llama : fix params struct slignment (#1936)

* Workaround struct misalignment during value-copy

Signed-off-by: mudler <mudler@localai.io>

* Move booleans at the bottom of the structure

Signed-off-by: mudler <mudler@localai.io>

* Add comment

Signed-off-by: mudler <mudler@localai.io>

---------

Signed-off-by: mudler <mudler@localai.io>

.flake8

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

.gitignore

@@ -34,8 +34,10 @@ models/*
 /perplexity
 /embedding
 /train-text-from-scratch
+/simple
 /benchmark-matmult
 /vdot
+/server
 /Pipfile
 /libllama.so
 

.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 +70,7 @@ 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)
@@ -238,6 +239,9 @@ 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)
@@ -461,8 +465,11 @@ 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>)
+    target_link_libraries(ggml_shared PUBLIC Threads::Threads ${LLAMA_EXTRA_LIBS})
 endif()
 
 add_library(llama
@@ -488,9 +495,18 @@ 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")
+
+    if (BUILD_SHARED_LIBS)
+        set_property(TARGET ggml_shared PROPERTY CUDA_ARCHITECTURES "native")
+        set_property(TARGET ggml_shared PROPERTY CUDA_SELECT_NVCC_ARCH_FLAGS "Auto")
+    endif()
+
+    set_property(TARGET llama PROPERTY CUDA_ARCHITECTURES "native")
 endif()
 
 

Makefile

@@ -3,6 +3,8 @@ BUILD_TARGETS = main quantize quantize-stats perplexity embedding vdot train-tex
 
 ifdef LLAMA_BUILD_SERVER
 	BUILD_TARGETS += server
+	LLAMA_SERVER_VERBOSE ?= 1
+server: private CXXFLAGS += -DSERVER_VERBOSE=$(LLAMA_SERVER_VERBOSE)
 endif
 
 default: $(BUILD_TARGETS)
@@ -142,11 +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
+	LDFLAGS += -lopenblas
 endif # LLAMA_OPENBLAS
 
 ifdef LLAMA_BLIS
@@ -171,6 +169,9 @@ 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
@@ -254,7 +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
@@ -278,9 +279,6 @@ main: examples/main/main.cpp                                  build-info.h ggml.
 
 simple: examples/simple/simple.cpp                            build-info.h ggml.o llama.o common.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
-	@echo
-	@echo '====  Run ./simple -h for help.  ===='
-	@echo
 
 quantize: examples/quantize/quantize.cpp                      build-info.h ggml.o llama.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)

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

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
 
@@ -694,8 +698,9 @@ class LazyUnpickler(pickle.Unpickler):
         description = f'storage data_type={data_type} path-in-zip={filename} path={self.zip_file.filename}'
         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]
+    # @staticmethod
+    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)
 
@@ -812,7 +817,7 @@ def lazy_load_ggml_file(fp: io.BufferedReader, path: Path) -> ModelPlus:
     # Use mmap for the actual data to avoid race conditions with the file offset.
     off = fp.raw.tell()
     mapped = memoryview(mmap.mmap(fp.fileno(), 0, access=mmap.ACCESS_READ))
-    fp.raw.seek(off) # needed on Windows
+    fp.raw.seek(off)  # needed on Windows
 
     def read_tensor() -> None:  # this is a function so that variables captured in `load` don't change
         shape_len, name_len, ftype = struct.unpack("iii", must_read(fp, 12))
@@ -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

@@ -38,6 +38,7 @@ else()
     add_subdirectory(benchmark)
     add_subdirectory(baby-llama)
     add_subdirectory(train-text-from-scratch)
+    add_subdirectory(simple)
     if (LLAMA_METAL)
         add_subdirectory(metal)
     endif()

examples/common.cpp

@@ -106,9 +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;

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/main.cpp

@@ -354,7 +354,7 @@ int main(int argc, char ** argv) {
             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 %" PRIu64 "  token%s>>", skipped_tokens, skipped_tokens != 1 ? "s" : "");
+                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);

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));
-    ggml_metal_add_buffer(ctx_metal, "eval", ggml_get_mem_buffer(ctx_eval), ggml_get_mem_size(ctx_eval));
+    const size_t max_size_data = ggml_get_max_tensor_size(ctx_data);
+    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/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

@@ -1,37 +1,75 @@
 # 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)
-2. [Node JS Test](#node-js-test)
-3. [API Endpoints](#api-endpoints)
-4. [More examples](#more-examples)
-5. [Common Options](#common-options)
-6. [Performance Tuning and Memory Options](#performance-tuning-and-memory-options)
+-   `--threads N`, `-t N`: Set the number of threads to use during computation.
+-   `-m FNAME`, `--model FNAME`: Specify the path to the LLaMA model file (e.g., `models/7B/ggml-model.bin`).
+-   `-m ALIAS`, `--alias ALIAS`: Set an alias for the model. The alias will be returned in API responses.
+-   `-c N`, `--ctx-size N`: Set the size of the prompt context. The default is 512, but LLaMA models were built with a context of 2048, which will provide better results for longer input/inference.
+-   `-ngl N`, `--n-gpu-layers N`: When compiled with appropriate support (currently CLBlast or cuBLAS), this option allows offloading some layers to the GPU for computation. Generally results in increased performance.
+-   `-mg i, --main-gpu i`: When using multiple GPUs this option controls which GPU is used for small tensors for which the overhead of splitting the computation across all GPUs is not worthwhile. The GPU in question will use slightly more VRAM to store a scratch buffer for temporary results. By default GPU 0 is used. Requires cuBLAS.
+-   `-ts SPLIT, --tensor-split SPLIT`: When using multiple GPUs this option controls how large tensors should be split across all GPUs. `SPLIT` is a comma-separated list of non-negative values that assigns the proportion of data that each GPU should get in order. For example, "3,2" will assign 60% of the data to GPU 0 and 40% to GPU 1. By default the data is split in proportion to VRAM but this may not be optimal for performance. Requires cuBLAS.
+-   `-lv, --low-vram`: Do not allocate a VRAM scratch buffer for holding temporary results. Reduces VRAM usage at the cost of performance, particularly prompt processing speed. Requires cuBLAS.
+-   `-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`.
+-   `--embedding`: Enable embedding extraction, Default: disabled.
+
+## 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.):
-Make sure to build with the server option on
-```bash
-LLAMA_BUILD_SERVER=1 make
-```
+### 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
 
@@ -54,7 +92,6 @@ 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,
     });
 
@@ -73,247 +110,83 @@ 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. When 0, no tokens will be generated but the prompt is evaluated into the cache. (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. A space is inserted in the front like main.cpp does.
 
-`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.
+
+    Note that the special `BOS` token is not added in fron of the text and also a space character is not inserted automatically as it is for `/completion`.
+
+-   **POST** `/embedding`: Generate embedding of a given text just as [the embedding example](../embedding) does.
+
+    *Options:*
+
+    `content`: Set the text to process.
 
 ## 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.
-
-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();
+```sh
+node chat.mjs
 ```
 
-### 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
-
-```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
+```sh
+bash chat.sh
 ```
-
-### 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.
--   `-lv, --low-vram`: Do not allocate a VRAM scratch buffer for holding temporary results. Reduces VRAM usage at the cost of performance, particularly prompt processing speed. Requires cuBLAS.
--   `--embedding`: Enable the embedding mode. **Completion function doesn't work in this mode**.
--   `--host`: Set the hostname or ip address to listen. Default `127.0.0.1`;
--   `--port`: Set the port to listen. Default: `8080`.
-
-### 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/train-text-from-scratch/README.md

