Upload folder using huggingface_hub

.gitattributes

@@ -33,3 +33,5 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+stable-diffusion.cpp/ggml/examples/mnist/models/mnist/mnist_model.state_dict filter=lfs diff=lfs merge=lfs -text
+stable-diffusion.cpp/ggml/examples/mnist/models/mnist/t10k-images.idx3-ubyte filter=lfs diff=lfs merge=lfs -text

stable-diffusion.cpp/.dockerignore

@@ -0,0 +1,6 @@
+build*/
+test/
+
+.cache/
+*.swp
+models/

stable-diffusion.cpp/.github/workflows/build.yml

@@ -0,0 +1,201 @@
+name: CI
+
+on:
+  workflow_dispatch: # allows manual triggering
+    inputs:
+      create_release:
+        description: 'Create new release'
+        required: true
+        type: boolean
+  push:
+    branches:
+      - master
+      - ci
+    paths: ['.github/workflows/**', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu']
+  pull_request:
+    types: [opened, synchronize, reopened]
+    paths: ['**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu']
+
+env:
+  BRANCH_NAME: ${{ github.head_ref || github.ref_name }}
+
+jobs:
+  ubuntu-latest-cmake:
+    runs-on: ubuntu-latest
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v3
+        with:
+          submodules: recursive
+
+
+      - name: Dependencies
+        id: depends
+        run: |
+          sudo apt-get update
+          sudo apt-get install build-essential
+
+      - name: Build
+        id: cmake_build
+        run: |
+          mkdir build
+          cd build
+          cmake ..
+          cmake --build . --config Release
+
+      #- name: Test
+        #id: cmake_test
+        #run: |
+          #cd build
+          #ctest --verbose --timeout 900
+
+  macOS-latest-cmake:
+    runs-on: macos-latest
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v3
+        with:
+          submodules: recursive
+
+      - name: Dependencies
+        id: depends
+        continue-on-error: true
+        run: |
+          brew update
+
+      - name: Build
+        id: cmake_build
+        run: |
+          sysctl -a
+          mkdir build
+          cd build
+          cmake ..
+          cmake --build . --config Release
+
+      #- name: Test
+        #id: cmake_test
+        #run: |
+          #cd build
+          #ctest --verbose --timeout 900
+
+  windows-latest-cmake:
+    runs-on: windows-latest
+
+    strategy:
+      matrix:
+        include:
+          - build: 'noavx'
+            defines: '-DGGML_AVX=OFF -DGGML_AVX2=OFF -DGGML_FMA=OFF'
+          - build: 'avx2'
+            defines: '-DGGML_AVX2=ON'
+          - build: 'avx'
+            defines: '-DGGML_AVX2=OFF'
+          - build: 'avx512'
+            defines: '-DGGML_AVX512=ON'
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v3
+        with:
+          submodules: recursive
+
+      - name: Build
+        id: cmake_build
+        run: |
+          mkdir build
+          cd build
+          cmake .. ${{ matrix.defines }}
+          cmake --build . --config Release
+
+      - name: Check AVX512F support
+        id: check_avx512f
+        if: ${{ matrix.build == 'avx512' }}
+        continue-on-error: true
+        run: |
+          cd build
+          $vcdir = $(vswhere -latest -products * -requires Microsoft.VisualStudio.Component.VC.Tools.x86.x64 -property installationPath)
+          $msvc = $(join-path $vcdir $('VC\Tools\MSVC\'+$(gc -raw $(join-path $vcdir 'VC\Auxiliary\Build\Microsoft.VCToolsVersion.default.txt')).Trim()))
+          $cl =  $(join-path $msvc 'bin\Hostx64\x64\cl.exe')
+          echo 'int main(void){unsigned int a[4];__cpuid(a,7);return !(a[1]&65536);}' >> avx512f.c
+          & $cl /O2 /GS- /kernel avx512f.c /link /nodefaultlib /entry:main
+          .\avx512f.exe && echo "AVX512F: YES" && ( echo HAS_AVX512F=1 >> $env:GITHUB_ENV ) || echo "AVX512F: NO"
+
+      #- name: Test
+        #id: cmake_test
+        #run: |
+          #cd build
+          #ctest -C Release --verbose --timeout 900
+
+      - name: Get commit hash
+        id: commit
+        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
+        uses: pr-mpt/actions-commit-hash@v2
+
+      - name: Pack artifacts
+        id: pack_artifacts
+        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
+        run: |
+          Copy-Item ggml/LICENSE .\build\bin\Release\ggml.txt
+          Copy-Item LICENSE .\build\bin\Release\stable-diffusion.cpp.txt
+          7z a sd-${{ env.BRANCH_NAME }}-${{ steps.commit.outputs.short }}-bin-win-${{ matrix.build }}-x64.zip .\build\bin\Release\*
+
+      - name: Upload artifacts
+        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
+        uses: actions/upload-artifact@v3
+        with:
+          path: |
+            sd-${{ env.BRANCH_NAME }}-${{ steps.commit.outputs.short }}-bin-win-${{ matrix.build }}-x64.zip
+
+  release:
+    if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
+
+    runs-on: ubuntu-latest
+
+    needs:
+      - ubuntu-latest-cmake
+      - macOS-latest-cmake
+      - windows-latest-cmake
+
+    steps:
+      - name: Download artifacts
+        id: download-artifact
+        uses: actions/download-artifact@v3
+
+      - name: Get commit hash
+        id: commit
+        uses: pr-mpt/actions-commit-hash@v2
+
+      - name: Create release
+        id: create_release
+        uses: anzz1/action-create-release@v1
+        env:
+          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+        with:
+          tag_name: ${{ env.BRANCH_NAME }}-${{ steps.commit.outputs.short }}
+
+      - name: Upload release
+        id: upload_release
+        uses: actions/github-script@v3
+        with:
+          github-token: ${{secrets.GITHUB_TOKEN}}
+          script: |
+            const path = require('path');
+            const fs = require('fs');
+            const release_id = '${{ steps.create_release.outputs.id }}';
+            for (let file of await fs.readdirSync('./artifact')) {
+              if (path.extname(file) === '.zip') {
+                console.log('uploadReleaseAsset', file);
+                await github.repos.uploadReleaseAsset({
+                  owner: context.repo.owner,
+                  repo: context.repo.repo,
+                  release_id: release_id,
+                  name: file,
+                  data: await fs.readFileSync(`./artifact/${file}`)
+                });
+              }
+            }

stable-diffusion.cpp/.gitignore

@@ -0,0 +1,5 @@
+build*/
+test/
+
+.cache/
+*.swp

stable-diffusion.cpp/.gitmodules

@@ -0,0 +1,3 @@
+[submodule "ggml"]
+	path = ggml
+	url = https://github.com/leejet/ggml.git

stable-diffusion.cpp/CMakeLists.txt

@@ -0,0 +1,45 @@
+cmake_minimum_required(VERSION 3.12)
+project("stable-diffusion")
+
+set(CMAKE_EXPORT_COMPILE_COMMANDS ON)
+
+if (NOT XCODE AND NOT MSVC AND NOT CMAKE_BUILD_TYPE)
+    set(CMAKE_BUILD_TYPE Release CACHE STRING "Build type" FORCE)
+    set_property(CACHE CMAKE_BUILD_TYPE PROPERTY STRINGS "Debug" "Release" "MinSizeRel" "RelWithDebInfo")
+endif()
+
+set(CMAKE_LIBRARY_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/bin)
+set(CMAKE_RUNTIME_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/bin)
+
+if(CMAKE_SOURCE_DIR STREQUAL CMAKE_CURRENT_SOURCE_DIR)
+    set(SD_STANDALONE ON)
+else()
+    set(SD_STANDALONE OFF)
+endif()
+
+#
+# Option list
+#
+
+# general
+#option(SD_BUILD_TESTS                "sd: build tests"    ${SD_STANDALONE})
+option(SD_BUILD_EXAMPLES             "sd: build examples" ${SD_STANDALONE})
+option(BUILD_SHARED_LIBS             "sd: build shared libs" OFF)
+#option(SD_BUILD_SERVER               "sd: build server example"                           ON)
+
+
+# deps
+add_subdirectory(ggml)
+
+set(SD_LIB stable-diffusion)
+
+add_library(${SD_LIB} stable-diffusion.h stable-diffusion.cpp)
+target_link_libraries(${SD_LIB} PUBLIC ggml)
+target_include_directories(${SD_LIB} PUBLIC .)
+target_compile_features(${SD_LIB} PUBLIC cxx_std_11)
+
+
+if (SD_BUILD_EXAMPLES)
+    add_subdirectory(examples)
+endif()
+

stable-diffusion.cpp/Dockerfile

@@ -0,0 +1,17 @@
+ARG UBUNTU_VERSION=22.04
+
+FROM ubuntu:$UBUNTU_VERSION as build
+
+RUN apt-get update && apt-get install -y build-essential git cmake
+
+WORKDIR /sd.cpp
+
+COPY . .
+
+RUN mkdir build && cd build && cmake .. && cmake --build . --config Release
+
+FROM ubuntu:$UBUNTU_VERSION as runtime
+
+COPY --from=build /sd.cpp/build/bin/sd /sd
+
+ENTRYPOINT [ "/sd" ]

stable-diffusion.cpp/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2023 leejet
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

stable-diffusion.cpp/README.md

@@ -0,0 +1,198 @@
+<p align="center">
+  <img src="./assets/a%20lovely%20cat.png" width="256x">
+</p>
+
+# stable-diffusion.cpp
+
+Inference of [Stable Diffusion](https://github.com/CompVis/stable-diffusion) in pure C/C++
+
+## Features
+
+- Plain C/C++ implementation based on [ggml](https://github.com/ggerganov/ggml), working in the same way as [llama.cpp](https://github.com/ggerganov/llama.cpp)
+- 16-bit, 32-bit float support
+- 4-bit, 5-bit and 8-bit integer quantization support
+- Accelerated memory-efficient CPU inference
+    - Only requires ~2.3GB when using txt2img with fp16 precision to generate a 512x512 image
+- AVX, AVX2 and AVX512 support for x86 architectures
+- SD1.x and SD2.x support
+- Original `txt2img` and `img2img` mode
+- Negative prompt
+- [stable-diffusion-webui](https://github.com/AUTOMATIC1111/stable-diffusion-webui) style tokenizer (not all the features, only token weighting for now)
+- Sampling method
+    - `Euler A`
+    - `Euler`
+    - `Heun`
+    - `DPM2`
+    - `DPM++ 2M`
+    - [`DPM++ 2M v2`](https://github.com/AUTOMATIC1111/stable-diffusion-webui/discussions/8457)
+    - `DPM++ 2S a`
+- Cross-platform reproducibility (`--rng cuda`, consistent with the `stable-diffusion-webui GPU RNG`)
+- Embedds generation parameters into png output as webui-compatible text string
+- Supported platforms
+    - Linux
+    - Mac OS
+    - Windows
+    - Android (via Termux)
+
+### TODO
+
+- [ ] More sampling methods
+- [ ] GPU support
+- [ ] Make inference faster
+    - The current implementation of ggml_conv_2d is slow and has high memory usage
+- [ ] Continuing to reduce memory usage (quantizing the weights of ggml_conv_2d)
+- [ ] LoRA support
+- [ ] k-quants support
+
+## Usage
+
+### Get the Code
+
+```
+git clone --recursive https://github.com/leejet/stable-diffusion.cpp
+cd stable-diffusion.cpp
+```
+
+- If you have already cloned the repository, you can use the following command to update the repository to the latest code.
+
+```
+cd stable-diffusion.cpp
+git pull origin master
+git submodule init
+git submodule update
+```
+
+### Convert weights
+
+- download original weights(.ckpt or .safetensors). For example
+    - Stable Diffusion v1.4 from https://huggingface.co/CompVis/stable-diffusion-v-1-4-original
+    - Stable Diffusion v1.5 from https://huggingface.co/runwayml/stable-diffusion-v1-5
+    - Stable Diffuison v2.1 from https://huggingface.co/stabilityai/stable-diffusion-2-1
+
+    ```shell
+    curl -L -O https://huggingface.co/CompVis/stable-diffusion-v-1-4-original/resolve/main/sd-v1-4.ckpt
+    # curl -L -O https://huggingface.co/runwayml/stable-diffusion-v1-5/resolve/main/v1-5-pruned-emaonly.safetensors
+    # curl -L -O https://huggingface.co/stabilityai/stable-diffusion-2-1/blob/main/v2-1_768-nonema-pruned.safetensors
+    ```
+
+- convert weights to ggml model format
+
+    ```shell
+    cd models
+    pip install -r requirements.txt
+    python convert.py [path to weights] --out_type [output precision]
+    # For example, python convert.py sd-v1-4.ckpt --out_type f16
+    ```
+
+### Quantization
+
+You can specify the output model format using the --out_type parameter
+
+- `f16` for 16-bit floating-point
+- `f32` for 32-bit floating-point
+- `q8_0` for 8-bit integer quantization 
+- `q5_0` or `q5_1` for 5-bit integer quantization 
+- `q4_0` or `q4_1` for 4-bit integer quantization
+
+### Build
+
+#### Build from scratch
+
+```shell
+mkdir build
+cd build
+cmake ..
+cmake --build . --config Release
+```
+
+##### Using OpenBLAS
+
+```
+cmake .. -DGGML_OPENBLAS=ON
+cmake --build . --config Release
+```
+
+### Run
+
+```
+usage: ./bin/sd [arguments]
+
+arguments:
+  -h, --help                         show this help message and exit
+  -M, --mode [txt2img or img2img]    generation mode (default: txt2img)
+  -t, --threads N                    number of threads to use during computation (default: -1).
+                                     If threads <= 0, then threads will be set to the number of CPU physical cores
+  -m, --model [MODEL]                path to model
+  -i, --init-img [IMAGE]             path to the input image, required by img2img
+  -o, --output OUTPUT                path to write result image to (default: .\output.png)
+  -p, --prompt [PROMPT]              the prompt to render
+  -n, --negative-prompt PROMPT       the negative prompt (default: "")
+  --cfg-scale SCALE                  unconditional guidance scale: (default: 7.0)
+  --strength STRENGTH                strength for noising/unnoising (default: 0.75)
+                                     1.0 corresponds to full destruction of information in init image
+  -H, --height H                     image height, in pixel space (default: 512)
+  -W, --width W                      image width, in pixel space (default: 512)
+  --sampling-method {euler, euler_a, heun, dpm++2m, dpm++2mv2}
+                                     sampling method (default: "euler_a")
+  --steps  STEPS                     number of sample steps (default: 20)
+  --rng {std_default, cuda}          RNG (default: cuda)
+  -s SEED, --seed SEED               RNG seed (default: 42, use random seed for < 0)
+  -v, --verbose                      print extra info
+```
+
+#### txt2img example
+
+```
+./bin/sd -m ../models/sd-v1-4-ggml-model-f16.bin -p "a lovely cat"
+```
+
+Using formats of different precisions will yield results of varying quality.
+
+| f32  | f16  |q8_0  |q5_0  |q5_1  |q4_0  |q4_1  |
+| ----  |----  |----  |----  |----  |----  |----  |
+| ![](./assets/f32.png) |![](./assets/f16.png) |![](./assets/q8_0.png) |![](./assets/q5_0.png) |![](./assets/q5_1.png) |![](./assets/q4_0.png) |![](./assets/q4_1.png) |
+
+#### img2img example
+
+- `./output.png` is the image generated from the above txt2img pipeline
+
+
+```
+./bin/sd --mode img2img -m ../models/sd-v1-4-ggml-model-f16.bin -p "cat with blue eyes" -i ./output.png -o ./img2img_output.png --strength 0.4
+```
+
+<p align="center">
+  <img src="./assets/img2img_output.png" width="256x">
+</p>
+
+### Docker
+
+#### Building using Docker
+
+```shell
+docker build -t sd .
+```
+
+#### Run
+
+```shell
+docker run -v /path/to/models:/models -v /path/to/output/:/output sd [args...]
+# For example
+# docker run -v ./models:/models -v ./build:/output sd -m /models/sd-v1-4-ggml-model-f16.bin -p "a lovely cat" -v -o /output/output.png
+```
+
+## Memory/Disk Requirements
+
+| precision | f32  | f16  |q8_0  |q5_0  |q5_1  |q4_0  |q4_1  |
+| ----         | ----  |----  |----  |----  |----  |----  |----  |
+|  **Disk**        | 2.7G | 2.0G | 1.7G | 1.6G | 1.6G | 1.5G | 1.5G |
+|  **Memory**(txt2img - 512 x 512) | ~2.8G | ~2.3G | ~2.1G | ~2.0G | ~2.0G | ~2.0G | ~2.0G |
+
+
+## References
+
+- [ggml](https://github.com/ggerganov/ggml)
+- [stable-diffusion](https://github.com/CompVis/stable-diffusion)
+- [stable-diffusion-stability-ai](https://github.com/Stability-AI/stablediffusion)
+- [stable-diffusion-webui](https://github.com/AUTOMATIC1111/stable-diffusion-webui)
+- [k-diffusion](https://github.com/crowsonkb/k-diffusion)

stable-diffusion.cpp/examples/CMakeLists.txt

@@ -0,0 +1,8 @@
+# TODO: move into its own subdirectoy
+# TODO: make stb libs a target (maybe common)
+set(SD_TARGET sd)
+
+add_executable(${SD_TARGET} main.cpp stb_image.h stb_image_write.h)
+install(TARGETS ${SD_TARGET} RUNTIME)
+target_link_libraries(${SD_TARGET} PRIVATE stable-diffusion ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${SD_TARGET} PUBLIC cxx_std_11)

stable-diffusion.cpp/examples/main.cpp

@@ -0,0 +1,473 @@
+#include <stdio.h>
+#include <ctime>
+#include <fstream>
+#include <iostream>
+#include <random>
+#include <string>
+#include <thread>
+#include <unordered_set>
+
+#include "stable-diffusion.h"
+
+#define STB_IMAGE_IMPLEMENTATION
+#include "stb_image.h"
+
+#define STB_IMAGE_WRITE_IMPLEMENTATION
+#define STB_IMAGE_WRITE_STATIC
+#include "stb_image_write.h"
+
+#if defined(__APPLE__) && defined(__MACH__)
+#include <sys/sysctl.h>
+#include <sys/types.h>
+#endif
+
+#if !defined(_WIN32)
+#include <sys/ioctl.h>
+#include <unistd.h>
+#endif
+
+#define TXT2IMG "txt2img"
+#define IMG2IMG "img2img"
+
+// get_num_physical_cores is copy from
+// https://github.com/ggerganov/llama.cpp/blob/master/examples/common.cpp
+// LICENSE: https://github.com/ggerganov/llama.cpp/blob/master/LICENSE
+int32_t get_num_physical_cores() {
+#ifdef __linux__
+    // enumerate the set of thread siblings, num entries is num cores
+    std::unordered_set<std::string> siblings;
+    for (uint32_t cpu = 0; cpu < UINT32_MAX; ++cpu) {
+        std::ifstream thread_siblings("/sys/devices/system/cpu" + std::to_string(cpu) + "/topology/thread_siblings");
+        if (!thread_siblings.is_open()) {
+            break;  // no more cpus
+        }
+        std::string line;
+        if (std::getline(thread_siblings, line)) {
+            siblings.insert(line);
+        }
+    }
+    if (siblings.size() > 0) {
+        return static_cast<int32_t>(siblings.size());
+    }
+#elif defined(__APPLE__) && defined(__MACH__)
+    int32_t num_physical_cores;
+    size_t len = sizeof(num_physical_cores);
+    int result = sysctlbyname("hw.perflevel0.physicalcpu", &num_physical_cores, &len, NULL, 0);
+    if (result == 0) {
+        return num_physical_cores;
+    }
+    result = sysctlbyname("hw.physicalcpu", &num_physical_cores, &len, NULL, 0);
+    if (result == 0) {
+        return num_physical_cores;
+    }
+#elif defined(_WIN32)
+    // TODO: Implement
+#endif
+    unsigned int n_threads = std::thread::hardware_concurrency();
+    return n_threads > 0 ? (n_threads <= 4 ? n_threads : n_threads / 2) : 4;
+}
+
+const char* rng_type_to_str[] = {
+    "std_default",
+    "cuda",
+};
+
+// Names of the sampler method, same order as enum SampleMethod in stable-diffusion.h
+const char* sample_method_str[] = {
+    "euler_a",
+    "euler",
+    "heun",
+    "dpm2",
+    "dpm++2s_a",
+    "dpm++2m",
+    "dpm++2mv2"};
+
+// Names of the sigma schedule overrides, same order as Schedule in stable-diffusion.h
+const char* schedule_str[] = {
+    "default",
+    "discrete",
+    "karras"};
+
+struct Option {
+    int n_threads = -1;
+    std::string mode = TXT2IMG;
+    std::string model_path;
+    std::string output_path = "output.png";
+    std::string init_img;
+    std::string prompt;
+    std::string negative_prompt;
+    float cfg_scale = 7.0f;
+    int w = 512;
+    int h = 512;
+    SampleMethod sample_method = EULER_A;
+    Schedule schedule = DEFAULT;
+    int sample_steps = 20;
+    float strength = 0.75f;
+    RNGType rng_type = CUDA_RNG;
+    int64_t seed = 42;
+    bool verbose = false;
+
+    void print() {
+        printf("Option: \n");
+        printf("    n_threads:       %d\n", n_threads);
+        printf("    mode:            %s\n", mode.c_str());
+        printf("    model_path:      %s\n", model_path.c_str());
+        printf("    output_path:     %s\n", output_path.c_str());
+        printf("    init_img:        %s\n", init_img.c_str());
+        printf("    prompt:          %s\n", prompt.c_str());
+        printf("    negative_prompt: %s\n", negative_prompt.c_str());
+        printf("    cfg_scale:       %.2f\n", cfg_scale);
+        printf("    width:           %d\n", w);
+        printf("    height:          %d\n", h);
+        printf("    sample_method:   %s\n", sample_method_str[sample_method]);
+        printf("    schedule:        %s\n", schedule_str[schedule]);
+        printf("    sample_steps:    %d\n", sample_steps);
+        printf("    strength:        %.2f\n", strength);
+        printf("    rng:             %s\n", rng_type_to_str[rng_type]);
+        printf("    seed:            %ld\n", seed);
+    }
+};
+
+void print_usage(int argc, const char* argv[]) {
+    printf("usage: %s [arguments]\n", argv[0]);
+    printf("\n");
+    printf("arguments:\n");
+    printf("  -h, --help                         show this help message and exit\n");
+    printf("  -M, --mode [txt2img or img2img]    generation mode (default: txt2img)\n");
+    printf("  -t, --threads N                    number of threads to use during computation (default: -1).\n");
+    printf("                                     If threads <= 0, then threads will be set to the number of CPU physical cores\n");
+    printf("  -m, --model [MODEL]                path to model\n");
+    printf("  -i, --init-img [IMAGE]             path to the input image, required by img2img\n");
+    printf("  -o, --output OUTPUT                path to write result image to (default: .\\output.png)\n");
+    printf("  -p, --prompt [PROMPT]              the prompt to render\n");
+    printf("  -n, --negative-prompt PROMPT       the negative prompt (default: \"\")\n");
+    printf("  --cfg-scale SCALE                  unconditional guidance scale: (default: 7.0)\n");
+    printf("  --strength STRENGTH                strength for noising/unnoising (default: 0.75)\n");
+    printf("                                     1.0 corresponds to full destruction of information in init image\n");
+    printf("  -H, --height H                     image height, in pixel space (default: 512)\n");
+    printf("  -W, --width W                      image width, in pixel space (default: 512)\n");
+    printf("  --sampling-method {euler, euler_a, heun, dpm2, dpm++2s_a, dpm++2m, dpm++2mv2}\n");
+    printf("                                     sampling method (default: \"euler_a\")\n");
+    printf("  --steps  STEPS                     number of sample steps (default: 20)\n");
+    printf("  --rng {std_default, cuda}          RNG (default: cuda)\n");
+    printf("  -s SEED, --seed SEED               RNG seed (default: 42, use random seed for < 0)\n");
+    printf("  --schedule {discrete, karras}      Denoiser sigma schedule (default: discrete)\n");
+    printf("  -v, --verbose                      print extra info\n");
+}
+
+void parse_args(int argc, const char* argv[], Option* opt) {
+    bool invalid_arg = false;
+
+    for (int i = 1; i < argc; i++) {
+        std::string arg = argv[i];
+
+        if (arg == "-t" || arg == "--threads") {
+            if (++i >= argc) {
+                invalid_arg = true;
+                break;
+            }
+            opt->n_threads = std::stoi(argv[i]);
+        } else if (arg == "-M" || arg == "--mode") {
+            if (++i >= argc) {
+                invalid_arg = true;
+                break;
+            }
+            opt->mode = argv[i];
+
+        } else if (arg == "-m" || arg == "--model") {
+            if (++i >= argc) {
+                invalid_arg = true;
+                break;
+            }
+            opt->model_path = argv[i];
+        } else if (arg == "-i" || arg == "--init-img") {
+            if (++i >= argc) {
+                invalid_arg = true;
+                break;
+            }
+            opt->init_img = argv[i];
+        } else if (arg == "-o" || arg == "--output") {
+            if (++i >= argc) {
+                invalid_arg = true;
+                break;
+            }
+            opt->output_path = argv[i];
+        } else if (arg == "-p" || arg == "--prompt") {
+            if (++i >= argc) {
+                invalid_arg = true;
+                break;
+            }
+            opt->prompt = argv[i];
+        } else if (arg == "-n" || arg == "--negative-prompt") {
+            if (++i >= argc) {
+                invalid_arg = true;
+                break;
+            }
+            opt->negative_prompt = argv[i];
+        } else if (arg == "--cfg-scale") {
+            if (++i >= argc) {
+                invalid_arg = true;
+                break;
+            }
+            opt->cfg_scale = std::stof(argv[i]);
+        } else if (arg == "--strength") {
+            if (++i >= argc) {
+                invalid_arg = true;
+                break;
+            }
+            opt->strength = std::stof(argv[i]);
+        } else if (arg == "-H" || arg == "--height") {
+            if (++i >= argc) {
+                invalid_arg = true;
+                break;
+            }
+            opt->h = std::stoi(argv[i]);
+        } else if (arg == "-W" || arg == "--width") {
+            if (++i >= argc) {
+                invalid_arg = true;
+                break;
+            }
+            opt->w = std::stoi(argv[i]);
+        } else if (arg == "--steps") {
+            if (++i >= argc) {
+                invalid_arg = true;
+                break;
+            }
+            opt->sample_steps = std::stoi(argv[i]);
+        } else if (arg == "--rng") {
+            if (++i >= argc) {
+                invalid_arg = true;
+                break;
+            }
+            std::string rng_type_str = argv[i];
+            if (rng_type_str == "std_default") {
+                opt->rng_type = STD_DEFAULT_RNG;
+            } else if (rng_type_str == "cuda") {
+                opt->rng_type = CUDA_RNG;
+            } else {
+                invalid_arg = true;
+                break;
+            }
+        } else if (arg == "--schedule") {
+            if (++i >= argc) {
+                invalid_arg = true;
+                break;
+            }
+            const char* schedule_selected = argv[i];
+            int schedule_found = -1;
+            for (int d = 0; d < N_SCHEDULES; d++) {
+                if (!strcmp(schedule_selected, schedule_str[d])) {
+                    schedule_found = d;
+                }
+            }
+            if (schedule_found == -1) {
+                invalid_arg = true;
+                break;
+            }
+            opt->schedule = (Schedule)schedule_found;
+        } else if (arg == "-s" || arg == "--seed") {
+            if (++i >= argc) {
+                invalid_arg = true;
+                break;
+            }
+            opt->seed = std::stoll(argv[i]);
+        } else if (arg == "--sampling-method") {
+            if (++i >= argc) {
+                invalid_arg = true;
+                break;
+            }
+            const char* sample_method_selected = argv[i];
+            int sample_method_found = -1;
+            for (int m = 0; m < N_SAMPLE_METHODS; m++) {
+                if (!strcmp(sample_method_selected, sample_method_str[m])) {
+                    sample_method_found = m;
+                }
+            }
+            if (sample_method_found == -1) {
+                invalid_arg = true;
+                break;
+            }
+            opt->sample_method = (SampleMethod)sample_method_found;
+        } else if (arg == "-h" || arg == "--help") {
+            print_usage(argc, argv);
+            exit(0);
+        } else if (arg == "-v" || arg == "--verbose") {
+            opt->verbose = true;
+        } else {
+            fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
+            print_usage(argc, argv);
+            exit(1);
+        }
+        if (invalid_arg) {
+            fprintf(stderr, "error: invalid parameter for argument: %s\n", arg.c_str());
+            print_usage(argc, argv);
+            exit(1);
+        }
+    }
+
+    if (opt->n_threads <= 0) {
+        opt->n_threads = get_num_physical_cores();
+    }
+
+    if (opt->mode != TXT2IMG && opt->mode != IMG2IMG) {
+        fprintf(stderr, "error: invalid mode %s, must be one of ['%s', '%s']\n",
+                opt->mode.c_str(), TXT2IMG, IMG2IMG);
+        exit(1);
+    }
+
+    if (opt->prompt.length() == 0) {
+        fprintf(stderr, "error: the following arguments are required: prompt\n");
+        print_usage(argc, argv);
+        exit(1);
+    }
+
+    if (opt->model_path.length() == 0) {
+        fprintf(stderr, "error: the following arguments are required: model_path\n");
+        print_usage(argc, argv);
+        exit(1);
+    }
+
+    if (opt->mode == IMG2IMG && opt->init_img.length() == 0) {
+        fprintf(stderr, "error: when using the img2img mode, the following arguments are required: init-img\n");
+        print_usage(argc, argv);
+        exit(1);
+    }
+
+    if (opt->output_path.length() == 0) {
+        fprintf(stderr, "error: the following arguments are required: output_path\n");
+        print_usage(argc, argv);
+        exit(1);
+    }
+
+    if (opt->w <= 0 || opt->w % 64 != 0) {
+        fprintf(stderr, "error: the width must be a multiple of 64\n");
+        exit(1);
+    }
+
+    if (opt->h <= 0 || opt->h % 64 != 0) {
+        fprintf(stderr, "error: the height must be a multiple of 64\n");
+        exit(1);
+    }
+
+    if (opt->sample_steps <= 0) {
+        fprintf(stderr, "error: the sample_steps must be greater than 0\n");
+        exit(1);
+    }
+
+    if (opt->strength < 0.f || opt->strength > 1.f) {
+        fprintf(stderr, "error: can only work with strength in [0.0, 1.0]\n");
+        exit(1);
+    }
+
+    if (opt->seed < 0) {
+        srand((int)time(NULL));
+        opt->seed = rand();
+    }
+}
+
+std::string basename(const std::string& path) {
+    size_t pos = path.find_last_of('/');
+    if (pos != std::string::npos) {
+        return path.substr(pos + 1);
+    }
+    pos = path.find_last_of('\\');
+    if (pos != std::string::npos) {
+        return path.substr(pos + 1);
+    }
+    return path;
+}
+
+int main(int argc, const char* argv[]) {
+    Option opt;
+    parse_args(argc, argv, &opt);
+
+    if (opt.verbose) {
+        opt.print();
+        printf("%s", sd_get_system_info().c_str());
+        set_sd_log_level(SDLogLevel::DEBUG);
+    }
+
+    bool vae_decode_only = true;
+    std::vector<uint8_t> init_img;
+    if (opt.mode == IMG2IMG) {
+        vae_decode_only = false;
+
+        int c = 0;
+        unsigned char* img_data = stbi_load(opt.init_img.c_str(), &opt.w, &opt.h, &c, 3);
+        if (img_data == NULL) {
+            fprintf(stderr, "load image from '%s' failed\n", opt.init_img.c_str());
+            return 1;
+        }
+        if (c != 3) {
+            fprintf(stderr, "input image must be a 3 channels RGB image, but got %d channels\n", c);
+            free(img_data);
+            return 1;
+        }
+        if (opt.w <= 0 || opt.w % 64 != 0) {
+            fprintf(stderr, "error: the width of image must be a multiple of 64\n");
+            free(img_data);
+            return 1;
+        }
+        if (opt.h <= 0 || opt.h % 64 != 0) {
+            fprintf(stderr, "error: the height of image must be a multiple of 64\n");
+            free(img_data);
+            return 1;
+        }
+        init_img.assign(img_data, img_data + (opt.w * opt.h * c));
+    }
+
+    StableDiffusion sd(opt.n_threads, vae_decode_only, true, opt.rng_type);
+    if (!sd.load_from_file(opt.model_path, opt.schedule)) {
+        return 1;
+    }
+
+    std::vector<uint8_t> img;
+    if (opt.mode == TXT2IMG) {
+        img = sd.txt2img(opt.prompt,
+                         opt.negative_prompt,
+                         opt.cfg_scale,
+                         opt.w,
+                         opt.h,
+                         opt.sample_method,
+                         opt.sample_steps,
+                         opt.seed);
+    } else {
+        img = sd.img2img(init_img,
+                         opt.prompt,
+                         opt.negative_prompt,
+                         opt.cfg_scale,
+                         opt.w,
+                         opt.h,
+                         opt.sample_method,
+                         opt.sample_steps,
+                         opt.strength,
+                         opt.seed);
+    }
+
+    if (img.size() == 0) {
+        fprintf(stderr, "generate failed\n");
+        return 1;
+    }
+
+    std::string parameter_string = opt.prompt + "\n";
+    if (opt.negative_prompt.size() != 0) {
+        parameter_string += "Negative prompt: " + opt.negative_prompt + "\n";
+    }
+    parameter_string += "Steps: " + std::to_string(opt.sample_steps) + ", ";
+    parameter_string += "CFG scale: " + std::to_string(opt.cfg_scale) + ", ";
+    parameter_string += "Seed: " + std::to_string(opt.seed) + ", ";
+    parameter_string += "Size: " + std::to_string(opt.w) + "x" + std::to_string(opt.h) + ", ";
+    parameter_string += "Model: " + basename(opt.model_path) + ", ";
+    parameter_string += "RNG: " + std::string(rng_type_to_str[opt.rng_type]) + ", ";
+    parameter_string += "Sampler: " + std::string(sample_method_str[opt.sample_method]);
+    if (opt.schedule == KARRAS) {
+        parameter_string += " karras";
+    }
+    parameter_string += ", ";
+    parameter_string += "Version: stable-diffusion.cpp";
+
+    stbi_write_png(opt.output_path.c_str(), opt.w, opt.h, 3, img.data(), 0, parameter_string.c_str());
+    printf("save result image to '%s'\n", opt.output_path.c_str());
+
+    return 0;
+}

