ggml : optimize non-SIMD Q4_0 vector dot product (#703)

which allows us to use aligned_alloc or _aligned_malloc functions

.devops/main.Dockerfile

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

.dockerignore

@@ -21,4 +21,4 @@ models/*
 
 arm_neon.h
 compile_commands.json
-Dockerfile
+Dockerfile

.ecrc

@@ -0,0 +1,5 @@
+{
+  "Disable": {
+    "IndentSize": true
+  }
+}

.editorconfig

@@ -0,0 +1,19 @@
+# https://EditorConfig.org
+
+# Top-most EditorConfig file
+root = true
+
+# Unix-style newlines with a newline ending every file, utf-8 charset
+[*]
+end_of_line = lf
+insert_final_newline = true
+trim_trailing_whitespace = true
+charset = utf-8
+indent_style = space
+indent_size = 4
+
+[Makefile]
+indent_style = tab
+
+[prompts/*.txt]
+insert_final_newline = unset

.github/ISSUE_TEMPLATE/custom.md

@@ -22,9 +22,9 @@ Please provide a detailed written description of what you were trying to do, and
 
 # Current Behavior
 
-Please provide a detailed written description of what `llama.cpp` did, instead. 
+Please provide a detailed written description of what `llama.cpp` did, instead.
 
-# Environment and Context 
+# Environment and Context
 
 Please provide detailed information about your computer setup. This is important in case the issue is not reproducible except for under certain specific conditions.
 
@@ -133,7 +133,7 @@ llama_model_load: loading model part 8/8 from './models/65B/ggml-model-q4_0.bin.
 llama_model_load: .......................................................................................... done
 llama_model_load: model size =  4869.09 MB / num tensors = 723
 
-system_info: n_threads = 16 / 32 | AVX = 1 | AVX2 = 1 | AVX512 = 0 | FMA = 1 | NEON = 0 | ARM_FMA = 0 | F16C = 1 | FP16_VA = 0 | WASM_SIMD = 0 | BLAS = 0 | SSE3 = 1 | VSX = 0 | 
+system_info: n_threads = 16 / 32 | AVX = 1 | AVX2 = 1 | AVX512 = 0 | FMA = 1 | NEON = 0 | ARM_FMA = 0 | F16C = 1 | FP16_VA = 0 | WASM_SIMD = 0 | BLAS = 0 | SSE3 = 1 | VSX = 0 |
 
 main: prompt: 'Please close your issue when it has been answered.'
 main: number of tokens in prompt = 11
@@ -166,14 +166,14 @@ main:    total time = 246406.42 ms
 
  Performance counter stats for './main -m ./models/65B/ggml-model-q4_0.bin -t 16 -n 1024 -p Please close your issue when it has been answered.':
 
-        3636882.89 msec task-clock                #   14.677 CPUs utilized          
-             13509      context-switches          #    3.714 /sec                   
-              2436      cpu-migrations            #    0.670 /sec                   
-          10476679      page-faults               #    2.881 K/sec                  
+        3636882.89 msec task-clock                #   14.677 CPUs utilized
+             13509      context-switches          #    3.714 /sec
+              2436      cpu-migrations            #    0.670 /sec
+          10476679      page-faults               #    2.881 K/sec
     13133115082869      cycles                    #    3.611 GHz                      (16.77%)
        29314462753      stalled-cycles-frontend   #    0.22% frontend cycles idle     (16.76%)
     10294402631459      stalled-cycles-backend    #   78.39% backend cycles idle      (16.74%)
-    23479217109614      instructions              #    1.79  insn per cycle         
+    23479217109614      instructions              #    1.79  insn per cycle
                                                   #    0.44  stalled cycles per insn  (16.76%)
      2353072268027      branches                  #  647.002 M/sec                    (16.77%)
         1998682780      branch-misses             #    0.08% of all branches          (16.76%)

.github/workflows/docker.yml

@@ -60,4 +60,4 @@ jobs:
           push: ${{ github.event_name == 'push' }}
           platforms: linux/amd64,linux/arm64
           tags: "ghcr.io/ggerganov/llama.cpp:${{ matrix.config.tag }}"
-          file: ${{ matrix.config.dockerfile }}
+          file: ${{ matrix.config.dockerfile }}

.github/workflows/editorconfig.yml

@@ -0,0 +1,17 @@
+name: EditorConfig Checker
+
+on:
+  push:
+    branches:
+      - master
+  pull_request:
+    branches:
+      - master
+
+jobs:
+  editorconfig:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v3
+      - uses: editorconfig-checker/action-editorconfig-checker@main
+      - run: editorconfig-checker

.gitignore

@@ -19,6 +19,7 @@ models/*
 
 /main
 /quantize
+/quantize-stats
 /result
 /perplexity
 /embedding

CMakeLists.txt

@@ -56,6 +56,10 @@ option(LLAMA_AVX                    "llama: enable AVX"
 option(LLAMA_AVX2                   "llama: enable AVX2"                                    ON)
 option(LLAMA_AVX512                 "llama: enable AVX512"                                  OFF)
 option(LLAMA_FMA                    "llama: enable FMA"                                     ON)
+# in MSVC F16C is implied with AVX2/AVX512
+if (NOT MSVC)
+    option(LLAMA_F16C               "llama: enable F16C"                                    ON)
+endif()
 
 # 3rd party libs
 option(LLAMA_ACCELERATE             "llama: enable Accelerate framework"                    ON)
@@ -115,6 +119,7 @@ if (LLAMA_OPENBLAS)
 
         add_compile_definitions(GGML_USE_OPENBLAS)
         add_link_options(${BLAS_LIBRARIES})
+        set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} openblas)
     else()
         message(WARNING "OpenBLAS not found")
     endif()
@@ -139,6 +144,7 @@ if (LLAMA_ALL_WARNINGS)
             -Wpedantic
             -Wcast-qual
             -Wno-unused-function
+            -Wno-multichar
         )
     else()
         # todo : msvc
@@ -151,6 +157,10 @@ if (LLAMA_ALL_WARNINGS)
 
 endif()
 
+if (MSVC)
+    add_compile_definitions(_CRT_SECURE_NO_WARNINGS)
+endif()
+
 if (LLAMA_LTO)
     include(CheckIPOSupported)
     check_ipo_supported(RESULT result OUTPUT output)
@@ -201,7 +211,9 @@ elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "^(x86_64|i686|AMD64)$")
             add_compile_options(/arch:AVX)
         endif()
     else()
-        add_compile_options(-mf16c)
+        if (LLAMA_F16C)
+            add_compile_options(-mf16c)
+        endif()
         if (LLAMA_FMA)
             add_compile_options(-mfma)
         endif()
@@ -240,7 +252,9 @@ endif()
 
 add_library(llama
             llama.cpp
-            llama.h)
+            llama.h
+            llama_internal.h
+            llama_util.h)
 
 target_include_directories(llama PUBLIC .)
 target_compile_features(llama PUBLIC cxx_std_11) # don't bump

Makefile

@@ -37,7 +37,7 @@ LDFLAGS  =
 
 # warnings
 CFLAGS   += -Wall -Wextra -Wpedantic -Wcast-qual -Wdouble-promotion -Wshadow -Wstrict-prototypes -Wpointer-arith -Wno-unused-function
-CXXFLAGS += -Wall -Wextra -Wpedantic -Wcast-qual -Wno-unused-function
+CXXFLAGS += -Wall -Wextra -Wpedantic -Wcast-qual -Wno-unused-function -Wno-multichar
 
 # OS specific
 # TODO: support Windows
@@ -142,14 +142,14 @@ default: main quantize perplexity embedding
 ggml.o: ggml.c ggml.h
 	$(CC)  $(CFLAGS)   -c ggml.c -o ggml.o
 
-llama.o: llama.cpp llama.h
+llama.o: llama.cpp llama.h llama_util.h llama_internal.h
 	$(CXX) $(CXXFLAGS) -c llama.cpp -o llama.o
 
 common.o: examples/common.cpp examples/common.h
 	$(CXX) $(CXXFLAGS) -c examples/common.cpp -o common.o
 
 clean:
-	rm -vf *.o main quantize perplexity embedding
+	rm -vf *.o main quantize quantize-stats perplexity embedding benchmark-q4_0-matmult
 
 main: examples/main/main.cpp ggml.o llama.o common.o
 	$(CXX) $(CXXFLAGS) examples/main/main.cpp ggml.o llama.o common.o -o main $(LDFLAGS)
@@ -160,16 +160,26 @@ main: examples/main/main.cpp ggml.o llama.o common.o
 quantize: examples/quantize/quantize.cpp ggml.o llama.o
 	$(CXX) $(CXXFLAGS) examples/quantize/quantize.cpp ggml.o llama.o -o quantize $(LDFLAGS)
 
+quantize-stats: examples/quantize-stats/quantize-stats.cpp ggml.o llama.o
+	$(CXX) $(CXXFLAGS) examples/quantize-stats/quantize-stats.cpp ggml.o llama.o -o quantize-stats $(LDFLAGS)
+
 perplexity: examples/perplexity/perplexity.cpp ggml.o llama.o common.o
 	$(CXX) $(CXXFLAGS) examples/perplexity/perplexity.cpp ggml.o llama.o common.o -o perplexity $(LDFLAGS)
 
 embedding: examples/embedding/embedding.cpp ggml.o llama.o common.o
 	$(CXX) $(CXXFLAGS) examples/embedding/embedding.cpp ggml.o llama.o common.o -o embedding $(LDFLAGS)
 
+libllama.so: llama.o ggml.o
+	$(CXX) $(CXXFLAGS) -shared -fPIC -o libllama.so llama.o ggml.o $(LDFLAGS)
+
 #
 # Tests
 #
 
+benchmark: ggml.o
+	$(CXX) $(CXXFLAGS) examples/benchmark/benchmark-q4_0-matmult.c ggml.o -o benchmark-q4_0-matmult $(LDFLAGS)
+	./benchmark-q4_0-matmult
+
 .PHONY: tests
 tests:
 	bash ./tests/run-tests.sh

README.md

@@ -9,6 +9,7 @@ Inference of [LLaMA](https://arxiv.org/abs/2302.13971) model in pure C/C++
 
 **Hot topics:**
 
+- [Add GPU support to ggml](https://github.com/ggerganov/llama.cpp/discussions/915)
 - [Roadmap Apr 2023](https://github.com/ggerganov/llama.cpp/discussions/784)
 
 ## Description
@@ -42,11 +43,13 @@ New features will probably be added mostly through community contributions.
 - [X] [Chinese LLaMA / Alpaca](https://github.com/ymcui/Chinese-LLaMA-Alpaca)
 - [X] [Vigogne (French)](https://github.com/bofenghuang/vigogne)
 - [X] [Vicuna](https://github.com/ggerganov/llama.cpp/discussions/643#discussioncomment-5533894)
+- [X] [Koala](https://bair.berkeley.edu/blog/2023/04/03/koala/)
 
 **Bindings:**
 
 - Python: [abetlen/llama-cpp-python](https://github.com/abetlen/llama-cpp-python)
 - Go: [go-skynet/go-llama.cpp](https://github.com/go-skynet/go-llama.cpp)
+- Node.js: [hlhr202/llama-node](https://github.com/hlhr202/llama-node)
 
 **UI:**
 
@@ -147,21 +150,43 @@ https://user-images.githubusercontent.com/1991296/224442907-7693d4be-acaa-4e01-8
 
 ## Usage
 
-Here are the step for the LLaMA-7B model:
+Here are the step for the LLaMA-7B model.
+
+### Get the Code
 
 ```bash
-# build this repo
 git clone https://github.com/ggerganov/llama.cpp
 cd llama.cpp
-make
+```
 
-#For Windows and CMake, use the following command instead:
-cd <path_to_llama_folder>
-mkdir build
-cd build
-cmake ..
-cmake --build . --config Release
+### Build
 
+Note: For Windows, CMake or Zig can be used.
+
+1. Use `make`
+
+    ```bash
+    make
+    ```
+
+1. Use CMake
+
+    ```bash
+    mkdir build
+    cd build
+    cmake ..
+    cmake --build . --config Release
+    ```
+
+1. Use Zig
+
+    ```bash
+    zig build -Drelease-fast
+    ```
+
+### Prepare Data & Run
+
+```bash
 # obtain the original LLaMA model weights and place them in ./models
 ls ./models
 65B 30B 13B 7B tokenizer_checklist.chk tokenizer.model
@@ -179,8 +204,6 @@ python3 convert-pth-to-ggml.py models/7B/ 1
 ./main -m ./models/7B/ggml-model-q4_0.bin -n 128
 ```
 
-Currently, it's best to use Python 3.9 or Python 3.10, as `sentencepiece` has not yet published a wheel for Python 3.11.
-
 When running the larger models, make sure you have enough disk space to store all the intermediate files.
 
