTensor:
             if not isinstance(self.data_type, QuantizedDataType):
                 raise Exception(f"Can't turn an unquantized tensor into a quantized type ({data_type})")
             if self.data_type.have_g_idx:
-                sys.stderr.write("Error: Input uses the newer GPTQ-for-LLaMa format (using g_idx), which is not yet natively supported by GGML.  For now you can still convert this model by passing `--outtype f16` to dequantize, but that will result in a much larger output file for no quality benefit.\n")
+                sys.stderr.write(
+                    "Error: Input uses the newer GPTQ-for-LLaMa format (using g_idx), "
+                    "which is not yet natively supported by GGML. "
+                    "For now you can still convert this model by passing `--outtype f16` to dequantize, "
+                    "but that will result in a much larger output file for no quality benefit.\n")
                 sys.exit(1)
             assert not data_type.have_g_idx and self.data_type.have_addends and data_type.have_addends
 
@@ -695,7 +699,8 @@ class LazyUnpickler(pickle.Unpickler):
         return LazyStorage(load=load, kind=pid[1], description=description)
 
     # @staticmethod
-    def lazy_rebuild_tensor_v2(storage: Any, storage_offset: Any, size: Any, stride: Any,  # pyright: ignore[reportSelfClsParameterName]
+    def lazy_rebuild_tensor_v2(storage: Any, storage_offset: Any, size: Any, stride: Any,
+                               # pyright: ignore[reportSelfClsParameterName]
                                requires_grad: Any, backward_hooks: Any, metadata: Any = None) -> LazyTensor:
         assert isinstance(storage, LazyStorage)
 
@@ -1054,7 +1059,7 @@ def load_some_model(path: Path) -> ModelPlus:
         files = list(path.glob("model-00001-of-*.safetensors"))
         if not files:
             # Try the PyTorch patterns too, with lower priority
-            globs = ["consolidated.00.pth", "pytorch_model-00001-of-*.bin", "*.pt", "pytorch_model.bin" ]
+            globs = ["consolidated.00.pth", "pytorch_model-00001-of-*.bin", "*.pt", "pytorch_model.bin"]
             files = [file for glob in globs for file in path.glob(glob)]
         if not files:
             # Try GGML too, but with lower priority, since if both a non-GGML
@@ -1094,7 +1099,9 @@ def load_vocab(path: Path) -> SentencePieceVocab:
         elif path3.exists():
             path = path3
         else:
-            raise FileNotFoundError(f"Could not find tokenizer.model in {path} or its parent; if it's in another directory, pass the directory as --vocab-dir")
+            raise FileNotFoundError(
+                f"Could not find tokenizer.model in {path} or its parent; "
+                "if it's in another directory, pass the directory as --vocab-dir")
     added_tokens_path = path.parent / "added_tokens.json"
     print(f"Loading vocab file {path}")
     return SentencePieceVocab(path, added_tokens_path if added_tokens_path.exists() else None)
@@ -1110,7 +1117,9 @@ def default_outfile(model_paths: List[Path], params: Params) -> Path:
     }[params.file_type]
     ret = model_paths[0].parent / f"ggml-model-{namestr}.bin"
     if ret in model_paths:
-        sys.stderr.write(f"Error: Default output path ({ret}) would overwrite the input.  Please explicitly specify a path using --outfile.\n")
+        sys.stderr.write(
+            f"Error: Default output path ({ret}) would overwrite the input. "
+            "Please explicitly specify a path using --outfile.\n")
         sys.exit(1)
     return ret
 
@@ -1131,7 +1140,8 @@ def main(args_in: Optional[List[str]] = None) -> None:
     parser.add_argument("--outtype", choices=["f32", "f16", "q4_1", "q4_0"], help="output format (default: based on input)")
     parser.add_argument("--vocab-dir", type=Path, help="directory containing tokenizer.model, if separate from model file")
     parser.add_argument("--outfile", type=Path, help="path to write to; default: based on input")
-    parser.add_argument("model", type=Path, help="directory containing model file, or model file itself (*.pth, *.pt, *.bin)")
+    parser.add_argument("model", type=Path,
+                        help="directory containing model file, or model file itself (*.pth, *.pt, *.bin)")
     args = parser.parse_args(args_in)
 
     vocab: Vocab

examples/CMakeLists.txt

@@ -37,6 +37,8 @@ else()
     add_subdirectory(save-load-state)
     add_subdirectory(benchmark)
     add_subdirectory(baby-llama)
+    add_subdirectory(train-text-from-scratch)
+    add_subdirectory(simple)
     if (LLAMA_METAL)
         add_subdirectory(metal)
     endif()

examples/baby-llama/baby-llama.cpp

@@ -4,6 +4,10 @@
 #include <random>
 #include <cstring>
 
+#if defined(_MSC_VER)
+#pragma warning(disable: 4244 4267) // possible loss of data
+#endif
+
 float frand() {
     return (float)rand()/(float)RAND_MAX;
 }
@@ -79,34 +83,39 @@ 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 +157,8 @@ struct llama_hparams_lora {
     uint32_t n_rot   = 64;
     uint32_t n_lora  = 64;
 
-    bool operator!=(const llama_hparams & other) const {
+    bool operator!=(const llama_hparams_lora & other) const {
-        return memcmp(this, &other, sizeof(llama_hparams));
+        return memcmp(this, &other, sizeof(llama_hparams_lora)) != 0;
     }
 };
 
@@ -1465,7 +1474,7 @@ struct ggml_tensor * square_error_loss(struct ggml_context * ctx, struct ggml_te
 }
 
 struct ggml_tensor * cross_entropy_loss(struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * b) {
-    const float eps = 1e-3;
+    const float eps = 1e-3f;
     return
         ggml_sum(ctx,
             ggml_neg(ctx,

examples/benchmark/benchmark-matmult.cpp

@@ -16,6 +16,10 @@
 #include <iterator>
 #include <algorithm>
 
+#if defined(_MSC_VER)
+#pragma warning(disable: 4244 4267) // possible loss of data
+#endif
+
 float tensor_sum_elements(const ggml_tensor * tensor) {
     float sum = 0;
     if (tensor->type==GGML_TYPE_F32) {
@@ -29,9 +33,9 @@ float tensor_sum_elements(const ggml_tensor * tensor) {
 }
 
 void tensor_dump(const ggml_tensor * tensor, const char * name) {
-    printf("%15s: type = %i (%5s) ne = %5d x %5d x %5d, nb = (%5li, %5li, %5li) - ", name,
+    printf("%15s: type = %i (%5s) ne = %5" PRIi64 " x %5" PRIi64 " x %5" PRIi64 ", nb = (%5zi, %5zi, %5zi) - ", name,
         tensor->type, ggml_type_name(tensor->type),
-        (int) tensor->ne[0], (int) tensor->ne[1], (int) tensor->ne[2], tensor->nb[0], tensor->nb[1], tensor->nb[2]);
+        tensor->ne[0], tensor->ne[1], tensor->ne[2], tensor->nb[0], tensor->nb[1], tensor->nb[2]);
     float sum = tensor_sum_elements(tensor);
     printf("Sum of tensor %s is %6.2f\n", name, sum);
 }
@@ -120,7 +124,7 @@ int main(int argc, char ** argv)  {
     ctx_size += sizex*sizey*ggml_type_sizef(GGML_TYPE_F32); // BLAS
     ctx_size += 1024*1024*16;
 
-    printf("Allocating Memory of size %li bytes, %li MB\n",ctx_size, (ctx_size/1024/1024));
+    printf("Allocating Memory of size %zi bytes, %zi MB\n",ctx_size, (ctx_size/1024/1024));
 
     struct ggml_init_params params = {
         /*.mem_size   =*/ ctx_size,

examples/chat-vicuna.sh

@@ -0,0 +1,41 @@
+#!/bin/bash
+
+set -e
+
+cd "$(dirname "$0")/.." || exit
+
+MODEL="${MODEL:-./models/ggml-vic13b-uncensored-q5_0.bin}"
+PROMPT_TEMPLATE=${PROMPT_TEMPLATE:-./prompts/chat.txt}
+USER_NAME="### Human"
+AI_NAME="### Assistant"
+
+# Adjust to the number of CPU cores you want to use.
+N_THREAD="${N_THREAD:-8}"
+# Number of tokens to predict (made it larger than default because we want a long interaction)
+N_PREDICTS="${N_PREDICTS:-2048}"
+
+# Note: you can also override the generation options by specifying them on the command line:
+# For example, override the context size by doing: ./chatLLaMa --ctx_size 1024
+GEN_OPTIONS="${GEN_OPTIONS:---ctx_size 2048 --temp 0.7 --top_k 40 --top_p 0.5 --repeat_last_n 256 --batch_size 1024 --repeat_penalty 1.17647}"
+
+DATE_TIME=$(date +%H:%M)
+DATE_YEAR=$(date +%Y)
+
+PROMPT_FILE=$(mktemp -t llamacpp_prompt.XXXXXXX.txt)
+
+sed -e "s/\[\[USER_NAME\]\]/$USER_NAME/g" \
+    -e "s/\[\[AI_NAME\]\]/$AI_NAME/g" \
+    -e "s/\[\[DATE_TIME\]\]/$DATE_TIME/g" \
+    -e "s/\[\[DATE_YEAR\]\]/$DATE_YEAR/g" \
+     $PROMPT_TEMPLATE > $PROMPT_FILE
+
+# shellcheck disable=SC2086 # Intended splitting of GEN_OPTIONS
+./bin/main $GEN_OPTIONS \
+  --model "$MODEL" \
+  --threads "$N_THREAD" \
+  --n_predict "$N_PREDICTS" \
+  --color --interactive \
+  --file ${PROMPT_FILE} \
+  --reverse-prompt "### Human:" \
+  --in-prefix ' ' \
+  "$@"

examples/common.cpp

@@ -28,6 +28,10 @@
 #include <wchar.h>
 #endif
 
+#if defined(_MSC_VER)
+#pragma warning(disable: 4244 4267) // possible loss of data
+#endif
+
 int32_t get_num_physical_cores() {
 #ifdef __linux__
     // enumerate the set of thread siblings, num entries is num cores
@@ -102,9 +106,6 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
         }
 
         if (arg == "-s" || arg == "--seed") {
-#if defined(GGML_USE_CUBLAS)
-            fprintf(stderr, "WARNING: when using cuBLAS generation results are NOT guaranteed to be reproducible.\n");
-#endif
             if (++i >= argc) {
                 invalid_param = true;
                 break;
@@ -331,6 +332,12 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
             }
 #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;
@@ -367,7 +374,7 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
                 } else {
                     throw std::exception();
                 }
-            } catch (const std::exception &e) {
+            } catch (const std::exception&) {
                 invalid_param = true;
                 break;
             }
@@ -406,6 +413,14 @@ 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);
     }
@@ -479,6 +494,7 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     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");
@@ -528,6 +544,7 @@ struct llama_context * llama_init_from_gpt_params(const gpt_params & params) {
     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;
@@ -632,6 +649,9 @@ 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

@@ -30,6 +30,7 @@ struct gpt_params {
     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
@@ -112,7 +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/embedding/embedding.cpp

@@ -4,6 +4,10 @@
 
 #include <ctime>
 
+#if defined(_MSC_VER)
+#pragma warning(disable: 4244 4267) // possible loss of data
+#endif
+
 int main(int argc, char ** argv) {
     gpt_params params;
 

examples/jeopardy/graph.py

@@ -1,5 +1,5 @@
 import matplotlib.pyplot as plt
-import sys, os
+import os
 import csv
 
 labels = []
@@ -8,6 +8,7 @@ numEntries = 1
 
 rows = []
 
+
 def bar_chart(numbers, labels, pos):
     plt.bar(pos, numbers, color='blue')
     plt.xticks(ticks=pos, labels=labels)
@@ -16,6 +17,7 @@ def bar_chart(numbers, labels, pos):
     plt.ylabel("Questions Correct")
     plt.show()
 
+
 def calculatecorrect():
     directory = os.fsencode("./examples/jeopardy/results/")
     csv_reader = csv.reader(open("./examples/jeopardy/qasheet.csv", 'rt'), delimiter=',')
@@ -38,14 +40,13 @@ def calculatecorrect():
                     print(line)
                 else:
                     print("Correct answer: " + rows[i][2] + "\n")
-                    i+=1
+                    i += 1
                     print("Did the AI get the question right? (y/n)")
                     if input() == "y":
                         totalcorrect += 1
             numbers.append(totalcorrect)
 
 
-
 if __name__ == '__main__':
     calculatecorrect()
     pos = list(range(numEntries))

examples/main/README.md

@@ -288,5 +288,6 @@ These options provide extra functionality and customization when running the LLa
 -   `-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

@@ -23,11 +23,17 @@
 #include <unistd.h>
 #elif defined (_WIN32)
 #define WIN32_LEAN_AND_MEAN
+#ifndef NOMINMAX
 #define NOMINMAX
+#endif
 #include <windows.h>
 #include <signal.h>
 #endif
 
+#if defined(_MSC_VER)
+#pragma warning(disable: 4244 4267) // possible loss of data
+#endif
+
 static console_state con_st;
 static llama_context ** g_ctx;
 
@@ -81,6 +87,9 @@ 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);
@@ -328,9 +337,29 @@ 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 %zu 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)

examples/metal/metal.cpp

@@ -40,8 +40,10 @@ int main(int argc, char ** argv) {
     // 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));
+    const size_t max_size_data = ggml_get_max_tensor_size(ctx_data);
-    ggml_metal_add_buffer(ctx_metal, "eval", ggml_get_mem_buffer(ctx_eval), ggml_get_mem_size(ctx_eval));
+    const size_t max_size_eval = ggml_get_max_tensor_size(ctx_eval);
+    ggml_metal_add_buffer(ctx_metal, "data", ggml_get_mem_buffer(ctx_data), ggml_get_mem_size(ctx_data), max_size_data);
+    ggml_metal_add_buffer(ctx_metal, "eval", ggml_get_mem_buffer(ctx_eval), ggml_get_mem_size(ctx_eval), max_size_eval);
 
     // main
     {

examples/perplexity/perplexity.cpp

@@ -5,6 +5,10 @@
 #include <cmath>
 #include <ctime>
 
+#if defined(_MSC_VER)
+#pragma warning(disable: 4244 4267) // possible loss of data
+#endif
+
 std::vector<float> softmax(const std::vector<float>& logits) {
     std::vector<float> probs(logits.size());
     float max_logit = logits[0];

examples/quantize-stats/quantize-stats.cpp

@@ -19,6 +19,10 @@
 #include <thread>
 #include <mutex>
 
+#if defined(_MSC_VER)
+#pragma warning(disable: 4244 4267) // possible loss of data
+#endif
+
 struct quantize_stats_params {
     std::string model = "models/7B/ggml-model-f16.bin";
     bool verbose = false;

examples/quantize/quantize.cpp

@@ -4,43 +4,135 @@
 
 #include <cstdio>
 #include <cstring>
-#include <map>
+#include <vector>
 #include <string>
 
-static const std::map<std::string, llama_ftype> LLAMA_FTYPE_MAP = {
+struct quant_option {
-  {"q4_0",   LLAMA_FTYPE_MOSTLY_Q4_0},
+    std::string name;
-  {"q4_1",   LLAMA_FTYPE_MOSTLY_Q4_1},
+    llama_ftype ftype;
-  {"q5_0",   LLAMA_FTYPE_MOSTLY_Q5_0},
+    std::string desc;
-  {"q5_1",   LLAMA_FTYPE_MOSTLY_Q5_1},
-  {"q8_0",   LLAMA_FTYPE_MOSTLY_Q8_0},
-  {"q2_K",   LLAMA_FTYPE_MOSTLY_Q2_K},
-  {"q3_K",   LLAMA_FTYPE_MOSTLY_Q3_K_M},
-  {"q3_K_S", LLAMA_FTYPE_MOSTLY_Q3_K_S},
-  {"q3_K_M", LLAMA_FTYPE_MOSTLY_Q3_K_M},
-  {"q3_K_L", LLAMA_FTYPE_MOSTLY_Q3_K_L},
-  {"q4_K",   LLAMA_FTYPE_MOSTLY_Q4_K_M},
-  {"q4_K_S", LLAMA_FTYPE_MOSTLY_Q4_K_S},
-  {"q4_K_M", LLAMA_FTYPE_MOSTLY_Q4_K_M},
-  {"q5_K",   LLAMA_FTYPE_MOSTLY_Q5_K_M},
-  {"q5_K_S", LLAMA_FTYPE_MOSTLY_Q5_K_S},
-  {"q5_K_M", LLAMA_FTYPE_MOSTLY_Q5_K_M},
-  {"q6_K",   LLAMA_FTYPE_MOSTLY_Q6_K},
 };
 
-bool try_parse_ftype(const std::string & ftype_str, llama_ftype & ftype, std::string & ftype_str_out) {
+static const std::vector<struct quant_option> QUANT_OPTIONS = {
-    auto it = LLAMA_FTYPE_MAP.find(ftype_str);
+    {
-    if (it != LLAMA_FTYPE_MAP.end()) {
+        "Q4_0",
-        ftype = it->second;
+        LLAMA_FTYPE_MOSTLY_Q4_0,
-        ftype_str_out = it->first;
+        " 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++) {
+        for (auto & it : QUANT_OPTIONS) {
-            if (it->second == ftype_int) {
+            if (it.ftype == ftype_int) {
-                ftype = it->second;
+                ftype = it.ftype;
-                ftype_str_out = it->first;
+                ftype_str_out = it.name;
                 return true;
             }
         }
@@ -52,15 +144,15 @@ 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", 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, "Allowed quantization types:\n");
+    fprintf(stderr, "\nAllowed quantization types:\n");
-    for (auto it = LLAMA_FTYPE_MAP.begin(); it != LLAMA_FTYPE_MAP.end(); it++) {
+    for (auto & it : QUANT_OPTIONS) {
-        fprintf(stderr, "  type = \"%s\" or %d\n", it->first.c_str(), it->second);
+        printf("  %2d  or  %-6s : %s\n", it.ftype, it.name.c_str(), it.desc.c_str());
     }
     exit(1);
 }

examples/save-load-state/save-load-state.cpp

@@ -37,7 +37,7 @@ int main(int argc, char ** argv) {
     // init
     auto ctx = llama_init_from_file(params.model.c_str(), lparams);
     auto tokens = std::vector<llama_token>(params.n_ctx);
-    auto n_prompt_tokens = llama_tokenize(ctx, params.prompt.c_str(), tokens.data(), tokens.size(), true);
+    auto n_prompt_tokens = llama_tokenize(ctx, params.prompt.c_str(), tokens.data(), int(tokens.size()), true);
 
     if (n_prompt_tokens < 1) {
         fprintf(stderr, "%s : failed to tokenize prompt\n", __func__);

examples/server/CMakeLists.txt

@@ -1,6 +1,10 @@
 set(TARGET server)
+option(LLAMA_SERVER_VERBOSE "Build verbose logging option for Server" ON)
 include_directories(${CMAKE_CURRENT_SOURCE_DIR})
 add_executable(${TARGET} server.cpp json.hpp httplib.h)
+target_compile_definitions(${TARGET} PRIVATE
+    SERVER_VERBOSE=$<BOOL:${LLAMA_SERVER_VERBOSE}>
+)
 target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_11)
 if(TARGET BUILD_INFO)

examples/server/README.md

@@ -1,33 +1,74 @@
 # llama.cpp/example/server
 
-This example allow you to have a llama.cpp http server to interact from a web page or consume the API.
+This example demonstrates a simple HTTP API server to interact with llama.cpp.
 
-## Table of Contents
+Command line options:
 
-1. [Quick Start](#quick-start)
+-   `--threads N`, `-t N`: Set the number of threads to use during computation.
-2. [Node JS Test](#node-js-test)
+-   `-m FNAME`, `--model FNAME`: Specify the path to the LLaMA model file (e.g., `models/7B/ggml-model.bin`).
-3. [API Endpoints](#api-endpoints)
+-   `-m ALIAS`, `--alias ALIAS`: Set an alias for the model. The alias will be returned in API responses.
-4. [More examples](#more-examples)
+-   `-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.
-5. [Common Options](#common-options)
+-   `-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.
-6. [Performance Tuning and Memory Options](#performance-tuning-and-memory-options)
+-   `-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.
+-   `-b N`, `--batch-size N`: Set the batch size for prompt processing. Default: `512`.
+-   `--memory-f32`: Use 32-bit floats instead of 16-bit floats for memory key+value. Not recommended.
+-   `--mlock`: Lock the model in memory, preventing it from being swapped out when memory-mapped.
+-   `--no-mmap`: Do not memory-map the model. By default, models are mapped into memory, which allows the system to load only the necessary parts of the model as needed.
+-   `--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.
+-   `-to N`, `--timeout N`: Server read/write timeout in seconds. Default `600`.
+-   `--host`: Set the hostname or ip address to listen. Default `127.0.0.1`.
+-   `--port`: Set the port to listen. Default: `8080`.
+
+## Build
+
+Build llama.cpp with server from repository root with either make or CMake.
+
+- Using `make`:
+
+  ```bash
+  LLAMA_BUILD_SERVER=1 make
+  ```
+
+- Using `CMake`:
+
+  ```bash
+  mkdir build-server
+  cd build-server
+  cmake -DLLAMA_BUILD_SERVER=ON ..
+  cmake --build . --config Release
+  ```
 
 ## Quick Start
 
 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.):
+### Unix-based systems (Linux, macOS, etc.):
 
 ```bash
-./server -m models/7B/ggml-model.bin --ctx_size 2048
+./server -m models/7B/ggml-model.bin -c 2048
 ```
 
-#### Windows:
+### Windows:
 
 ```powershell
-server.exe -m models\7B\ggml-model.bin --ctx_size 2048
+server.exe -m models\7B\ggml-model.bin -c 2048
 ```
 
-That will start a server that by default listens on `127.0.0.1:8080`. You can consume the endpoints with Postman or NodeJS with axios library.
+The above command will start a server that by default listens on `127.0.0.1:8080`.
+You can consume the endpoints with Postman or NodeJS with axios library.
+
+## Testing with CURL
+
+Using [curl](https://curl.se/). On Windows `curl.exe` should be available in the base OS.
+
+```sh
+curl --request POST \
+    --url http://localhost:8080/completion \
+    --data '{"prompt": "Building a website can be done in 10 simple steps:","n_predict": 128}'
+```
 
 ## Node JS Test
 
@@ -50,7 +91,6 @@ const prompt = `Building a website can be done in 10 simple steps:`;
 async function Test() {
     let result = await axios.post("http://127.0.0.1:8080/completion", {
         prompt,
-        batch_size: 128,
         n_predict: 512,
     });
 
@@ -69,246 +109,75 @@ node .
 