@@ -4,7 +4,7 @@ Basic usage instructions:
 
 ```bash
 # get training data
-wget https://github.com/brunoklein99/deep-learning-notes/blob/master/shakespeare.txt
+wget https://raw.githubusercontent.com/brunoklein99/deep-learning-notes/master/shakespeare.txt
 
 # train
 ./bin/train-text-from-scratch \

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.cu

@@ -13,6 +13,10 @@
 #include "ggml-cuda.h"
 #include "ggml.h"
 
+#if defined(_MSC_VER)
+#pragma warning(disable: 4244 4267) // possible loss of data
+#endif
+
 static_assert(sizeof(half) == sizeof(ggml_fp16_t), "wrong fp16 size");
 
 #define CUDA_CHECK(err)                                                                 \
@@ -46,7 +50,15 @@ static_assert(sizeof(half) == sizeof(ggml_fp16_t), "wrong fp16 size");
     } while (0)
 #endif // CUDART_VERSION >= 11
 
-typedef void (*dequantize_kernel_t)(const void * vx, const int ib, const int iqs, float & v0, float & v1);
+#ifdef GGML_CUDA_DMMV_F16
+typedef half dfloat; // dequantize float
+typedef half2 dfloat2;
+#else
+typedef float dfloat; // dequantize float
+typedef float2 dfloat2;
+#endif //GGML_CUDA_DMMV_F16
+
+typedef void (*dequantize_kernel_t)(const void * vx, const int ib, const int iqs, dfloat2 & v);
 typedef void (*to_fp32_cuda_t)(const void * x, float * y, int k, cudaStream_t stream);
 typedef void (*dot_kernel_k_t)(const void * vx, const int ib, const int iqs, const float * y, float & v);
 typedef void (*cpy_kernel_t)(const char * cx, char * cdst);
@@ -230,82 +242,106 @@ static __global__ void rms_norm_f32(const float * x, float * dst, const int ncol
     }
 }
 
-static __device__ void dequantize_q4_0(const void * vx, const int ib, const int iqs, float & v0, float & v1){
+static __device__ __forceinline__ void dequantize_q4_0(const void * vx, const int ib, const int iqs, dfloat2 & v){
     const block_q4_0 * x = (const block_q4_0 *) vx;
 
-    const float d = x[ib].d;
+    const dfloat d = x[ib].d;
 
-    const uint8_t vui = x[ib].qs[iqs];
+    const int vui = x[ib].qs[iqs];
 
-    const int8_t vi0 = vui & 0xF;
-    const int8_t vi1 = vui >> 4;
+    v.x = vui & 0xF;
+    v.y = vui >> 4;
 
-    v0 = (vi0 - 8)*d;
-    v1 = (vi1 - 8)*d;
+#ifdef GGML_CUDA_DMMV_F16
+    v = __hsub2(v, {8.0f, 8.0f});
+    v = __hmul2(v, {d, d});
+#else
+    v.x = (v.x - 8.0f) * d;
+    v.y = (v.y - 8.0f) * d;
+#endif // GGML_CUDA_DMMV_F16
 }
 
-static __device__ void dequantize_q4_1(const void * vx, const int ib, const int iqs, float & v0, float & v1){
+static __device__ __forceinline__ void dequantize_q4_1(const void * vx, const int ib, const int iqs, dfloat2 & v){
     const block_q4_1 * x = (const block_q4_1 *) vx;
 
-    const float d = x[ib].d;
-    const float m = x[ib].m;
+    const dfloat d = x[ib].d;
+    const dfloat m = x[ib].m;
 
-    const uint8_t vui = x[ib].qs[iqs];
+    const int vui = x[ib].qs[iqs];
 
-    const int8_t vi0 = vui & 0xF;
-    const int8_t vi1 = vui >> 4;
+    v.x = vui & 0xF;
+    v.y = vui >> 4;
 
-    v0 = vi0*d + m;
-    v1 = vi1*d + m;
+#ifdef GGML_CUDA_DMMV_F16
+    v = __hmul2(v, {d, d});
+    v = __hadd2(v, {m, m});
+#else
+    v.x = (v.x * d) + m;
+    v.y = (v.y * d) + m;
+#endif // GGML_CUDA_DMMV_F16
 }
 
-static __device__ void dequantize_q5_0(const void * vx, const int ib, const int iqs, float & v0, float & v1){
+static __device__ __forceinline__ void dequantize_q5_0(const void * vx, const int ib, const int iqs, dfloat2 & v){
     const block_q5_0 * x = (const block_q5_0 *) vx;
 
-    const float d = x[ib].d;
+    const dfloat d = x[ib].d;
 
     uint32_t qh;
     memcpy(&qh, x[ib].qh, sizeof(qh));
 
-    const uint8_t xh_0 = ((qh >> (iqs +  0)) << 4) & 0x10;
-    const uint8_t xh_1 = ((qh >> (iqs + 12))     ) & 0x10;
+    const int xh_0 = ((qh >> (iqs +  0)) << 4) & 0x10;
+    const int xh_1 = ((qh >> (iqs + 12))     ) & 0x10;
 
-    const int32_t x0 = ((x[ib].qs[iqs] & 0xf) | xh_0) - 16;
-    const int32_t x1 = ((x[ib].qs[iqs] >>  4) | xh_1) - 16;
+    v.x = ((x[ib].qs[iqs] & 0xf) | xh_0);
+    v.y = ((x[ib].qs[iqs] >>  4) | xh_1);
 
-    v0 = x0*d;
-    v1 = x1*d;
+#ifdef GGML_CUDA_DMMV_F16
+    v = __hsub2(v, {16.0f, 16.0f});
+    v = __hmul2(v, {d, d});
+#else
+    v.x = (v.x - 16.0f) * d;
+    v.y = (v.y - 16.0f) * d;
+#endif // GGML_CUDA_DMMV_F16
 }
 
-static __device__ void dequantize_q5_1(const void * vx, const int ib, const int iqs, float & v0, float & v1){
+static __device__ __forceinline__ void dequantize_q5_1(const void * vx, const int ib, const int iqs, dfloat2 & v){
     const block_q5_1 * x = (const block_q5_1 *) vx;
 
-    const float d = x[ib].d;
-    const float m = x[ib].m;
+    const dfloat d = x[ib].d;
+    const dfloat m = x[ib].m;
 
     uint32_t qh;
     memcpy(&qh, x[ib].qh, sizeof(qh));
 
-    const uint8_t xh_0 = ((qh >> (iqs +  0)) << 4) & 0x10;
-    const uint8_t xh_1 = ((qh >> (iqs + 12))     ) & 0x10;
+    const int xh_0 = ((qh >> (iqs +  0)) << 4) & 0x10;
+    const int xh_1 = ((qh >> (iqs + 12))     ) & 0x10;
 
-    const int32_t x0 = ((x[ib].qs[iqs] & 0xf) | xh_0);
-    const int32_t x1 = ((x[ib].qs[iqs] >>  4) | xh_1);
+    v.x = ((x[ib].qs[iqs] & 0xf) | xh_0);
+    v.y = ((x[ib].qs[iqs] >>  4) | xh_1);
 
-    v0 = x0*d + m;
-    v1 = x1*d + m;
+#ifdef GGML_CUDA_DMMV_F16
+    v = __hmul2(v, {d, d});
+    v = __hadd2(v, {m, m});
+#else
+    v.x = (v.x * d) + m;
+    v.y = (v.y * d) + m;
+#endif // GGML_CUDA_DMMV_F16
 }
 
-static __device__ void dequantize_q8_0(const void * vx, const int ib, const int iqs, float & v0, float & v1){
+static __device__ __forceinline__ void dequantize_q8_0(const void * vx, const int ib, const int iqs, dfloat2 & v){
     const block_q8_0 * x = (const block_q8_0 *) vx;
 
-    const float d = x[ib].d;
+    const dfloat d = x[ib].d;
 
-    const int8_t vi0 = x[ib].qs[iqs + 0];
-    const int8_t vi1 = x[ib].qs[iqs + 1];
+    v.x = x[ib].qs[iqs + 0];
+    v.y = x[ib].qs[iqs + 1];
 
-    v0 = vi0*d;
-    v1 = vi1*d;
+#ifdef GGML_CUDA_DMMV_F16
+    v = __hmul2(v, {d, d});
+#else
+    v.x *= d;
+    v.y *= d;
+#endif // GGML_CUDA_DMMV_F16
 }
 
 //================================== k-quants
@@ -479,15 +515,15 @@ static __global__ void dequantize_mul_mat_vec_q2_k(const void * vx, const float
 
     const block_q2_K * x = (const block_q2_K *)vx + ib0;
 
-    const int tid = threadIdx.x/K_QUANTS_PER_ITERATION;  // 0...31
-    const int ix  = threadIdx.x%K_QUANTS_PER_ITERATION;  // 0
+    const int tid = threadIdx.x/K_QUANTS_PER_ITERATION;  // 0...31 or 0...15
+    const int ix  = threadIdx.x%K_QUANTS_PER_ITERATION;  // 0 or 0,1
 
     const int step = 16/K_QUANTS_PER_ITERATION;
 
-    const int im = tid/step;      // 0 or 1. 0 computes 0..., 1 computes 128...
-    const int in = tid - step*im; // 0...7
+    const int im = tid/step;                             // 0 or 1. 0 computes 0..., 1 computes 128...
+    const int in = tid - step*im;                        // 0...15 or 0...7
 
-    const int l0 = K_QUANTS_PER_ITERATION*in;        // 0...14 in steps of 4
+    const int l0 = K_QUANTS_PER_ITERATION*in;            // 0...15 or 0...14 in steps of 2
     const int q_offset = 32*im + l0;
     const int s_offset = 8*im;
     const int y_offset = 128*im + l0;
@@ -542,27 +578,30 @@ static __global__ void dequantize_mul_mat_vec_q2_k(const void * vx, const float
     }
 }
 