stable-diffusion.cpp/examples/stb_image_write.h

@@ -0,0 +1,1741 @@
+/* stb_image_write - v1.16 - public domain - http://nothings.org/stb
+   writes out PNG/BMP/TGA/JPEG/HDR images to C stdio - Sean Barrett 2010-2015
+                                     no warranty implied; use at your own risk
+
+   Before #including,
+
+       #define STB_IMAGE_WRITE_IMPLEMENTATION
+
+   in the file that you want to have the implementation.
+
+   Will probably not work correctly with strict-aliasing optimizations.
+
+ABOUT:
+
+   This header file is a library for writing images to C stdio or a callback.
+
+   The PNG output is not optimal; it is 20-50% larger than the file
+   written by a decent optimizing implementation; though providing a custom
+   zlib compress function (see STBIW_ZLIB_COMPRESS) can mitigate that.
+   This library is designed for source code compactness and simplicity,
+   not optimal image file size or run-time performance.
+
+BUILDING:
+
+   You can #define STBIW_ASSERT(x) before the #include to avoid using assert.h.
+   You can #define STBIW_MALLOC(), STBIW_REALLOC(), and STBIW_FREE() to replace
+   malloc,realloc,free.
+   You can #define STBIW_MEMMOVE() to replace memmove()
+   You can #define STBIW_ZLIB_COMPRESS to use a custom zlib-style compress function
+   for PNG compression (instead of the builtin one), it must have the following signature:
+   unsigned char * my_compress(unsigned char *data, int data_len, int *out_len, int quality);
+   The returned data will be freed with STBIW_FREE() (free() by default),
+   so it must be heap allocated with STBIW_MALLOC() (malloc() by default),
+
+UNICODE:
+
+   If compiling for Windows and you wish to use Unicode filenames, compile
+   with
+       #define STBIW_WINDOWS_UTF8
+   and pass utf8-encoded filenames. Call stbiw_convert_wchar_to_utf8 to convert
+   Windows wchar_t filenames to utf8.
+
+USAGE:
+
+   There are five functions, one for each image file format:
+
+     int stbi_write_png(char const *filename, int w, int h, int comp, const void *data, int stride_in_bytes);
+     int stbi_write_bmp(char const *filename, int w, int h, int comp, const void *data);
+     int stbi_write_tga(char const *filename, int w, int h, int comp, const void *data);
+     int stbi_write_jpg(char const *filename, int w, int h, int comp, const void *data, int quality);
+     int stbi_write_hdr(char const *filename, int w, int h, int comp, const float *data);
+
+     void stbi_flip_vertically_on_write(int flag); // flag is non-zero to flip data vertically
+
+   There are also five equivalent functions that use an arbitrary write function. You are
+   expected to open/close your file-equivalent before and after calling these:
+
+     int stbi_write_png_to_func(stbi_write_func *func, void *context, int w, int h, int comp, const void  *data, int stride_in_bytes);
+     int stbi_write_bmp_to_func(stbi_write_func *func, void *context, int w, int h, int comp, const void  *data);
+     int stbi_write_tga_to_func(stbi_write_func *func, void *context, int w, int h, int comp, const void  *data);
+     int stbi_write_hdr_to_func(stbi_write_func *func, void *context, int w, int h, int comp, const float *data);
+     int stbi_write_jpg_to_func(stbi_write_func *func, void *context, int x, int y, int comp, const void *data, int quality);
+
+   where the callback is:
+      void stbi_write_func(void *context, void *data, int size);
+
+   You can configure it with these global variables:
+      int stbi_write_tga_with_rle;             // defaults to true; set to 0 to disable RLE
+      int stbi_write_png_compression_level;    // defaults to 8; set to higher for more compression
+      int stbi_write_force_png_filter;         // defaults to -1; set to 0..5 to force a filter mode
+
+
+   You can define STBI_WRITE_NO_STDIO to disable the file variant of these
+   functions, so the library will not use stdio.h at all. However, this will
+   also disable HDR writing, because it requires stdio for formatted output.
+
+   Each function returns 0 on failure and non-0 on success.
+
+   The functions create an image file defined by the parameters. The image
+   is a rectangle of pixels stored from left-to-right, top-to-bottom.
+   Each pixel contains 'comp' channels of data stored interleaved with 8-bits
+   per channel, in the following order: 1=Y, 2=YA, 3=RGB, 4=RGBA. (Y is
+   monochrome color.) The rectangle is 'w' pixels wide and 'h' pixels tall.
+   The *data pointer points to the first byte of the top-left-most pixel.
+   For PNG, "stride_in_bytes" is the distance in bytes from the first byte of
+   a row of pixels to the first byte of the next row of pixels.
+
+   PNG creates output files with the same number of components as the input.
+   The BMP format expands Y to RGB in the file format and does not
+   output alpha.
+
+   PNG supports writing rectangles of data even when the bytes storing rows of
+   data are not consecutive in memory (e.g. sub-rectangles of a larger image),
+   by supplying the stride between the beginning of adjacent rows. The other
+   formats do not. (Thus you cannot write a native-format BMP through the BMP
+   writer, both because it is in BGR order and because it may have padding
+   at the end of the line.)
+
+   PNG allows you to set the deflate compression level by setting the global
+   variable 'stbi_write_png_compression_level' (it defaults to 8).
+
+   HDR expects linear float data. Since the format is always 32-bit rgb(e)
+   data, alpha (if provided) is discarded, and for monochrome data it is
+   replicated across all three channels.
+
+   TGA supports RLE or non-RLE compressed data. To use non-RLE-compressed
+   data, set the global variable 'stbi_write_tga_with_rle' to 0.
+
+   JPEG does ignore alpha channels in input data; quality is between 1 and 100.
+   Higher quality looks better but results in a bigger image.
+   JPEG baseline (no JPEG progressive).
+
+CREDITS:
+
+
+   Sean Barrett           -    PNG/BMP/TGA
+   Baldur Karlsson        -    HDR
+   Jean-Sebastien Guay    -    TGA monochrome
+   Tim Kelsey             -    misc enhancements
+   Alan Hickman           -    TGA RLE
+   Emmanuel Julien        -    initial file IO callback implementation
+   Jon Olick              -    original jo_jpeg.cpp code
+   Daniel Gibson          -    integrate JPEG, allow external zlib
+   Aarni Koskela          -    allow choosing PNG filter
+
+   bugfixes:
+      github:Chribba
+      Guillaume Chereau
+      github:jry2
+      github:romigrou
+      Sergio Gonzalez
+      Jonas Karlsson
+      Filip Wasil
+      Thatcher Ulrich
+      github:poppolopoppo
+      Patrick Boettcher
+      github:xeekworx
+      Cap Petschulat
+      Simon Rodriguez
+      Ivan Tikhonov
+      github:ignotion
+      Adam Schackart
+      Andrew Kensler
+
+LICENSE
+
+  See end of file for license information.
+
+*/
+
+#ifndef INCLUDE_STB_IMAGE_WRITE_H
+#define INCLUDE_STB_IMAGE_WRITE_H
+
+#include <stdlib.h>
+
+// if STB_IMAGE_WRITE_STATIC causes problems, try defining STBIWDEF to 'inline' or 'static inline'
+#ifndef STBIWDEF
+#ifdef STB_IMAGE_WRITE_STATIC
+#define STBIWDEF  static
+#else
+#ifdef __cplusplus
+#define STBIWDEF  extern "C"
+#else
+#define STBIWDEF  extern
+#endif
+#endif
+#endif
+
+#ifndef STB_IMAGE_WRITE_STATIC  // C++ forbids static forward declarations
+STBIWDEF int stbi_write_tga_with_rle;
+STBIWDEF int stbi_write_png_compression_level;
+STBIWDEF int stbi_write_force_png_filter;
+#endif
+
+#ifndef STBI_WRITE_NO_STDIO
+STBIWDEF int stbi_write_png(char const *filename, int w, int h, int comp, const void  *data, int stride_in_bytes, const char* parameters = NULL);
+STBIWDEF int stbi_write_bmp(char const *filename, int w, int h, int comp, const void  *data);
+STBIWDEF int stbi_write_tga(char const *filename, int w, int h, int comp, const void  *data);
+STBIWDEF int stbi_write_hdr(char const *filename, int w, int h, int comp, const float *data);
+STBIWDEF int stbi_write_jpg(char const *filename, int x, int y, int comp, const void  *data, int quality);
+
+#ifdef STBIW_WINDOWS_UTF8
+STBIWDEF int stbiw_convert_wchar_to_utf8(char *buffer, size_t bufferlen, const wchar_t* input);
+#endif
+#endif
+
+typedef void stbi_write_func(void *context, void *data, int size);
+
+STBIWDEF int stbi_write_png_to_func(stbi_write_func *func, void *context, int w, int h, int comp, const void  *data, int stride_in_bytes);
+STBIWDEF int stbi_write_bmp_to_func(stbi_write_func *func, void *context, int w, int h, int comp, const void  *data);
+STBIWDEF int stbi_write_tga_to_func(stbi_write_func *func, void *context, int w, int h, int comp, const void  *data);
+STBIWDEF int stbi_write_hdr_to_func(stbi_write_func *func, void *context, int w, int h, int comp, const float *data);
+STBIWDEF int stbi_write_jpg_to_func(stbi_write_func *func, void *context, int x, int y, int comp, const void  *data, int quality);
+
+STBIWDEF void stbi_flip_vertically_on_write(int flip_boolean);
+
+#endif//INCLUDE_STB_IMAGE_WRITE_H
+
+#ifdef STB_IMAGE_WRITE_IMPLEMENTATION
+
+#ifdef _WIN32
+   #ifndef _CRT_SECURE_NO_WARNINGS
+   #define _CRT_SECURE_NO_WARNINGS
+   #endif
+   #ifndef _CRT_NONSTDC_NO_DEPRECATE
+   #define _CRT_NONSTDC_NO_DEPRECATE
+   #endif
+#endif
+
+#ifndef STBI_WRITE_NO_STDIO
+#include <stdio.h>
+#endif // STBI_WRITE_NO_STDIO
+
+#include <stdarg.h>
+#include <stdlib.h>
+#include <string.h>
+#include <math.h>
+
+#if defined(STBIW_MALLOC) && defined(STBIW_FREE) && (defined(STBIW_REALLOC) || defined(STBIW_REALLOC_SIZED))
+// ok
+#elif !defined(STBIW_MALLOC) && !defined(STBIW_FREE) && !defined(STBIW_REALLOC) && !defined(STBIW_REALLOC_SIZED)
+// ok
+#else
+#error "Must define all or none of STBIW_MALLOC, STBIW_FREE, and STBIW_REALLOC (or STBIW_REALLOC_SIZED)."
+#endif
+
+#ifndef STBIW_MALLOC
+#define STBIW_MALLOC(sz)        malloc(sz)
+#define STBIW_REALLOC(p,newsz)  realloc(p,newsz)
+#define STBIW_FREE(p)           free(p)
+#endif
+
+#ifndef STBIW_REALLOC_SIZED
+#define STBIW_REALLOC_SIZED(p,oldsz,newsz) STBIW_REALLOC(p,newsz)
+#endif
+
+
+#ifndef STBIW_MEMMOVE
+#define STBIW_MEMMOVE(a,b,sz) memmove(a,b,sz)
+#endif
+
+
+#ifndef STBIW_ASSERT
+#include <assert.h>
+#define STBIW_ASSERT(x) assert(x)
+#endif
+
+#define STBIW_UCHAR(x) (unsigned char) ((x) & 0xff)
+
+#ifdef STB_IMAGE_WRITE_STATIC
+static int stbi_write_png_compression_level = 8;
+static int stbi_write_tga_with_rle = 1;
+static int stbi_write_force_png_filter = -1;
+#else
+int stbi_write_png_compression_level = 8;
+int stbi_write_tga_with_rle = 1;
+int stbi_write_force_png_filter = -1;
+#endif
+
+static int stbi__flip_vertically_on_write = 0;
+
+STBIWDEF void stbi_flip_vertically_on_write(int flag)
+{
+   stbi__flip_vertically_on_write = flag;
+}
+
+typedef struct
+{
+   stbi_write_func *func;
+   void *context;
+   unsigned char buffer[64];
+   int buf_used;
+} stbi__write_context;
+
+// initialize a callback-based context
+static void stbi__start_write_callbacks(stbi__write_context *s, stbi_write_func *c, void *context)
+{
+   s->func    = c;
+   s->context = context;
+}
+
+#ifndef STBI_WRITE_NO_STDIO
+
+static void stbi__stdio_write(void *context, void *data, int size)
+{
+   fwrite(data,1,size,(FILE*) context);
+}
+
+#if defined(_WIN32) && defined(STBIW_WINDOWS_UTF8)
+#ifdef __cplusplus
+#define STBIW_EXTERN extern "C"
+#else
+#define STBIW_EXTERN extern
+#endif
+STBIW_EXTERN __declspec(dllimport) int __stdcall MultiByteToWideChar(unsigned int cp, unsigned long flags, const char *str, int cbmb, wchar_t *widestr, int cchwide);
+STBIW_EXTERN __declspec(dllimport) int __stdcall WideCharToMultiByte(unsigned int cp, unsigned long flags, const wchar_t *widestr, int cchwide, char *str, int cbmb, const char *defchar, int *used_default);
+
+STBIWDEF int stbiw_convert_wchar_to_utf8(char *buffer, size_t bufferlen, const wchar_t* input)
+{
+   return WideCharToMultiByte(65001 /* UTF8 */, 0, input, -1, buffer, (int) bufferlen, NULL, NULL);
+}
+#endif
+
+static FILE *stbiw__fopen(char const *filename, char const *mode)
+{
+   FILE *f;
+#if defined(_WIN32) && defined(STBIW_WINDOWS_UTF8)
+   wchar_t wMode[64];
+   wchar_t wFilename[1024];
+   if (0 == MultiByteToWideChar(65001 /* UTF8 */, 0, filename, -1, wFilename, sizeof(wFilename)/sizeof(*wFilename)))
+      return 0;
+
+   if (0 == MultiByteToWideChar(65001 /* UTF8 */, 0, mode, -1, wMode, sizeof(wMode)/sizeof(*wMode)))
+      return 0;
+
+#if defined(_MSC_VER) && _MSC_VER >= 1400
+   if (0 != _wfopen_s(&f, wFilename, wMode))
+      f = 0;
+#else
+   f = _wfopen(wFilename, wMode);
+#endif
+
+#elif defined(_MSC_VER) && _MSC_VER >= 1400
+   if (0 != fopen_s(&f, filename, mode))
+      f=0;
+#else
+   f = fopen(filename, mode);
+#endif
+   return f;
+}
+
+static int stbi__start_write_file(stbi__write_context *s, const char *filename)
+{
+   FILE *f = stbiw__fopen(filename, "wb");
+   stbi__start_write_callbacks(s, stbi__stdio_write, (void *) f);
+   return f != NULL;
+}
+
+static void stbi__end_write_file(stbi__write_context *s)
+{
+   fclose((FILE *)s->context);
+}
+
+#endif // !STBI_WRITE_NO_STDIO
+
+typedef unsigned int stbiw_uint32;
+typedef int stb_image_write_test[sizeof(stbiw_uint32)==4 ? 1 : -1];
+
+static void stbiw__writefv(stbi__write_context *s, const char *fmt, va_list v)
+{
+   while (*fmt) {
+      switch (*fmt++) {
+         case ' ': break;
+         case '1': { unsigned char x = STBIW_UCHAR(va_arg(v, int));
+                     s->func(s->context,&x,1);
+                     break; }
+         case '2': { int x = va_arg(v,int);
+                     unsigned char b[2];
+                     b[0] = STBIW_UCHAR(x);
+                     b[1] = STBIW_UCHAR(x>>8);
+                     s->func(s->context,b,2);
+                     break; }
+         case '4': { stbiw_uint32 x = va_arg(v,int);
+                     unsigned char b[4];
+                     b[0]=STBIW_UCHAR(x);
+                     b[1]=STBIW_UCHAR(x>>8);
+                     b[2]=STBIW_UCHAR(x>>16);
+                     b[3]=STBIW_UCHAR(x>>24);
+                     s->func(s->context,b,4);
+                     break; }
+         default:
+            STBIW_ASSERT(0);
+            return;
+      }
+   }
+}
+
+static void stbiw__writef(stbi__write_context *s, const char *fmt, ...)
+{
+   va_list v;
+   va_start(v, fmt);
+   stbiw__writefv(s, fmt, v);
+   va_end(v);
+}
+
+static void stbiw__write_flush(stbi__write_context *s)
+{
+   if (s->buf_used) {
+      s->func(s->context, &s->buffer, s->buf_used);
+      s->buf_used = 0;
+   }
+}
+
+static void stbiw__putc(stbi__write_context *s, unsigned char c)
+{
+   s->func(s->context, &c, 1);
+}
+
+static void stbiw__write1(stbi__write_context *s, unsigned char a)
+{
+   if ((size_t)s->buf_used + 1 > sizeof(s->buffer))
+      stbiw__write_flush(s);
+   s->buffer[s->buf_used++] = a;
+}
+
+static void stbiw__write3(stbi__write_context *s, unsigned char a, unsigned char b, unsigned char c)
+{
+   int n;
+   if ((size_t)s->buf_used + 3 > sizeof(s->buffer))
+      stbiw__write_flush(s);
+   n = s->buf_used;
+   s->buf_used = n+3;
+   s->buffer[n+0] = a;
+   s->buffer[n+1] = b;
+   s->buffer[n+2] = c;
+}
+
+static void stbiw__write_pixel(stbi__write_context *s, int rgb_dir, int comp, int write_alpha, int expand_mono, unsigned char *d)
+{
+   unsigned char bg[3] = { 255, 0, 255}, px[3];
+   int k;
+
+   if (write_alpha < 0)
+      stbiw__write1(s, d[comp - 1]);
+
+   switch (comp) {
+      case 2: // 2 pixels = mono + alpha, alpha is written separately, so same as 1-channel case
+      case 1:
+         if (expand_mono)
+            stbiw__write3(s, d[0], d[0], d[0]); // monochrome bmp
+         else
+            stbiw__write1(s, d[0]);  // monochrome TGA
+         break;
+      case 4:
+         if (!write_alpha) {
+            // composite against pink background
+            for (k = 0; k < 3; ++k)
+               px[k] = bg[k] + ((d[k] - bg[k]) * d[3]) / 255;
+            stbiw__write3(s, px[1 - rgb_dir], px[1], px[1 + rgb_dir]);
+            break;
+         }
+         /* FALLTHROUGH */
+      case 3:
+         stbiw__write3(s, d[1 - rgb_dir], d[1], d[1 + rgb_dir]);
+         break;
+   }
+   if (write_alpha > 0)
+      stbiw__write1(s, d[comp - 1]);
+}
+
+static void stbiw__write_pixels(stbi__write_context *s, int rgb_dir, int vdir, int x, int y, int comp, void *data, int write_alpha, int scanline_pad, int expand_mono)
+{
+   stbiw_uint32 zero = 0;
+   int i,j, j_end;
+
+   if (y <= 0)
+      return;
+
+   if (stbi__flip_vertically_on_write)
+      vdir *= -1;
+
+   if (vdir < 0) {
+      j_end = -1; j = y-1;
+   } else {
+      j_end =  y; j = 0;
+   }
+
+   for (; j != j_end; j += vdir) {
+      for (i=0; i < x; ++i) {
+         unsigned char *d = (unsigned char *) data + (j*x+i)*comp;
+         stbiw__write_pixel(s, rgb_dir, comp, write_alpha, expand_mono, d);
+      }
+      stbiw__write_flush(s);
+      s->func(s->context, &zero, scanline_pad);
+   }
+}
+
+static int stbiw__outfile(stbi__write_context *s, int rgb_dir, int vdir, int x, int y, int comp, int expand_mono, void *data, int alpha, int pad, const char *fmt, ...)
+{
+   if (y < 0 || x < 0) {
+      return 0;
+   } else {
+      va_list v;
+      va_start(v, fmt);
+      stbiw__writefv(s, fmt, v);
+      va_end(v);
+      stbiw__write_pixels(s,rgb_dir,vdir,x,y,comp,data,alpha,pad, expand_mono);
+      return 1;
+   }
+}
+
+static int stbi_write_bmp_core(stbi__write_context *s, int x, int y, int comp, const void *data)
+{
+   if (comp != 4) {
+      // write RGB bitmap
+      int pad = (-x*3) & 3;
+      return stbiw__outfile(s,-1,-1,x,y,comp,1,(void *) data,0,pad,
+              "11 4 22 4" "4 44 22 444444",
+              'B', 'M', 14+40+(x*3+pad)*y, 0,0, 14+40,  // file header
+               40, x,y, 1,24, 0,0,0,0,0,0);             // bitmap header
+   } else {
+      // RGBA bitmaps need a v4 header
+      // use BI_BITFIELDS mode with 32bpp and alpha mask
+      // (straight BI_RGB with alpha mask doesn't work in most readers)
+      return stbiw__outfile(s,-1,-1,x,y,comp,1,(void *)data,1,0,
+         "11 4 22 4" "4 44 22 444444 4444 4 444 444 444 444",
+         'B', 'M', 14+108+x*y*4, 0, 0, 14+108, // file header
+         108, x,y, 1,32, 3,0,0,0,0,0, 0xff0000,0xff00,0xff,0xff000000u, 0, 0,0,0, 0,0,0, 0,0,0, 0,0,0); // bitmap V4 header
+   }
+}
+
+STBIWDEF int stbi_write_bmp_to_func(stbi_write_func *func, void *context, int x, int y, int comp, const void *data)
+{
+   stbi__write_context s = { 0 };
+   stbi__start_write_callbacks(&s, func, context);
+   return stbi_write_bmp_core(&s, x, y, comp, data);
+}
+
+#ifndef STBI_WRITE_NO_STDIO
+STBIWDEF int stbi_write_bmp(char const *filename, int x, int y, int comp, const void *data)
+{
+   stbi__write_context s = { 0 };
+   if (stbi__start_write_file(&s,filename)) {
+      int r = stbi_write_bmp_core(&s, x, y, comp, data);
+      stbi__end_write_file(&s);
+      return r;
+   } else
+      return 0;
+}
+#endif //!STBI_WRITE_NO_STDIO
+
+static int stbi_write_tga_core(stbi__write_context *s, int x, int y, int comp, void *data)
+{
+   int has_alpha = (comp == 2 || comp == 4);
+   int colorbytes = has_alpha ? comp-1 : comp;
+   int format = colorbytes < 2 ? 3 : 2; // 3 color channels (RGB/RGBA) = 2, 1 color channel (Y/YA) = 3
+
+   if (y < 0 || x < 0)
+      return 0;
+
+   if (!stbi_write_tga_with_rle) {
+      return stbiw__outfile(s, -1, -1, x, y, comp, 0, (void *) data, has_alpha, 0,
+         "111 221 2222 11", 0, 0, format, 0, 0, 0, 0, 0, x, y, (colorbytes + has_alpha) * 8, has_alpha * 8);
+   } else {
+      int i,j,k;
+      int jend, jdir;
+
+      stbiw__writef(s, "111 221 2222 11", 0,0,format+8, 0,0,0, 0,0,x,y, (colorbytes + has_alpha) * 8, has_alpha * 8);
+
+      if (stbi__flip_vertically_on_write) {
+         j = 0;
+         jend = y;
+         jdir = 1;
+      } else {
+         j = y-1;
+         jend = -1;
+         jdir = -1;
+      }
+      for (; j != jend; j += jdir) {
+         unsigned char *row = (unsigned char *) data + j * x * comp;
+         int len;
+
+         for (i = 0; i < x; i += len) {
+            unsigned char *begin = row + i * comp;
+            int diff = 1;
+            len = 1;
+
+            if (i < x - 1) {
+               ++len;
+               diff = memcmp(begin, row + (i + 1) * comp, comp);
+               if (diff) {
+                  const unsigned char *prev = begin;
+                  for (k = i + 2; k < x && len < 128; ++k) {
+                     if (memcmp(prev, row + k * comp, comp)) {
+                        prev += comp;
+                        ++len;
+                     } else {
+                        --len;
+                        break;
+                     }
+                  }
+               } else {
+                  for (k = i + 2; k < x && len < 128; ++k) {
+                     if (!memcmp(begin, row + k * comp, comp)) {
+                        ++len;
+                     } else {
+                        break;
+                     }
+                  }
+               }
+            }
+
+            if (diff) {
+               unsigned char header = STBIW_UCHAR(len - 1);
+               stbiw__write1(s, header);
+               for (k = 0; k < len; ++k) {
+                  stbiw__write_pixel(s, -1, comp, has_alpha, 0, begin + k * comp);
+               }
+            } else {
+               unsigned char header = STBIW_UCHAR(len - 129);
+               stbiw__write1(s, header);
+               stbiw__write_pixel(s, -1, comp, has_alpha, 0, begin);
+            }
+         }
+      }
+      stbiw__write_flush(s);
+   }
+   return 1;
+}
+
+STBIWDEF int stbi_write_tga_to_func(stbi_write_func *func, void *context, int x, int y, int comp, const void *data)
+{
+   stbi__write_context s = { 0 };
+   stbi__start_write_callbacks(&s, func, context);
+   return stbi_write_tga_core(&s, x, y, comp, (void *) data);
+}
+
+#ifndef STBI_WRITE_NO_STDIO
+STBIWDEF int stbi_write_tga(char const *filename, int x, int y, int comp, const void *data)
+{
+   stbi__write_context s = { 0 };
+   if (stbi__start_write_file(&s,filename)) {
+      int r = stbi_write_tga_core(&s, x, y, comp, (void *) data);
+      stbi__end_write_file(&s);
+      return r;
+   } else
+      return 0;
+}
+#endif
+
+// *************************************************************************************************
+// Radiance RGBE HDR writer
+// by Baldur Karlsson
+
+#define stbiw__max(a, b)  ((a) > (b) ? (a) : (b))
+
+#ifndef STBI_WRITE_NO_STDIO
+
+static void stbiw__linear_to_rgbe(unsigned char *rgbe, float *linear)
+{
+   int exponent;
+   float maxcomp = stbiw__max(linear[0], stbiw__max(linear[1], linear[2]));
+
+   if (maxcomp < 1e-32f) {
+      rgbe[0] = rgbe[1] = rgbe[2] = rgbe[3] = 0;
+   } else {
+      float normalize = (float) frexp(maxcomp, &exponent) * 256.0f/maxcomp;
+
+      rgbe[0] = (unsigned char)(linear[0] * normalize);
+      rgbe[1] = (unsigned char)(linear[1] * normalize);
+      rgbe[2] = (unsigned char)(linear[2] * normalize);
+      rgbe[3] = (unsigned char)(exponent + 128);
+   }
+}
+
+static void stbiw__write_run_data(stbi__write_context *s, int length, unsigned char databyte)
+{
+   unsigned char lengthbyte = STBIW_UCHAR(length+128);
+   STBIW_ASSERT(length+128 <= 255);
+   s->func(s->context, &lengthbyte, 1);
+   s->func(s->context, &databyte, 1);
+}
+
+static void stbiw__write_dump_data(stbi__write_context *s, int length, unsigned char *data)
+{
+   unsigned char lengthbyte = STBIW_UCHAR(length);
+   STBIW_ASSERT(length <= 128); // inconsistent with spec but consistent with official code
+   s->func(s->context, &lengthbyte, 1);
+   s->func(s->context, data, length);
+}
+
+static void stbiw__write_hdr_scanline(stbi__write_context *s, int width, int ncomp, unsigned char *scratch, float *scanline)
+{
+   unsigned char scanlineheader[4] = { 2, 2, 0, 0 };
+   unsigned char rgbe[4];
+   float linear[3];
+   int x;
+
+   scanlineheader[2] = (width&0xff00)>>8;
+   scanlineheader[3] = (width&0x00ff);
+
+   /* skip RLE for images too small or large */
+   if (width < 8 || width >= 32768) {
+      for (x=0; x < width; x++) {
+         switch (ncomp) {
+            case 4: /* fallthrough */
+            case 3: linear[2] = scanline[x*ncomp + 2];
+                    linear[1] = scanline[x*ncomp + 1];
+                    linear[0] = scanline[x*ncomp + 0];
+                    break;
+            default:
+                    linear[0] = linear[1] = linear[2] = scanline[x*ncomp + 0];
+                    break;
+         }
+         stbiw__linear_to_rgbe(rgbe, linear);
+         s->func(s->context, rgbe, 4);
+      }
+   } else {
+      int c,r;
+      /* encode into scratch buffer */
+      for (x=0; x < width; x++) {
+         switch(ncomp) {
+            case 4: /* fallthrough */
+            case 3: linear[2] = scanline[x*ncomp + 2];
+                    linear[1] = scanline[x*ncomp + 1];
+                    linear[0] = scanline[x*ncomp + 0];
+                    break;
+            default:
+                    linear[0] = linear[1] = linear[2] = scanline[x*ncomp + 0];
+                    break;
+         }
+         stbiw__linear_to_rgbe(rgbe, linear);
+         scratch[x + width*0] = rgbe[0];
+         scratch[x + width*1] = rgbe[1];
+         scratch[x + width*2] = rgbe[2];
+         scratch[x + width*3] = rgbe[3];
+      }
+
+      s->func(s->context, scanlineheader, 4);
+
+      /* RLE each component separately */
+      for (c=0; c < 4; c++) {
+         unsigned char *comp = &scratch[width*c];
+
+         x = 0;
+         while (x < width) {
+            // find first run
+            r = x;
+            while (r+2 < width) {
+               if (comp[r] == comp[r+1] && comp[r] == comp[r+2])
+                  break;
+               ++r;
+            }
+            if (r+2 >= width)
+               r = width;
+            // dump up to first run
+            while (x < r) {
+               int len = r-x;
+               if (len > 128) len = 128;
+               stbiw__write_dump_data(s, len, &comp[x]);
+               x += len;
+            }
+            // if there's a run, output it
+            if (r+2 < width) { // same test as what we break out of in search loop, so only true if we break'd
+               // find next byte after run
+               while (r < width && comp[r] == comp[x])
+                  ++r;
+               // output run up to r
+               while (x < r) {
+                  int len = r-x;
+                  if (len > 127) len = 127;
+                  stbiw__write_run_data(s, len, comp[x]);
+                  x += len;
+               }
+            }
+         }
+      }
+   }
+}
+
+static int stbi_write_hdr_core(stbi__write_context *s, int x, int y, int comp, float *data)
+{
+   if (y <= 0 || x <= 0 || data == NULL)
+      return 0;
+   else {
+      // Each component is stored separately. Allocate scratch space for full output scanline.
+      unsigned char *scratch = (unsigned char *) STBIW_MALLOC(x*4);
+      int i, len;
+      char buffer[128];
+      char header[] = "#?RADIANCE\n# Written by stb_image_write.h\nFORMAT=32-bit_rle_rgbe\n";
+      s->func(s->context, header, sizeof(header)-1);
+
+#ifdef __STDC_LIB_EXT1__
+      len = sprintf_s(buffer, sizeof(buffer), "EXPOSURE=          1.0000000000000\n\n-Y %d +X %d\n", y, x);
+#else
+      len = sprintf(buffer, "EXPOSURE=          1.0000000000000\n\n-Y %d +X %d\n", y, x);
+#endif
+      s->func(s->context, buffer, len);
+
+      for(i=0; i < y; i++)
+         stbiw__write_hdr_scanline(s, x, comp, scratch, data + comp*x*(stbi__flip_vertically_on_write ? y-1-i : i));
+      STBIW_FREE(scratch);
+      return 1;
+   }
+}
+
+STBIWDEF int stbi_write_hdr_to_func(stbi_write_func *func, void *context, int x, int y, int comp, const float *data)
+{
+   stbi__write_context s = { 0 };
+   stbi__start_write_callbacks(&s, func, context);
+   return stbi_write_hdr_core(&s, x, y, comp, (float *) data);
+}
+
+STBIWDEF int stbi_write_hdr(char const *filename, int x, int y, int comp, const float *data)
+{
+   stbi__write_context s = { 0 };
+   if (stbi__start_write_file(&s,filename)) {
+      int r = stbi_write_hdr_core(&s, x, y, comp, (float *) data);
+      stbi__end_write_file(&s);
+      return r;
+   } else
+      return 0;
+}
+#endif // STBI_WRITE_NO_STDIO
+
+
+//////////////////////////////////////////////////////////////////////////////
+//
+// PNG writer
+//
+
+#ifndef STBIW_ZLIB_COMPRESS
+// stretchy buffer; stbiw__sbpush() == vector<>::push_back() -- stbiw__sbcount() == vector<>::size()
+#define stbiw__sbraw(a) ((int *) (void *) (a) - 2)
+#define stbiw__sbm(a)   stbiw__sbraw(a)[0]
+#define stbiw__sbn(a)   stbiw__sbraw(a)[1]
+
+#define stbiw__sbneedgrow(a,n)  ((a)==0 || stbiw__sbn(a)+n >= stbiw__sbm(a))
+#define stbiw__sbmaybegrow(a,n) (stbiw__sbneedgrow(a,(n)) ? stbiw__sbgrow(a,n) : 0)
+#define stbiw__sbgrow(a,n)  stbiw__sbgrowf((void **) &(a), (n), sizeof(*(a)))
+
+#define stbiw__sbpush(a, v)      (stbiw__sbmaybegrow(a,1), (a)[stbiw__sbn(a)++] = (v))
+#define stbiw__sbcount(a)        ((a) ? stbiw__sbn(a) : 0)
+#define stbiw__sbfree(a)         ((a) ? STBIW_FREE(stbiw__sbraw(a)),0 : 0)
+
+static void *stbiw__sbgrowf(void **arr, int increment, int itemsize)
+{
+   int m = *arr ? 2*stbiw__sbm(*arr)+increment : increment+1;
+   void *p = STBIW_REALLOC_SIZED(*arr ? stbiw__sbraw(*arr) : 0, *arr ? (stbiw__sbm(*arr)*itemsize + sizeof(int)*2) : 0, itemsize * m + sizeof(int)*2);
+   STBIW_ASSERT(p);
+   if (p) {
+      if (!*arr) ((int *) p)[1] = 0;
+      *arr = (void *) ((int *) p + 2);
+      stbiw__sbm(*arr) = m;
+   }
+   return *arr;
+}
+
+static unsigned char *stbiw__zlib_flushf(unsigned char *data, unsigned int *bitbuffer, int *bitcount)
+{
+   while (*bitcount >= 8) {
+      stbiw__sbpush(data, STBIW_UCHAR(*bitbuffer));
+      *bitbuffer >>= 8;
+      *bitcount -= 8;
+   }
+   return data;
+}
+
+static int stbiw__zlib_bitrev(int code, int codebits)
+{
+   int res=0;
+   while (codebits--) {
+      res = (res << 1) | (code & 1);
+      code >>= 1;
+   }
+   return res;
+}
+
+static unsigned int stbiw__zlib_countm(unsigned char *a, unsigned char *b, int limit)
+{
+   int i;
+   for (i=0; i < limit && i < 258; ++i)
+      if (a[i] != b[i]) break;
+   return i;
+}
+
+static unsigned int stbiw__zhash(unsigned char *data)
+{
+   stbiw_uint32 hash = data[0] + (data[1] << 8) + (data[2] << 16);
+   hash ^= hash << 3;
+   hash += hash >> 5;
+   hash ^= hash << 4;
+   hash += hash >> 17;
+   hash ^= hash << 25;
+   hash += hash >> 6;
+   return hash;
+}
+
+#define stbiw__zlib_flush() (out = stbiw__zlib_flushf(out, &bitbuf, &bitcount))
+#define stbiw__zlib_add(code,codebits) \
+      (bitbuf |= (code) << bitcount, bitcount += (codebits), stbiw__zlib_flush())
+#define stbiw__zlib_huffa(b,c)  stbiw__zlib_add(stbiw__zlib_bitrev(b,c),c)
+// default huffman tables
+#define stbiw__zlib_huff1(n)  stbiw__zlib_huffa(0x30 + (n), 8)
+#define stbiw__zlib_huff2(n)  stbiw__zlib_huffa(0x190 + (n)-144, 9)
+#define stbiw__zlib_huff3(n)  stbiw__zlib_huffa(0 + (n)-256,7)
+#define stbiw__zlib_huff4(n)  stbiw__zlib_huffa(0xc0 + (n)-280,8)
+#define stbiw__zlib_huff(n)  ((n) <= 143 ? stbiw__zlib_huff1(n) : (n) <= 255 ? stbiw__zlib_huff2(n) : (n) <= 279 ? stbiw__zlib_huff3(n) : stbiw__zlib_huff4(n))
+#define stbiw__zlib_huffb(n) ((n) <= 143 ? stbiw__zlib_huff1(n) : stbiw__zlib_huff2(n))
+
+#define stbiw__ZHASH   16384
+
+#endif // STBIW_ZLIB_COMPRESS
+
+STBIWDEF unsigned char * stbi_zlib_compress(unsigned char *data, int data_len, int *out_len, int quality)
+{
+#ifdef STBIW_ZLIB_COMPRESS
+   // user provided a zlib compress implementation, use that
+   return STBIW_ZLIB_COMPRESS(data, data_len, out_len, quality);
+#else // use builtin
+   static unsigned short lengthc[] = { 3,4,5,6,7,8,9,10,11,13,15,17,19,23,27,31,35,43,51,59,67,83,99,115,131,163,195,227,258, 259 };
+   static unsigned char  lengtheb[]= { 0,0,0,0,0,0,0, 0, 1, 1, 1, 1, 2, 2, 2, 2, 3, 3, 3, 3, 4, 4, 4,  4,  5,  5,  5,  5,  0 };
+   static unsigned short distc[]   = { 1,2,3,4,5,7,9,13,17,25,33,49,65,97,129,193,257,385,513,769,1025,1537,2049,3073,4097,6145,8193,12289,16385,24577, 32768 };
+   static unsigned char  disteb[]  = { 0,0,0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7,8,8,9,9,10,10,11,11,12,12,13,13 };
+   unsigned int bitbuf=0;
+   int i,j, bitcount=0;
+   unsigned char *out = NULL;
+   unsigned char ***hash_table = (unsigned char***) STBIW_MALLOC(stbiw__ZHASH * sizeof(unsigned char**));
+   if (hash_table == NULL)
+      return NULL;
+   if (quality < 5) quality = 5;
+
+   stbiw__sbpush(out, 0x78);   // DEFLATE 32K window
+   stbiw__sbpush(out, 0x5e);   // FLEVEL = 1
+   stbiw__zlib_add(1,1);  // BFINAL = 1
+   stbiw__zlib_add(1,2);  // BTYPE = 1 -- fixed huffman
+
+   for (i=0; i < stbiw__ZHASH; ++i)
+      hash_table[i] = NULL;
+
+   i=0;
+   while (i < data_len-3) {
+      // hash next 3 bytes of data to be compressed
+      int h = stbiw__zhash(data+i)&(stbiw__ZHASH-1), best=3;
+      unsigned char *bestloc = 0;
+      unsigned char **hlist = hash_table[h];
+      int n = stbiw__sbcount(hlist);
+      for (j=0; j < n; ++j) {
+         if (hlist[j]-data > i-32768) { // if entry lies within window
+            int d = stbiw__zlib_countm(hlist[j], data+i, data_len-i);
+            if (d >= best) { best=d; bestloc=hlist[j]; }
+         }
+      }
+      // when hash table entry is too long, delete half the entries
+      if (hash_table[h] && stbiw__sbn(hash_table[h]) == 2*quality) {
+         STBIW_MEMMOVE(hash_table[h], hash_table[h]+quality, sizeof(hash_table[h][0])*quality);
+         stbiw__sbn(hash_table[h]) = quality;
+      }
+      stbiw__sbpush(hash_table[h],data+i);
+
+      if (bestloc) {
+         // "lazy matching" - check match at *next* byte, and if it's better, do cur byte as literal
+         h = stbiw__zhash(data+i+1)&(stbiw__ZHASH-1);
+         hlist = hash_table[h];
+         n = stbiw__sbcount(hlist);
+         for (j=0; j < n; ++j) {
+            if (hlist[j]-data > i-32767) {
+               int e = stbiw__zlib_countm(hlist[j], data+i+1, data_len-i-1);
+               if (e > best) { // if next match is better, bail on current match
+                  bestloc = NULL;
+                  break;
+               }
+            }
+         }
+      }
+
+      if (bestloc) {
+         int d = (int) (data+i - bestloc); // distance back
+         STBIW_ASSERT(d <= 32767 && best <= 258);
+         for (j=0; best > lengthc[j+1]-1; ++j);
+         stbiw__zlib_huff(j+257);
+         if (lengtheb[j]) stbiw__zlib_add(best - lengthc[j], lengtheb[j]);
+         for (j=0; d > distc[j+1]-1; ++j);
+         stbiw__zlib_add(stbiw__zlib_bitrev(j,5),5);
+         if (disteb[j]) stbiw__zlib_add(d - distc[j], disteb[j]);
+         i += best;
+      } else {
+         stbiw__zlib_huffb(data[i]);
+         ++i;
+      }
+   }
+   // write out final bytes
+   for (;i < data_len; ++i)
+      stbiw__zlib_huffb(data[i]);
+   stbiw__zlib_huff(256); // end of block
+   // pad with 0 bits to byte boundary
+   while (bitcount)
+      stbiw__zlib_add(0,1);
+
+   for (i=0; i < stbiw__ZHASH; ++i)
+      (void) stbiw__sbfree(hash_table[i]);
+   STBIW_FREE(hash_table);
+
+   // store uncompressed instead if compression was worse
+   if (stbiw__sbn(out) > data_len + 2 + ((data_len+32766)/32767)*5) {
+      stbiw__sbn(out) = 2;  // truncate to DEFLATE 32K window and FLEVEL = 1
+      for (j = 0; j < data_len;) {
+         int blocklen = data_len - j;
+         if (blocklen > 32767) blocklen = 32767;
+         stbiw__sbpush(out, data_len - j == blocklen); // BFINAL = ?, BTYPE = 0 -- no compression
+         stbiw__sbpush(out, STBIW_UCHAR(blocklen)); // LEN
+         stbiw__sbpush(out, STBIW_UCHAR(blocklen >> 8));
+         stbiw__sbpush(out, STBIW_UCHAR(~blocklen)); // NLEN
+         stbiw__sbpush(out, STBIW_UCHAR(~blocklen >> 8));
+         memcpy(out+stbiw__sbn(out), data+j, blocklen);
+         stbiw__sbn(out) += blocklen;
+         j += blocklen;
+      }
+   }
+
+   {
+      // compute adler32 on input
+      unsigned int s1=1, s2=0;
+      int blocklen = (int) (data_len % 5552);
+      j=0;
+      while (j < data_len) {
+         for (i=0; i < blocklen; ++i) { s1 += data[j+i]; s2 += s1; }
+         s1 %= 65521; s2 %= 65521;
+         j += blocklen;
+         blocklen = 5552;
+      }
+      stbiw__sbpush(out, STBIW_UCHAR(s2 >> 8));
+      stbiw__sbpush(out, STBIW_UCHAR(s2));
+      stbiw__sbpush(out, STBIW_UCHAR(s1 >> 8));
+      stbiw__sbpush(out, STBIW_UCHAR(s1));
+   }
+   *out_len = stbiw__sbn(out);
+   // make returned pointer freeable