 ## API Endpoints
 
-You can interact with this API Endpoints. This implementations just support chat style interaction.
+-   **POST** `/completion`: Given a prompt, it returns the predicted completion.
 
--   **POST** `hostname:port/completion`: Setting up the Llama Context to begin the completions tasks.
+    *Options:*
 
-*Options:*
+    `temperature`: Adjust the randomness of the generated text (default: 0.8).
 
-`batch_size`: Set the batch size for prompt processing (default: 512).
+    `top_k`: Limit the next token selection to the K most probable tokens (default: 40).
 
-`temperature`: Adjust the randomness of the generated text (default: 0.8).
+    `top_p`: Limit the next token selection to a subset of tokens with a cumulative probability above a threshold P (default: 0.9).
 
-`top_k`: Limit the next token selection to the K most probable tokens (default: 40).
+    `n_predict`: Set the number of tokens to predict when generating text. **Note:** May exceed the set limit slightly if the last token is a partial multibyte character.  (default: 128, -1 = infinity).
 
-`top_p`: Limit the next token selection to a subset of tokens with a cumulative probability above a threshold P (default: 0.9).
+    `n_keep`: Specify the number of tokens from the initial prompt to retain when the model resets its internal context.
+    By default, this value is set to 0 (meaning no tokens are kept). Use `-1` to retain all tokens from the initial prompt.
 
-`n_predict`: Set the number of tokens to predict when generating text (default: 128, -1 = infinity).
+    `stream`: It allows receiving each predicted token in real-time instead of waiting for the completion to finish. To enable this, set to `true`.
 
-`threads`: Set the number of threads to use during computation.
+    `prompt`: Provide a prompt. Internally, the prompt is compared, and it detects if a part has already been evaluated, and the remaining part will be evaluate.
 
-`n_keep`: Specify the number of tokens from the initial prompt to retain when the model resets its internal context. By default, this value is set to 0 (meaning no tokens are kept). Use `-1` to retain all tokens from the initial prompt.
+    `stop`: Specify a JSON array of stopping strings.
+    These words will not be included in the completion, so make sure to add them to the prompt for the next iteration (default: []).
 
-`as_loop`: It allows receiving each predicted token in real-time instead of waiting for the completion to finish. To enable this, set to `true`.
+    `tfs_z`: Enable tail free sampling with parameter z (default: 1.0, 1.0 = disabled).
 
-`interactive`: It allows interacting with the completion, and the completion stops as soon as it encounters a `stop word`. To enable this, set to `true`.
+    `typical_p`: Enable locally typical sampling with parameter p (default: 1.0, 1.0 = disabled).
 
-`prompt`: Provide a prompt. Internally, the prompt is compared, and it detects if a part has already been evaluated, and the remaining part will be evaluate.
+    `repeat_penalty`: Control the repetition of token sequences in the generated text (default: 1.1).
 
-`stop`: Specify the words or characters that indicate a stop. These words will not be included in the completion, so make sure to add them to the prompt for the next iteration.
+    `repeat_last_n`: Last n tokens to consider for penalizing repetition (default: 64, 0 = disabled, -1 = ctx-size).
 
-`exclude`: Specify the words or characters you do not want to appear in the completion. These words will not be included in the completion, so make sure to add them to the prompt for the next iteration.
+    `penalize_nl`: Penalize newline tokens when applying the repeat penalty (default: true).
 
--   **POST** `hostname:port/embedding`: Generate embedding of a given text
+    `presence_penalty`: Repeat alpha presence penalty (default: 0.0, 0.0 = disabled).
 
-*Options:*
+    `frequency_penalty`: Repeat alpha frequency penalty (default: 0.0, 0.0 = disabled);
 
-`content`: Set the text to get generate the embedding.
+    `mirostat`: Enable Mirostat sampling, controlling perplexity during text generation (default: 0, 0 = disabled, 1 = Mirostat, 2 = Mirostat 2.0).
 
-`threads`: Set the number of threads to use during computation.
+    `mirostat_tau`: Set the Mirostat target entropy, parameter tau (default: 5.0).
 
-To use this endpoint, you need to start the server with the `--embedding` option added.
+    `mirostat_eta`: Set the Mirostat learning rate, parameter eta (default: 0.1).
 
--   **POST** `hostname:port/tokenize`: Tokenize a given text
+    `seed`: Set the random number generator (RNG) seed (default: -1, < 0 = random seed).
 
-*Options:*
+    `ignore_eos`: Ignore end of stream token and continue generating (default: false).
 
-`content`: Set the text to tokenize.
+    `logit_bias`: Modify the likelihood of a token appearing in the generated text completion. For example, use `"logit_bias": [[15043,1.0]]` to increase the likelihood of the token 'Hello', or `"logit_bias": [[15043,-1.0]]` to decrease its likelihood. Setting the value to false, `"logit_bias": [[15043,false]]` ensures that the token `Hello` is never produced (default: []).
 
--   **GET** `hostname:port/next-token`: Receive the next token predicted, execute this request in a loop. Make sure set `as_loop` as `true` in the completion request.
+-   **POST** `/tokenize`: Tokenize a given text.
 
-*Options:*
+    *Options:*
 
-`stop`: Set `hostname:port/next-token?stop=true` to stop the token generation.
+    `content`: Set the text to tokenize.
 
 ## More examples
 
 ### Interactive mode
 
-This mode allows interacting in a chat-like manner. It is recommended for models designed as assistants such as `Vicuna`, `WizardLM`, `Koala`, among others. Make sure to add the correct stop word for the corresponding model.
+Check the sample in [chat.mjs](chat.mjs).
+Run with NodeJS version 16 or later:
 
-The prompt should be generated by you, according to the model's guidelines. You should keep adding the model's completions to the context as well.
+```sh
-
+node chat.mjs
-This example works well for `Vicuna - version 1`.
-
-```javascript
-const axios = require("axios");
-
-let prompt = `A chat between a curious human and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the human's questions.
-### Human: Hello, Assistant.
-### Assistant: Hello. How may I help you today?
-### Human: Please tell me the largest city in Europe.
-### Assistant: Sure. The largest city in Europe is Moscow, the capital of Russia.`;
-
-async function ChatCompletion(answer) {
-    // the user's next question to the prompt
-    prompt += `\n### Human: ${answer}\n`
-
-    result = await axios.post("http://127.0.0.1:8080/completion", {
-        prompt,
-        batch_size: 128,
-        temperature: 0.2,
-        top_k: 40,
-        top_p: 0.9,
-        n_keep: -1,
-        n_predict: 2048,
-        stop: ["\n### Human:"], // when detect this, stop completion
-        exclude: ["### Assistant:"], // no show in the completion
-        threads: 8,
-        as_loop: true, // use this to request the completion token by token
-        interactive: true, // enable the detection of a stop word
-    });
-
-    // create a loop to receive every token predicted
-    // note: this operation is blocking, avoid use this in a ui thread
-
-    let message = "";
-    while (true) {
-        // you can stop the inference adding '?stop=true' like this http://127.0.0.1:8080/next-token?stop=true
-        result = await axios.get("http://127.0.0.1:8080/next-token");
-        process.stdout.write(result.data.content);
-        message += result.data.content;
-
-        // to avoid an infinite loop
-        if (result.data.stop) {
-            console.log("Completed");
-            // make sure to add the completion to the prompt.
-            prompt += `### Assistant: ${message}`;
-            break;
-        }
-    }
-}
-
-// This function should be called every time a question to the model is needed.
-async function Test() {
-    // the server can't inference in paralell
-    await ChatCompletion("Write a long story about a time magician in a fantasy world");
-    await ChatCompletion("Summary the story");
-}
-
-Test();
 ```
 
-### Alpaca example
+Another sample in [chat.sh](chat.sh).
+Requires [bash](https://www.gnu.org/software/bash/), [curl](https://curl.se) and [jq](https://jqlang.github.io/jq/).
+Run with bash:
 
-**Temporaly note:** no tested, if you have the model, please test it and report me some issue
+```sh
-
+bash chat.sh
-```javascript
-const axios = require("axios");
-
-let prompt = `Below is an instruction that describes a task. Write a response that appropriately completes the request.
-`;
-
-async function DoInstruction(instruction) {
-    prompt += `\n\n### Instruction:\n\n${instruction}\n\n### Response:\n\n`;
-    result = await axios.post("http://127.0.0.1:8080/completion", {
-        prompt,
-        batch_size: 128,
-        temperature: 0.2,
-        top_k: 40,
-        top_p: 0.9,
-        n_keep: -1,
-        n_predict: 2048,
-        stop: ["### Instruction:\n\n"], // when detect this, stop completion
-        exclude: [], // no show in the completion
-        threads: 8,
-        as_loop: true, // use this to request the completion token by token
-        interactive: true, // enable the detection of a stop word
-    });
-
-    // create a loop to receive every token predicted
-    // note: this operation is blocking, avoid use this in a ui thread
-
-    let message = "";
-    while (true) {
-        result = await axios.get("http://127.0.0.1:8080/next-token");
-        process.stdout.write(result.data.content);
-        message += result.data.content;
-
-        // to avoid an infinite loop
-        if (result.data.stop) {
-            console.log("Completed");
-            // make sure to add the completion and the user's next question to the prompt.
-            prompt += message;
-            break;
-        }
-    }
-}
-
-// This function should be called every time a instruction to the model is needed.
-DoInstruction("Destroy the world"); // as joke
 ```
-
-### Embeddings
-
-First, run the server with `--embedding` option:
-
-```bash
-server -m models/7B/ggml-model.bin --ctx_size 2048 --embedding
-```
-
-Run this code in NodeJS:
-
-```javascript
-const axios = require('axios');
-
-async function Test() {
-    let result = await axios.post("http://127.0.0.1:8080/embedding", {
-        content: `Hello`,
-        threads: 5
-    });
-    // print the embedding array
-    console.log(result.data.embedding);
-}
-
-Test();
-```
-
-### Tokenize
-
-Run this code in NodeJS:
-
-```javascript
-const axios = require('axios');
-
-async function Test() {
-    let result = await axios.post("http://127.0.0.1:8080/tokenize", {
-        content: `Hello`
-    });
-    // print the embedding array
-    console.log(result.data.tokens);
-}
-
-Test();
-```
-
-## Common Options
-
--   `-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.
--   `--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`.
-
-### RNG Seed
-
--   `-s SEED, --seed SEED`: Set the random number generator (RNG) seed (default: -1, < 0 = random seed).
-
-The RNG seed is used to initialize the random number generator that influences the text generation process. By setting a specific seed value, you can obtain consistent and reproducible results across multiple runs with the same input and settings. This can be helpful for testing, debugging, or comparing the effects of different options on the generated text to see when they diverge. If the seed is set to a value less than 0, a random seed will be used, which will result in different outputs on each run.
-
-## Performance Tuning and Memory Options
-
-### No Memory Mapping
-
--   `--no-mmap`: Do not memory-map the model. By default, models are mapped into memory, which allows the system to load only the necessary parts of the model as needed. However, if the model is larger than your total amount of RAM or if your system is low on available memory, using mmap might increase the risk of pageouts, negatively impacting performance.
-
-### Memory Float 32
-
--   `--memory-f32`: Use 32-bit floats instead of 16-bit floats for memory key+value. This doubles the context memory requirement but does not appear to increase generation quality in a measurable way. Not recommended.
-
-## Limitations:
-
--   The actual implementation of llama.cpp need a `llama-state` for handle multiple contexts and clients, but this could require more powerful hardware.

examples/server/chat.mjs

@@ -0,0 +1,89 @@
+import * as readline from 'node:readline'
+import { stdin, stdout } from 'node:process'
+
+const API_URL = 'http://127.0.0.1:8080'
+
+const chat = [
+    {
+        human: "Hello, Assistant.",
+        assistant: "Hello. How may I help you today?"
+    },
+    {
+        human: "Please tell me the largest city in Europe.",
+        assistant: "Sure. The largest city in Europe is Moscow, the capital of Russia."
+    },
+]
+
+const instruction = `A chat between a curious human and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the human's questions.`
+
+function format_prompt(question) {
+    return `${instruction}\n${
+        chat.map(m =>`### Human: ${m.human}\n### Assistant: ${m.assistant}`).join("\n")
+    }\n### Human: ${question}\n### Assistant:`
+}
+
+async function tokenize(content) {
+    const result = await fetch(`${API_URL}/tokenize`, {
+        method: 'POST',
+        body: JSON.stringify({ content })
+    })
+
+    if (!result.ok) {
+        return []
+    }
+
+    return await result.json().tokens
+}
+
+const n_keep = await tokenize(instruction).length
+
+async function chat_completion(question) {
+    const result = await fetch(`${API_URL}/completion`, {
+        method: 'POST',
+        body: JSON.stringify({
+            prompt: format_prompt(question),
+            temperature: 0.2,
+            top_k: 40,
+            top_p: 0.9,
+            n_keep: n_keep,
+            n_predict: 256,
+            stop: ["\n### Human:"], // stop completion after generating this
+            stream: true,
+        })
+    })
+
+    if (!result.ok) {
+        return
+    }
+
+    let answer = ''
+
+    for await (var chunk of result.body) {
+        const t = Buffer.from(chunk).toString('utf8')
+        if (t.startsWith('data: ')) {
+            const message = JSON.parse(t.substring(6))
+            answer += message.content
+            process.stdout.write(message.content)
+            if (message.stop) {
+                if (message.truncated) {
+                    chat.shift()
+                }
+                break
+            }
+        }
+    }
+
+    process.stdout.write('\n')
+    chat.push({ human: question, assistant: answer.trimStart() })
+}
+
+const rl = readline.createInterface({ input: stdin, output: stdout });
+
+const readlineQuestion = (rl, query, options) => new Promise((resolve, reject) => {
+    rl.question(query, options, resolve)
+});
+
+while(true) {
+    const question = await readlineQuestion(rl, '> ')
+    await chat_completion(question)
+}

examples/server/chat.sh

@@ -0,0 +1,77 @@
+#!/bin/bash
+
+API_URL="${API_URL:-http://127.0.0.1:8080}"
+
+CHAT=(
+    "Hello, Assistant."
+    "Hello. How may I help you today?"
+    "Please tell me the largest city in Europe."
+    "Sure. The largest city in Europe is Moscow, the capital of Russia."
+)
+
+INSTRUCTION="A chat between a curious human and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the human's questions."
+
+trim() {
+    shopt -s extglob
+    set -- "${1##+([[:space:]])}"
+    printf "%s" "${1%%+([[:space:]])}"
+}
+
+trim_trailing() {
+    shopt -s extglob
+    printf "%s" "${1%%+([[:space:]])}"
+}
+
+format_prompt() {
+    echo -n "${INSTRUCTION}"
+    printf "\n### Human: %s\n### Assistant: %s" "${CHAT[@]}" "$1"
+}
+
+tokenize() {
+    curl \
+        --silent \
+        --request POST \
+        --url "${API_URL}/tokenize" \
+        --data-raw "$(jq -ns --arg content "$1" '{content:$content}')" \
+    | jq '.tokens[]'
+}
+
+N_KEEP=$(tokenize "${INSTRUCTION}" | wc -l)
+
+chat_completion() {
+    PROMPT="$(trim_trailing "$(format_prompt "$1")")"
+    DATA="$(echo -n "$PROMPT" | jq -Rs --argjson n_keep $N_KEEP '{
+        prompt: .,
+        temperature: 0.2,
+        top_k: 40,
+        top_p: 0.9,
+        n_keep: $n_keep,
+        n_predict: 256,
+        stop: ["\n### Human:"],
+        stream: true
+    }')"
+
+    ANSWER=''
+
+    while IFS= read -r LINE; do
+        if [[ $LINE = data:* ]]; then
+            CONTENT="$(echo "${LINE:5}" | jq -r '.content')"
+            printf "%s" "${CONTENT}"
+            ANSWER+="${CONTENT}"
+        fi
+    done < <(curl \
+        --silent \
+        --no-buffer \
+        --request POST \
+        --url "${API_URL}/completion" \
+        --data-raw "${DATA}")
+
+    printf "\n"
+
+    CHAT+=("$1" "$(trim "$ANSWER")")
+}
+
+while true; do
+    read -r -e -p "> " QUESTION
+    chat_completion "${QUESTION}"
+done

examples/simple/CMakeLists.txt

@@ -0,0 +1,7 @@
+set(TARGET simple)
+add_executable(${TARGET} simple.cpp)
+target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_11)
+if(TARGET BUILD_INFO)
+  add_dependencies(${TARGET} BUILD_INFO)
+endif()

examples/simple/simple.cpp

@@ -0,0 +1,177 @@
+#ifndef _GNU_SOURCE
+#define _GNU_SOURCE
+#endif
+
+#include "common.h"
+#include "llama.h"
+#include "build-info.h"
+
+#include <cassert>
+#include <cinttypes>
+#include <cmath>
+#include <cstdio>
+#include <cstring>
+#include <ctime>
+#include <fstream>
+#include <iostream>
+#include <string>
+#include <vector>
+
+#if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__))
+#include <signal.h>
+#include <unistd.h>
+#elif defined (_WIN32)
+#define WIN32_LEAN_AND_MEAN
+#define NOMINMAX
+#include <windows.h>
+#include <signal.h>
+#endif
+
+
+
+int main(int argc, char ** argv)
+{
+    gpt_params params;
+
+    //---------------------------------
+    // Print help :
+    //---------------------------------
+
+    if ( argc == 1 || argv[1][0] == '-' )
+    {
+        printf( "usage: %s MODEL_PATH [PROMPT]\n" , argv[0] );
+        return 1 ;
+    }
+
+    //---------------------------------
+    // Load parameters :
+    //---------------------------------
+
+    if ( argc >= 2 )
+    {
+        params.model = argv[1];
+    }
+
+    if ( argc >= 3 )
+    {
+        params.prompt = argv[2];
+    }
+
+    if ( params.prompt.empty() )
+    {
+        params.prompt = "Hello my name is";
+    }
+
+    //---------------------------------
+    // Init LLM :
+    //---------------------------------
+
+    llama_init_backend();
+
+    llama_context * ctx ;
+
+    ctx = llama_init_from_gpt_params( params );
+
+    if ( ctx == NULL )
+    {
+        fprintf( stderr , "%s: error: unable to load model\n" , __func__ );
+        return 1;
+    }
+
+    //---------------------------------
+    // Tokenize the prompt :
+    //---------------------------------
+
+    std::vector<llama_token> tokens_list;
+    tokens_list = ::llama_tokenize( ctx , params.prompt , true );
+
+    const int max_context_size     = llama_n_ctx( ctx );
+    const int max_tokens_list_size = max_context_size - 4 ;
+
+    if ( (int)tokens_list.size() > max_tokens_list_size )
+    {
+        fprintf( stderr , "%s: error: prompt too long (%d tokens, max %d)\n" ,
+             __func__ , (int)tokens_list.size() , max_tokens_list_size );
+        return 1;
+    }
+
+    fprintf( stderr, "\n\n" );
+
+    // Print the tokens from the prompt :
+
+    for( auto id : tokens_list )
+    {
+        printf( "%s" , llama_token_to_str( ctx , id ) );
+    }
+
+    fflush(stdout);
+
+
+    //---------------------------------
+    // Main prediction loop :
+    //---------------------------------
+
+    // The LLM keeps a contextual cache memory of previous token evaluation.
+    // Usually, once this cache is full, it is required to recompute a compressed context based on previous
+    // tokens (see "infinite text generation via context swapping" in the main example), but in this minimalist
+    // example, we will just stop the loop once this cache is full or once an end of stream is detected.
+
+    while ( llama_get_kv_cache_token_count( ctx ) < max_context_size )
+    {
+        //---------------------------------
+        // Evaluate the tokens :
+        //---------------------------------
+
+        if ( llama_eval( ctx , tokens_list.data() , tokens_list.size() , llama_get_kv_cache_token_count( ctx ) , params.n_threads ) )
+        {
+            fprintf( stderr,  "%s : failed to eval\n" , __func__ );
+            return 1;
+        }
+
+        tokens_list.clear();
+
+        //---------------------------------
+        // Select the best prediction :
+        //---------------------------------
+
+        llama_token new_token_id = 0;
+
+        auto logits  = llama_get_logits( ctx );
+        auto n_vocab = llama_n_vocab( ctx ); // the size of the LLM vocabulary (in tokens)
+
+        std::vector<llama_token_data> candidates;
+        candidates.reserve( n_vocab );
+
+        for( llama_token token_id = 0 ; token_id < n_vocab ; token_id++ )
+        {
+            candidates.emplace_back( llama_token_data{ token_id , logits[ token_id ] , 0.0f } );
+        }
+
+        llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false };
+
+        // Select it using the "Greedy sampling" method :
+        new_token_id = llama_sample_token_greedy( ctx , &candidates_p );
+
+
+        // is it an end of stream ?
+        if ( new_token_id == llama_token_eos() )
+        {
+            fprintf(stderr, " [end of text]\n");
+            break;
+        }
+
+        // Print the new token :
+        printf( "%s" , llama_token_to_str( ctx , new_token_id ) );
+        fflush( stdout );
+
+        // Push this new token for next evaluation :
+        tokens_list.push_back( new_token_id );
+
+    } // wend of main loop
+
+    llama_free( ctx );
+
+    return 0;
+}
+
+// EOF

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://raw.githubusercontent.com/brunoklein99/deep-learning-notes/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.nix