-static __global__ void dequantize_mul_mat_vec_q3_k(const void * vx, const float * yy, float * dst, const int ncols) {
+static __global__ void dequantize_mul_mat_vec_q3_k(const void * vx, const float * yy, float * dst, const int ncols, int nrows) {
 
     const uint16_t kmask1 = 0x0303;
     const uint16_t kmask2 = 0x0f0f;
 
-    const int row = blockIdx.x;
+    const int row = blockIdx.y*blockDim.y + threadIdx.y;
+    if (row > nrows) return;
+
     const int num_blocks_per_row = ncols / QK_K;
     const int ib0 = row*num_blocks_per_row;
 
     const block_q3_K * x = (const block_q3_K *)vx + ib0;
 
-    const int tid = threadIdx.x/2;  // 0...15
-    const int ix  = threadIdx.x%2;  // 0, 1
+    const int tid = threadIdx.x/K_QUANTS_PER_ITERATION;  // 0...31 or 0...16
+    const int ix  = threadIdx.x%K_QUANTS_PER_ITERATION;  // 0 or 0,1
 
-    const int n  = 2;           // iterations in the inner loop
-    const int im = tid/8;       // 0 or 1. 0 computes 0..., 1 computes 128...
-    const int in = tid - 8*im;  // 0...7
+    const int n  = K_QUANTS_PER_ITERATION;               // iterations in the inner loop
+    const int step = 16/K_QUANTS_PER_ITERATION;
+    const int im = tid/step;                             // 0 or 1. 0 computes 0..., 1 computes 128...
+    const int in = tid - step*im;                        // 0....15 or 0...7
 
     const uint8_t m = 1 << (4*im);
 
-    const int l0 = n*in;        // 0...28 in steps of 4
+    const int l0 = n*in;                                 // 0...15 or 0...14 in steps of 2
     const int q_offset =  32*im + l0;
     const int y_offset = 128*im + l0;
 
@@ -573,7 +612,7 @@ static __global__ void dequantize_mul_mat_vec_q3_k(const void * vx, const float
 
     float tmp = 0; // partial sum for thread in warp
 
-    for (int i = ix; i < num_blocks_per_row; i += 2) {
+    for (int i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) {
 
         const float   * y  = yy + i * QK_K + y_offset;
         const uint8_t * q = x[i].qs + q_offset;
@@ -614,22 +653,25 @@ static __global__ void dequantize_mul_mat_vec_q3_k(const void * vx, const float
     }
 }
 
-static __global__ void dequantize_mul_mat_vec_q4_k(const void * vx, const float * yy, float * dst, const int ncols) {
+static __global__ void dequantize_mul_mat_vec_q4_k(const void * vx, const float * yy, float * dst, const int ncols, int nrows) {
 
     const uint16_t kmask1 = 0x3f3f;
     const uint16_t kmask2 = 0x0f0f;
     const uint16_t kmask3 = 0xc0c0;
 
-    const int row = blockIdx.x;
+    const int row = blockIdx.y*blockDim.y + threadIdx.y;
+    if (row > nrows) return;
     const int num_blocks_per_row = ncols / QK_K;
     const int ib0 = row*num_blocks_per_row;
 
-    const int tid = threadIdx.x/2;  // 0...15
-    const int ix  = threadIdx.x%2;
+    const int tid = threadIdx.x/K_QUANTS_PER_ITERATION;  // 0...31 or 0...16
+    const int ix  = threadIdx.x%K_QUANTS_PER_ITERATION;  // 0 or 0,1
 
-    const int il  = tid/4;     // 0...3
-    const int ir  = tid - 4*il;// 0...3
-    const int n   = 4;
+    const int step = 8/K_QUANTS_PER_ITERATION;           // 8 or 4
+
+    const int il  = tid/step;                            // 0...3
+    const int ir  = tid - step*il;                       // 0...7 or 0...3
+    const int n   = 2 * K_QUANTS_PER_ITERATION;          // 2 or 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;
@@ -645,7 +687,7 @@ static __global__ void dequantize_mul_mat_vec_q4_k(const void * vx, const float
 
     float tmp = 0; // partial sum for thread in warp
 
-    for (int i = ix; i < num_blocks_per_row; i += 2) {
+    for (int i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) {
 
         const uint8_t * q1 = x[i].qs + q_offset;
         const uint8_t * q2 = q1 + 64;
@@ -700,7 +742,7 @@ static __global__ void dequantize_mul_mat_vec_q5_k(const void * vx, const float
 
     const int il  = tid/4;     // 0...3
     const int ir  = tid - 4*il;// 0...3
-    const int n   = 4;
+    const int n   = 2;
 
     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;
@@ -739,11 +781,16 @@ static __global__ void dequantize_mul_mat_vec_q5_k(const void * vx, const float
         float4 sum = {0.f, 0.f, 0.f, 0.f};
         float smin = 0;
         for (int l = 0; l < n; ++l) {
-            sum.x += y1[l+ 0] * ((ql1[l] & 0xF) + (qh[l] & (hm1 << 0) ? 16 : 0));
-            sum.y += y1[l+32] * ((ql1[l] >>  4) + (qh[l] & (hm1 << 1) ? 16 : 0));
-            sum.z += y2[l+ 0] * ((ql2[l] & 0xF) + (qh[l] & (hm2 << 0) ? 16 : 0));
-            sum.w += y2[l+32] * ((ql2[l] >>  4) + (qh[l] & (hm2 << 1) ? 16 : 0));
-            smin += y1[l] * sc[2] + y1[l+32] * sc[3] + y2[l] * sc[6] + y2[l+32] * sc[7];
+            sum.x += y1[l+ 0] * ((ql1[l+ 0] & 0xF) + (qh[l+ 0] & (hm1 << 0) ? 16 : 0))
+                   + y1[l+16] * ((ql1[l+16] & 0xF) + (qh[l+16] & (hm1 << 0) ? 16 : 0));
+            sum.y += y1[l+32] * ((ql1[l+ 0] >>  4) + (qh[l+ 0] & (hm1 << 1) ? 16 : 0))
+                   + y1[l+48] * ((ql1[l+16] >>  4) + (qh[l+16] & (hm1 << 1) ? 16 : 0));
+            sum.z += y2[l+ 0] * ((ql2[l+ 0] & 0xF) + (qh[l+ 0] & (hm2 << 0) ? 16 : 0))
+                   + y2[l+16] * ((ql2[l+16] & 0xF) + (qh[l+16] & (hm2 << 0) ? 16 : 0));
+            sum.w += y2[l+32] * ((ql2[l+ 0] >>  4) + (qh[l+ 0] & (hm2 << 1) ? 16 : 0))
+                   + y2[l+48] * ((ql2[l+16] >>  4) + (qh[l+16] & (hm2 << 1) ? 16 : 0));
+            smin += (y1[l] + y1[l+16]) * sc[2] + (y1[l+32] + y1[l+48]) * sc[3]
+                  + (y2[l] + y2[l+16]) * sc[6] + (y2[l+32] + y2[l+48]) * sc[7];
         }
         tmp += dall * (sum.x * sc[0] + sum.y * sc[1] + sum.z * sc[4] + sum.w * sc[5]) - dmin * smin;
 
@@ -839,11 +886,12 @@ static __global__ void dequantize_mul_mat_vec_q6_k(const void * vx, const float
     }
 }
 
-static __device__ void convert_f16(const void * vx, const int ib, const int iqs, float & v0, float & v1){
+static __device__ void convert_f16(const void * vx, const int ib, const int iqs, dfloat2 & v){
     const half * x = (const half *) vx;
 
-    v0 = __half2float(x[ib + iqs + 0]);
-    v1 = __half2float(x[ib + iqs + 1]);
+    // automatic half -> float type cast if dfloat == float
+    v.x = x[ib + iqs + 0];
+    v.y = x[ib + iqs + 1];
 }
 
 template <int qk, int qr, dequantize_kernel_t dequantize_kernel>
@@ -860,13 +908,15 @@ static __global__ void dequantize_block(const void * vx, float * y, const int k)
     const int y_offset = qr == 1 ? 1 : qk/2;
 
     // dequantize
-    float & v0 = y[iybs + iqs + 0];
-    float & v1 = y[iybs + iqs + y_offset];
-    dequantize_kernel(vx, ib, iqs, v0, v1);
+    dfloat2 v;
+    dequantize_kernel(vx, ib, iqs, v);
+
+    y[iybs + iqs + 0]        = v.x;
+    y[iybs + iqs + y_offset] = v.y;
 }
 
 template <int qk, int qr, dequantize_kernel_t dequantize_kernel>
-static __global__ void dequantize_mul_mat_vec(const void * vx, const float * y, float * dst, const int ncols, const int nrows) {
+static __global__ void dequantize_mul_mat_vec(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows) {
     // qk = quantized weights per x block
     // qr = number of quantized weights per data value in x block
     const int row = blockIdx.y*blockDim.y + threadIdx.y;
@@ -881,7 +931,12 @@ static __global__ void dequantize_mul_mat_vec(const void * vx, const float * y,