 ### Memory/Disk Requirements
@@ -242,7 +265,7 @@ There 26 letters in the English Alphabet
 The majority (54%) are using public transit. This includes buses, trams and metros with over 100 lines throughout the city which make it very accessible for tourists to navigate around town as well as locals who commute by tram or metro on a daily basis
 > List 5 words that start with "ca".
 cadaver, cauliflower, cabbage (vegetable), catalpa (tree) and Cailleach.
-> 
+>
 ```
 
 ### Using [GPT4All](https://github.com/nomic-ai/gpt4all)
@@ -253,19 +276,19 @@ cadaver, cauliflower, cabbage (vegetable), catalpa (tree) and Cailleach.
 convert the model from the old format to the new format with [./migrate-ggml-2023-03-30-pr613.py](./migrate-ggml-2023-03-30-pr613.py):
 
   ```bash
-  python3 convert-gpt4all-to-ggml.py models/gpt4all-7B/gpt4all-lora-quantized.bin ./models/tokenizer.model 
+  python3 convert-gpt4all-to-ggml.py models/gpt4all-7B/gpt4all-lora-quantized.bin ./models/tokenizer.model
   python3 migrate-ggml-2023-03-30-pr613.py models/gpt4all-7B/gpt4all-lora-quantized.bin models/gpt4all-7B/gpt4all-lora-quantized-new.bin
   ```
-  
+
 - You can now use the newly generated `gpt4all-lora-quantized-new.bin` model in exactly the same way as all other models
 - The original model is saved in the same folder with a suffix `.orig`
 
 ### Obtaining and verifying the Facebook LLaMA original model and Stanford Alpaca model data
 
 - **Under no circumstances share IPFS, magnet links, or any other links to model downloads anywhere in this respository, including in issues, discussions or pull requests. They will be immediately deleted.**
-- The LLaMA models are officially distributed by Facebook and will **never** be provided through this repository. 
+- The LLaMA models are officially distributed by Facebook and will **never** be provided through this repository.
 - Refer to [Facebook's LLaMA repository](https://github.com/facebookresearch/llama/pull/73/files) if you need to request access to the model data.
-- Please verify the sha256 checksums of all downloaded model files to confirm that you have the correct model data files before creating an issue relating to your model files.
+- Please verify the [sha256 checksums](SHA256SUMS) of all downloaded model files to confirm that you have the correct model data files before creating an issue relating to your model files.
 - The following command will verify if you have all possible latest files in your self-installed `./models` subdirectory:
 
   `sha256sum --ignore-missing -c SHA256SUMS` on Linux
@@ -283,7 +306,7 @@ convert the model from the old format to the new format with [./migrate-ggml-202
   - GPT-3.5 / InstructGPT / ChatGPT:
     - [Aligning language models to follow instructions](https://openai.com/research/instruction-following)
     - [Training language models to follow instructions with human feedback](https://arxiv.org/abs/2203.02155)
-    
+
 ### Perplexity (Measuring model quality)
 
 You can use the `perplexity` example to measure perplexity over the given prompt.  For more background,

build.zig

@@ -1,14 +1,14 @@
 const std = @import("std");
 
-pub fn build(b: *std.Build) void {
+pub fn build(b: *std.build.Builder) void {
     const target = b.standardTargetOptions(.{});
-    const optimize = b.standardOptimizeOption(.{});
+    const optimize = b.standardReleaseOptions();
+    const want_lto = b.option(bool, "lto", "Want -fLTO");
 
-    const lib = b.addStaticLibrary(.{
-        .name = "llama",
-        .target = target,
-        .optimize = optimize,
-    });
+    const lib = b.addStaticLibrary("llama", null);
+    lib.want_lto = want_lto;
+    lib.setTarget(target);
+    lib.setBuildMode(optimize);
     lib.linkLibCpp();
     lib.addIncludePath(".");
     lib.addIncludePath("examples");
@@ -17,11 +17,11 @@ pub fn build(b: *std.Build) void {
     }, &.{"-std=c11"});
     lib.addCSourceFiles(&.{
         "llama.cpp",
-        "examples/common.cpp",
     }, &.{"-std=c++11"});
     lib.install();
 
-    const build_args = .{ .b = b, .lib = lib, .target = target, .optimize = optimize };
+    const build_args = .{ .b = b, .lib = lib, .target = target, .optimize = optimize, .want_lto = want_lto };
+
     const exe = build_example("main", build_args);
     _ = build_example("quantize", build_args);
     _ = build_example("perplexity", build_args);
@@ -42,18 +42,17 @@ pub fn build(b: *std.Build) void {
 fn build_example(comptime name: []const u8, args: anytype) *std.build.LibExeObjStep {
     const b = args.b;
     const lib = args.lib;
-    const target = args.target;
-    const optimize = args.optimize;
+    const want_lto = args.want_lto;
 
-    const exe = b.addExecutable(.{
-        .name = name,
-        .target = target,
-        .optimize = optimize,
-    });
+    const exe = b.addExecutable(name, null);
+    exe.want_lto = want_lto;
+    lib.setTarget(args.target);
+    lib.setBuildMode(args.optimize);
     exe.addIncludePath(".");
     exe.addIncludePath("examples");
     exe.addCSourceFiles(&.{
         std.fmt.comptimePrint("examples/{s}/{s}.cpp", .{name, name}),
+        "examples/common.cpp",
     }, &.{"-std=c++11"});
     exe.linkLibrary(lib);
     exe.install();

examples/CMakeLists.txt

@@ -31,6 +31,7 @@ if (EMSCRIPTEN)
 else()
     add_subdirectory(main)
     add_subdirectory(quantize)
+    add_subdirectory(quantize-stats)
     add_subdirectory(perplexity)
     add_subdirectory(embedding)
 endif()

examples/Miku.sh

@@ -19,15 +19,15 @@ GEN_OPTIONS=(--batch_size 1024
 --top_p 0.5)
 
 if [ -n "$N_THREAD" ]; then
-	GEN_OPTIONS+=(--threads "$N_THREAD")
+    GEN_OPTIONS+=(--threads "$N_THREAD")
 fi
 
 ./main "${GEN_OPTIONS[@]}" \
-	--model "$MODEL" \
-	--n_predict "$N_PREDICTS" \
-	--color --interactive \
-	--reverse-prompt "${USER_NAME}:" \
-	--prompt "
+    --model "$MODEL" \
+    --n_predict "$N_PREDICTS" \
+    --color --interactive \
+    --reverse-prompt "${USER_NAME}:" \
+    --prompt "
 This is a transcript of a 1000 page, never ending conversation between ${USER_NAME} and the cute and helpful AI assistant ${AI_NAME}. ${AI_NAME} is a girl who is an AI running on the users computer.
 ${AI_NAME} can think for herself without the user seeing her thoughts by adding a /think prefix to her output. She uses this to reason about the world and to think about what she should say next.
 ${AI_NAME} is always coherent and makes sense, but if she isn't sure if what she is saying is correct she will ask the user for help.

examples/alpaca.sh

@@ -7,4 +7,4 @@
 cd `dirname $0`
 cd ..
 
-./main -m ./models/ggml-alpaca-7b-q4.bin --color -f ./prompts/alpaca.txt -ins -b 256 --top_k 10000 --temp 0.2 --repeat_penalty 1 -t 7
+./main -m ./models/ggml-alpaca-7b-q4.bin --color -f ./prompts/alpaca.txt --ctx_size 2048 -n -1 -ins -b 256 --top_k 10000 --temp 0.2 --repeat_penalty 1 -t 7

examples/benchmark/benchmark-q4_0-matmult.c

@@ -0,0 +1,270 @@
+/*
+    License: MIT License
+
+    Changelog:
+    - 2023-03-31 Initial version by Sebastian Apel (https://github.com/SebastianApel)
+
+*/
+
+#include <locale.h>
+#include "ggml.h"
+#include <assert.h>
+#include <math.h>
+#include <cstring>
+#include <cstdio>
+#include <cinttypes>
+#include <unordered_map>
+#include <queue>
+#include <string.h>
+#include <cassert>
+#include <fstream>
+#include <string>
+#include <iterator>
+#include <algorithm>
+
+float tensor_sum_elements(struct ggml_tensor * tensor) {
+    float sum = 0;
+    if (tensor->type==6) {
+        for (int j = 0; j < tensor->ne[1]; j++) {
+            for (int k = 0; k < tensor->ne[0]; k++) {
+                sum +=  ((float *) tensor->data)[j*tensor->ne[0]+k];
+            }
+        }
+    }
+    return sum;
+}
+
+
+/*
+    These are mapping to unknown
+    GGML_TYPE_I8,
+    GGML_TYPE_I16,
+    GGML_TYPE_I32,
+    GGML_TYPE_COUNT,
+*/
+
+#define TENSOR_TYPE_AS_STR(TYPE) TYPE == GGML_TYPE_F32 ? "FP32" : TYPE == GGML_TYPE_F16 ? "FP16" : TYPE == GGML_TYPE_Q4_0 ? "Q4_0" : TYPE == GGML_TYPE_Q4_1 ? "Q4_1" : "UNKNOWN"
+
+#define TENSOR_DUMP(TENSOR) printf("%15s: type = %i (%5s) ne = %5d x %5d x %5d, nb = (%5li, %5li, %5li) - ", #TENSOR, \
+        TENSOR->type,TENSOR_TYPE_AS_STR(TENSOR->type),\
+        TENSOR->ne[0], TENSOR->ne[1], TENSOR->ne[2], TENSOR->nb[0], TENSOR->nb[1], TENSOR->nb[2]); \
+    { float sum = tensor_sum_elements(TENSOR); printf("Sum of tensor %s is %6.2f\n",#TENSOR, sum); }
+
+struct benchmark_params_struct {
+    int32_t n_threads     = 1;
+    int32_t n_iterations  = 10;
+};
+
+void print_usage(int /*argc*/, char ** argv, struct benchmark_params_struct params) {
+    fprintf(stderr, "usage: %s [options]\n", argv[0]);
+    fprintf(stderr, "\n");
+    fprintf(stderr, "options:\n");
+    fprintf(stderr, "  -h, --help            show this help message and exit\n");
+    fprintf(stderr, "  -t N, --threads N     number of threads to use during computation (default: %d)\n", params.n_threads);
+    fprintf(stderr, "  -i N, --iter N     number of iterations to use during computation (default: %d)\n", params.n_iterations);
+    fprintf(stderr, "\n");
+}
+
+int main(int argc, char ** argv)  {
+
+
+    struct benchmark_params_struct benchmark_params;
+
+    bool invalid_param = false;
+    std::string arg;
+    for (int i = 1; i < argc; i++) {
+        arg = argv[i];
+
+        if (arg == "-t" || arg == "--threads") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            benchmark_params.n_threads = std::stoi(argv[i]);
+        } else if (arg == "-i" || arg == "--iter") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            benchmark_params.n_iterations = std::stoi(argv[i]);
+        }  else if (arg == "-h" || arg == "--help") {
+            print_usage(argc, argv, benchmark_params);
+            exit(0);
+        }
+        if (invalid_param) {
+            fprintf(stderr, "error: invalid parameter for argument: %s\n", arg.c_str());
+            print_usage(argc, argv, benchmark_params);
+            exit(1);
+        }
+    }
+
+
+    // create the ggml context
+    printf("Starting Test\n");
+
+
+
+    struct ggml_context * ctx;
+    //const int sizex = 4096;
+    //const int sizey = 11008;
+
+#undef VERBOSE_DEBUGGING
+#ifndef VERBOSE_DEBUGGING
+    const int sizey = 4096;
+    const int sizex = 11008;
+    const int sizez = 128;
+#else
+    /* Working - let's increase size */
+    const int sizey = 1;
+    const int sizex = (8*32);
+    const int sizez = 1;
+
+    /*const int sizey = 1;
+    const int sizex = 3*(8*32);
+    const int sizez = 1;*/
+#endif
+
+    //printf("Memsize required = %i\n", sizex*sizex);
+    ggml_type wtype = GGML_TYPE_F32;
+
+    size_t ctx_size = 0;
+    ctx_size += sizex*sizey*ggml_type_sizef(wtype);
+    ctx_size += sizex*sizey*ggml_type_sizef(wtype);
+    ctx_size += sizex*sizey*ggml_type_sizef(GGML_TYPE_F32);
+    ctx_size += sizex*sizeof(float);
+    ctx_size += 1024*1024*100;
+
+    printf("Allocating Memory of size %li byes, %li MB\n",ctx_size, (ctx_size/1024/1024));
+
+    struct ggml_init_params params = {
+        /*.mem_size   =*/ ctx_size,
+        /*.mem_buffer =*/ NULL,
+        /* no_alloc   =*/ 0
+    };
+
+    ctx = ggml_init(params);
+    if (!ctx) {
+        fprintf(stderr, "%s: ggml_init() failed\n", __func__);
+        return false;
+    }
+
+
+    printf("Creating new tensors\n");
+    // printf("Creating new tensor m1\n");
+    struct ggml_tensor * m11 = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, sizex, sizey);
+    ggml_set_f32(m11, 1.0f);
+
+    // printf("Creating new tensor m1\n");
+    struct ggml_tensor * m12 = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, sizex, sizey);
+    ggml_set_f32(m12, 1.5f);
+
+    // printf("Creating new tensor m2\n");
+    struct ggml_tensor * m2 = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, sizex, sizez);
+    ggml_set_f32(m2, 2.0f);
+
+    printf("\n------ Test 1 - Matrix Mult via F32 code ------------------------------------------------------------------------------\n");
+    // printf("Creating new tensor m11xm2\n");
+    struct ggml_tensor * m11xm2 = ggml_mul_mat(ctx, m11, m2);
+
+    // printf("Creating compute graph\n");
+    struct ggml_cgraph gf = ggml_build_forward(m11xm2);
+
+    gf.n_threads=benchmark_params.n_threads;
+    printf("cgraph->n_threads=%i\n",gf.n_threads);
+
+    TENSOR_DUMP(m11);
+    TENSOR_DUMP(m2);
+
+    ggml_graph_compute(ctx, &gf);
+
+    TENSOR_DUMP(gf.nodes[0]);
+
+    printf("\n------ Test 2 - Matrix Mult via Q4_0 code ------------------------------------------------------------------------------\n");
+
+    int32_t nelements = sizex*sizey;
+    int32_t ne[2] = { sizex, sizey };
+
+    std::vector<int64_t> hist_cur(1 << 4, 0);
+
+    // Set up a the benchmark matrices
+    // printf("Creating new tensor q11 & Running quantize\n");