@@ -48,6 +48,19 @@
           '';
           meta.mainProgram = "llama";
         };
+        apps.llama-server = {
+          type = "app";
+          program = "${self.packages.${system}.default}/bin/llama-server";
+        };
+        apps.llama-embedding = {
+          type = "app";
+          program = "${self.packages.${system}.default}/bin/embedding";
+        };
+        apps.llama = {
+          type = "app";
+          program = "${self.packages.${system}.default}/bin/llama";
+        };
+        apps.default = self.apps.${system}.llama;
         devShells.default = pkgs.mkShell {
           packages = with pkgs; [
             cmake

ggml-cuda.h

@@ -24,11 +24,14 @@ 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_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

@@ -41,12 +41,15 @@ void ggml_metal_free(struct ggml_metal_context * ctx);
 // - 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
+// - max_size specifies the maximum size of a tensor and is used to create shared views such
+//   that it is guaranteed that the tensor will fit in at least one of the views
 //
 bool ggml_metal_add_buffer(
         struct ggml_metal_context * ctx,
                        const char * name,
                              void * data,
-                           size_t   size);
+                           size_t   size,
+                           size_t   max_size);
 
 // set data from host memory into the device
 void ggml_metal_set_tensor(struct ggml_metal_context * ctx, struct ggml_tensor * t);
@@ -55,6 +58,7 @@ void ggml_metal_set_tensor(struct ggml_metal_context * ctx, struct ggml_tensor *
 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

ggml-metal.m

@@ -52,18 +52,25 @@ struct ggml_metal_context {
     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(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(alibi_f32);
     GGML_METAL_DECL_KERNEL(cpy_f32_f16);
     GGML_METAL_DECL_KERNEL(cpy_f32_f32);
+    GGML_METAL_DECL_KERNEL(cpy_f16_f16);
 
 #undef GGML_METAL_DECL_KERNEL
 };
@@ -86,6 +93,7 @@ struct ggml_metal_context * ggml_metal_init(void) {
 
     ctx->device = MTLCreateSystemDefaultDevice();
     ctx->queue  = [ctx->device newCommandQueue];
+    ctx->n_buffers = 0;
 
     // determine if we can use MPS
     if (MPSSupportsMTLDevice(ctx->device)) {
@@ -152,22 +160,37 @@ struct ggml_metal_context * ggml_metal_init(void) {
         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(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(alibi_f32);
         GGML_METAL_ADD_KERNEL(cpy_f32_f16);
         GGML_METAL_ADD_KERNEL(cpy_f32_f32);
+        GGML_METAL_ADD_KERNEL(cpy_f16_f16);
 
 #undef GGML_METAL_ADD_KERNEL
     }
 
+    fprintf(stderr, "%s: recommendedMaxWorkingSetSize = %8.2f MB\n", __func__, ctx->device.recommendedMaxWorkingSetSize / 1024.0 / 1024.0);
+    fprintf(stderr, "%s: hasUnifiedMemory             = %s\n",       __func__, ctx->device.hasUnifiedMemory ? "true" : "false");
+    if (ctx->device.maxTransferRate != 0) {
+        fprintf(stderr, "%s: maxTransferRate              = %8.2f MB/s\n", __func__, ctx->device.maxTransferRate / 1024.0 / 1024.0);
+    } else {
+        fprintf(stderr, "%s: maxTransferRate              = built-in GPU\n", __func__);
+    }
+
     return ctx;
 }
 
@@ -184,10 +207,13 @@ void ggml_metal_free(struct ggml_metal_context * ctx) {
 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);
 
+    const int64_t tsize = ggml_nbytes(t);
+
+    // find the view that contains the tensor fully
     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) {
+        if (ioffs >= 0 && ioffs + tsize <= (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);
@@ -205,7 +231,8 @@ bool ggml_metal_add_buffer(
         struct ggml_metal_context * ctx,
                      const char * name,
                            void * data,
-                         size_t   size) {
+                         size_t   size,
+                         size_t   max_size) {
     if (ctx->n_buffers >= GGML_METAL_MAX_BUFFERS) {
         fprintf(stderr, "%s: too many buffers\n", __func__);
         return false;
@@ -222,31 +249,69 @@ bool ggml_metal_add_buffer(
             }
         }
 
-        size_t page_size = getpagesize();
+        const size_t size_page = getpagesize();
-        size_t aligned_size = size;
+
-        if ((aligned_size % page_size) != 0) {
+        size_t size_aligned = size;
-            aligned_size += (page_size - (aligned_size % page_size));
+        if ((size_aligned % size_page) != 0) {
+            size_aligned += (size_page - (size_aligned % size_page));
         }
 
+        // the buffer fits into the max buffer size allowed by the device
+        if (size_aligned <= ctx->device.maxBufferLength) {
             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) {
+            ctx->buffers[ctx->n_buffers].metal = [ctx->device newBufferWithBytesNoCopy:data length:size_aligned options:MTLResourceStorageModeShared deallocator:nil];
-            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);
+                fprintf(stderr, "%s: failed to allocate '%-16s' buffer, size = %8.2f MB\n", __func__, name, size_aligned / 1024.0 / 1024.0);
                 return false;
+            }
+
+            fprintf(stderr, "%s: allocated '%-16s' buffer, size = %8.2f MB", __func__, name, size_aligned / 1024.0 / 1024.0);
+
+            ++ctx->n_buffers;
         } else {
-            fprintf(stderr, "%s: allocated '%-16s' buffer, size = %8.2f MB\n", __func__, name, aligned_size / 1024.0 / 1024.0);
+            // this overlap between the views will guarantee that the tensor with the maximum size will fully fit into
+            // one of the views
+            const size_t size_ovlp = ((max_size + size_page - 1) / size_page + 1) * size_page; // round-up 2 pages just in case
+            const size_t size_step = ctx->device.maxBufferLength - size_ovlp;
+            const size_t size_view = ctx->device.maxBufferLength;
+
+            for (size_t i = 0; i < size; i += size_step) {
+                const size_t size_step_aligned = (i + size_view <= size) ? size_view : (size_aligned - i);
+
+                ctx->buffers[ctx->n_buffers].name = name;
+                ctx->buffers[ctx->n_buffers].data = (void *) ((uint8_t *) data + i);
+                ctx->buffers[ctx->n_buffers].size = size_step_aligned;
+
+                ctx->buffers[ctx->n_buffers].metal = [ctx->device newBufferWithBytesNoCopy:(void *) ((uint8_t *) data + i) length:size_step_aligned 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, size_step_aligned / 1024.0 / 1024.0);
+                    return false;
+                }
+
+                fprintf(stderr, "%s: allocated '%-16s' buffer, size = %8.2f MB, offs = %12ld", __func__, name, size_step_aligned / 1024.0 / 1024.0, i);
+                if (i + size_step < size) {
+                    fprintf(stderr, "\n");
                 }
 
                 ++ctx->n_buffers;
             }
+        }
+
+        fprintf(stderr, ", (%8.2f / %8.2f)",
+                ctx->device.currentAllocatedSize / 1024.0 / 1024.0,
+                ctx->device.recommendedMaxWorkingSetSize / 1024.0 / 1024.0);
+
+        if (ctx->device.currentAllocatedSize > ctx->device.recommendedMaxWorkingSetSize) {
+            fprintf(stderr, ", warning: current allocated size is greater than the recommended max working set size\n");
+        } else {
+            fprintf(stderr, "\n");
+        }
+    }
 
     return true;
 }
@@ -278,15 +343,40 @@ void ggml_metal_graph_compute(
                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  = [ctx->queue commandBuffer];
+            id<MTLCommandBuffer> command_buffer = command_buffers[cb_idx];
+
             id<MTLComputeCommandEncoder> encoder = nil;
 
-    for (int i = 0; i < gf->n_nodes; ++i) {
+            const int node_start =                                      (cb_idx + 0) * n_nodes_per_cb;
-        //metal_printf("%s: encoding node %3d, op = %8s\n", __func__, i, ggml_op_name(gf->nodes[i]->op));
+            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;
@@ -574,6 +664,15 @@ void ggml_metal_graph_compute(
                                             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);
@@ -583,6 +682,15 @@ void ggml_metal_graph_compute(
                                             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);
@@ -599,7 +707,6 @@ void ggml_metal_graph_compute(
                                         }
                                 };
 
-
                                 [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];
@@ -619,15 +726,14 @@ void ggml_metal_graph_compute(
                                 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) {
+                                }
+                                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 if (src0t == GGML_TYPE_Q4_K) {
-                            [encoder setThreadgroupMemoryLength:nth0*nth1*sizeof(float) atIndex:0];
-                            [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne11, 1) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
-                        } else if (src0t == GGML_TYPE_Q6_K) {
-                            [encoder setThreadgroupMemoryLength:nth0*nth1*sizeof(float) atIndex:0];
-                            [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne11, 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)];
@@ -645,7 +751,9 @@ void ggml_metal_graph_compute(
                                 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");
                             }
@@ -683,6 +791,70 @@ void ggml_metal_graph_compute(
 
                             [encoder dispatchThreadgroups:MTLSizeMake(nrows, 1, 1) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
                         } break;
+                    case GGML_OP_NORM:
+                        {
+                            if (encoder == nil) {
+                                encoder = [command_buffer computeCommandEncoder];
+                            }
+
+                            const float eps = 1e-5f;
+
+                            const int nth = 256;
+
+                            [encoder setComputePipelineState:ctx->pipeline_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_ALIBI:
+                        {
+                            if (encoder == nil) {
+                                encoder = [command_buffer computeCommandEncoder];
+                            }
+
+                            GGML_ASSERT((src0t == GGML_TYPE_F32));
+
+                            const int   n_past   = ((int32_t *) src1->data)[0]; UNUSED(n_past);
+                            const int   n_head   = ((int32_t *) src1->data)[1];
+                            const float max_bias = ((float *)   src1->data)[2];
+
+                            if (__builtin_popcount(n_head) != 1) {
+                                GGML_ASSERT(false && "only power-of-two n_head implemented");
+                            }
+
+                            const int n_heads_log2_floor = 1 << (int) floor(log2(n_head));
+                            const float m0 = powf(2.0f, -(max_bias) / n_heads_log2_floor);
+
+                            [encoder setComputePipelineState:ctx->pipeline_alibi_f32];
+                            [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:&m0  length:sizeof(    float) atIndex:18];
+                            const int nth = 32;
+                            [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
+                        } break;
                     case GGML_OP_ROPE:
                         {
                             if (encoder == nil) {
@@ -736,6 +908,14 @@ void ggml_metal_graph_compute(
                                             default: GGML_ASSERT(false && "not implemented");
                                         };
                                     } break;
+                                case GGML_TYPE_F16:
+                                    {
+                                        switch (dstt) {
+                                            case GGML_TYPE_F16: [encoder setComputePipelineState:ctx->pipeline_cpy_f16_f16]; break;
+                                            case GGML_TYPE_F32: GGML_ASSERT(false && "cpy_f16_f32 not implemented"); break;
+                                            default: GGML_ASSERT(false && "not implemented");
+                                        };
+                                    } break;
                                 default: GGML_ASSERT(false && "not implemented");
                             }
 
@@ -772,12 +952,21 @@ void ggml_metal_graph_compute(
             }
 
             [command_buffer commit];
-    [command_buffer waitUntilCompleted];
+        });
+    }
 
-    {
+    // wait for all threads to finish
-        const double time_elapsed = [command_buffer GPUEndTime] - [command_buffer GPUStartTime];
+    dispatch_barrier_sync(queue, ^{});
-        UNUSED(time_elapsed);
 
-        metal_printf("%s: time elapsed = %f ms\n", __func__, time_elapsed * 1000.0);
+    [command_buffers[n_cb - 1] waitUntilCompleted];
+
+    // check status of command buffers
+    // needed to detect if the device ran out-of-memory for example (#1881)
+    for (int i = 0; i < n_cb; i++) {
+        MTLCommandBufferStatus status = (MTLCommandBufferStatus) [command_buffers[i] status];
+        if (status != MTLCommandBufferStatusCompleted) {
+            fprintf(stderr, "%s: command buffer %d failed with status %lu\n", __func__, i, status);
+            GGML_ASSERT(false);
+        }
     }
 }

ggml-metal.metal

@@ -256,6 +256,72 @@ kernel void kernel_get_rows_q4_1(
                        (device float *) ((device char *)  dst + i*nb1), ne00);
 }
 
+kernel void kernel_norm(
+        device const  void * src0,
+        device       float * dst,
+        constant   int64_t & ne00,
+        constant  uint64_t & nb01,
+        constant     float & eps,
+        threadgroup float  * sum [[threadgroup(0)]],
+        uint tgpig[[threadgroup_position_in_grid]],
+        uint tpitg[[thread_position_in_threadgroup]],
+        uint   ntg[[threads_per_threadgroup]]) {
+    device const float * x = (device const float *) ((device const char *) src0 + tgpig*nb01);
+    // MEAN
+    // parallel sum
+    sum[tpitg] = 0.0f;
+    for (int i00 = tpitg; i00 < ne00; i00 += ntg) {
+        sum[tpitg] += x[i00];
+    }
+    // reduce
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+    for (uint i = ntg/2; i > 0; i /= 2) {
+        if (tpitg < i) {
+            sum[tpitg] += sum[tpitg + i];
+        }
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+    }
+    // broadcast
+    if (tpitg == 0) {
+        sum[0] /= ne00;
+    }
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+    const float mean  = sum[0];
+
+    // recenter
+    device float * y = dst + tgpig*ne00;
+    for (int i00 = tpitg; i00 < ne00; i00 += ntg) {
+        y[i00] = x[i00] - mean;
+    }
+
+    // VARIANCE
+    // parallel sum
+    sum[tpitg] = 0.0f;
+    for (int i00 = tpitg; i00 < ne00; i00 += ntg) {
+        sum[tpitg] += y[i00] * y[i00];
+    }
+    // reduce
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+    for (uint i = ntg/2; i > 0; i /= 2) {
+        if (tpitg < i) {
+            sum[tpitg] += sum[tpitg + i];
+        }
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+    }
+    // broadcast
+    if (tpitg == 0) {
+        sum[0] /= ne00;
+    }
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+    const float variance = sum[0];
+
+    const float scale = 1.0f/sqrt(variance + eps);
+    for (int i00 = tpitg; i00 < ne00; i00 += ntg) {
+        y[i00] = y[i00] * scale;
+    }
+}
+
+
 kernel void kernel_rms_norm(
         device const  void * src0,
         device       float * dst,
@@ -304,34 +370,22 @@ kernel void kernel_mul_mat_q4_0_f32(
         device const float * src1,
         device       float * dst,
         constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant  uint64_t & nb00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
         constant   int64_t & ne10,
-        constant   int64_t & ne11,
-        constant  uint64_t & nb10,
-        constant  uint64_t & nb11,
-        constant  uint64_t & nb12,
         constant   int64_t & ne0,
-        constant   int64_t & ne1,
         threadgroup float  * sum [[threadgroup(0)]],
         uint2 tgpig[[threadgroup_position_in_grid]],
-        uint2  tpig[[thread_position_in_grid]],
         uint2 tpitg[[thread_position_in_threadgroup]],
         uint2  tptg[[threads_per_threadgroup]]) {
     const int nb = ne00/QK4_0;
 
-    const int8_t m8 = 8;
-
     const int64_t r0 = tgpig.x;
     const int64_t r1 = tgpig.y;
 
     device const block_q4_0 * x = (device const block_q4_0 *) src0 + r0*nb;
     device const float      * y = (device const float      *) src1 + r1*ne10;
 
-    const uint nth = tptg.x*tptg.y;
+    const int nth = tptg.x*tptg.y;
-    const uint ith = tptg.y*tpitg.x + tpitg.y;
+    const int ith = tptg.y*tpitg.x + tpitg.y;
 
     const int ix = tpitg.y/4;           // 0 or 1
     const int iy = tpitg.y - 4*ix;      // 0...3
@@ -351,47 +405,32 @@ kernel void kernel_mul_mat_q4_0_f32(
 
         for (int j = 0; j < 4; ++j) {
 
-            acc[0] += yl[j+ 0] * ((int8_t)(xl[j] & 0xF) - m8);
+            acc[0] += yl[j] * (xl[j] & 0xF) + yl[j+16] * (xl[j] >> 4);
-            acc[1] += yl[j+16] * ((int8_t)(xl[j] >>  4) - m8);
+            acc[1] += yl[j] + yl[j+16];
 
         }
 
-        sumf += d * (acc[0] + acc[1]);
+        sumf += d * (acc[0] - 8.f*acc[1]);
     }
 
     sum[ith] = sumf;
 
     //
     // Accumulate the sum from all threads in the threadgroup
-    // This version is slightly faster than the commented out one below,
-    // which I copy-pasted from ggerganov's q4_0 dot product for metal.
     //
     threadgroup_barrier(mem_flags::mem_threadgroup);
     if (ith%4 == 0) {
-        for (int i = 1; i < 4; ++i) sum[ith] += sum[ith + i];
+        sum[ith] += sum[ith+1] + sum[ith+2] + sum[ith+3];
     }
     threadgroup_barrier(mem_flags::mem_threadgroup);
     if (ith%16 == 0) {
-        for (int i = 4; i < 16; i += 4) sum[ith] += sum[ith + i];
+        sum[ith] += sum[ith+4] + sum[ith+8] + sum[ith+12];
     }
     threadgroup_barrier(mem_flags::mem_threadgroup);
     if (ith == 0) {
-        for (int i = 16; i < nth; i += 16) sum[0] += sum[i];
+        for (uint i = 16; i < nth; i += 16) sum[0] += sum[i];
         dst[r1*ne0 + r0] = sum[0];
     }
-
-    //// accumulate the sum from all threads in the threadgroup
-    //threadgroup_barrier(mem_flags::mem_threadgroup);
-    //for (uint i = nth/2; i > 0; i /= 2) {
-    //    if (ith < i) {
-    //        sum[ith] += sum[ith + i];
-    //    }
-    //    threadgroup_barrier(mem_flags::mem_threadgroup);
-    //}
-
-    //if (ith == 0) {
-    //    dst[r1*ne0 + r0] = sum[0];
-    //}
 }
 
 kernel void kernel_mul_mat_q4_1_f32(
@@ -399,20 +438,10 @@ kernel void kernel_mul_mat_q4_1_f32(
         device const float * src1,
         device       float * dst,
         constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant  uint64_t & nb00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
         constant   int64_t & ne10,
-        constant   int64_t & ne11,
-        constant  uint64_t & nb10,
-        constant  uint64_t & nb11,
-        constant  uint64_t & nb12,
         constant   int64_t & ne0,
-        constant   int64_t & ne1,
         threadgroup float  * sum [[threadgroup(0)]],
         uint2 tgpig[[threadgroup_position_in_grid]],
-        uint2  tpig[[thread_position_in_grid]],
         uint2 tpitg[[thread_position_in_threadgroup]],
         uint2  tptg[[threads_per_threadgroup]]) {
     const int nb = ne00/QK4_1;
@@ -460,11 +489,11 @@ kernel void kernel_mul_mat_q4_1_f32(
     //
     threadgroup_barrier(mem_flags::mem_threadgroup);
     if (ith%4 == 0) {
-        for (int i = 1; i < 4; ++i) sum[ith] += sum[ith + i];
+        sum[ith] += sum[ith+1] + sum[ith+2] + sum[ith+3];
     }
     threadgroup_barrier(mem_flags::mem_threadgroup);
     if (ith%16 == 0) {
-        for (int i = 4; i < 16; i += 4) sum[ith] += sum[ith + i];
+        sum[ith] += sum[ith+4] + sum[ith+8] + sum[ith+12];
     }
     threadgroup_barrier(mem_flags::mem_threadgroup);
     if (ith == 0) {
@@ -522,6 +551,48 @@ kernel void kernel_mul_mat_f16_f32(
     }
 }
 
+kernel void kernel_alibi_f32(
+        device const float * src0,
+        device       float * dst,
+        constant   int64_t & ne00,
+        constant   int64_t & ne01,
+        constant   int64_t & ne02,
+        constant   int64_t & ne03,
+        constant  uint64_t & nb00,
+        constant  uint64_t & nb01,
+        constant  uint64_t & nb02,
+        constant  uint64_t & nb03,
+        constant   int64_t & ne0,
+        constant   int64_t & ne1,
+        constant   int64_t & ne2,
+        constant   int64_t & ne3,
+        constant  uint64_t & nb0,
+        constant  uint64_t & nb1,
+        constant  uint64_t & nb2,
+        constant  uint64_t & nb3,
+        constant      float & m0,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3   ntg[[threads_per_threadgroup]]) {
+    const int64_t i03 = tgpig[2];
+    const int64_t i02 = tgpig[1];
+    const int64_t i01 = tgpig[0];
+
+    const int64_t n = i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
+
+    const int64_t i3 = n / (ne2*ne1*ne0);
+    const int64_t i2 = (n - i3*ne2*ne1*ne0) / (ne1*ne0);
+    const int64_t i1 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0) / ne0;
+    const int64_t i0 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0 - i1*ne0);
+
+    device float * dst_data = (device float *) ((device char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
+    float m_k = pow(m0, i2 + 1);
+    for (int64_t i00 = tpitg.x; i00 < ne00; i00 += ntg.x) {
+        device const float * src = (device float *)((device char *) src0 + i03*nb03 + i02*nb02 + i01*nb01 + i00*nb00);
+        dst_data[i00] = src[0] + m_k * (i00 - ne00 + 1);
+    }
+}
+
 kernel void kernel_rope(
         device const  void * src0,
         device       float * dst,
@@ -577,6 +648,47 @@ kernel void kernel_rope(
     }
 }
 