     const int vals_per_iter = iter_stride / WARP_SIZE; // num quantized vals per thread and i iter
     const int y_offset = qr == 1 ? 1 : qk/2;
 
-    float tmp = 0.0f; // partial sum for thread in warp
+// partial sum for each thread
+#ifdef GGML_CUDA_DMMV_F16
+    half2 tmp = {0.0f, 0.0f}; // two sums for f16 to take advantage of half2 intrinsics
+#else
+    float tmp = 0.0f;
+#endif // GGML_CUDA_DMMV_F16
 
     for (int i = 0; i < ncols; i += iter_stride) {
         const int col = i + vals_per_iter*tid;
@@ -895,14 +950,21 @@ static __global__ void dequantize_mul_mat_vec(const void * vx, const float * y,
             // process 2 vals per j iter
 
             // dequantize
-            float v0, v1;
-            dequantize_kernel(vx, ib, iqs + j/qr, v0, v1);
             // for qr = 2 the iqs needs to increase by 1 per j iter because 2 weights per data val
+            dfloat2 v;
+            dequantize_kernel(vx, ib, iqs + j/qr, v);
 
             // matrix multiplication
-            tmp += v0 * y[iybs + iqs + j/qr + 0];
-            tmp += v1 * y[iybs + iqs + j/qr + y_offset];
             // for qr = 2 the y index needs to increase by 1 per j iter because of y_offset = qk/2
+#ifdef GGML_CUDA_DMMV_F16
+            tmp += __hmul2(v, {
+                y[iybs + iqs + j/qr + 0],
+                y[iybs + iqs + j/qr + y_offset]
+            });
+#else
+            tmp += v.x * y[iybs + iqs + j/qr + 0];
+            tmp += v.y * y[iybs + iqs + j/qr + y_offset];
+#endif // GGML_CUDA_DMMV_F16
         }
     }
 
@@ -914,7 +976,11 @@ static __global__ void dequantize_mul_mat_vec(const void * vx, const float * y,
     }
 
     if (tid == 0) {
+#ifdef GGML_CUDA_DMMV_F16
+        dst[row] = tmp.x + tmp.y;
+#else
         dst[row] = tmp;
+#endif // GGML_CUDA_DMMV_F16
     }
 }
 
@@ -1209,7 +1275,7 @@ static void dequantize_row_q6_K_cuda(const void * vx, float * y, const int k, cu
     dequantize_block_q6_K<<<nb, 64, 0, stream>>>(vx, y);
 }
 
-static void dequantize_mul_mat_vec_q4_0_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
+static void dequantize_mul_mat_vec_q4_0_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
     GGML_ASSERT(ncols % GGML_CUDA_DMMV_X == 0);
     const int block_num_y = (nrows + GGML_CUDA_DMMV_Y - 1) / GGML_CUDA_DMMV_Y;
     const dim3 block_nums(1, block_num_y, 1);
@@ -1218,7 +1284,7 @@ static void dequantize_mul_mat_vec_q4_0_cuda(const void * vx, const float * y, f
         <<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
 }
 
-static void dequantize_mul_mat_vec_q4_1_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
+static void dequantize_mul_mat_vec_q4_1_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
     GGML_ASSERT(ncols % GGML_CUDA_DMMV_X == 0);
     const int block_num_y = (nrows + GGML_CUDA_DMMV_Y - 1) / GGML_CUDA_DMMV_Y;
     const dim3 block_nums(1, block_num_y, 1);
@@ -1227,7 +1293,7 @@ static void dequantize_mul_mat_vec_q4_1_cuda(const void * vx, const float * y, f
         <<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
 }
 
-static void dequantize_mul_mat_vec_q5_0_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
+static void dequantize_mul_mat_vec_q5_0_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
     GGML_ASSERT(ncols % GGML_CUDA_DMMV_X == 0);
     const int block_num_y = (nrows + GGML_CUDA_DMMV_Y - 1) / GGML_CUDA_DMMV_Y;
     const dim3 block_nums(1, block_num_y, 1);
@@ -1236,7 +1302,7 @@ static void dequantize_mul_mat_vec_q5_0_cuda(const void * vx, const float * y, f
         <<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
 }
 
-static void dequantize_mul_mat_vec_q5_1_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
+static void dequantize_mul_mat_vec_q5_1_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
     GGML_ASSERT(ncols % GGML_CUDA_DMMV_X == 0);
     const int block_num_y = (nrows + GGML_CUDA_DMMV_Y - 1) / GGML_CUDA_DMMV_Y;
     const dim3 block_nums(1, block_num_y, 1);
@@ -1245,7 +1311,7 @@ static void dequantize_mul_mat_vec_q5_1_cuda(const void * vx, const float * y, f
         <<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
 }
 
-static void dequantize_mul_mat_vec_q8_0_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
+static void dequantize_mul_mat_vec_q8_0_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
     GGML_ASSERT(ncols % GGML_CUDA_DMMV_X == 0);
     const int block_num_y = (nrows + GGML_CUDA_DMMV_Y - 1) / GGML_CUDA_DMMV_Y;
     const dim3 block_nums(1, block_num_y, 1);
@@ -1256,7 +1322,7 @@ static void dequantize_mul_mat_vec_q8_0_cuda(const void * vx, const float * y, f
 
 static void dequantize_mul_mat_vec_q2_K_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
     GGML_ASSERT(ncols % QK_K == 0);
-    const int ny = 2;
+    const int ny = 2; // very slightly faster than 1 even when K_QUANTS_PER_ITERATION = 2
     const int block_num_y = (nrows + ny - 1) / ny;
     const dim3 block_nums(1, block_num_y, 1);
     const dim3 block_dims(32, ny, 1);
@@ -1265,14 +1331,20 @@ static void dequantize_mul_mat_vec_q2_K_cuda(const void * vx, const float * y, f
 
 static void dequantize_mul_mat_vec_q3_K_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
     GGML_ASSERT(ncols % QK_K == 0);
-    const dim3 block_dims(32, 1, 1);
-    dequantize_mul_mat_vec_q3_k<<<nrows, block_dims, 0, stream>>>(vx, y, dst, ncols);
+    const int ny = 2 / K_QUANTS_PER_ITERATION;
+    const int block_num_y = (nrows + ny - 1) / ny;
+    const dim3 block_nums(1, block_num_y, 1);
+    const dim3 block_dims(32, ny, 1);
+    dequantize_mul_mat_vec_q3_k<<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
 }
 
 static void dequantize_mul_mat_vec_q4_K_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
     GGML_ASSERT(ncols % QK_K == 0);
-    const dim3 block_dims(32, 1, 1);
-    dequantize_mul_mat_vec_q4_k<<<nrows, block_dims, 0, stream>>>(vx, y, dst, ncols);
+    const int ny = 2 / K_QUANTS_PER_ITERATION;
+    const int block_num_y = (nrows + ny - 1) / ny;
+    const dim3 block_nums(1, block_num_y, 1);
+    const dim3 block_dims(32, ny, 1);
+    dequantize_mul_mat_vec_q4_k<<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
 }
 
 static void dequantize_mul_mat_vec_q5_K_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
@@ -1295,7 +1367,7 @@ static void convert_fp16_to_fp32_cuda(const void * vx, float * y, const int k, c
     dequantize_block<1, 1, convert_f16><<<num_blocks, CUDA_DEQUANTIZE_BLOCK_SIZE, 0, stream>>>(vx, y, k);
 }
 
-static void convert_mul_mat_vec_f16_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
+static void convert_mul_mat_vec_f16_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
     GGML_ASSERT(ncols % GGML_CUDA_DMMV_X == 0);
     const int block_num_y = (nrows + GGML_CUDA_DMMV_Y - 1) / GGML_CUDA_DMMV_Y;
     const dim3 block_nums(1, block_num_y, 1);
@@ -1463,19 +1535,13 @@ static void * g_scratch_buffer = nullptr;
 static size_t g_scratch_size = 1024*1024*1024; // 1 GB by default
 static size_t g_scratch_offset = 0;
 
-#define GGML_CUDA_MAX_STREAMS 8 // Set this to 1 for reproducible matrix multiplication.
-#define GGML_CUDA_MAX_EVENTS 64
-
 static int g_device_count = -1;
 static int g_main_device = 0;
 static float g_tensor_split[GGML_CUDA_MAX_DEVICES] = {0};
 
 static cublasHandle_t g_cublas_handles[GGML_CUDA_MAX_DEVICES] = {nullptr};
 