+    struct ggml_tensor * q11 = ggml_new_tensor_2d(ctx, GGML_TYPE_Q4_0, sizex, sizey);
+    ggml_quantize_q4_0((const float *) m11->data, q11->data, nelements, ne[0], hist_cur.data());
+
+    // Set up a the compute graph
+    // printf("Creating new tensor q31\n");
+    struct ggml_tensor * q31 = ggml_mul_mat(ctx, q11, m2);
+
+    // printf("Creating compute graph\n");
+    struct ggml_cgraph gf31 = ggml_build_forward(q31);
+    gf31.n_threads=benchmark_params.n_threads;
+
+    // Set up a second graph computation to make sure we override the CPU cache lines
+    // printf("Creating new tensor q12 & Running quantize\n");
+    struct ggml_tensor * q12 = ggml_new_tensor_2d(ctx, GGML_TYPE_Q4_0, sizex, sizey);
+    ggml_quantize_q4_0((const float *) m12->data, q12->data, nelements, ne[0], hist_cur.data());
+
+    // printf("Creating new tensor q32\n");
+    struct ggml_tensor * q32 = ggml_mul_mat(ctx, q12, m2);
+
+    //printf("Creating compute graph\n");
+    struct ggml_cgraph gf32 = ggml_build_forward(q32);
+    gf32.n_threads=benchmark_params.n_threads;
+    printf("cgraph->n_threads=%i\n",gf31.n_threads);
+
+    const int dimx = sizex;
+    const int dimy = sizey;
+    const int dimz = sizez;
+    long long int flops_per_dot_product = dimy + dimy;
+    long long int flops_per_matrix = flops_per_dot_product * dimx * dimz; ;
+    printf("Matrix Multiplication of (%i,%i,%i) x (%i,%i,%i) - aboout %6.2f gFLOPS\n\n", sizex, sizey, 1, sizex, sizez, 1, 1.0f*flops_per_matrix / 1000 / 1000 / 1000);
+
+
+    // Let's use the F32 result from above as a reference for the q4_0 multiplication
+    float sum_of_F32_reference = tensor_sum_elements(gf.nodes[0]);
+
+
+    printf("Iteration;NThreads; SizeX; SizeY; SizeZ; Required_FLOPS; Elapsed_u_Seconds; FLOPS_per_u_Second\n");
+    printf("==============================================================================================\n");
+
+    for (int i=0;i<benchmark_params.n_iterations ;i++) {
+
+        long long int start = ggml_time_us();
+        //printf("Running ggml_graph_compute\n");
+        ggml_graph_compute(ctx, &gf31);
+        long long int stop = ggml_time_us();
+        long long int usec = stop-start;
+        float sec = usec/1000000;
+        float flops_per_usec = (1.0f*flops_per_matrix)/usec;
+        printf("%9i;%8i;%6i;%6i;%6i;%15lli;%18lli;%19.2f\n",
+            i,
+            gf31.n_threads,
+            sizex, sizey, sizez, flops_per_matrix,
+            usec,flops_per_usec);
+
+#ifdef VERBOSE_DEBUGGING
+        TENSOR_DUMP("res",gf31.nodes[0])
+#endif
+
+        // Check that the matrix multiplication result is in the right ballpark
+        // We cannot use the exact value from the F32 multiplication because the quantizuation will be slightly different
+        float sum_of_Q4_result = tensor_sum_elements(gf31.nodes[0]);
+        float delta = abs(sum_of_Q4_result - sum_of_F32_reference);
+        float allowed_delta = (sum_of_F32_reference) / 1000 / 1000; //  Let's accept an epsilon of 10^-6
+
+        if (delta > allowed_delta)  {
+            printf("\nABORT - ERROR in Matrix Multiplication result - expected %6.2f, got %6.2f (delta %6.2f > allowed_delta %6.2f)\n",
+                sum_of_F32_reference,
+                sum_of_Q4_result,
+                delta,
+                allowed_delta
+            );
+            exit(0);
+        }
+
+        // Running a different graph computation to make sure we override the CPU cache lines
+        ggml_graph_compute(ctx, &gf32);
+
+    }
+
+}

examples/common.cpp

@@ -1,7 +1,5 @@
 #include "common.h"
 
-#include "ggml.h"
-
 #include <cassert>
 #include <cstring>
 #include <fstream>
@@ -16,12 +14,19 @@
 #endif
 
 #if defined (_WIN32)
+#include <fcntl.h>
+#include <io.h>
 #pragma comment(lib,"kernel32.lib")
 extern "C" __declspec(dllimport) void* __stdcall GetStdHandle(unsigned long nStdHandle);
 extern "C" __declspec(dllimport) int __stdcall GetConsoleMode(void* hConsoleHandle, unsigned long* lpMode);
 extern "C" __declspec(dllimport) int __stdcall SetConsoleMode(void* hConsoleHandle, unsigned long dwMode);
 extern "C" __declspec(dllimport) int __stdcall SetConsoleCP(unsigned int wCodePageID);
 extern "C" __declspec(dllimport) int __stdcall SetConsoleOutputCP(unsigned int wCodePageID);
+extern "C" __declspec(dllimport) int __stdcall WideCharToMultiByte(unsigned int CodePage, unsigned long dwFlags,
+                                                                   const wchar_t * lpWideCharStr, int cchWideChar,
+                                                                   char * lpMultiByteStr, int cbMultiByte,
+                                                                   const char * lpDefaultChar, bool * lpUsedDefaultChar);
+#define CP_UTF8 65001
 #endif
 
 bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
@@ -154,6 +159,8 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
             params.use_color = true;
         } else if (arg == "--mlock") {
             params.use_mlock = true;
+        } else if (arg == "--no-mmap") {
+            params.use_mmap = false;
         } else if (arg == "--mtest") {
             params.mem_test = true;
         } else if (arg == "--verbose-prompt") {
@@ -233,9 +240,12 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     fprintf(stderr, "  -b N, --batch_size N  batch size for prompt processing (default: %d)\n", params.n_batch);
     fprintf(stderr, "  --perplexity          compute perplexity over the prompt\n");
     fprintf(stderr, "  --keep                number of tokens to keep from the initial prompt (default: %d, -1 = all)\n", params.n_keep);
-    if (ggml_mlock_supported()) {
+    if (llama_mlock_supported()) {
         fprintf(stderr, "  --mlock               force system to keep model in RAM rather than swapping or compressing\n");
     }
+    if (llama_mmap_supported()) {
+        fprintf(stderr, "  --no-mmap             do not memory-map model (slower load but may reduce pageouts if not using mlock)\n");
+    }
     fprintf(stderr, "  --mtest               compute maximum memory usage\n");
     fprintf(stderr, "  --verbose-prompt      print prompt before generation\n");
     fprintf(stderr, "  -m FNAME, --model FNAME\n");
@@ -307,12 +317,20 @@ void win32_console_init(bool enable_color) {
             SetConsoleMode(hConOut, dwMode | 0x4); // ENABLE_VIRTUAL_TERMINAL_PROCESSING (0x4)
         }
         // Set console output codepage to UTF8
-        SetConsoleOutputCP(65001); // CP_UTF8
+        SetConsoleOutputCP(CP_UTF8);
     }
     void* hConIn = GetStdHandle((unsigned long)-10); // STD_INPUT_HANDLE (-10)
     if (hConIn && hConIn != (void*)-1 && GetConsoleMode(hConIn, &dwMode)) {
-        // Set console input codepage to UTF8
-        SetConsoleCP(65001); // CP_UTF8
+        // Set console input codepage to UTF16
+        _setmode(_fileno(stdin), _O_WTEXT);
     }
 }
+
+// Convert a wide Unicode string to an UTF8 string
+void win32_utf8_encode(const std::wstring & wstr, std::string & str) {
+    int size_needed = WideCharToMultiByte(CP_UTF8, 0, &wstr[0], (int)wstr.size(), NULL, 0, NULL, NULL);
+    std::string strTo(size_needed, 0);
+    WideCharToMultiByte(CP_UTF8, 0, &wstr[0], (int)wstr.size(), &strTo[0], size_needed, NULL, NULL);
+    str = strTo;
+}
 #endif

examples/common.h

@@ -47,6 +47,7 @@ struct gpt_params {
     bool instruct          = false; // instruction mode (used for Alpaca models)
     bool ignore_eos        = false; // do not stop generating after eos
     bool perplexity        = false; // compute perplexity over the prompt
+    bool use_mmap          = true;  // use mmap for faster loads
     bool use_mlock         = false; // use mlock to keep model in memory
     bool mem_test          = false; // compute maximum memory usage
     bool verbose_prompt    = false; // print prompt tokens before generation
@@ -92,4 +93,5 @@ void set_console_color(console_state & con_st, console_color_t color);
 
 #if defined (_WIN32)
 void win32_console_init(bool enable_color);
+void win32_utf8_encode(const std::wstring & wstr, std::string & str);
 #endif

examples/embedding/README.md

@@ -1,3 +1,3 @@
-# embedding
-
-TODO
+# embedding
+
+TODO

examples/embedding/embedding.cpp

@@ -38,6 +38,7 @@ int main(int argc, char ** argv) {
         lparams.seed       = params.seed;
         lparams.f16_kv     = params.memory_f16;
         lparams.logits_all = params.perplexity;
+        lparams.use_mmap   = params.use_mmap;
         lparams.use_mlock  = params.use_mlock;
         lparams.embedding  = params.embedding;
 

examples/gpt4all.sh

@@ -10,6 +10,6 @@ cd ..
 ./main --color --instruct --threads 4 \
        --model ./models/gpt4all-7B/gpt4all-lora-quantized.bin \
        --file ./prompts/alpaca.txt \
-       --batch_size 8 --ctx_size 2048 \
+       --batch_size 8 --ctx_size 2048 -n -1 \
        --repeat_last_n 64 --repeat_penalty 1.3 \
        --n_predict 128 --temp 0.1 --top_k 40 --top_p 0.95

examples/main/README.md

@@ -1,3 +1,3 @@
-# main
-
-TODO
+# main
+
+TODO

examples/main/main.cpp

@@ -1,3 +1,8 @@
+// Defines sigaction on msys:
+#ifndef _GNU_SOURCE
+#define _GNU_SOURCE
+#endif
+
 #include "common.h"
 #include "llama.h"
 
@@ -97,6 +102,7 @@ int main(int argc, char ** argv) {
         lparams.n_parts    = params.n_parts;
         lparams.seed       = params.seed;
         lparams.f16_kv     = params.memory_f16;
+        lparams.use_mmap   = params.use_mmap;
         lparams.use_mlock  = params.use_mlock;
 
         ctx = llama_init_from_file(params.model.c_str(), lparams);
@@ -162,7 +168,7 @@ int main(int argc, char ** argv) {
     }
 
     // enable interactive mode if reverse prompt or interactive start is specified
-    if (params.antiprompt.size() != 0 || params.interactive_start) { 
+    if (params.antiprompt.size() != 0 || params.interactive_start) {
         params.interactive = true;
     }
 
@@ -386,10 +392,19 @@ int main(int argc, char ** argv) {
                 std::string line;
                 bool another_line = true;
                 do {
+#if defined(_WIN32)
+                    std::wstring wline;
+                    if (!std::getline(std::wcin, wline)) {
+                        // input stream is bad or EOF received
+                        return 0;
+                    }
+                    win32_utf8_encode(wline, line);
+#else
                     if (!std::getline(std::cin, line)) {
                         // input stream is bad or EOF received
                         return 0;
                     }
+#endif
                     if (line.empty() || line.back() != '\\') {
                         another_line = false;
                     } else {

examples/perplexity/README.md

@@ -1,3 +1,3 @@
-# perplexity
-
-TODO
+# perplexity
+
+TODO

examples/perplexity/perplexity.cpp

@@ -115,6 +115,7 @@ int main(int argc, char ** argv) {
         lparams.seed       = params.seed;
         lparams.f16_kv     = params.memory_f16;
         lparams.logits_all = params.perplexity;
+        lparams.use_mmap   = params.use_mmap;
         lparams.use_mlock  = params.use_mlock;
         lparams.embedding  = params.embedding;
 

examples/quantize-stats/CMakeLists.txt

@@ -0,0 +1,4 @@
+set(TARGET quantize-stats)
+add_executable(${TARGET} quantize-stats.cpp)
+target_link_libraries(${TARGET} PRIVATE llama ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_11)

examples/quantize-stats/quantize-stats.cpp

@@ -0,0 +1,354 @@
+#include "ggml.h"
+#include "llama.h"
+#include "llama_internal.h"
+
+#include <algorithm>
+#include <cassert>
+#include <cinttypes>
+#include <cmath>
+#include <cstdio>
+#include <cstring>
+#include <map>
+#include <numeric>
+#include <regex>
+#include <string>
+#include <unordered_map>
+#include <vector>
+
+static const char * type_strs[] = { "q4_0", "q4_1", "i8", "i16", "i32", "f16", "f32"  };
+static_assert(sizeof(type_strs) == GGML_TYPE_COUNT * sizeof(char *), "Incomplete type list");
+
+struct quantize_stats_params {
+    std::string model = "models/7B/ggml-model-f16.bin";
+    bool verbose = false;
+    bool per_layer_stats = false;
+    bool print_histogram = false;
+    bool reference = false;
+    std::vector<std::string> include_layers;
+    std::vector<std::string> exclude_layers;
+    std::vector<enum ggml_type> include_types;
+};
+
+const int64_t SCRATCH_ELEMENTS = 32*32;
+const size_t HISTOGRAM_BUCKETS = 150;
+const double HISTOGRAM_RANGE = 0.03;
+
+struct error_stats {
+    size_t num_samples;
+    double total_error;
+    double max_error;
+    uint64_t error_histogram[HISTOGRAM_BUCKETS];
+};
+
+
+void quantize_stats_print_usage(int /*argc*/, char ** argv) {
+    quantize_stats_params params;
+    fprintf(stderr, "usage: %s [options]\n", argv[0]);
+    fprintf(stderr, "\n");