+kernel void kernel_cpy_f16_f16(
+        device const half * src0,
+        device       half * dst,
+        constant   int64_t & ne00,
+        constant   int64_t & ne01,
+        constant   int64_t & ne02,
+        constant   int64_t & ne03,
+        constant  uint64_t & nb00,
+        constant  uint64_t & nb01,
+        constant  uint64_t & nb02,
+        constant  uint64_t & nb03,
+        constant   int64_t & ne0,
+        constant   int64_t & ne1,
+        constant   int64_t & ne2,
+        constant   int64_t & ne3,
+        constant  uint64_t & nb0,
+        constant  uint64_t & nb1,
+        constant  uint64_t & nb2,
+        constant  uint64_t & nb3,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3   ntg[[threads_per_threadgroup]]) {
+    const int64_t i03 = tgpig[2];
+    const int64_t i02 = tgpig[1];
+    const int64_t i01 = tgpig[0];
+
+    const int64_t n = i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
+
+    const int64_t i3 = n / (ne2*ne1*ne0);
+    const int64_t i2 = (n - i3*ne2*ne1*ne0) / (ne1*ne0);
+    const int64_t i1 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0) / ne0;
+    const int64_t i0 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0 - i1*ne0);
+
+    device half * dst_data = (device half *) ((device char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
+
+    for (int64_t i00 = tpitg.x; i00 < ne00; i00 += ntg.x) {
+        device const half * src = (device half *)((device char *) src0 + i03*nb03 + i02*nb02 + i01*nb01 + i00*nb00);
+        dst_data[i00] = src[0];
+    }
+}
+
 kernel void kernel_cpy_f32_f16(
         device const float * src0,
         device        half * dst,
@@ -671,6 +783,15 @@ typedef struct {
     half d;           // super-block scale for quantized scales
     half dmin;        // super-block scale for quantized mins
 } block_q2_k;
+// 84 bytes / block
+
+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
+    half d;                    // super-block scale
+} block_q3_k;
+// 110 bytes / block
 
 typedef struct {
     half d;             // super-block scale for quantized scales
@@ -678,6 +799,16 @@ typedef struct {
     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;
+// 144 bytes / block
+
+typedef struct {
+    half d;                      // super-block scale for quantized scales
+    half 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;
+// 176 bytes / block
 
 typedef struct {
     uint8_t ql[QK_K/2];      // quants, lower 4 bits
@@ -685,16 +816,19 @@ typedef struct {
     int8_t  scales[QK_K/16]; // scales, quantized with 8 bits
     half d;                  // super-block scale
 } block_q6_k;
+// 210 bytes / block
 
 static inline uchar4 get_scale_min_k4(int j, device const uint8_t * q) {
     uchar4 r;
     if (j < 4) {
-        r[0] = q[j+0] & 63; r[1] = q[j+4] & 63;
+        r[0] = q[j+0] & 63;
-        r[2] = q[j+1] & 63; r[3] = q[j+5] & 63;
+        r[2] = q[j+1] & 63;
+        r[1] = q[j+4] & 63;
+        r[3] = q[j+5] & 63;
     } else {
         r[0] = (q[j+4] & 0xF) | ((q[j-4] >> 6) << 4);
-        r[1] = (q[j+4] >>  4) | ((q[j-0] >> 6) << 4);
         r[2] = (q[j+5] & 0xF) | ((q[j-3] >> 6) << 4);
+        r[1] = (q[j+4] >>  4) | ((q[j-0] >> 6) << 4);
         r[3] = (q[j+5] >>  4) | ((q[j+1] >> 6) << 4);
     }
     return r;
@@ -735,10 +869,65 @@ static void dequantize_row_q2_k(device const block_q2_k * x, device float * y, i
     }
 }
 
+static void dequantize_row_q3_k(device const block_q3_k * x, device float * y, int k) {
+    assert(k % QK_K == 0);
+    const int nb = k / QK_K;
+
+    const uint16_t kmask1 = 0x0303;
+    const uint16_t kmask2 = 0x0f0f;
+
+    uint16_t aux[8];
+    thread const int8_t * scales = (thread const int8_t*)aux;
+
+    for (int i = 0; i < nb; i++) {
+
+        const float d_all = (float)(x[i].d);
+
+        device const uint8_t * q = x[i].qs;
+        device const uint8_t * h = x[i].hmask;
+        uint8_t m = 1;
+
+        device const uint16_t * a = (device const uint16_t *)x[i].scales;
+        aux[0] = (a[0] & kmask2) | (((a[4] >> 0) & kmask1) << 4);
+        aux[1] = (a[1] & kmask2) | (((a[5] >> 0) & kmask1) << 4);
+        aux[2] = (a[2] & kmask2) | (((a[4] >> 2) & kmask1) << 4);
+        aux[3] = (a[3] & kmask2) | (((a[5] >> 2) & kmask1) << 4);
+        aux[4] = ((a[0] >> 4) & kmask2) | (((a[4] >> 4) & kmask1) << 4);
+        aux[5] = ((a[1] >> 4) & kmask2) | (((a[5] >> 4) & kmask1) << 4);
+        aux[6] = ((a[2] >> 4) & kmask2) | (((a[4] >> 6) & kmask1) << 4);
+        aux[7] = ((a[3] >> 4) & kmask2) | (((a[5] >> 6) & kmask1) << 4);
+
+        int is = 0;
+        float dl;
+        for (int n = 0; n < QK_K; n += 128) {
+            int shift = 0;
+            for (int j = 0; j < 4; ++j) {
+
+                dl = d_all * (scales[is++] - 32);
+                for (int l = 0; l < 16; ++l) {
+                    *y++ = dl * ((int8_t)((q[l+ 0] >> shift) & 3) - ((h[l+ 0] & m) ? 0 : 4));
+                }
+
+                dl = d_all * (scales[is++] - 32);
+                for (int l = 0; l < 16; ++l) {
+                    *y++ = dl * ((int8_t)((q[l+16] >> shift) & 3) - ((h[l+16] & m) ? 0 : 4));
+                }
+
+                shift += 2;
+                m <<= 1;
+            }
+            q += 32;
+        }
+
+    }
+
+}
+
 static void dequantize_row_q4_k(device const block_q4_k * x, device float * y, int k) {
     assert(k % QK_K == 0);
     const int nb = k / QK_K;
 
+
     for (int i = 0; i < nb; i++) {
 
         const float d = x[i].d;
@@ -760,6 +949,33 @@ static void dequantize_row_q4_k(device const block_q4_k * x, device float * y, i
     }
 }
 
+static void dequantize_row_q5_k(device const block_q5_k * x, device float * y, int k) {
+    assert(k % QK_K == 0);
+    const int nb = k / QK_K;
+
+   for (int i = 0; i < nb; i++) {
+
+        const float d = (float)(x[i].d);
+        const float min = (float)(x[i].dmin);
+
+        device const uint8_t * ql = x[i].qs;
+        device const uint8_t * qh = x[i].qh;
+
+        int is = 0;
+        uint8_t u1 = 1, u2 = 2;
+        for (int j = 0; j < QK_K; j += 64) {
+            const uchar4 sc = get_scale_min_k4(is, x[i].scales);
+            const float d1 = d * sc[0]; const float m1 = min * sc[1];
+            const float d2 = d * sc[2]; const float m2 = min * sc[3];
+            for (int l = 0; l < 32; ++l) *y++ = d1 * ((ql[l] & 0xF) + (qh[l] & u1 ? 16 : 0)) - m1;
+            for (int l = 0; l < 32; ++l) *y++ = d2 * ((ql[l]  >> 4) + (qh[l] & u2 ? 16 : 0)) - m2;
+            ql += 32; is += 2;
+            u1 <<= 2; u2 <<= 2;
+        }
+    }
+
+}
+
 static void dequantize_row_q6_k(device const block_q6_k * x, device float * y, int k) {
     assert(k % QK_K == 0);
     const int nb = k / QK_K;
@@ -808,6 +1024,22 @@ kernel void kernel_get_rows_q2_k(
                        (device float *) ((device char *)  dst + i*nb1), ne00);
 }
 
+kernel void kernel_get_rows_q3_k(
+        device const  void * src0,
+        device const   int * src1,
+        device       float * dst,
+        constant   int64_t & ne00,
+        constant  uint64_t & nb01,
+        constant  uint64_t & nb1,
+        uint tpig[[thread_position_in_grid]]) {
+    const int i = tpig;
+    const int r = ((device int32_t *) src1)[i];
+
+    dequantize_row_q3_k(
+            (device const block_q3_k *) ((device char *) src0 + r*nb01),
+                       (device float *) ((device char *)  dst + i*nb1), ne00);
+}
+
 kernel void kernel_get_rows_q4_k(
         device const  void * src0,
         device const   int * src1,
@@ -824,6 +1056,22 @@ kernel void kernel_get_rows_q4_k(
                        (device float *) ((device char *)  dst + i*nb1), ne00);
 }
 
+kernel void kernel_get_rows_q5_k(
+        device const  void * src0,
+        device const   int * src1,
+        device       float * dst,
+        constant   int64_t & ne00,
+        constant  uint64_t & nb01,
+        constant  uint64_t & nb1,
+        uint tpig[[thread_position_in_grid]]) {
+    const int i = tpig;
+    const int r = ((device int32_t *) src1)[i];
+
+    dequantize_row_q5_k(
+            (device const block_q5_k *) ((device char *) src0 + r*nb01),
+                       (device float *) ((device char *)  dst + i*nb1), ne00);
+}
+
 kernel void kernel_get_rows_q6_k(
         device const  void * src0,
         device const   int * src1,
@@ -847,20 +1095,10 @@ kernel void kernel_mul_mat_q2_k_f32(
         device const float * src1,
         device       float * dst,
         constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant  uint64_t & nb00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
         constant   int64_t & ne10,
-        constant   int64_t & ne11,
-        constant  uint64_t & nb10,
-        constant  uint64_t & nb11,
-        constant  uint64_t & nb12,
         constant   int64_t & ne0,
-        constant   int64_t & ne1,
         threadgroup float  * sum [[threadgroup(0)]],
         uint2 tgpig[[threadgroup_position_in_grid]],
-        uint2  tpig[[thread_position_in_grid]],               // we don't use this for now
         uint2 tpitg[[thread_position_in_threadgroup]],
         uint2  tptg[[threads_per_threadgroup]]) {
 
@@ -875,7 +1113,6 @@ kernel void kernel_mul_mat_q2_k_f32(
     const int nth = tptg.x*tptg.y;
     const int ith = tptg.y*tpitg.x + tpitg.y;
 
-
     const int tid = tpitg.y;    // 0...16
     const int il  = tid/4;      // 0...3
     const int ir  = tid%4;      // 0...3
@@ -885,35 +1122,54 @@ kernel void kernel_mul_mat_q2_k_f32(
     const int n   = 8;
     const int is  = 4*il + (n*ir)/16;
 
+    const int y_offset = 64*il + n*ir;
+    const int q_offset = 32*ip + n*ir;
+
     sum[ith] = 0.0f;
 
     float sumf = 0;
     for (int i = tpitg.x; i < nb; i += tptg.x) {
 
-        device const uint8_t * q = x[i].qs + 32*ip + n*ir;
+        device const uint8_t * q = x[i].qs + q_offset;
         device const uint8_t * scales = x[i].scales + is;
 
         uint8_t d1 = scales[0] & 0xF;
-        uint8_t m1 = scales[0] >>  4;
         uint8_t d2 = scales[2] & 0xF;
+        uint8_t m1 = scales[0] >>  4;
         uint8_t m2 = scales[2] >>  4;
 
-        device const float   * y = yy + i*QK_K + 64*il + n*ir;
+        device const float   * y = yy + i*QK_K + y_offset;
+
+        //float4 s = {0.f, 0.f, 0.f, 0.f};
+        float2 s = {0.f, 0.f};
+        float smin = 0;
+        for (int l = 0; l < n; ++l) {
+            s[0] += y[l+ 0] * ((q[l] >> shift1) & 3);
+            s[1] += y[l+32] * ((q[l] >> shift2) & 3);
+            smin += y[l+ 0] * m1 + y[l+32] * m2;
+        }
 
         const float dall = (float)x[i].d;
         const float dmin = (float)x[i].dmin;
 
-        float4 s = {0.f, 0.f, 0.f, 0.f};
+        sumf += dall * (s[0] * d1 + s[1] * d2) - dmin * smin;
-        for (int l = 0; l < n; ++l) {
-            s[0] += y[l+ 0] * ((q[l] >> shift1) & 3); s[1] += y[l+ 0];
-            s[2] += y[l+32] * ((q[l] >> shift2) & 3); s[3] += y[l+32];
-        }
-        sumf += dall * (s[0] * d1 + s[2] * d2) - dmin * (s[1] * m1 + s[3] * m2);
-
 
     }
     sum[ith] = sumf;
 
+    //int mask1 = (ith%4 == 0);
+    //int mask2 = (ith%16 == 0);
+
+    //threadgroup_barrier(mem_flags::mem_threadgroup);
+    //for (int i = 1; i < 4; ++i) sum[ith] += mask1 * sum[ith + i];
+    //threadgroup_barrier(mem_flags::mem_threadgroup);
+    //for (int i = 4; i < 16; i += 4) sum[ith] += mask2 * sum[ith + i];
+    //threadgroup_barrier(mem_flags::mem_threadgroup);
+    //if (ith == 0) {
+    //    for (int i = 16; i < nth; i += 16) sum[0] += sum[i];
+    //    dst[r1*ne0 + r0] = sum[0];
+    //}
+
     //
     // Accumulate the sum from all threads in the threadgroup
     // This version is slightly faster than the commented out one below,
@@ -932,19 +1188,109 @@ kernel void kernel_mul_mat_q2_k_f32(
         for (int i = 16; i < nth; i += 16) sum[0] += sum[i];
         dst[r1*ne0 + r0] = sum[0];
     }
+}
 
-    //// accumulate the sum from all threads in the threadgroup
+kernel void kernel_mul_mat_q3_k_f32(
-    //threadgroup_barrier(mem_flags::mem_threadgroup);
+        device const  void * src0,
-    //for (uint i = nth/2; i > 0; i /= 2) {
+        device const float * src1,
-    //    if (ith < i) {
+        device       float * dst,
-    //        sum[ith] += sum[ith + i];
+        constant   int64_t & ne00,
-    //    }
+        constant   int64_t & ne10,
-    //    threadgroup_barrier(mem_flags::mem_threadgroup);
+        constant   int64_t & ne0,
-    //}
+        constant   int64_t & ne1,
+        threadgroup float  * sum [[threadgroup(0)]],
+        uint2 tgpig[[threadgroup_position_in_grid]],
+        uint2 tpitg[[thread_position_in_threadgroup]],
+        uint2  tptg[[threads_per_threadgroup]]) {
+
+    const uint16_t kmask1 = 0x0303;
+    const uint16_t kmask2 = 0x0f0f;
+
+    const uint8_t m3 = 3;
+    const int8_t  m4 = 4;
+
+    const int nb = ne00/QK_K;
+
+    const int64_t r0 = tgpig.x;
+    const int64_t r1 = tgpig.y;
+
+    device const block_q3_k * x = (device const block_q3_k *) src0 + r0*nb;
+    device const float     * yy = (device const float      *) src1 + r1*ne10;
+
+    const int nth = tptg.x*tptg.y;
+    const int ith = tptg.y*tpitg.x + tpitg.y;
+
+    const int tid = tpitg.y;        // expecting 16
+    const int ip  = tid/8;          // 0 or 1
+    const int il  = tid/2 - 4*ip;   // 0...3
+    const int ir  = tid%2;
+    const int n   = 8;
+    const int l0  = n*ir;
+
+    const uint8_t m = 1 << (4*ip + il);
+
+    const int shift = 2*il;
+
+    const uint16_t s_shift1 = 4*ip;
+    const uint16_t s_shift2 = s_shift1 + 2*(il/2);
+    const int ik = 4 + (il%2);
+
+    const int q_offset = 32*ip + l0;
+    const int y_offset = 128*ip + 32*il + l0;
+
+    //float sumf = 0;
+    float sumf1 = 0, sumf2 = 0;
+    for (int i = tpitg.x; i < nb; i += tptg.x) {
+
+        const float d_all = (float)(x[i].d);
+
+        device const uint8_t * q = x[i].qs + q_offset;
+        device const uint8_t * h = x[i].hmask + l0;
+        device const float   * y = yy + i * QK_K + y_offset;
+
+        device const uint16_t * a = (device const uint16_t *)x[i].scales;
+        const char2 scales = as_type<char2>((uint16_t)(((a[il] >> s_shift1) & kmask2) | (((a[ik] >> s_shift2) & kmask1) << 4)));
+
+        float s = 0;
+        for (int l = 0; l < n; ++l) {
+            s += y[l+ 0] * ((int8_t)((q[l+ 0] >> shift) & m3) - ((h[l+ 0] & m) ? 0 : m4));
+        }
+        float d = d_all * s;
+        sumf1 += d * scales[0];
+        sumf2 += d;
+        //sumf += d_all * s * (scales[0] - 32);
+
+        s = 0;
+        for (int l = 0; l < n; ++l) {
+            s += y[l+16] * ((int8_t)((q[l+16] >> shift) & m3) - ((h[l+16] & m) ? 0 : m4));
+        }
+        d = d_all * s;
+        sumf1 += d * scales[1];
+        sumf2 += d;
+        //sumf += d_all * s * (scales[1] - 32);
+
+    }
+
+    //sum[ith] = sumf;
+    sum[ith] = sumf1 - 32.f*sumf2;
+
+    //
+    // Accumulate the sum from all threads in the threadgroup
+    //
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+    if (ith%4 == 0) {
+        for (int i = 1; i < 4; ++i) sum[ith] += sum[ith + i];
+    }
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+    if (ith%16 == 0) {
+        for (int i = 4; i < 16; i += 4) sum[ith] += sum[ith + i];
+    }
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+    if (ith == 0) {
+        for (int i = 16; i < nth; i += 16) sum[0] += sum[i];
+        dst[r1*ne0 + r0] = sum[0];
+    }
 
-    //if (ith == 0) {
-    //    dst[r1*ne0 + r0] = sum[0];
-    //}
 }
 
 kernel void kernel_mul_mat_q4_k_f32(
@@ -952,23 +1298,17 @@ kernel void kernel_mul_mat_q4_k_f32(
         device const float * src1,
         device       float * dst,
         constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant  uint64_t & nb00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
         constant   int64_t & ne10,
-        constant   int64_t & ne11,
-        constant  uint64_t & nb10,
-        constant  uint64_t & nb11,
-        constant  uint64_t & nb12,
         constant   int64_t & ne0,
-        constant   int64_t & ne1,
         threadgroup float  * sum [[threadgroup(0)]],
         uint2 tgpig[[threadgroup_position_in_grid]],
-        uint2  tpig[[thread_position_in_grid]],               // we don't use this for now
         uint2 tpitg[[thread_position_in_threadgroup]],
         uint2  tptg[[threads_per_threadgroup]]) {
 
+    const uint16_t kmask1 = 0x3f3f;
+    const uint16_t kmask2 = 0x0f0f;
+    const uint16_t kmask3 = 0xc0c0;
+
     const int nb = ne00/QK_K;
 
     const int64_t r0 = tgpig.x;
@@ -977,37 +1317,55 @@ kernel void kernel_mul_mat_q4_k_f32(
     device const block_q4_k * x = (device const block_q4_k *) src0 + r0*nb;
     device const float     * yy = (device const float      *) src1 + r1*ne10;
 
-    const uint nth = tptg.x*tptg.y;
+    const int nth = tptg.x*tptg.y;
-    const uint ith = tptg.y*tpitg.x + tpitg.y;
+    const int ith = tptg.y*tpitg.x + tpitg.y;
 
     const int tid = tpitg.y;   // 0...16
     const int il  = tid/4;     // 0...3
-    const int ir  = tid%4;     // 0...3
+    const int ir  = tid - 4*il;// 0...3
-    const int n   = 8;
+    const int n   = 4;
-    const int is  = 2*il;
+
+    const int im = il/2;  // 0 or 1. 0 computes 0,32 + 128,160, 1 computes 64,96 + 192,224
+    const int in = il%2;
+
+    const int l0 = n*(2*ir + in);
+    const int q_offset = 32*im + l0;
+    const int y_offset = 64*im + l0;
 
     sum[ith] = 0.0f;
 
+    uchar2 sc1, sc2, sc3, sc4;
+
     float sumf = 0;
     for (int i = tpitg.x; i < nb; i += tptg.x) {
 
-        device const uint8_t * q = (x + i)->qs + 32*il + n*ir;
+        device const uint8_t * q1 = (x + i)->qs + q_offset;
-        device const float   * y = yy + i*QK_K + 64*il + n*ir;
+        device const uint8_t * q2 = q1 + 64;
-        device const uint8_t * scales = (x + i)->scales;
+        device const float   * y1 = yy + i*QK_K + y_offset;
+        device const float   * y2 = y1 + 128;
 
         const float dall = (float)((x + i)->d);
         const float dmin = (float)((x + i)->dmin);
 
-        const uchar4 sc = get_scale_min_k4(is, scales);
+        device const uint16_t * a = (device const uint16_t *)(x + i)->scales;
+        sc1 = as_type<uchar2>((uint16_t)(a[im+0] & kmask1));
+        sc2 = as_type<uchar2>((uint16_t)(a[im+2] & kmask1));
+        sc3 = as_type<uchar2>((uint16_t)(((a[im+4] >> 0) & kmask2) | ((a[im+0] & kmask3) >> 2)));
+        sc4 = as_type<uchar2>((uint16_t)(((a[im+4] >> 4) & kmask2) | ((a[im+2] & kmask3) >> 2)));
 
         float4 s = {0.f, 0.f, 0.f, 0.f};
+        float smin = 0;
         for (int l = 0; l < n; ++l) {
-            s[0] += y[l+ 0] * (q[l] & 0xF); s[1] += y[l+ 0];
+
-            s[2] += y[l+32] * (q[l] >>  4); s[3] += y[l+32];
+            s[0] += y1[l] * (q1[l] & 0xF); s[1] += y1[l+32] * (q1[l] >> 4);
-        }
+            s[2] += y2[l] * (q2[l] & 0xF); s[3] += y2[l+32] * (q2[l] >> 4);
-        sumf += dall * (s[0] * sc[0] + s[2] * sc[2]) - dmin * (s[1] * sc[1] + s[3] * sc[3]);
+            smin += y1[l] * sc2[0] + y1[l+32] * sc2[1] + y2[l] * sc4[0] + y2[l+32] * sc4[1];
 
         }
+        sumf += dall * (s[0] * sc1[0] + s[1] * sc1[1] + s[2] * sc3[0] + s[3] * sc3[1]) - dmin * smin;
+
+    }
+
     sum[ith] = sumf;
 