-static cudaStream_t g_cudaStreams_main[GGML_CUDA_MAX_DEVICES][GGML_CUDA_MAX_STREAMS] = { nullptr };
-
-static cudaStream_t g_cudaStreams_memcpy_src1[GGML_CUDA_MAX_DEVICES][GGML_CUDA_MAX_STREAMS] = { nullptr };
-static cudaEvent_t g_cudaEvents_memcpy_src1[GGML_CUDA_MAX_DEVICES][GGML_CUDA_MAX_EVENTS] = { nullptr };
+static cudaStream_t g_cudaStreams_main[GGML_CUDA_MAX_DEVICES] = { nullptr };
 
 void ggml_init_cublas() {
     static bool initialized = false;
@@ -1499,15 +1565,8 @@ void ggml_init_cublas() {
         for (int id = 0; id < g_device_count; ++id) {
             CUDA_CHECK(cudaSetDevice(id));
 
-            // create streams
-            for (int i = 0; i < GGML_CUDA_MAX_STREAMS; ++i) {
-                CUDA_CHECK(cudaStreamCreateWithFlags(&g_cudaStreams_main[id][i], cudaStreamNonBlocking));
-                CUDA_CHECK(cudaStreamCreateWithFlags(&g_cudaStreams_memcpy_src1[id][i], cudaStreamNonBlocking));
-            }
-            // create events
-            for (int i = 0; i < GGML_CUDA_MAX_EVENTS; ++i) {
-                CUDA_CHECK(cudaEventCreateWithFlags(&g_cudaEvents_memcpy_src1[id][i], cudaEventDisableTiming));
-            }
+            // create main stream
+            CUDA_CHECK(cudaStreamCreateWithFlags(&g_cudaStreams_main[id], cudaStreamNonBlocking));
 
             // create cublas handle
             CUBLAS_CHECK(cublasCreate(&g_cublas_handles[id]));
@@ -1723,21 +1782,40 @@ inline void ggml_cuda_op_dequantize_mul_mat_vec(
     const int64_t ne00 = src0->ne[0];
     const int64_t nrows = i01_high - i01_low;
 
+// on some GPUs it is faster to convert src1 to half and to use half precision intrinsics
+#ifdef GGML_CUDA_DMMV_F16
+    size_t ash;
+    dfloat * src1_dfloat = nullptr; // dfloat == half
+
+    bool src1_convert_f16 = src0->type == GGML_TYPE_Q4_0 || src0->type == GGML_TYPE_Q4_1 ||
+        src0->type == GGML_TYPE_Q5_0 || src0->type == GGML_TYPE_Q5_1 ||
+        src0->type == GGML_TYPE_Q8_0 || src0->type == GGML_TYPE_F16;
+
+    if (src1_convert_f16) {
+        src1_dfloat = (half *) ggml_cuda_pool_malloc(ne00*sizeof(half), &ash);
+        ggml_cpy_f32_f16_cuda((char *) src1_ddf_i, (char *) src1_dfloat, ne00,
+                                ne00, 1, sizeof(float), 0, 0,
+                                ne00, 1, sizeof(half),  0, 0, cudaStream_main);
+    }
+#else
+    dfloat * src1_dfloat = src1_ddf_i; // dfloat == float, no conversion
+#endif // GGML_CUDA_DMMV_F16
+
     switch (src0->type) {
         case GGML_TYPE_Q4_0:
-            dequantize_mul_mat_vec_q4_0_cuda(src0_ddq_i, src1_ddf_i, dst_ddf_i, ne00, nrows, cudaStream_main);
+            dequantize_mul_mat_vec_q4_0_cuda(src0_ddq_i, src1_dfloat, dst_ddf_i, ne00, nrows, cudaStream_main);
             break;
         case GGML_TYPE_Q4_1:
-            dequantize_mul_mat_vec_q4_1_cuda(src0_ddq_i, src1_ddf_i, dst_ddf_i, ne00, nrows, cudaStream_main);
+            dequantize_mul_mat_vec_q4_1_cuda(src0_ddq_i, src1_dfloat, dst_ddf_i, ne00, nrows, cudaStream_main);
             break;
         case GGML_TYPE_Q5_0:
-            dequantize_mul_mat_vec_q5_0_cuda(src0_ddq_i, src1_ddf_i, dst_ddf_i, ne00, nrows, cudaStream_main);
+            dequantize_mul_mat_vec_q5_0_cuda(src0_ddq_i, src1_dfloat, dst_ddf_i, ne00, nrows, cudaStream_main);
             break;
         case GGML_TYPE_Q5_1:
-            dequantize_mul_mat_vec_q5_1_cuda(src0_ddq_i, src1_ddf_i, dst_ddf_i, ne00, nrows, cudaStream_main);
+            dequantize_mul_mat_vec_q5_1_cuda(src0_ddq_i, src1_dfloat, dst_ddf_i, ne00, nrows, cudaStream_main);
             break;
         case GGML_TYPE_Q8_0:
-            dequantize_mul_mat_vec_q8_0_cuda(src0_ddq_i, src1_ddf_i, dst_ddf_i, ne00, nrows, cudaStream_main);
+            dequantize_mul_mat_vec_q8_0_cuda(src0_ddq_i, src1_dfloat, dst_ddf_i, ne00, nrows, cudaStream_main);
             break;
         case GGML_TYPE_Q2_K:
             dequantize_mul_mat_vec_q2_K_cuda(src0_ddq_i, src1_ddf_i, dst_ddf_i, ne00, nrows, cudaStream_main);
@@ -1755,7 +1833,7 @@ inline void ggml_cuda_op_dequantize_mul_mat_vec(
             dequantize_mul_mat_vec_q6_K_cuda(src0_ddq_i, src1_ddf_i, dst_ddf_i, ne00, nrows, cudaStream_main);
             break;
         case GGML_TYPE_F16:
-            convert_mul_mat_vec_f16_cuda(src0_ddq_i, src1_ddf_i, dst_ddf_i, ne00, nrows, cudaStream_main);
+            convert_mul_mat_vec_f16_cuda(src0_ddq_i, src1_dfloat, dst_ddf_i, ne00, nrows, cudaStream_main);
             break;
         default:
             GGML_ASSERT(false);
@@ -1763,6 +1841,12 @@ inline void ggml_cuda_op_dequantize_mul_mat_vec(
     }
     CUDA_CHECK(cudaGetLastError());
 
+#ifdef GGML_CUDA_DMMV_F16
+    if (src1_convert_f16) {
+        ggml_cuda_pool_free(src1_dfloat, ash);
+    }
+#endif // GGML_CUDA_DMMV_F16
+
     (void) src1;
     (void) dst;
     (void) src0_ddf_i;
@@ -1974,6 +2058,12 @@ static void ggml_cuda_op(const ggml_tensor * src0, const ggml_tensor * src1, ggm
     size_t src1_asf[GGML_CUDA_MAX_DEVICES] = {0};
     size_t  dst_asf[GGML_CUDA_MAX_DEVICES] = {0};
 
+    // if multiple GPUs are used they need to wait for the main GPU to finish
+    if (split && g_device_count > 1) {
+        CUDA_CHECK(cudaSetDevice(g_main_device));
+        CUDA_CHECK(cudaDeviceSynchronize());
+    }
+
     for (int id = 0; id < g_device_count; ++id) {
         if (!split && id != g_main_device) {
             continue;
@@ -2072,9 +2162,7 @@ static void ggml_cuda_op(const ggml_tensor * src0, const ggml_tensor * src1, ggm
                 }
                 const int64_t i11 = i13*ne12 + i12;
 
-                cudaStream_t cudaStream_main        =        g_cudaStreams_main[id][i0 % GGML_CUDA_MAX_STREAMS];
-                cudaStream_t cudaStream_memcpy_src1 = g_cudaStreams_memcpy_src1[id][i0 % GGML_CUDA_MAX_STREAMS];
-                cudaEvent_t  cudaEvent_memcpy_src1  =  g_cudaEvents_memcpy_src1[id][i0 % GGML_CUDA_MAX_EVENTS];
+                cudaStream_t cudaStream_main = g_cudaStreams_main[id];
 
                 // for split tensors the data begins at i0 == i0_offset_low
                 char  * src0_ddq_i = src0_ddq[id] + (i0 - i0_offset_low)*src0_stride*src0_ts/src0_bs;
@@ -2102,14 +2190,14 @@ static void ggml_cuda_op(const ggml_tensor * src0, const ggml_tensor * src1, ggm
                     if (src1->backend == GGML_BACKEND_CPU) {
                         GGML_ASSERT(!flatten_rows || nrows0 == ggml_nrows(src1));
                         int64_t nrows1 = flatten_rows ? nrows0 : ne11;
-                        CUDA_CHECK(ggml_cuda_cpy_tensor_2d(src1_ddf_i, src1, i03, i02, 0, nrows1, cudaStream_memcpy_src1));
+                        CUDA_CHECK(ggml_cuda_cpy_tensor_2d(src1_ddf_i, src1, i03, i02, 0, nrows1, cudaStream_main));
                     } else if (src1->backend == GGML_BACKEND_GPU && src1_is_contiguous) {
                         if (id != g_main_device) {
                             GGML_ASSERT(!flatten_rows);
                             float * src1_ddf_i_source = (float *) src1_extra->data_device[g_main_device];
                             src1_ddf_i_source += i11*src1_stride;
                             CUDA_CHECK(cudaMemcpyAsync(src1_ddf_i, src1_ddf_i_source, src1_stride*sizeof(float),
-                                                    cudaMemcpyDeviceToDevice, cudaStream_memcpy_src1));
+                                                    cudaMemcpyDeviceToDevice, cudaStream_main));
                         }
                     } else if (src1_on_device && !src1_is_contiguous) {
                         GGML_ASSERT(!split);
@@ -2118,7 +2206,6 @@ static void ggml_cuda_op(const ggml_tensor * src0, const ggml_tensor * src1, ggm
                         GGML_ASSERT(false);
                     }
                 }
-                CUDA_CHECK(cudaEventRecord(cudaEvent_memcpy_src1, cudaStream_memcpy_src1));
 
                 if (!src0_on_device || !src0_is_contiguous) {
                     if (src0_is_f32) {
@@ -2134,9 +2221,6 @@ static void ggml_cuda_op(const ggml_tensor * src0, const ggml_tensor * src1, ggm
                     CUDA_CHECK(cudaGetLastError());
                 }
 
-                // wait with main stream until src1 memcpy is done
-                CUDA_CHECK(cudaStreamWaitEvent(cudaStream_main, cudaEvent_memcpy_src1, 0));
-
                 // do the computation
                 op(src0, src1, dst, src0_ddq_i, src0_ddf_i, src1_ddf_i, dst_ddf_i, i02, i01_low, i01_high, i11, cudaStream_main);
 
@@ -2174,8 +2258,13 @@ static void ggml_cuda_op(const ggml_tensor * src0, const ggml_tensor * src1, ggm
 
     // wait until each device is finished, then free their buffers
     for (int id = 0; id < g_device_count; ++id) {
+        if (src0_asq[id] == 0 && src0_asf[id] == 0 && src1_asf[id] == 0 && dst_asf[id] == 0) {
+            continue;
+        }
+
         CUDA_CHECK(cudaSetDevice(id));
         CUDA_CHECK(cudaDeviceSynchronize());
+
         if (src0_asq[id] > 0) {
             ggml_cuda_pool_free(src0_ddq[id], src0_asq[id]);
         }
@@ -2241,7 +2330,7 @@ void ggml_cuda_mul_mat_vec_p021(const ggml_tensor * src0, const ggml_tensor * sr
     const int64_t ne02 = src0->ne[2];
 
     CUDA_CHECK(cudaSetDevice(g_main_device));
-    cudaStream_t cudaStream_main = g_cudaStreams_main[g_main_device][0];
+    cudaStream_t cudaStream_main = g_cudaStreams_main[g_main_device];
 
     struct ggml_tensor_extra_gpu * src0_extra = (ggml_tensor_extra_gpu *) src0->extra;
     void * src0_ddq = src0_extra->data_device[g_main_device];
@@ -2253,8 +2342,6 @@ void ggml_cuda_mul_mat_vec_p021(const ggml_tensor * src0, const ggml_tensor * sr
     float * dst_ddf = (float *) dst_extra->data_device[g_main_device];
 
     ggml_mul_mat_p021_f16_f32_cuda(src0_ddq, src1_ddf, dst_ddf, ne00, ne01, ne02, cudaStream_main);
-
-    CUDA_CHECK(cudaDeviceSynchronize());
 }
 
 void ggml_cuda_mul_mat_vec_nc(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst){
@@ -2272,7 +2359,7 @@ void ggml_cuda_mul_mat_vec_nc(const ggml_tensor * src0, const ggml_tensor * src1
     const int64_t nb02 = src0->nb[2];
 
     CUDA_CHECK(cudaSetDevice(g_main_device));
-    cudaStream_t cudaStream_main = g_cudaStreams_main[g_main_device][0];
+    cudaStream_t cudaStream_main = g_cudaStreams_main[g_main_device];
 
     struct ggml_tensor_extra_gpu * src0_extra = (ggml_tensor_extra_gpu *) src0->extra;
     void * src0_ddq = src0_extra->data_device[g_main_device];
@@ -2287,8 +2374,6 @@ void ggml_cuda_mul_mat_vec_nc(const ggml_tensor * src0, const ggml_tensor * src1
     const int channel_stride_x = nb02 / sizeof(half);
 
     ggml_mul_mat_vec_nc_f16_f32_cuda(src0_ddq, src1_ddf, dst_ddf, ne00, ne01, row_stride_x, ne02, channel_stride_x, cudaStream_main);
-
-    CUDA_CHECK(cudaDeviceSynchronize());
 }
 
 void ggml_cuda_mul_mat(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
@@ -2344,7 +2429,7 @@ void ggml_cuda_cpy(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tens
     const int64_t nb12 = src1->nb[2];
 
     CUDA_CHECK(cudaSetDevice(g_main_device));
-    cudaStream_t cudaStream_main = g_cudaStreams_main[g_main_device][0];
+    cudaStream_t cudaStream_main = g_cudaStreams_main[g_main_device];
 
     const struct ggml_tensor_extra_gpu * src0_extra = (ggml_tensor_extra_gpu *) src0->extra;
     const struct ggml_tensor_extra_gpu * src1_extra = (ggml_tensor_extra_gpu *) src1->extra;
@@ -2362,8 +2447,6 @@ void ggml_cuda_cpy(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tens
         GGML_ASSERT(false);
     }
 
-    CUDA_CHECK(cudaDeviceSynchronize());
-
     (void) dst;
 }
 

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);

ggml-metal.m

@@ -57,6 +57,7 @@ struct ggml_metal_context {
     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);
@@ -66,8 +67,10 @@ struct ggml_metal_context {
     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
 };
@@ -162,6 +165,7 @@ struct ggml_metal_context * ggml_metal_init(void) {
         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);
@@ -171,12 +175,22 @@ struct ggml_metal_context * ggml_metal_init(void) {
         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;
 }
 
@@ -193,10 +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);
@@ -214,7 +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;
@@ -231,30 +249,68 @@ bool ggml_metal_add_buffer(
             }
         }
 
-        size_t page_size = getpagesize();
-        size_t aligned_size = size;
-        if ((aligned_size % page_size) != 0) {
-            aligned_size += (page_size - (aligned_size % page_size));
+        const size_t size_page = getpagesize();
+
+        size_t size_aligned = size;
+        if ((size_aligned % size_page) != 0) {
+            size_aligned += (size_page - (size_aligned % size_page));
         }
 
-        ctx->buffers[ctx->n_buffers].name = name;
-        ctx->buffers[ctx->n_buffers].data = data;
-        ctx->buffers[ctx->n_buffers].size = size;
+        // 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) {
-            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];
+            ctx->buffers[ctx->n_buffers].metal = [ctx->device newBufferWithBytesNoCopy:data length:size_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, aligned_size / 1024.0 / 1024.0);
-            return false;
+            if (ctx->buffers[ctx->n_buffers].metal == nil) {
+                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;
+            }
         }
 
-        ++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;
@@ -735,6 +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) {
@@ -788,6 +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");
                             }
 
@@ -831,4 +959,14 @@ void ggml_metal_graph_compute(
     dispatch_barrier_sync(queue, ^{});
 
     [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,
@@ -485,6 +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,
@@ -540,6 +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,

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));

ggml.h

@@ -303,6 +303,7 @@ extern "C" {
         GGML_OP_STEP,
         GGML_OP_RELU,
         GGML_OP_GELU,
+        GGML_OP_GELU_QUICK,
         GGML_OP_SILU,
         GGML_OP_SILU_BACK,
         GGML_OP_NORM, // normalize
@@ -331,12 +332,15 @@ extern "C" {
         GGML_OP_ROPE_BACK,
         GGML_OP_ALIBI,
         GGML_OP_CLAMP,
-        GGML_OP_CONV_1D_1S,
-        GGML_OP_CONV_1D_2S,
+        GGML_OP_CONV_1D_S1_PH,
+        GGML_OP_CONV_1D_S2_PH,
+        GGML_OP_CONV_2D_SK_P0,
 
         GGML_OP_FLASH_ATTN,
         GGML_OP_FLASH_FF,
         GGML_OP_FLASH_ATTN_BACK,