+    fprintf(stderr, "options:\n");
+    fprintf(stderr, "  -h, --help            show this help message and exit\n");
+    fprintf(stderr, "  -m FNAME, --model FNAME\n");
+    fprintf(stderr, "                        model path (default: %s)\n", params.model.c_str());
+    fprintf(stderr, "  -r, --reference\n");
+    fprintf(stderr, "                        use reference implementation (default: false)\n");
+    fprintf(stderr, "  -v, --verbose\n");
+    fprintf(stderr, "                        verbose output (default: false)\n");
+    fprintf(stderr, "  -p, --per-layer-stats\n");
+    fprintf(stderr, "                        print stats per layer (default: false)\n");
+    fprintf(stderr, "  --histogram\n");
+    fprintf(stderr, "                        print error histogram (default: false)\n");
+    fprintf(stderr, "  -l LAYER, --include-layer LAYER\n");
+    fprintf(stderr, "                        only test layers matching pattern\n");
+    fprintf(stderr, "  -L LAYER, --exclude-layer LAYER\n");
+    fprintf(stderr, "                        exclude layers matching pattern\n");
+    fprintf(stderr, "  -t TYPE, --type TYPE\n");
+    fprintf(stderr, "                        only test given type (q4_0, q4_1)\n");
+    fprintf(stderr, "\n");
+}
+
+// Check if a layer is included/excluded by command line
+bool layer_included(const quantize_stats_params params, const std::string & layer) {
+    for (const auto& excluded : params.exclude_layers) {
+        if (std::regex_search(layer, std::regex(excluded))) {
+            return false;
+        }
+    }
+    for (const auto& included : params.include_layers) {
+        if (std::regex_search(layer, std::regex(included))) {
+            return true;
+        }
+    }
+    return params.include_layers.empty();
+}
+
+// Update error statistics given vectors with the before/after result of quantization
+void update_error_stats(int64_t nelements, const float * input, const float * output, error_stats & stats) {
+    for (int64_t i = 0; i < nelements; i++) {
+        double diff = input[i] - output[i];
+        stats.total_error += diff * diff;
+        stats.max_error = fmax(fabs(diff), stats.max_error);
+        stats.error_histogram[std::max(std::min((size_t) floor(fabs(diff) / HISTOGRAM_RANGE * HISTOGRAM_BUCKETS), HISTOGRAM_BUCKETS-1), (size_t) 0)]++;
+    }
+    stats.num_samples += nelements;
+}
+
+double find_quantile(const error_stats & stats, double quantile) {
+    double sum = std::accumulate(std::begin(stats.error_histogram), std::end(stats.error_histogram), 0.0);
+
+    double accum = 0;
+    for (size_t i = 0; i < HISTOGRAM_BUCKETS; i++) {
+        accum += stats.error_histogram[i];
+        if (accum >= sum*quantile) {
+            return (i+1) * HISTOGRAM_RANGE / HISTOGRAM_BUCKETS;
+        }
+    }
+    return INFINITY;
+}
+
+void print_error_stats(const std::string & name, const error_stats & stats, bool print_histogram) {
+    double rmse = sqrt(stats.total_error / (double) stats.num_samples);
+    double median = find_quantile(stats, .5);
+    double pct95 = find_quantile(stats, .95);
+    printf("%-50s: rmse %.8f, maxerr %.8f, 95pct<%.4f, median<%.4f\n", name.c_str(), rmse, stats.max_error, pct95, median);
+    if (print_histogram) {
+        printf("Error distribution:\n");
+        for (size_t i = 0; i < HISTOGRAM_BUCKETS; i++) {
+            double lower = i * HISTOGRAM_RANGE / HISTOGRAM_BUCKETS;
+            double upper = (i+1) * HISTOGRAM_RANGE / HISTOGRAM_BUCKETS;
+            if (i == HISTOGRAM_BUCKETS -1) upper = INFINITY;
+            printf("[%3.4f, %3.4f): %11" PRIu64 "\n", lower, upper, stats.error_histogram[i]);
+        }
+    }
+}
+
+// copied from ggml.h - verify that we can access this as a flat array
+static bool tensor_is_contiguous(const struct ggml_tensor * tensor) {
+    static_assert(GGML_MAX_DIMS == 4, "GGML_MAX_DIMS is not 4 - update this function");
+
+    return
+        tensor->nb[0] == ggml_type_size(tensor->type) &&
+        tensor->nb[1] == (tensor->nb[0]*tensor->ne[0])/ggml_blck_size(tensor->type) &&
+        tensor->nb[2] == tensor->nb[1]*tensor->ne[1] &&
+        tensor->nb[3] == tensor->nb[2]*tensor->ne[2];
+}
+
+// Run quantization function for a single layer and update error stats
+void test_roundtrip_on_layer(
+        std::string & name,
+        bool print_layer_stats,
+        const quantize_fns_t & qfns,
+        bool use_reference,
+        const ggml_tensor * layer,
+        float * input_scratch,
+        char *quantized_scratch,
+        float * output_scratch,
+        error_stats & total_error) {
+
+    assert(tensor_is_contiguous(layer));
+    error_stats layer_error {};
+    int64_t nelements = ggml_nelements(layer);
+
+    for (int64_t offset = 0; offset < nelements; offset += SCRATCH_ELEMENTS) {
+        int64_t chunk_size = std::min(SCRATCH_ELEMENTS, nelements - offset);
+
+        if (layer->type == GGML_TYPE_F16) {
+            for (int i = 0; i < chunk_size; i++) {
+                input_scratch[i] = ggml_get_f32_1d(layer, i + offset);
+            }
+        } else {
+            input_scratch = ggml_get_data_f32(layer) + offset;
+        }
+
+        if (use_reference) {
+            qfns.quantize_row_q_reference(input_scratch, quantized_scratch, chunk_size);
+        } else {
+            qfns.quantize_row_q(input_scratch, quantized_scratch, chunk_size);
+        }
+        qfns.dequantize_row_q(quantized_scratch, output_scratch, chunk_size);
+
+        update_error_stats(chunk_size, input_scratch, output_scratch, total_error);
+        if (print_layer_stats) {
+            update_error_stats(chunk_size, input_scratch, output_scratch, layer_error);
+        }
+    }
+    if (print_layer_stats) {
+        print_error_stats(name, layer_error, false);
+    }
+}
+
+int main(int argc, char ** argv) {
+    ggml_time_init();
+
+    quantize_stats_params params;
+
+    // read command line
+
+    bool invalid_param = false;
+    std::string arg;
+    for (int i = 1; i < argc; i++) {
+        arg = argv[i];
+
+        if (arg == "-h" || arg == "--help") {
+            quantize_stats_print_usage(argc, argv);
+            exit(0);
+        } else if (arg == "-r" || arg == "--reference") {
+            params.reference = true;
+        } else if (arg == "-v") {
+            params.verbose = true;
+        } else if (arg == "-p" || arg == "--per-layer-stats") {
+            params.per_layer_stats = true;
+        } else if (arg == "--histogram") {
+            params.print_histogram = true;
+        } else if (arg == "-m" || arg == "--model") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.model = argv[i];
+        } else if (arg == "-l" || arg == "--include-layer") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.include_layers.push_back(argv[i]);
+        } else if (arg == "-L" || arg == "--exclude-layer") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.exclude_layers.push_back(argv[i]);
+        } else if (arg == "-t" || arg == "--type") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            int j;
+            for (j = 0; j < GGML_TYPE_COUNT && strcmp(argv[i], type_strs[j]) != 0; j++) {
+                // find match
+            }
+            if (j < GGML_TYPE_COUNT) {
+                params.include_types.push_back((ggml_type) j);
+            } else {
+                fprintf(stderr, "error: %s not in list of types\n", argv[i]);
+                invalid_param = true;
+            }
+        } else {
+            fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
+            quantize_stats_print_usage(argc, argv);
+            return 1;
+        }
+    }
+    if (invalid_param) {
+        fprintf(stderr, "error: invalid parameter for argument: %s\n", arg.c_str());
+        quantize_stats_print_usage(argc, argv);
+        return 1;
+    }
+
+    // load the model
+    fprintf(stderr, "Loading model\n");
+
+    const int64_t t_main_start_us = ggml_time_us();
+    llama_context * ctx;
+
+    {
+        auto lparams = llama_context_default_params();
+
+        lparams.n_ctx      = 256;
+        lparams.n_parts    = 1;
+        lparams.seed       = 1;
+        lparams.f16_kv     = false;
+        lparams.use_mlock  = false;
+
+        ctx = llama_init_from_file(params.model.c_str(), lparams);
+
+        if (ctx == NULL) {
+            fprintf(stderr, "%s: error: failed to load model '%s'\n", __func__, params.model.c_str());
+            return 1;
+        }
+    }
+
+    const auto &tensors = llama_internal_get_tensor_map(ctx);
+
+    // check layer tensors
+    int included_layers = 0;
+    int64_t max_nelements = 0;
+    bool is_f16 = false;
+    for (const auto& kv_tensor : tensors) {
+        if (!layer_included(params, kv_tensor.first)) {
+            continue;
+        }
+        if (params.verbose) {
+            printf("%s: type %s, size %" PRId64 "\n", kv_tensor.first.c_str(), type_strs[kv_tensor.second->type], ggml_nelements(kv_tensor.second));
+        }
+        if (kv_tensor.second->type == GGML_TYPE_F16) {
+            is_f16 = true;
+        } else if (kv_tensor.second->type != GGML_TYPE_F32) {
+            fprintf(stderr, "%s: error: Quantization should be tested with a float model, "
+                "this model contains already quantized layers (%s is type %d)\n", __func__, kv_tensor.first.c_str(), kv_tensor.second->type);
+            llama_free(ctx);
+            return 1;
+        }
+        included_layers++;
+        max_nelements = std::max(max_nelements, ggml_nelements(kv_tensor.second));
+    }
+
+    if (is_f16) {
+        printf("note: source model is f16\n");
+    }
+    printf("testing %d layers with max size %" PRId64 "\n", included_layers, max_nelements);
+    // allocate scratch space
+    std::vector<float> input_scratch(SCRATCH_ELEMENTS);
+    std::vector<char> quantized_scratch(SCRATCH_ELEMENTS*4);
+    std::vector<float> output_scratch(SCRATCH_ELEMENTS);
+
+    // loop throught quantization types
+    for (int i = 0; i < GGML_TYPE_COUNT; i++) {
+        if (!params.include_types.empty() && std::find(params.include_types.begin(), params.include_types.end(), i) == params.include_types.end()) {
+            continue;
+        }
+        quantize_fns_t qfns = ggml_internal_get_quantize_fn(i);
+        if (qfns.quantize_row_q && qfns.dequantize_row_q) {
+            if (params.verbose) {
+                printf("testing %s ...\n",  type_strs[i]);
+            }
+
+            error_stats global_stats {};
+
+            for (const auto& kv_tensor : tensors) {
+                if (!layer_included(params, kv_tensor.first)) {
+                    continue;
+                }
+                if (params.verbose) {
+                    printf("  %s ...\n",  kv_tensor.first.c_str());
+                }
+                std::string layer_name { type_strs[i] };
+                layer_name += "::" + kv_tensor.first;
+                test_roundtrip_on_layer(
+                        layer_name,
+                        params.per_layer_stats,
+                        qfns,
+                        params.reference,
+                        kv_tensor.second,
+                        input_scratch.data(),
+                        quantized_scratch.data(),
+                        output_scratch.data(),
+                        global_stats
+                );
+            }
+
+            print_error_stats(type_strs[i], global_stats, params.print_histogram);
+        }
+    }
+
+
+    llama_free(ctx);
+    // report timing
+    {
+        const int64_t t_main_end_us = ggml_time_us();
+
+        printf("\n");
+        printf("%s:    total time = %8.2f ms\n", __func__, (t_main_end_us - t_main_start_us)/1000.0);
+    }
+
+    return 0;
+}

examples/quantize/quantize.cpp

@@ -5,15 +5,15 @@
 #include <string>
 
 // usage:
-//  ./llama-quantize models/llama/ggml-model.bin models/llama/ggml-model-quant.bin type
+//  ./quantize models/llama/ggml-model.bin models/llama/ggml-model-quant.bin type
 //
 int main(int argc, char ** argv) {
     ggml_time_init();
 
     if (argc != 4) {
         fprintf(stderr, "usage: %s model-f32.bin model-quant.bin type\n", argv[0]);
-        fprintf(stderr, "  type = 2 - q4_0\n");
-        fprintf(stderr, "  type = 3 - q4_1\n");
+        fprintf(stderr, "  type = %d - q4_0\n", LLAMA_FTYPE_MOSTLY_Q4_0);
+        fprintf(stderr, "  type = %d - q4_1\n", LLAMA_FTYPE_MOSTLY_Q4_1);
         return 1;
     }
 
@@ -27,7 +27,7 @@ int main(int argc, char ** argv) {
     const std::string fname_inp = argv[1];
     const std::string fname_out = argv[2];
 
-    const int itype = atoi(argv[3]);
+    const enum llama_ftype ftype = (enum llama_ftype)atoi(argv[3]);
 
     const int64_t t_main_start_us = ggml_time_us();
 