     //
@@ -1043,25 +1401,114 @@ kernel void kernel_mul_mat_q4_k_f32(
     //}
 }
 
+kernel void kernel_mul_mat_q5_k_f32(
+        device const  void * src0,
+        device const float * src1,
+        device       float * dst,
+        constant   int64_t & ne00,
+        constant   int64_t & ne10,
+        constant   int64_t & ne0,
+        threadgroup float  * sum [[threadgroup(0)]],
+        uint2 tgpig[[threadgroup_position_in_grid]],
+        uint2 tpitg[[thread_position_in_threadgroup]],
+        uint2  tptg[[threads_per_threadgroup]]) {
+
+    const uint16_t kmask1 = 0x3f3f;
+    const uint16_t kmask2 = 0x0f0f;
+    const uint16_t kmask3 = 0xc0c0;
+
+    const int nb = ne00/QK_K;
+
+    const int64_t r0 = tgpig.x;
+    const int64_t r1 = tgpig.y;
+
+    device const block_q5_k * x = (device const block_q5_k *) src0 + r0*nb;
+    device const float     * yy = (device const float      *) src1 + r1*ne10;
+
+    const int nth = tptg.x*tptg.y;
+    const int ith = tptg.y*tpitg.x + tpitg.y;
+
+    const int tid = tpitg.y;   // 0...16
+    const int il  = tid/4;     // 0...3
+    const int ir  = tid - 4*il;// 0...3
+    const int n   = 4;
+
+    const int im = il/2;  // 0 or 1. 0 computes 0,32 + 128,160, 1 computes 64,96 + 192,224
+    const int in = il%2;
+
+    const int l0 = n*(2*ir + in);
+    const int q_offset = 32*im + l0;
+    const int y_offset = 64*im + l0;
+
+    const uint8_t hm1 = 1u << (2*im);
+    const uint8_t hm2 = hm1 << 1;
+    const uint8_t hm3 = hm1 << 4;
+    const uint8_t hm4 = hm2 << 4;
+
+    uchar2 sc1, sc2, sc3, sc4;
+
+    float sumf = 0;
+    for (int i = tpitg.x; i < nb; i += tptg.x) {
+
+        device const uint8_t * q1 = (x + i)->qs + q_offset;
+        device const uint8_t * q2 = q1 + 64;
+        device const uint8_t * qh = (x + i)->qh + l0;
+        device const float   * y1 = yy + i*QK_K + y_offset;
+        device const float   * y2 = y1 + 128;
+
+        const float dall = (float)((x + i)->d);
+        const float dmin = (float)((x + i)->dmin);
+
+        device const uint16_t * a = (device const uint16_t *)(x + i)->scales;
+        sc1 = as_type<uchar2>((uint16_t)(a[im+0] & kmask1));
+        sc2 = as_type<uchar2>((uint16_t)(a[im+2] & kmask1));
+        sc3 = as_type<uchar2>((uint16_t)(((a[im+4] >> 0) & kmask2) | ((a[im+0] & kmask3) >> 2)));
+        sc4 = as_type<uchar2>((uint16_t)(((a[im+4] >> 4) & kmask2) | ((a[im+2] & kmask3) >> 2)));
+
+        float4 s = {0.f, 0.f, 0.f, 0.f};
+        float smin = 0;
+        for (int l = 0; l < n; ++l) {
+
+            s[0] += y1[l+ 0] * ((q1[l] & 0xF) + (qh[l] & hm1 ? 16 : 0));
+            s[1] += y1[l+32] * ((q1[l] >>  4) + (qh[l] & hm2 ? 16 : 0));
+            s[2] += y2[l+ 0] * ((q2[l] & 0xF) + (qh[l] & hm3 ? 16 : 0));
+            s[3] += y2[l+32] * ((q2[l] >>  4) + (qh[l] & hm4 ? 16 : 0));
+            smin += y1[l] * sc2[0] + y1[l+32] * sc2[1] + y2[l] * sc4[0] + y2[l+32] * sc4[1];
+
+        }
+        sumf += dall * (s[0] * sc1[0] + s[1] * sc1[1] + s[2] * sc3[0] + s[3] * sc3[1]) - dmin * smin;
+
+    }
+    sum[ith] = sumf;
+
+    //
+    // Accumulate the sum from all threads in the threadgroup
+    //
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+    if (ith%4 == 0) {
+        sum[ith] += sum[ith+1] + sum[ith+2] + sum[ith+3];
+    }
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+    if (ith%16 == 0) {
+        sum[ith] += sum[ith+4] + sum[ith+8] + sum[ith+12];
+    }
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+    if (ith == 0) {
+        for (int i = 16; i < nth; i += 16) sum[0] += sum[i];
+        dst[r1*ne0 + r0] = sum[0];
+    }
+
+}
+
 kernel void kernel_mul_mat_q6_k_f32(
         device const  void * src0,
         device const float * src1,
         device       float * dst,
         constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant  uint64_t & nb00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
         constant   int64_t & ne10,
-        constant   int64_t & ne11,
-        constant  uint64_t & nb10,
-        constant  uint64_t & nb11,
-        constant  uint64_t & nb12,
         constant   int64_t & ne0,
-        constant   int64_t & ne1,
         threadgroup float  * sum [[threadgroup(0)]],
         uint2 tgpig[[threadgroup_position_in_grid]],
-        uint2  tpig[[thread_position_in_grid]],               // we don't use this for now
         uint2 tpitg[[thread_position_in_threadgroup]],
         uint2  tptg[[threads_per_threadgroup]]) {
 
@@ -1078,24 +1525,29 @@ kernel void kernel_mul_mat_q6_k_f32(
     device const block_q6_k * x = (device const block_q6_k *) src0 + r0*nb;
     device const float     * yy = (device const float      *) src1 + r1*ne10;
 
-    const uint nth = tptg.x*tptg.y;
+    const int nth = tptg.x*tptg.y;
-    const uint ith = tptg.y*tpitg.x + tpitg.y;
+    const int ith = tptg.y*tpitg.x + tpitg.y;
 
-    const int step = QK_K / tptg.y;     // we expect this to be 16
+    // Note: we absolutely assume that tptg.y = 16 and QK_K = 256!
-    const int iqs  = step * tpitg.y;    // 0...240 in steps of 16
+    const int iqs  = 16 * tpitg.y;
     const int ip   = iqs / 128;         // 0 or 1
     const int il   = (iqs - 128*ip)/16; // 0...7
     const int n    = 4;
-    const int is   = 8*ip + (n*il)/16;
+    const int l0   = n*il;
+    const int is   = 8*ip + l0/16;
+
+    const int y_offset = 128*ip + l0;
+    const int q_offset_l = 64*ip + l0;
+    const int q_offset_h = 32*ip + l0;
 
     float sumf = 0;
     for (int i = tpitg.x; i < nb; i += tptg.x) {
 
-        device const uint8_t * ql = x[i].ql + 64*ip + n*il;
+        device const uint8_t * ql = x[i].ql + q_offset_l;
-        device const uint8_t * qh = x[i].qh + 32*ip + n*il;
+        device const uint8_t * qh = x[i].qh + q_offset_h;
         device const int8_t  * sc = x[i].scales + is;
 
-        device const float * y = yy + i * QK_K + 128*ip + n*il;
+        device const float * y = yy + i * QK_K + y_offset;
 
         const float dall = x[i].d;
 

ggml-opencl.cpp

@@ -15,7 +15,11 @@
 
 #include "ggml.h"
 
-#define CL_DMMV_BLOCK_SIZE 32;
+#if defined(_MSC_VER)
+#pragma warning(disable: 4244 4267) // possible loss of data
+#endif
+
+#define CL_DMMV_BLOCK_SIZE 32
 
 #define MULTILINE_QUOTE(...) #__VA_ARGS__
 static std::string program_source = MULTILINE_QUOTE(
@@ -59,6 +63,46 @@ struct __attribute__ ((packed)) block_q8_0
     int8_t qs[QK8_0];
 };
 
+struct __attribute__((packed)) block_q2_K
+{
+    uint8_t scales[16];
+    uint8_t qs[64];
+    half d;
+    half dmin;
+};
+
+struct __attribute__((packed)) block_q3_K
+{
+    uint8_t hmask[32];
+    uint8_t qs[64];
+    uint8_t scales[12];
+    half d;
+};
+
+struct __attribute__((packed)) block_q4_K
+{
+    half d;
+    half dmin;
+    uint8_t scales[12];
+    uint8_t qs[128];
+};
+
+struct __attribute__((packed)) block_q5_K
+{
+    half d;
+    half dmin;
+    uint8_t scales[12];
+    uint8_t qh[32];
+    uint8_t qs[128];
+};
+
+struct __attribute__((packed)) block_q6_K
+{
+    uint8_t ql[128];
+    uint8_t qh[64];
+    int8_t scales[16];
+    half d;
+};
 
 __kernel void convert_fp16_to_fp32(__global half* x, __global float* y) {
     const uint i = get_global_id(0);
@@ -131,8 +175,314 @@ void convert_f16(__global half* x, const int ib, const int iqs, float* v0, float
     *v0 = vload_half(0, &x[ib + 0]);
     *v1 = vload_half(0, &x[ib + 1]);
 }
+
+inline void get_scale_min_k4(int j, const __global uint8_t *q, uint8_t *d, uint8_t *m)
+{
+    if (j < 4)
+    {
+        *d = q[j] & 63;
+        *m = q[j + 4] & 63;
+    }
+    else
+    {
+        *d = (q[j + 4] & 0xF) | ((q[j - 4] >> 6) << 4);
+        *m = (q[j + 4] >> 4) | ((q[j - 0] >> 6) << 4);
+    }
+}
+
+__kernel void dequantize_block_q2_K(__global const struct block_q2_K *x, __global float *yy)
+{
+    const int i = get_group_id(0);
+    const int tid = get_local_id(0);
+    const int n = tid / 32;
+    const int l = tid - 32 * n;
+    const int is = 8 * n + l / 16;
+
+    const uint8_t q = x[i].qs[32 * n + l];
+    __global float *y = yy + i * 256 + 128 * n;
+
+    const float dall = vload_half(0, &x[i].d);
+    const float dmin = vload_half(0, &x[i].dmin);
+
+    y[l + 0] = dall * (x[i].scales[is + 0] & 0xF) * ((q >> 0) & 3) - dmin * (x[i].scales[is + 0] >> 4);
+    y[l + 32] = dall * (x[i].scales[is + 2] & 0xF) * ((q >> 2) & 3) - dmin * (x[i].scales[is + 2] >> 4);
+    y[l + 64] = dall * (x[i].scales[is + 4] & 0xF) * ((q >> 4) & 3) - dmin * (x[i].scales[is + 4] >> 4);
+    y[l + 96] = dall * (x[i].scales[is + 6] & 0xF) * ((q >> 6) & 3) - dmin * (x[i].scales[is + 6] >> 4);
+}
+
+__kernel void dequantize_block_q3_K(__global const struct block_q3_K *x, __global float *yy)
+{
+    int r = get_local_id(0) / 4;
+    int i = get_group_id(0);
+    int tid = r / 2;
+    int is0 = r % 2;
+    int l0 = 16 * is0 + 4 * (get_local_id(0) % 4);
+    int n = tid / 4;
+    int j = tid - 4 * n;
+
+    uint8_t m = 1 << (4 * n + j);
+    int is = 8 * n + 2 * j + is0;
+    int shift = 2 * j;
+
+    int8_t us = is < 4 ? (x[i].scales[is - 0] & 0xF) | (((x[i].scales[is + 8] >> 0) & 3) << 4)
+              : is < 8 ? (x[i].scales[is - 0] & 0xF) | (((x[i].scales[is + 4] >> 2) & 3) << 4)
+              : is < 12  ? (x[i].scales[is - 8] >> 4) | (((x[i].scales[is + 0] >> 4) & 3) << 4)
+              : (x[i].scales[is - 8] >> 4) | (((x[i].scales[is - 4] >> 6) & 3) << 4);
+    float d_all = vload_half(0, &x[i].d);
+    float dl = d_all * (us - 32);
+
+    __global float *y = yy + i * 256 + 128 * n + 32 * j;
+    const __global uint8_t *q = x[i].qs + 32 * n;
+    const __global uint8_t *hm = x[i].hmask;
+
+    for (int l = l0; l < l0 + 4; ++l)
+        y[l] = dl * ((int8_t)((q[l] >> shift) & 3) - ((hm[l] & m) ? 0 : 4));
+}
+
+__kernel void dequantize_block_q4_K(__global const struct block_q4_K *x, __global float *yy)
+{
+    const int i = get_group_id(0);
+    const int tid = get_local_id(0);
+    const int il = tid / 8;
+    const int ir = tid % 8;
+    const int is = 2 * il;
+    const int n = 4;
+
+    __global float *y = yy + i * 256 + 64 * il + n * ir;
+
+    const float dall = vload_half(0, &x[i].d);
+    const float dmin = vload_half(0, &x[i].dmin);
+
+    __global const uint8_t *q = x[i].qs + 32 * il + n * ir;
+
+    uint8_t sc, m;
+    get_scale_min_k4(is + 0, x[i].scales, &sc, &m);
+    float d1 = dall * sc;
+    float m1 = dmin * m;
+    get_scale_min_k4(is + 1, x[i].scales, &sc, &m);
+    float d2 = dall * sc;
+    float m2 = dmin * m;
+    for (int l = 0; l < n; ++l)
+    {
+        y[l + 0] = d1 * (q[l] & 0xF) - m1;
+        y[l + 32] = d2 * (q[l] >> 4) - m2;
+    }
+}
+
+__kernel void dequantize_block_q5_K(__global const struct block_q5_K *x, __global float *yy)
+{
+    const int i = get_group_id(0);
+    const int tid = get_local_id(0);
+    const int il = tid / 16;
+    const int ir = tid % 16;
+    const int is = 2 * il;
+
+    __global float *y = yy + i * 256 + 64 * il + 2 * ir;
+
+    const float dall = vload_half(0, &x[i].d);
+    const float dmin = vload_half(0, &x[i].dmin);
+
+    __global const uint8_t *ql = x[i].qs + 32 * il + 2 * ir;
+    __global const uint8_t *qh = x[i].qh + 2 * ir;
+
+    uint8_t sc, m;
+    get_scale_min_k4(is + 0, x[i].scales, &sc, &m);
+    const float d1 = dall * sc;
+    const float m1 = dmin * m;
+    get_scale_min_k4(is + 1, x[i].scales, &sc, &m);
+    const float d2 = dall * sc;
+    const float m2 = dmin * m;
+
+    uint8_t hm = 1 << (2 * il);
+    y[0] = d1 * ((ql[0] & 0xF) + (qh[0] & hm ? 16 : 0)) - m1;
+    y[1] = d1 * ((ql[1] & 0xF) + (qh[1] & hm ? 16 : 0)) - m1;
+    hm <<= 1;
+    y[32] = d2 * ((ql[0] >> 4) + (qh[0] & hm ? 16 : 0)) - m2;
+    y[33] = d2 * ((ql[1] >> 4) + (qh[1] & hm ? 16 : 0)) - m2;
+}
+
+__kernel void dequantize_block_q6_K(__global const struct block_q6_K *x, __global float *yy)
+{
+    const int i = get_group_id(0);
+    const int tid = get_local_id(0);
+    const int ip = tid / 32;
+    const int il = tid - 32 * ip;
+    const int is = 8 * ip + il / 16;
+
+    __global float *y = yy + i * 256 + 128 * ip + il;
+
+    const float d = vload_half(0, &x[i].d);
+
+    __global const uint8_t *ql = x[i].ql + 64 * ip + il;
+    const uint8_t qh = x[i].qh[32 * ip + il];
+    __global const int8_t *sc = x[i].scales + is;
+
+    y[0] = d * sc[0] * ((int8_t)((ql[0] & 0xF) | (((qh >> 0) & 3) << 4)) - 32);
+    y[32] = d * sc[2] * ((int8_t)((ql[32] & 0xF) | (((qh >> 2) & 3) << 4)) - 32);
+    y[64] = d * sc[4] * ((int8_t)((ql[0] >> 4) | (((qh >> 4) & 3) << 4)) - 32);
+    y[96] = d * sc[6] * ((int8_t)((ql[32] >> 4) | (((qh >> 6) & 3) << 4)) - 32);
+}
+
+
+void vec_dot_q2_K(__global const struct block_q2_K* x, const int ib, const int iqs, const __global float *yy, float *result) {
+
+    int n = iqs / 128;
+    int r = iqs - 128 * n;
+    int l = r / 8;
+
+    __global const float *y = yy + 128 * n + l;
+    __global const uint8_t *q = x[ib].qs + 32 * n + l;
+    __global const uint8_t *s = x[ib].scales + 8 * n;
+
+    const float dall = vload_half(0, &x[ib].d);
+    const float dmin = vload_half(0, &x[ib].dmin);
+
+    float sum = y[  0] * (dall * ((s[0] & 0xF) * ((q[ 0] >> 0) & 3)) - dmin * (s[0] >> 4))
+              + y[ 32] * (dall * ((s[2] & 0xF) * ((q[ 0] >> 2) & 3)) - dmin * (s[2] >> 4))
+              + y[ 64] * (dall * ((s[4] & 0xF) * ((q[ 0] >> 4) & 3)) - dmin * (s[4] >> 4))
+              + y[ 96] * (dall * ((s[6] & 0xF) * ((q[ 0] >> 6) & 3)) - dmin * (s[6] >> 4))
+              + y[ 16] * (dall * ((s[1] & 0xF) * ((q[16] >> 0) & 3)) - dmin * (s[1] >> 4))
+              + y[ 48] * (dall * ((s[3] & 0xF) * ((q[16] >> 2) & 3)) - dmin * (s[3] >> 4))
+              + y[ 80] * (dall * ((s[5] & 0xF) * ((q[16] >> 4) & 3)) - dmin * (s[5] >> 4))
+              + y[112] * (dall * ((s[7] & 0xF) * ((q[16] >> 6) & 3)) - dmin * (s[7] >> 4));
+
+    *result = sum;
+}
+
+void vec_dot_q3_K(__global const struct block_q3_K* x, const int ib, const int iqs, const __global float *yy, float *result) {
+
+    const uint32_t kmask1 = 0x03030303;
+    const uint32_t kmask2 = 0x0f0f0f0f;
+
+    uint32_t aux[3];
+    uint32_t utmp[4];
+
+    int n = iqs/128;
+    int r = iqs - 128*n;
+    int l = r/8;
+
+    __global const float   * y = yy + 128*n + l;
+    __global const uint8_t * q = x[ib].qs + 32*n + l;
+    __global const uint8_t * hm = x[ib].hmask + l;
+    const int8_t * s = (const int8_t *)utmp + 8*n;
+
+    aux[0] = x[ib].scales[0] | x[ib].scales[1] << 8 | x[ib].scales[2] << 16 | x[ib].scales[3] << 24;
+    aux[1] = x[ib].scales[4] | x[ib].scales[5] << 8 | x[ib].scales[6] << 16 | x[ib].scales[7] << 24;
+    aux[2] = x[ib].scales[8] | x[ib].scales[9] << 8 | x[ib].scales[10] << 16 | x[ib].scales[11] << 24;
+
+    utmp[3] = ((aux[1] >> 4) & kmask2) | (((aux[2] >> 6) & kmask1) << 4);
+    utmp[2] = ((aux[0] >> 4) & kmask2) | (((aux[2] >> 4) & kmask1) << 4);
+    utmp[1] = (aux[1] & kmask2) | (((aux[2] >> 2) & kmask1) << 4);
+    utmp[0] = (aux[0] & kmask2) | (((aux[2] >> 0) & kmask1) << 4);
+
+    const float dall = vload_half(0, &x[ib].d);
+    const uint8_t m = 1 << (4*n);
+
+    float sum = y[  0] * (s[0] - 32) * (((q[ 0] >> 0) & 3) - (hm[ 0] & (m << 0) ? 0 : 4))
+              + y[ 32] * (s[2] - 32) * (((q[ 0] >> 2) & 3) - (hm[ 0] & (m << 1) ? 0 : 4))
+              + y[ 64] * (s[4] - 32) * (((q[ 0] >> 4) & 3) - (hm[ 0] & (m << 2) ? 0 : 4))
+              + y[ 96] * (s[6] - 32) * (((q[ 0] >> 6) & 3) - (hm[ 0] & (m << 3) ? 0 : 4))
+              + y[ 16] * (s[1] - 32) * (((q[16] >> 0) & 3) - (hm[16] & (m << 0) ? 0 : 4))
+              + y[ 48] * (s[3] - 32) * (((q[16] >> 2) & 3) - (hm[16] & (m << 1) ? 0 : 4))
+              + y[ 80] * (s[5] - 32) * (((q[16] >> 4) & 3) - (hm[16] & (m << 2) ? 0 : 4))
+              + y[112] * (s[7] - 32) * (((q[16] >> 6) & 3) - (hm[16] & (m << 3) ? 0 : 4));
+
+    *result = sum * dall;
+
+}
+
+void vec_dot_q4_K(__global const struct block_q4_K* x, const int ib, const int iqs, const __global float *yy, float *result) {
+
+    const int j  = iqs / 64;        // j  is in 0...3
+    const int ir = (iqs - 64*j)/2;  // ir is in 0...28 in steps of 4
+    const int is = 2*j;             // is is in 0...6 in steps of 2
+
+    __global const float   * y = yy + 64*j + ir;
+    __global const uint8_t * q = x[ib].qs + 32*j + ir;
+
+    const float dall = vload_half(0, &x[ib].d);
+    const float dmin = vload_half(0, &x[ib].dmin);
+
+    uint8_t sc, m;
+    get_scale_min_k4(is + 0, x[ib].scales, &sc, &m);
+    const float d1 = dall * sc;
+    const float m1 = dmin * m;
+    get_scale_min_k4(is + 1, x[ib].scales, &sc, &m);
+    const float d2 = dall * sc;
+    const float m2 = dmin * m;
+
+    float sum = 0;
+    for (int k = 0; k < 4; ++k) {
+        sum += y[k +  0] * (d1 * (q[k] & 0xF) - m1);
+        sum += y[k + 32] * (d2 * (q[k] >>  4) - m2);
+    }
+
+    *result = sum;
+}
+
+void vec_dot_q5_K(__global const struct block_q5_K* x, const int ib, const int iqs, const __global float *yy, float *result) {
+
+    const int j  = iqs / 64;
+    const int ir = (iqs - 64*j)/2;
+    const int is = 2*j;
+
+    __global const float   * y  = yy + 64*j + ir;
+    __global const uint8_t * ql = x[ib].qs + 32*j + ir;
+    __global const uint8_t * qh = x[ib].qh + ir;
+
+    const float dall = vload_half(0, &x[ib].d);
+    const float dmin = vload_half(0, &x[ib].dmin);
+
+    uint8_t sc, m;
+    get_scale_min_k4(is + 0, x[ib].scales, &sc, &m);
+    const float d1 = dall * sc;
+    const float m1 = dmin * m;
+    get_scale_min_k4(is + 1, x[ib].scales, &sc, &m);
+    const float d2 = dall * sc;
+    const float m2 = dmin * m;
+
+    uint8_t hm  = 1 << is;
+    float sum = 0;
+    for (int k = 0; k < 4; ++k) {
+        sum += y[k +  0] * (d1 * ((ql[k] & 0xF) + (qh[k] & hm ? 16 : 0)) - m1);
+    }
+    hm <<= 1;
+    for (int k = 0; k < 4; ++k) {
+        sum += y[k + 32] * (d2 * ((ql[k] >>  4) + (qh[k] & hm ? 16 : 0)) - m2);
+    }
+    *result = sum;
+
+}
+
+void vec_dot_q6_K(__global const struct block_q6_K* x, const int ib, const int iqs, const __global float *yy, float *result) {
+
+
+    const int ip = iqs / 128;        // 0 or 1
+    const int il = (iqs - 128*ip)/8; // 0...15
+    const int is = 8*ip;
+
+    __global const float * y = yy + 128*ip + il;
+
+    const float d = vload_half(0, &x[ib].d);
+
+    __global const uint8_t * ql = x[ib].ql + 64*ip + il;
+    __global const uint8_t * qh = x[ib].qh + 32*ip + il;
+    __global const int8_t  * sc = x[ib].scales + is;
+
+    *result = y[  0] * d * sc[0] * ((int8_t)((ql[ 0] & 0xF) | (((qh[ 0] >> 0) & 3) << 4)) - 32)
+           + y[ 32] * d * sc[2] * ((int8_t)((ql[32] & 0xF) | (((qh[ 0] >> 2) & 3) << 4)) - 32)
+           + y[ 64] * d * sc[4] * ((int8_t)((ql[ 0]  >> 4) | (((qh[ 0] >> 4) & 3) << 4)) - 32)
+           + y[ 96] * d * sc[6] * ((int8_t)((ql[32]  >> 4) | (((qh[ 0] >> 6) & 3) << 4)) - 32)
+           + y[ 16] * d * sc[1] * ((int8_t)((ql[16] & 0xF) | (((qh[16] >> 0) & 3) << 4)) - 32)
+           + y[ 48] * d * sc[3] * ((int8_t)((ql[48] & 0xF) | (((qh[16] >> 2) & 3) << 4)) - 32)
+           + y[ 80] * d * sc[5] * ((int8_t)((ql[16]  >> 4) | (((qh[16] >> 4) & 3) << 4)) - 32)
+           + y[112] * d * sc[7] * ((int8_t)((ql[48]  >> 4) | (((qh[16] >> 6) & 3) << 4)) - 32);
+
+}
+
 );
 