+   STBIW_MEMMOVE(stbiw__sbraw(out), out, *out_len);
+   return (unsigned char *) stbiw__sbraw(out);
+#endif // STBIW_ZLIB_COMPRESS
+}
+
+static unsigned int stbiw__crc32(unsigned char *buffer, int len)
+{
+#ifdef STBIW_CRC32
+    return STBIW_CRC32(buffer, len);
+#else
+   static unsigned int crc_table[256] =
+   {
+      0x00000000, 0x77073096, 0xEE0E612C, 0x990951BA, 0x076DC419, 0x706AF48F, 0xE963A535, 0x9E6495A3,
+      0x0eDB8832, 0x79DCB8A4, 0xE0D5E91E, 0x97D2D988, 0x09B64C2B, 0x7EB17CBD, 0xE7B82D07, 0x90BF1D91,
+      0x1DB71064, 0x6AB020F2, 0xF3B97148, 0x84BE41DE, 0x1ADAD47D, 0x6DDDE4EB, 0xF4D4B551, 0x83D385C7,
+      0x136C9856, 0x646BA8C0, 0xFD62F97A, 0x8A65C9EC, 0x14015C4F, 0x63066CD9, 0xFA0F3D63, 0x8D080DF5,
+      0x3B6E20C8, 0x4C69105E, 0xD56041E4, 0xA2677172, 0x3C03E4D1, 0x4B04D447, 0xD20D85FD, 0xA50AB56B,
+      0x35B5A8FA, 0x42B2986C, 0xDBBBC9D6, 0xACBCF940, 0x32D86CE3, 0x45DF5C75, 0xDCD60DCF, 0xABD13D59,
+      0x26D930AC, 0x51DE003A, 0xC8D75180, 0xBFD06116, 0x21B4F4B5, 0x56B3C423, 0xCFBA9599, 0xB8BDA50F,
+      0x2802B89E, 0x5F058808, 0xC60CD9B2, 0xB10BE924, 0x2F6F7C87, 0x58684C11, 0xC1611DAB, 0xB6662D3D,
+      0x76DC4190, 0x01DB7106, 0x98D220BC, 0xEFD5102A, 0x71B18589, 0x06B6B51F, 0x9FBFE4A5, 0xE8B8D433,
+      0x7807C9A2, 0x0F00F934, 0x9609A88E, 0xE10E9818, 0x7F6A0DBB, 0x086D3D2D, 0x91646C97, 0xE6635C01,
+      0x6B6B51F4, 0x1C6C6162, 0x856530D8, 0xF262004E, 0x6C0695ED, 0x1B01A57B, 0x8208F4C1, 0xF50FC457,
+      0x65B0D9C6, 0x12B7E950, 0x8BBEB8EA, 0xFCB9887C, 0x62DD1DDF, 0x15DA2D49, 0x8CD37CF3, 0xFBD44C65,
+      0x4DB26158, 0x3AB551CE, 0xA3BC0074, 0xD4BB30E2, 0x4ADFA541, 0x3DD895D7, 0xA4D1C46D, 0xD3D6F4FB,
+      0x4369E96A, 0x346ED9FC, 0xAD678846, 0xDA60B8D0, 0x44042D73, 0x33031DE5, 0xAA0A4C5F, 0xDD0D7CC9,
+      0x5005713C, 0x270241AA, 0xBE0B1010, 0xC90C2086, 0x5768B525, 0x206F85B3, 0xB966D409, 0xCE61E49F,
+      0x5EDEF90E, 0x29D9C998, 0xB0D09822, 0xC7D7A8B4, 0x59B33D17, 0x2EB40D81, 0xB7BD5C3B, 0xC0BA6CAD,
+      0xEDB88320, 0x9ABFB3B6, 0x03B6E20C, 0x74B1D29A, 0xEAD54739, 0x9DD277AF, 0x04DB2615, 0x73DC1683,
+      0xE3630B12, 0x94643B84, 0x0D6D6A3E, 0x7A6A5AA8, 0xE40ECF0B, 0x9309FF9D, 0x0A00AE27, 0x7D079EB1,
+      0xF00F9344, 0x8708A3D2, 0x1E01F268, 0x6906C2FE, 0xF762575D, 0x806567CB, 0x196C3671, 0x6E6B06E7,
+      0xFED41B76, 0x89D32BE0, 0x10DA7A5A, 0x67DD4ACC, 0xF9B9DF6F, 0x8EBEEFF9, 0x17B7BE43, 0x60B08ED5,
+      0xD6D6A3E8, 0xA1D1937E, 0x38D8C2C4, 0x4FDFF252, 0xD1BB67F1, 0xA6BC5767, 0x3FB506DD, 0x48B2364B,
+      0xD80D2BDA, 0xAF0A1B4C, 0x36034AF6, 0x41047A60, 0xDF60EFC3, 0xA867DF55, 0x316E8EEF, 0x4669BE79,
+      0xCB61B38C, 0xBC66831A, 0x256FD2A0, 0x5268E236, 0xCC0C7795, 0xBB0B4703, 0x220216B9, 0x5505262F,
+      0xC5BA3BBE, 0xB2BD0B28, 0x2BB45A92, 0x5CB36A04, 0xC2D7FFA7, 0xB5D0CF31, 0x2CD99E8B, 0x5BDEAE1D,
+      0x9B64C2B0, 0xEC63F226, 0x756AA39C, 0x026D930A, 0x9C0906A9, 0xEB0E363F, 0x72076785, 0x05005713,
+      0x95BF4A82, 0xE2B87A14, 0x7BB12BAE, 0x0CB61B38, 0x92D28E9B, 0xE5D5BE0D, 0x7CDCEFB7, 0x0BDBDF21,
+      0x86D3D2D4, 0xF1D4E242, 0x68DDB3F8, 0x1FDA836E, 0x81BE16CD, 0xF6B9265B, 0x6FB077E1, 0x18B74777,
+      0x88085AE6, 0xFF0F6A70, 0x66063BCA, 0x11010B5C, 0x8F659EFF, 0xF862AE69, 0x616BFFD3, 0x166CCF45,
+      0xA00AE278, 0xD70DD2EE, 0x4E048354, 0x3903B3C2, 0xA7672661, 0xD06016F7, 0x4969474D, 0x3E6E77DB,
+      0xAED16A4A, 0xD9D65ADC, 0x40DF0B66, 0x37D83BF0, 0xA9BCAE53, 0xDEBB9EC5, 0x47B2CF7F, 0x30B5FFE9,
+      0xBDBDF21C, 0xCABAC28A, 0x53B39330, 0x24B4A3A6, 0xBAD03605, 0xCDD70693, 0x54DE5729, 0x23D967BF,
+      0xB3667A2E, 0xC4614AB8, 0x5D681B02, 0x2A6F2B94, 0xB40BBE37, 0xC30C8EA1, 0x5A05DF1B, 0x2D02EF8D
+   };
+
+   unsigned int crc = ~0u;
+   int i;
+   for (i=0; i < len; ++i)
+      crc = (crc >> 8) ^ crc_table[buffer[i] ^ (crc & 0xff)];
+   return ~crc;
+#endif
+}
+
+#define stbiw__wpng4(o,a,b,c,d) ((o)[0]=STBIW_UCHAR(a),(o)[1]=STBIW_UCHAR(b),(o)[2]=STBIW_UCHAR(c),(o)[3]=STBIW_UCHAR(d),(o)+=4)
+#define stbiw__wp32(data,v) stbiw__wpng4(data, (v)>>24,(v)>>16,(v)>>8,(v));
+#define stbiw__wptag(data,s) stbiw__wpng4(data, s[0],s[1],s[2],s[3])
+
+static void stbiw__wpcrc(unsigned char **data, int len)
+{
+   unsigned int crc = stbiw__crc32(*data - len - 4, len+4);
+   stbiw__wp32(*data, crc);
+}
+
+static unsigned char stbiw__paeth(int a, int b, int c)
+{
+   int p = a + b - c, pa = abs(p-a), pb = abs(p-b), pc = abs(p-c);
+   if (pa <= pb && pa <= pc) return STBIW_UCHAR(a);
+   if (pb <= pc) return STBIW_UCHAR(b);
+   return STBIW_UCHAR(c);
+}
+
+// @OPTIMIZE: provide an option that always forces left-predict or paeth predict
+static void stbiw__encode_png_line(unsigned char *pixels, int stride_bytes, int width, int height, int y, int n, int filter_type, signed char *line_buffer)
+{
+   static int mapping[] = { 0,1,2,3,4 };
+   static int firstmap[] = { 0,1,0,5,6 };
+   int *mymap = (y != 0) ? mapping : firstmap;
+   int i;
+   int type = mymap[filter_type];
+   unsigned char *z = pixels + stride_bytes * (stbi__flip_vertically_on_write ? height-1-y : y);
+   int signed_stride = stbi__flip_vertically_on_write ? -stride_bytes : stride_bytes;
+
+   if (type==0) {
+      memcpy(line_buffer, z, width*n);
+      return;
+   }
+
+   // first loop isn't optimized since it's just one pixel
+   for (i = 0; i < n; ++i) {
+      switch (type) {
+         case 1: line_buffer[i] = z[i]; break;
+         case 2: line_buffer[i] = z[i] - z[i-signed_stride]; break;
+         case 3: line_buffer[i] = z[i] - (z[i-signed_stride]>>1); break;
+         case 4: line_buffer[i] = (signed char) (z[i] - stbiw__paeth(0,z[i-signed_stride],0)); break;
+         case 5: line_buffer[i] = z[i]; break;
+         case 6: line_buffer[i] = z[i]; break;
+      }
+   }
+   switch (type) {
+      case 1: for (i=n; i < width*n; ++i) line_buffer[i] = z[i] - z[i-n]; break;
+      case 2: for (i=n; i < width*n; ++i) line_buffer[i] = z[i] - z[i-signed_stride]; break;
+      case 3: for (i=n; i < width*n; ++i) line_buffer[i] = z[i] - ((z[i-n] + z[i-signed_stride])>>1); break;
+      case 4: for (i=n; i < width*n; ++i) line_buffer[i] = z[i] - stbiw__paeth(z[i-n], z[i-signed_stride], z[i-signed_stride-n]); break;
+      case 5: for (i=n; i < width*n; ++i) line_buffer[i] = z[i] - (z[i-n]>>1); break;
+      case 6: for (i=n; i < width*n; ++i) line_buffer[i] = z[i] - stbiw__paeth(z[i-n], 0,0); break;
+   }
+}
+
+STBIWDEF unsigned char *stbi_write_png_to_mem(const unsigned char *pixels, int stride_bytes, int x, int y, int n, int *out_len, const char* parameters)
+{
+   int force_filter = stbi_write_force_png_filter;
+   int param_length = 0;
+   int ctype[5] = { -1, 0, 4, 2, 6 };
+   unsigned char sig[8] = { 137,80,78,71,13,10,26,10 };
+   unsigned char *out,*o, *filt, *zlib;
+   signed char *line_buffer;
+   int j,zlen;
+
+   if (stride_bytes == 0)
+      stride_bytes = x * n;
+
+   if (force_filter >= 5) {
+      force_filter = -1;
+   }
+
+   filt = (unsigned char *) STBIW_MALLOC((x*n+1) * y); if (!filt) return 0;
+   line_buffer = (signed char *) STBIW_MALLOC(x * n); if (!line_buffer) { STBIW_FREE(filt); return 0; }
+   for (j=0; j < y; ++j) {
+      int filter_type;
+      if (force_filter > -1) {
+         filter_type = force_filter;
+         stbiw__encode_png_line((unsigned char*)(pixels), stride_bytes, x, y, j, n, force_filter, line_buffer);
+      } else { // Estimate the best filter by running through all of them:
+         int best_filter = 0, best_filter_val = 0x7fffffff, est, i;
+         for (filter_type = 0; filter_type < 5; filter_type++) {
+            stbiw__encode_png_line((unsigned char*)(pixels), stride_bytes, x, y, j, n, filter_type, line_buffer);
+
+            // Estimate the entropy of the line using this filter; the less, the better.
+            est = 0;
+            for (i = 0; i < x*n; ++i) {
+               est += abs((signed char) line_buffer[i]);
+            }
+            if (est < best_filter_val) {
+               best_filter_val = est;
+               best_filter = filter_type;
+            }
+         }
+         if (filter_type != best_filter) {  // If the last iteration already got us the best filter, don't redo it
+            stbiw__encode_png_line((unsigned char*)(pixels), stride_bytes, x, y, j, n, best_filter, line_buffer);
+            filter_type = best_filter;
+         }
+      }
+      // when we get here, filter_type contains the filter type, and line_buffer contains the data
+      filt[j*(x*n+1)] = (unsigned char) filter_type;
+      STBIW_MEMMOVE(filt+j*(x*n+1)+1, line_buffer, x*n);
+   }
+   STBIW_FREE(line_buffer);
+   zlib = stbi_zlib_compress(filt, y*( x*n+1), &zlen, stbi_write_png_compression_level);
+   STBIW_FREE(filt);
+   if (!zlib) return 0;
+
+   if(parameters != NULL) {
+      param_length = strlen(parameters);
+      param_length += strlen("parameters") + 1; // For the name and the null-byte
+   }
+
+   // each tag requires 12 bytes of overhead
+   out = (unsigned char *) STBIW_MALLOC(8 + 12+13 + 12+zlen + 12 + ((parameters)?(param_length+12):0));
+   if (!out) return 0;
+   *out_len = 8 + 12+13 + 12+zlen + 12 + ((parameters)?(param_length+12):0);
+
+   o=out;
+   STBIW_MEMMOVE(o,sig,8); o+= 8;
+   stbiw__wp32(o, 13); // header length
+   stbiw__wptag(o, "IHDR");
+   stbiw__wp32(o, x);
+   stbiw__wp32(o, y);
+   *o++ = 8;
+   *o++ = STBIW_UCHAR(ctype[n]);
+   *o++ = 0;
+   *o++ = 0;
+   *o++ = 0;
+   stbiw__wpcrc(&o,13);
+
+   if(parameters != NULL) {
+      stbiw__wp32(o, param_length);
+      stbiw__wptag(o, "tEXt");
+      STBIW_MEMMOVE(o, "parameters", strlen("parameters"));
+      o+=strlen("parameters");
+      *o++ = 0; // Null pyte separator
+      STBIW_MEMMOVE(o, parameters, strlen(parameters));
+      o+=strlen(parameters);
+      stbiw__wpcrc(&o, param_length);
+   }
+
+   stbiw__wp32(o, zlen);
+   stbiw__wptag(o, "IDAT");
+   STBIW_MEMMOVE(o, zlib, zlen);
+   o += zlen;
+   STBIW_FREE(zlib);
+   stbiw__wpcrc(&o, zlen);
+
+   stbiw__wp32(o,0);
+   stbiw__wptag(o, "IEND");
+   stbiw__wpcrc(&o,0);
+
+   STBIW_ASSERT(o == out + *out_len);
+
+   return out;
+}
+
+#ifndef STBI_WRITE_NO_STDIO
+STBIWDEF int stbi_write_png(char const *filename, int x, int y, int comp, const void *data, int stride_bytes, const char* parameters)
+{
+   FILE *f;
+   int len;
+   unsigned char *png = stbi_write_png_to_mem((const unsigned char *) data, stride_bytes, x, y, comp, &len, parameters);
+   if (png == NULL) return 0;
+
+   f = stbiw__fopen(filename, "wb");
+   if (!f) { STBIW_FREE(png); return 0; }
+   fwrite(png, 1, len, f);
+   fclose(f);
+   STBIW_FREE(png);
+   return 1;
+}
+#endif
+
+STBIWDEF int stbi_write_png_to_func(stbi_write_func *func, void *context, int x, int y, int comp, const void *data, int stride_bytes)
+{
+   int len;
+   unsigned char *png = stbi_write_png_to_mem((const unsigned char *) data, stride_bytes, x, y, comp, &len, NULL);
+   if (png == NULL) return 0;
+   func(context, png, len);
+   STBIW_FREE(png);
+   return 1;
+}
+
+
+/* ***************************************************************************
+ *
+ * JPEG writer
+ *
+ * This is based on Jon Olick's jo_jpeg.cpp:
+ * public domain Simple, Minimalistic JPEG writer - http://www.jonolick.com/code.html
+ */
+
+static const unsigned char stbiw__jpg_ZigZag[] = { 0,1,5,6,14,15,27,28,2,4,7,13,16,26,29,42,3,8,12,17,25,30,41,43,9,11,18,
+      24,31,40,44,53,10,19,23,32,39,45,52,54,20,22,33,38,46,51,55,60,21,34,37,47,50,56,59,61,35,36,48,49,57,58,62,63 };
+
+static void stbiw__jpg_writeBits(stbi__write_context *s, int *bitBufP, int *bitCntP, const unsigned short *bs) {
+   int bitBuf = *bitBufP, bitCnt = *bitCntP;
+   bitCnt += bs[1];
+   bitBuf |= bs[0] << (24 - bitCnt);
+   while(bitCnt >= 8) {
+      unsigned char c = (bitBuf >> 16) & 255;
+      stbiw__putc(s, c);
+      if(c == 255) {
+         stbiw__putc(s, 0);
+      }
+      bitBuf <<= 8;
+      bitCnt -= 8;
+   }
+   *bitBufP = bitBuf;
+   *bitCntP = bitCnt;
+}
+
+static void stbiw__jpg_DCT(float *d0p, float *d1p, float *d2p, float *d3p, float *d4p, float *d5p, float *d6p, float *d7p) {
+   float d0 = *d0p, d1 = *d1p, d2 = *d2p, d3 = *d3p, d4 = *d4p, d5 = *d5p, d6 = *d6p, d7 = *d7p;
+   float z1, z2, z3, z4, z5, z11, z13;
+
+   float tmp0 = d0 + d7;
+   float tmp7 = d0 - d7;
+   float tmp1 = d1 + d6;
+   float tmp6 = d1 - d6;
+   float tmp2 = d2 + d5;
+   float tmp5 = d2 - d5;
+   float tmp3 = d3 + d4;
+   float tmp4 = d3 - d4;
+
+   // Even part
+   float tmp10 = tmp0 + tmp3;   // phase 2
+   float tmp13 = tmp0 - tmp3;
+   float tmp11 = tmp1 + tmp2;
+   float tmp12 = tmp1 - tmp2;
+
+   d0 = tmp10 + tmp11;       // phase 3
+   d4 = tmp10 - tmp11;
+
+   z1 = (tmp12 + tmp13) * 0.707106781f; // c4
+   d2 = tmp13 + z1;       // phase 5
+   d6 = tmp13 - z1;
+
+   // Odd part
+   tmp10 = tmp4 + tmp5;       // phase 2
+   tmp11 = tmp5 + tmp6;
+   tmp12 = tmp6 + tmp7;
+
+   // The rotator is modified from fig 4-8 to avoid extra negations.
+   z5 = (tmp10 - tmp12) * 0.382683433f; // c6
+   z2 = tmp10 * 0.541196100f + z5; // c2-c6
+   z4 = tmp12 * 1.306562965f + z5; // c2+c6
+   z3 = tmp11 * 0.707106781f; // c4
+
+   z11 = tmp7 + z3;      // phase 5
+   z13 = tmp7 - z3;
+
+   *d5p = z13 + z2;         // phase 6
+   *d3p = z13 - z2;
+   *d1p = z11 + z4;
+   *d7p = z11 - z4;
+
+   *d0p = d0;  *d2p = d2;  *d4p = d4;  *d6p = d6;
+}
+
+static void stbiw__jpg_calcBits(int val, unsigned short bits[2]) {
+   int tmp1 = val < 0 ? -val : val;
+   val = val < 0 ? val-1 : val;
+   bits[1] = 1;
+   while(tmp1 >>= 1) {
+      ++bits[1];
+   }
+   bits[0] = val & ((1<<bits[1])-1);
+}
+
+static int stbiw__jpg_processDU(stbi__write_context *s, int *bitBuf, int *bitCnt, float *CDU, int du_stride, float *fdtbl, int DC, const unsigned short HTDC[256][2], const unsigned short HTAC[256][2]) {
+   const unsigned short EOB[2] = { HTAC[0x00][0], HTAC[0x00][1] };
+   const unsigned short M16zeroes[2] = { HTAC[0xF0][0], HTAC[0xF0][1] };
+   int dataOff, i, j, n, diff, end0pos, x, y;
+   int DU[64];
+
+   // DCT rows
+   for(dataOff=0, n=du_stride*8; dataOff<n; dataOff+=du_stride) {
+      stbiw__jpg_DCT(&CDU[dataOff], &CDU[dataOff+1], &CDU[dataOff+2], &CDU[dataOff+3], &CDU[dataOff+4], &CDU[dataOff+5], &CDU[dataOff+6], &CDU[dataOff+7]);
+   }
+   // DCT columns
+   for(dataOff=0; dataOff<8; ++dataOff) {
+      stbiw__jpg_DCT(&CDU[dataOff], &CDU[dataOff+du_stride], &CDU[dataOff+du_stride*2], &CDU[dataOff+du_stride*3], &CDU[dataOff+du_stride*4],
+                     &CDU[dataOff+du_stride*5], &CDU[dataOff+du_stride*6], &CDU[dataOff+du_stride*7]);
+   }
+   // Quantize/descale/zigzag the coefficients
+   for(y = 0, j=0; y < 8; ++y) {
+      for(x = 0; x < 8; ++x,++j) {
+         float v;
+         i = y*du_stride+x;
+         v = CDU[i]*fdtbl[j];
+         // DU[stbiw__jpg_ZigZag[j]] = (int)(v < 0 ? ceilf(v - 0.5f) : floorf(v + 0.5f));
+         // ceilf() and floorf() are C99, not C89, but I /think/ they're not needed here anyway?
+         DU[stbiw__jpg_ZigZag[j]] = (int)(v < 0 ? v - 0.5f : v + 0.5f);
+      }
+   }
+
+   // Encode DC
+   diff = DU[0] - DC;
+   if (diff == 0) {
+      stbiw__jpg_writeBits(s, bitBuf, bitCnt, HTDC[0]);
+   } else {
+      unsigned short bits[2];
+      stbiw__jpg_calcBits(diff, bits);
+      stbiw__jpg_writeBits(s, bitBuf, bitCnt, HTDC[bits[1]]);
+      stbiw__jpg_writeBits(s, bitBuf, bitCnt, bits);
+   }
+   // Encode ACs
+   end0pos = 63;
+   for(; (end0pos>0)&&(DU[end0pos]==0); --end0pos) {
+   }
+   // end0pos = first element in reverse order !=0
+   if(end0pos == 0) {
+      stbiw__jpg_writeBits(s, bitBuf, bitCnt, EOB);
+      return DU[0];
+   }
+   for(i = 1; i <= end0pos; ++i) {
+      int startpos = i;
+      int nrzeroes;
+      unsigned short bits[2];
+      for (; DU[i]==0 && i<=end0pos; ++i) {
+      }
+      nrzeroes = i-startpos;
+      if ( nrzeroes >= 16 ) {
+         int lng = nrzeroes>>4;
+         int nrmarker;
+         for (nrmarker=1; nrmarker <= lng; ++nrmarker)
+            stbiw__jpg_writeBits(s, bitBuf, bitCnt, M16zeroes);
+         nrzeroes &= 15;
+      }
+      stbiw__jpg_calcBits(DU[i], bits);
+      stbiw__jpg_writeBits(s, bitBuf, bitCnt, HTAC[(nrzeroes<<4)+bits[1]]);
+      stbiw__jpg_writeBits(s, bitBuf, bitCnt, bits);
+   }
+   if(end0pos != 63) {
+      stbiw__jpg_writeBits(s, bitBuf, bitCnt, EOB);
+   }
+   return DU[0];
+}
+
+static int stbi_write_jpg_core(stbi__write_context *s, int width, int height, int comp, const void* data, int quality) {
+   // Constants that don't pollute global namespace
+   static const unsigned char std_dc_luminance_nrcodes[] = {0,0,1,5,1,1,1,1,1,1,0,0,0,0,0,0,0};
+   static const unsigned char std_dc_luminance_values[] = {0,1,2,3,4,5,6,7,8,9,10,11};
+   static const unsigned char std_ac_luminance_nrcodes[] = {0,0,2,1,3,3,2,4,3,5,5,4,4,0,0,1,0x7d};
+   static const unsigned char std_ac_luminance_values[] = {
+      0x01,0x02,0x03,0x00,0x04,0x11,0x05,0x12,0x21,0x31,0x41,0x06,0x13,0x51,0x61,0x07,0x22,0x71,0x14,0x32,0x81,0x91,0xa1,0x08,
+      0x23,0x42,0xb1,0xc1,0x15,0x52,0xd1,0xf0,0x24,0x33,0x62,0x72,0x82,0x09,0x0a,0x16,0x17,0x18,0x19,0x1a,0x25,0x26,0x27,0x28,
+      0x29,0x2a,0x34,0x35,0x36,0x37,0x38,0x39,0x3a,0x43,0x44,0x45,0x46,0x47,0x48,0x49,0x4a,0x53,0x54,0x55,0x56,0x57,0x58,0x59,
+      0x5a,0x63,0x64,0x65,0x66,0x67,0x68,0x69,0x6a,0x73,0x74,0x75,0x76,0x77,0x78,0x79,0x7a,0x83,0x84,0x85,0x86,0x87,0x88,0x89,
+      0x8a,0x92,0x93,0x94,0x95,0x96,0x97,0x98,0x99,0x9a,0xa2,0xa3,0xa4,0xa5,0xa6,0xa7,0xa8,0xa9,0xaa,0xb2,0xb3,0xb4,0xb5,0xb6,
+      0xb7,0xb8,0xb9,0xba,0xc2,0xc3,0xc4,0xc5,0xc6,0xc7,0xc8,0xc9,0xca,0xd2,0xd3,0xd4,0xd5,0xd6,0xd7,0xd8,0xd9,0xda,0xe1,0xe2,
+      0xe3,0xe4,0xe5,0xe6,0xe7,0xe8,0xe9,0xea,0xf1,0xf2,0xf3,0xf4,0xf5,0xf6,0xf7,0xf8,0xf9,0xfa
+   };
+   static const unsigned char std_dc_chrominance_nrcodes[] = {0,0,3,1,1,1,1,1,1,1,1,1,0,0,0,0,0};
+   static const unsigned char std_dc_chrominance_values[] = {0,1,2,3,4,5,6,7,8,9,10,11};
+   static const unsigned char std_ac_chrominance_nrcodes[] = {0,0,2,1,2,4,4,3,4,7,5,4,4,0,1,2,0x77};
+   static const unsigned char std_ac_chrominance_values[] = {
+      0x00,0x01,0x02,0x03,0x11,0x04,0x05,0x21,0x31,0x06,0x12,0x41,0x51,0x07,0x61,0x71,0x13,0x22,0x32,0x81,0x08,0x14,0x42,0x91,
+      0xa1,0xb1,0xc1,0x09,0x23,0x33,0x52,0xf0,0x15,0x62,0x72,0xd1,0x0a,0x16,0x24,0x34,0xe1,0x25,0xf1,0x17,0x18,0x19,0x1a,0x26,
+      0x27,0x28,0x29,0x2a,0x35,0x36,0x37,0x38,0x39,0x3a,0x43,0x44,0x45,0x46,0x47,0x48,0x49,0x4a,0x53,0x54,0x55,0x56,0x57,0x58,
+      0x59,0x5a,0x63,0x64,0x65,0x66,0x67,0x68,0x69,0x6a,0x73,0x74,0x75,0x76,0x77,0x78,0x79,0x7a,0x82,0x83,0x84,0x85,0x86,0x87,
+      0x88,0x89,0x8a,0x92,0x93,0x94,0x95,0x96,0x97,0x98,0x99,0x9a,0xa2,0xa3,0xa4,0xa5,0xa6,0xa7,0xa8,0xa9,0xaa,0xb2,0xb3,0xb4,
+      0xb5,0xb6,0xb7,0xb8,0xb9,0xba,0xc2,0xc3,0xc4,0xc5,0xc6,0xc7,0xc8,0xc9,0xca,0xd2,0xd3,0xd4,0xd5,0xd6,0xd7,0xd8,0xd9,0xda,
+      0xe2,0xe3,0xe4,0xe5,0xe6,0xe7,0xe8,0xe9,0xea,0xf2,0xf3,0xf4,0xf5,0xf6,0xf7,0xf8,0xf9,0xfa
+   };
+   // Huffman tables
+   static const unsigned short YDC_HT[256][2] = { {0,2},{2,3},{3,3},{4,3},{5,3},{6,3},{14,4},{30,5},{62,6},{126,7},{254,8},{510,9}};
+   static const unsigned short UVDC_HT[256][2] = { {0,2},{1,2},{2,2},{6,3},{14,4},{30,5},{62,6},{126,7},{254,8},{510,9},{1022,10},{2046,11}};
+   static const unsigned short YAC_HT[256][2] = {
+      {10,4},{0,2},{1,2},{4,3},{11,4},{26,5},{120,7},{248,8},{1014,10},{65410,16},{65411,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
+      {12,4},{27,5},{121,7},{502,9},{2038,11},{65412,16},{65413,16},{65414,16},{65415,16},{65416,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
+      {28,5},{249,8},{1015,10},{4084,12},{65417,16},{65418,16},{65419,16},{65420,16},{65421,16},{65422,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
+      {58,6},{503,9},{4085,12},{65423,16},{65424,16},{65425,16},{65426,16},{65427,16},{65428,16},{65429,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
+      {59,6},{1016,10},{65430,16},{65431,16},{65432,16},{65433,16},{65434,16},{65435,16},{65436,16},{65437,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
+      {122,7},{2039,11},{65438,16},{65439,16},{65440,16},{65441,16},{65442,16},{65443,16},{65444,16},{65445,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
+      {123,7},{4086,12},{65446,16},{65447,16},{65448,16},{65449,16},{65450,16},{65451,16},{65452,16},{65453,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
+      {250,8},{4087,12},{65454,16},{65455,16},{65456,16},{65457,16},{65458,16},{65459,16},{65460,16},{65461,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
+      {504,9},{32704,15},{65462,16},{65463,16},{65464,16},{65465,16},{65466,16},{65467,16},{65468,16},{65469,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
+      {505,9},{65470,16},{65471,16},{65472,16},{65473,16},{65474,16},{65475,16},{65476,16},{65477,16},{65478,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
+      {506,9},{65479,16},{65480,16},{65481,16},{65482,16},{65483,16},{65484,16},{65485,16},{65486,16},{65487,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
+      {1017,10},{65488,16},{65489,16},{65490,16},{65491,16},{65492,16},{65493,16},{65494,16},{65495,16},{65496,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
+      {1018,10},{65497,16},{65498,16},{65499,16},{65500,16},{65501,16},{65502,16},{65503,16},{65504,16},{65505,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
+      {2040,11},{65506,16},{65507,16},{65508,16},{65509,16},{65510,16},{65511,16},{65512,16},{65513,16},{65514,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
+      {65515,16},{65516,16},{65517,16},{65518,16},{65519,16},{65520,16},{65521,16},{65522,16},{65523,16},{65524,16},{0,0},{0,0},{0,0},{0,0},{0,0},
+      {2041,11},{65525,16},{65526,16},{65527,16},{65528,16},{65529,16},{65530,16},{65531,16},{65532,16},{65533,16},{65534,16},{0,0},{0,0},{0,0},{0,0},{0,0}
+   };
+   static const unsigned short UVAC_HT[256][2] = {
+      {0,2},{1,2},{4,3},{10,4},{24,5},{25,5},{56,6},{120,7},{500,9},{1014,10},{4084,12},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
+      {11,4},{57,6},{246,8},{501,9},{2038,11},{4085,12},{65416,16},{65417,16},{65418,16},{65419,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
+      {26,5},{247,8},{1015,10},{4086,12},{32706,15},{65420,16},{65421,16},{65422,16},{65423,16},{65424,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
+      {27,5},{248,8},{1016,10},{4087,12},{65425,16},{65426,16},{65427,16},{65428,16},{65429,16},{65430,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
+      {58,6},{502,9},{65431,16},{65432,16},{65433,16},{65434,16},{65435,16},{65436,16},{65437,16},{65438,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
+      {59,6},{1017,10},{65439,16},{65440,16},{65441,16},{65442,16},{65443,16},{65444,16},{65445,16},{65446,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
+      {121,7},{2039,11},{65447,16},{65448,16},{65449,16},{65450,16},{65451,16},{65452,16},{65453,16},{65454,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
+      {122,7},{2040,11},{65455,16},{65456,16},{65457,16},{65458,16},{65459,16},{65460,16},{65461,16},{65462,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
+      {249,8},{65463,16},{65464,16},{65465,16},{65466,16},{65467,16},{65468,16},{65469,16},{65470,16},{65471,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
+      {503,9},{65472,16},{65473,16},{65474,16},{65475,16},{65476,16},{65477,16},{65478,16},{65479,16},{65480,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
+      {504,9},{65481,16},{65482,16},{65483,16},{65484,16},{65485,16},{65486,16},{65487,16},{65488,16},{65489,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
+      {505,9},{65490,16},{65491,16},{65492,16},{65493,16},{65494,16},{65495,16},{65496,16},{65497,16},{65498,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
+      {506,9},{65499,16},{65500,16},{65501,16},{65502,16},{65503,16},{65504,16},{65505,16},{65506,16},{65507,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
+      {2041,11},{65508,16},{65509,16},{65510,16},{65511,16},{65512,16},{65513,16},{65514,16},{65515,16},{65516,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
+      {16352,14},{65517,16},{65518,16},{65519,16},{65520,16},{65521,16},{65522,16},{65523,16},{65524,16},{65525,16},{0,0},{0,0},{0,0},{0,0},{0,0},
+      {1018,10},{32707,15},{65526,16},{65527,16},{65528,16},{65529,16},{65530,16},{65531,16},{65532,16},{65533,16},{65534,16},{0,0},{0,0},{0,0},{0,0},{0,0}
+   };
+   static const int YQT[] = {16,11,10,16,24,40,51,61,12,12,14,19,26,58,60,55,14,13,16,24,40,57,69,56,14,17,22,29,51,87,80,62,18,22,
+                             37,56,68,109,103,77,24,35,55,64,81,104,113,92,49,64,78,87,103,121,120,101,72,92,95,98,112,100,103,99};
+   static const int UVQT[] = {17,18,24,47,99,99,99,99,18,21,26,66,99,99,99,99,24,26,56,99,99,99,99,99,47,66,99,99,99,99,99,99,
+                              99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99};
+   static const float aasf[] = { 1.0f * 2.828427125f, 1.387039845f * 2.828427125f, 1.306562965f * 2.828427125f, 1.175875602f * 2.828427125f,
+                                 1.0f * 2.828427125f, 0.785694958f * 2.828427125f, 0.541196100f * 2.828427125f, 0.275899379f * 2.828427125f };
+
+   int row, col, i, k, subsample;
+   float fdtbl_Y[64], fdtbl_UV[64];
+   unsigned char YTable[64], UVTable[64];
+
+   if(!data || !width || !height || comp > 4 || comp < 1) {
+      return 0;
+   }
+
+   quality = quality ? quality : 90;
+   subsample = quality <= 90 ? 1 : 0;
+   quality = quality < 1 ? 1 : quality > 100 ? 100 : quality;
+   quality = quality < 50 ? 5000 / quality : 200 - quality * 2;
+
+   for(i = 0; i < 64; ++i) {
+      int uvti, yti = (YQT[i]*quality+50)/100;
+      YTable[stbiw__jpg_ZigZag[i]] = (unsigned char) (yti < 1 ? 1 : yti > 255 ? 255 : yti);
+      uvti = (UVQT[i]*quality+50)/100;
+      UVTable[stbiw__jpg_ZigZag[i]] = (unsigned char) (uvti < 1 ? 1 : uvti > 255 ? 255 : uvti);
+   }
+
+   for(row = 0, k = 0; row < 8; ++row) {
+      for(col = 0; col < 8; ++col, ++k) {
+         fdtbl_Y[k]  = 1 / (YTable [stbiw__jpg_ZigZag[k]] * aasf[row] * aasf[col]);
+         fdtbl_UV[k] = 1 / (UVTable[stbiw__jpg_ZigZag[k]] * aasf[row] * aasf[col]);
+      }
+   }
+
+   // Write Headers
+   {
+      static const unsigned char head0[] = { 0xFF,0xD8,0xFF,0xE0,0,0x10,'J','F','I','F',0,1,1,0,0,1,0,1,0,0,0xFF,0xDB,0,0x84,0 };
+      static const unsigned char head2[] = { 0xFF,0xDA,0,0xC,3,1,0,2,0x11,3,0x11,0,0x3F,0 };
+      const unsigned char head1[] = { 0xFF,0xC0,0,0x11,8,(unsigned char)(height>>8),STBIW_UCHAR(height),(unsigned char)(width>>8),STBIW_UCHAR(width),
+                                      3,1,(unsigned char)(subsample?0x22:0x11),0,2,0x11,1,3,0x11,1,0xFF,0xC4,0x01,0xA2,0 };
+      s->func(s->context, (void*)head0, sizeof(head0));
+      s->func(s->context, (void*)YTable, sizeof(YTable));
+      stbiw__putc(s, 1);
+      s->func(s->context, UVTable, sizeof(UVTable));
+      s->func(s->context, (void*)head1, sizeof(head1));
+      s->func(s->context, (void*)(std_dc_luminance_nrcodes+1), sizeof(std_dc_luminance_nrcodes)-1);
+      s->func(s->context, (void*)std_dc_luminance_values, sizeof(std_dc_luminance_values));
+      stbiw__putc(s, 0x10); // HTYACinfo
+      s->func(s->context, (void*)(std_ac_luminance_nrcodes+1), sizeof(std_ac_luminance_nrcodes)-1);
+      s->func(s->context, (void*)std_ac_luminance_values, sizeof(std_ac_luminance_values));
+      stbiw__putc(s, 1); // HTUDCinfo
+      s->func(s->context, (void*)(std_dc_chrominance_nrcodes+1), sizeof(std_dc_chrominance_nrcodes)-1);
+      s->func(s->context, (void*)std_dc_chrominance_values, sizeof(std_dc_chrominance_values));
+      stbiw__putc(s, 0x11); // HTUACinfo
+      s->func(s->context, (void*)(std_ac_chrominance_nrcodes+1), sizeof(std_ac_chrominance_nrcodes)-1);
+      s->func(s->context, (void*)std_ac_chrominance_values, sizeof(std_ac_chrominance_values));
+      s->func(s->context, (void*)head2, sizeof(head2));
+   }
+
+   // Encode 8x8 macroblocks
+   {
+      static const unsigned short fillBits[] = {0x7F, 7};
+      int DCY=0, DCU=0, DCV=0;
+      int bitBuf=0, bitCnt=0;
+      // comp == 2 is grey+alpha (alpha is ignored)
+      int ofsG = comp > 2 ? 1 : 0, ofsB = comp > 2 ? 2 : 0;
+      const unsigned char *dataR = (const unsigned char *)data;
+      const unsigned char *dataG = dataR + ofsG;
+      const unsigned char *dataB = dataR + ofsB;
+      int x, y, pos;
+      if(subsample) {
+         for(y = 0; y < height; y += 16) {
+            for(x = 0; x < width; x += 16) {
+               float Y[256], U[256], V[256];
+               for(row = y, pos = 0; row < y+16; ++row) {
+                  // row >= height => use last input row
+                  int clamped_row = (row < height) ? row : height - 1;
+                  int base_p = (stbi__flip_vertically_on_write ? (height-1-clamped_row) : clamped_row)*width*comp;
+                  for(col = x; col < x+16; ++col, ++pos) {
+                     // if col >= width => use pixel from last input column
+                     int p = base_p + ((col < width) ? col : (width-1))*comp;
+                     float r = dataR[p], g = dataG[p], b = dataB[p];
+                     Y[pos]= +0.29900f*r + 0.58700f*g + 0.11400f*b - 128;
+                     U[pos]= -0.16874f*r - 0.33126f*g + 0.50000f*b;
+                     V[pos]= +0.50000f*r - 0.41869f*g - 0.08131f*b;
+                  }
+               }
+               DCY = stbiw__jpg_processDU(s, &bitBuf, &bitCnt, Y+0,   16, fdtbl_Y, DCY, YDC_HT, YAC_HT);
+               DCY = stbiw__jpg_processDU(s, &bitBuf, &bitCnt, Y+8,   16, fdtbl_Y, DCY, YDC_HT, YAC_HT);
+               DCY = stbiw__jpg_processDU(s, &bitBuf, &bitCnt, Y+128, 16, fdtbl_Y, DCY, YDC_HT, YAC_HT);
+               DCY = stbiw__jpg_processDU(s, &bitBuf, &bitCnt, Y+136, 16, fdtbl_Y, DCY, YDC_HT, YAC_HT);
+
+               // subsample U,V
+               {
+                  float subU[64], subV[64];
+                  int yy, xx;
+                  for(yy = 0, pos = 0; yy < 8; ++yy) {
+                     for(xx = 0; xx < 8; ++xx, ++pos) {
+                        int j = yy*32+xx*2;
+                        subU[pos] = (U[j+0] + U[j+1] + U[j+16] + U[j+17]) * 0.25f;
+                        subV[pos] = (V[j+0] + V[j+1] + V[j+16] + V[j+17]) * 0.25f;
+                     }
+                  }
+                  DCU = stbiw__jpg_processDU(s, &bitBuf, &bitCnt, subU, 8, fdtbl_UV, DCU, UVDC_HT, UVAC_HT);
+                  DCV = stbiw__jpg_processDU(s, &bitBuf, &bitCnt, subV, 8, fdtbl_UV, DCV, UVDC_HT, UVAC_HT);
+               }
+            }
+         }
+      } else {
+         for(y = 0; y < height; y += 8) {
+            for(x = 0; x < width; x += 8) {
+               float Y[64], U[64], V[64];
+               for(row = y, pos = 0; row < y+8; ++row) {
+                  // row >= height => use last input row
+                  int clamped_row = (row < height) ? row : height - 1;
+                  int base_p = (stbi__flip_vertically_on_write ? (height-1-clamped_row) : clamped_row)*width*comp;
+                  for(col = x; col < x+8; ++col, ++pos) {
+                     // if col >= width => use pixel from last input column
+                     int p = base_p + ((col < width) ? col : (width-1))*comp;
+                     float r = dataR[p], g = dataG[p], b = dataB[p];
+                     Y[pos]= +0.29900f*r + 0.58700f*g + 0.11400f*b - 128;
+                     U[pos]= -0.16874f*r - 0.33126f*g + 0.50000f*b;
+                     V[pos]= +0.50000f*r - 0.41869f*g - 0.08131f*b;
+                  }
+               }
+
+               DCY = stbiw__jpg_processDU(s, &bitBuf, &bitCnt, Y, 8, fdtbl_Y,  DCY, YDC_HT, YAC_HT);
+               DCU = stbiw__jpg_processDU(s, &bitBuf, &bitCnt, U, 8, fdtbl_UV, DCU, UVDC_HT, UVAC_HT);
+               DCV = stbiw__jpg_processDU(s, &bitBuf, &bitCnt, V, 8, fdtbl_UV, DCV, UVDC_HT, UVAC_HT);
+            }
+         }
+      }
+
+      // Do the bit alignment of the EOI marker
+      stbiw__jpg_writeBits(s, &bitBuf, &bitCnt, fillBits);
+   }
+
+   // EOI
+   stbiw__putc(s, 0xFF);
+   stbiw__putc(s, 0xD9);
+
+   return 1;
+}
+
+STBIWDEF int stbi_write_jpg_to_func(stbi_write_func *func, void *context, int x, int y, int comp, const void *data, int quality)
+{
+   stbi__write_context s = { 0 };
+   stbi__start_write_callbacks(&s, func, context);
+   return stbi_write_jpg_core(&s, x, y, comp, (void *) data, quality);
+}
+
+
+#ifndef STBI_WRITE_NO_STDIO
+STBIWDEF int stbi_write_jpg(char const *filename, int x, int y, int comp, const void *data, int quality)
+{
+   stbi__write_context s = { 0 };
+   if (stbi__start_write_file(&s,filename)) {
+      int r = stbi_write_jpg_core(&s, x, y, comp, data, quality);
+      stbi__end_write_file(&s);
+      return r;
+   } else
+      return 0;
+}
+#endif
+
+#endif // STB_IMAGE_WRITE_IMPLEMENTATION
+
+/* Revision history
+      1.16  (2021-07-11)
+             make Deflate code emit uncompressed blocks when it would otherwise expand
+             support writing BMPs with alpha channel
+      1.15  (2020-07-13) unknown
+      1.14  (2020-02-02) updated JPEG writer to downsample chroma channels
+      1.13
+      1.12
+      1.11  (2019-08-11)
+
+      1.10  (2019-02-07)
+             support utf8 filenames in Windows; fix warnings and platform ifdefs
+      1.09  (2018-02-11)
+             fix typo in zlib quality API, improve STB_I_W_STATIC in C++
+      1.08  (2018-01-29)
+             add stbi__flip_vertically_on_write, external zlib, zlib quality, choose PNG filter
+      1.07  (2017-07-24)
+             doc fix
+      1.06 (2017-07-23)
+             writing JPEG (using Jon Olick's code)
+      1.05   ???
+      1.04 (2017-03-03)
+             monochrome BMP expansion
+      1.03   ???
+      1.02 (2016-04-02)
+             avoid allocating large structures on the stack
+      1.01 (2016-01-16)
+             STBIW_REALLOC_SIZED: support allocators with no realloc support
+             avoid race-condition in crc initialization
+             minor compile issues
+      1.00 (2015-09-14)
+             installable file IO function
+      0.99 (2015-09-13)
+             warning fixes; TGA rle support
+      0.98 (2015-04-08)
+             added STBIW_MALLOC, STBIW_ASSERT etc
+      0.97 (2015-01-18)
+             fixed HDR asserts, rewrote HDR rle logic
+      0.96 (2015-01-17)
+             add HDR output
+             fix monochrome BMP
+      0.95 (2014-08-17)
+             add monochrome TGA output
+      0.94 (2014-05-31)
+             rename private functions to avoid conflicts with stb_image.h
+      0.93 (2014-05-27)
+             warning fixes
+      0.92 (2010-08-01)
+             casts to unsigned char to fix warnings
+      0.91 (2010-07-17)
+             first public release
+      0.90   first internal release
+*/
+
+/*
+------------------------------------------------------------------------------
+This software is available under 2 licenses -- choose whichever you prefer.
+------------------------------------------------------------------------------
+ALTERNATIVE A - MIT License
+Copyright (c) 2017 Sean Barrett
+Permission is hereby granted, free of charge, to any person obtaining a copy of
+this software and associated documentation files (the "Software"), to deal in
+the Software without restriction, including without limitation the rights to
+use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies
+of the Software, and to permit persons to whom the Software is furnished to do
+so, subject to the following conditions:
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.
+------------------------------------------------------------------------------
+ALTERNATIVE B - Public Domain (www.unlicense.org)
+This is free and unencumbered software released into the public domain.
+Anyone is free to copy, modify, publish, use, compile, sell, or distribute this
+software, either in source code form or as a compiled binary, for any purpose,
+commercial or non-commercial, and by any means.
+In jurisdictions that recognize copyright laws, the author or authors of this
+software dedicate any and all copyright interest in the software to the public
+domain. We make this dedication for the benefit of the public at large and to
+the detriment of our heirs and successors. We intend this dedication to be an
+overt act of relinquishment in perpetuity of all present and future rights to
+this software under copyright law.
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION
+WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
+------------------------------------------------------------------------------
+*/