@@ -37,7 +37,7 @@ int main(int argc, char ** argv) {
     {
         const int64_t t_start_us = ggml_time_us();
 
-        if (llama_model_quantize(fname_inp.c_str(), fname_out.c_str(), itype)) {
+        if (llama_model_quantize(fname_inp.c_str(), fname_out.c_str(), ftype)) {
             fprintf(stderr, "%s: failed to quantize model from '%s'\n", __func__, fname_inp.c_str());
             return 1;
         }

flake.nix

@@ -28,8 +28,9 @@
           ];
           installPhase = ''
             mkdir -p $out/bin
-            mv bin/main $out/bin/llama
-            mv bin/quantize $out/bin/quantize
+            mv bin/* $out/bin/
+            mv $out/bin/main $out/bin/llama
+
             echo "#!${llama-python}/bin/python" > $out/bin/convert-pth-to-ggml
             cat ${./convert-pth-to-ggml.py} >> $out/bin/convert-pth-to-ggml
             chmod +x $out/bin/convert-pth-to-ggml

ggml.c

@@ -1,4 +1,4 @@
-// Defines CLOCK_MONOTONIC and asprintf on Linux
+// Defines CLOCK_MONOTONIC on Linux
 #define _GNU_SOURCE
 
 #include "ggml.h"
@@ -26,14 +26,9 @@
 #define static_assert(cond, msg) struct global_scope_noop_trick
 #endif
 
-#if defined _MSC_VER || defined(__MINGW32__)
+#if defined(_WIN32)
 
-#if !defined(__MINGW32__)
-#include <Windows.h>
-#else
-// ref: https://github.com/ggerganov/whisper.cpp/issues/168
 #include <windows.h>
-#endif
 
 typedef volatile LONG atomic_int;
 typedef atomic_int atomic_bool;
@@ -55,6 +50,7 @@ typedef HANDLE pthread_t;
 
 typedef DWORD thread_ret_t;
 static int pthread_create(pthread_t* out, void* unused, thread_ret_t(*func)(void*), void* arg) {
+    (void) unused;
     HANDLE handle = CreateThread(NULL, 0, (LPTHREAD_START_ROUTINE) func, arg, 0, NULL);
     if (handle == NULL)
     {
@@ -66,6 +62,7 @@ static int pthread_create(pthread_t* out, void* unused, thread_ret_t(*func)(void
 }
 
 static int pthread_join(pthread_t thread, void* unused) {
+    (void) unused;
     return (int) WaitForSingleObject(thread, INFINITE);
 }
 
@@ -97,17 +94,6 @@ typedef void* thread_ret_t;
 #define static_assert(cond, msg) _Static_assert(cond, msg)
 #endif
 
-#define GGML_MLOCK_SUPPORT 0
-
-#ifdef __has_include
-    #if __has_include(<sys/mman.h>)
-        #undef GGML_MLOCK_SUPPORT
-        #define GGML_MLOCK_SUPPORT 1
-        #include <sys/mman.h>
-    #endif
-#endif
-
-
 /*#define GGML_PERF*/
 #define GGML_DEBUG 0
 #define GGML_GELU_FP16
@@ -128,6 +114,14 @@ typedef void* thread_ret_t;
     #define GGML_MEM_ALIGN 16
 #endif
 
+#if defined(_MSC_VER) || defined(__MINGW32__)
+#define GGML_ALIGNED_MALLOC(size)  _aligned_malloc(size, GGML_MEM_ALIGN)
+#define GGML_ALIGNED_FREE(ptr)     _aligned_free(ptr)
+#else
+#define GGML_ALIGNED_MALLOC(size)  aligned_alloc(GGML_MEM_ALIGN, size)
+#define GGML_ALIGNED_FREE(ptr)     free(ptr)
+#endif
+
 #define UNUSED(x) (void)(x)
 #define SWAP(x, y, T) do { T SWAP = x; x = y; y = SWAP; } while (0)
 
@@ -242,12 +236,12 @@ static inline float fp32_from_bits(uint32_t w) {
 }
 
 static inline uint32_t fp32_to_bits(float f) {
-	union {
-		float as_value;
-		uint32_t as_bits;
-	} fp32;
-	fp32.as_value = f;
-	return fp32.as_bits;
+    union {
+        float as_value;
+        uint32_t as_bits;
+    } fp32;
+    fp32.as_value = f;
+    return fp32.as_bits;
 }
 
 static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
@@ -610,10 +604,7 @@ static void quantize_row_q4_0(const float * restrict x, void * restrict vy, int
         for (int l = 0; l < 2; l++) amaxv[4*l] = vmaxq_f32(amaxv[4*l], amaxv[4*l+2]);
         for (int l = 0; l < 1; l++) amaxv[8*l] = vmaxq_f32(amaxv[8*l], amaxv[8*l+4]);
 
-        // absolute max
-        const float amax = MAX(
-                MAX(vgetq_lane_f32(amaxv[0], 0), vgetq_lane_f32(amaxv[0], 1)),
-                MAX(vgetq_lane_f32(amaxv[0], 2), vgetq_lane_f32(amaxv[0], 3)));
+        const float amax = vmaxvq_f32(amaxv[0]);
 
         const float d = amax / ((1 << 3) - 1);
         const float id = d ? 1.0f/d : 0.0f;
@@ -935,7 +926,7 @@ static void quantize_row_q4_1(const float * restrict x, void * restrict vy, int
         float32x4_t minv[8];
         float32x4_t maxv[8];
 
-        for (int l = 0; l < 8; l++) srcv[l] = vld1q_f32(x + i*32 + 4*l);
+        for (int l = 0; l < 8; l++) srcv[l] = vld1q_f32(x + i*QK + 4*l);
 
         for (int l = 0; l < 4; l++) minv[2*l] = vminq_f32(srcv[2*l], srcv[2*l + 1]);
         for (int l = 0; l < 2; l++) minv[4*l] = vminq_f32(minv[4*l], minv[4*l + 2]);
@@ -958,7 +949,8 @@ static void quantize_row_q4_1(const float * restrict x, void * restrict vy, int
 
         for (int l = 0; l < 8; l++) {
             const float32x4_t v  = vmulq_n_f32(vsubq_f32(srcv[l], minv0), id);
-            const int32x4_t   vi = vcvtq_s32_f32(v);
+            const float32x4_t vf = vaddq_f32(v, vdupq_n_f32(0.5f)); // needed to round to nearest
+            const int32x4_t   vi = vcvtq_s32_f32(vf);
 
             y[i].qs[2*l + 0] = vgetq_lane_s32(vi, 0) | (vgetq_lane_s32(vi, 1) << 4);
             y[i].qs[2*l + 1] = vgetq_lane_s32(vi, 2) | (vgetq_lane_s32(vi, 3) << 4);
@@ -1897,7 +1889,7 @@ static void ggml_vec_dot_q4_0(const int n, float * restrict s, const void * rest
         sum1 += x1->d * y1->d * (vgetq_lane_s32(p_1, 0) + vgetq_lane_s32(p_1, 1) + vgetq_lane_s32(p_1, 2) + vgetq_lane_s32(p_1, 3));
 #endif
 #else
-	    const int16x8_t pl0l = vmull_s8(vget_low_s8 (v0_0ls), vget_low_s8 (v1_0ls));
+        const int16x8_t pl0l = vmull_s8(vget_low_s8 (v0_0ls), vget_low_s8 (v1_0ls));
         const int16x8_t pl0h = vmull_s8(vget_high_s8(v0_0ls), vget_high_s8(v1_0ls));
 
         const int16x8_t ph0l = vmull_s8(vget_low_s8 (v0_0hs), vget_low_s8 (v1_0hs));
@@ -1962,7 +1954,7 @@ static void ggml_vec_dot_q4_0(const int n, float * restrict s, const void * rest
     // Initialize accumulator with zeros
     __m256 acc = _mm256_setzero_ps();
 
-    /* Prepare the constants we will need during execution */        
+    /* Prepare the constants we will need during execution */
     const __m256i lowMask = _mm256_set1_epi8( 0xF );
     const __m256i offset_8 = _mm256_set1_epi16( 8 );
 
@@ -1972,61 +1964,59 @@ static void ggml_vec_dot_q4_0(const int n, float * restrict s, const void * rest
 
     // Main loop
     for (int i = 0; i < nb; i+=UNROLL_COUNT) {
-
-        // This loop will be unrolled by the compiler    
+        // This loop will be unrolled by the compiler
         for (int u=0;u<UNROLL_COUNT;u++)  {
-            /* Compute combined scale for the block */ 
-            const __m256 scale = _mm256_mul_ps( 
-                    _mm256_broadcast_ss( &x[i+u].d ), 
-                    _mm256_broadcast_ss( &y[i+u].d ) ); 
+            /* Compute combined scale for the block */
+            const __m256 scale = _mm256_mul_ps(
+                    _mm256_broadcast_ss( &x[i+u].d ),
+                    _mm256_broadcast_ss( &y[i+u].d ) );
 
-            /* get input from x 
-               Input: 32 Nibbles (16 bytes) at *x[i+u] 
-               Output: 2 vectors with 16 values of type int16_t (x_high_q, x_low_q) */             
-                                      
-            /* Load 16 bytes from memory */  
-            const __m128i tmp_x = _mm_loadu_si128( ( const __m128i* ) x[i+u].qs); 
-            /* Expand bytes into uint16_t values */                                 
-            const __m256i bytes_x = _mm256_cvtepu8_epi16(tmp_x); 
+            /* get input from x
+               Input: 32 Nibbles (16 bytes) at *x[i+u]
+               Output: 2 vectors with 16 values of type int16_t (x_high_q, x_low_q) */
+
+            /* Load 16 bytes from memory */
+            const __m128i tmp_x = _mm_loadu_si128( ( const __m128i* ) x[i+u].qs);
+            /* Expand bytes into uint16_t values */
+            const __m256i bytes_x = _mm256_cvtepu8_epi16(tmp_x);
             /* Unpack values into individual bytes */
             __m256i x_low_q = _mm256_and_si256( lowMask, bytes_x );
             const __m256i pre_shift_x_high_q = _mm256_andnot_si256( lowMask, bytes_x );
-            __m256i x_high_q = _mm256_srli_epi16( pre_shift_x_high_q, 4 );            
+            __m256i x_high_q = _mm256_srli_epi16( pre_shift_x_high_q, 4 );
             /* Now we have two vectors with bytes in [ 0 .. 15 ] interval.  Offset them into [ -8 .. +7 ] interval.  */
-            x_high_q = _mm256_sub_epi16( x_high_q, offset_8 ); 
-            x_low_q = _mm256_sub_epi16( x_low_q, offset_8 ); 
+            x_high_q = _mm256_sub_epi16( x_high_q, offset_8 );
+            x_low_q = _mm256_sub_epi16( x_low_q, offset_8 );
 
-            /* get input from y 
-               Input: 32 Nibbles (16 bytes) at *y[i+u] 
-               Output: 2 vectors with 16 values of type int16_t (y_high_q, y_low_q) */             
+            /* get input from y
+               Input: 32 Nibbles (16 bytes) at *y[i+u]
+               Output: 2 vectors with 16 values of type int16_t (y_high_q, y_low_q) */
 
-            /* Load 16 bytes from memory */  
-            const __m128i tmp_y = _mm_loadu_si128( (const __m128i* ) y[i+u].qs); 
-            /* Expand bytes into uint16_t values */     
-            const __m256i bytes_y = _mm256_cvtepu8_epi16(tmp_y); 
+            /* Load 16 bytes from memory */
+            const __m128i tmp_y = _mm_loadu_si128( (const __m128i* ) y[i+u].qs);
+            /* Expand bytes into uint16_t values */
+            const __m256i bytes_y = _mm256_cvtepu8_epi16(tmp_y);
             /* Unpack values into individual bytes */
-            const __m256i pre_shift_y_high_q = _mm256_andnot_si256( lowMask, bytes_y ); 
-            __m256i y_high_q = _mm256_srli_epi16( pre_shift_y_high_q, 4 ); 
-            __m256i y_low_q = _mm256_and_si256( lowMask, bytes_y ); 
+            const __m256i pre_shift_y_high_q = _mm256_andnot_si256( lowMask, bytes_y );
+            __m256i y_high_q = _mm256_srli_epi16( pre_shift_y_high_q, 4 );
+            __m256i y_low_q = _mm256_and_si256( lowMask, bytes_y );
             /* Now we have two vectors with bytes in [ 0 .. 15 ] interval.  Offset them into [ -8 .. +7 ] interval.  */
-            y_high_q = _mm256_sub_epi16( y_high_q, offset_8 ); 
-            y_low_q = _mm256_sub_epi16( y_low_q, offset_8 ); 
+            y_high_q = _mm256_sub_epi16( y_high_q, offset_8 );
+            y_low_q = _mm256_sub_epi16( y_low_q, offset_8 );
 
-            /* Compute products of int16_t integers, add pairwise, store as int32_t */     
-            __m256i xy_high_q = _mm256_madd_epi16( x_high_q, y_high_q ); 
-            __m256i xy_low_q = _mm256_madd_epi16( x_low_q, y_low_q ); 
+            /* Compute products of int16_t integers, add pairwise, store as int32_t */
+            __m256i xy_high_q = _mm256_madd_epi16( x_high_q, y_high_q );
+            __m256i xy_low_q = _mm256_madd_epi16( x_low_q, y_low_q );
 
-            /* Accumulate the products of int32_t integers -> we now have a vector of 8 int_32t */ 
-            __m256i xy_q = _mm256_add_epi32( xy_high_q, xy_low_q ); 
+            /* Accumulate the products of int32_t integers -> we now have a vector of 8 int_32t */
+            __m256i xy_q = _mm256_add_epi32( xy_high_q, xy_low_q );
 
-            /* Convert to vectore of 8 int32_t to 8 floats */ 
-            __m256 q = _mm256_cvtepi32_ps( xy_q ); 
+            /* Convert to vectore of 8 int32_t to 8 floats */
+            __m256 q = _mm256_cvtepi32_ps( xy_q );
 
-            /* Multiply q with scale and accumulate */ 
-            acc = _mm256_fmadd_ps( scale, q, acc );    
+            /* Multiply q with scale and accumulate */
+            acc = _mm256_fmadd_ps( scale, q, acc );
         }
-       
-    }   
+    }
 
     // Return horizontal sum of the acc vector
     __m128 res = _mm256_extractf128_ps( acc, 1 );
@@ -2087,18 +2077,18 @@ static void ggml_vec_dot_q4_0(const int n, float * restrict s, const void * rest
     float sum1 = 0.0f;
 
     for (int i = 0; i < nb; i += 2) {
-        const block_q4_0 * restrict x0 = &px[i + 0];
-        const block_q4_0 * restrict y0 = &py[i + 0];
-        const block_q4_0 * restrict x1 = &px[i + 1];
-        const block_q4_0 * restrict y1 = &py[i + 1];
+        const block_q4_0 * restrict x0 = &x[i + 0];
+        const block_q4_0 * restrict y0 = &y[i + 0];
+        const block_q4_0 * restrict x1 = &x[i + 1];
+        const block_q4_0 * restrict y1 = &y[i + 1];
 
         const v128_t m4b = wasm_u8x16_splat(0xf);
         const v128_t s8b = wasm_i8x16_splat(0x8);
 
-        const v128_t v0_0 = wasm_v128_load(x0.qs);
-        const v128_t v0_1 = wasm_v128_load(y0.qs);
-        const v128_t v1_0 = wasm_v128_load(x1.qs);
-        const v128_t v1_1 = wasm_v128_load(y1.qs);
+        const v128_t v0_0 = wasm_v128_load(x0->qs);
+        const v128_t v0_1 = wasm_v128_load(y0->qs);
+        const v128_t v1_0 = wasm_v128_load(x1->qs);