+
 std::string dequant_template = MULTILINE_QUOTE(
 __kernel void KERNEL_NAME(__global X_TYPE* x, __global float* y) {
     const int i = get_group_id(0)*get_local_size(0) + get_local_id(0)*2;
@@ -160,7 +510,7 @@ __kernel void KERNEL_NAME(__global X_TYPE* x, __global float* y) {
 std::string dequant_mul_mat_vec_template = MULTILINE_QUOTE(
 __kernel void KERNEL_NAME(__global X_TYPE* x, __local float* tmp, __global float* y, __global float* dst, const int ncols) {
     const int block_size = get_local_size(0);
-    const int row = get_global_id(0) / block_size;
+    const int row = get_group_id(0);
     const int tid = get_local_id(0);
 
     const uint qk = QUANT_K;
@@ -199,6 +549,45 @@ __kernel void KERNEL_NAME(__global X_TYPE* x, __local float* tmp, __global float
 }
 );
 
+std::string dequant_mul_mat_vec_k_template = MULTILINE_QUOTE(
+__kernel void KERNEL_NAME(__global X_TYPE* x, __local float* tmp, __global float* y, __global float* dst, const int ncols) {
+    const int block_size = get_local_size(0);
+    const int row = get_group_id(0);
+    const int tid = get_local_id(0);
+
+    const int iter_stride = 256;
+    const int vals_per_iter = iter_stride / block_size;
+    const int num_blocks_per_row = ncols / 256;
+    const int ib0 = row*num_blocks_per_row;
+
+    tmp[tid] = 0;
+
+    for (int i = 0; i < ncols; i += iter_stride) {
+        const int col = i + vals_per_iter*tid;
+        const int ib = ib0 + col/256; // x block index
+        const int iqs = col%256; // x quant index
+        const int iybs = col - col%256; // y block start index
+
+        // dequantize
+        float v;
+        DOT_KERNEL(x, ib, iqs, y + iybs, &v);
+        tmp[tid] += v;
+    }
+
+    // sum up partial sums and write back result
+    barrier(CLK_LOCAL_MEM_FENCE);
+    for (int s=block_size/2; s>0; s>>=1) {
+        if (tid < s) {
+            tmp[tid] += tmp[tid + s];
+        }
+        barrier(CLK_LOCAL_MEM_FENCE);
+    }
+    if (tid == 0) {
+        dst[row] = tmp[0];
+    }
+}
+);
+
 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);
@@ -260,6 +649,18 @@ std::array<std::string, 2> mul_str_values = {
     "mul_f32", "float"
 };
 
+std::array<std::string, 3> dmmv_k_str_keys = {
+    "KERNEL_NAME", "X_TYPE", "DOT_KERNEL"
+};
+
+std::array<std::string, 15> dmmv_k_str_values = {
+    "dequantize_mul_mat_vec_q2_K", "struct block_q2_K", "vec_dot_q2_K",
+    "dequantize_mul_mat_vec_q3_K", "struct block_q3_K", "vec_dot_q3_K",
+    "dequantize_mul_mat_vec_q4_K", "struct block_q4_K", "vec_dot_q4_K",
+    "dequantize_mul_mat_vec_q5_K", "struct block_q5_K", "vec_dot_q5_K",
+    "dequantize_mul_mat_vec_q6_K", "struct block_q6_K", "vec_dot_q6_K",
+};
+
 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) {
@@ -289,6 +690,14 @@ std::string generate_kernels() {
         }
         src << mul_kernel << '\n';
     }
+    for (size_t i = 0; i < dmmv_k_str_values.size(); i += dmmv_k_str_keys.size()) {
+        std::string dmmv_k_kernel = dequant_mul_mat_vec_k_template;
+        for (size_t j = 0; j < dmmv_k_str_keys.size(); j++) {
+            replace(dmmv_k_kernel, dmmv_k_str_keys[j], dmmv_k_str_values[i + j]);
+        }
+        src << dmmv_k_kernel << '\n';
+    }
+
     return src.str();
 }
 
@@ -300,6 +709,8 @@ 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 dequantize_block_q2_k_cl, dequantize_block_q3_k_cl, dequantize_block_q4_k_cl, dequantize_block_q5_k_cl, dequantize_block_q6_k_cl;
+static cl_kernel dequantize_mul_mat_vec_q2_K_cl, dequantize_mul_mat_vec_q3_K_cl, dequantize_mul_mat_vec_q4_K_cl, dequantize_mul_mat_vec_q5_K_cl, dequantize_mul_mat_vec_q6_K_cl;
 static cl_kernel mul_f32_cl;
 static bool fp16_support;
 
@@ -529,6 +940,12 @@ void ggml_cl_init(void) {
     CL_CHECK((dequantize_row_q5_0_cl = clCreateKernel(program, "dequantize_row_q5_0", &err), err));
     CL_CHECK((dequantize_row_q5_1_cl = clCreateKernel(program, "dequantize_row_q5_1", &err), err));
     CL_CHECK((dequantize_row_q8_0_cl = clCreateKernel(program, "dequantize_row_q8_0", &err), err));
+    CL_CHECK((dequantize_row_q8_0_cl = clCreateKernel(program, "dequantize_row_q8_0", &err), err));
+    CL_CHECK((dequantize_block_q2_k_cl = clCreateKernel(program, "dequantize_block_q2_K", &err), err));
+    CL_CHECK((dequantize_block_q3_k_cl = clCreateKernel(program, "dequantize_block_q3_K", &err), err));
+    CL_CHECK((dequantize_block_q4_k_cl = clCreateKernel(program, "dequantize_block_q4_K", &err), err));
+    CL_CHECK((dequantize_block_q5_k_cl = clCreateKernel(program, "dequantize_block_q5_K", &err), err));
+    CL_CHECK((dequantize_block_q6_k_cl = clCreateKernel(program, "dequantize_block_q6_K", &err), err));
 
     // dequant mul mat kernel
     CL_CHECK((dequantize_mul_mat_vec_q4_0_cl = clCreateKernel(program, "dequantize_mul_mat_vec_q4_0", &err), err));
@@ -537,6 +954,11 @@ 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));
+    CL_CHECK((dequantize_mul_mat_vec_q2_K_cl = clCreateKernel(program, "dequantize_mul_mat_vec_q2_K", &err), err));
+    CL_CHECK((dequantize_mul_mat_vec_q3_K_cl = clCreateKernel(program, "dequantize_mul_mat_vec_q3_K", &err), err));
+    CL_CHECK((dequantize_mul_mat_vec_q4_K_cl = clCreateKernel(program, "dequantize_mul_mat_vec_q4_K", &err), err));
+    CL_CHECK((dequantize_mul_mat_vec_q5_K_cl = clCreateKernel(program, "dequantize_mul_mat_vec_q5_K", &err), err));
+    CL_CHECK((dequantize_mul_mat_vec_q6_K_cl = clCreateKernel(program, "dequantize_mul_mat_vec_q6_K", &err), err));
 
     // mul kernel
     CL_CHECK((mul_f32_cl = clCreateKernel(program, "mul_f32", &err), err));
@@ -554,6 +976,16 @@ static cl_kernel* ggml_get_to_fp32_cl(ggml_type type) {
             return &dequantize_row_q5_1_cl;
         case GGML_TYPE_Q8_0:
             return &dequantize_row_q8_0_cl;
+        case GGML_TYPE_Q2_K:
+            return &dequantize_block_q2_k_cl;
+        case GGML_TYPE_Q3_K:
+            return &dequantize_block_q3_k_cl;
+        case GGML_TYPE_Q4_K:
+            return &dequantize_block_q4_k_cl;
+        case GGML_TYPE_Q5_K:
+            return &dequantize_block_q5_k_cl;
+        case GGML_TYPE_Q6_K:
+            return &dequantize_block_q6_k_cl;
         case GGML_TYPE_F16:
             return &convert_row_f16_cl;
         default:
@@ -561,6 +993,50 @@ static cl_kernel* ggml_get_to_fp32_cl(ggml_type type) {
     }
 }
 
+static size_t ggml_cl_global_denom(ggml_type type) {
+    switch (type) {
+        case GGML_TYPE_Q4_0:
+        case GGML_TYPE_Q4_1:
+        case GGML_TYPE_Q5_0:
+        case GGML_TYPE_Q5_1:
+        case GGML_TYPE_Q8_0:
+            return 1;
+        case GGML_TYPE_Q2_K:
+        case GGML_TYPE_Q3_K:
+            return 4;
+        case GGML_TYPE_Q4_K:
+            return 8;
+        case GGML_TYPE_Q5_K:
+        case GGML_TYPE_Q6_K:
+            return 4;
+        case GGML_TYPE_F16:
+        default:
+            return 1;
+    }
+}
+
+static size_t ggml_cl_local_size(ggml_type type) {
+    switch (type) {
+        case GGML_TYPE_Q4_0:
+        case GGML_TYPE_Q4_1:
+        case GGML_TYPE_Q5_0:
+        case GGML_TYPE_Q5_1:
+        case GGML_TYPE_Q8_0:
+            return 0;
+        case GGML_TYPE_Q2_K:
+        case GGML_TYPE_Q3_K:
+            return 64;
+        case GGML_TYPE_Q4_K:
+            return 32;
+        case GGML_TYPE_Q5_K:
+        case GGML_TYPE_Q6_K:
+            return 64;
+        case GGML_TYPE_F16:
+        default:
+            return 0;
+    }
+}
+
 static cl_kernel* ggml_get_dequantize_mul_mat_vec_cl(ggml_type type) {
     switch (type) {
         case GGML_TYPE_Q4_0:
@@ -575,6 +1051,16 @@ static cl_kernel* ggml_get_dequantize_mul_mat_vec_cl(ggml_type type) {
             return &dequantize_mul_mat_vec_q8_0_cl;
         case GGML_TYPE_F16:
             return &convert_mul_mat_vec_f16_cl;
+        case GGML_TYPE_Q2_K:
+            return &dequantize_mul_mat_vec_q2_K_cl;
+        case GGML_TYPE_Q3_K:
+            return &dequantize_mul_mat_vec_q3_K_cl;
+        case GGML_TYPE_Q4_K:
+            return &dequantize_mul_mat_vec_q4_K_cl;
+        case GGML_TYPE_Q5_K:
+            return &dequantize_mul_mat_vec_q5_K_cl;
+        case GGML_TYPE_Q6_K:
+            return &dequantize_mul_mat_vec_q6_K_cl;
         default:
             return nullptr;
     }
@@ -1017,6 +1503,9 @@ static void ggml_cl_mul_mat_q_f32(const ggml_tensor * src0, const ggml_tensor *
     cl_kernel* dmmv = ggml_get_dequantize_mul_mat_vec_cl(type);
     GGML_ASSERT(to_fp32_cl != nullptr);
 
+    const size_t global_denom = ggml_cl_global_denom(type);
+    const size_t local = ggml_cl_local_size(type);
+
     size_t ev_idx = 0;
     std::vector<cl_event> events;
 
@@ -1049,10 +1538,10 @@ static void ggml_cl_mul_mat_q_f32(const ggml_tensor * src0, const ggml_tensor *
                 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;
+                const size_t global = x_ne / global_denom;
                 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(clEnqueueNDRangeKernel(queue, *to_fp32_cl, 1, NULL, &global, NULL, events.size(), !events.empty() ? events.data() : NULL, NULL));
+                CL_CHECK(clEnqueueNDRangeKernel(queue, *to_fp32_cl, 1, NULL, &global, local > 0 ? &local : 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));
@@ -1167,7 +1656,7 @@ 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];
@@ -1179,6 +1668,7 @@ void ggml_cl_transform_tensor(ggml_tensor * tensor) {
     size_t q_size;
     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++) {
@@ -1190,35 +1680,5 @@ void ggml_cl_transform_tensor(ggml_tensor * tensor) {
     CL_CHECK(clFinish(queue));
 
     tensor->data = dst;
-    tensor->backend = GGML_BACKEND_GPU;
+    GGML_ASSERT(tensor->backend == GGML_BACKEND_GPU);
-}
-
-void ggml_cl_load_data(const char * fname, struct ggml_tensor * tensor, const size_t offset) {
-    cl_int err;
-    FILE * fp = fopen(fname, "rb");
-
-    const size_t size = ggml_nbytes(tensor);
-
-    cl_mem dst;
-    CL_CHECK((dst = clCreateBuffer(context, CL_MEM_READ_ONLY, size, nullptr, &err), err));
-    void * buf_host = malloc(size);
-
-#ifdef _WIN32
-    int ret = _fseeki64(fp, (__int64) offset, SEEK_SET);
-#else
-    int ret = fseek(fp, (long) offset, SEEK_SET);
-#endif
-    GGML_ASSERT(ret == 0); // same
-
-    size_t ret2 = fread(buf_host, size, 1, fp);
-    if (ret2 != 1) {
-        fprintf(stderr, "unexpectedly reached end of file");
-        exit(1);
-    }
-
-    clEnqueueWriteBuffer(queue, dst, CL_TRUE, 0, size, buf_host, 0, nullptr, nullptr);
-
-    tensor->data = dst;
-    free(buf_host);
-    fclose(fp);
 }

ggml-opencl.h

@@ -18,8 +18,7 @@ void   ggml_cl_host_free(void * ptr);
 
 void ggml_cl_free_data(const struct ggml_tensor* tensor);
 
-void ggml_cl_transform_tensor(struct ggml_tensor * tensor);
+void ggml_cl_transform_tensor(void * data, struct ggml_tensor * tensor);
-void ggml_cl_load_data(const char * fname, struct ggml_tensor * tensor, size_t offset);
 
 #ifdef  __cplusplus
 }

ggml.h

@@ -296,6 +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,
@@ -309,6 +310,7 @@ extern "C" {
         GGML_OP_RMS_NORM_BACK,
 
         GGML_OP_MUL_MAT,
+        GGML_OP_OUT_PROD,
 
         GGML_OP_SCALE,
         GGML_OP_SET,
@@ -324,6 +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,
@@ -333,10 +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,
     };
 
@@ -478,6 +485,7 @@ extern "C" {
 
     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);
@@ -492,8 +500,9 @@ extern "C" {
     GGML_API size_t  ggml_set_scratch (struct ggml_context * ctx, struct ggml_scratch scratch);
     GGML_API void    ggml_set_no_alloc(struct ggml_context * ctx, bool no_alloc);
 
-    GGML_API void *  ggml_get_mem_buffer(struct ggml_context * ctx);
+    GGML_API void *  ggml_get_mem_buffer     (const struct ggml_context * ctx);
-    GGML_API size_t  ggml_get_mem_size  (struct ggml_context * ctx);
+    GGML_API size_t  ggml_get_mem_size       (const struct ggml_context * ctx);
+    GGML_API size_t  ggml_get_max_tensor_size(const struct ggml_context * ctx);
 
     GGML_API struct ggml_tensor * ggml_new_tensor(
             struct ggml_context * ctx,
@@ -574,6 +583,11 @@ 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,
@@ -645,6 +659,11 @@ 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);
@@ -698,14 +717,22 @@ extern "C" {
             struct ggml_tensor  * a,
             struct ggml_tensor  * b);
 
-    // A: m rows, n columns
+    // A: n columns, m rows
-    // B: p rows, n columns (i.e. we transpose it internally)
+    // 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
     //
@@ -916,6 +943,17 @@ 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
@@ -982,6 +1020,14 @@ 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,
@@ -1005,6 +1051,19 @@ 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
     //
@@ -1099,6 +1158,8 @@ 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;
@@ -1123,6 +1184,49 @@ 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
@@ -1131,6 +1235,27 @@ 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
     //

llama.cpp

@@ -19,6 +19,11 @@
 #ifdef GGML_USE_METAL
 #include "ggml-metal.h"
 #endif
+#ifdef GGML_USE_K_QUANTS
+#ifndef QK_K
+#define QK_K 256
+#endif
+#endif
 
 #include <array>
 #include <ctime>
@@ -40,6 +45,10 @@
 #include <sstream>
 #include <numeric>
 
+#if defined(_MSC_VER)
+#pragma warning(disable: 4244 4267) // possible loss of data
+#endif
+
 #define LLAMA_USE_SCRATCH
 #define LLAMA_MAX_SCRATCH_BUFFERS 16
 
@@ -165,6 +174,11 @@ struct llama_kv_cache {
         if (ctx) {
             ggml_free(ctx);
         }
+
+#ifdef GGML_USE_CUBLAS
+        ggml_cuda_free_data(k);
+        ggml_cuda_free_data(v);
+#endif // GGML_USE_CUBLAS
     }
 };
 
@@ -224,6 +238,7 @@ struct llama_model {
         for (size_t i = 0; i < tensors_by_name.size(); ++i) {
             ggml_cuda_free_data(tensors_by_name[i].second);
         }
+        ggml_cuda_free_scratch();
 #elif defined(GGML_USE_CLBLAST)
         for (size_t i = 0; i < tensors_by_name.size(); ++i) {
             ggml_cl_free_data(tensors_by_name[i].second);
@@ -716,6 +731,9 @@ struct llama_model_loader {
 
     struct ggml_tensor * get_tensor_for(llama_load_tensor & lt, ggml_backend backend) {
         struct ggml_tensor * tensor;
+        if (backend != GGML_BACKEND_CPU) {
+            ggml_set_no_alloc(ggml_ctx, true);
+        }
         if (lt.ne.size() == 2) {
             tensor = ggml_new_tensor_2d(ggml_ctx, lt.type, lt.ne.at(0), lt.ne.at(1));
         } else {
@@ -725,6 +743,9 @@ struct llama_model_loader {
         ggml_set_name(tensor, lt.name.c_str());
         LLAMA_ASSERT(lt.ggml_tensor == NULL); // if this fails, we called get_tensor twice on the same tensor
 