stable-diffusion.cpp/ggml/.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

stable-diffusion.cpp/ggml/.github/workflows/ci.yml

@@ -0,0 +1,137 @@
+name: CI
+
+on:
+  push:
+    branches: [ master ]
+  pull_request:
+    branches: [ master ]
+
+jobs:
+  test-ubuntu-opencl:
+    runs-on: ubuntu-latest
+    env:
+      GGML_NLOOP: 3
+      GGML_NITER: 1
+      GGML_N_THREADS: 2
+
+    steps:
+    - uses: actions/checkout@v3
+
+    - name: Dependencies
+      run: |
+        wget -O- https://apt.repos.intel.com/intel-gpg-keys/GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB | gpg --dearmor | sudo tee /usr/share/keyrings/oneapi-archive-keyring.gpg > /dev/null
+        echo "deb [signed-by=/usr/share/keyrings/oneapi-archive-keyring.gpg] https://apt.repos.intel.com/oneapi all main" | sudo tee /etc/apt/sources.list.d/oneAPI.list  
+        sudo apt-get update
+        sudo apt-get install -y --no-install-recommends llvm intel-oneapi-runtime-opencl intel-oneapi-runtime-compilers libclblast-dev
+    - name: Create Build Environment
+      run: mkdir build
+
+    - name: Configure CMake
+      working-directory: ./build
+      run: cmake -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ -DGGML_TEST_COVERAGE=ON -DGGML_CLBLAST=ON ..
+
+    - name: Build
+      working-directory: ./build
+      run: make
+
+    - name: Test
+      working-directory: ./build
+      run: ctest --verbose --timeout 900
+
+    - name: Test Coverage
+      working-directory: ./build
+      run: |
+        llvm-profdata merge -sparse tests/*.profraw -o ggml.profdata
+        llvm-cov      report ./bin/test-grad0 -instr-profile=ggml.profdata
+        llvm-cov      report ./bin/test-opt   -instr-profile=ggml.profdata
+  test-macos-metal:
+    runs-on: macos-13
+    env:
+      GGML_NLOOP: 3
+      GGML_NITER: 1
+      GGML_N_THREADS: 2
+
+    steps:
+    - uses: actions/checkout@v3
+
+    - name: Create Build Environment
+      run: mkdir build
+
+    - name: Configure CMake
+      working-directory: ./build
+      run: cmake -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ -DGGML_TEST_COVERAGE=ON -DGGML_METAL=ON ..
+
+    - name: Build
+      working-directory: ./build
+      run: make
+
+    - name: Test
+      working-directory: ./build
+      run: ctest --verbose --timeout 900
+
+    - name: Test Coverage
+      working-directory: ./build
+      run: |
+        xcrun llvm-profdata merge -sparse tests/*.profraw -o ggml.profdata
+        xcrun llvm-cov      report ./bin/test-grad0 -instr-profile=ggml.profdata
+        xcrun llvm-cov      report ./bin/test-opt   -instr-profile=ggml.profdata
+
+  build:
+
+    strategy:
+      matrix:
+        os: [ubuntu-latest, macos-latest]
+
+    runs-on: ${{ matrix.os }}
+
+    env:
+      GGML_NLOOP: 3
+      GGML_NITER: 1
+
+    steps:
+    - uses: actions/checkout@v3
+
+    - name: Dependencies for Ubuntu
+      if: matrix.os == 'ubuntu-latest'
+      run: |
+        sudo apt-get update
+        sudo apt-get install llvm
+
+    - name: Set GGML_N_THREADS for Ubuntu
+      run: echo "GGML_N_THREADS=2" >> $GITHUB_ENV
+      if: matrix.os == 'ubuntu-latest'
+
+    - name: Set GGML_N_THREADS for MacOS
+      run: echo "GGML_N_THREADS=2" >> $GITHUB_ENV
+      if: matrix.os == 'macos-latest'
+
+    - name: Create Build Environment
+      run: mkdir build
+
+    - name: Configure CMake
+      working-directory: ./build
+      run: cmake -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ -DGGML_TEST_COVERAGE=ON ..
+
+    - name: Build
+      working-directory: ./build
+      run: make
+
+    - name: Test
+      working-directory: ./build
+      run: ctest --verbose --timeout 900
+
+    - name: Test Coverage for Ubuntu
+      if: matrix.os == 'ubuntu-latest'
+      working-directory: ./build
+      run: |
+        llvm-profdata merge -sparse tests/*.profraw -o ggml.profdata
+        llvm-cov      report ./bin/test-grad0 -instr-profile=ggml.profdata
+        llvm-cov      report ./bin/test-opt   -instr-profile=ggml.profdata
+
+    - name: Test Coverage for MacOS
+      if: matrix.os == 'macos-latest'
+      working-directory: ./build
+      run: |
+        xcrun llvm-profdata merge -sparse tests/*.profraw -o ggml.profdata
+        xcrun llvm-cov      report ./bin/test-grad0 -instr-profile=ggml.profdata
+        xcrun llvm-cov      report ./bin/test-opt   -instr-profile=ggml.profdata

stable-diffusion.cpp/ggml/.gitignore

@@ -0,0 +1,37 @@
+build/
+build-debug/
+build-release/
+build-sanitize-addr/
+build-sanitize-thread/
+build-cov/
+build-ci-debug/
+build-ci-release/
+build-cublas/
+out/
+tmp/
+models/
+models-mnt
+
+compile_commands.json
+CMakeSettings.json
+.vs/
+.vscode/
+.clangd
+
+.exrc
+.cache
+.DS_Store
+.stablelm
+.gpt-2
+
+src/arm_neon.h
+tests/arm_neon.h
+
+zig-out/
+zig-cache/
+
+*.dot
+
+*.sw?
+
+__pycache__/

stable-diffusion.cpp/ggml/CMakeLists.txt

@@ -0,0 +1,197 @@
+cmake_minimum_required (VERSION 3.3)
+project(ggml VERSION 0.1.0)
+
+set(CMAKE_EXPORT_COMPILE_COMMANDS "on")
+set(CMAKE_RUNTIME_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/bin)
+set(CMAKE_INSTALL_RPATH "${CMAKE_INSTALL_PREFIX}/lib")
+
+if(CMAKE_SOURCE_DIR STREQUAL CMAKE_CURRENT_SOURCE_DIR)
+    set(GGML_STANDALONE ON)
+    include(cmake/GitVars.cmake)
+    include(cmake/BuildTypes.cmake)
+else()
+    set(GGML_STANDALONE OFF)
+endif()
+
+if (EMSCRIPTEN)
+    set(BUILD_SHARED_LIBS_DEFAULT OFF)
+else()
+    if (MINGW)
+        set(BUILD_SHARED_LIBS_DEFAULT OFF)
+    else()
+        set(BUILD_SHARED_LIBS_DEFAULT ON)
+    endif()
+endif()
+
+# options
+
+option(BUILD_SHARED_LIBS            "ggml: build shared libs" ${BUILD_SHARED_LIBS_DEFAULT})
+
+option(GGML_ALL_WARNINGS            "ggml: enable all compiler warnings"                   ON)
+option(GGML_ALL_WARNINGS_3RD_PARTY  "ggml: enable all compiler warnings in 3rd party libs" OFF)
+
+option(GGML_SANITIZE_THREAD         "ggml: enable thread sanitizer"    OFF)
+option(GGML_SANITIZE_ADDRESS        "ggml: enable address sanitizer"   OFF)
+option(GGML_SANITIZE_UNDEFINED      "ggml: enable undefined sanitizer" OFF)
+
+option(GGML_BUILD_TESTS             "ggml: build tests"    ${GGML_STANDALONE})
+option(GGML_BUILD_EXAMPLES          "ggml: build examples" ${GGML_STANDALONE})
+
+option(GGML_TEST_COVERAGE           "ggml: enable test coverage" OFF)
+
+option(GGML_PERF                    "ggml: enable perf timings"          OFF)
+option(GGML_NO_ACCELERATE           "ggml: disable Accelerate framework" OFF)
+option(GGML_OPENBLAS                "ggml: use OpenBLAS"                 OFF)
+option(GGML_CLBLAST                 "ggml: use clBLAST"                  OFF)
+option(GGML_CUBLAS                  "ggml: use cuBLAS"                   OFF)
+option(GGML_METAL                   "ggml: use Metal"                    OFF)
+
+# sanitizers
+
+if (GGML_SANITIZE_THREAD)
+    set(CMAKE_C_FLAGS   "${CMAKE_C_FLAGS}   -fsanitize=thread")
+    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fsanitize=thread")
+endif()
+
+if (GGML_SANITIZE_ADDRESS)
+    set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS}     -fsanitize=address -fno-omit-frame-pointer")
+    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fsanitize=address -fno-omit-frame-pointer")
+endif()
+
+if (GGML_SANITIZE_UNDEFINED)
+    set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS}     -fsanitize=undefined")
+    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fsanitize=undefined")
+endif()
+
+# instruction set specific
+option(GGML_AVX                     "ggml: enable AVX"                                     ON)
+option(GGML_AVX2                    "ggml: enable AVX2"                                    ON)
+option(GGML_AVX512                  "ggml: enable AVX512"                                  OFF)
+option(GGML_AVX512_VBMI             "ggml: enable AVX512-VBMI"                             OFF)
+option(GGML_AVX512_VNNI             "ggml: enable AVX512-VNNI"                             OFF)
+option(GGML_FMA                     "ggml: enable FMA"                                     ON)
+# in MSVC F16C is implied with AVX2/AVX512
+if (NOT MSVC)
+    option(GGML_F16C                "ggml: enable F16C"                                    ON)
+endif()
+
+#set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -ffast-math")
+#set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -march=native")
+#set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -mcpu=native")
+
+# warning flags
+
+if (GGML_ALL_WARNINGS)
+    if (NOT MSVC)
+        set(c_flags   -Wall -Wpedantic -Wformat=2 -Wno-unused -Wstrict-prototypes)
+        set(cxx_flags -Wall -Wpedantic -Wformat=2)
+    else()
+        # todo : windows
+    endif()
+
+    add_compile_options(
+        "$<$<COMPILE_LANGUAGE:C>:${c_flags}>"
+        "$<$<COMPILE_LANGUAGE:CXX>:${cxx_flags}>"
+    )
+endif()
+
+if (NOT MSVC)
+    add_compile_options(
+        "$<$<COMPILE_LANGUAGE:C>:-Werror=vla>"
+        "$<$<COMPILE_LANGUAGE:CXX>:-Werror=vla>"
+        "$<$<COMPILE_LANGUAGE:CUDA>:-Xcompiler;-Werror=vla>"
+    )
+endif()
+
+#
+# POSIX conformance
+#
+
+# clock_gettime came in POSIX.1b (1993)
+# CLOCK_MONOTONIC came in POSIX.1-2001 / SUSv3 as optional
+# posix_memalign came in POSIX.1-2001 / SUSv3
+# M_PI is an XSI extension since POSIX.1-2001 / SUSv3, came in XPG1 (1985)
+add_compile_definitions(_XOPEN_SOURCE=600)
+
+# Somehow in OpenBSD whenever POSIX conformance is specified
+# some string functions rely on locale_t availability,
+# which was introduced in POSIX.1-2008, forcing us to go higher
+if (CMAKE_SYSTEM_NAME MATCHES "OpenBSD")
+    remove_definitions(-D_XOPEN_SOURCE=600)
+    add_compile_definitions(_XOPEN_SOURCE=700)
+endif()
+
+# Data types, macros and functions related to controlling CPU affinity
+# are available on Linux through GNU extensions in libc
+if (CMAKE_SYSTEM_NAME MATCHES "Linux")
+    add_compile_definitions(_GNU_SOURCE)
+endif()
+
+# RLIMIT_MEMLOCK came in BSD, is not specified in POSIX.1,
+# and on macOS its availability depends on enabling Darwin extensions
+# similarly on DragonFly, enabling BSD extensions is necessary
+if (CMAKE_SYSTEM_NAME MATCHES "Darwin")
+    add_compile_definitions(_DARWIN_C_SOURCE)
+endif()
+if (CMAKE_SYSTEM_NAME MATCHES "DragonFly")
+    add_compile_definitions(_DARWIN_C_SOURCE)
+endif()
+
+# alloca is a non-standard interface that is not visible on BSDs when
+# POSIX conformance is specified, but not all of them provide a clean way
+# to enable it in such cases
+if (CMAKE_SYSTEM_NAME MATCHES "FreeBSD")
+    add_compile_definitions(__BSD_VISIBLE)
+endif()
+if (CMAKE_SYSTEM_NAME MATCHES "NetBSD")
+    add_compile_definitions(_NETBSD_SOURCE)
+endif()
+if (CMAKE_SYSTEM_NAME MATCHES "OpenBSD")
+    add_compile_definitions(_BSD_SOURCE)
+endif()
+
+if (WHISPER_PERF)
+    set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DGGML_PERF)
+endif()
+
+# dependencies
+
+set(CMAKE_C_STANDARD   11)
+set(CMAKE_CXX_STANDARD 11)
+
+find_package(Threads REQUIRED)
+
+# main
+
+if (NOT CMAKE_BUILD_TYPE AND NOT CMAKE_CONFIGURATION_TYPES)
+    set(CMAKE_BUILD_TYPE Release CACHE STRING "Build type" FORCE)
+    set_property(CACHE CMAKE_BUILD_TYPE PROPERTY STRINGS "Debug" "Release" "RelWithDebInfo")
+endif ()
+
+if (GGML_BUILD_TESTS)
+    if (GGML_TEST_COVERAGE)
+        if (CMAKE_C_COMPILER_ID MATCHES "Clang")
+            set(CMAKE_C_FLAGS   "${CMAKE_C_FLAGS}   -fprofile-instr-generate -fcoverage-mapping")
+            set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fprofile-instr-generate -fcoverage-mapping")
+        else()
+            message(WARNING "Test coverage is only supported for Clang")
+        endif()
+    endif()
+endif()
+
+add_subdirectory(src)
+
+if (GGML_BUILD_TESTS)
+    enable_testing()
+    add_subdirectory(tests)
+endif ()
+
+if (GGML_BUILD_EXAMPLES)
+    add_subdirectory(examples)
+endif ()
+
+configure_file(${CMAKE_CURRENT_SOURCE_DIR}/ggml.pc.in
+               ${CMAKE_CURRENT_BINARY_DIR}/ggml.pc
+               @ONLY)
+install(FILES ${CMAKE_CURRENT_BINARY_DIR}/ggml.pc
+        DESTINATION share/pkgconfig)

stable-diffusion.cpp/ggml/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2022 Georgi Gerganov
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

stable-diffusion.cpp/ggml/README.md

@@ -0,0 +1,140 @@
+# ggml
+
+[Roadmap](https://github.com/users/ggerganov/projects/7) / [Manifesto](https://github.com/ggerganov/llama.cpp/discussions/205)
+
+Tensor library for machine learning
+
+***Note that this project is under active development. \
+Some of the development is currently happening in the [llama.cpp](https://github.com/ggerganov/llama.cpp) and [whisper.cpp](https://github.com/ggerganov/whisper.cpp) repos***
+
+## Features
+
+- Written in C
+- 16-bit float support
+- Integer quantization support (4-bit, 5-bit, 8-bit, etc.)
+- Automatic differentiation
+- ADAM and L-BFGS optimizers
+- Optimized for Apple Silicon
+- On x86 architectures utilizes AVX / AVX2 intrinsics
+- On ppc64 architectures utilizes VSX intrinsics
+- No third-party dependencies
+- Zero memory allocations during runtime
+
+## Updates
+
+- [X] Example of GPT-2 inference [examples/gpt-2](https://github.com/ggerganov/ggml/tree/master/examples/gpt-2)
+- [X] Example of GPT-J inference [examples/gpt-j](https://github.com/ggerganov/ggml/tree/master/examples/gpt-j)
+- [X] Example of Whisper inference [examples/whisper](https://github.com/ggerganov/ggml/tree/master/examples/whisper)
+- [X] Support 4-bit integer quantization https://github.com/ggerganov/ggml/pull/27
+- [X] Example of Cerebras-GPT inference [examples/gpt-2](https://github.com/ggerganov/ggml/tree/master/examples/gpt-2)
+- [ ] Example of FLAN-T5 inference https://github.com/ggerganov/ggml/pull/12
+- [X] Example of LLaMA inference [ggerganov/llama.cpp](https://github.com/ggerganov/llama.cpp)
+- [X] Example of LLaMA training [ggerganov/llama.cpp/examples/baby-llama](https://github.com/ggerganov/llama.cpp/tree/master/examples/baby-llama)
+- [X] Example of Falcon inference [cmp-nct/ggllm.cpp](https://github.com/cmp-nct/ggllm.cpp)
+- [X] Example of BLOOM inference [NouamaneTazi/bloomz.cpp](https://github.com/NouamaneTazi/bloomz.cpp)
+- [X] Example of RWKV inference [saharNooby/rwkv.cpp](https://github.com/saharNooby/rwkv.cpp)
+- [X] Example of SAM inference [examples/sam](https://github.com/ggerganov/ggml/tree/master/examples/sam)
+- [X] Idea for GPU support: https://github.com/ggerganov/llama.cpp/discussions/915
+- [X] Example of StableLM (GPT-NeoX) inference [examples/gpt-neox](https://github.com/ggerganov/ggml/tree/master/examples/gpt-neox)
+- [X] Example of BERT inference [skeskinen/bert.cpp](https://github.com/skeskinen/bert.cpp)
+- [X] Example of 💫 StarCoder inference [examples/starcoder](https://github.com/ggerganov/ggml/tree/master/examples/starcoder)
+- [X] Example of MPT inference [examples/mpt](https://github.com/ggerganov/ggml/tree/master/examples/mpt)
+- [X] Example of Replit inference [examples/replit](https://github.com/ggerganov/ggml/tree/master/examples/replit)
+- [X] Example of BioGPT inference [PABannier/biogpt.cpp](https://github.com/PABannier/biogpt.cpp)
+- [X] Example of Encodec inference [PABannier/encodec.cpp](https://github.com/PABannier/encodec.cpp)
+- [X] Example of CLIP inference [monatis/clip.cpp](https://github.com/monatis/clip.cpp)
+- [X] Example of MiniGPT4 inference [Maknee/minigpt4.cpp](https://github.com/Maknee/minigpt4.cpp)
+- [X] Example of ChatGLM inference [li-plus/chatglm.cpp](https://github.com/li-plus/chatglm.cpp)
+- [X] Example of Stable Diffusion inference [leejet/stable-diffusion.cpp](https://github.com/leejet/stable-diffusion.cpp)
+- [X] Example of Qwen inference [QwenLM/qwen.cpp](https://github.com/QwenLM/qwen.cpp)
+
+## Whisper inference (example)
+
+With ggml you can efficiently run [Whisper](examples/whisper) inference on the CPU.
+
+Memory requirements:
+
+| Model  | Disk   | Mem     |
+| ---    | ---    | ---     |
+| tiny   |  75 MB | ~280 MB |
+| base   | 142 MB | ~430 MB |
+| small  | 466 MB | ~1.0 GB |
+| medium | 1.5 GB | ~2.6 GB |
+| large  | 2.9 GB | ~4.7 GB |
+
+## GPT inference (example)
+
+With ggml you can efficiently run [GPT-2](examples/gpt-2) and [GPT-J](examples/gpt-j) inference on the CPU.
+
+Here is how to run the example programs:
+
+```bash
+# Build ggml + examples
+git clone https://github.com/ggerganov/ggml
+cd ggml
+mkdir build && cd build
+cmake ..
+make -j4 gpt-2 gpt-j
+
+# Run the GPT-2 small 117M model
+../examples/gpt-2/download-ggml-model.sh 117M
+./bin/gpt-2 -m models/gpt-2-117M/ggml-model.bin -p "This is an example"
+
+# Run the GPT-J 6B model (requires 12GB disk space and 16GB CPU RAM)
+../examples/gpt-j/download-ggml-model.sh 6B
+./bin/gpt-j -m models/gpt-j-6B/ggml-model.bin -p "This is an example"
+
+# Install Python dependencies
+python3 -m pip install -r ../requirements.txt
+
+# Run the Cerebras-GPT 111M model
+# Download from: https://huggingface.co/cerebras
+python3 ../examples/gpt-2/convert-cerebras-to-ggml.py /path/to/Cerebras-GPT-111M/
+./bin/gpt-2 -m /path/to/Cerebras-GPT-111M/ggml-model-f16.bin -p "This is an example"
+```
+
+The inference speeds that I get for the different models on my 32GB MacBook M1 Pro are as follows:
+
+| Model | Size  | Time / Token |
+| ---   | ---   | ---    |
+| GPT-2 |  117M |   5 ms |
+| GPT-2 |  345M |  12 ms |
+| GPT-2 |  774M |  23 ms |
+| GPT-2 | 1558M |  42 ms |
+| ---   | ---   | ---    |
+| GPT-J |    6B | 125 ms |
+
+For more information, checkout the corresponding programs in the [examples](examples) folder.
+
+## Using Metal (only with GPT-2)
+
+For GPT-2 models, offloading to GPU is possible. Note that it will not improve inference performances but will reduce power consumption and free up the CPU for other tasks.
+
+To enable GPU offloading on MacOS:
+
+```bash
+cmake -DGGML_METAL=ON -DBUILD_SHARED_LIBS=Off ..
+
+# add -ngl 1
+./bin/gpt-2 -t 4 -ngl 100 -m models/gpt-2-117M/ggml-model.bin -p "This is an example"
+```
+
+## Using cuBLAS
+
+```bash
+# fix the path to point to your CUDA compiler
+cmake -DGGML_CUBLAS=ON -DCMAKE_CUDA_COMPILER=/usr/local/cuda-12.1/bin/nvcc ..
+```
+
+## Using clBLAST
+
+```bash
+cmake -DGGML_CLBLAST=ON ..
+```
+
+## Resources
+
+- [GGML - Large Language Models for Everyone](https://github.com/rustformers/llm/blob/main/crates/ggml/README.md): a description of the GGML format provided by the maintainers of the `llm` Rust crate, which provides Rust bindings for GGML
+- [marella/ctransformers](https://github.com/marella/ctransformers): Python bindings for GGML models.
+- [go-skynet/go-ggml-transformers.cpp](https://github.com/go-skynet/go-ggml-transformers.cpp): Golang bindings for GGML models
+- [smspillaz/ggml-gobject](https://github.com/smspillaz/ggml-gobject): GObject-introspectable wrapper for use of GGML on the GNOME platform.

stable-diffusion.cpp/ggml/build.zig

@@ -0,0 +1,158 @@
+const std = @import("std");
+const builtin = @import("builtin");
+
+// Zig Version: 0.11.0
+// Zig Build Command: zig build
+// Zig Run Command: zig build -h
+//     zig build run_dolly-v2
+//     zig build run_gpt-2
+//     zig build run_gpt-j
+//     zig build run_gpt-neox
+//     zig build run_mnist
+//     zig build run_mpt
+//     zig build run_replit
+//     zig build run_starcoder
+//     zig build run_test-grad0
+//     zig build run_test-mul-mat0
+//     zig build run_test-mul-mat2
+//     zig build run_test-opt
+//     zig build run_test-vec1
+//     zig build run_test0
+//     zig build run_test1
+//     zig build run_test2
+//     zig build run_test3
+//     zig build run_zig_test0
+//     zig build run_zig_test1
+//     zig build run_zig_test2
+//     zig build run_zig_test3
+pub fn build(b: *std.build.Builder) void {
+    const target = b.standardTargetOptions(.{});
+    const optimize = b.standardOptimizeOption(.{});
+    const lib = b.addStaticLibrary(.{
+        .name = "ggml",
+        .target = target,
+        .optimize = optimize,
+    });
+    lib.addIncludePath(.{ .path = "./include" });
+    lib.addIncludePath(.{ .path = "./include/ggml" });
+    lib.addCSourceFiles(&.{
+        "src/ggml.c",
+    }, &.{"-std=c11"});
+    lib.linkLibC();
+    lib.linkLibCpp();
+    b.installArtifact(lib);
+
+    // examples
+    const examples = .{
+        "dolly-v2",
+        "gpt-2",
+        "gpt-j",
+        "gpt-neox",
+        "mnist",
+        "mpt",
+        "replit",
+        "starcoder",
+        // "whisper",
+    };
+    inline for (examples) |name| {
+        const exe = b.addExecutable(.{
+            .name = name,
+            .target = target,
+            .optimize = optimize,
+        });
+        exe.addIncludePath(.{ .path = "./include" });
+        exe.addIncludePath(.{ .path = "./include/ggml" });
+        exe.addIncludePath(.{ .path = "./examples" });
+        // exe.addIncludePath("./examples/whisper");
+        exe.addCSourceFiles(&.{
+            std.fmt.comptimePrint("examples/{s}/main.cpp", .{name}),
+            "examples/common.cpp",
+            "examples/common-ggml.cpp",
+            // "examples/whisper/whisper.cpp",
+        }, &.{"-std=c++11"});
+        exe.linkLibrary(lib);
+        b.installArtifact(exe);
+        const run_cmd = b.addRunArtifact(exe);
+        run_cmd.step.dependOn(b.getInstallStep());
+        if (b.args) |args| run_cmd.addArgs(args);
+        const run_step = b.step("run_" ++ name, "Run examples");
+        run_step.dependOn(&run_cmd.step);
+    }
+
+    // tests
+    const tests = if (builtin.target.cpu.arch == .x86_64) .{
+        // "test-blas0",
+        // "test-grad0",
+        "test-mul-mat0",
+        // "test-mul-mat1",
+        "test-mul-mat2",
+        // "test-opt",
+        // "test-svd0",
+        // "test-vec0",
+        "test-vec1",
+        // "test-vec2",
+        "test0",
+        "test1",
+        "test2",
+        "test3",
+    } else .{
+        // "test-blas0",
+        // "test-grad0",
+        "test-mul-mat0",
+        // "test-mul-mat1",
+        "test-mul-mat2",
+        // "test-opt",
+        // "test-svd0",
+        // "test-vec0",
+        // "test-vec1",
+        // "test-vec2",
+        "test0",
+        "test1",
+        "test2",
+        "test3",
+    };
+    inline for (tests) |name| {
+        const exe = b.addExecutable(.{
+            .name = name,
+            .target = target,
+            .optimize = optimize,
+        });
+        exe.addIncludePath(.{ .path = "./include" });
+        exe.addIncludePath(.{ .path = "./include/ggml" });
+        exe.addCSourceFiles(&.{
+            std.fmt.comptimePrint("tests/{s}.c", .{name}),
+        }, &.{"-std=c11"});
+        exe.linkLibrary(lib);
+        b.installArtifact(exe);
+        const run_cmd = b.addRunArtifact(exe);
+        run_cmd.step.dependOn(b.getInstallStep());
+        if (b.args) |args| run_cmd.addArgs(args);
+        const run_step = b.step("run_" ++ name, "Run tests");
+        run_step.dependOn(&run_cmd.step);
+    }
+
+    // zig_tests
+    const zig_tests = .{
+        "test0",
+        "test1",
+        "test2",
+        "test3",
+    };
+    inline for (zig_tests) |name| {
+        const exe = b.addExecutable(.{
+            .name = name,
+            .root_source_file = .{ .path = std.fmt.comptimePrint("tests/{s}.zig", .{name}) },
+            .target = target,
+            .optimize = optimize,
+        });
+        exe.addIncludePath(.{ .path = "./include" });
+        exe.addIncludePath(.{ .path = "./include/ggml" });
+        exe.linkLibrary(lib);
+        b.installArtifact(exe);
+        const run_cmd = b.addRunArtifact(exe);
+        run_cmd.step.dependOn(b.getInstallStep());
+        if (b.args) |args| run_cmd.addArgs(args);
+        const run_step = b.step("run_zig_" ++ name, "Run zig_tests");
+        run_step.dependOn(&run_cmd.step);
+    }
+}