+        const v128_t v1_1 = wasm_v128_load(y1->qs);
 
         // 4-bit -> 8-bit
         const v128_t v0_0l = wasm_v128_and(v0_0, m4b);
@@ -2170,18 +2160,20 @@ static void ggml_vec_dot_q4_0(const int n, float * restrict s, const void * rest
         const uint8_t * restrict p0 = x[i].qs;
         const uint8_t * restrict p1 = y[i].qs;
 
+        int sumi = 0;
         for (int j = 0; j < QK/2; j++) {
             const uint8_t v0 = p0[j];
             const uint8_t v1 = p1[j];
 
-            const float f0 = d0*((int8_t) (v0 & 0xf) - 8);
-            const float f1 = d0*((int8_t) (v0 >> 4)  - 8);
+            const int8_t i0 = (int8_t) (v0 & 0xf) - 8;
+            const int8_t i1 = (int8_t) (v0 >> 4)  - 8;
 
-            const float f2 = d1*((int8_t) (v1 & 0xf) - 8);
-            const float f3 = d1*((int8_t) (v1 >> 4)  - 8);
+            const int8_t i2 = (int8_t) (v1 & 0xf) - 8;
+            const int8_t i3 = (int8_t) (v1 >> 4)  - 8;
 
-            sumf += f0*f2 + f1*f3;
+            sumi += i0*i2 + i1*i3;
         }
+        sumf += d0 * d1 * sumi;
     }
 #endif
 
@@ -2578,29 +2570,26 @@ inline static void ggml_vec_norm_inv_f32(const int n, float * s, const float * x
 //
 
 static const int GGML_BLCK_SIZE[GGML_TYPE_COUNT] = {
-    QK,
-    QK,
-    1,
-    1,
-    1,
-    1,
-    1,
+    [GGML_TYPE_F32]  = 1,
+    [GGML_TYPE_F16]  = 1,
+    [GGML_TYPE_Q4_0] = QK,
+    [GGML_TYPE_Q4_1] = QK,
+    [GGML_TYPE_I8]   = 1,
+    [GGML_TYPE_I16]  = 1,
+    [GGML_TYPE_I32]  = 1,
 };
-
-static_assert(GGML_TYPE_COUNT == 7, "GGML_TYPE_COUNT != 5");
+static_assert(GGML_TYPE_COUNT == 7, "GGML_BLCK_SIZE is outdated");
 
 static const size_t GGML_TYPE_SIZE[GGML_TYPE_COUNT] = {
-    sizeof(block_q4_0),
-    sizeof(block_q4_1),
-    sizeof(int8_t ),
-    sizeof(int16_t),
-    sizeof(int32_t),
-    sizeof(ggml_fp16_t),
-    sizeof(float  ),
+    [GGML_TYPE_F32]  = sizeof(float),
+    [GGML_TYPE_F16]  = sizeof(ggml_fp16_t),
+    [GGML_TYPE_Q4_0] = sizeof(block_q4_0),
+    [GGML_TYPE_Q4_1] = sizeof(block_q4_1),
+    [GGML_TYPE_I8]   = sizeof(int8_t),
+    [GGML_TYPE_I16]  = sizeof(int16_t),
+    [GGML_TYPE_I32]  = sizeof(int32_t),
 };
-
-// don't forget to update the array above when adding new types
-static_assert(GGML_TYPE_COUNT == 7, "GGML_TYPE_COUNT != 5");
+static_assert(GGML_TYPE_COUNT == 7, "GGML_TYPE_SIZE is outdated");
 
 static const char * GGML_OP_LABEL[GGML_OP_COUNT] = {
     "NONE",
@@ -2629,6 +2618,7 @@ static const char * GGML_OP_LABEL[GGML_OP_COUNT] = {
 
     "SCALE",
     "CPY",
+    "CONT",
     "RESHAPE",
     "VIEW",
     "PERMUTE",
@@ -2644,7 +2634,7 @@ static const char * GGML_OP_LABEL[GGML_OP_COUNT] = {
     "FLASH_FF",
 };
 
-static_assert(GGML_OP_COUNT == 35, "GGML_OP_COUNT != 35");
+static_assert(GGML_OP_COUNT == 36, "GGML_OP_COUNT != 36");
 
 static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
     "none",
@@ -2673,6 +2663,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
 
     "x*v",
     "x-\\>y",
+    "cont(x)",
     "reshape(x)",
     "view(x)",
     "permute(x)",
@@ -2688,22 +2679,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
     "flash_ff(x)",
 };
 
-static_assert(GGML_OP_COUNT == 35, "GGML_OP_COUNT != 35");
-
-//
-// ggml object
-//
-
-struct ggml_object {
-    size_t offs;
-    size_t size;
-
-    struct ggml_object * next;
-
-    char padding[8];
-};
-
-static const size_t GGML_OBJECT_SIZE = sizeof(struct ggml_object);
+static_assert(GGML_OP_COUNT == 36, "GGML_OP_COUNT != 36");
 
 static_assert(sizeof(struct ggml_object)%GGML_MEM_ALIGN == 0, "ggml_object size must be a multiple of GGML_MEM_ALIGN");
 static_assert(sizeof(struct ggml_tensor)%GGML_MEM_ALIGN == 0, "ggml_tensor size must be a multiple of GGML_MEM_ALIGN");
@@ -2716,7 +2692,6 @@ struct ggml_context {
     size_t mem_size;
     void * mem_buffer;
     bool   mem_buffer_owned;
-    bool   mem_buffer_mlocked;
     bool   no_alloc;
 
     int    n_objects;
@@ -3001,9 +2976,8 @@ struct ggml_context * ggml_init(struct ggml_init_params params) {
 
     *ctx = (struct ggml_context) {
         /*.mem_size           =*/ params.mem_size,
-        /*.mem_buffer         =*/ params.mem_buffer ? params.mem_buffer : malloc(params.mem_size),
+        /*.mem_buffer         =*/ params.mem_buffer ? params.mem_buffer : GGML_ALIGNED_MALLOC(params.mem_size),
         /*.mem_buffer_owned   =*/ params.mem_buffer ? false : true,
-        /*.mem_buffer_mlocked =*/ false,
         /*.no_alloc           =*/ params.no_alloc,
         /*.n_objects          =*/ 0,
         /*.objects_begin      =*/ NULL,
@@ -3036,16 +3010,8 @@ void ggml_free(struct ggml_context * ctx) {
             GGML_PRINT_DEBUG("%s: context %d with %d objects has been freed. memory used = %zu\n",
                     __func__, i, ctx->n_objects, ctx->objects_end->offs + ctx->objects_end->size);
 
-#if GGML_MLOCK_SUPPORT
-            if (ctx->mem_buffer_mlocked) {
-                if (munlock(ctx->mem_buffer, ctx->mem_size)) {
-                    fprintf(stderr, "%s: failed to munlock buffer: %s\n", __func__, strerror(errno));
-                }
-            }
-#endif
-
             if (ctx->mem_buffer_owned) {
-                free(ctx->mem_buffer);
+                GGML_ALIGNED_FREE(ctx->mem_buffer);
             }
 
             found = true;
@@ -3072,48 +3038,6 @@ size_t ggml_set_scratch(struct ggml_context * ctx, struct ggml_scratch scratch)
     return result;
 }
 
-#ifdef __APPLE__
-#define MLOCK_SUGGESTION \
-    "Try increasing the sysctl values 'vm.user_wire_limit' and 'vm.global_user_wire_limit' and/or " \
-    "decreasing 'vm.global_no_user_wire_amount'.  Also try increasing RLIMIT_MLOCK (ulimit -l).\n"
-#else
-#define MLOCK_SUGGESTION \
-    "Try increasing RLIMIT_MLOCK ('ulimit -l' as root).\n"
-#endif
-
-bool ggml_mlock_supported(void) {
-    return GGML_MLOCK_SUPPORT;
-}
-
-bool ggml_mlock(
-        struct ggml_context * ctx,
-        const void *opt_extra_addr,
-        size_t opt_extra_len,
-        char **err_p) {
-    // TODO: Use SetProcessWorkingSetSize() + VirtualLock() on WIN32
-#if GGML_MLOCK_SUPPORT
-    if (ctx->mem_buffer_mlocked) {
-        return true;
-    }
-    if (mlock(ctx->mem_buffer, ctx->mem_size) ||
-        (opt_extra_len &&
-         mlock(opt_extra_addr, opt_extra_len))) {
-        if ((*err_p = malloc(1024))) {
-            snprintf(*err_p, 1024,
-                     "failed to mlock %zu-byte buffer: %s\n" MLOCK_SUGGESTION,
-                     ctx->mem_size + opt_extra_len,
-                     strerror(errno));
-        }
-        return false;
-    }
-    ctx->mem_buffer_mlocked = true;
-    return true;
-#else // GGML_MLOCK_SUPPORT
-    *err_p = strdup("can't mlock because it's not supported on this system");
-    return false;
-#endif // GGML_MLOCK_SUPPORT
-}
-
 ////////////////////////////////////////////////////////////////////////////////
 
 struct ggml_tensor * ggml_new_tensor_impl(
@@ -4388,6 +4312,41 @@ struct ggml_tensor * ggml_cpy_inplace(
     return ggml_cpy_impl(ctx, a, b, true);
 }
 
+// ggml_cont
+
+struct ggml_tensor * ggml_cont_impl(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        bool inplace) {
+    bool is_node = false;
+
+    if (!inplace && a->grad) {
+        GGML_ASSERT(false); // TODO: implement backward
+        is_node = true;
+    }
+
+    struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
+
+    result->op   = GGML_OP_CONT;
+    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
+    result->src0 = a;
+    result->src1 = NULL;
+
+    return result;
+}
+
+struct ggml_tensor * ggml_cont(
+        struct ggml_context * ctx,
+        struct ggml_tensor * a) {
+    return ggml_cont_impl(ctx, a, false);
+}
+
+struct ggml_tensor * ggml_cont_inplace(
+        struct ggml_context * ctx,
+        struct ggml_tensor * a) {
+    return ggml_cont_impl(ctx, a, true);
+}
+
 // ggml_reshape
 
 struct ggml_tensor * ggml_reshape(
@@ -4930,6 +4889,85 @@ static void ggml_compute_forward_dup_f16(
 
     // TODO: add more special-case implementations for tensor shapes/strides that can benefit from memcpy
 
+    if (ggml_is_contiguous(dst)) {
+        if (src0->nb[0] == sizeof(ggml_fp16_t)) {
+            if (dst->type == GGML_TYPE_F16) {
+                size_t id = 0;
+                const size_t rs = ne00*nb00;
+
+                for (int i03 = 0; i03 < ne03; i03++) {
+                    for (int i02 = 0; i02 < ne02; i02++) {
+                        for (int i01 = 0; i01 < ne01; i01++) {
+                            const char * src0_ptr = (char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03;
+                            char * dst_ptr = (char *) dst->data + id*rs;
+
+                            memcpy(dst_ptr, src0_ptr, rs);
+
+                            id++;
+                        }
+                    }
+                }
+            } else if (dst->type == GGML_TYPE_F32) {
+                size_t id = 0;
+                float * dst_ptr = (float *) dst->data;
+
+                for (int i03 = 0; i03 < ne03; i03++) {
+                    for (int i02 = 0; i02 < ne02; i02++) {
+                        for (int i01 = 0; i01 < ne01; i01++) {
+                            for (int i00 = 0; i00 < ne00; i00++) {
+                                const ggml_fp16_t * src0_ptr = (ggml_fp16_t *) ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03);
+
+                                dst_ptr[id] = GGML_FP16_TO_FP32(*src0_ptr);
+                                id++;
+                            }
+                        }
+                    }
+                }
+            } else {
+                GGML_ASSERT(false); // TODO: implement
+            }
+        } else {
+            //printf("%s: this is not optimal - fix me\n", __func__);
+
+            if (dst->type == GGML_TYPE_F32) {
+                size_t id = 0;
+                float * dst_ptr = (float *) dst->data;
+
+                for (int i03 = 0; i03 < ne03; i03++) {
+                    for (int i02 = 0; i02 < ne02; i02++) {
+                        for (int i01 = 0; i01 < ne01; i01++) {
+                            for (int i00 = 0; i00 < ne00; i00++) {
+                                const ggml_fp16_t * src0_ptr = (ggml_fp16_t *) ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03);
+
+                                dst_ptr[id] = GGML_FP16_TO_FP32(*src0_ptr);
+                                id++;
+                            }
+                        }
+                    }
+                }
+            } else if (dst->type == GGML_TYPE_F16) {
+                size_t id = 0;
+                ggml_fp16_t * dst_ptr = (ggml_fp16_t *) dst->data;
+
+                for (int i03 = 0; i03 < ne03; i03++) {
+                    for (int i02 = 0; i02 < ne02; i02++) {
+                        for (int i01 = 0; i01 < ne01; i01++) {
+                            for (int i00 = 0; i00 < ne00; i00++) {
+                                const ggml_fp16_t * src0_ptr = (ggml_fp16_t *) ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03);
+
+                                dst_ptr[id] = *src0_ptr;
+                                id++;
+                            }
+                        }
+                    }
+                }
+            } else {
+                GGML_ASSERT(false); // TODO: implement
+            }
+        }
+        return;
+    }
+
     // dst counters
     int64_t i10 = 0;
     int64_t i11 = 0;
@@ -5024,6 +5062,105 @@ static void ggml_compute_forward_dup_f32(
         return;
     }
 
+    if (src0->type == dst->type &&
+        src0->ne[0] == dst->ne[0] &&
+        src0->nb[0] == GGML_TYPE_SIZE[src0->type] && dst->nb[0] == GGML_TYPE_SIZE[dst->type]) {
+        // copy by rows
+        const size_t rs = ne00*nb00;
+        for (int64_t i03 = 0; i03 < ne03; i03++) {
+            for (int64_t i02 = 0; i02 < ne02; i02++) {
+                for (int64_t i01 = 0; i01 < ne01; i01++) {
+                    memcpy(
+                        ((char *)  dst->data + i01*nb1  + i02*nb2  + i03*nb3),
+                        ((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03),
+                        rs);
+                }
+            }
+        }
+        return;
+    }
+
+    if (ggml_is_contiguous(dst)) {
+        // TODO: simplify
+        if (src0->nb[0] == sizeof(float)) {
+            if (dst->type == GGML_TYPE_F32) {
+                size_t id = 0;
+                const size_t rs = ne00*nb00;
+
+                for (int i03 = 0; i03 < ne03; i03++) {
+                    for (int i02 = 0; i02 < ne02; i02++) {
+                        for (int i01 = 0; i01 < ne01; i01++) {
+                            const char * src0_ptr = (char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03;