+        if (backend != GGML_BACKEND_CPU) {
+            ggml_set_no_alloc(ggml_ctx, use_mmap);
+        }
         tensor->backend = backend;
         lt.ggml_tensor = tensor;
         num_ggml_tensors_created++;
@@ -740,6 +761,7 @@ struct llama_model_loader {
     void load_all_data(llama_progress_callback progress_callback, void *  progress_callback_user_data, llama_mlock * lmlock) {
         size_t data_size = 0;
         size_t prefetch_size = 0;
+        size_t lock_size = 0;
         for (const llama_load_tensor & lt : tensors_map.tensors) {
             data_size += lt.size;
             if (lt.ggml_tensor->backend == GGML_BACKEND_CPU) {
@@ -749,11 +771,6 @@ struct llama_model_loader {
 
         if (use_mmap) {
             mapping.reset(new llama_mmap(&file_loaders.at(0)->file, prefetch_size));
-            if (!lmlock) {
-                // Don't call the callback since the actual loading will be lazy
-                // and we can't measure it.
-                progress_callback = NULL;
-            }
             if (lmlock) {
                 lmlock->init(mapping->addr);
             }
@@ -761,20 +778,49 @@ struct llama_model_loader {
 
         size_t done_size = 0;
         for (llama_load_tensor & lt : tensors_map.tensors) {
-            if (lt.ggml_tensor->backend != GGML_BACKEND_CPU) {
-                continue;
-            }
             if (progress_callback) {
                 progress_callback((float) done_size / data_size, progress_callback_user_data);
             }
             LLAMA_ASSERT(lt.ggml_tensor); // unused tensors should have been caught by load_data already
             lt.data = (uint8_t *) lt.ggml_tensor->data;
-            load_data_for(lt);
+
-            lt.ggml_tensor->data = lt.data;
+            // allocate temp buffer if not using mmap
-            done_size += lt.size;
+            if (!use_mmap && lt.data == NULL) {
-            if (use_mmap && lmlock) {
+                GGML_ASSERT(lt.ggml_tensor->backend != GGML_BACKEND_CPU);
-                lmlock->grow_to(done_size);
+                lt.data = (uint8_t*)malloc(ggml_nbytes(lt.ggml_tensor));
             }
+
+            load_data_for(lt);
+
+            switch(lt.ggml_tensor->backend) {
+                case GGML_BACKEND_CPU:
+                    lt.ggml_tensor->data = lt.data;
+                    if (use_mmap && lmlock) {
+                        lock_size += lt.size;
+                        lmlock->grow_to(lock_size);
+                    }
+                    break;
+#if defined(GGML_USE_CUBLAS)
+                case GGML_BACKEND_GPU:
+                case GGML_BACKEND_GPU_SPLIT:
+                    ggml_cuda_transform_tensor(lt.data, lt.ggml_tensor);
+                    if (!use_mmap) {
+                        free(lt.data);
+                    }
+                    break;
+#elif defined(GGML_USE_CLBLAST)
+                case GGML_BACKEND_GPU:
+                    ggml_cl_transform_tensor(lt.data, lt.ggml_tensor);
+                    if (!use_mmap) {
+                        free(lt.data);
+                    }
+                    break;
+#endif
+                default:
+                    continue;
+            }
+
+            done_size += lt.size;
         }
     }
 
@@ -845,7 +891,8 @@ static bool kv_cache_init(
         const struct llama_hparams & hparams,
              struct llama_kv_cache & cache,
                          ggml_type   wtype,
-                               int   n_ctx) {
+                               int   n_ctx,
+                               int   n_gpu_layers) {
     const int n_embd  = hparams.n_embd;
     const int n_layer = hparams.n_layer;
 
@@ -853,6 +900,7 @@ static bool kv_cache_init(
     const int64_t n_elements = n_embd*n_mem;
 
     cache.buf.resize(2u*n_elements*ggml_type_size(wtype) + 2u*MB);
+    cache.n = 0;
 
     struct ggml_init_params params;
     params.mem_size   = cache.buf.size;
@@ -871,6 +919,16 @@ static bool kv_cache_init(
     ggml_set_name(cache.k, "cache_k");
     ggml_set_name(cache.v, "cache_v");
 
+    (void) n_gpu_layers;
+#ifdef GGML_USE_CUBLAS
+    if (n_gpu_layers > n_layer + 1) {
+        ggml_cuda_assign_buffers_no_scratch(cache.v);
+    }
+    if (n_gpu_layers > n_layer + 2) {
+        ggml_cuda_assign_buffers_no_scratch(cache.k);
+    }
+#endif // GGML_USE_CUBLAS
+
     return true;
 }
 
@@ -881,6 +939,7 @@ struct llama_context_params llama_context_default_params() {
         /*.gpu_layers                  =*/ 0,
         /*.main_gpu                    =*/ 0,
         /*.tensor_split                =*/ {0},
+        /*.low_vram                    =*/ false,
         /*.seed                        =*/ -1,
         /*.f16_kv                      =*/ true,
         /*.logits_all                  =*/ false,
@@ -990,6 +1049,7 @@ static void llama_model_load_internal(
         int n_gpu_layers,
         int main_gpu,
         const float * tensor_split,
+        bool low_vram,
         ggml_type memory_type,
         bool use_mmap,
         bool use_mlock,
@@ -1014,6 +1074,12 @@ static void llama_model_load_internal(
             case 40: model.type = e_model::MODEL_13B; break;
             case 60: model.type = e_model::MODEL_30B; break;
             case 80: model.type = e_model::MODEL_65B; break;
+            default:
+                {
+                    if (hparams.n_layer < 32) {
+                        model.type = e_model::MODEL_7B;
+                    }
+                } break;
         }
 
         hparams.n_ctx = n_ctx;
@@ -1109,18 +1175,34 @@ static void llama_model_load_internal(
         ml->ggml_ctx = ctx;
 
         model.tok_embeddings = ml->get_tensor("tok_embeddings.weight", {n_embd, n_vocab}, GGML_BACKEND_CPU);
-        model.norm           = ml->get_tensor("norm.weight",           {n_embd},          GGML_BACKEND_CPU);
 
         // "output" tensor
         {
+            ggml_backend backend_norm;
             ggml_backend backend_output;
             if (n_gpu_layers > int(n_layer)) { // NOLINT
+                // norm is not performance relevant on its own but keeping it in VRAM reduces data copying
+                // on Windows however this is detrimental unless everything is on the GPU
+#ifndef _WIN32
+                backend_norm = low_vram ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD;
+#else
+                backend_norm = low_vram || n_gpu_layers <= (int) n_layer + 2 ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD;
+#endif // _WIN32
+
                 backend_output = LLAMA_BACKEND_OFFLOAD_SPLIT;
             } else {
+                backend_norm = GGML_BACKEND_CPU;
                 backend_output = GGML_BACKEND_CPU;
             }
 
+            model.norm   = ml->get_tensor("norm.weight",   {n_embd},          backend_norm);
             model.output = ml->get_tensor("output.weight", {n_embd, n_vocab}, backend_output);
+            if (backend_norm == GGML_BACKEND_GPU) {
+                vram_weights += ggml_nbytes(model.norm);
+            }
+            if (backend_output == GGML_BACKEND_GPU_SPLIT) {
+                vram_weights += ggml_nbytes(model.output);
+            }
         }
 
         const int i_gpu_start = n_layer - n_gpu_layers;
@@ -1150,7 +1232,7 @@ static void llama_model_load_internal(
             if (backend == GGML_BACKEND_GPU) {
                 vram_weights +=
                     ggml_nbytes(layer.attention_norm) + ggml_nbytes(layer.wq) + ggml_nbytes(layer.wk)             +
-                    ggml_nbytes(layer.wv)             + ggml_nbytes(layer.wo) + ggml_nbytes(layer.attention_norm) +
+                    ggml_nbytes(layer.wv)             + ggml_nbytes(layer.wo) + ggml_nbytes(layer.ffn_norm) +
                     ggml_nbytes(layer.w1)             + ggml_nbytes(layer.w2) + ggml_nbytes(layer.w3);
             }
         }
@@ -1178,23 +1260,49 @@ static void llama_model_load_internal(
                 mem_required / 1024.0 / 1024.0, mem_required_state / 1024.0 / 1024.0);
 
         (void) vram_scratch;
+        (void) n_batch;
 #ifdef GGML_USE_CUBLAS
+        if (low_vram) {
+            fprintf(stderr, "%s: not allocating a VRAM scratch buffer due to low VRAM option\n", __func__);
+            ggml_cuda_set_scratch_size(0); // disable scratch
+        } else {
             vram_scratch = n_batch * MB;
             ggml_cuda_set_scratch_size(vram_scratch);
             if (n_gpu_layers > 0) {
-            fprintf(stderr, "%s: allocating batch_size x 1 MB = %ld MB VRAM for the scratch buffer\n",
+                fprintf(stderr, "%s: allocating batch_size x 1 MB = %zd MB VRAM for the scratch buffer\n",
                         __func__, vram_scratch / MB);
             }
+        }
 #endif // GGML_USE_CUBLAS
 #if defined(GGML_USE_CUBLAS) || defined(GGML_USE_CLBLAST)
         const int n_gpu = std::min(n_gpu_layers, int(hparams.n_layer));
 
-        fprintf(stderr, "%s: offloading %d layers to GPU\n", __func__, n_gpu);
+        fprintf(stderr, "%s: offloading %d repeating layers to GPU\n", __func__, n_gpu);
         if (n_gpu_layers > (int) hparams.n_layer) {
-            fprintf(stderr, "%s: offloading output layer to GPU\n", __func__);
+            fprintf(stderr, "%s: offloading non-repeating layers to GPU\n", __func__);
         }
+        size_t vram_kv_cache = 0;
+        if (n_gpu_layers > (int) hparams.n_layer + 1) {
+            if (low_vram) {
+                fprintf(stderr, "%s: cannot offload v cache to GPU due to low VRAM option\n", __func__);
+            } else {
+                fprintf(stderr, "%s: offloading v cache to GPU\n", __func__);
+                vram_kv_cache += MEM_REQ_KV_SELF().at(model.type) / 2;
+            }
+        }
+        if (n_gpu_layers > (int) hparams.n_layer + 2) {
+            if (low_vram) {
+                fprintf(stderr, "%s: cannot offload k cache to GPU due to low VRAM option\n", __func__);
+            } else {
+                fprintf(stderr, "%s: offloading k cache to GPU\n", __func__);
+                vram_kv_cache += MEM_REQ_KV_SELF().at(model.type) / 2;
+            }
+        }
+        const int max_offloadable_layers = low_vram ? hparams.n_layer + 1 : hparams.n_layer + 3;
+        fprintf(stderr, "%s: offloaded %d/%d layers to GPU\n",
+                __func__, std::min(n_gpu_layers, max_offloadable_layers), hparams.n_layer + 3);
         fprintf(stderr, "%s: total VRAM used: %zu MB\n",
-                __func__, (vram_weights + vram_scratch + MB - 1) / MB); // round up
+                __func__, (vram_weights + vram_scratch + vram_kv_cache + MB - 1) / MB); // round up
 #else
         (void) n_gpu_layers;
 #endif
@@ -1205,58 +1313,15 @@ static void llama_model_load_internal(
         model.tensors_by_name.emplace_back(lt.name, lt.ggml_tensor);
     }
 
-    ml->load_all_data(progress_callback, progress_callback_user_data, use_mlock ? &model.mlock_mmap : NULL);
+    (void) tensor_split;
-
 #if defined(GGML_USE_CUBLAS)
     {
         ggml_cuda_set_tensor_split(tensor_split);
-
-        size_t done_size = 0;
-        size_t data_size = 0;
-        for (llama_load_tensor & lt : ml->tensors_map.tensors) {
-            data_size += lt.size;
-            if (lt.ggml_tensor->backend == GGML_BACKEND_CPU) {
-                done_size += lt.size;
     }
-        }
-        for (llama_load_tensor & lt : ml->tensors_map.tensors) {
-            ggml_backend backend = lt.ggml_tensor->backend;
-            if (backend != GGML_BACKEND_GPU && backend != GGML_BACKEND_GPU_SPLIT) {
-                continue;
-            }
-            if (progress_callback) {
-                progress_callback((float) done_size / data_size, progress_callback_user_data);
-            }
-            ggml_cuda_load_data(fname.c_str(), lt.ggml_tensor, lt.shards.at(0).file_off);
-            done_size += lt.size;
-        }
-    }
-#elif defined(GGML_USE_CLBLAST)
-    {
-        size_t done_size = 0;
-        size_t data_size = 0;
-        for (llama_load_tensor & lt : ml->tensors_map.tensors) {
-            data_size += lt.size;
-            if (lt.ggml_tensor->backend == GGML_BACKEND_CPU) {
-                done_size += lt.size;
-            }
-        }
-        for (llama_load_tensor & lt : ml->tensors_map.tensors) {
-            if (lt.ggml_tensor->backend != GGML_BACKEND_GPU) {
-                continue;
-            }
-            if (progress_callback) {
-                progress_callback((float) done_size / data_size, progress_callback_user_data);
-            }
-            ggml_cl_load_data(fname.c_str(), lt.ggml_tensor, lt.shards.at(0).file_off);
-            done_size += lt.size;
-        }
-    }
-#else
-    (void) n_batch;
-    (void) tensor_split;
 #endif
 
+    ml->load_all_data(progress_callback, progress_callback_user_data, use_mlock ? &model.mlock_mmap : NULL);
+
     if (progress_callback) {
         progress_callback(1.0f, progress_callback_user_data);
     }
@@ -1277,6 +1342,7 @@ static bool llama_model_load(
         int n_gpu_layers,
         int main_gpu,
         float * tensor_split,
+        bool low_vram,
         ggml_type memory_type,
         bool use_mmap,
         bool use_mlock,
@@ -1284,7 +1350,7 @@ static bool llama_model_load(
         llama_progress_callback progress_callback,
         void *progress_callback_user_data) {
     try {
-        llama_model_load_internal(fname, model, vocab, n_ctx, n_batch, n_gpu_layers, main_gpu, tensor_split, memory_type,
+        llama_model_load_internal(fname, model, vocab, n_ctx, n_batch, n_gpu_layers, main_gpu, tensor_split, low_vram, memory_type,
                                   use_mmap, use_mlock, vocab_only, progress_callback, progress_callback_user_data);
         return true;
     } catch (const std::exception & err) {
@@ -1360,12 +1426,33 @@ static bool llama_eval_internal(
     const int i_gpu_start = n_layer - n_gpu_layers;
     (void) i_gpu_start;
 
+    // offload functions set the tensor output backend to GPU
+    // tensors are GPU-accelerated if any input or the output has been offloaded
+    //
+    // with the low VRAM option VRAM scratch is disabled in llama_load_model_internal
+    // in that case ggml_cuda_assign_buffers has no effect
+    offload_func_t offload_func_nr = llama_nop; // nr = non-repeating
+    offload_func_t offload_func_kq = llama_nop;
+    offload_func_t offload_func_v  = llama_nop;
+
+#ifdef GGML_USE_CUBLAS
+        if (n_gpu_layers > n_layer) {
+            offload_func_nr = ggml_cuda_assign_buffers;
+        }
+        if (n_gpu_layers > n_layer + 1) {
+            offload_func_v  = ggml_cuda_assign_buffers;
+        }
+        if (n_gpu_layers > n_layer + 2) {
+            offload_func_kq = ggml_cuda_assign_buffers;
+        }
+#endif // GGML_USE_CUBLAS
+
     for (int il = 0; il < n_layer; ++il) {
         offload_func_t offload_func = llama_nop;
 
 #ifdef GGML_USE_CUBLAS
         if (il >= i_gpu_start) {
-            offload_func = ggml_cuda_assign_buffers; // sets the output backend to GPU
+            offload_func = ggml_cuda_assign_buffers;
         }
 #endif // GGML_USE_CUBLAS
 
@@ -1388,31 +1475,42 @@ static bool llama_eval_internal(
         // self-attention
         {
             // compute Q and K and RoPE them
-            struct ggml_tensor * tmpq = ggml_mul_mat(ctx0, model.layers[il].wq, cur);
-            // offload_func(tmpq);
-            ggml_set_name(tmpq, "tmpq");
-
             struct ggml_tensor * tmpk = ggml_mul_mat(ctx0, model.layers[il].wk, cur);
-            // offload_func(tmpk);
+            offload_func_kq(tmpk);
             ggml_set_name(tmpk, "tmpk");
 
+            struct ggml_tensor * tmpq = ggml_mul_mat(ctx0, model.layers[il].wq, cur);
+            offload_func_kq(tmpq);
+            ggml_set_name(tmpq, "tmpq");
+
             struct ggml_tensor * Kcur = ggml_rope_inplace(ctx0, ggml_reshape_3d(ctx0, tmpk, n_embd/n_head, n_head, N), n_past, n_rot, 0);
+            offload_func_kq(Kcur);
             ggml_set_name(Kcur, "Kcur");
 
             struct ggml_tensor * Qcur = ggml_rope_inplace(ctx0, ggml_reshape_3d(ctx0, tmpq, n_embd/n_head, n_head, N), n_past, n_rot, 0);
+            offload_func_kq(Qcur);
             ggml_set_name(Qcur, "Qcur");
 
             // store key and value to memory
             {
                 // compute the transposed [N, n_embd] V matrix
-                struct ggml_tensor * Vcur = ggml_transpose(ctx0, ggml_reshape_2d(ctx0, ggml_mul_mat(ctx0, model.layers[il].wv, cur), n_embd, N));
+
+                struct ggml_tensor * tmpv = ggml_mul_mat(ctx0, model.layers[il].wv, cur);
+                offload_func_v(tmpv);
+                ggml_set_name(tmpv, "tmpv");
+
+                struct ggml_tensor * Vcur = ggml_transpose(ctx0, ggml_reshape_2d(ctx0, tmpv, n_embd, N));
+                offload_func_v(Vcur);
                 ggml_set_name(Vcur, "Vcur");
 
                 struct ggml_tensor * k = ggml_view_1d(ctx0, kv_self.k, N*n_embd, (ggml_element_size(kv_self.k)*n_embd)*(il*n_ctx + n_past));
+                offload_func_kq(k);
                 ggml_set_name(k, "k");
+
                 struct ggml_tensor * v = ggml_view_2d(ctx0, kv_self.v, N, n_embd,
                         (   n_ctx)*ggml_element_size(kv_self.v),
                         (il*n_ctx)*ggml_element_size(kv_self.v)*n_embd + n_past*ggml_element_size(kv_self.v));
+                offload_func_v(v);
                 ggml_set_name(v, "v");
 
                 // important: storing RoPE-ed version of K in the KV cache!
@@ -1424,6 +1522,7 @@ static bool llama_eval_internal(
                 ggml_permute(ctx0,
                         Qcur,
                         0, 2, 1, 3);
+            offload_func_kq(Q);
             ggml_set_name(Q, "Q");
 
             struct ggml_tensor * K =
@@ -1432,10 +1531,12 @@ static bool llama_eval_internal(
                             ggml_view_1d(ctx0, kv_self.k, (n_past + N)*n_embd, il*n_ctx*ggml_element_size(kv_self.k)*n_embd),
                             n_embd/n_head, n_head, n_past + N),
                         0, 2, 1, 3);
+            offload_func_kq(K);
             ggml_set_name(K, "K");
 
             // K * Q
             struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q);
+            offload_func_kq(KQ);
             ggml_set_name(KQ, "KQ");
 
             // KQ_scaled = KQ / sqrt(n_embd/n_head)
@@ -1444,14 +1545,17 @@ static bool llama_eval_internal(
 
             // KQ_scaled shape [n_past + N, N, n_head, 1]
             struct ggml_tensor * KQ_scaled = ggml_scale_inplace(ctx0, KQ, KQ_scale);
+            offload_func_kq(KQ_scaled);
             ggml_set_name(KQ_scaled, "KQ_scaled");
 
             // KQ_masked = mask_past(KQ_scaled)
             struct ggml_tensor * KQ_masked = ggml_diag_mask_inf_inplace(ctx0, KQ_scaled, n_past);
+            offload_func_kq(KQ_masked);
             ggml_set_name(KQ_masked, "KQ_masked");
 
             // KQ = soft_max(KQ_masked)
             struct ggml_tensor * KQ_soft_max = ggml_soft_max_inplace(ctx0, KQ_masked);
+            offload_func_v(KQ_soft_max);
             ggml_set_name(KQ_soft_max, "KQ_soft_max");
 
             // split cached V into n_head heads
@@ -1461,10 +1565,12 @@ static bool llama_eval_internal(
                         n_ctx*ggml_element_size(kv_self.v),
                         n_ctx*ggml_element_size(kv_self.v)*n_embd/n_head,
                         il*n_ctx*ggml_element_size(kv_self.v)*n_embd);
+            offload_func_v(V);
             ggml_set_name(V, "V");
 
 #if 1
             struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V, KQ_soft_max);
+            offload_func_v(KQV);
             ggml_set_name(KQV, "KQV");
 #else
             // make V contiguous in memory to speed up the matmul, however we waste time on the copy
@@ -1476,12 +1582,14 @@ static bool llama_eval_internal(
 
             // KQV_merged = KQV.permute(0, 2, 1, 3)
             struct ggml_tensor * KQV_merged = ggml_permute(ctx0, KQV, 0, 2, 1, 3);
+            offload_func_v(KQV_merged);
             ggml_set_name(KQV_merged, "KQV_merged");
 
             // cur = KQV_merged.contiguous().view(n_embd, N)
             cur = ggml_cpy(ctx0,
                     KQV_merged,
                     ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, N));
+            offload_func_v(cur);
             ggml_set_name(cur, "KQV_merged_contiguous");
 
             // projection (no bias)
@@ -1493,7 +1601,6 @@ static bool llama_eval_internal(
         }
 
         lctx.use_buf(ctx0, 1);
-        //ggml_cuda_set_scratch(1);
 
         struct ggml_tensor * inpFF = ggml_add(ctx0, cur, inpSA);
         offload_func(inpFF);
@@ -1523,7 +1630,7 @@ static bool llama_eval_internal(
                     model.layers[il].w1,
                     cur);
             offload_func(cur);
-            ggml_set_name(cur, "result_w2");
+            ggml_set_name(cur, "result_w1");
 
             // SILU activation
             cur = ggml_silu(ctx0, cur);
@@ -1551,32 +1658,20 @@ static bool llama_eval_internal(
     }
 
     lctx.use_buf(ctx0, 0);
-    //ggml_cuda_set_scratch(0);
 
     // used at the end to optionally extract the embeddings
     struct ggml_tensor * embeddings = NULL;
 
-    offload_func_t offload_func = llama_nop;
-
-#ifdef GGML_USE_CUBLAS
-        if (n_gpu_layers > n_layer) {
-            offload_func = ggml_cuda_assign_buffers; // sets the output backend to GPU
-        }
-#endif // GGML_USE_CUBLAS
 