+        GGML_OP_WIN_PART,
+        GGML_OP_WIN_UNPART,
 
         GGML_OP_MAP_UNARY,
         GGML_OP_MAP_BINARY,
@@ -500,8 +504,9 @@ 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 size_t  ggml_get_mem_size  (struct ggml_context * ctx);
+    GGML_API void *  ggml_get_mem_buffer     (const 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,
@@ -556,8 +561,8 @@ extern "C" {
     GGML_API void *  ggml_get_data    (const struct ggml_tensor * tensor);
     GGML_API float * ggml_get_data_f32(const struct ggml_tensor * tensor);
 
-    GGML_API const char * ggml_get_name(const struct ggml_tensor * tensor);
-    GGML_API void         ggml_set_name(struct ggml_tensor * tensor, const char * name);
+    GGML_API const char *         ggml_get_name(const struct ggml_tensor * tensor);
+    GGML_API struct ggml_tensor * ggml_set_name(struct ggml_tensor * tensor, const char * name);
 
     //
     // operations on tensors with backpropagation
@@ -610,24 +615,47 @@ extern "C" {
             struct ggml_tensor  * a,
             struct ggml_tensor  * b);
 
+    GGML_API struct ggml_tensor * ggml_sub_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b);
+
     GGML_API struct ggml_tensor * ggml_mul(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
             struct ggml_tensor  * b);
 
+    GGML_API struct ggml_tensor * ggml_mul_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b);
+
     GGML_API struct ggml_tensor * ggml_div(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
             struct ggml_tensor  * b);
 
+    GGML_API struct ggml_tensor * ggml_div_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b);
+
     GGML_API struct ggml_tensor * ggml_sqr(
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
 
+    GGML_API struct ggml_tensor * ggml_sqr_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
     GGML_API struct ggml_tensor * ggml_sqrt(
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
 
+    GGML_API struct ggml_tensor * ggml_sqrt_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
     GGML_API struct ggml_tensor * ggml_log(
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
@@ -667,31 +695,67 @@ extern "C" {
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
 
+    GGML_API struct ggml_tensor * ggml_abs_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
     GGML_API struct ggml_tensor * ggml_sgn(
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
 
+    GGML_API struct ggml_tensor * ggml_sgn_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
     GGML_API struct ggml_tensor * ggml_neg(
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
 
+    GGML_API struct ggml_tensor * ggml_neg_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
     GGML_API struct ggml_tensor * ggml_step(
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
 
+    GGML_API struct ggml_tensor * ggml_step_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
     GGML_API struct ggml_tensor * ggml_relu(
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
 
+    GGML_API struct ggml_tensor * ggml_relu_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
     // TODO: double-check this computation is correct
     GGML_API struct ggml_tensor * ggml_gelu(
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
 
+    GGML_API struct ggml_tensor * ggml_gelu_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_gelu_quick(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_gelu_quick_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
     GGML_API struct ggml_tensor * ggml_silu(
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
 
+    GGML_API struct ggml_tensor * ggml_silu_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
     // a - x
     // b - dy
     GGML_API struct ggml_tensor * ggml_silu_back(
@@ -705,10 +769,18 @@ extern "C" {
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
 
+    GGML_API struct ggml_tensor * ggml_norm_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
     GGML_API struct ggml_tensor * ggml_rms_norm(
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
 
+    GGML_API struct ggml_tensor * ggml_rms_norm_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
     // a - x
     // b - dy
     GGML_API struct ggml_tensor * ggml_rms_norm_back(
@@ -998,16 +1070,55 @@ extern "C" {
             float                 min,
             float                 max);
 
-    // padding = 1
+    // TODO: implement general-purpose convolutions
+    // GGML_API struct ggml_tensor * ggml_conv_1d(
+    //        struct ggml_context * ctx,
+    //        struct ggml_tensor  * a,
+    //        struct ggml_tensor  * b,
+    //        int                   s0
+    //        int                   p0,
+    //        int                   d0);
+    //
+    // GGML_API struct ggml_tensor * ggml_conv_2d(
+    //        struct ggml_context * ctx,
+    //        struct ggml_tensor  * a,
+    //        struct ggml_tensor  * b,
+    //        int                   s0,
+    //        int                   s1,
+    //        int                   p0,
+    //        int                   p1,
+    //        int                   d0,
+    //        int                   d1);
+
+    // padding = half
     // TODO: we don't support extra parameters for now
     //       that's why we are hard-coding the stride, padding, and dilation
     //       not great ..
-    GGML_API struct ggml_tensor * ggml_conv_1d_1s(
+    // example:
+    // a:      3   80  768    1
+    // b:   3000   80    1    1
+    // res: 3000  768    1    1
+    // used in whisper
+    GGML_API struct ggml_tensor * ggml_conv_1d_s1_ph(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
             struct ggml_tensor  * b);
 
-    GGML_API struct ggml_tensor * ggml_conv_1d_2s(
+    // used in whisper
+    GGML_API struct ggml_tensor * ggml_conv_1d_s2_ph(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b);
+
+    // kernel size is a->ne[0] x a->ne[1]
+    // stride is equal to kernel size
+    // padding is zero
+    // example:
+    // a:     16   16    3  768
+    // b:   1024 1024    3    1
+    // res:   64   64  768    1
+    // used in sam
+    GGML_API struct ggml_tensor * ggml_conv_2d_sk_p0(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
             struct ggml_tensor  * b);
@@ -1035,6 +1146,26 @@ extern "C" {
             struct ggml_tensor  * c0,
             struct ggml_tensor  * c1);
 
+    // partition into non-overlapping windows with padding if needed
+    // example:
+    // a:   768   64   64    1
+    // w:    14
+    // res: 768   14   14    25
+    // used in sam
+    GGML_API struct ggml_tensor * ggml_win_part(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            int                   w);
+
+    // reverse of ggml_win_part
+    // used in sam
+    GGML_API struct ggml_tensor * ggml_win_unpart(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            int                   w0,
+            int                   h0,
+            int                   w);
+
     // Mapping operations
     typedef void (*ggml_unary_op_f32_t)(const int, float *, const float *);
     typedef void (*ggml_binary_op_f32_t)(const int, float *, const float *, const float *);

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>
@@ -886,6 +891,7 @@ 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;
@@ -904,6 +910,7 @@ 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);
@@ -918,21 +925,21 @@ static bool kv_cache_init(
 
 struct llama_context_params llama_context_default_params() {
     struct llama_context_params result = {
+        /*.seed                        =*/ -1,
         /*.n_ctx                       =*/ 512,
         /*.n_batch                     =*/ 512,
         /*.gpu_layers                  =*/ 0,
         /*.main_gpu                    =*/ 0,
         /*.tensor_split                =*/ {0},
+        /*.progress_callback           =*/ nullptr,
+        /*.progress_callback_user_data =*/ nullptr,
         /*.low_vram                    =*/ false,
-        /*.seed                        =*/ -1,
         /*.f16_kv                      =*/ true,
         /*.logits_all                  =*/ false,
         /*.vocab_only                  =*/ false,
         /*.use_mmap                    =*/ true,
         /*.use_mlock                   =*/ false,
         /*.embedding                   =*/ false,
-        /*.progress_callback           =*/ nullptr,
-        /*.progress_callback_user_data =*/ nullptr,
     };
 
     return result;
@@ -1253,7 +1260,7 @@ static void llama_model_load_internal(
             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);
             }
         }
@@ -1613,7 +1620,7 @@ 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);
@@ -1650,11 +1657,7 @@ static bool llama_eval_internal(
     {
         cur = ggml_rms_norm(ctx0, inpL);
         offload_func_nr(cur);
-        ggml_set_name(cur, "rms_norm_inpL");
-
-        cur = ggml_rms_norm(ctx0, cur);
-        offload_func_nr(cur);
-        ggml_set_name(cur, "rms_norm_after");
+        ggml_set_name(cur, "rms_norm_2");
 
         // cur = cur*norm(broadcasted)
         cur = ggml_mul(ctx0, cur, model.norm);
@@ -2489,8 +2492,23 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
         } else {
             new_type = quantized_type;
 #ifdef GGML_USE_K_QUANTS
+            if (quantized_type == GGML_TYPE_Q2_K || quantized_type == GGML_TYPE_Q3_K || quantized_type == GGML_TYPE_Q4_K ||
+                quantized_type == GGML_TYPE_Q5_K || quantized_type == GGML_TYPE_Q6_K) {
+                int nx = tensor.ne.at(0);
+                int ny = tensor.ne.at(1);
+                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;
+                int nx = tensor.ne.at(0);
+                int ny = tensor.ne.at(1);
+                if (nx % QK_K == 0 && ny % QK_K == 0) {
+                    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;
@@ -2694,16 +2712,21 @@ struct llama_context * llama_init_from_file(
         // 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_ptr  = ctx->model.mapping->addr;
+            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_ptr  = ggml_get_mem_buffer(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__);               \
@@ -2711,12 +2734,13 @@ struct llama_context * llama_init_from_file(
         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, "eval", ctx->buf_compute.addr, ctx->buf_compute.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, "kv",   ctx->model.kv_self.buf.addr, ctx->model.kv_self.buf.size));
-        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, "scr1", ctx->buf_scratch[1].addr,    ctx->buf_scratch[1].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, "kv",   ctx->model.kv_self.buf.addr, ctx->model.kv_self.buf.size, 0));
+
+        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
@@ -3102,9 +3126,7 @@ size_t llama_copy_state_data(struct llama_context * ctx, uint8_t * dst) {
         if (kv_size) {
             const size_t elt_size = ggml_element_size(kv_self.k);
 
-            char buffer[4096];
-
-            ggml_context * cpy_ctx = ggml_init({ sizeof(buffer), buffer, /* no_alloc */ true });
+            ggml_context * cpy_ctx = ggml_init({ 4096, NULL, /* no_alloc */ true });
             ggml_cgraph gf{};
             gf.n_threads = 1;
 
@@ -3210,9 +3232,7 @@ size_t llama_set_state_data(struct llama_context * ctx, uint8_t * src) {
 
             const size_t elt_size = ggml_element_size(kv_self.k);
 
-            char buffer[4096];
-
-            ggml_context * cpy_ctx = ggml_init({ sizeof(buffer), buffer, /* no_alloc */ true });
+            ggml_context * cpy_ctx = ggml_init({ 4096, NULL, /* no_alloc */ true });
             ggml_cgraph gf{};
             gf.n_threads = 1;
 
@@ -3447,9 +3467,12 @@ 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: 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);
-    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:      sample time = %8.2f ms / %5d runs   (%8.2f ms per token, %8.2f tokens per second)\n",
+            __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: 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

@@ -71,28 +71,27 @@ extern "C" {
 
     typedef void (*llama_progress_callback)(float progress, void *ctx);
 
-    struct llama_context_params {
+   struct llama_context_params {
+        int seed;                              // RNG seed, -1 for random
         int n_ctx;                             // text context
         int n_batch;                           // prompt processing batch size
         int n_gpu_layers;                      // number of layers to store in VRAM
         int main_gpu;                          // the GPU that is used for scratch and small tensors
         float tensor_split[LLAMA_MAX_DEVICES]; // how to split layers across multiple GPUs
-        bool low_vram;                         // if true, reduce VRAM usage at the cost of performance
-        int seed;                              // RNG seed, -1 for random
+        // called with a progress value between 0 and 1, pass NULL to disable
+        llama_progress_callback progress_callback;
+        // context pointer passed to the progress callback
+        void * progress_callback_user_data;
 
+        // Keep the booleans together to avoid misalignment during copy-by-value.
+        bool low_vram;   // if true, reduce VRAM usage at the cost of performance
         bool f16_kv;     // use fp16 for KV cache
         bool logits_all; // the llama_eval() call computes all logits, not just the last one
         bool vocab_only; // only load the vocabulary, no weights
         bool use_mmap;   // use mmap if possible
         bool use_mlock;  // force system to keep model in RAM
         bool embedding;  // embedding mode only
-
-        // called with a progress value between 0 and 1, pass NULL to disable
-        llama_progress_callback progress_callback;
-        // context pointer passed to the progress callback
-        void * progress_callback_user_data;
     };
-
     // model file types
     enum llama_ftype {
         LLAMA_FTYPE_ALL_F32              = 0,

scripts/verify-checksum-models.py

@@ -1,9 +1,10 @@
 import os
 import hashlib
 
+
 def sha256sum(file):
     block_size = 16 * 1024 * 1024  # 16 MB block size
-    b  = bytearray(block_size)
+    b = bytearray(block_size)
     file_hash = hashlib.sha256()
     mv = memoryview(b)
     with open(file, 'rb', buffering=0) as f:
@@ -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))