+                            char * dst_ptr = (char *) dst->data + id*rs;
+
+                            memcpy(dst_ptr, src0_ptr, rs);
+
+                            id++;
+                        }
+                    }
+                }
+            } else if (dst->type == GGML_TYPE_F16) {
+                size_t id = 0;
+                ggml_fp16_t * dst_ptr = (ggml_fp16_t *) dst->data;
+
+                for (int i03 = 0; i03 < ne03; i03++) {
+                    for (int i02 = 0; i02 < ne02; i02++) {
+                        for (int i01 = 0; i01 < ne01; i01++) {
+                            for (int i00 = 0; i00 < ne00; i00++) {
+                                const float * src0_ptr = (float *) ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03);
+
+                                dst_ptr[id] = GGML_FP32_TO_FP16(*src0_ptr);
+                                id++;
+                            }
+                        }
+                    }
+                }
+            } else {
+                GGML_ASSERT(false); // TODO: implement
+            }
+        } else {
+            //printf("%s: this is not optimal - fix me\n", __func__);
+
+            if (dst->type == GGML_TYPE_F32) {
+                size_t id = 0;
+                float * dst_ptr = (float *) dst->data;
+
+                for (int i03 = 0; i03 < ne03; i03++) {
+                    for (int i02 = 0; i02 < ne02; i02++) {
+                        for (int i01 = 0; i01 < ne01; i01++) {
+                            for (int i00 = 0; i00 < ne00; i00++) {
+                                const float * src0_ptr = (float *) ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03);
+
+                                dst_ptr[id] = *src0_ptr;
+                                id++;
+                            }
+                        }
+                    }
+                }
+            } else if (dst->type == GGML_TYPE_F16) {
+                size_t id = 0;
+                ggml_fp16_t * dst_ptr = (ggml_fp16_t *) dst->data;
+
+                for (int i03 = 0; i03 < ne03; i03++) {
+                    for (int i02 = 0; i02 < ne02; i02++) {
+                        for (int i01 = 0; i01 < ne01; i01++) {
+                            for (int i00 = 0; i00 < ne00; i00++) {
+                                const float * src0_ptr = (float *) ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03);
+
+                                dst_ptr[id] = GGML_FP32_TO_FP16(*src0_ptr);
+                                id++;
+                            }
+                        }
+                    }
+                }
+            } else {
+                GGML_ASSERT(false); // TODO: implement
+            }
+        }
+
+        return;
+    }
+
     // dst counters
     int64_t i10 = 0;
     int64_t i11 = 0;
@@ -5144,14 +5281,18 @@ static void ggml_compute_forward_add_f32(
     GGML_ASSERT(nb00 == sizeof(float));
 
     if (nb10 == sizeof(float)) {
-        const int j0 = (n/nth)*ith;
-        const int j1 = ith == nth - 1 ? n : (n/nth)*(ith + 1);
-
-        for (int j = j0; j < j1; j++) {
+        for (int j = ith; j < n; j += nth) {
+#ifdef GGML_USE_ACCELERATE
+            vDSP_vadd(
+                    (float *) ((char *) src0->data + j*nb01), 1,
+                    (float *) ((char *) src1->data + j*nb11), 1,
+                    (float *) ((char *) dst->data  + j*nb1),  1, nc);
+#else
             ggml_vec_add_f32(nc,
                     (float *) ((char *) dst->data  + j*nb1),
                     (float *) ((char *) src0->data + j*nb01),
                     (float *) ((char *) src1->data + j*nb11));
+#endif
         }
     } else {
         // src1 is not contiguous
@@ -6304,7 +6445,7 @@ static void ggml_compute_forward_mul_mat_f32(
                 cblas_sgemm(CblasRowMajor, CblasNoTrans, CblasTrans,
                         ne11, ne01, ne10,
                         1.0f,    y, ne10,
-                                 x, ne10,
+                                 x, ne00,
                         0.0f,    d, ne01);
             }
         }
@@ -6476,7 +6617,7 @@ static void ggml_compute_forward_mul_mat_f16_f32(
                 cblas_sgemm(CblasRowMajor, CblasNoTrans, CblasTrans,
                         ne11, ne01, ne10,
                         1.0f,    y, ne10,
-                                 x, ne10,
+                                 x, ne00,
                         0.0f,    d, ne01);
             }
         }
@@ -6564,29 +6705,27 @@ static void ggml_compute_forward_mul_mat_f16_f32(
     //}
 }
 
-typedef void (*dequantize_row_q_t)(const void * restrict x, float * restrict y, int k);
-typedef void (*quantize_row_q_t)(const float * restrict x, void * restrict y, int k);
-typedef void (*vec_dot_q_t)(const int n, float * restrict s, const void * restrict x, const void * restrict y);
-
-typedef struct {
-    dequantize_row_q_t dequantize_row_q;
-    quantize_row_q_t   quantize_row_q;
-    vec_dot_q_t        vec_dot_q;
-} quantize_fns_t;
-
 static const quantize_fns_t quantize_fns[GGML_TYPE_COUNT] = {
     [GGML_TYPE_Q4_0] = {
-        .dequantize_row_q = dequantize_row_q4_0,
-        .quantize_row_q   = quantize_row_q4_0,
-        .vec_dot_q        = ggml_vec_dot_q4_0,
+        .dequantize_row_q         = dequantize_row_q4_0,
+        .quantize_row_q           = quantize_row_q4_0,
+        .quantize_row_q_reference = (quantize_row_q_t) quantize_row_q4_0_reference,
+        .vec_dot_q                = ggml_vec_dot_q4_0,
     },
     [GGML_TYPE_Q4_1] = {
-        .dequantize_row_q = dequantize_row_q4_1,
-        .quantize_row_q   = quantize_row_q4_1,
-        .vec_dot_q        = ggml_vec_dot_q4_1,
+        .dequantize_row_q         = dequantize_row_q4_1,
+        .quantize_row_q           = quantize_row_q4_1,
+        .quantize_row_q_reference = (quantize_row_q_t) quantize_row_q4_1_reference,
+        .vec_dot_q                = ggml_vec_dot_q4_1,
     },
 };
 
+// For internal test use
+quantize_fns_t ggml_internal_get_quantize_fn(size_t i) {
+    GGML_ASSERT(i < GGML_TYPE_COUNT);
+    return quantize_fns[i];
+}
+
 static void ggml_compute_forward_mul_mat_q_f32(
         const struct ggml_compute_params * params,
         const struct ggml_tensor * src0,
@@ -6691,7 +6830,7 @@ static void ggml_compute_forward_mul_mat_q_f32(
                 cblas_sgemm(CblasRowMajor, CblasNoTrans, CblasTrans,
                         ne11, ne01, ne10,
                         1.0f,    y, ne10,
-                                 x, ne10,
+                                 x, ne00,
                         0.0f,    d, ne01);
             }
         }
@@ -6901,6 +7040,15 @@ static void ggml_compute_forward_cpy(
     ggml_compute_forward_dup(params, src0, dst);
 }
 
+// ggml_compute_forward_cont
+
+static void ggml_compute_forward_cont(
+        const struct ggml_compute_params * params,
+        const struct ggml_tensor * src0,
+        struct ggml_tensor * dst) {
+    ggml_compute_forward_dup(params, src0, dst);
+}
+
 // ggml_compute_forward_reshape
 
 static void ggml_compute_forward_reshape(
@@ -8731,6 +8879,10 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
             {
                 ggml_compute_forward_cpy(params, tensor->src0, tensor);
             } break;
+        case GGML_OP_CONT:
+            {
+                ggml_compute_forward_cont(params, tensor->src0, tensor);
+            } break;
         case GGML_OP_RESHAPE:
             {
                 ggml_compute_forward_reshape(params, tensor->src0, tensor);
@@ -8975,8 +9127,9 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
                     src1->grad =
                         ggml_add_impl(ctx,
                                 src1->grad,
-                                // TODO: fix transpose, the node will break the graph connections
-                                ggml_mul_mat(ctx, ggml_transpose(ctx, src0), tensor->grad),
+                                ggml_mul_mat(ctx,
+                                    ggml_cont(ctx, ggml_transpose(ctx, src0)),
+                                    tensor->grad),
                                 inplace);
                 }
             } break;
@@ -8988,6 +9141,10 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
             {
                 GGML_ASSERT(false); // TODO: not implemented
             } break;
+        case GGML_OP_CONT:
+            {
+                GGML_ASSERT(false); // TODO: not implemented
+            } break;
         case GGML_OP_RESHAPE:
             {
                 GGML_ASSERT(false); // TODO: not implemented
@@ -9442,6 +9599,7 @@ void ggml_graph_compute(struct ggml_context * ctx, struct ggml_cgraph * cgraph)
                         node->n_tasks = n_threads;
                     } break;
                 case GGML_OP_CPY:
+                case GGML_OP_CONT:
                 case GGML_OP_RESHAPE:
                 case GGML_OP_VIEW:
                 case GGML_OP_PERMUTE:

ggml.h

@@ -198,13 +198,14 @@ struct ggml_object;
 struct ggml_context;
 
 enum ggml_type {
-    GGML_TYPE_Q4_0,
-    GGML_TYPE_Q4_1,
+    // explicitly numbered values are used in llama.cpp files
+    GGML_TYPE_F32  = 0,
+    GGML_TYPE_F16  = 1,
+    GGML_TYPE_Q4_0 = 2,
+    GGML_TYPE_Q4_1 = 3,
     GGML_TYPE_I8,
     GGML_TYPE_I16,
     GGML_TYPE_I32,
-    GGML_TYPE_F16,
-    GGML_TYPE_F32,
     GGML_TYPE_COUNT,
 };
 
@@ -236,6 +237,7 @@ enum ggml_op {
 
     GGML_OP_SCALE,
     GGML_OP_CPY,
+    GGML_OP_CONT,
     GGML_OP_RESHAPE,
     GGML_OP_VIEW,
     GGML_OP_PERMUTE,
@@ -253,6 +255,19 @@ enum ggml_op {
     GGML_OP_COUNT,
 };
 
+
+// ggml object
+struct ggml_object {
+    size_t offs;
+    size_t size;
+
+    struct ggml_object * next;
+
+    char padding[8];
+};
+
+static const size_t GGML_OBJECT_SIZE = sizeof(struct ggml_object);
+
 // n-dimensional tensor
 struct ggml_tensor {
     enum ggml_type type;
@@ -344,13 +359,6 @@ size_t ggml_used_mem(const struct ggml_context * ctx);
 
 size_t ggml_set_scratch(struct ggml_context * ctx, struct ggml_scratch scratch);
 
-bool ggml_mlock_supported(void);
-bool ggml_mlock(
-        struct ggml_context * ctx,
-        const void *opt_extra_addr,
-        size_t opt_extra_len,
-        char **err_p);
-
 struct ggml_tensor * ggml_new_tensor(
         struct ggml_context * ctx,
         enum   ggml_type type,
@@ -519,6 +527,11 @@ struct ggml_tensor * ggml_cpy(
         struct ggml_tensor  * a,
         struct ggml_tensor  * b);
 
+// make contiguous
+struct ggml_tensor * ggml_cont(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a);
+
 // return view(a), b specifies the new shape
 // TODO: when we start computing gradient, make a copy instead of view
 struct ggml_tensor * ggml_reshape(
@@ -783,6 +796,30 @@ int ggml_cpu_has_blas(void);
 int ggml_cpu_has_sse3(void);
 int ggml_cpu_has_vsx(void);
 
+
+//
+// Internal types and functions exposed for tests and benchmarks
+//
+
+#ifdef  __cplusplus
+// restrict not standard in C++
+#define GGML_RESTRICT
+#else
+#define GGML_RESTRICT restrict
+#endif
+typedef void (*dequantize_row_q_t)(const void * GGML_RESTRICT x, float * GGML_RESTRICT y, int k);
+typedef void (*quantize_row_q_t)(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int k);
+typedef void (*vec_dot_q_t)(const int n, float * GGML_RESTRICT s, const void * GGML_RESTRICT x, const void * GGML_RESTRICT y);
+
+typedef struct {
+    dequantize_row_q_t dequantize_row_q;
+    quantize_row_q_t   quantize_row_q;
+    quantize_row_q_t   quantize_row_q_reference;
+    vec_dot_q_t        vec_dot_q;
+} quantize_fns_t;
+
+quantize_fns_t ggml_internal_get_quantize_fn(size_t i);
+
 #ifdef  __cplusplus
 }
 #endif

llama.h

@@ -55,6 +55,7 @@ extern "C" {
         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
 
@@ -64,8 +65,20 @@ extern "C" {
         void * progress_callback_user_data;
     };
 
+    // model file types
+    enum llama_ftype {
+        LLAMA_FTYPE_ALL_F32     = 0,
+        LLAMA_FTYPE_MOSTLY_F16  = 1,  // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q4_0 = 2,  // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q4_1 = 3,  // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q4_1_SOME_F16 = 4, // tok_embeddings.weight and output.weight are F16
+    };
+
     LLAMA_API struct llama_context_params llama_context_default_params();
 
+    LLAMA_API bool llama_mmap_supported();
+    LLAMA_API bool llama_mlock_supported();
+
     // Various functions for loading a ggml llama model.
     // Allocate (almost) all memory needed for the model.
     // Return NULL on failure
@@ -81,7 +94,7 @@ extern "C" {
     LLAMA_API int llama_model_quantize(
             const char * fname_inp,
             const char * fname_out,
-                   int   itype);
+      enum llama_ftype   ftype);
 