     // norm
     {
         cur = ggml_rms_norm(ctx0, inpL);
-        offload_func(cur);
+        offload_func_nr(cur);
-        ggml_set_name(cur, "rms_norm_inpL");
+        ggml_set_name(cur, "rms_norm_2");
-
-        cur = ggml_rms_norm(ctx0, cur);
-        offload_func(cur);
-        ggml_set_name(cur, "rms_norm_after");
 
         // cur = cur*norm(broadcasted)
         cur = ggml_mul(ctx0, cur, model.norm);
-        offload_func(cur);
+        // offload_func_nr(cur); // TODO CPU + GPU mirrored backend
         ggml_set_name(cur, "result_norm");
 
         embeddings = cur;
@@ -2184,6 +2279,10 @@ llama_token llama_sample_token_mirostat_v2(struct llama_context * ctx, llama_tok
         return -log2f(candidate.p) > *mu;
     }));
 
+    if (candidates->size == 0) {
+        candidates->size = 1;
+    }
+
     // Normalize the probabilities of the remaining words
     llama_sample_softmax(ctx, candidates);
 
@@ -2321,7 +2420,10 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
         case LLAMA_FTYPE_MOSTLY_Q5_0: quantized_type = GGML_TYPE_Q5_0; break;
         case LLAMA_FTYPE_MOSTLY_Q5_1: quantized_type = GGML_TYPE_Q5_1; break;
         case LLAMA_FTYPE_MOSTLY_Q8_0: quantized_type = GGML_TYPE_Q8_0; break;
+        case LLAMA_FTYPE_MOSTLY_F16: quantized_type = GGML_TYPE_F16; break;
+        case LLAMA_FTYPE_ALL_F32: quantized_type = GGML_TYPE_F32; break;
 
+#ifdef GGML_USE_K_QUANTS
         // K-quants
         case LLAMA_FTYPE_MOSTLY_Q2_K:   quantized_type = GGML_TYPE_Q2_K; break;
         case LLAMA_FTYPE_MOSTLY_Q3_K_S:
@@ -2332,6 +2434,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
         case LLAMA_FTYPE_MOSTLY_Q5_K_S:
         case LLAMA_FTYPE_MOSTLY_Q5_K_M: quantized_type = GGML_TYPE_Q5_K; break;
         case LLAMA_FTYPE_MOSTLY_Q6_K:   quantized_type = GGML_TYPE_Q6_K; break;
+#endif
         default: throw std::runtime_error(format("invalid output file type %d\n", ftype));
     }
 
@@ -2343,6 +2446,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
                                                                             /*vocab_only*/ false));
     llama_file_saver file_saver(fname_out.c_str(), model_loader->file_loaders.at(0).get(), params->ftype);
 
+#ifdef GGML_USE_K_QUANTS
     int n_attention_wv    = 0;
     int n_feed_forward_w2 = 0;
     for (auto& tensor : model_loader->tensors_map.tensors) {
@@ -2356,6 +2460,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
 
     int i_attention_wv = 0;
     int i_feed_forward_w2 = 0;
+#endif
 
     size_t total_size_org = 0;
     size_t total_size_new = 0;
@@ -2381,12 +2486,8 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
 
         // quantize only 2D tensors
         quantize &= (tensor.ne.size() == 2);
-
+        quantize &= params->quantize_output_tensor || tensor.name != "output.weight";
-        // uncomment this to keep the output layer in FP16
+        quantize &= quantized_type != tensor.type;
-        if (!params->quantize_output_tensor && tensor.name == "output.weight") {
-           quantize = false;
-        }
-        quantize = quantize && quantized_type != tensor.type;
 
         enum ggml_type new_type;
         void * new_data;
@@ -2400,31 +2501,39 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
             printf("size = %8.3f MB\n", tensor.size/1024.0/1024.0);
         } else {
             new_type = quantized_type;
-            // TODO: temporary disabled until Metal / OpenCL support is available
+#ifdef GGML_USE_K_QUANTS
-            //       ref: https://github.com/ggerganov/llama.cpp/issues/1711
+            if (quantized_type == GGML_TYPE_Q2_K || quantized_type == GGML_TYPE_Q3_K || quantized_type == GGML_TYPE_Q4_K ||
-            //if (tensor.name == "output.weight") {
+                quantized_type == GGML_TYPE_Q5_K || quantized_type == GGML_TYPE_Q6_K) {
-            //    new_type = GGML_TYPE_Q6_K;
+                int nx = tensor.ne.at(0);
-            //}
+                int ny = tensor.ne.at(0);
-            if (tensor.name.find("attention.wv.weight") != std::string::npos) {
+                if (nx % QK_K != 0 || ny % QK_K != 0) {
+                    fprintf(stderr, "\n\n========================= Tensor sizes %d x %d are not divisible by %d\n",nx,ny,QK_K);
+                    fprintf(stderr, "This is required to be able to use k-quants for now!\n");
+                    fprintf(stderr, "========================================================================================\n\n");
+                    throw std::runtime_error("Unsupported tensor size encountered\n");
+                }
+            }
+            if (tensor.name == "output.weight") {
+               new_type = GGML_TYPE_Q6_K;
+            } else if (tensor.name.find("attention.wv.weight") != std::string::npos) {
                 if      (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M || ftype == LLAMA_FTYPE_MOSTLY_Q2_K) new_type = GGML_TYPE_Q4_K;
                 else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L) new_type = GGML_TYPE_Q5_K;
                 else if ((ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M || ftype == LLAMA_FTYPE_MOSTLY_Q5_K_M) &&
                          (i_attention_wv < n_attention_wv/8 || i_attention_wv >= 7*n_attention_wv/8 ||
                          (i_attention_wv - n_attention_wv/8)%3 == 2)) new_type = GGML_TYPE_Q6_K;
                 ++i_attention_wv;
-            }
+            } else if (tensor.name.find("feed_forward.w2.weight") != std::string::npos) {
-            if (tensor.name.find("feed_forward.w2.weight") != std::string::npos) {
                 if      (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M || ftype == LLAMA_FTYPE_MOSTLY_Q2_K) new_type = GGML_TYPE_Q4_K;
                 else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L) new_type = GGML_TYPE_Q5_K;
                 else if ((ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M || ftype == LLAMA_FTYPE_MOSTLY_Q5_K_M) &&
                          (i_feed_forward_w2 < n_feed_forward_w2/8 || i_feed_forward_w2 >= 7*n_feed_forward_w2/8 ||
                          (i_feed_forward_w2 - n_feed_forward_w2/8)%3 == 2)) new_type = GGML_TYPE_Q6_K;
                 ++i_feed_forward_w2;
-            }
+            } else if (tensor.name.find("attention.wo.weight") != std::string::npos) {
-            if (tensor.name.find("attention.wo.weight") != std::string::npos) {
                 if      (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M || ftype == LLAMA_FTYPE_MOSTLY_Q2_K) new_type = GGML_TYPE_Q4_K;
                 else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L) new_type = GGML_TYPE_Q5_K;
             }
+#endif
 
             float * f32_data;
             size_t nelements = tensor.ne.at(0) * tensor.ne.at(1);
@@ -2541,7 +2650,7 @@ struct llama_model * llama_load_model_from_file(
     ggml_type memory_type = params.f16_kv ? GGML_TYPE_F16 : GGML_TYPE_F32;
 
     if (!llama_model_load(path_model, *model, model->vocab, params.n_ctx, params.n_batch, params.n_gpu_layers,
-                params.main_gpu, params.tensor_split, memory_type, params.use_mmap, params.use_mlock,
+                params.main_gpu, params.tensor_split, params.low_vram, memory_type, params.use_mmap, params.use_mlock,
                 params.vocab_only, params.progress_callback, params.progress_callback_user_data)) {
         delete model;
         fprintf(stderr, "%s: failed to load model\n", __func__);
@@ -2593,7 +2702,7 @@ struct llama_context * llama_new_context_with_model(
 
     // reserve memory for context buffers
     if (!params.vocab_only) {
-        if (!kv_cache_init(ctx->model.hparams, ctx->kv_self, memory_type, ctx->model.hparams.n_ctx)) {
+        if (!kv_cache_init(ctx->model.hparams, ctx->kv_self, memory_type, ctx->model.hparams.n_ctx, params.n_gpu_layers)) {
             fprintf(stderr, "%s: kv_cache_init() failed for self-attention cache\n", __func__);
             llama_free(ctx);
             return nullptr;
@@ -2628,16 +2737,21 @@ struct llama_context * llama_new_context_with_model(
         // this allocates all Metal resources and memory buffers
         ctx->ctx_metal = ggml_metal_init();
 
-        void *data_ptr = NULL;
+        void * data_ptr  = NULL;
         size_t data_size = 0;
+
         if (params.use_mmap) {
             data_ptr  = ctx->model.mapping->addr;
-            data_size= ctx->model.mapping->size;
+            data_size = ctx->model.mapping->size;
         } else {
             data_ptr  = ggml_get_mem_buffer(ctx->model.ctx);
-            data_size= ggml_get_mem_size(ctx->model.ctx);
+            data_size = ggml_get_mem_size  (ctx->model.ctx);
         }
 
+        const size_t max_size = ggml_get_max_tensor_size(ctx->model.ctx);
+
+        printf("%s: max tensor size = %8.2f MB\n", __func__, max_size/1024.0/1024.0);
+
 #define LLAMA_METAL_CHECK_BUF(result)                                          \
     if (!(result)) {                                                           \
         fprintf(stderr, "%s: failed to add buffer\n", __func__);               \
@@ -2645,12 +2759,13 @@ struct llama_context * llama_new_context_with_model(
         return NULL;                                                           \
     }
 
-        LLAMA_METAL_CHECK_BUF(ggml_metal_add_buffer(ctx->ctx_metal, "data", data_ptr, data_size));
+        LLAMA_METAL_CHECK_BUF(ggml_metal_add_buffer(ctx->ctx_metal, "data", data_ptr, data_size, max_size));
-        LLAMA_METAL_CHECK_BUF(ggml_metal_add_buffer(ctx->ctx_metal, "eval", ctx->buf_compute.addr, ctx->buf_compute.size));
 
-        LLAMA_METAL_CHECK_BUF(ggml_metal_add_buffer(ctx->ctx_metal, "kv",   ctx->kv_self.buf.addr, ctx->kv_self.buf.size));
+        LLAMA_METAL_CHECK_BUF(ggml_metal_add_buffer(ctx->ctx_metal, "eval", ctx->buf_compute.addr, ctx->buf_compute.size, 0));
-        LLAMA_METAL_CHECK_BUF(ggml_metal_add_buffer(ctx->ctx_metal, "scr0", ctx->buf_scratch[0].addr,    ctx->buf_scratch[0].size));
+        LLAMA_METAL_CHECK_BUF(ggml_metal_add_buffer(ctx->ctx_metal, "kv",   ctx->kv_self.buf.addr, ctx->kv_self.buf.size, 0));
-        LLAMA_METAL_CHECK_BUF(ggml_metal_add_buffer(ctx->ctx_metal, "scr1", ctx->buf_scratch[1].addr,    ctx->buf_scratch[1].size));
+
+        LLAMA_METAL_CHECK_BUF(ggml_metal_add_buffer(ctx->ctx_metal, "scr0", ctx->buf_scratch[0].addr, ctx->buf_scratch[0].size, 0));
+        LLAMA_METAL_CHECK_BUF(ggml_metal_add_buffer(ctx->ctx_metal, "scr1", ctx->buf_scratch[1].addr, ctx->buf_scratch[1].size, 0));
 #undef LLAMA_METAL_CHECK_BUF
     }
 #endif
@@ -3353,6 +3468,19 @@ int llama_n_embd(const struct llama_context * ctx) {
     return ctx->model.hparams.n_embd;
 }
 
+int llama_get_vocab(
+        const struct llama_context * ctx,
+        const char * * strings,
+        float  * scores,
+        int capacity) {
+    int n = std::min(capacity, (int) ctx->vocab.id_to_token.size());
+    for (int i = 0; i<n; ++i) {
+        strings[i] = ctx->vocab.id_to_token[i].tok.c_str();
+        scores[i]  = ctx->vocab.id_to_token[i].score;
+    }
+    return n;
+}
+
 float * llama_get_logits(struct llama_context * ctx) {
     return ctx->logits.data();
 }
@@ -3391,9 +3519,12 @@ void llama_print_timings(struct llama_context * ctx) {
 
     fprintf(stderr, "\n");
     fprintf(stderr, "%s:        load time = %8.2f ms\n", __func__, ctx->t_load_us / 1000.0);
-    fprintf(stderr, "%s:      sample time = %8.2f ms / %5d runs   (%8.2f ms per token)\n", __func__, 1e-3 * ctx->t_sample_us, n_sample, 1e-3 * ctx->t_sample_us / n_sample);
+    fprintf(stderr, "%s:      sample time = %8.2f ms / %5d runs   (%8.2f ms per token, %8.2f tokens per second)\n",
-    fprintf(stderr, "%s: prompt eval time = %8.2f ms / %5d tokens (%8.2f ms per token)\n", __func__, 1e-3 * ctx->t_p_eval_us, n_p_eval, 1e-3 * ctx->t_p_eval_us / n_p_eval);
+            __func__, 1e-3 * ctx->t_sample_us, n_sample, 1e-3 * ctx->t_sample_us / n_sample, 1e6 / ctx->t_sample_us * n_sample);
-    fprintf(stderr, "%s:        eval time = %8.2f ms / %5d runs   (%8.2f ms per token)\n", __func__, 1e-3 * ctx->t_eval_us,   n_eval,   1e-3 * ctx->t_eval_us   / n_eval);
+    fprintf(stderr, "%s: prompt eval time = %8.2f ms / %5d tokens (%8.2f ms per token, %8.2f tokens per second)\n",
+            __func__, 1e-3 * ctx->t_p_eval_us, n_p_eval, 1e-3 * ctx->t_p_eval_us / n_p_eval, 1e6 / ctx->t_p_eval_us * n_p_eval);
+    fprintf(stderr, "%s:        eval time = %8.2f ms / %5d runs   (%8.2f ms per token, %8.2f tokens per second)\n",
+            __func__, 1e-3 * ctx->t_eval_us,   n_eval,   1e-3 * ctx->t_eval_us   / n_eval,   1e6 / ctx->t_eval_us   * n_eval);
     fprintf(stderr, "%s:       total time = %8.2f ms\n", __func__, (t_end_us - ctx->t_start_us)/1000.0);
 }
 

llama.h

@@ -78,6 +78,7 @@ extern "C" {
         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
@@ -237,6 +238,14 @@ 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
@@ -252,9 +261,9 @@ extern "C" {
     LLAMA_API const char * llama_token_to_str(const struct llama_context * ctx, llama_token token);
 
     // Special tokens
-    LLAMA_API llama_token llama_token_bos();
+    LLAMA_API llama_token llama_token_bos();  // beginning-of-sentence
-    LLAMA_API llama_token llama_token_eos();
+    LLAMA_API llama_token llama_token_eos();  // end-of-sentence
-    LLAMA_API llama_token llama_token_nl();
+    LLAMA_API llama_token llama_token_nl();   // next-line
 
     // Sampling functions
 

pocs/vdot/vdot.cpp

@@ -10,6 +10,10 @@
 
 #include <ggml.h>
 
+#if defined(_MSC_VER)
+#pragma warning(disable: 4244 4267) // possible loss of data
+#endif
+
 constexpr int kVecSize = 1 << 18;
 
 float drawFromGaussianPdf(std::mt19937& rndm) {

scripts/verify-checksum-models.py

@@ -1,6 +1,7 @@
 import os
 import hashlib
 
+
 def sha256sum(file):
     block_size = 16 * 1024 * 1024  # 16 MB block size
     b = bytearray(block_size)
@@ -15,6 +16,7 @@ def sha256sum(file):
 
     return file_hash.hexdigest()
 
+
 # Define the path to the llama directory (parent folder of script directory)
 llama_path = os.path.abspath(os.path.join(os.path.dirname(__file__), os.pardir))
 

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

@@ -9,12 +9,15 @@
 #include <string>
 #include <vector>
 
+#if defined(_MSC_VER)
+#pragma warning(disable: 4244 4267) // possible loss of data
+#endif
 
-const float MAX_QUANTIZATION_REFERENCE_ERROR = 0.0001;
+const float MAX_QUANTIZATION_REFERENCE_ERROR = 0.0001f;
-const float MAX_QUANTIZATION_TOTAL_ERROR = 0.002;
+const float MAX_QUANTIZATION_TOTAL_ERROR = 0.002f;
-const float MAX_QUANTIZATION_TOTAL_ERROR_2BITS = 0.0075;
+const float MAX_QUANTIZATION_TOTAL_ERROR_2BITS = 0.0075f;
-const float MAX_QUANTIZATION_TOTAL_ERROR_3BITS = 0.0040;
+const float MAX_QUANTIZATION_TOTAL_ERROR_3BITS = 0.0040f;
-const float MAX_DOT_PRODUCT_ERROR = 0.02;
+const float MAX_DOT_PRODUCT_ERROR = 0.02f;
 
 const char* RESULT_STR[] = {"ok", "FAILED"};
 

tests/test-quantize-perf.cpp

@@ -13,6 +13,10 @@
 #include <string>
 #include <vector>
 
+#if defined(_MSC_VER)
+#pragma warning(disable: 4244 4267) // possible loss of data
+#endif
+
 #define MAX_ALIGNMENT 64
 #define QK 32
 #define WARMUP 5

tests/test-sampling.cpp

@@ -176,27 +176,27 @@ void test_frequency_presence_penalty(
 int main(void) {
     ggml_time_init();
 
-    test_top_k({0.1, 0.2, 0.3, 0.4}, {0.4}, 1);
+    test_top_k({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f}, 1);
-    test_top_k({0.1, 0.2, 0.3, 0.4}, {0.4, 0.3, 0.2}, 3);
+    test_top_k({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f, 0.3f, 0.2f}, 3);
 
-    test_top_p({0.1, 0.2, 0.3, 0.4}, {0.4}, 0);
+    test_top_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f}, 0);
-    test_top_p({0.1, 0.2, 0.3, 0.4}, {0.4, 0.3}, 0.7);
+    test_top_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f, 0.3f}, 0.7f);
-    test_top_p({0.1, 0.2, 0.3, 0.4}, {0.4, 0.3, 0.2, 0.1}, 1);
+    test_top_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f, 0.3f, 0.2f, 0.1f}, 1);
 
-    test_tfs({0.1, 0.15, 0.2, 0.25, 0.3}, {0.3}, 0.25);
+    test_tfs({0.1f, 0.15f, 0.2f, 0.25f, 0.3f}, {0.3f}, 0.25f);
-    test_tfs({0.1, 0.15, 0.2, 0.25, 0.3}, {0.3, 0.25}, 0.75);
+    test_tfs({0.1f, 0.15f, 0.2f, 0.25f, 0.3f}, {0.3f, 0.25f}, 0.75f);
-    test_tfs({0.1, 0.15, 0.2, 0.25, 0.3}, {0.3, 0.25}, 0.99);
+    test_tfs({0.1f, 0.15f, 0.2f, 0.25f, 0.3f}, {0.3f, 0.25f}, 0.99f);
 
-    test_typical({0.97, 0.01, 0.01, 0.01}, {0.97}, 0.5);
+    test_typical({0.97f, 0.01f, 0.01f, 0.01f}, {0.97f}, 0.5f);
-    test_typical({0.4, 0.2, 0.2, 0.2}, {0.2, 0.2, 0.2}, 0.5);
+    test_typical({0.4f, 0.2f, 0.2f, 0.2f}, {0.2f, 0.2f, 0.2f}, 0.5f);
 
-    test_repetition_penalty({0.2, 0.2, 0.2, 0.2, 0.2}, {0}, {0.25, 0.25, 0.25, 0.25, 0}, 50.0);
+    test_repetition_penalty({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0}, {0.25f, 0.25f, 0.25f, 0.25f, 0}, 50.0f);
-    test_repetition_penalty({0.2, 0.2, 0.2, 0.2, 0.2}, {0, 1, 2}, {0.5, 0.5, 0, 0, 0}, 50.0);
+    test_repetition_penalty({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0, 1, 2}, {0.5f, 0.5f, 0, 0, 0}, 50.0f);
-    test_repetition_penalty({0.2, 0.2, 0.2, 0.2, 0.2}, {0, 1, 2, 0, 0}, {0.5, 0.5, 0, 0, 0}, 50.0);
+    test_repetition_penalty({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0, 1, 2, 0, 0}, {0.5f, 0.5f, 0, 0, 0}, 50.0f);
 
-    test_frequency_presence_penalty({0.2, 0.2, 0.2, 0.2, 0.2}, {0},             {0.249997, 0.249997, 0.249997, 0.249997, 0.000011}, 5.0, 5.0);
+    test_frequency_presence_penalty({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0},             {0.249997f, 0.249997f, 0.249997f, 0.249997f, 0.000011f}, 5.0f, 5.0f);
-    test_frequency_presence_penalty({0.2, 0.2, 0.2, 0.2, 0.2}, {0, 1, 2},       {0.499966, 0.499966, 0.000023, 0.000023, 0.000023}, 5.0, 5.0);
+    test_frequency_presence_penalty({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0, 1, 2},       {0.499966f, 0.499966f, 0.000023f, 0.000023f, 0.000023f}, 5.0f, 5.0f);
-    test_frequency_presence_penalty({0.2, 0.2, 0.2, 0.2, 0.2}, {0, 1, 2, 0, 0}, {0.499977, 0.499977, 0.000023, 0.000023, 0.000000}, 5.0, 5.0);
+    test_frequency_presence_penalty({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0, 1, 2, 0, 0}, {0.499977f, 0.499977f, 0.000023f, 0.000023f, 0.000000f}, 5.0f, 5.0f);
 
     printf("OK\n");
 }

tests/test-tokenizer-0.cpp

@@ -53,7 +53,7 @@ int main(int argc, char **argv) {
 
     for (const auto & test_kv : k_tests()) {
         std::vector<llama_token> res(test_kv.first.size());
-        const int n = llama_tokenize(ctx, test_kv.first.c_str(), res.data(), res.size(), true);
+        const int n = llama_tokenize(ctx, test_kv.first.c_str(), res.data(), int(res.size()), true);
         res.resize(n);
 
         bool correct = res.size() == test_kv.second.size();