stable-diffusion.cpp/ggml/ci/run.sh

@@ -0,0 +1,334 @@
+#/bin/bash
+#
+# sample usage:
+#
+# mkdir tmp
+#
+# # CPU-only build
+# bash ./ci/run.sh ./tmp/results ./tmp/mnt
+#
+# # with CUDA support
+# GG_BUILD_CUDA=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt
+#
+
+if [ -z "$2" ]; then
+    echo "usage: $0 <output-dir> <mnt-dir>"
+    exit 1
+fi
+
+mkdir -p "$1"
+mkdir -p "$2"
+
+OUT=$(realpath "$1")
+MNT=$(realpath "$2")
+
+rm -v $OUT/*.log
+rm -v $OUT/*.exit
+rm -v $OUT/*.md
+
+sd=`dirname $0`
+cd $sd/../
+SRC=`pwd`
+
+## helpers
+
+# download a file if it does not exist or if it is outdated
+function gg_wget {
+    local out=$1
+    local url=$2
+
+    local cwd=`pwd`
+
+    mkdir -p $out
+    cd $out
+
+    # should not re-download if file is the same
+    wget -nv -N $url
+
+    cd $cwd
+}
+
+function gg_printf {
+    printf -- "$@" >> $OUT/README.md
+}
+
+function gg_run {
+    ci=$1
+
+    set -o pipefail
+    set -x
+
+    gg_run_$ci | tee $OUT/$ci.log
+    cur=$?
+    echo "$cur" > $OUT/$ci.exit
+
+    set +x
+    set +o pipefail
+
+    gg_sum_$ci
+
+    ret=$((ret | cur))
+}
+
+## ci
+
+# ctest_debug
+
+function gg_run_ctest_debug {
+    cd ${SRC}
+
+    rm -rf build-ci-debug && mkdir build-ci-debug && cd build-ci-debug
+
+    set -e
+
+    (time cmake -DCMAKE_BUILD_TYPE=Debug ..     ) 2>&1 | tee -a $OUT/${ci}-cmake.log
+    (time make -j                               ) 2>&1 | tee -a $OUT/${ci}-make.log
+
+    (time ctest --output-on-failure -E test-opt ) 2>&1 | tee -a $OUT/${ci}-ctest.log
+
+    set +e
+}
+
+function gg_sum_ctest_debug {
+    gg_printf '### %s\n\n' "${ci}"
+
+    gg_printf 'Runs ctest in debug mode\n'
+    gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
+    gg_printf '```\n'
+    gg_printf '%s\n' "$(cat $OUT/${ci}-ctest.log)"
+    gg_printf '```\n'
+    gg_printf '\n'
+}
+
+# ctest_release
+
+function gg_run_ctest_release {
+    cd ${SRC}
+
+    rm -rf build-ci-release && mkdir build-ci-release && cd build-ci-release
+
+    set -e
+
+    (time cmake -DCMAKE_BUILD_TYPE=Release ..   ) 2>&1 | tee -a $OUT/${ci}-cmake.log
+    (time make -j                               ) 2>&1 | tee -a $OUT/${ci}-make.log
+
+    if [ -z $GG_BUILD_LOW_PERF ]; then
+        (time ctest --output-on-failure ) 2>&1 | tee -a $OUT/${ci}-ctest.log
+    else
+        (time ctest --output-on-failure -E test-opt ) 2>&1 | tee -a $OUT/${ci}-ctest.log
+    fi
+
+    set +e
+}
+
+function gg_sum_ctest_release {
+    gg_printf '### %s\n\n' "${ci}"
+
+    gg_printf 'Runs ctest in release mode\n'
+    gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
+    gg_printf '```\n'
+    gg_printf '%s\n' "$(cat $OUT/${ci}-ctest.log)"
+    gg_printf '```\n'
+}
+
+# gpt_2
+
+function gg_run_gpt_2 {
+    cd ${SRC}
+
+    gg_wget models-mnt/gpt-2 https://huggingface.co/ggerganov/ggml/resolve/main/ggml-model-gpt-2-117M.bin
+
+    cd build-ci-release
+
+    set -e
+
+    model="../models-mnt/gpt-2/ggml-model-gpt-2-117M.bin"
+    prompts="../examples/prompts/gpt-2.txt"
+
+    (time ./bin/gpt-2 --model ${model} -s 1234 -n 64 -tt ${prompts}                       ) 2>&1 | tee -a $OUT/${ci}-tg.log
+    (time ./bin/gpt-2 --model ${model} -s 1234 -n 64 -p "I believe the meaning of life is") 2>&1 | tee -a $OUT/${ci}-tg.log
+
+    (time ./bin/gpt-2-batched --model ${model} -s 1234 -n 64 -np 8 -p "I believe the meaning of life is") 2>&1 | tee -a $OUT/${ci}-tg.log
+
+    set +e
+}
+
+function gg_sum_gpt_2 {
+    gg_printf '### %s\n\n' "${ci}"
+
+    gg_printf 'Runs short GPT-2 text generation\n'
+    gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
+    gg_printf '```\n'
+    gg_printf '%s\n' "$(cat $OUT/${ci}-tg.log)"
+    gg_printf '```\n'
+}
+
+# mnist
+
+function gg_run_mnist {
+    cd ${SRC}
+
+    cd build-ci-release
+
+    set -e
+
+    mkdir -p models/mnist
+    python3 ../examples/mnist/convert-h5-to-ggml.py ../examples/mnist/models/mnist/mnist_model.state_dict
+
+    model_f32="./models/mnist/ggml-model-f32.bin"
+    samples="../examples/mnist/models/mnist/t10k-images.idx3-ubyte"
+
+    # first command runs and exports "mnist.ggml", the second command runs the exported model
+
+    (time ./bin/mnist     ${model_f32} ${samples} ) 2>&1 | tee -a $OUT/${ci}-mnist.log
+    (time ./bin/mnist-cpu ./mnist.ggml ${samples} ) 2>&1 | tee -a $OUT/${ci}-mnist.log
+
+    set +e
+}
+
+function gg_sum_mnist {
+    gg_printf '### %s\n\n' "${ci}"
+
+    gg_printf 'MNIST\n'
+    gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
+    gg_printf '```\n'
+    gg_printf '%s\n' "$(cat $OUT/${ci}-mnist.log)"
+    gg_printf '```\n'
+}
+
+# whisper
+
+function gg_run_whisper {
+    cd ${SRC}
+
+    gg_wget models-mnt/whisper/ https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-base.en.bin
+    gg_wget models-mnt/whisper/ https://github.com/ggerganov/whisper.cpp/raw/master/samples/jfk.wav
+
+    cd build-ci-release
+
+    set -e
+
+    path_models="../models-mnt/whisper/"
+    model_f16="${path_models}/ggml-base.en.bin"
+    audio_0="${path_models}/jfk.wav"
+
+    (time ./bin/whisper -m ${model_f16} -f ${audio_0} ) 2>&1 | tee -a $OUT/${ci}-main.log
+
+    grep -q "And so my fellow Americans" $OUT/${ci}-main.log
+
+    set +e
+}
+
+function gg_sum_whisper {
+    gg_printf '### %s\n\n' "${ci}"
+
+    gg_printf 'Runs short Whisper transcription\n'
+    gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
+    gg_printf '```\n'
+    gg_printf '%s\n' "$(cat $OUT/${ci}-main.log)"
+    gg_printf '```\n'
+}
+
+# sam
+
+function gg_run_sam {
+    cd ${SRC}
+
+    gg_wget models-mnt/sam/ https://dl.fbaipublicfiles.com/segment_anything/sam_vit_b_01ec64.pth
+    gg_wget models-mnt/sam/ https://raw.githubusercontent.com/YavorGIvanov/sam.cpp/ceafb7467bff7ec98e0c4f952e58a9eb8fd0238b/img.jpg
+
+    cd build-ci-release
+
+    set -e
+
+    path_models="../models-mnt/sam/"
+    model_f16="${path_models}/ggml-model-f16.bin"
+    img_0="${path_models}/img.jpg"
+
+    python3 ../examples/sam/convert-pth-to-ggml.py ${path_models}/sam_vit_b_01ec64.pth ${path_models}/ 1
+
+    (time ./bin/sam -m ${model_f16} -i ${img_0} ) 2>&1 | tee -a $OUT/${ci}-main.log
+
+    grep -q "bbox (371, 436), (144, 168)" $OUT/${ci}-main.log
+
+    set +e
+}
+
+function gg_sum_sam {
+    gg_printf '### %s\n\n' "${ci}"
+
+    gg_printf 'Run SAM\n'
+    gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
+    gg_printf '```\n'
+    gg_printf '%s\n' "$(cat $OUT/${ci}-main.log)"
+    gg_printf '```\n'
+}
+
+# mpt
+
+function gg_run_mpt {
+    cd ${SRC}
+
+    gg_wget models-mnt/mpt/7B/ https://huggingface.co/mosaicml/mpt-7b/raw/main/config.json
+    gg_wget models-mnt/mpt/7B/ https://huggingface.co/mosaicml/mpt-7b/raw/main/tokenizer.json
+    gg_wget models-mnt/mpt/7B/ https://huggingface.co/mosaicml/mpt-7b/raw/main/tokenizer_config.json
+    gg_wget models-mnt/mpt/7B/ https://huggingface.co/mosaicml/mpt-7b/raw/main/pytorch_model.bin.index.json
+    gg_wget models-mnt/mpt/7B/ https://huggingface.co/mosaicml/mpt-7b/raw/main/configuration_mpt.py
+    gg_wget models-mnt/mpt/7B/ https://huggingface.co/mosaicml/mpt-7b/resolve/main/pytorch_model-00001-of-00002.bin
+    gg_wget models-mnt/mpt/7B/ https://huggingface.co/mosaicml/mpt-7b/resolve/main/pytorch_model-00002-of-00002.bin
+
+    cd build-ci-release
+
+    set -e
+
+    path_models="../models-mnt/mpt/7B"
+    model_f16="${path_models}/ggml-model-f16.bin"
+    model_q4_0="${path_models}/ggml-model-q4_0.bin"
+
+    python3 ../examples/mpt/convert-h5-to-ggml.py ${path_models} 1
+    ./bin/mpt-quantize ${model_f16} ${model_q4_0} q4_0
+
+    (time ./bin/mpt --model ${model_f16}  -s 1234 -n 64 -p "I believe the meaning of life is") 2>&1 | tee -a $OUT/${ci}-tg.log
+    (time ./bin/mpt --model ${model_q4_0} -s 1234 -n 64 -p "I believe the meaning of life is") 2>&1 | tee -a $OUT/${ci}-tg.log
+
+    set +e
+}
+
+function gg_sum_mpt {
+    gg_printf '### %s\n\n' "${ci}"
+
+    gg_printf 'Runs short MPT text generation\n'
+    gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
+    gg_printf '```\n'
+    gg_printf '%s\n' "$(cat $OUT/${ci}-tg.log)"
+    gg_printf '```\n'
+}
+
+## main
+
+if [ -z $GG_BUILD_LOW_PERF ]; then
+    rm -rf ${SRC}/models-mnt
+
+    mnt_models=${MNT}/models
+    mkdir -p ${mnt_models}
+    ln -sfn ${mnt_models} ${SRC}/models-mnt
+fi
+
+python3 -m pip install -r ${SRC}/requirements.txt
+
+ret=0
+
+test $ret -eq 0 && gg_run ctest_debug
+test $ret -eq 0 && gg_run ctest_release
+test $ret -eq 0 && gg_run gpt_2
+test $ret -eq 0 && gg_run mnist
+test $ret -eq 0 && gg_run whisper
+test $ret -eq 0 && gg_run sam
+
+if [ -z $GG_BUILD_LOW_PERF ]; then
+    if [ -z ${GG_BUILD_VRAM_GB} ] || [ ${GG_BUILD_VRAM_GB} -ge 16 ]; then
+        test $ret -eq 0 && gg_run mpt
+    fi
+fi
+
+exit $ret

stable-diffusion.cpp/ggml/cmake/BuildTypes.cmake

@@ -0,0 +1,54 @@
+# Add new build types
+
+# ReleaseGG - Release with enabled asserts
+
+SET(CMAKE_CXX_FLAGS_RELEASEGG
+    "-O3"
+    CACHE STRING "Flags used by the c++ compiler during release builds with enabled asserts."
+    FORCE )
+SET(CMAKE_C_FLAGS_RELEASEGG
+    "-O3"
+    CACHE STRING "Flags used by the compiler during release builds with enabled asserts."
+    FORCE )
+SET(CMAKE_EXE_LINKER_FLAGS_RELEASEGG
+    ""
+    CACHE STRING "Flags used for linking binaries during release builds with enabled asserts."
+    FORCE )
+SET(CMAKE_SHARED_LINKER_FLAGS_RELEASEGG
+    ""
+    CACHE STRING "Flags used by the shared libraries linker during release builds with enabled asserts."
+    FORCE )
+MARK_AS_ADVANCED(
+    CMAKE_CXX_FLAGS_RELEASEGG
+    CMAKE_C_FLAGS_RELEASEGG
+    CMAKE_EXE_LINKER_FLAGS_RELEASEGG
+    CMAKE_SHARED_LINKER_FLAGS_RELEASEGG )
+
+# RelWithDebInfoGG - RelWithDebInfo with enabled asserts
+
+SET(CMAKE_CXX_FLAGS_RELWITHDEBINFOGG
+    "-O2 -g"
+    CACHE STRING "Flags used by the c++ compiler during release builds with debug symbols and enabled asserts."
+    FORCE )
+SET(CMAKE_C_FLAGS_RELWITHDEBINFOGG
+    "-O2 -g"
+    CACHE STRING "Flags used by the compiler during release builds with debug symbols and enabled asserts."
+    FORCE )
+SET(CMAKE_EXE_LINKER_FLAGS_RELWITHDEBINFOGG
+    ""
+    CACHE STRING "Flags used for linking binaries during release builds with debug symbols and enabled asserts."
+    FORCE )
+SET(CMAKE_SHARED_LINKER_FLAGS_RELWITHDEBINFOGG
+    ""
+    CACHE STRING "Flags used by the shared libraries linker during release builds with debug symbols and enabled asserts."
+    FORCE )
+MARK_AS_ADVANCED(
+    CMAKE_CXX_FLAGS_RELWITHDEBINFOGG
+    CMAKE_C_FLAGS_RELWITHDEBINFOGG
+    CMAKE_EXE_LINKER_FLAGS_RELWITHDEBINFOGG
+    CMAKE_SHARED_LINKER_FLAGS_RELWITHDEBINFOGG )
+
+if (NOT XCODE AND NOT MSVC AND NOT CMAKE_BUILD_TYPE)
+    set(CMAKE_BUILD_TYPE Release CACHE STRING "Build type" FORCE)
+    set_property(CACHE CMAKE_BUILD_TYPE PROPERTY STRINGS "Debug" "Release" "MinSizeRel" "RelWithDebInfo" "ReleaseGG" "RelWithDebInfoGG")
+endif()

stable-diffusion.cpp/ggml/cmake/GitVars.cmake

@@ -0,0 +1,22 @@
+find_package(Git)
+
+# the commit's SHA1
+execute_process(COMMAND
+    "${GIT_EXECUTABLE}" describe --match=NeVeRmAtCh --always --abbrev=8
+    WORKING_DIRECTORY "${CMAKE_SOURCE_DIR}"
+    OUTPUT_VARIABLE GIT_SHA1
+    ERROR_QUIET OUTPUT_STRIP_TRAILING_WHITESPACE)
+
+# the date of the commit
+execute_process(COMMAND
+    "${GIT_EXECUTABLE}" log -1 --format=%ad --date=local
+    WORKING_DIRECTORY "${CMAKE_SOURCE_DIR}"
+    OUTPUT_VARIABLE GIT_DATE
+    ERROR_QUIET OUTPUT_STRIP_TRAILING_WHITESPACE)
+
+# the subject of the commit
+execute_process(COMMAND
+    "${GIT_EXECUTABLE}" log -1 --format=%s
+    WORKING_DIRECTORY "${CMAKE_SOURCE_DIR}"
+    OUTPUT_VARIABLE GIT_COMMIT_SUBJECT
+    ERROR_QUIET OUTPUT_STRIP_TRAILING_WHITESPACE)

stable-diffusion.cpp/ggml/examples/CMakeLists.txt

@@ -0,0 +1,30 @@
+if (GGML_ALL_WARNINGS)
+  if (NOT MSVC)
+      set(cxx_flags
+          # TODO(marella): Add other warnings.
+          -Wpedantic
+          -Wunused-variable
+          -Wno-unused-function
+          -Wno-multichar
+      )
+      add_compile_options("$<$<COMPILE_LANGUAGE:CXX>:${cxx_flags}>")
+  endif()
+endif()
+
+add_library(common STATIC common.cpp)
+target_include_directories(common PUBLIC ${CMAKE_CURRENT_SOURCE_DIR})
+
+add_library(common-ggml STATIC common-ggml.cpp)
+target_link_libraries(common-ggml PRIVATE ggml)
+target_include_directories(common-ggml PUBLIC ${CMAKE_CURRENT_SOURCE_DIR})
+
+add_subdirectory(gpt-2)
+add_subdirectory(gpt-j)
+add_subdirectory(whisper)
+add_subdirectory(mnist)
+add_subdirectory(gpt-neox)
+add_subdirectory(dolly-v2)
+add_subdirectory(replit)
+add_subdirectory(mpt)
+add_subdirectory(starcoder)
+add_subdirectory(sam)

stable-diffusion.cpp/ggml/examples/common-ggml.cpp

@@ -0,0 +1,246 @@
+#include "common-ggml.h"
+
+#include <regex>
+#include <map>
+
+static const std::map<std::string, enum ggml_ftype> GGML_FTYPE_MAP = {
+    {"q4_0", GGML_FTYPE_MOSTLY_Q4_0},
+    {"q4_1", GGML_FTYPE_MOSTLY_Q4_1},
+    {"q5_0", GGML_FTYPE_MOSTLY_Q5_0},
+    {"q5_1", GGML_FTYPE_MOSTLY_Q5_1},
+    {"q8_0", GGML_FTYPE_MOSTLY_Q8_0},
+};
+
+void ggml_print_ftypes(FILE * fp) {
+    for (auto it = GGML_FTYPE_MAP.begin(); it != GGML_FTYPE_MAP.end(); it++) {
+        fprintf(fp, "  type = \"%s\" or %d\n", it->first.c_str(), it->second);
+    }
+}
+
+enum ggml_ftype ggml_parse_ftype(const char * str) {
+    enum ggml_ftype ftype;
+    if (str[0] == 'q') {
+        const auto it = GGML_FTYPE_MAP.find(str);
+        if (it == GGML_FTYPE_MAP.end()) {
+            fprintf(stderr, "%s: unknown ftype '%s'\n", __func__, str);
+            return GGML_FTYPE_UNKNOWN;
+        }
+        ftype = it->second;
+    } else {
+        ftype = (enum ggml_ftype) atoi(str);
+    }
+
+    return ftype;
+}
+
+bool ggml_common_quantize_0(
+        std::ifstream & finp,
+        std::ofstream & fout,
+        const ggml_ftype ftype,
+        const std::vector<std::string> & to_quant,
+        const std::vector<std::string> & to_skip) {
+
+    ggml_type qtype = GGML_TYPE_F32;
+
+    switch (ftype) {
+        case GGML_FTYPE_MOSTLY_Q4_0: qtype = GGML_TYPE_Q4_0; break;
+        case GGML_FTYPE_MOSTLY_Q4_1: qtype = GGML_TYPE_Q4_1; break;
+        case GGML_FTYPE_MOSTLY_Q5_0: qtype = GGML_TYPE_Q5_0; break;
+        case GGML_FTYPE_MOSTLY_Q5_1: qtype = GGML_TYPE_Q5_1; break;
+        case GGML_FTYPE_MOSTLY_Q8_0: qtype = GGML_TYPE_Q8_0; break;
+        case GGML_FTYPE_UNKNOWN:
+        case GGML_FTYPE_ALL_F32:
+        case GGML_FTYPE_MOSTLY_F16:
+        case GGML_FTYPE_MOSTLY_Q4_1_SOME_F16:
+        case GGML_FTYPE_MOSTLY_Q2_K:
+        case GGML_FTYPE_MOSTLY_Q3_K:
+        case GGML_FTYPE_MOSTLY_Q4_K:
+        case GGML_FTYPE_MOSTLY_Q5_K:
+        case GGML_FTYPE_MOSTLY_Q6_K:
+                {
+                    fprintf(stderr, "%s: invalid model type %d\n", __func__, ftype);
+                    return false;
+                }
+    };
+
+    if (!ggml_is_quantized(qtype)) {
+        fprintf(stderr, "%s: invalid quantization type %d (%s)\n", __func__, qtype, ggml_type_name(qtype));
+        return false;
+    }
+
+    size_t total_size_org = 0;
+    size_t total_size_new = 0;
+
+    std::vector<float> work;
+
+    std::vector<uint8_t>     data_u8;
+    std::vector<ggml_fp16_t> data_f16;
+    std::vector<float>       data_f32;
+
+    std::vector<int64_t> hist_all(1 << 4, 0);
+
+    while (true) {
+        int32_t n_dims;
+        int32_t length;
+        int32_t ttype;
+
+        finp.read(reinterpret_cast<char *>(&n_dims), sizeof(n_dims));
+        finp.read(reinterpret_cast<char *>(&length), sizeof(length));
+        finp.read(reinterpret_cast<char *>(&ttype),  sizeof(ttype));
+
+        if (finp.eof()) {
+            break;
+        }
+
+        int32_t nelements = 1;
+        int32_t ne[4] = { 1, 1, 1, 1 };
+        for (int i = 0; i < n_dims; ++i) {
+            finp.read (reinterpret_cast<char *>(&ne[i]), sizeof(ne[i]));
+            nelements *= ne[i];
+        }
+
+        std::string name(length, 0);
+        finp.read (&name[0], length);
+
+        printf("%64s - [%5d, %5d, %5d], type = %6s ", name.data(), ne[0], ne[1], ne[2], ggml_type_name((ggml_type) ttype));
+
+        bool quantize = false;
+
+        // check if we should quantize this tensor
+        for (const auto & s : to_quant) {
+            if (std::regex_match(name, std::regex(s))) {
+                quantize = true;
+                break;
+            }
+        }
+
+        // check if we should skip this tensor
+        for (const auto & s : to_skip) {
+            if (std::regex_match(name, std::regex(s))) {
+                quantize = false;
+                break;
+            }
+        }
+
+        // quantize only 2D tensors
+        quantize &= (n_dims == 2);
+
+        if (quantize) {
+            if (ttype != GGML_TYPE_F32 && ttype != GGML_TYPE_F16) {
+                fprintf(stderr, "%s: unsupported ttype %d (%s) for integer quantization\n", __func__, ttype, ggml_type_name((ggml_type) ttype));
+                return false;
+            }
+
+            if (ttype == GGML_TYPE_F16) {
+                data_f16.resize(nelements);
+                finp.read(reinterpret_cast<char *>(data_f16.data()), nelements * sizeof(ggml_fp16_t));
+                data_f32.resize(nelements);
+                for (int i = 0; i < nelements; ++i) {
+                    data_f32[i] = ggml_fp16_to_fp32(data_f16[i]);
+                }
+            } else {
+                data_f32.resize(nelements);
+                finp.read(reinterpret_cast<char *>(data_f32.data()), nelements * sizeof(float));
+            }
+
+            ttype = qtype;
+        } else {
+            const int bpe = (ttype == 0) ? sizeof(float) : sizeof(uint16_t);
+
+            data_u8.resize(nelements*bpe);
+            finp.read(reinterpret_cast<char *>(data_u8.data()), nelements * bpe);
+        }
+
+        fout.write(reinterpret_cast<char *>(&n_dims), sizeof(n_dims));
+        fout.write(reinterpret_cast<char *>(&length), sizeof(length));
+        fout.write(reinterpret_cast<char *>(&ttype),  sizeof(ttype));
+        for (int i = 0; i < n_dims; ++i) {
+            fout.write(reinterpret_cast<char *>(&ne[i]), sizeof(ne[i]));
+        }
+        fout.write(&name[0], length);
+
+        if (quantize) {
+            work.resize(nelements); // for quantization
+
+            size_t cur_size = 0;
+            std::vector<int64_t> hist_cur(1 << 4, 0);
+
+            switch ((ggml_type) ttype) {
+                case GGML_TYPE_Q4_0:
+                    {
+                        cur_size = ggml_quantize_q4_0(data_f32.data(), work.data(), nelements, ne[0], hist_cur.data());
+                    } break;
+                case GGML_TYPE_Q4_1:
+                    {
+                        cur_size = ggml_quantize_q4_1(data_f32.data(), work.data(), nelements, ne[0], hist_cur.data());
+                    } break;
+                case GGML_TYPE_Q5_0:
+                    {
+                        cur_size = ggml_quantize_q5_0(data_f32.data(), work.data(), nelements, ne[0], hist_cur.data());
+                    } break;
+                case GGML_TYPE_Q5_1:
+                    {
+                        cur_size = ggml_quantize_q5_1(data_f32.data(), work.data(), nelements, ne[0], hist_cur.data());
+                    } break;
+                case GGML_TYPE_Q8_0:
+                    {
+                        cur_size = ggml_quantize_q8_0(data_f32.data(), work.data(), nelements, ne[0], hist_cur.data());
+                    } break;
+                case GGML_TYPE_F32:
+                case GGML_TYPE_F16:
+                case GGML_TYPE_I8:
+                case GGML_TYPE_I16:
+                case GGML_TYPE_I32:
+                case GGML_TYPE_Q8_1:
+                case GGML_TYPE_Q2_K:
+                case GGML_TYPE_Q3_K:
+                case GGML_TYPE_Q4_K:
+                case GGML_TYPE_Q5_K:
+                case GGML_TYPE_Q6_K:
+                case GGML_TYPE_Q8_K:
+                case GGML_TYPE_COUNT:
+                    {
+                        fprintf(stderr, "%s: unsupported quantization type %d (%s)\n", __func__, ttype, ggml_type_name((ggml_type) ttype));
+                        return false;
+                    }
+            }
+
+            fout.write(reinterpret_cast<char *>(work.data()), cur_size);
+            total_size_new += cur_size;
+
+            printf("size = %8.2f MB -> %8.2f MB | hist: ", nelements * sizeof(float)/1024.0/1024.0, cur_size/1024.0/1024.0);
+            for (int i = 0; i < (int) hist_cur.size(); ++i) {
+                hist_all[i] += hist_cur[i];
+            }
+
+            for (int i = 0; i < (int) hist_cur.size(); ++i) {
+                printf("%5.3f ", hist_cur[i] / (float)nelements);
+            }
+            printf("\n");
+        } else {
+            printf("size = %8.3f MB\n", data_u8.size()/1024.0/1024.0);
+            fout.write(reinterpret_cast<char *>(data_u8.data()), data_u8.size());
+            total_size_new += data_u8.size();
+        }
+
+        total_size_org += nelements * sizeof(float);
+    }
+
+    printf("%s: model size  = %8.2f MB\n", __func__, total_size_org/1024.0/1024.0);
+    printf("%s: quant size  = %8.2f MB | ftype = %d (%s)\n", __func__, total_size_new/1024.0/1024.0, ftype, ggml_type_name(qtype));
+
+    {
+        int64_t sum_all = 0;
+        for (int i = 0; i < (int) hist_all.size(); ++i) {
+            sum_all += hist_all[i];
+        }
+
+        printf("%s: hist: ", __func__);
+        for (int i = 0; i < (int) hist_all.size(); ++i) {
+            printf("%5.3f ", hist_all[i] / (float)sum_all);
+        }
+        printf("\n");
+    }
+
+    return true;
+}

stable-diffusion.cpp/ggml/examples/common-ggml.h

@@ -0,0 +1,18 @@
+#pragma once
+
+#include "ggml.h"
+
+#include <fstream>
+#include <vector>
+#include <string>
+
+enum ggml_ftype ggml_parse_ftype(const char * str);
+
+void ggml_print_ftypes(FILE * fp = stderr);
+
+bool ggml_common_quantize_0(
+        std::ifstream & finp,
+        std::ofstream & fout,
+        const ggml_ftype ftype,
+        const std::vector<std::string> & to_quant,
+        const std::vector<std::string> & to_skip);

stable-diffusion.cpp/ggml/examples/common.cpp

@@ -0,0 +1,817 @@
+#define _USE_MATH_DEFINES // for M_PI
+
+#include "common.h"
+
+// third-party utilities
+// use your favorite implementations
+#define DR_WAV_IMPLEMENTATION
+#include "dr_wav.h"
+
+#include <cmath>
+#include <cstring>
+#include <fstream>
+#include <regex>
+#include <locale>
+#include <codecvt>
+#include <sstream>
+
+#if defined(_MSC_VER)
+#pragma warning(disable: 4244 4267) // possible loss of data
+#endif
+
+// Function to check if the next argument exists
+std::string get_next_arg(int& i, int argc, char** argv, const std::string& flag, gpt_params& params) {
+    if (i + 1 < argc && argv[i + 1][0] != '-') {
+        return argv[++i];
+    } else {
+        fprintf(stderr, "error: %s requires one argument.\n", flag.c_str());
+        gpt_print_usage(argc, argv, params);
+        exit(0);
+    }
+}
+
+bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
+    for (int i = 1; i < argc; i++) {
+        std::string arg = argv[i];
+
+        if (arg == "-s" || arg == "--seed") {
+            params.seed = std::stoi(get_next_arg(i, argc, argv, arg, params));
+        } else if (arg == "-t" || arg == "--threads") {
+            params.n_threads = std::stoi(get_next_arg(i, argc, argv, arg, params));
+        } else if (arg == "-p" || arg == "--prompt") {
+            params.prompt = get_next_arg(i, argc, argv, arg, params);
+        } else if (arg == "-n" || arg == "--n_predict") {
+            params.n_predict = std::stoi(get_next_arg(i, argc, argv, arg, params));
+        } else if (arg == "-np" || arg == "--n_parallel") {
+            params.n_parallel = std::stoi(get_next_arg(i, argc, argv, arg, params));
+        } else if (arg == "--top_k") {
+            params.top_k = std::stoi(get_next_arg(i, argc, argv, arg, params));
+        } else if (arg == "--top_p") {
+            params.top_p = std::stof(get_next_arg(i, argc, argv, arg, params));
+        } else if (arg == "--temp") {
+            params.temp = std::stof(get_next_arg(i, argc, argv, arg, params));
+        } else if (arg == "--repeat-last-n") {
+            params.repeat_last_n = std::stoi(get_next_arg(i, argc, argv, arg, params));
+        } else if (arg == "--repeat-penalty") {
+            params.repeat_penalty = std::stof(get_next_arg(i, argc, argv, arg, params));
+        } else if (arg == "-b" || arg == "--batch_size") {
+            params.n_batch= std::stoi(get_next_arg(i, argc, argv, arg, params));
+        } else if (arg == "-c" || arg == "--context") {
+            params.n_ctx= std::stoi(get_next_arg(i, argc, argv, arg, params));
+        } else if (arg == "-ngl" || arg == "--gpu-layers" || arg == "--n-gpu-layers") {
+            params.n_gpu_layers = std::stoi(get_next_arg(i, argc, argv, arg, params));
+        } else if (arg == "--ignore-eos") {
+            params.ignore_eos = true;
+        } else if (arg == "-m" || arg == "--model") {
+            params.model = get_next_arg(i, argc, argv, arg, params);
+        } else if (arg == "-i" || arg == "--interactive") {
+            params.interactive = true;
+        } else if (arg == "-ip" || arg == "--interactive-port") {
+            params.interactive = true;
+            params.interactive_port = std::stoi(get_next_arg(i, argc, argv, arg, params));
+        } else if (arg == "-h" || arg == "--help") {
+            gpt_print_usage(argc, argv, params);
+            exit(0);
+        } else if (arg == "-f" || arg == "--file") {
+            get_next_arg(i, argc, argv, arg, params);
+            std::ifstream file(argv[i]);
+            if (!file) {
+                fprintf(stderr, "error: failed to open file '%s'\n", argv[i]);
+                break;
+            }
+            std::copy(std::istreambuf_iterator<char>(file), std::istreambuf_iterator<char>(), back_inserter(params.prompt));
+            if (params.prompt.back() == '\n') {
+                params.prompt.pop_back();
+            }
+        } else if (arg == "-tt" || arg == "--token_test") {
+            params.token_test = get_next_arg(i, argc, argv, arg, params);
+        }
+        else {
+            fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
+            gpt_print_usage(argc, argv, params);
+            exit(0);
+        }
+    }
+
+    return true;
+}
+
+void gpt_print_usage(int /*argc*/, char ** argv, const gpt_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, "  -s SEED, --seed SEED  RNG seed (default: -1)\n");
+    fprintf(stderr, "  -t N, --threads N     number of threads to use during computation (default: %d)\n", params.n_threads);
+    fprintf(stderr, "  -p PROMPT, --prompt PROMPT\n");
+    fprintf(stderr, "                        prompt to start generation with (default: random)\n");
+    fprintf(stderr, "  -f FNAME, --file FNAME\n");
+    fprintf(stderr, "                        load prompt from a file\n");
+    fprintf(stderr, "  -tt TOKEN_TEST, --token_test TOKEN_TEST\n");
+    fprintf(stderr, "                        test tokenization\n");
+    fprintf(stderr, "  -n N, --n_predict N   number of tokens to predict (default: %d)\n", params.n_predict);
+    fprintf(stderr, "  --top_k N             top-k sampling (default: %d)\n", params.top_k);
+    fprintf(stderr, "  --top_p N             top-p sampling (default: %.1f)\n", params.top_p);
+    fprintf(stderr, "  --temp N              temperature (default: %.1f)\n", params.temp);
+    fprintf(stderr, "  --repeat-last-n N     last n tokens to consider for penalize (default: %d, 0 = disabled)\n", params.repeat_last_n);
+    fprintf(stderr, "  --repeat-penalty N    penalize repeat sequence of tokens (default: %.2f, 1.0 = disabled)\n", (double)params.repeat_penalty);
+    fprintf(stderr, "  -b N, --batch_size N  batch size for prompt processing (default: %d)\n", params.n_batch);
+    fprintf(stderr, "  -c N, --context N     context / KV cache size (default: %d)\n", params.n_ctx);
+    fprintf(stderr, "  --ignore-eos          ignore EOS token during generation\n");
+    fprintf(stderr, "  -ngl N, --gpu-layers N  number of layers to offload to GPU on supported models (default: %d)\n", params.n_gpu_layers);
+    fprintf(stderr, "  -m FNAME, --model FNAME\n");
+    fprintf(stderr, "                        model path (default: %s)\n", params.model.c_str());
+    fprintf(stderr, "\n");
+}
+
+std::string gpt_random_prompt(std::mt19937 & rng) {
+    const int r = rng() % 10;
+    switch (r) {
+        case 0: return "So";
+        case 1: return "Once upon a time";
+        case 2: return "When";
+        case 3: return "The";
+        case 4: return "After";
+        case 5: return "If";
+        case 6: return "import";
+        case 7: return "He";
+        case 8: return "She";
+        case 9: return "They";
+        default: return "To";
+    }
+
+    return "The";
+}
+
+std::string trim(const std::string & s) {
+    std::regex e("^\\s+|\\s+$");
+    return std::regex_replace(s, e, "");
+}
+
+std::string replace(const std::string & s, const std::string & from, const std::string & to) {
+    std::string result = s;
+    size_t pos = 0;
+    while ((pos = result.find(from, pos)) != std::string::npos) {
+        result.replace(pos, from.length(), to);
+        pos += to.length();
+    }
+    return result;
+}
+
+void gpt_vocab::add_special_token(const std::string & token) {
+    special_tokens.push_back(token);
+}
+
+std::map<std::string, int32_t> json_parse(const std::string & fname) {
+    std::map<std::string, int32_t> result;
+
+    // read file into string
+    std::string json;
+    {
+        std::ifstream ifs(fname);
+        if (!ifs) {
+            fprintf(stderr, "Failed to open %s\n", fname.c_str());
+            exit(1);
+        }
+
+        json = std::string((std::istreambuf_iterator<char>(ifs)),
+                (std::istreambuf_iterator<char>()));
+    }
+
+    if (json[0] != '{') {
+        return result;
+    }
+
+    // parse json
+    {
+        bool has_key  = false;
+        bool in_token = false;
+
+        std::string str_key = "";
+        std::string str_val = "";
+
+        int n = json.size();
+        for (int i = 1; i < n; ++i) {
+            if (!in_token) {
+                if (json[i] == ' ') continue;
+                if (json[i] == '"') {
+                    in_token = true;
+                    continue;
+                }
+            } else {
+                if (json[i] == '\\' && i+1 < n) {
+                    if (has_key == false) {
+                        str_key += json[i];
+                    } else {
+                        str_val += json[i];
+                    }
+                    ++i;
+                } else if (json[i] == '"') {
+                    if (has_key == false) {
+                        has_key = true;
+                        ++i;
+                        while (json[i] == ' ') ++i;
+                        ++i; // :
+                        while (json[i] == ' ') ++i;
+                        if (json[i] != '\"') {
+                            while (json[i] != ',' && json[i] != '}') {
+                                str_val += json[i++];
+                            }
+                            has_key = false;
+                        } else {
+                            in_token = true;
+                            continue;
+                        }
+                    } else {
+                        has_key = false;
+                    }
+
+                    str_key = ::replace(str_key, "\\u0120", " " ); // \u0120 -> space
+                    str_key = ::replace(str_key, "\\u010a", "\n"); // \u010a -> new line
+                    str_key = ::replace(str_key, "\\\"",    "\""); // \\\"   -> "
+
+                    try {
+                        result[str_key] = std::stoi(str_val);
+                    } catch (...) {
+                        //fprintf(stderr, "%s: ignoring key '%s' with value '%s'\n", fname.c_str(), str_key.c_str(), str_val.c_str());
+
+                    }
+                    str_key = "";
+                    str_val = "";
+                    in_token = false;
+                    continue;
+                }
+                if (has_key == false) {
+                    str_key += json[i];
+                } else {
+                    str_val += json[i];
+                }
+            }
+        }
+    }
+
+    return result;
+}
+
+std::string convert_to_utf8(const std::wstring & input) {
+    std::wstring_convert<std::codecvt_utf8<wchar_t>> converter;
+    return converter.to_bytes(input);
+}
+
+
+std::wstring convert_to_wstring(const std::string & input) {
+    std::wstring_convert<std::codecvt_utf8<wchar_t>> converter;
+    return converter.from_bytes(input);
+}
+
+void gpt_split_words(std::string str, std::vector<std::string>& words) {
+    const std::string pattern = R"('s|'t|'re|'ve|'m|'ll|'d| ?[[:alpha:]]+| ?[[:digit:]]+| ?[^\s[:alpha:][:digit:]]+|\s+(?!\S)|\s+)";
+    const std::regex re(pattern);
+    std::smatch m;
+
+    while (std::regex_search(str, m, re)) {
+        for (auto x : m) {
+            words.push_back(x);
+        }
+        str = m.suffix();
+    }
+}
+
+std::vector<gpt_vocab::id> gpt_tokenize(const gpt_vocab & vocab, const std::string & text) {
+    std::vector<std::string> words;
+
+    // first split the text into words
+    {
+        std::string str = text;
+
+        // Generate the subpattern from the special_tokens vector if it's not empty
+        if (!vocab.special_tokens.empty()) {
+            const std::regex escape(R"([\[\\\^\$\.\|\?\*\+\(\)\{\}])");
+            std::string special_tokens_subpattern;
+            for (const auto & token : vocab.special_tokens) {
+                if (!special_tokens_subpattern.empty()) {
+                    special_tokens_subpattern += "|";
+                }
+                special_tokens_subpattern += std::regex_replace(token, escape, R"(\$&)");
+            }
+
+            std::regex re(special_tokens_subpattern);
+            std::smatch m;
+            // Split the text by special tokens.
+            while (std::regex_search(str, m, re)) {
+                // Split the substrings in-between special tokens into words.
+                gpt_split_words(m.prefix(), words);
+                // Add matched special tokens as words.
+                for (auto x : m) {
+                    words.push_back(x);
+                }
+                str = m.suffix();
+            }
+            // Remaining text without special tokens will be handled below.
+        }
+
+        gpt_split_words(str, words);
+    }
+
+    // find the longest token that forms each word in words:
+    std::vector<gpt_vocab::id> tokens;
+    for (const auto & word : words) {
+        for (int i = 0; i < (int) word.size(); ){
+            for (int j = word.size() - 1; j >= i; j--){
+                auto cand = word.substr(i, j-i+1);
+                auto it = vocab.token_to_id.find(cand);
+                if (it != vocab.token_to_id.end()){ // word.substr(i, j-i+1) in vocab
+                    tokens.push_back(it->second);
+                    i = j + 1;
+                    break;
+                }
+                else if (j == i){ // word.substr(i, 1) has no matching
+                    fprintf(stderr, "%s: unknown token '%s'\n", __func__, word.substr(i, 1).data());
+                    i++;
+                }
+            }
+        }
+    }
+
+    return tokens;
+}
+
+std::vector<gpt_vocab::id> parse_tokens_from_string(const std::string& input, char delimiter) {
+    std::vector<gpt_vocab::id> output;
+    std::stringstream ss(input);
+    std::string token;
+
+    while (std::getline(ss, token, delimiter)) {
+        output.push_back(std::stoi(token));
+    }
+
+    return output;
+}
+
+std::map<std::string, std::vector<gpt_vocab::id>> extract_tests_from_file(const std::string & fpath_test){
+    if (fpath_test.empty()){
+        fprintf(stderr, "%s : No test file found.\n", __func__);
+        return std::map<std::string, std::vector<gpt_vocab::id>>();
+    }
+
+    std::map<std::string, std::vector<gpt_vocab::id>> tests;
+
+    auto fin = std::ifstream(fpath_test, std::ios_base::in);
+    const char * delimeter = " => ";
+    const char del_tok = ',';
+    std::string line;
+    while (std::getline(fin, line)) {
+        size_t delimiterPos = line.find(delimeter);
+        if (delimiterPos != std::string::npos) {
+            std::string text = line.substr(0, delimiterPos);
+            std::string s_tokens = line.substr(delimiterPos + std::strlen(delimeter));
+            tests[text] = parse_tokens_from_string(s_tokens, del_tok);
+        }
+    }
+    return tests;
+}
+
+void test_gpt_tokenizer(gpt_vocab & vocab, const std::string & fpath_test){
+    std::map<std::string, std::vector<gpt_vocab::id>> tests = extract_tests_from_file(fpath_test);
+
+    size_t n_fails = 0;
+
+    for (const auto & test : tests) {
+        std::vector<gpt_vocab::id> tokens = gpt_tokenize(vocab, test.first);
+
+        if (tokens != test.second){
+            n_fails++;
+
+            // print out failure cases
+            fprintf(stderr, "%s : failed test: '%s'\n", __func__, test.first.c_str());
+            fprintf(stderr, "%s : tokens in hf:   ", __func__);
+            for (const auto & t : test.second) {
+                fprintf(stderr, "%s(%d), ", vocab.id_to_token[t].c_str(), t);
+            }
+            fprintf(stderr, "\n");
+            fprintf(stderr, "%s : tokens in ggml: ", __func__);
+            for (const auto & t : tokens) {
+                fprintf(stderr, "%s(%d), ", vocab.id_to_token[t].c_str(), t);
+            }
+            fprintf(stderr, "\n");
+        }
+    }
+
+    fprintf(stderr, "%s : %zu tests failed out of %zu tests.\n", __func__, n_fails, tests.size());
+}
+
+bool gpt_vocab_init(const std::string & fname, gpt_vocab & vocab) {
+    printf("%s: loading vocab from '%s'\n", __func__, fname.c_str());
+
+    vocab.token_to_id = ::json_parse(fname);
+
+    for (const auto & kv : vocab.token_to_id) {
+        vocab.id_to_token[kv.second] = kv.first;
+    }
+
+    printf("%s: vocab size = %d\n", __func__, (int) vocab.token_to_id.size());
+
+    // print the vocabulary
+    //for (auto kv : vocab.token_to_id) {
+    //    printf("'%s' -> %d\n", kv.first.data(), kv.second);
+    //}
+
+    return true;
+}
+
+gpt_vocab::id gpt_sample_top_k_top_p(
+        const gpt_vocab & vocab,
+        const float * logits,
+        int    top_k,
+        double top_p,
+        double temp,
+        std::mt19937 & rng) {
+    int n_logits = vocab.id_to_token.size();
+
+    std::vector<std::pair<double, gpt_vocab::id>> logits_id;
+    logits_id.reserve(n_logits);
+
+    {
+        const double scale = 1.0/temp;
+        for (int i = 0; i < n_logits; ++i) {
+            logits_id.push_back(std::make_pair(logits[i]*scale, i));
+        }
+    }
+
+    // find the top K tokens
+    std::partial_sort(
+            logits_id.begin(),
+            logits_id.begin() + top_k, logits_id.end(),
+            [](const std::pair<double, gpt_vocab::id> & a, const std::pair<double, gpt_vocab::id> & b) {
+        return a.first > b.first;
+    });
+
+    logits_id.resize(top_k);
+
+    double maxl = -INFINITY;
+    for (const auto & kv : logits_id) {
+        maxl = std::max(maxl, kv.first);
+    }
+
+    // compute probs for the top K tokens
+    std::vector<double> probs;
+    probs.reserve(logits_id.size());
+
+    double sum = 0.0;
+    for (const auto & kv : logits_id) {
+        double p = exp(kv.first - maxl);
+        probs.push_back(p);
+        sum += p;
+    }
+
+    // normalize the probs
+    for (auto & p : probs) {
+        p /= sum;
+    }
+
+    if (top_p < 1.0f) {
+        double cumsum = 0.0f;
+        for (int i = 0; i < top_k; i++) {
+            cumsum += probs[i];
+            if (cumsum >= top_p) {
+                top_k = i + 1;
+                probs.resize(top_k);
+                logits_id.resize(top_k);
+                break;
+            }
+        }
+
+        cumsum = 1.0/cumsum;
+        for (int i = 0; i < (int) probs.size(); i++) {
+            probs[i] *= cumsum;
+        }
+    }
+
+    //printf("\n");
+    //for (int i = 0; i < (int) probs.size(); i++) {
+    //    printf("%d: '%s' %f\n", i, vocab.id_to_token.at(logits_id[i].second).c_str(), probs[i]);
+    //}
+    //exit(0);
+
+    std::discrete_distribution<> dist(probs.begin(), probs.end());
+    int idx = dist(rng);
+
+    return logits_id[idx].second;
+}
+
+gpt_vocab::id gpt_sample_top_k_top_p_repeat(
+        const gpt_vocab & vocab,
+        const float * logits,
+        const int32_t * last_n_tokens_data,
+        size_t last_n_tokens_data_size,
+        int    top_k,
+        double top_p,
+        double temp,
+        int repeat_last_n,
+        float repeat_penalty,
+        std::mt19937 & rng) {
+
+    int n_logits = vocab.id_to_token.size();
+
+    const auto * plogits = logits;
+
+    const auto last_n_tokens = std::vector<int32_t>(last_n_tokens_data, last_n_tokens_data + last_n_tokens_data_size);
+
+    if (temp <= 0) {
+        // select the token with the highest logit directly
+        float max_logit = plogits[0];
+        gpt_vocab::id max_id = 0;
+
+        for (int i = 1; i < n_logits; ++i) {
+            if (plogits[i] > max_logit) {
+                max_logit = plogits[i];
+                max_id = i;
+            }
+        }
+        return max_id;
+    }
+
+
+    std::vector<std::pair<double, gpt_vocab::id>> logits_id;
+    logits_id.reserve(n_logits);
+
+    {
+        const float scale = 1.0f/temp;
+        for (int i = 0; i < n_logits; ++i) {
+            // repetition penalty from ctrl paper (https://arxiv.org/abs/1909.05858)
+            // credit https://github.com/facebookresearch/llama/compare/main...shawwn:llama:main
+            if (repeat_last_n > 0 && std::find(last_n_tokens.end()-repeat_last_n, last_n_tokens.end(), i) != last_n_tokens.end()) {
+                // if score < 0 then repetition penalty has to multiplied to reduce the previous token probability
+                if (plogits[i] < 0.0f) {
+                    logits_id.push_back(std::make_pair(plogits[i]*scale*repeat_penalty, i));
+                } else {
+                    logits_id.push_back(std::make_pair(plogits[i]*scale/repeat_penalty, i));
+                }
+            } else {
+                logits_id.push_back(std::make_pair(plogits[i]*scale, i));
+            }
+        }
+    }
+
+    // find the top K tokens
+    std::partial_sort(
+            logits_id.begin(),
+            logits_id.begin() + top_k, logits_id.end(),
+            [](const std::pair<double, gpt_vocab::id> & a, const std::pair<double, gpt_vocab::id> & b) {
+        return a.first > b.first;
+    });
+
+    logits_id.resize(top_k);
+
+    double maxl = -INFINITY;
+    for (const auto & kv : logits_id) {
+        maxl = std::max(maxl, kv.first);
+    }
+
+    // compute probs for the top K tokens
+    std::vector<double> probs;
+    probs.reserve(logits_id.size());
+
+    double sum = 0.0;
+    for (const auto & kv : logits_id) {
+        double p = exp(kv.first - maxl);
+        probs.push_back(p);
+        sum += p;
+    }
+
+    // normalize the probs
+    for (auto & p : probs) {
+        p /= sum;
+    }
+
+    if (top_p < 1.0f) {
+        double cumsum = 0.0f;
+        for (int i = 0; i < top_k; i++) {
+            cumsum += probs[i];
+            if (cumsum >= top_p) {
+                top_k = i + 1;
+                probs.resize(top_k);
+                logits_id.resize(top_k);
+                break;
+            }
+        }
+
+        cumsum = 1.0/cumsum;
+        for (int i = 0; i < (int) probs.size(); i++) {
+            probs[i] *= cumsum;
+        }
+    }
+
+//    printf("\n");
+//    for (int i = 0; i < (int) probs.size(); i++) {
+//    for (int i = 0; i < 10; i++) {
+//        printf("%d: '%s' %f\n", i, vocab.id_to_token.at(logits_id[i].second).c_str(), probs[i]);
+//    }
+
+    std::discrete_distribution<> dist(probs.begin(), probs.end());
+    int idx = dist(rng);
+
+    return logits_id[idx].second;
+
+}
+
+bool read_wav(const std::string & fname, std::vector<float>& pcmf32, std::vector<std::vector<float>>& pcmf32s, bool stereo) {
+    drwav wav;
+    std::vector<uint8_t> wav_data; // used for pipe input from stdin
+
+    if (fname == "-") {
+        {
+            uint8_t buf[1024];
+            while (true)
+            {
+                const size_t n = fread(buf, 1, sizeof(buf), stdin);
+                if (n == 0) {
+                    break;
+                }
+                wav_data.insert(wav_data.end(), buf, buf + n);
+            }
+        }
+
+        if (drwav_init_memory(&wav, wav_data.data(), wav_data.size(), nullptr) == false) {
+            fprintf(stderr, "error: failed to open WAV file from stdin\n");
+            return false;
+        }
+
+        fprintf(stderr, "%s: read %zu bytes from stdin\n", __func__, wav_data.size());
+    }
+    else if (drwav_init_file(&wav, fname.c_str(), nullptr) == false) {
+        fprintf(stderr, "error: failed to open '%s' as WAV file\n", fname.c_str());
+        return false;
+    }
+
+    if (wav.channels != 1 && wav.channels != 2) {
+        fprintf(stderr, "%s: WAV file '%s' must be mono or stereo\n", __func__, fname.c_str());
+        return false;
+    }
+
+    if (stereo && wav.channels != 2) {
+        fprintf(stderr, "%s: WAV file '%s' must be stereo for diarization\n", __func__, fname.c_str());
+        return false;
+    }
+
+    if (wav.sampleRate != COMMON_SAMPLE_RATE) {
+        fprintf(stderr, "%s: WAV file '%s' must be %i kHz\n", __func__, fname.c_str(), COMMON_SAMPLE_RATE/1000);
+        return false;
+    }
+
+    if (wav.bitsPerSample != 16) {
+        fprintf(stderr, "%s: WAV file '%s' must be 16-bit\n", __func__, fname.c_str());
+        return false;
+    }
+
+    const uint64_t n = wav_data.empty() ? wav.totalPCMFrameCount : wav_data.size()/(wav.channels*wav.bitsPerSample/8);
+
+    std::vector<int16_t> pcm16;
+    pcm16.resize(n*wav.channels);
+    drwav_read_pcm_frames_s16(&wav, n, pcm16.data());
+    drwav_uninit(&wav);
+
+    // convert to mono, float
+    pcmf32.resize(n);
+    if (wav.channels == 1) {
+        for (uint64_t i = 0; i < n; i++) {
+            pcmf32[i] = float(pcm16[i])/32768.0f;
+        }
+    } else {
+        for (uint64_t i = 0; i < n; i++) {
+            pcmf32[i] = float(pcm16[2*i] + pcm16[2*i + 1])/65536.0f;
+        }
+    }
+
+    if (stereo) {
+        // convert to stereo, float
+        pcmf32s.resize(2);
+
+        pcmf32s[0].resize(n);
+        pcmf32s[1].resize(n);
+        for (uint64_t i = 0; i < n; i++) {
+            pcmf32s[0][i] = float(pcm16[2*i])/32768.0f;
+            pcmf32s[1][i] = float(pcm16[2*i + 1])/32768.0f;
+        }
+    }
+
+    return true;
+}
+
+void high_pass_filter(std::vector<float> & data, float cutoff, float sample_rate) {
+    const float rc = 1.0f / (2.0f * M_PI * cutoff);
+    const float dt = 1.0f / sample_rate;
+    const float alpha = dt / (rc + dt);
+
+    float y = data[0];
+
+    for (size_t i = 1; i < data.size(); i++) {
+        y = alpha * (y + data[i] - data[i - 1]);
+        data[i] = y;
+    }
+}
+
+bool vad_simple(std::vector<float> & pcmf32, int sample_rate, int last_ms, float vad_thold, float freq_thold, bool verbose) {
+    const int n_samples      = pcmf32.size();
+    const int n_samples_last = (sample_rate * last_ms) / 1000;
+
+    if (n_samples_last >= n_samples) {
+        // not enough samples - assume no speech
+        return false;
+    }
+
+    if (freq_thold > 0.0f) {
+        high_pass_filter(pcmf32, freq_thold, sample_rate);
+    }
+
+    float energy_all  = 0.0f;
+    float energy_last = 0.0f;
+
+    for (int i = 0; i < n_samples; i++) {
+        energy_all += fabsf(pcmf32[i]);
+
+        if (i >= n_samples - n_samples_last) {
+            energy_last += fabsf(pcmf32[i]);
+        }
+    }
+
+    energy_all  /= n_samples;
+    energy_last /= n_samples_last;
+
+    if (verbose) {
+        fprintf(stderr, "%s: energy_all: %f, energy_last: %f, vad_thold: %f, freq_thold: %f\n", __func__, energy_all, energy_last, vad_thold, freq_thold);
+    }
+
+    if (energy_last > vad_thold*energy_all) {
+        return false;
+    }
+
+    return true;
+}
+
+float similarity(const std::string & s0, const std::string & s1) {
+    const size_t len0 = s0.size() + 1;
+    const size_t len1 = s1.size() + 1;
+
+    std::vector<int> col(len1, 0);
+    std::vector<int> prevCol(len1, 0);
+
+    for (size_t i = 0; i < len1; i++) {
+        prevCol[i] = i;
+    }
+
+    for (size_t i = 0; i < len0; i++) {
+        col[0] = i;
+        for (size_t j = 1; j < len1; j++) {
+            col[j] = std::min(std::min(1 + col[j - 1], 1 + prevCol[j]), prevCol[j - 1] + (i > 0 && s0[i - 1] == s1[j - 1] ? 0 : 1));
+        }
+        col.swap(prevCol);
+    }
+
+    const float dist = prevCol[len1 - 1];
+
+    return 1.0f - (dist / std::max(s0.size(), s1.size()));
+}
+
+bool sam_params_parse(int argc, char ** argv, sam_params & params) {
+    for (int i = 1; i < argc; i++) {
+        std::string arg = argv[i];
+
+        if (arg == "-s" || arg == "--seed") {
+            params.seed = std::stoi(argv[++i]);
+        } else if (arg == "-t" || arg == "--threads") {
+            params.n_threads = std::stoi(argv[++i]);
+        } else if (arg == "-m" || arg == "--model") {
+            params.model = argv[++i];
+        } else if (arg == "-i" || arg == "--inp") {
+            params.fname_inp = argv[++i];
+        } else if (arg == "-o" || arg == "--out") {
+            params.fname_out = argv[++i];
+        } else if (arg == "-h" || arg == "--help") {
+            sam_print_usage(argc, argv, params);
+            exit(0);
+        } else {
+            fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
+            sam_print_usage(argc, argv, params);
+            exit(0);
+        }
+    }
+
+    return true;
+}
+
+void sam_print_usage(int /*argc*/, char ** argv, const sam_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, "  -s SEED, --seed SEED  RNG seed (default: -1)\n");
+    fprintf(stderr, "  -t N, --threads N     number of threads to use during computation (default: %d)\n", params.n_threads);
+    fprintf(stderr, "  -m FNAME, --model FNAME\n");
+    fprintf(stderr, "                        model path (default: %s)\n", params.model.c_str());
+    fprintf(stderr, "  -i FNAME, --inp FNAME\n");
+    fprintf(stderr, "                        input file (default: %s)\n", params.fname_inp.c_str());
+    fprintf(stderr, "  -o FNAME, --out FNAME\n");
+    fprintf(stderr, "                        output file (default: %s)\n", params.fname_out.c_str());
+    fprintf(stderr, "\n");
+}

stable-diffusion.cpp/ggml/examples/common.h

@@ -0,0 +1,179 @@
+// Various helper functions and utilities
+
+#pragma once
+
+#include <string>
+#include <map>
+#include <vector>
+#include <random>
+#include <thread>
+
+#define COMMON_SAMPLE_RATE 16000
+
+//
+// GPT CLI argument parsing
+//
+
+struct gpt_params {
+    int32_t seed         = -1;   // RNG seed
+    int32_t n_threads    = std::min(4, (int32_t) std::thread::hardware_concurrency());
+    int32_t n_predict    = 200;  // new tokens to predict
+    int32_t n_parallel   = 1;    // number of parallel streams
+    int32_t n_batch      = 8;    // batch size for prompt processing
+    int32_t n_ctx        = 2048; // context size (this is the KV cache max size)
+    int32_t n_gpu_layers = 0;    // number of layers to offlload to the GPU
+
+    bool ignore_eos = false; // ignore EOS token when generating text
+
+    // sampling parameters
+    int32_t top_k          = 40;
+    float   top_p          = 0.9f;
+    float   temp           = 0.9f;
+    int32_t repeat_last_n  = 64;
+    float   repeat_penalty = 1.00f;
+
+    std::string model      = "models/gpt-2-117M/ggml-model.bin"; // model path
+    std::string prompt     = "";
+    std::string token_test = "";
+
+    bool    interactive      = false;
+    int32_t interactive_port = -1;
+};
+
+bool gpt_params_parse(int argc, char ** argv, gpt_params & params);
+
+void gpt_print_usage(int argc, char ** argv, const gpt_params & params);
+
+std::string gpt_random_prompt(std::mt19937 & rng);
+
+//
+// Vocab utils
+//
+
+std::string trim(const std::string & s);
+
+std::string replace(
+        const std::string & s,
+        const std::string & from,
+        const std::string & to);
+
+struct gpt_vocab {
+    using id    = int32_t;
+    using token = std::string;
+
+    std::map<token, id> token_to_id;
+    std::map<id, token> id_to_token;
+    std::vector<std::string> special_tokens;
+
+    void add_special_token(const std::string & token);
+};
+
+// poor-man's JSON parsing
+std::map<std::string, int32_t> json_parse(const std::string & fname);
+
+std::string convert_to_utf8(const std::wstring & input);
+
+std::wstring convert_to_wstring(const std::string & input);
+
+void gpt_split_words(std::string str, std::vector<std::string>& words);
+
+// split text into tokens
+//
+// ref: https://github.com/openai/gpt-2/blob/a74da5d99abaaba920de8131d64da2862a8f213b/src/encoder.py#L53
+//
+// Regex (Python):
+// r"""'s|'t|'re|'ve|'m|'ll|'d| ?\p{L}+| ?\p{N}+| ?[^\s\p{L}\p{N}]+|\s+(?!\S)|\s+"""
+//
+// Regex (C++):
+// R"('s|'t|'re|'ve|'m|'ll|'d| ?[[:alpha:]]+| ?[[:digit:]]+| ?[^\s[:alpha:][:digit:]]+|\s+(?!\S)|\s+)"
+//
+std::vector<gpt_vocab::id> gpt_tokenize(const gpt_vocab & vocab, const std::string & text);
+
+// test outputs of gpt_tokenize
+//
+//   - compare with tokens generated by the huggingface tokenizer
+//   - test cases are chosen based on the model's main language (under 'prompt' directory)
+//   - if all sentences are tokenized identically, print 'All tests passed.'
+//   - otherwise, print sentence, huggingface tokens, ggml tokens
+//
+void test_gpt_tokenizer(gpt_vocab & vocab, const std::string & fpath_test);
+
+// load the tokens from encoder.json
+bool gpt_vocab_init(const std::string & fname, gpt_vocab & vocab);
+
+// sample next token given probabilities for each embedding
+//
+//   - consider only the top K tokens
+//   - from them, consider only the top tokens with cumulative probability > P
+//
+// TODO: not sure if this implementation is correct
+// TODO: temperature is not implemented
+//
+gpt_vocab::id gpt_sample_top_k_top_p(
+        const gpt_vocab & vocab,
+        const float * logits,
+        int    top_k,
+        double top_p,
+        double temp,
+        std::mt19937 & rng);
+
+gpt_vocab::id gpt_sample_top_k_top_p_repeat(
+        const gpt_vocab & vocab,
+        const float * logits,
+        const int32_t * last_n_tokens_data,
+        size_t last_n_tokens_data_size,
+        int    top_k,
+        double top_p,
+        double temp,
+        int repeat_last_n,
+        float repeat_penalty,
+        std::mt19937 & rng);
+
+//
+// Audio utils
+//
+
+// Read WAV audio file and store the PCM data into pcmf32
+// The sample rate of the audio must be equal to COMMON_SAMPLE_RATE
+// If stereo flag is set and the audio has 2 channels, the pcmf32s will contain 2 channel PCM
+bool read_wav(
+        const std::string & fname,
+        std::vector<float> & pcmf32,
+        std::vector<std::vector<float>> & pcmf32s,
+        bool stereo);
+
+// Apply a high-pass frequency filter to PCM audio
+// Suppresses frequencies below cutoff Hz
+void high_pass_filter(
+        std::vector<float> & data,
+        float cutoff,
+        float sample_rate);
+
+// Basic voice activity detection (VAD) using audio energy adaptive threshold
+bool vad_simple(
+        std::vector<float> & pcmf32,
+        int   sample_rate,
+        int   last_ms,
+        float vad_thold,
+        float freq_thold,
+        bool  verbose);
+
+// compute similarity between two strings using Levenshtein distance
+float similarity(const std::string & s0, const std::string & s1);
+
+//
+// SAM argument parsing
+//
+
+struct sam_params {
+    int32_t seed      = -1; // RNG seed
+    int32_t n_threads = std::min(4, (int32_t) std::thread::hardware_concurrency());
+
+    std::string model     = "models/sam-vit-b/ggml-model-f16.bin"; // model path
+    std::string fname_inp = "img.jpg";
+    std::string fname_out = "img.out";
+};
+
+bool sam_params_parse(int argc, char ** argv, sam_params & params);
+
+void sam_print_usage(int argc, char ** argv, const sam_params & params);

stable-diffusion.cpp/ggml/examples/dolly-v2/CMakeLists.txt

@@ -0,0 +1,13 @@
+#
+# dollyv2
+
+set(TEST_TARGET dollyv2)
+add_executable(${TEST_TARGET} main.cpp)
+target_link_libraries(${TEST_TARGET} PRIVATE ggml common common-ggml)
+
+#
+# dollyv2-quantize
+
+set(TEST_TARGET dollyv2-quantize)
+add_executable(${TEST_TARGET} quantize.cpp)
+target_link_libraries(${TEST_TARGET} PRIVATE ggml common common-ggml)

stable-diffusion.cpp/ggml/examples/dolly-v2/README.md

@@ -0,0 +1,187 @@
+# Dolly-V2
+
+Transformer architecture: GPT-NeoX
+
+Modeled from examples/stablelm
+
+Ref: https://github.com/databrickslabs/dolly
+
+Ref: https://github.com/stability-AI/stableLM/#stablelm-alpha
+
+## Usage
+
+```bash
+# get the repo and build it
+git clone https://github.com/ggerganov/ggml
+cd ggml
+mkdir build && cd build
+cmake ..
+make -j
+
+# get the Dolly-V2 3B model
+git clone https://huggingface.co/databricks/dolly-v2-3b
+
+# install Python dependencies
+python3 -m pip install -r ../requirements.txt
+
+# convert model to FP16
+python3 ../examples/dolly-v2/convert-h5-to-ggml.py ./dolly-v2-3b/ 1
+
+# run inference using FP16 precision
+./bin/dollyv2 -m ./dolly-v2-3b/ggml-model-f16.bin -p "State the meaning of life." -t 6 -n 64
+
+main: seed = 1683218142
+dollyv2_model_load: loading model from './dolly-v2-3b/ggml-model-f16.bin' - please wait ...
+dollyv2_model_load: n_vocab = 50280
+dollyv2_model_load: n_ctx   = 2048
+dollyv2_model_load: n_embd  = 2560
+dollyv2_model_load: n_head  = 32
+dollyv2_model_load: n_layer = 32
+dollyv2_model_load: n_rot   = 20
+dollyv2_model_load: ftype   = 1
+dollyv2_model_load: ggml ctx size = 7374.91 MB
+dollyv2_model_load: memory_size =   640.00 MB, n_mem = 65536
+dollyv2_model_load: ................................................ done
+dollyv2_model_load: model size =  5295.10 MB / num tensors = 388
+main: number of tokens in prompt = 32
+main: token[0] =  30003, Below
+main: token[1] =    310,  is
+main: token[2] =    271,  an
+main: token[3] =   9775,  instruction
+main: token[4] =    326,  that
+main: token[5] =   8631,  describes
+main: token[6] =    247,  a
+main: token[7] =   4836,  task
+main: token[8] =    964, .
+main: token[9] =  19566,  Write
+main: token[10] =    247,  a
+main: token[11] =   2380,  response
+main: token[12] =    326,  that
+main: token[13] =  20420,  appropriately
+main: token[14] =  29141,  completes
+main: token[15] =    253,  the
+main: token[16] =   2748,  request
+main: token[17] =    964, .
+main: token[18] =    187, 
+
+main: token[19] =    187, 
+
+main: token[20] =  50278, ### Instruction:
+main: token[21] =    187, 
+
+main: token[22] =   5443, State
+main: token[23] =    253,  the
+main: token[24] =   4495,  meaning
+main: token[25] =    273,  of
+main: token[26] =   1495,  life
+main: token[27] =    964, .
+main: token[28] =    187, 
+
+main: token[29] =    187, 
+
+main: token[30] =  50279, ### Response:
+main: token[31] =    187, 
+
+
+Below is an instruction that describes a task. Write a response that appropriately completes the request.
+
+### Instruction:
+State the meaning of life.
+
+### Response:
+The meaning of life is to love and be loved.
+
+### End
+
+main: mem per token = 16136720 bytes
+main:     load time =  2202.58 ms
+main:   sample time =     2.57 ms
+main:  predict time =  1497.14 ms / 33.27 ms per token
+main:    total time =  6187.27 ms
+```
+
+## 5-bit integer quantization mode
+
+```bash
+# quantize the model to 5-bits using Q5_0 quantization
+./bin/dollyv2-quantize ./dolly-v2-3b/ggml-model-f16.bin ./dolly-v2-3b/ggml-model-q5_0.bin q5_0
+
+# run the quantized model
+./bin/dollyv2 -m ./dolly-v2-3b/ggml-model-q5_0.bin -p "State the meaning of life." -t 6 -n 64
+
+main: seed = 1683218518
+dollyv2_model_load: loading model from './dolly-v2-3b/ggml-model-q5_0.bin' - please wait ...
+dollyv2_model_load: n_vocab = 50280
+dollyv2_model_load: n_ctx   = 2048
+dollyv2_model_load: n_embd  = 2560
+dollyv2_model_load: n_head  = 32
+dollyv2_model_load: n_layer = 32
+dollyv2_model_load: n_rot   = 20
+dollyv2_model_load: ftype   = 8
+dollyv2_model_load: ggml ctx size = 3902.68 MB
+dollyv2_model_load: memory_size =   640.00 MB, n_mem = 65536
+dollyv2_model_load: ................................................ done
+dollyv2_model_load: model size =  1822.87 MB / num tensors = 388
+main: number of tokens in prompt = 32
+main: token[0] =  30003, Below
+main: token[1] =    310,  is
+main: token[2] =    271,  an
+main: token[3] =   9775,  instruction
+main: token[4] =    326,  that
+main: token[5] =   8631,  describes
+main: token[6] =    247,  a
+main: token[7] =   4836,  task
+main: token[8] =    964, .
+main: token[9] =  19566,  Write
+main: token[10] =    247,  a
+main: token[11] =   2380,  response
+main: token[12] =    326,  that
+main: token[13] =  20420,  appropriately
+main: token[14] =  29141,  completes
+main: token[15] =    253,  the
+main: token[16] =   2748,  request
+main: token[17] =    964, .
+main: token[18] =    187, 
+
+main: token[19] =    187, 
+
+main: token[20] =  50278, ### Instruction:
+main: token[21] =    187, 
+
+main: token[22] =   5443, State
+main: token[23] =    253,  the
+main: token[24] =   4495,  meaning
+main: token[25] =    273,  of
+main: token[26] =   1495,  life
+main: token[27] =    964, .
+main: token[28] =    187, 
+
+main: token[29] =    187, 
+
+main: token[30] =  50279, ### Response:
+main: token[31] =    187, 
+
+
+Below is an instruction that describes a task. Write a response that appropriately completes the request.
+
+### Instruction:
+State the meaning of life.
+
+### Response:
+The meaning of life is the discovery of the true self.
+
+### End
+
+main: mem per token = 16127760 bytes
+main:     load time =  1011.09 ms
+main:   sample time =     2.79 ms
+main:  predict time =  1271.62 ms / 27.64 ms per token
+main:    total time =  2802.51 ms
+```
+
+## Notes
+
+- No guarantees for correctness
+- The tokenizer is currently hacked - probably works only for English
+- Non-parallel residual is not supported
+- Contributions and improvements are welcome

stable-diffusion.cpp/ggml/examples/dolly-v2/convert-h5-to-ggml.py

@@ -0,0 +1,116 @@
+import sys
+import struct
+import json
+import numpy as np
+
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+if len(sys.argv) < 3:
+    print("Usage: convert-h5-to-ggml.py dir-model [use-f32]\n")
+    print("  ftype == 0 -> float32")
+    print("  ftype == 1 -> float16")
+    sys.exit(1)
+
+# output in the same directory as the model
+dir_model = sys.argv[1]
+fname_out = sys.argv[1] + "/ggml-model.bin"
+
+with open(dir_model + "/tokenizer.json", "r", encoding="utf-8") as f:
+    encoder = json.load(f)
+
+with open(dir_model + "/config.json", "r", encoding="utf-8") as f:
+    hparams = json.load(f)
+
+# possible data types
+#   ftype == 0 -> float32
+#   ftype == 1 -> float16
+#
+# map from ftype to string
+ftype_str = ["f32", "f16"]
+
+ftype = 1
+if len(sys.argv) > 2:
+    ftype = int(sys.argv[2])
+    if ftype < 0 or ftype > 1:
+        print("Invalid ftype: " + str(ftype))
+        sys.exit(1)
+    fname_out = sys.argv[1] + "/ggml-model-" + ftype_str[ftype] + ".bin"
+
+
+tokenizer = AutoTokenizer.from_pretrained(dir_model)
+model = AutoModelForCausalLM.from_pretrained(dir_model, low_cpu_mem_usage=True)
+#print (model)
+
+#print(tokenizer.encode('I believe the meaning of life is'))
+
+list_vars = model.state_dict()
+for name in list_vars.keys():
+    print(name, list_vars[name].shape, list_vars[name].dtype)
+
+fout = open(fname_out, "wb")
+
+print(hparams)
+
+fout.write(struct.pack("i", 0x67676d6c)) # magic: ggml in hex
+fout.write(struct.pack("i", hparams["vocab_size"]))
+fout.write(struct.pack("i", hparams["max_position_embeddings"]))
+fout.write(struct.pack("i", hparams["hidden_size"]))
+fout.write(struct.pack("i", hparams["num_attention_heads"]))
+fout.write(struct.pack("i", hparams["num_hidden_layers"]))
+fout.write(struct.pack("i", int(hparams["rotary_pct"]*(hparams["hidden_size"]//hparams["num_attention_heads"]))))
+fout.write(struct.pack("i", hparams["use_parallel_residual"]))
+fout.write(struct.pack("i", ftype))
+
+# TODO: temporary hack to not deal with implementing the tokenizer
+dot_token = tokenizer.encode('.')[0]
+for i in range(hparams["vocab_size"]):
+    text = tokenizer.decode([dot_token, i]).encode('utf-8')
+    # remove the first byte (it's always '.')
+    text = text[1:]
+    fout.write(struct.pack("i", len(text)))
+    fout.write(text)
+
+for name in list_vars.keys():
+    data = list_vars[name].squeeze().numpy()
+    print("Processing variable: " + name + " with shape: ", data.shape)
+
+    # we don't need these
+    if name.endswith(".attention.masked_bias") or     \
+       name.endswith(".attention.bias") or \
+       name.endswith(".attention.rotary_emb.inv_freq"):
+        print("  Skipping variable: " + name)
+        continue
+
+    n_dims = len(data.shape);
+
+    # ftype == 0 -> float32, ftype == 1 -> float16
+    ftype_cur = 0;
+    if ftype != 0:
+        if name[-7:] == ".weight" and n_dims == 2:
+            print("  Converting to float16")
+            data = data.astype(np.float16)
+            ftype_cur = 1
+        else:
+            print("  Converting to float32")
+            data = data.astype(np.float32)
+            ftype_cur = 0
+    else:
+        if data.dtype != np.float32:
+            print("  Converting to float32")
+            data = data.astype(np.float32)
+            ftype_cur = 0
+
+    # header
+    str = name.encode('utf-8')
+    fout.write(struct.pack("iii", n_dims, len(str), ftype_cur))
+    for i in range(n_dims):
+        fout.write(struct.pack("i", data.shape[n_dims - 1 - i]))
+    fout.write(str);
+
+    # data
+    data.tofile(fout)
+
+fout.close()
+
+print("Done. Output file: " + fname_out)
+print("")

stable-diffusion.cpp/ggml/examples/dolly-v2/main.cpp

@@ -0,0 +1,969 @@
+#include "ggml/ggml.h"
+
+#include "common.h"
+#include "common-ggml.h"
+
+#include <cassert>
+#include <cmath>
+#include <cstdio>
+#include <cstring>
+#include <cinttypes>
+#include <fstream>
+#include <iostream>
+#include <map>
+#include <string>
+#include <vector>
+
+#if !defined(_WIN32)
+#define DOLLY_INTERACTIVE_PORT
+#endif
+
+#if defined(DOLLY_INTERACTIVE_PORT)
+#include <arpa/inet.h>
+#include <netinet/in.h>
+#include <sys/socket.h>
+#include <unistd.h>
+#endif
+
+#if defined(_MSC_VER)
+#pragma warning(disable: 4244 4267) // possible loss of data
+#endif
+
+// default hparams (Dolly-V2 3B)
+struct dollyv2_hparams {
+    int32_t n_vocab = 50254; // tokenizer.vocab_size
+    int32_t n_ctx   = 2048;  // model.config.max_position_embeddings
+    int32_t n_embd  = 2560;  // model.config.hidden_size
+    int32_t n_head  = 32;    // model.config.num_attention_heads
+    int32_t n_layer = 32;    // model.config.num_hidden_layers
+    int32_t n_rot   = 20;    // rotary_pct[25%] * (n_embd / n_head)
+    int32_t par_res = 1; // 1 = true, 0 = false
+    int32_t ftype   = GGML_FTYPE_MOSTLY_F16;
+    float   eps     = 1e-5f;
+};
+
+const std::string INSTRUCTION_KEY = "### Instruction:";
+const std::string RESPONSE_KEY    = "### Response:";
+const std::string END_KEY         = "### End";
+const std::string INTRO_BLURB     = "Below is an instruction that describes a task. Write a response that appropriately completes the request.";
+
+// dollyv2 prompt format
+std::string prompt_for_generation(const std::string& instruction) {
+    return INTRO_BLURB + "\n\n" + INSTRUCTION_KEY + "\n" + instruction + "\n\n" + RESPONSE_KEY + "\n";
+}
+
+struct dollyv2_layer {
+    // pre normalization
+    struct ggml_tensor * ln_1_g;
+    struct ggml_tensor * ln_1_b;
+
+    // attention
+    struct ggml_tensor * c_attn_attn_w;
+    struct ggml_tensor * c_attn_attn_b;
+
+    struct ggml_tensor * c_attn_proj_w;
+    struct ggml_tensor * c_attn_proj_b;
+
+    // post normalization
+    struct ggml_tensor * ln_2_g;
+    struct ggml_tensor * ln_2_b;
+
+    // ff
+    struct ggml_tensor * c_mlp_fc_w;
+    struct ggml_tensor * c_mlp_fc_b;
+
+    struct ggml_tensor * c_mlp_proj_w;
+    struct ggml_tensor * c_mlp_proj_b;
+};
+
+struct dollyv2_model {
+    dollyv2_hparams hparams;
+
+    // normalization
+    struct ggml_tensor * ln_f_g;
+    struct ggml_tensor * ln_f_b;
+
+    struct ggml_tensor * wte; // position embedding
+
+    struct ggml_tensor * lmh_g; // language model head
+    //struct ggml_tensor * lmh_b; // language model bias
+
+    std::vector<dollyv2_layer> layers;
+
+    // key + value memory
+    struct ggml_tensor * memory_k;
+    struct ggml_tensor * memory_v;
+
+    //
+    struct ggml_context * ctx;
+    std::map<std::string, struct ggml_tensor *> tensors;
+};
+
+// load the model's weights from a file
+bool dollyv2_model_load(const std::string & fname, dollyv2_model & model, gpt_vocab & vocab) {
+    printf("%s: loading model from '%s' - please wait ...\n", __func__, fname.c_str());
+
+    auto fin = std::ifstream(fname, std::ios::binary);
+    if (!fin) {
+        fprintf(stderr, "%s: failed to open '%s'\n", __func__, fname.c_str());
+        return false;
+    }
+
+    // verify magic
+    {
+        uint32_t magic;
+        fin.read((char *) &magic, sizeof(magic));
+        if (magic != GGML_FILE_MAGIC) {
+            fprintf(stderr, "%s: invalid model file '%s' (bad magic)\n", __func__, fname.c_str());
+            return false;
+        }
+    }
+
+    // load hparams
+    {
+        auto & hparams = model.hparams;
+
+        fin.read((char *) &hparams.n_vocab, sizeof(hparams.n_vocab));
+        fin.read((char *) &hparams.n_ctx,   sizeof(hparams.n_ctx));
+        fin.read((char *) &hparams.n_embd,  sizeof(hparams.n_embd));
+        fin.read((char *) &hparams.n_head,  sizeof(hparams.n_head));
+        fin.read((char *) &hparams.n_layer, sizeof(hparams.n_layer));
+        fin.read((char *) &hparams.n_rot,   sizeof(hparams.n_rot));
+        fin.read((char *) &hparams.par_res, sizeof(hparams.par_res));
+        fin.read((char *) &hparams.ftype,   sizeof(hparams.ftype));
+
+        const int32_t qntvr = hparams.ftype / GGML_QNT_VERSION_FACTOR;
+
+        printf("%s: n_vocab = %d\n", __func__, hparams.n_vocab);
+        printf("%s: n_ctx   = %d\n", __func__, hparams.n_ctx);
+        printf("%s: n_embd  = %d\n", __func__, hparams.n_embd);
+        printf("%s: n_head  = %d\n", __func__, hparams.n_head);
+        printf("%s: n_layer = %d\n", __func__, hparams.n_layer);
+        printf("%s: n_rot   = %d\n", __func__, hparams.n_rot);
+        printf("%s: par_res = %d\n", __func__, hparams.par_res);
+        printf("%s: ftype   = %d\n", __func__, hparams.ftype);
+        printf("%s: qntvr   = %d\n", __func__, qntvr);
+
+        hparams.ftype %= GGML_QNT_VERSION_FACTOR;
+    }
+
+    // load vocab
+    {
+        const int32_t n_vocab = model.hparams.n_vocab;
+
+        std::string word;
+        std::vector<char> buf(128);
+
+        for (int i = 0; i < n_vocab; i++) {
+            uint32_t len;
+            fin.read((char *) &len, sizeof(len));
+
+            buf.resize(len);
+            fin.read((char *) buf.data(), len);
+            word.assign(buf.data(), len);
+
+            vocab.token_to_id[word] = i;
+            vocab.id_to_token[i] = word;
+        }
+
+        vocab.add_special_token("### End");
+        vocab.add_special_token("### Instruction:");
+        vocab.add_special_token("### Response:");
+    }
+
+    // for the big tensors, we have the option to store the data in 16-bit floats or quantized
+    // in order to save memory and also to speed up the computation
+    ggml_type wtype = ggml_ftype_to_ggml_type((ggml_ftype) (model.hparams.ftype));
+    if (wtype == GGML_TYPE_COUNT) {
+        fprintf(stderr, "%s: invalid model file '%s' (bad ftype value %d)\n",
+                __func__, fname.c_str(), model.hparams.ftype);
+        return false;
+    }
+
+    auto & ctx = model.ctx;
+
+    size_t ctx_size = 0;
+
+    {
+        const auto & hparams = model.hparams;
+
+        const int n_embd  = hparams.n_embd;
+        const int n_layer = hparams.n_layer;
+        const int n_ctx   = hparams.n_ctx;
+        const int n_vocab = hparams.n_vocab;
+
+        ctx_size += n_embd*ggml_type_sizef(GGML_TYPE_F32); // ln_f_g
+        ctx_size += n_embd*ggml_type_sizef(GGML_TYPE_F32); // ln_f_b
+
+        ctx_size += n_embd*n_vocab*ggml_type_sizef(wtype); // wte
+
+        ctx_size += n_embd*n_vocab*ggml_type_sizef(wtype);           // lmh_g
+        //ctx_size +=        n_vocab*ggml_type_sizef(GGML_TYPE_F32); // lmh_b
+
+        ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // ln_1_g
+        ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // ln_1_b
+
+        ctx_size += n_layer*(3*n_embd*n_embd*ggml_type_sizef(wtype));         // c_attn_attn_w
+        ctx_size += n_layer*(       3*n_embd*ggml_type_sizef(GGML_TYPE_F32)); // c_attn_attn_b
+
+        ctx_size += n_layer*(n_embd*n_embd*ggml_type_sizef(wtype));         // c_attn_proj_w
+        ctx_size += n_layer*(n_embd*n_embd*ggml_type_sizef(GGML_TYPE_F32)); // c_attn_proj_b
+
+        ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // ln_2_g
+        ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // ln_2_b
+
+        ctx_size += n_layer*(4*n_embd*n_embd*ggml_type_sizef(wtype));         // c_mlp_fc_w
+        ctx_size += n_layer*(       4*n_embd*ggml_type_sizef(GGML_TYPE_F32)); // c_mlp_fc_b
+
+        ctx_size += n_layer*(4*n_embd*n_embd*ggml_type_sizef(wtype));         // c_mlp_proj_w
+        ctx_size += n_layer*(         n_embd*ggml_type_sizef(GGML_TYPE_F32)); // c_mlp_proj_b
+
+        ctx_size += n_ctx*n_layer*n_embd*ggml_type_sizef(GGML_TYPE_F32); // memory_k
+        ctx_size += n_ctx*n_layer*n_embd*ggml_type_sizef(GGML_TYPE_F32); // memory_v
+
+        ctx_size += (6 + 16*n_layer)*512; // object overhead
+
+        printf("%s: ggml ctx size = %6.2f MB\n", __func__, ctx_size/(1024.0*1024.0));
+    }
+
+    // create the ggml context
+    {
+        struct ggml_init_params params = {
+            /*.mem_size   =*/ ctx_size,
+            /*.mem_buffer =*/ NULL,
+            /*.no_alloc   =*/ false,
+        };
+
+        model.ctx = ggml_init(params);
+        if (!model.ctx) {
+            fprintf(stderr, "%s: ggml_init() failed\n", __func__);
+            return false;
+        }
+    }
+
+    // prepare memory for the weights
+    {
+        const auto & hparams = model.hparams;
+
+        const int n_embd  = hparams.n_embd;
+        const int n_layer = hparams.n_layer;
+        const int n_vocab = hparams.n_vocab;
+
+        model.layers.resize(n_layer);
+
+        model.wte    = ggml_new_tensor_2d(ctx, wtype,         n_embd, n_vocab);
+
+        model.ln_f_g = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
+        model.ln_f_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
+
+        model.lmh_g  = ggml_new_tensor_2d(ctx, wtype,         n_embd, n_vocab);
+        //model.lmh_b  = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_vocab);
+
+        // map by name
+        model.tensors["gpt_neox.embed_in.weight"] = model.wte;
+
+        model.tensors["gpt_neox.final_layer_norm.weight"] = model.ln_f_g;
+        model.tensors["gpt_neox.final_layer_norm.bias"]   = model.ln_f_b;
+
+        model.tensors["embed_out.weight"] = model.lmh_g;
+        //model.tensors["lm_head.bias"]   = model.lmh_b;
+
+        for (int i = 0; i < n_layer; ++i) {
+            auto & layer = model.layers[i];
+
+            layer.ln_1_g          = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_embd);
+            layer.ln_1_b          = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_embd);
+
+            layer.c_attn_attn_w   = ggml_new_tensor_2d(ctx, wtype,           n_embd, 3*n_embd);
+            layer.c_attn_attn_b   = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 3*n_embd);
+
+            layer.c_attn_proj_w   = ggml_new_tensor_2d(ctx, wtype,           n_embd,   n_embd);
+            layer.c_attn_proj_b   = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_embd);
+
+            layer.ln_2_g          = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_embd);
+            layer.ln_2_b          = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_embd);
+
+            layer.c_mlp_fc_w      = ggml_new_tensor_2d(ctx, wtype,           n_embd, 4*n_embd);
+            layer.c_mlp_fc_b      = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 4*n_embd);
+
+            layer.c_mlp_proj_w    = ggml_new_tensor_2d(ctx, wtype,         4*n_embd,   n_embd);
+            layer.c_mlp_proj_b    = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_embd);
+
+            // map by name
+
+            // unmapped: attention.rotary_emb, mlp.act
+
+            model.tensors["gpt_neox.layers." + std::to_string(i) + ".input_layernorm.weight"] = layer.ln_1_g;
+            model.tensors["gpt_neox.layers." + std::to_string(i) + ".input_layernorm.bias"]   = layer.ln_1_b;
+
+            model.tensors["gpt_neox.layers." + std::to_string(i) + ".attention.query_key_value.weight"] = layer.c_attn_attn_w;
+            model.tensors["gpt_neox.layers." + std::to_string(i) + ".attention.query_key_value.bias"]   = layer.c_attn_attn_b;
+
+            model.tensors["gpt_neox.layers." + std::to_string(i) + ".attention.dense.weight"] = layer.c_attn_proj_w;
+            model.tensors["gpt_neox.layers." + std::to_string(i) + ".attention.dense.bias"]   = layer.c_attn_proj_b;
+
+            model.tensors["gpt_neox.layers." + std::to_string(i) + ".post_attention_layernorm.weight"] = layer.ln_2_g;
+            model.tensors["gpt_neox.layers." + std::to_string(i) + ".post_attention_layernorm.bias"]   = layer.ln_2_b;
+
+            model.tensors["gpt_neox.layers." + std::to_string(i) + ".mlp.dense_h_to_4h.weight"] = layer.c_mlp_fc_w;
+            model.tensors["gpt_neox.layers." + std::to_string(i) + ".mlp.dense_h_to_4h.bias"]   = layer.c_mlp_fc_b;
+
+            model.tensors["gpt_neox.layers." + std::to_string(i) + ".mlp.dense_4h_to_h.weight"] = layer.c_mlp_proj_w;
+            model.tensors["gpt_neox.layers." + std::to_string(i) + ".mlp.dense_4h_to_h.bias"]   = layer.c_mlp_proj_b;
+        }
+    }
+
+    // key + value memory
+    {
+        const auto & hparams = model.hparams;
+
+        const int n_embd  = hparams.n_embd;
+        const int n_layer = hparams.n_layer;
+        const int n_ctx   = hparams.n_ctx;
+
+        const int64_t n_mem      = n_layer*n_ctx;
+        const int64_t n_elements = n_embd*n_mem;
+
+        model.memory_k = ggml_new_tensor_1d(ctx, GGML_TYPE_F16, n_elements);
+        model.memory_v = ggml_new_tensor_1d(ctx, GGML_TYPE_F16, n_elements);
+
+        const size_t memory_size = ggml_nbytes(model.memory_k) + ggml_nbytes(model.memory_v);
+
+        printf("%s: memory_size = %8.2f MB, n_mem = %" PRId64 "\n", __func__, memory_size/1024.0/1024.0, n_mem);
+    }
+
+    // load weights
+    {
+        int n_tensors = 0;
+        size_t total_size = 0;
+
+        printf("%s: ", __func__);
+
+        while (true) {
+            int32_t n_dims;
+            int32_t length;
+            int32_t ttype;
+
+            fin.read(reinterpret_cast<char *>(&n_dims), sizeof(n_dims));
+            fin.read(reinterpret_cast<char *>(&length), sizeof(length));
+            fin.read(reinterpret_cast<char *>(&ttype),  sizeof(ttype));
+
+            if (fin.eof()) {
+                break;
+            }
+
+            int32_t nelements = 1;
+            int32_t ne[2] = { 1, 1 };
+            for (int i = 0; i < n_dims; ++i) {
+                fin.read(reinterpret_cast<char *>(&ne[i]), sizeof(ne[i]));
+                nelements *= ne[i];
+            }
+
+            std::string name(length, 0);
+            fin.read(&name[0], length);
+
+            if (model.tensors.find(name) == model.tensors.end()) {
+                fprintf(stderr, "%s: unknown tensor '%s' in model file\n", __func__, name.c_str());
+                return false;
+            }
+
+            auto tensor = model.tensors[name];
+            if (ggml_nelements(tensor) != nelements) {
+                fprintf(stderr, "%s: tensor '%s' has wrong size in model file\n", __func__, name.c_str());
+                return false;
+            }
+
+            if (tensor->ne[0] != ne[0] || tensor->ne[1] != ne[1]) {
+                fprintf(stderr, "%s: tensor '%s' has wrong shape in model file: got [%5d, %5d], expected [%5d, %5d]\n",
+                        __func__, name.c_str(), (int) tensor->ne[0], (int) tensor->ne[1], ne[0], ne[1]);
+                return false;
+            }
+
+            // for debugging
+            if (0) {
+                printf("%24s - [%5d, %5d], type = %6s, %6.2f MB, %9zu bytes\n", name.c_str(), ne[0], ne[1], ggml_type_name(ggml_type(ttype)), ggml_nbytes(tensor)/1024.0/1024.0, ggml_nbytes(tensor));
+            }
+
+            const size_t bpe = ggml_type_size(ggml_type(ttype));
+
+            if ((nelements*bpe)/ggml_blck_size(tensor->type) != ggml_nbytes(tensor)) {
+                fprintf(stderr, "%s: tensor '%s' has wrong size in model file: got %zu, expected %zu\n",
+                        __func__, name.c_str(), ggml_nbytes(tensor), nelements*bpe);
+                return false;
+            }
+
+            fin.read(reinterpret_cast<char *>(tensor->data), ggml_nbytes(tensor));
+
+            total_size += ggml_nbytes(tensor);
+            if (++n_tensors % 8 == 0) {
+                printf(".");
+                fflush(stdout);
+            }
+        }
+
+        printf(" done\n");
+
+        printf("%s: model size = %8.2f MB / num tensors = %d\n", __func__, total_size/1024.0/1024.0, n_tensors);
+    }
+
+    fin.close();
+
+    return true;
+}
+
+// feed-forward network
+ggml_tensor * gpt_neox_ff(
+        const dollyv2_layer & layer,
+        ggml_context        * ctx0,
+        ggml_tensor         * inp,
+        float                 eps) {
+    ggml_tensor * cur = ggml_norm(ctx0, inp, eps);
+
+    cur = ggml_add(ctx0,
+        ggml_mul(ctx0,
+            ggml_repeat(ctx0, layer.ln_2_g, cur),
+            cur),
+        ggml_repeat(ctx0, layer.ln_2_b, cur));
+
+    cur = ggml_mul_mat(ctx0,
+            layer.c_mlp_fc_w,
+            cur);
+
+    cur = ggml_add(ctx0,
+            ggml_repeat(ctx0, layer.c_mlp_fc_b, cur),
+            cur);
+
+    // GELU activation
+    cur = ggml_gelu(ctx0, cur);
+
+    // projection
+    // cur = proj_w*cur + proj_b
+    cur = ggml_mul_mat(ctx0,
+            layer.c_mlp_proj_w,
+            cur);
+
+    cur = ggml_add(ctx0,
+            ggml_repeat(ctx0, layer.c_mlp_proj_b, cur),
+            cur);
+    return cur;
+}
+
+// evaluate the transformer
+//
+//   - model:     the model
+//   - n_threads: number of threads to use
+//   - n_past:    the context size so far
+//   - embd_inp:  the embeddings of the tokens in the context
+//   - embd_w:    the predicted logits for the next token
+//
+bool dollyv2_eval(
+        const dollyv2_model & model,
+        const int n_threads,
+        const int n_past,
+        const std::vector<gpt_vocab::id> & embd_inp,
+              std::vector<float>         & embd_w,
+              size_t                     & mem_per_token) {
+    const int N = embd_inp.size();
+
+    const auto & hparams = model.hparams;
+
+    const int n_embd  = hparams.n_embd;
+    const int n_layer = hparams.n_layer;
+    const int n_ctx   = hparams.n_ctx;
+    const int n_head  = hparams.n_head;
+    const int n_vocab = hparams.n_vocab;
+    const int n_rot   = hparams.n_rot;
+
+    static size_t buf_size = 256u*1024*1024;
+    static void * buf = malloc(buf_size);
+
+    if (mem_per_token > 0 && mem_per_token*N > buf_size) {
+        const size_t buf_size_new = 1.1*(mem_per_token*N); // add 10% to account for ggml object overhead
+        //printf("\n%s: reallocating buffer from %zu to %zu bytes\n", __func__, buf_size, buf_size_new);
+
+        // reallocate
+        buf_size = buf_size_new;
+        buf = realloc(buf, buf_size);
+        if (buf == nullptr) {
+            fprintf(stderr, "%s: failed to allocate %zu bytes\n", __func__, buf_size);
+            return false;
+        }
+    }
+
+    struct ggml_init_params params = {
+        /*.mem_size   =*/ buf_size,
+        /*.mem_buffer =*/ buf,
+        /*.no_alloc   =*/ false,
+    };
+
+    struct ggml_context * ctx0 = ggml_init(params);
+    struct ggml_cgraph gf = { };
+
+    // KQ_pos - contains the positions
+    struct ggml_tensor * KQ_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N);
+    int * data = (int *) KQ_pos->data;
+    for (int i = 0; i < N; ++i) {
+        data[i] = n_past + i;
+    }
+
+    struct ggml_tensor * embd = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N);
+    memcpy(embd->data, embd_inp.data(), N*ggml_element_size(embd));
+
+    // wte
+    struct ggml_tensor * inpL = ggml_get_rows(ctx0, model.wte, embd);
+
+    for (int il = 0; il < n_layer; ++il) {
+        struct ggml_tensor * cur;
+
+        // self-attention
+        {
+            {
+                cur = ggml_norm(ctx0, inpL, hparams.eps);
+
+                cur = ggml_add(ctx0,
+                        ggml_mul(ctx0,
+                            ggml_repeat(ctx0, model.layers[il].ln_1_g, cur),
+                            cur),
+                        ggml_repeat(ctx0, model.layers[il].ln_1_b, cur));
+            }
+
+            // compute QKV
+            {
+                cur = ggml_mul_mat(ctx0,
+                        model.layers[il].c_attn_attn_w,
+                        cur);
+
+                cur = ggml_add(ctx0,
+                        ggml_repeat(ctx0, model.layers[il].c_attn_attn_b, cur),
+                        cur);
+            }
+
+            struct ggml_tensor * Qcur = ggml_cont(ctx0, ggml_view_3d(ctx0, cur, n_embd/n_head, n_head, N, cur->nb[1]/n_head, cur->nb[1], 0*sizeof(float)*n_embd/n_head));
+            struct ggml_tensor * Kcur = ggml_cont(ctx0, ggml_view_3d(ctx0, cur, n_embd/n_head, n_head, N, cur->nb[1]/n_head, cur->nb[1], 1*sizeof(float)*n_embd/n_head));
+            struct ggml_tensor * Vcur = ggml_cont(ctx0, ggml_view_3d(ctx0, cur, n_embd/n_head, n_head, N, cur->nb[1]/n_head, cur->nb[1], 2*sizeof(float)*n_embd/n_head));
+
+            // using mode = 2 for GPT-NeoX mode
+            Qcur = ggml_rope_inplace(ctx0, Qcur, KQ_pos, n_rot, 2, 0);
+            Kcur = ggml_rope_inplace(ctx0, Kcur, KQ_pos, n_rot, 2, 0);
+
+            // store key and value to memory
+            {
+                Vcur = ggml_transpose(ctx0, ggml_reshape_2d(ctx0, Vcur, n_embd, N));
+
+                struct ggml_tensor * k = ggml_view_1d(ctx0, model.memory_k, N*n_embd, (ggml_element_size(model.memory_k)*n_embd)*(il*n_ctx + n_past));
+                struct ggml_tensor * v = ggml_view_2d(ctx0, model.memory_v, N, n_embd,
+                        (   n_ctx)*ggml_element_size(model.memory_v),
+                        (il*n_ctx)*ggml_element_size(model.memory_v)*n_embd + n_past*ggml_element_size(model.memory_v));
+
+                ggml_build_forward_expand(&gf, ggml_cpy(ctx0, Kcur, k));
+                ggml_build_forward_expand(&gf, ggml_cpy(ctx0, Vcur, v));
+            }
+
+            // Q = Qcur.contiguous().view(n_embd/n_head, n_head, N).permute(0, 2, 1, 3)
+            struct ggml_tensor * Q =
+                ggml_permute(ctx0,
+                        Qcur,
+                        0, 2, 1, 3);
+
+            // K = Kmem.view(n_embd/n_head, n_head, n_past + N).permute(0, 2, 1, 3)
+            struct ggml_tensor * K =
+                ggml_permute(ctx0,
+                        ggml_reshape_3d(ctx0,
+                            ggml_view_1d(ctx0, model.memory_k, (n_past + N)*n_embd, il*n_ctx*ggml_element_size(model.memory_k)*n_embd),
+                            n_embd/n_head, n_head, n_past + N),
+                        0, 2, 1, 3);
+
+            // K * Q
+            struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q);
+
+            // KQ_scaled = KQ / sqrt(n_embd/n_head)
+            struct ggml_tensor * KQ_scaled =
+                ggml_scale_inplace(ctx0,
+                        KQ,
+                        ggml_new_f32(ctx0, 1.0f/sqrt(float(n_embd)/n_head))
+                        );
+
+            // KQ_masked = mask_past(KQ_scaled)
+            struct ggml_tensor * KQ_masked = ggml_diag_mask_inf_inplace(ctx0, KQ_scaled, n_past);
+
+            // KQ = soft_max(KQ_masked)
+            struct ggml_tensor * KQ_soft_max = ggml_soft_max_inplace(ctx0, KQ_masked);
+
+            // V_trans = Vmem.view(n_embd/n_head, n_head, n_past + N).permute(1, 2, 0, 3).contiguous()
+            struct ggml_tensor * V =
+                ggml_view_3d(ctx0, model.memory_v,
+                        n_past + N, n_embd/n_head, n_head,
+                        n_ctx*ggml_element_size(model.memory_v),
+                        n_ctx*ggml_element_size(model.memory_v)*n_embd/n_head,
+                        il*n_ctx*ggml_element_size(model.memory_v)*n_embd);
+
+            // KQV = transpose(V) * KQ_soft_max
+            struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V, KQ_soft_max);
+
+            // KQV_merged = KQV.permute(0, 2, 1, 3)
+            struct ggml_tensor * KQV_merged = ggml_permute(ctx0, KQV, 0, 2, 1, 3);
+
+            // cur = KQV_merged.contiguous().view(n_embd, N)
+            cur = ggml_cpy(ctx0,
+                    KQV_merged,
+                    ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, N));
+
+            // projection
+            {
+                cur = ggml_mul_mat(ctx0,
+                        model.layers[il].c_attn_proj_w,
+                        cur);
+
+                cur = ggml_add(ctx0, ggml_repeat(ctx0, model.layers[il].c_attn_proj_b, cur), cur);
+            }
+        }
+
+        if (hparams.par_res == 0) {
+            struct ggml_tensor * inpFF = ggml_add(ctx0, cur, inpL);
+
+            cur = gpt_neox_ff(model.layers[il], ctx0, inpFF, hparams.eps);
+
+            // input for next layer
+            inpL = ggml_add(ctx0, cur, inpFF);
+        } else {
+            struct ggml_tensor * inpFF = cur;
+
+            // this is independent of the self-attention result, so it could be done in parallel to the self-attention
+            // note here we pass inpL instead of cur
+            cur = gpt_neox_ff(model.layers[il], ctx0, inpL, hparams.eps);
+
+            // layer input + FF
+            cur  = ggml_add(ctx0, cur, inpFF);
+
+            // input for next layer
+            inpL = ggml_add(ctx0, cur, inpL);
+        }
+
+    }
+
+    // norm
+    {
+        inpL = ggml_norm(ctx0, inpL, hparams.eps);
+
+        // inpL = ln_f_g*inpL + ln_f_b
+        inpL = ggml_add(ctx0,
+                ggml_mul(ctx0,
+                    ggml_repeat(ctx0, model.ln_f_g, inpL),
+                    inpL),
+                ggml_repeat(ctx0, model.ln_f_b, inpL));
+    }
+
+    // lm_head
+    {
+        inpL = ggml_mul_mat(ctx0, model.lmh_g, inpL);
+
+        //inpL = ggml_add(ctx0,
+        //        ggml_repeat(ctx0, model.lmh_b, inpL),
+        //        inpL);
+    }
+
+    // logits -> probs
+    //inpL = ggml_soft_max_inplace(ctx0, inpL);
+
+    // run the computation
+    ggml_build_forward_expand(&gf, inpL);
+    ggml_graph_compute_with_ctx(ctx0, &gf, n_threads);
+
+    //if (n_past%100 == 0) {
+    //    ggml_graph_print   (&gf);
+    //    ggml_graph_dump_dot(&gf, NULL, "gpt-2.dot");
+    //}
+
+    //embd_w.resize(n_vocab*N);
+    //memcpy(embd_w.data(), ggml_get_data(inpL), sizeof(float)*n_vocab*N);
+
+    // return result for just the last token
+    embd_w.resize(n_vocab);
+    memcpy(embd_w.data(), (float *) ggml_get_data(inpL) + (n_vocab*(N-1)), sizeof(float)*n_vocab);
+
+    if (mem_per_token == 0) {
+        mem_per_token = ggml_used_mem(ctx0)/N;
+    }
+    //printf("used_mem = %zu\n", ggml_used_mem(ctx0));
+
+    ggml_free(ctx0);
+
+    return true;
+}
+
+std::string execute_prompt(
+        const dollyv2_model &model,
+        gpt_vocab &vocab,
+        const std::string &prompt,
+        gpt_params &params,
+        std::mt19937 &rng,
+        int64_t t_load_us,
+        int64_t t_sample_us,
+        int64_t t_predict_us,
+        size_t mem_per_token,
+        int n_past,
+        bool stream_response_to_cout = false) {
+    std::string output = "";
+    std::vector<float> logits;
+
+    // tokenize the prompt
+    std::vector<gpt_vocab::id> embd_inp = ::gpt_tokenize(vocab, prompt);
+
+    params.n_predict = std::min(params.n_predict, model.hparams.n_ctx - (int)embd_inp.size());
+
+    printf("%s: number of tokens in prompt = %zu\n", __func__, embd_inp.size());
+    for (size_t i = 0; i < embd_inp.size(); i++) {
+        printf("%s: token[%zu] = %6d, %s\n", __func__, i, embd_inp[i], vocab.id_to_token.at(embd_inp[i]).c_str());
+    }
+    printf("\n");
+
+    std::vector<gpt_vocab::id> embd;
+
+    dollyv2_eval(model, params.n_threads, 0, {0, 1, 2, 3}, logits, mem_per_token);
+
+    const int32_t end_token = vocab.token_to_id["### End"];
+
+    for (size_t i = embd.size(); i < embd_inp.size() + params.n_predict; i++) {
+        // predict
+        if (embd.size() > 0) {
+            const int64_t t_start_us = ggml_time_us();
+
+            if (!dollyv2_eval(model, params.n_threads, n_past, embd, logits, mem_per_token)) {
+                printf("Failed to predict\n");
+                return output;
+            }
+
+            t_predict_us += ggml_time_us() - t_start_us;
+        }
+
+        n_past += embd.size();
+        embd.clear();
+
+        if (i >= embd_inp.size()) {
+            // sample next token
+            const int top_k = params.top_k;
+            const float top_p = params.top_p;
+            const float temp = params.temp;
+
+            const int n_vocab = model.hparams.n_vocab;
+
+            gpt_vocab::id id = 0;
+
+            {
+                const int64_t t_start_sample_us = ggml_time_us();
+
+                id = gpt_sample_top_k_top_p(vocab, logits.data() + (logits.size() - n_vocab), top_k, top_p, temp, rng);
+
+                t_sample_us += ggml_time_us() - t_start_sample_us;
+            }
+
+            // add it to the context
+            embd.push_back(id);
+        } else {
+            // if here, it means we are still processing the input prompt
+            for (size_t k = i; k < embd_inp.size(); k++) {
+                embd.push_back(embd_inp[k]);
+                if (int32_t(embd.size()) > params.n_batch) {
+                    break;
+                }
+            }
+            i += embd.size() - 1;
+        }
+
+        // display text
+        for (auto id : embd) {
+            output += vocab.id_to_token[id];
+            if (stream_response_to_cout) {
+                printf("%s", vocab.id_to_token[id].c_str());
+            }
+        }
+        if (stream_response_to_cout) {
+            fflush(stdout);
+        }
+
+        // end of text token
+        if (embd.back() == 0 || (end_token > 0 && embd.back() == end_token)) {
+            return output;
+        }
+    }
+    return output;
+}
+
+#if defined(DOLLY_INTERACTIVE_PORT)
+int setup_port(const int port) {
+    int sockfd = socket(AF_INET, SOCK_STREAM, 0);
+    if (sockfd < 0) {
+        fprintf(stderr, "%s: Failed to create new socket\n", __func__);
+        return -1;
+    }
+
+    sockaddr_in servaddr;
+    std::memset(&servaddr, 0, sizeof(servaddr));
+
+    servaddr.sin_family = AF_INET;
+    servaddr.sin_addr.s_addr = htonl(INADDR_ANY);
+    servaddr.sin_port = htons(port);
+
+    if (bind(sockfd, (struct sockaddr *)&servaddr, sizeof(servaddr)) < 0) {
+        fprintf(stderr, "%s: Failed to bind to port %i\n", __func__, port);
+        return -1;
+    }
+
+    if (listen(sockfd, 10) < 0) {
+        fprintf(stderr, "%s: Failed to listen to socket on port %i\n", __func__, port);
+        return -1;
+    }
+    return sockfd;
+}
+
+std::string read_from_port(int sockfd, int clientfd) {
+    if (clientfd < 0) {
+        fprintf(stderr, "%s: Failed to accept new connection\n", __func__);
+        return "";
+    }
+
+    char buffer[4096];
+    std::memset(buffer, 0, sizeof(buffer));
+
+    if (read(clientfd, buffer, sizeof(buffer)) < 0) {
+        fprintf(stderr, "%s: Failed to read from client\n", __func__);
+    } else {
+        std::cout << "Received: " << buffer;
+        return std::string(buffer);
+    }
+    return std::string("");
+}
+#endif
+
+int main(int argc, char ** argv) {
+    ggml_time_init();
+
+    const int64_t t_main_start_us = ggml_time_us();
+
+    gpt_params params;
+    params.model = "models/dolly-v2-3b/ggml-model-f16.bin";
+
+    if (gpt_params_parse(argc, argv, params) == false) {
+        return 1;
+    }
+
+    if (params.seed < 0) {
+        params.seed = time(NULL);
+    }
+
+    printf("%s: seed = %d\n", __func__, params.seed);
+
+    std::mt19937 rng(params.seed);
+
+    int64_t t_load_us = 0;
+    int64_t t_sample_us = 0;
+    int64_t t_predict_us = 0;
+
+    // determine the required inference memory per token:
+    size_t mem_per_token = 0;
+
+    int n_past = 0;
+
+    gpt_vocab vocab;
+    dollyv2_model model;
+
+    // load the model
+    {
+        const int64_t t_start_us = ggml_time_us();
+
+        if (!dollyv2_model_load(params.model, model, vocab)) {
+            fprintf(stderr, "%s: failed to load model from '%s'\n", __func__, params.model.c_str());
+            return 1;
+        }
+
+        t_load_us = ggml_time_us() - t_start_us;
+
+        test_gpt_tokenizer(vocab, params.token_test);
+    }
+
+#if defined(DOLLY_INTERACTIVE_PORT)
+    int sockfd = -1;
+    if (params.interactive_port != -1) {
+        sockfd = setup_port(params.interactive_port);
+        if (sockfd == -1) {
+            return 1;
+        }
+        fprintf(stdout, "Model is ready on port %i\n", params.interactive_port);
+        fflush(stdout);
+    }
+#endif
+
+    if (params.interactive || params.interactive_port != -1) {
+        while (true) {
+            std::string prompt_input;
+#if defined(DOLLY_INTERACTIVE_PORT)
+            int clientfd = -1;
+            if (params.interactive_port != -1) {
+                sockaddr_in clientaddr;
+                socklen_t clientaddrlen = sizeof(clientaddr);
+                clientfd = accept(sockfd, (struct sockaddr *)&clientaddr, &clientaddrlen);
+                prompt_input = read_from_port(sockfd, clientfd);
+            } else
+#endif
+            {
+                printf("Please enter your quesiton:\n>");
+                fflush(stdout);
+
+                std::getline(std::cin, prompt_input);
+            }
+
+            if (strcmp(prompt_input.c_str(), "exit") == 0) {
+                break;
+            }
+
+            const std::string prompt = prompt_for_generation(prompt_input);
+            // call the model
+            const std::string response = execute_prompt(model, vocab, prompt, params, rng, t_load_us, t_sample_us, t_predict_us, mem_per_token, n_past, true);
+
+#if defined(DOLLY_INTERACTIVE_PORT)
+            if (params.interactive_port != -1) {
+                if (write(clientfd, response.c_str(), response.size()) < 0) {
+                    fprintf(stderr, "%s: Failed to write answer '%s' to client\n", __func__, response.c_str());
+                }
+
+                if (close(clientfd) < 0) {
+                    fprintf(stderr, "%s: Failed to close client socket\n", __func__);
+                }
+            } else
+#endif
+            {
+                printf("%s\n\n", response.c_str());
+            }
+            fflush(stdout);
+        }
+    } else {
+        if (params.prompt.empty()) {
+            params.prompt = gpt_random_prompt(rng);
+        }
+
+        const std::string prompt = prompt_for_generation(params.prompt);
+        execute_prompt(model, vocab, prompt, params, rng, t_load_us, t_sample_us, t_predict_us, mem_per_token, n_past, true);
+    }
+
+    // report timing
+    {
+        const int64_t t_main_end_us = ggml_time_us();
+
+        printf("\n\n");
+        printf("%s: mem per token = %8zu bytes\n", __func__, mem_per_token);
+        printf("%s:     load time = %8.2f ms\n", __func__, t_load_us / 1000.0f);
+        printf("%s:   sample time = %8.2f ms\n", __func__, t_sample_us / 1000.0f);
+        printf("%s:  predict time = %8.2f ms / %.2f ms per token\n", __func__, t_predict_us / 1000.0f, t_predict_us / 1000.0f / n_past);
+        printf("%s:    total time = %8.2f ms\n", __func__, (t_main_end_us - t_main_start_us) / 1000.0f);
+    }
+
+    ggml_free(model.ctx);
+
+#if defined(DOLLY_INTERACTIVE_PORT)
+    if (params.interactive_port != -1 && close(sockfd) < 0) {
+        fprintf(stderr, "%s: Failed to close server socket\n", __func__);
+    }
+#endif
+
+    return 0;
+}

stable-diffusion.cpp/ggml/examples/dolly-v2/quantize.cpp

@@ -0,0 +1,178 @@
+#include "ggml/ggml.h"
+
+#include "common.h"
+#include "common-ggml.h"
+
+#include <cassert>
+#include <cmath>
+#include <cstdio>
+#include <cstring>
+#include <fstream>
+#include <map>
+#include <string>
+#include <vector>
+#include <regex>
+
+// default hparams (dollyv2 3B)
+struct dollyv2_hparams {
+    int32_t n_vocab = 50254; // tokenizer.vocab_size
+    int32_t n_ctx   = 2048;  // model.config.max_position_embeddings
+    int32_t n_embd  = 2560;  // model.config.hidden_size
+    int32_t n_head  = 32;    // model.config.num_attention_heads
+    int32_t n_layer = 32;    // model.config.num_hidden_layers
+    int32_t n_rot   = 20;    // rotary_pct[25%] * (n_embd / n_head)
+    int32_t par_res = 1; // 1 = true, 0 = false
+    int32_t ftype   = GGML_FTYPE_MOSTLY_F16;
+};
+
+// quantize a model
+bool dollyv2_model_quantize(const std::string & fname_inp, const std::string & fname_out, ggml_ftype ftype) {
+    gpt_vocab vocab;
+
+    printf("%s: loading model from '%s'\n", __func__, fname_inp.c_str());
+
+    auto finp = std::ifstream(fname_inp, std::ios::binary);
+    if (!finp) {
+        fprintf(stderr, "%s: failed to open '%s' for reading\n", __func__, fname_inp.c_str());
+        return false;
+    }
+
+    auto fout = std::ofstream(fname_out, std::ios::binary);
+    if (!fout) {
+        fprintf(stderr, "%s: failed to open '%s' for writing\n", __func__, fname_out.c_str());
+        return false;
+    }
+
+    // verify magic
+    {
+        uint32_t magic;
+        finp.read((char *) &magic, sizeof(magic));
+        if (magic != GGML_FILE_MAGIC) {
+            fprintf(stderr, "%s: invalid model file '%s' (bad magic)\n", __func__, fname_inp.c_str());
+            return false;
+        }
+
+        fout.write((char *) &magic, sizeof(magic));
+    }
+
+    dollyv2_hparams hparams;
+
+    // load hparams
+    {
+        finp.read((char *) &hparams.n_vocab, sizeof(hparams.n_vocab));
+        finp.read((char *) &hparams.n_ctx,   sizeof(hparams.n_ctx));
+        finp.read((char *) &hparams.n_embd,  sizeof(hparams.n_embd));
+        finp.read((char *) &hparams.n_head,  sizeof(hparams.n_head));
+        finp.read((char *) &hparams.n_layer, sizeof(hparams.n_layer));
+        finp.read((char *) &hparams.n_rot,   sizeof(hparams.n_rot));
+        finp.read((char *) &hparams.par_res, sizeof(hparams.par_res));
+        finp.read((char *) &hparams.ftype,   sizeof(hparams.ftype));
+
+        const int32_t qntvr_src =    hparams.ftype / GGML_QNT_VERSION_FACTOR;
+        const int32_t ftype_dst = GGML_QNT_VERSION * GGML_QNT_VERSION_FACTOR + ftype;
+
+        printf("%s: n_vocab     = %d\n", __func__, hparams.n_vocab);
+        printf("%s: n_ctx       = %d\n", __func__, hparams.n_ctx);
+        printf("%s: n_embd      = %d\n", __func__, hparams.n_embd);
+        printf("%s: n_head      = %d\n", __func__, hparams.n_head);
+        printf("%s: n_layer     = %d\n", __func__, hparams.n_layer);
+        printf("%s: par_res     = %d\n", __func__, hparams.par_res);
+        printf("%s: ftype (src) = %d\n", __func__, hparams.ftype);
+        printf("%s: qntvr (src) = %d\n", __func__, qntvr_src);
+        printf("%s: ftype (dst) = %d\n", __func__, ftype_dst);
+        printf("%s: qntvr (dst) = %d\n", __func__, GGML_QNT_VERSION);
+
+        fout.write((char *) &hparams.n_vocab, sizeof(hparams.n_vocab));
+        fout.write((char *) &hparams.n_ctx,   sizeof(hparams.n_ctx));
+        fout.write((char *) &hparams.n_embd,  sizeof(hparams.n_embd));
+        fout.write((char *) &hparams.n_head,  sizeof(hparams.n_head));
+        fout.write((char *) &hparams.n_layer, sizeof(hparams.n_layer));
+        fout.write((char *) &hparams.n_rot,   sizeof(hparams.n_rot));
+        fout.write((char *) &hparams.par_res, sizeof(hparams.par_res));
+        fout.write((char *) &ftype_dst,       sizeof(ftype_dst));
+    }
+
+    // load vocab
+    {
+        const int32_t n_vocab = hparams.n_vocab;
+
+        std::string word;
+        for (int i = 0; i < n_vocab; i++) {
+            uint32_t len;
+            finp.read ((char *) &len, sizeof(len));
+            fout.write((char *) &len, sizeof(len));
+
+            word.resize(len);
+            finp.read ((char *) word.data(), len);
+            fout.write((char *) word.data(), len);
+
+            vocab.token_to_id[word] = i;
+            vocab.id_to_token[i] = word;
+        }
+    }
+
+    // regexes of tensor names to be quantized
+    const std::vector<std::string> to_quant = {
+        ".*weight",
+    };
+
+    if (!ggml_common_quantize_0(finp, fout, ftype, to_quant, {})) {
+        fprintf(stderr, "%s: failed to quantize model '%s'\n", __func__, fname_inp.c_str());
+        return false;
+    }
+
+    finp.close();
+    fout.close();
+
+    return true;
+}
+
+// usage:
+//  ./dollyv2-quantize models/dolly-v2-3B/ggml-model.bin models/dolly-v2-3B/ggml-model-quant.bin type
+//
+int main(int argc, char ** argv) {
+    if (argc != 4) {
+        fprintf(stderr, "usage: %s model-f32.bin model-quant.bin type\n", argv[0]);
+        ggml_print_ftypes(stderr);
+        return 1;
+    }
+
+    // needed to initialize f16 tables
+    {
+        struct ggml_init_params params = { 0, NULL, false };
+        struct ggml_context * ctx = ggml_init(params);
+        ggml_free(ctx);
+    }
+
+    const std::string fname_inp = argv[1];
+    const std::string fname_out = argv[2];
+
+    const ggml_ftype ftype = ggml_parse_ftype(argv[3]);
+
+    const int64_t t_main_start_us = ggml_time_us();
+
+    int64_t t_quantize_us = 0;
+
+    // load the model
+    {
+        const int64_t t_start_us = ggml_time_us();
+
+        if (!dollyv2_model_quantize(fname_inp, fname_out, ggml_ftype(ftype))) {
+            fprintf(stderr, "%s: failed to quantize model from '%s'\n", __func__, fname_inp.c_str());
+            return 1;
+        }
+
+        t_quantize_us = ggml_time_us() - t_start_us;
+    }
+
+    // report timing
+    {
+        const int64_t t_main_end_us = ggml_time_us();
+
+        printf("\n");
+        printf("%s: quantize time = %8.2f ms\n", __func__, t_quantize_us/1000.0f);
+        printf("%s:    total time = %8.2f ms\n", __func__, (t_main_end_us - t_main_start_us)/1000.0f);
+    }
+
+    return 0;
+}

stable-diffusion.cpp/ggml/examples/gpt-2/CMakeLists.txt

@@ -0,0 +1,36 @@
+#
+# gpt-2
+
+set(TEST_TARGET gpt-2)
+add_executable(${TEST_TARGET} main.cpp)
+target_link_libraries(${TEST_TARGET} PRIVATE ggml common common-ggml)
+
+#
+# gpt-2-quantize
+
+set(TEST_TARGET gpt-2-quantize)
+add_executable(${TEST_TARGET} quantize.cpp)
+target_link_libraries(${TEST_TARGET} PRIVATE ggml common common-ggml)
+
+#
+# gpt-2-batched
+
+set(TEST_TARGET gpt-2-batched)
+add_executable(${TEST_TARGET} main-batched.cpp)
+target_link_libraries(${TEST_TARGET} PRIVATE ggml common common-ggml)
+
+
+#
+# For GPU offloading
+
+if (GGML_CUBLAS)
+    add_compile_definitions(GGML_USE_CUBLAS)
+endif()
+
+if (GGML_CLBLAST)
+    add_compile_definitions(GGML_USE_CLBLAST)
+endif()
+
+if (GGML_METAL)
+    add_compile_definitions(GGML_USE_METAL)
+endif()

stable-diffusion.cpp/ggml/examples/gpt-2/README.md

@@ -0,0 +1,225 @@
+# gpt-2
+
+This is a C++ example running GPT-2 inference using the [ggml](https://github.com/ggerganov/ggml) library.
+
+The program runs on the CPU - no video card is required.
+
+The [Cerebras-GPT](https://huggingface.co/cerebras) models are also supported.
+
+The example supports the following GPT-2 models:
+
+| Model | Description  | Disk Size |
+| ---   | ---          | ---       |
+| 117M  | Small model  | 240 MB    |
+| 345M  | Medium model | 680 MB    |
+| 774M  | Large model  | 1.5 GB    |
+| 1558M | XL model     | 3.0 GB    |
+
+Sample performance on MacBook M1 Pro:
+
+| Model | Size  | Time / Token |
+| ---   | ---   | ---    |
+| GPT-2 |  117M |   5 ms |
+| GPT-2 |  345M |  12 ms |
+| GPT-2 |  774M |  23 ms |
+| GPT-2 | 1558M |  42 ms |
+
+*TODO: add tables for Cerebras-GPT models*
+
+Sample output:
+
+```
+$ ./bin/gpt-2 -h
+usage: ./bin/gpt-2 [options]
+
+options:
+  -h, --help            show this help message and exit
+  -s SEED, --seed SEED  RNG seed (default: -1)
+  -t N, --threads N     number of threads to use during computation (default: 8)
+  -p PROMPT, --prompt PROMPT
+                        prompt to start generation with (default: random)
+  -n N, --n_predict N   number of tokens to predict (default: 200)
+  --top_k N             top-k sampling (default: 40)
+  --top_p N             top-p sampling (default: 0.9)
+  --temp N              temperature (default: 1.0)
+  -b N, --batch_size N  batch size for prompt processing (default: 8)
+  -m FNAME, --model FNAME
+                        model path (default: models/gpt-2-117M/ggml-model.bin)
+
+$ ./bin/gpt-2
+gpt2_model_load: loading model from 'models/gpt-2-117M/ggml-model.bin'
+gpt2_model_load: n_vocab = 50257
+gpt2_model_load: n_ctx   = 1024
+gpt2_model_load: n_embd  = 768
+gpt2_model_load: n_head  = 12
+gpt2_model_load: n_layer = 12
+gpt2_model_load: f16     = 1
+gpt2_model_load: ggml ctx size = 311.12 MB
+gpt2_model_load: memory size =    72.00 MB, n_mem = 12288
+gpt2_model_load: model size  =   239.08 MB
+main: number of tokens in prompt = 1
+
+So this is going to be the end of the line for us.
+
+If the Dolphins continue to do their business, it's possible that the team could make a bid to bring in new defensive coordinator Scott Linehan.
+
+Linehan's job is a little daunting, but he's a great coach and an excellent coach. I don't believe we're going to make the playoffs.
+
+We're going to have to work hard to keep our heads down and get ready to go.<|endoftext|>
+
+main: mem per token =  2048612 bytes
+main:     load time =   106.32 ms
+main:   sample time =     7.10 ms
+main:  predict time =   506.40 ms / 5.06 ms per token
+main:    total time =   629.84 ms
+```
+
+## Downloading and converting the original models (GPT-2)
+
+You can download the original model files using the [download-model.sh](download-model.sh) Bash script. The models are
+in Tensorflow format, so in order to use them with ggml, you need to convert them to appropriate format. This is done
+via the [convert-ckpt-to-ggml.py](convert-ckpt-to-ggml.py) python script.
+
+Here is the entire process for the GPT-2 117M model (download from official site + conversion):
+
+```
+cd ggml/build
+../examples/gpt-2/download-model.sh 117M
+
+Downloading model 117M ...
+models/gpt-2-117M/checkpoint                      100%[=============================>]      77  --.-KB/s    in 0s
+models/gpt-2-117M/encoder.json                    100%[=============================>]   1018K  1.20MB/s    in 0.8s
+models/gpt-2-117M/hparams.json                    100%[=============================>]      90  --.-KB/s    in 0s
+models/gpt-2-117M/model.ckpt.data-00000-of-00001  100%[=============================>] 474.70M  1.21MB/s    in 8m 39s
+models/gpt-2-117M/model.ckpt.index                100%[=============================>]   5.09K  --.-KB/s    in 0s
+models/gpt-2-117M/model.ckpt.meta                 100%[=============================>] 460.11K   806KB/s    in 0.6s
+models/gpt-2-117M/vocab.bpe                       100%[=============================>] 445.62K   799KB/s    in 0.6s
+Done! Model '117M' saved in 'models/gpt-2-117M/'
+
+Run the convert-ckpt-to-ggml.py script to convert the model to ggml format.
+
+  python /Users/john/ggml/examples/gpt-2/convert-ckpt-to-ggml.py models/gpt-2-117M/ 1
+
+```
+
+This conversion requires that you have python and Tensorflow installed on your computer. Still, if you want to avoid
+this, you can download the already converted ggml models as described below.
+
+## Downloading and converting the original models (Cerebras-GPT)
+
+Clone the respective repository from here: https://huggingface.co/cerebras
+
+Use the [convert-cerebras-to-ggml.py](convert-cerebras-to-ggml.py) script to convert the model to `ggml` format:
+
+```
+cd ggml/build
+git clone https://huggingface.co/cerebras/Cerebras-GPT-111M models/
+python ../examples/gpt-2/convert-cerebras-to-ggml.py models/Cerebras-GPT-111M/
+
+```
+
+## Downloading the ggml model directly (GPT-2)
+
+For convenience, I will be hosting the converted ggml model files in order to make it easier to run the examples. This
+way, you can directly download a single binary file and start using it. No python or Tensorflow is required.
+
+Here is how to get the 117M ggml model:
+
+```
+cd ggml/build
+../examples/gpt-2/download-ggml-model.sh 117M
+
+Downloading ggml model 117M ...
+models/gpt-2-117M/ggml-model.bin         100%[===============================>] 239.58M  8.52MB/s    in 28s
+Done! Model '117M' saved in 'models/gpt-2-117M/ggml-model.bin'
+You can now use it like this:
+
+  $ ./bin/gpt-2 -m models/gpt-2-117M/ggml-model.bin -p "This is an example"
+
+```
+
+At some point, I might decide to stop hosting these models. So in that case, simply revert to the manual process above.
+
+## Quantizing the models
+
+You can also try to quantize the `ggml` models via 4-bit integer quantization.
+Keep in mind that for smaller models, this will render them completely useless.
+You generally want to quantize larger models.
+
+```
+# quantize GPT-2 F16 to Q4_0 (faster but less precise)
+./bin/gpt-2-quantize models/gpt-2-1558M/ggml-model-f16.bin models/gpt-2-1558M/ggml-model-q4_0.bin 2
+./bin/gpt-2 -m models/gpt-2-1558M/ggml-model-q4_0.bin -p "This is an example"
+
+# quantize Cerebras F16 to Q4_1 (slower but more precise)
+./bin/gpt-2-quantize models/Cerebras-GPT-6.7B/ggml-model-f16.bin models/Cerebras-GPT-6.7B/ggml-model-q4_1.bin 3
+./bin/gpt-2 -m models/Cerebras-GPT-6.7B/ggml-model-q4_1.bin -p "This is an example"
+
+```
+
+## Batched generation example
+
+You can try the batched generation from a given prompt using the gpt-2-batched binary.
+
+Sample output:
+
+```
+$ gpt-2-batched -np 5 -m models/gpt-2-117M/ggml-model.bin -p "Hello my name is" -n 50
+
+main: seed = 1697037431
+gpt2_model_load: loading model from 'models/gpt-2-117M/ggml-model.bin'
+gpt2_model_load: n_vocab = 50257
+gpt2_model_load: n_ctx   = 1024
+gpt2_model_load: n_embd  = 768
+gpt2_model_load: n_head  = 12
+gpt2_model_load: n_layer = 12
+gpt2_model_load: ftype   = 1
+gpt2_model_load: qntvr   = 0
+gpt2_model_load: ggml tensor size    = 320 bytes
+gpt2_model_load: backend buffer size = 312.72 MB
+ggml_init_cublas: found 1 CUDA devices:
+  Device 0: NVIDIA GeForce GTX 1660, compute capability 7.5
+gpt2_model_load: using CPU backend
+gpt2_model_load: memory size =    72.00 MB, n_mem = 12288
+gpt2_model_load: model size  =   239.08 MB
+extract_tests_from_file : No test file found.
+test_gpt_tokenizer : 0 tests failed out of 0 tests.
+main: compute buffer size: 3.26 MB
+
+
+main: generating 5 sequences ...
+main: prompt: 'Hello my name is'
+main: number of tokens in prompt = 4, first 8 tokens: 15496 616 1438 318
+
+
+sequence 0:
+
+Hello my name is John. You can call me any way you want, if you want, but for my very first date, I will be on the phone with you. We're both in our early 20s, but I feel like it's all
+
+sequence 1:
+
+Hello my name is Robert, and I want to say that we're proud to have your company here on the world's largest platform for sharing your stories with us. This is a huge opportunity for our community. We have hundreds of people on this team and
+
+sequence 2:
+
+Hello my name is Jack. I'm the one who created you.
+
+Jack is a boy with a big smile and a big heart. He is a handsome guy. He loves the outdoors and loves the people he meets. He wants to be a
+
+sequence 3:
+
+Hello my name is John. I am a Canadian citizen with a large number of family in Quebec and I am interested in studying. My aim is to take up a post in the Journal of the International Academy of Sciences of Canada which I am currently finishing.
+
+sequence 4:
+
+Hello my name is Dan. I am an entrepreneur. I am a great father. I am a great husband. I am a great husband. I am a great dad. And I am a great husband.
+
+I love my life. I love
+
+
+
+main:     load time =   880.80 ms
+main:   sample time =    91.43 ms
+main:  predict time =  2518.29 ms
+main:    total time =  3544.32 ms
+```

stable-diffusion.cpp/ggml/examples/gpt-2/convert-cerebras-to-ggml.py

@@ -0,0 +1,183 @@
+# Convert Cerebras models to ggml format
+#
+# ref: https://www.cerebras.net/blog/cerebras-gpt-a-family-of-open-compute-efficient-large-language-models/
+#
+
+import sys
+import struct
+import json
+import torch
+import numpy as np
+import re
+
+from transformers import AutoModelForCausalLM
+
+# ref: https://github.com/openai/gpt-2/blob/master/src/encoder.py
+def bytes_to_unicode():
+    """
+    Returns list of utf-8 byte and a corresponding list of unicode strings.
+    The reversible bpe codes work on unicode strings.
+    This means you need a large # of unicode characters in your vocab if you want to avoid UNKs.
+    When you're at something like a 10B token dataset you end up needing around 5K for decent coverage.
+    This is a signficant percentage of your normal, say, 32K bpe vocab.
+    To avoid that, we want lookup tables between utf-8 bytes and unicode strings.
+    And avoids mapping to whitespace/control characters the bpe code barfs on.
+    """
+    bs = list(range(ord("!"), ord("~")+1))+list(range(ord("¡"), ord("¬")+1))+list(range(ord("®"), ord("ÿ")+1))
+    cs = bs[:]
+    n = 0
+    for b in range(2**8):
+        if b not in bs:
+            bs.append(b)
+            cs.append(2**8+n)
+            n += 1
+    cs = [chr(n) for n in cs]
+    return dict(zip(bs, cs))
+
+if len(sys.argv) < 2:
+    print("Usage: convert-cerebras-to-ggml.py dir-model [use-f32]\n")
+    sys.exit(1)
+
+# output in the same directory as the model
+dir_model = sys.argv[1]
+fname_out = sys.argv[1] + "/ggml-model-f16.bin"
+
+with open(dir_model + "/vocab.json", "r", encoding="utf-8") as f:
+    encoder = json.load(f)
+
+with open(dir_model + "/config.json", "r", encoding="utf-8") as f:
+    hparams = json.load(f)
+
+# use 16-bit or 32-bit floats
+use_f16 = True
+if len(sys.argv) > 2:
+    use_f16 = False
+    fname_out = sys.argv[1] + "/ggml-model-f32.bin"
+
+model = AutoModelForCausalLM.from_pretrained(dir_model, low_cpu_mem_usage=True)
+#print (model)
+
+list_vars = model.state_dict()
+#print (list_vars)
+
+print(hparams)
+
+fout = open(fname_out, "wb")
+
+fout.write(struct.pack("i", 0x67676d6c)) # magic: ggml in hex
+fout.write(struct.pack("i", hparams["vocab_size"]))
+fout.write(struct.pack("i", hparams["n_positions"]))
+fout.write(struct.pack("i", hparams["n_embd"]))
+fout.write(struct.pack("i", hparams["n_head"]))
+fout.write(struct.pack("i", hparams["n_layer"]))
+fout.write(struct.pack("i", use_f16))
+
+byte_encoder = bytes_to_unicode()
+byte_decoder = {v:k for k, v in byte_encoder.items()}
+
+fout.write(struct.pack("i", len(encoder)))
+
+for key in encoder:
+    text = bytearray([byte_decoder[c] for c in key])
+    fout.write(struct.pack("i", len(text)))
+    fout.write(text)
+
+for name in list_vars.keys():
+    data = list_vars[name].squeeze().numpy()
+    print("Processing variable: " + name + " with shape: ", data.shape)
+
+    # rename headers to keep compatibility
+    if name == "transformer.ln_f.weight":
+        name = "model/ln_f/g"
+    elif name == "transformer.ln_f.bias":
+        name = "model/ln_f/b"
+    elif name == "transformer.wte.weight":
+        name = "model/wte"
+    elif name == "transformer.wpe.weight":
+        name = "model/wpe"
+    elif name == "lm_head.weight":
+        name = "model/lm_head"
+    elif re.match(r"transformer.h\.\d+\.ln_1\.weight", name):
+        i = re.findall("\d+", name)[0]
+        name = f"model/h{i}/ln_1/g"
+    elif re.match(r"transformer.h\.\d+\.ln_1\.bias", name):
+        i = re.findall("\d+", name)[0]
+        name = f"model/h{i}/ln_1/b"
+    elif re.match(r"transformer.h\.\d+\.attn\.c_attn\.weight", name):
+        i = re.findall("\d+", name)[0]
+        name = f"model/h{i}/attn/c_attn/w"
+    elif re.match(r"transformer.h\.\d+\.attn\.c_attn\.bias", name):
+        i = re.findall("\d+", name)[0]
+        name = f"model/h{i}/attn/c_attn/b"
+    elif re.match(r"transformer.h\.\d+\.attn\.c_proj\.weight", name):
+        i = re.findall("\d+", name)[0]
+        name = f"model/h{i}/attn/c_proj/w"
+    elif re.match(r"transformer.h.\d+.attn.c_proj.bias", name):
+        i = re.findall("\d+", name)[0]
+        name = f"model/h{i}/attn/c_proj/b"
+    elif re.match(r"transformer.h.\d+.ln_2.weight", name):
+        i = re.findall("\d+", name)[0]
+        name = f"model/h{i}/ln_2/g"
+    elif re.match(r"transformer.h.\d+.ln_2.bias", name):
+        i = re.findall("\d+", name)[0]
+        name = f"model/h{i}/ln_2/b"
+    elif re.match(r"transformer.h.\d+.mlp.c_fc.weight", name):
+        i = re.findall("\d+", name)[0]
+        name = f"model/h{i}/mlp/c_fc/w"
+    elif re.match(r"transformer.h.\d+.mlp.c_fc.bias", name):
+        i = re.findall("\d+", name)[0]
+        name = f"model/h{i}/mlp/c_fc/b"
+    elif re.match(r"transformer.h.\d+.mlp.c_proj.weight", name):
+        i = re.findall("\d+", name)[0]
+        name = f"model/h{i}/mlp/c_proj/w"
+    elif re.match(r"transformer.h.\d+.mlp.c_proj.bias", name):
+        i = re.findall("\d+", name)[0]
+        name = f"model/h{i}/mlp/c_proj/b"
+    else:
+        print("Unrecognized variable name. %s", name)
+
+    # we don't need these
+    if name.endswith("attn.masked_bias") or name.endswith(".attn.bias"):
+        print("  Skipping variable: " + name)
+        continue
+
+    n_dims = len(data.shape);
+
+    # ftype == 0 -> float32, ftype == 1 -> float16
+    ftype = 0;
+    if use_f16:
+        if (name == "model/wte" or name == "model/lm_head" or name[-2:] == "/g" or name[-2:] == "/w") and n_dims == 2:
+            print("  Converting to float16")
+            data = data.astype(np.float16)
+            ftype = 1
+        else:
+            print("  Converting to float32")
+            data = data.astype(np.float32)
+            ftype = 0
+
+    # for efficiency - transpose the projection matrices
+    # "model/h.*/attn/c_attn/w"
+    # "model/h.*/attn/c_proj/w"
+    # "model/h.*/mlp/c_fc/w"
+    # "model/h.*/mlp/c_proj/w"
+    if name[-14:] == "/attn/c_attn/w" or \
+       name[-14:] == "/attn/c_proj/w" or \
+       name[-11:] == "/mlp/c_fc/w" or \
+       name[-13:] == "/mlp/c_proj/w":
+        print("  Transposing")
+        data = data.transpose()
+
+    # header
+    str = name.encode('utf-8')
+    fout.write(struct.pack("iii", n_dims, len(str), ftype))
+    for i in range(n_dims):
+        fout.write(struct.pack("i", data.shape[n_dims - 1 - i]))
+    fout.write(str);
+
+    # data
+    data.tofile(fout)
+
+fout.close()
+
+print("Done. Output file: " + fname_out)
+print("")

stable-diffusion.cpp/ggml/examples/gpt-2/convert-ckpt-to-ggml.py

@@ -0,0 +1,159 @@
+# Convert a model checkpoint to a ggml compatible file
+#
+# Load the model using TensorFlow.
+# Iterate over all variables and write them to a binary file.
+#
+# For each variable, write the following:
+#   - Number of dimensions (int)
+#   - Name length (int)
+#   - Dimensions (int[n_dims])
+#   - Name (char[name_length])
+#   - Data (float[n_dims])
+#
+# By default, the bigger matrices are converted to 16-bit floats.
+# This can be disabled by adding the "use-f32" CLI argument.
+#
+# At the start of the ggml file we write the model parameters
+# and vocabulary.
+#
+
+import sys
+import json
+import struct
+import numpy as np
+import tensorflow as tf
+
+# ref: https://github.com/openai/gpt-2/blob/master/src/encoder.py
+def bytes_to_unicode():
+    """
+    Returns list of utf-8 byte and a corresponding list of unicode strings.
+    The reversible bpe codes work on unicode strings.
+    This means you need a large # of unicode characters in your vocab if you want to avoid UNKs.
+    When you're at something like a 10B token dataset you end up needing around 5K for decent coverage.
+    This is a signficant percentage of your normal, say, 32K bpe vocab.
+    To avoid that, we want lookup tables between utf-8 bytes and unicode strings.
+    And avoids mapping to whitespace/control characters the bpe code barfs on.
+    """
+    bs = list(range(ord("!"), ord("~")+1))+list(range(ord("¡"), ord("¬")+1))+list(range(ord("®"), ord("ÿ")+1))
+    cs = bs[:]
+    n = 0
+    for b in range(2**8):
+        if b not in bs:
+            bs.append(b)
+            cs.append(2**8+n)
+            n += 1
+    cs = [chr(n) for n in cs]
+    return dict(zip(bs, cs))
+
+# helper method to convert a numpy array to different float types
+def convert_to_ftype(data, ftype):
+    # fp16
+    if ftype == 1:
+        return data.astype(np.float16)
+
+    assert False, "Invalid ftype: " + str(ftype)
+
+if len(sys.argv) < 3:
+    print("Usage: convert-ckpt-to-ggml.py dir-model ftype\n")
+    print("  ftype == 0 -> float32")
+    print("  ftype == 1 -> float16")
+    sys.exit(1)
+
+# output in the same directory as the model
+dir_model = sys.argv[1]
+fname_out = sys.argv[1] + "/ggml-model.bin"
+
+with open(dir_model + "/encoder.json", "r", encoding="utf-8") as f:
+    encoder = json.load(f)
+
+with open(dir_model + "/hparams.json", "r", encoding="utf-8") as f:
+    hparams = json.load(f)
+
+# possible data types
+#   ftype == 0 -> float32
+#   ftype == 1 -> float16
+#
+# map from ftype to string
+ftype_str = ["f32", "f16"]
+
+ftype = 1
+if len(sys.argv) > 2:
+    ftype = int(sys.argv[2])
+    if ftype < 0 or ftype > 1:
+        print("Invalid ftype: " + str(ftype))
+        sys.exit(1)
+    fname_out = sys.argv[1] + "/ggml-model-" + ftype_str[ftype] + ".bin"
+
+list_vars = tf.train.list_variables(dir_model)
+
+fout = open(fname_out, "wb")
+
+fout.write(struct.pack("i", 0x67676d6c)) # magic: ggml in hex
+fout.write(struct.pack("i", hparams["n_vocab"]))
+fout.write(struct.pack("i", hparams["n_ctx"]))
+fout.write(struct.pack("i", hparams["n_embd"]))
+fout.write(struct.pack("i", hparams["n_head"]))
+fout.write(struct.pack("i", hparams["n_layer"]))
+fout.write(struct.pack("i", ftype))
+
+byte_encoder = bytes_to_unicode()
+byte_decoder = {v:k for k, v in byte_encoder.items()}
+
+fout.write(struct.pack("i", len(encoder)))
+
+for key in encoder:
+    text = bytearray([byte_decoder[c] for c in key])
+    fout.write(struct.pack("i", len(text)))
+    fout.write(text)
+
+for name, shape in list_vars:
+    print("Processing variable: " + name + " with shape: ", shape)
+
+    data = tf.train.load_variable(dir_model, name).squeeze()
+    n_dims = len(data.shape);
+
+    # for efficiency - transpose the projection matrices
+    # "model/h.*/attn/c_attn/w"
+    # "model/h.*/attn/c_proj/w"
+    # "model/h.*/mlp/c_fc/w"
+    # "model/h.*/mlp/c_proj/w"
+    if name[-14:] == "/attn/c_attn/w" or \
+       name[-14:] == "/attn/c_proj/w" or \
+       name[-11:] == "/mlp/c_fc/w" or \
+       name[-13:] == "/mlp/c_proj/w":
+        print("  Transposing")
+        data = data.transpose()
+
+    dshape = data.shape
+
+    ftype_cur = 0
+    if ftype != 0:
+        # match name:
+        #  "model/wte"
+        #  "model/h.*/attn/c_attn/w"
+        #  "model/h.*/attn/c_proj/w"
+        #  "model/h.*/mlp/c_fc/w"
+        #  "model/h.*/mlp/c_proj/w"
+        if name == "model/wte" or name[-2:] == "/w":
+            print("  Converting to " + ftype_str[ftype])
+            data = convert_to_ftype(data, ftype)
+            ftype_cur = ftype
+        else:
+            print("  Converting to float32")
+            data = data.astype(np.float32)
+            ftype_cur = 0
+
+    # header
+    str = name.encode('utf-8')
+    fout.write(struct.pack("iii", n_dims, len(str), ftype_cur))
+    for i in range(n_dims):
+        fout.write(struct.pack("i", dshape[n_dims - 1 - i]))
+    fout.write(str);
+
+    # data
+    data.tofile(fout)
+
+fout.close()
+
+print("Done. Output file: " + fname_out)
+print("")

stable-diffusion.cpp/ggml/examples/gpt-2/convert-h5-to-ggml.py

@@ -0,0 +1,195 @@
+# Convert GPT-2 h5 transformer model to ggml format
+#
+# Load the model using GPT2Model.
+# Iterate over all variables and write them to a binary file.
+#
+# For each variable, write the following:
+#   - Number of dimensions (int)
+#   - Name length (int)
+#   - Dimensions (int[n_dims])
+#   - Name (char[name_length])
+#   - Data (float[n_dims])
+#
+# By default, the bigger matrices are converted to 16-bit floats.
+# This can be disabled by adding the "use-f32" CLI argument.
+#
+# At the start of the ggml file we write the model parameters
+# and vocabulary.
+#
+
+import sys
+import struct
+import json
+import numpy as np
+import re
+
+from transformers import GPT2Model
+
+# ref: https://github.com/openai/gpt-2/blob/master/src/encoder.py
+def bytes_to_unicode():
+    """
+    Returns list of utf-8 byte and a corresponding list of unicode strings.
+    The reversible bpe codes work on unicode strings.
+    This means you need a large # of unicode characters in your vocab if you want to avoid UNKs.
+    When you're at something like a 10B token dataset you end up needing around 5K for decent coverage.
+    This is a signficant percentage of your normal, say, 32K bpe vocab.
+    To avoid that, we want lookup tables between utf-8 bytes and unicode strings.
+    And avoids mapping to whitespace/control characters the bpe code barfs on.
+    """
+    bs = list(range(ord("!"), ord("~")+1))+list(range(ord("¡"), ord("¬")+1))+list(range(ord("®"), ord("ÿ")+1))
+    cs = bs[:]
+    n = 0
+    for b in range(2**8):
+        if b not in bs:
+            bs.append(b)
+            cs.append(2**8+n)
+            n += 1
+    cs = [chr(n) for n in cs]
+    return dict(zip(bs, cs))
+
+if len(sys.argv) < 2:
+    print("Usage: convert-h5-to-ggml.py dir-model [use-f32]\n")
+    sys.exit(1)
+
+# output in the same directory as the model
+dir_model = sys.argv[1]
+fname_out = sys.argv[1] + "/ggml-model.bin"
+
+with open(dir_model + "/vocab.json", "r", encoding="utf-8") as f:
+    encoder = json.load(f)
+
+with open(dir_model + "/added_tokens.json", "r", encoding="utf-8") as f:
+    encoder_added = json.load(f)
+
+with open(dir_model + "/config.json", "r", encoding="utf-8") as f:
+    hparams = json.load(f)
+
+# use 16-bit or 32-bit floats
+use_f16 = True
+if len(sys.argv) > 2:
+    use_f16 = False
+    fname_out = sys.argv[1] + "/ggml-model-f32.bin"
+
+model = GPT2Model.from_pretrained(dir_model, low_cpu_mem_usage=True)
+#print (model)
+
+list_vars = model.state_dict()
+#print (list_vars)
+
+fout = open(fname_out, "wb")
+
+fout.write(struct.pack("i", 0x67676d6c)) # magic: ggml in hex
+fout.write(struct.pack("i", hparams["vocab_size"]))
+fout.write(struct.pack("i", hparams["n_positions"]))
+fout.write(struct.pack("i", hparams["n_embd"]))
+fout.write(struct.pack("i", hparams["n_head"]))
+fout.write(struct.pack("i", hparams["n_layer"]))
+#fout.write(struct.pack("i", hparams["rotary_dim"]))
+fout.write(struct.pack("i", use_f16))
+
+byte_encoder = bytes_to_unicode()
+byte_decoder = {v:k for k, v in byte_encoder.items()}
+
+fout.write(struct.pack("i", len(encoder) + len(encoder_added)))
+
+for key in encoder:
+    text = bytearray([byte_decoder[c] for c in key])
+    fout.write(struct.pack("i", len(text)))
+    fout.write(text)
+
+for key in encoder_added:
+    text = bytearray([byte_decoder[c] for c in key])
+    fout.write(struct.pack("i", len(text)))
+    fout.write(text)
+
+for name in list_vars.keys():
+    data = list_vars[name].squeeze().numpy()
+    print("Processing variable: " + name + " with shape: ", data.shape)
+
+    # we don't need these
+    if name.endswith("attn.masked_bias") or name.endswith(".attn.bias"):
+        print("  Skipping variable: " + name)
+        continue
+
+    n_dims = len(data.shape);
+
+    # ftype == 0 -> float32, ftype == 1 -> float16
+    ftype = 0;
+    if use_f16:
+        if name[-7:] == ".weight" and n_dims == 2:
+            print("  Converting to float16")
+            data = data.astype(np.float16)
+            ftype = 1
+        else:
+            print("  Converting to float32")
+            data = data.astype(np.float32)
+            ftype = 0
+
+    # for efficiency - transpose these matrices:
+    #  "transformer.h.*.mlp.c_proj.weight
+    if name.endswith(".mlp.c_proj.weight"):
+        print("  Transposing")
+        data = data.transpose()
+
+    # rename headers to keep compatibility
+    if name == "ln_f.weight":
+        name = "model/ln_f/g"
+    elif name == "ln_f.bias":
+        name = "model/ln_f/b"
+    elif name == "wte.weight":
+        name = "model/wte"
+    elif name == "wpe.weight":
+        name = "model/wpe"
+    elif re.match(r"h\.\d+\.ln_1\.weight", name):
+        i = re.findall("\d+", name)[0]
+        name = f"model/h{i}/ln_1/g"
+    elif re.match(r"h\.\d+\.ln_1\.bias", name):
+        i = re.findall("\d+", name)[0]
+        name = f"model/h{i}/ln_1/b"
+    elif re.match(r"h\.\d+\.attn\.c_attn\.weight", name):
+        i = re.findall("\d+", name)[0]
+        name = f"model/h{i}/attn/c_attn/w"
+    elif re.match(r"h\.\d+\.attn\.c_attn\.bias", name):
+        i = re.findall("\d+", name)[0]
+        name = f"model/h{i}/attn/c_attn/b"
+    elif re.match(r"h\.\d+\.attn\.c_proj\.weight", name):
+        i = re.findall("\d+", name)[0]
+        name = f"model/h{i}/attn/c_proj/w"
+    elif re.match(r"h.\d+.attn.c_proj.bias", name):
+        i = re.findall("\d+", name)[0]
+        name = f"model/h{i}/attn/c_proj/b"
+    elif re.match(r"h.\d+.ln_2.weight", name):
+        i = re.findall("\d+", name)[0]
+        name = f"model/h{i}/ln_2/g"
+    elif re.match(r"h.\d+.ln_2.bias", name):
+        i = re.findall("\d+", name)[0]
+        name = f"model/h{i}/ln_2/b"
+    elif re.match(r"h.\d+.mlp.c_fc.weight", name):
+        i = re.findall("\d+", name)[0]
+        name = f"model/h{i}/mlp/c_fc/w"
+    elif re.match(r"h.\d+.mlp.c_fc.bias", name):
+        i = re.findall("\d+", name)[0]
+        name = f"model/h{i}/mlp/c_fc/b"
+    elif re.match(r"h.\d+.mlp.c_proj.weight", name):
+        i = re.findall("\d+", name)[0]
+        name = f"model/h{i}/mlp/c_proj/w"
+    elif re.match(r"h.\d+.mlp.c_proj.bias", name):
+        i = re.findall("\d+", name)[0]
+        name = f"model/h{i}/mlp/c_proj/b"
+    else:
+        print("Unrecognized variable name. %s", name)
+
+    str = name.encode('utf-8')
+
+    fout.write(struct.pack("iii", n_dims, len(str), ftype))
+    for i in range(n_dims):
+        fout.write(struct.pack("i", data.shape[n_dims - 1 - i]))
+    fout.write(str);
+
+    # data
+    data.tofile(fout)
+
+fout.close()
+
+print("Done. Output file: " + fname_out)
+print("")

stable-diffusion.cpp/ggml/examples/gpt-2/download-ggml-model.sh

@@ -0,0 +1,69 @@
+#!/bin/bash
+
+# This script downloads GPT-2 model files that have already been converted to ggml format.
+# This way you don't have to convert them yourself.
+#
+# If you want to download the original GPT-2 model files, use the "download-model.sh" script instead.
+
+#src="https://ggml.ggerganov.com"
+#pfx="ggml-model-gpt-2"
+
+src="https://huggingface.co/ggerganov/ggml"
+pfx="resolve/main/ggml-model-gpt-2"
+
+ggml_path=$(dirname $(realpath $0))
+
+# GPT-2 models
+models=( "117M" "345M" "774M" "1558M" )
+
+# list available models
+function list_models {
+    printf "\n"
+    printf "  Available models:"
+    for model in "${models[@]}"; do
+        printf " $model"
+    done
+    printf "\n\n"
+}
+
+if [ "$#" -ne 1 ]; then
+    printf "Usage: $0 <model>\n"
+    list_models
+
+    exit 1
+fi
+
+model=$1
+
+if [[ ! " ${models[@]} " =~ " ${model} " ]]; then
+    printf "Invalid model: $model\n"
+    list_models
+
+    exit 1
+fi
+
+# download ggml model
+
+printf "Downloading ggml model $model ...\n"
+
+mkdir -p models/gpt-2-$model
+
+if [ -x "$(command -v wget)" ]; then
+    wget --quiet --show-progress -O models/gpt-2-$model/ggml-model.bin $src/$pfx-$model.bin
+elif [ -x "$(command -v curl)" ]; then
+    curl -L --output models/gpt-2-$model/ggml-model.bin $src/$pfx-$model.bin
+else
+    printf "Either wget or curl is required to download models.\n"
+    exit 1
+fi
+
+if [ $? -ne 0 ]; then
+    printf "Failed to download ggml model $model \n"
+    printf "Please try again later or download the original GPT-2 model files and convert them yourself.\n"
+    exit 1
+fi
+
+printf "Done! Model '$model' saved in 'models/gpt-2-$model/ggml-model.bin'\n"
+printf "You can now use it like this:\n\n"
+printf "  $ ./bin/gpt-2 -m models/gpt-2-$model/ggml-model.bin -p \"This is an example\"\n"
+printf "\n"

stable-diffusion.cpp/ggml/examples/gpt-2/download-model.sh

@@ -0,0 +1,48 @@
+#!/bin/bash
+
+ggml_path=$(dirname $(realpath $0))
+
+# GPT-2 models
+models=( "117M" "345M" "774M" "1558M" )
+
+# list available models
+function list_models {
+    printf "\n"
+    printf "  Available models:"
+    for model in "${models[@]}"; do
+        printf " $model"
+    done
+    printf "\n\n"
+}
+
+if [ "$#" -ne 1 ]; then
+    printf "Usage: $0 <model>\n"
+    list_models
+
+    exit 1
+fi
+
+model=$1
+
+if [[ ! " ${models[@]} " =~ " ${model} " ]]; then
+    printf "Invalid model: $model\n"
+    list_models
+
+    exit 1
+fi
+
+# download model
+
+printf "Downloading model $model ...\n"
+
+mkdir -p models/gpt-2-$model
+
+for file in checkpoint encoder.json hparams.json model.ckpt.data-00000-of-00001 model.ckpt.index model.ckpt.meta vocab.bpe; do
+    wget --quiet --show-progress -O models/gpt-2-$model/$file https://openaipublic.blob.core.windows.net/gpt-2/models/$model/$file
+done
+
+printf "Done! Model '$model' saved in 'models/gpt-2-$model/'\n\n"
+printf "Run the convert-ckpt-to-ggml.py script to convert the model to ggml format.\n"
+printf "\n"
+printf "  python $ggml_path/convert-ckpt-to-ggml.py models/gpt-2-$model/\n"
+printf "\n"