     // Returns the KV cache that will contain the context for the
     // ongoing prediction with the model.
@@ -166,4 +179,4 @@ extern "C" {
 }
 #endif
 
-#endif
+#endif // LLAMA_H

llama_internal.h

@@ -0,0 +1,12 @@
+// Internal header to be included by llama.cpp and tests/benchmarks only.
+
+#ifndef LLAMA_INTERNAL_H
+#define LLAMA_INTERNAL_H
+
+#include <vector>
+#include <string>
+struct ggml_tensor;
+
+std::vector<std::pair<std::string, struct ggml_tensor *>>& llama_internal_get_tensor_map(struct llama_context * ctx);
+
+#endif // LLAMA_INTERNAL_H

llama_util.h

@@ -0,0 +1,389 @@
+// Internal header to be included only by llama.cpp.
+// Contains wrappers around OS interfaces.
+
+#ifndef LLAMA_UTIL_H
+#define LLAMA_UTIL_H
+
+#include <cstdio>
+#include <cstdint>
+#include <cerrno>
+#include <cstring>
+#include <cstdarg>
+#include <cstdlib>
+#include <climits>
+
+#include <string>
+#include <vector>
+
+#ifdef __has_include
+    #if __has_include(<unistd.h>)
+        #include <unistd.h>
+        #if defined(_POSIX_MAPPED_FILES)
+            #include <sys/mman.h>
+        #endif
+    #endif
+#endif
+
+#if defined(_WIN32)
+    #define WIN32_LEAN_AND_MEAN
+    #ifndef NOMINMAX
+        #define NOMINMAX
+    #endif
+    #include <windows.h>
+    #include <io.h>
+    #include <stdio.h> // for _fseeki64
+#endif
+
+#define LLAMA_ASSERT(x) \
+    do { \
+        if (!(x)) { \
+            fprintf(stderr, "LLAMA_ASSERT: %s:%d: %s\n", __FILE__, __LINE__, #x); \
+            abort(); \
+        } \
+    } while (0)
+
+#ifdef __GNUC__
+__attribute__((format(printf, 1, 2)))
+#endif
+static std::string format(const char * fmt, ...) {
+    va_list ap, ap2;
+    va_start(ap, fmt);
+    va_copy(ap2, ap);
+    int size = vsnprintf(NULL, 0, fmt, ap);
+    LLAMA_ASSERT(size >= 0 && size < INT_MAX);
+    std::vector<char> buf(size + 1);
+    int size2 = vsnprintf(buf.data(), size + 1, fmt, ap2);
+    LLAMA_ASSERT(size2 == size);
+    va_end(ap2);
+    va_end(ap);
+    return std::string(buf.data(), size);
+};
+
+struct llama_file {
+    // use FILE * so we don't have to re-open the file to mmap
+    FILE * fp;
+    size_t size;
+
+    llama_file(const char * fname, const char * mode) {
+        fp = std::fopen(fname, mode);
+        if (fp == NULL) {
+            throw format("failed to open %s: %s", fname, std::strerror(errno));
+        }
+        seek(0, SEEK_END);
+        size = tell();
+        seek(0, SEEK_SET);
+    }
+
+    size_t tell() const {
+#ifdef _WIN32
+        __int64 ret = _ftelli64(fp);
+#else
+        long ret = std::ftell(fp);
+#endif
+        LLAMA_ASSERT(ret != -1); // this really shouldn't fail
+        return (size_t) ret;
+    }
+
+    void seek(size_t offset, int whence) {
+#ifdef _WIN32
+        int ret = _fseeki64(fp, (__int64) offset, whence);
+#else
+        int ret = std::fseek(fp, (long) offset, whence);
+#endif
+        LLAMA_ASSERT(ret == 0); // same
+    }
+
+    void read_raw(void * ptr, size_t size) {
+        if (size == 0) {
+            return;
+        }
+        errno = 0;
+        std::size_t ret = std::fread(ptr, size, 1, fp);
+        if (ferror(fp)) {
+            throw format("read error: %s", strerror(errno));
+        }
+        if (ret != 1) {
+            throw std::string("unexpectedly reached end of file");
+        }
+    }
+
+    std::uint32_t read_u32() {
+        std::uint32_t ret;
+        read_raw(&ret, sizeof(ret));
+        return ret;
+    }
+
+    std::string read_string(std::uint32_t len) {
+        std::vector<char> chars(len);
+        read_raw(chars.data(), len);
+        return std::string(chars.data(), len);
+    }
+
+    void write_raw(const void * ptr, size_t size) {
+        if (size == 0) {
+            return;
+        }
+        errno = 0;
+        size_t ret = std::fwrite(ptr, size, 1, fp);
+        if (ret != 1) {
+            throw format("write error: %s", strerror(errno));
+        }
+    }
+
+    void write_u32(std::uint32_t val) {
+        write_raw(&val, sizeof(val));
+    }
+
+    ~llama_file() {
+        if (fp) {
+            std::fclose(fp);
+        }
+    }
+};
+
+#if defined(_WIN32)
+static std::string llama_format_win_err(DWORD err) {
+    LPSTR buf;
+    size_t size = FormatMessageA(FORMAT_MESSAGE_ALLOCATE_BUFFER | FORMAT_MESSAGE_FROM_SYSTEM | FORMAT_MESSAGE_IGNORE_INSERTS,
+                                 NULL, err, MAKELANGID(LANG_NEUTRAL, SUBLANG_DEFAULT), (LPSTR)&buf, 0, NULL);
+    if (!size) {
+        return "FormatMessageA failed";
+    }
+    std::string ret(buf, size);
+    LocalFree(buf);
+    return ret;
+}
+#endif
+
+struct llama_mmap {
+    void * addr;
+    size_t size;
+
+    llama_mmap(const llama_mmap &) = delete;
+
+#ifdef _POSIX_MAPPED_FILES
+    static constexpr bool SUPPORTED = true;
+
+    llama_mmap(struct llama_file * file) {
+        size = file->size;
+        int fd = fileno(file->fp);
+        int flags = MAP_SHARED;
+#ifdef __linux__
+        flags |= MAP_POPULATE;
+#endif
+        addr = mmap(NULL, file->size, PROT_READ, flags, fd, 0);
+        close(fd);
+        if (addr == MAP_FAILED) {
+            throw format("mmap failed: %s", strerror(errno));
+        }
+
+        // Advise the kernel to preload the mapped memory
+        if (madvise(addr, file->size, MADV_WILLNEED)) {
+            fprintf(stderr, "warning: madvise(.., MADV_WILLNEED) failed: %s\n",
+                    strerror(errno));
+        }
+    }
+
+    ~llama_mmap() {
+        munmap(addr, size);
+    }
+#elif defined(_WIN32)
+    static constexpr bool SUPPORTED = true;
+
+    llama_mmap(struct llama_file * file) {
+        size = file->size;
+
+        HANDLE hFile = (HANDLE) _get_osfhandle(_fileno(file->fp));
+
+        HANDLE hMapping = CreateFileMappingA(hFile, NULL, PAGE_READONLY, 0, 0, NULL);
+        DWORD error = GetLastError();
+        CloseHandle(hFile);
+
+        if (hMapping == NULL) {
+            throw format("CreateFileMappingA failed: %s", llama_format_win_err(error).c_str());
+        }
+
+        addr = MapViewOfFile(hMapping, FILE_MAP_READ, 0, 0, 0);
+        error = GetLastError();
+        CloseHandle(hMapping);
+
+        if (addr == NULL) {
+            throw format("MapViewOfFile failed: %s", llama_format_win_err(error).c_str());
+        }
+
+        #if _WIN32_WINNT >= _WIN32_WINNT_WIN8
+        // Advise the kernel to preload the mapped memory
+        WIN32_MEMORY_RANGE_ENTRY range;
+        range.VirtualAddress = addr;
+        range.NumberOfBytes = (SIZE_T)size;
+        if (!PrefetchVirtualMemory(GetCurrentProcess(), 1, &range, 0)) {
+            fprintf(stderr, "warning: PrefetchVirtualMemory failed: %s\n",
+                    llama_format_win_err(GetLastError()).c_str());
+        }
+        #else
+        #pragma message("warning: You are building for pre-Windows 8; prefetch not supported")
+        #endif // _WIN32_WINNT >= _WIN32_WINNT_WIN8
+    }
+
+    ~llama_mmap() {
+        if (!UnmapViewOfFile(addr)) {
+            fprintf(stderr, "warning: UnmapViewOfFile failed: %s\n",
+                    llama_format_win_err(GetLastError()).c_str());
+        }
+    }
+#else
+    static constexpr bool SUPPORTED = false;
+
+    llama_mmap(struct llama_file *) {
+        throw std::string("mmap not supported");
+    }
+#endif
+};
+
+// Represents some region of memory being locked using mlock or VirtualLock;
+// will automatically unlock on destruction.
+struct llama_mlock {
+    void * addr = NULL;
+    size_t size = 0;
+    bool failed_already = false;
+
+    llama_mlock() {}
+    llama_mlock(const llama_mlock &) = delete;
+
+    ~llama_mlock() {
+        if (size) {
+            raw_unlock(addr, size);
+        }
+    }
+
+    void init(void * addr) {
+        LLAMA_ASSERT(this->addr == NULL && this->size == 0);
+        this->addr = addr;
+    }
+
+    void grow_to(size_t target_size) {
+        LLAMA_ASSERT(addr);
+        if (failed_already) {
+            return;
+        }
+        size_t granularity = lock_granularity();
+        target_size = (target_size + granularity - 1) & ~(granularity - 1);
+        if (target_size > size) {
+            if (raw_lock((uint8_t *) addr + size, target_size - size)) {
+                size = target_size;
+            } else {
+                failed_already = true;
+            }
+        }
+    }
+
+#ifdef _POSIX_MEMLOCK_RANGE
+    static constexpr bool SUPPORTED = true;
+
+    size_t lock_granularity() {
+        return (size_t) sysconf(_SC_PAGESIZE);
+    }
+
+    #ifdef __APPLE__
+        #define MLOCK_SUGGESTION \
+            "Try increasing the sysctl values 'vm.user_wire_limit' and 'vm.global_user_wire_limit' and/or " \
+            "decreasing 'vm.global_no_user_wire_amount'.  Also try increasing RLIMIT_MLOCK (ulimit -l).\n"
+    #else
+        #define MLOCK_SUGGESTION \
+            "Try increasing RLIMIT_MLOCK ('ulimit -l' as root).\n"
+    #endif
+
+    bool raw_lock(const void * addr, size_t size) {
+        if (!mlock(addr, size)) {
+            return true;
+        } else {
+            fprintf(stderr, "warning: failed to mlock %zu-byte buffer (after previously locking %zu bytes): %s\n" MLOCK_SUGGESTION,
+                    size, this->size, std::strerror(errno));
+            return false;
+        }
+    }
+
+    #undef MLOCK_SUGGESTION
+
+    void raw_unlock(void * addr, size_t size) {
+        if (munlock(addr, size)) {
+            fprintf(stderr, "warning: failed to munlock buffer: %s\n", std::strerror(errno));
+        }
+    }
+#elif defined(_WIN32)
+    static constexpr bool SUPPORTED = true;
+
+    size_t lock_granularity() {
+        SYSTEM_INFO si;
+        GetSystemInfo(&si);
+        return (size_t) si.dwPageSize;
+    }
+
+    bool raw_lock(void * addr, size_t size) {
+        for (int tries = 1; ; tries++) {
+            if (VirtualLock(addr, size)) {
+                return true;
+            }
+            if (tries == 2) {
+                fprintf(stderr, "warning: failed to VirtualLock %zu-byte buffer (after previously locking %zu bytes): %s\n",
+                        size, this->size, llama_format_win_err(GetLastError()).c_str());
+                return false;
+            }
+
+            // It failed but this was only the first try; increase the working
+            // set size and try again.
+            SIZE_T min_ws_size, max_ws_size;
+            if (!GetProcessWorkingSetSize(GetCurrentProcess(), &min_ws_size, &max_ws_size)) {
+                fprintf(stderr, "warning: GetProcessWorkingSetSize failed: %s\n",
+                        llama_format_win_err(GetLastError()).c_str());
+                return false;
+            }
+            // Per MSDN: "The maximum number of pages that a process can lock
+            // is equal to the number of pages in its minimum working set minus
+            // a small overhead."
+            // Hopefully a megabyte is enough overhead:
+            size_t increment = size + 1048576;
+            // The minimum must be <= the maximum, so we need to increase both:
+            min_ws_size += increment;
+            max_ws_size += increment;
+            if (!SetProcessWorkingSetSize(GetCurrentProcess(), min_ws_size, max_ws_size)) {
+                fprintf(stderr, "warning: SetProcessWorkingSetSize failed: %s\n",
+                        llama_format_win_err(GetLastError()).c_str());
+                return false;
+            }
+        }
+    }
+
+    void raw_unlock(void * addr, size_t size) {
+        if (!VirtualUnlock(addr, size)) {
+            fprintf(stderr, "warning: failed to VirtualUnlock buffer: %s\n",
+                    llama_format_win_err(GetLastError()).c_str());
+        }
+    }
+#else
+    static constexpr bool SUPPORTED = false;
+
+    void raw_lock(const void * addr, size_t size) {
+        fprintf(stderr, "warning: mlock not supported on this system\n");
+    }
+
+    void raw_unlock(const void * addr, size_t size) {}
+#endif
+};
+
+// Replacement for std::vector<uint8_t> that doesn't require zero-initialization.
+struct llama_buffer {
+    uint8_t * addr = NULL;
+    size_t size = 0;
+
+    void resize(size_t size) {
+        delete[] addr;
+        addr = new uint8_t[size];
+        this->size = size;
+    }
+
+    ~llama_buffer() {
+        delete[] addr;
+    }
+};
+#endif

prompts/chat-with-bob.txt

@@ -4,4 +4,4 @@ User: Hello, Bob.
 Bob: Hello. How may I help you today?
 User: Please tell me the largest city in Europe.
 Bob: Sure. The largest city in Europe is Moscow, the capital of Russia.
-User:
+User:

prompts/reason-act.txt

@@ -15,4 +15,4 @@ Answer: The calculate tool says it is 9.3333333333
 Question: What is capital of france?
 Thought: Do I need to use an action? No, I know the answer
 Answer: Paris is the capital of France
-Question:
+Question: