Croco.Cpp (CCPP) - Readme to be updated :

Unroll DISCLAIMER:

Croco.Cpp is a fork of KoboldCPP, already known as KoboldCPP Frankenstein, Frankenfork, or shortened in KCPP-F. The namechange is due to my boredom with the Frankenstein marker I myself initiated a year ago. As usual, the Croco.Cpp builds are NOT supported by the KoboldCPP (KCPP) team, Github, or Discord channel. They are for greedy-test and amusement only. Any potential support found them is a courtesy, not a due. My CCPP version number bumps as soon as the version number in the official experimental branch bumps in the following way x.xxx (ex : 1.80.1) : (KCPP)x.xxx.(CCPP)xx. They are not "upgrades" over the official version. And they might be bugged at time: only the official KCPP releases are to be considered correctly numbered, reliable and "fixed". The LllamaCPP version + the additional PRs integrated follow my CCPP versioning in the title, so everybody knows what version they deal with. Important : New models sometimes integrated in my builds (like recently Mistral Nemo, which posed problems for several users) are for personal testing only, and CAN'T be fixed if they fail because their support come from third party PRs coming from LlamaCPP merged "savagely" in my builds, sometimes before even being merged on LlamaCPP master.

Presentation :

Croco.Cpp (CCPP) is a fork of the experimental branch of KoboldCPP (KCPP), mainly aimed at NVidia Cuda users (I'm myself using Ampere GPUs, it MIGHT support the other backends also, everything is compîled but Hipblas/ROCm, but it's not tested), with a few modifications accordingly to my own needs :

• A more cluttered GUI that I had to enlarge to put all my mess.
• More context steps in GUI, as well as more Blas Batch Size (supports MMVQ 1-8 for example).
• Physical Blas Batch Size Exposed and configurable.
• 20 or so different modes of quantization for the context cache (F16, around 15 KV modes with Flash Attention, 7 quantum legacy K cache modes without Flash Attention for models like Gemma).
• KV cache supports IQ4_NL and Q6_0 (except for Gemma), thanks to Ikawrakow.
• Supports inference for new quants made by Ikawrakow (Q6_0 legacy for irregularly shaped tensors ; IQ_2K, 3K, 4K, 5K, 6K (first gen) ; IQ2_KS, 4_KSS, 4_KS (second gen) ; IQ2_KT, 3_KT, 4_KT (Trellis).
• A dozen or so commits taken from Ikawrakow's IK_Llama.CPP for performances (notably on Gemma). That includes a few more GGML ops.
• A slightly different benchmark (one flag per column instead of a single flag space).
• 10 Stories slots instead of 6 in the web-interface (KLite).
• Often some PRs unsupported/not yet supported in KCPP (I look especially at Cuda and KV cache related PRs).
• More infos displayed in the CLI, without activating debug mode.
• Smartcontext instead of contextshift by default in GUI for compatibility with Gemma.
• Support the edition of NORM_EPS_RMS value.
• More logging out of debug mode.
• Support EmphasisFSM by Yoshku to handle the "" and ** formatting in KCPP and SillyTavern (mostly, if you have troubles of chat (thoughts, actions, dialogues) formatting, and anti-slop doesn't cut it for your needs somehow).
• Since 1.71010, an enhanced model layers autoloader on GPU (which is less and less cluttered and bugged lol), based on Concedo's code and Pyroserenus formulas, but different from Henky's subsequent commit on KCPP-official. It's compatible with KV_Quants, accounts for FA, MMQ, LowVram, works in single and multi-GPU (up to 16?), is accessible in CLI and GUI modes, and can be configured easily in tandem with tensor split for an entirely customized loading accordingly to one's rig and needs.

Recommanded settings for Commande Line Interface / GUI :

--flashattention (except for Gemma?)
--blastbatchsize 128 (256 for Gemma)
--usecublas mmq (for NVidia users, MMQ mode is faster)

Check the help section (koboldcpp.exe --help or python koboldcpp.py --help) for more infos.

Croco.Cpp specifics :

Unroll the 26 KV cache options (all should be considered experimental except F16, KV Q8_0, and KV Q4_0)

With Flash Attention :

• F16 -> Fullproof (the usual KV quant since the beginning of LCPP/KCPP)
• K F16 with : V Q8_0, Q6_0 (experimental), Q5_1, Q5_0, Q4_0
• K Q8_0 with : V F16, Q8_0 (stable, my current main, part of the LCPP/KCPP main triplet), Q6_0 (experimental), Q5_1 (maybe unstable), Q5_0 (maybe unstable), the rest is untested beyond benches), Q4_0 (maybe stable)
• K Q5_1 with : V Q5_0, Q4_0
• K Q5_0 with : V Q4_0
• KV Q4_0 (quite stable, if we consider that it's part of the LCPP/KCPP main triplet) Works in command line, normally also via the GUI, and normally saves on .KCPPS config files.
• KV IQ4_NL (with -1% perplexity compared to Q4_0).

Without Flash Attention nor MMQ (for models like Gemma) :

• K F16 with KQ8_0, Q5_1, Q5_0, Q4_1, and Q4_0.
• K Q6_0 and IQ4_NL to be tested, might not work.

Unroll the options to set KV Quants (obsolete)

KCPP official KV quantized modes (modes 1 and 2 require Flash Attention) :

0 = 1616/F16 (16 BPW), 1 = FA8080/KVq8_0 (8.5 BPW), 2 = FA4040/KVq4_0 (4.5BPW),

CCPP unofficial KV quantized modes (require flash attention) :

3 = FA1680/Kf16-Vq8_0 (12.25BPW), 4 = FA1651/Kf16-Vq5_1 (11BPW), 5 = FA1650/Kf16-Vq5_0 (10.75BPW), 6 = FA1641/Kf16-Vq4_1 (10.5BPW), 7 = FA1640/Kf16-Vq4_0 (10.25BPW), 8 = FA8051/Kq8_0-Vq5_1 (7.25BPW), 9 = FA8050/Kq8_0-Vq5_0 (7BPW), 10 = FA8041/Kq8_0-Vq4_1 (6.75BPW), 11 = FA8040/Kq8_0-Vq4_0 (6.5BPW), 12 = FA5151/KVq5_1 (6BPW), 13 = FA5150/Kq5_1-Vq5_0 (5.75BPW), 14 = FA5141/Kq5_1-Vq4_1 (5.5BPW), 15 = FA5140/Kq5_1-Vq4_0 (5.25BPW), 16 = FA5050/Kq5_0-Vq5_0 (5.5BPW), 17 = FA5041/Kq5_0-Vq4_1 (5.25BPW), 18 = FA5040/Kq5_0-Vq4_0 (5BPW), 19 = FA4141/Kq4_1-Vq4_1 (5BPW), 20 = FA4140/Kq4_1-Vq4_0 (4.75BPW)

21 = 1616/F16 (16 BPW), (same as 0, I just used it for the GUI slider).

22 = 8016/Kq8_0, Vf16 (12.25BPW), FA and no-FA both

23 = 5116/Kq5_1-Vf16 (11BPW), no-FA 24 = 5016/Kq5_1-Vf16 (10.75BPW), no-FA 25 = 4116/Kq4_1-Vf16 (10.50BPW), no-FA 26 = 4016/Kq4_0-Vf16 (10.25BPW), no-FA

choices=[0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,21,22,23,24,25,26], default=0)

Unroll the details about Emphasisfsm by Yoshqu

The common problem during text generiation are misplaced emphasis characters.

*looks at you "why* this is here?"

while it should be

*looks at you* "why this is here?"

This emphasisfsm solves this by simple (and fast) grammar expressed by deterministic finite state machine.

Single letters are not practical in LLMs as tokens often contains more than one.

Emphasisfsm uses LLM tokens as its alphabet making it very fast.

Those are only most obvious examples. There are more, eg. ' "***' is a valid token to transition from qout to star. and '*this' is vaild for quot->none or none->quot.

Usage

To support variety of GUIs this extension shamefully exploits GBNF grammar string. This is not a proper GBNF grammar, it only uses the field which is easily editable in most GUIs

emphasisfsm "_bias_[D][_emph1_][,_emphn_]"

Empty string emphasisfsm is disabled. The easiest way to enable is to

emphasisfsm "-20"

which defaults to

emphasisfsm "-20 \" \" * *"

(no debug, only * and " are considered)

how it works

Main loop is extended from:

• retrieve logits
• sample logits, select token (top_k and friends)
• output token

to

• retrieve logits
• ban forbidden emphasisfsm transitions from current state (stetting their logits low)
• sample logits, select token (top_k and friends)
• emphasisfsm trasition on selected token
• output token

TODO

• find split utf8 letters over more than one token (i don't plant to support it, but warning would be nice)
• banning end tokens generation inside of emphasis - forcing LLM to finsh his 'thought' ?

Meta-Llama-3-8B stats for default (" *) emphasisfsm

empcats_gen: ban bias: -17.500000
empcats_gen: emphasis indifferent tokens: 126802
empcats_gen: tokens for emphasis '"' '"': 1137
empcats_gen: tokens for emphasis '*' '*': 315
empcats_gen: always banned tokens: 2
empcats_gen: total tokens: 128256

Always banned tokens are :

' "*"',  ' "*"'

Tests

emphasisfsm "-20 1 1 2 2 3 3 4 4 5 5 6 6 7 7 8 8 9 9 0 0"

This forces that every digit is a citation, so example text completion looks like:

Give me math vector of random numbers.Here is a 3-dimensional math vector with random numbers:
Vector:
[
    3.445,
    -5.117,
    7.992
]

There is no other digit between two 3, two 4, two 5 and so on....

Unroll the Dry sampler recommanded settings

Multiplier : 0.8 Base : 1.75 Allowed length : 2 Range : Uses the repetition penalty range (usual parameter is 2048) Usual sequence breakers : '\n', ':', '"', '*'

#Croco.Cpp notes :

• I often mislabel the Cuda specifics of the builds. Here's the right nomenclature : Cuda 12.2 arch 60617075 : Cu12.2_SMC2_Ar60617075_DmmvX64Y2_MMY2_KQIt2 Cuda 12.1 arch 52617075 : CuCML_ArCML_SMC2_DmmvX32Y1 (CML : CMakeList) Cuda 11.4.4/11.5 arch 35375052 : CuCML_ArCML_SMC2_DmmvX32Y1

koboldcpp-experimental

KoboldCpp-experimental is a sligthly extended KoboldCpp with custom functionality.

koboldcpp

KoboldCpp is an easy-to-use AI text-generation software for GGML and GGUF models, inspired by the original KoboldAI. It's a single self-contained distributable from Concedo, that builds off llama.cpp, and adds a versatile KoboldAI API endpoint, additional format support, Stable Diffusion image generation, speech-to-text, backward compatibility, as well as a fancy UI with persistent stories, editing tools, save formats, memory, world info, author's note, characters, scenarios and everything KoboldAI and KoboldAI Lite have to offer.

Windows Usage (Precompiled Binary, Recommended)

• Windows binaries are provided in the form of koboldcpp.exe, which is a pyinstaller wrapper containing all necessary files. Download the latest koboldcpp.exe release here
• To run, simply execute koboldcpp.exe.
• Launching with no command line arguments displays a GUI containing a subset of configurable settings. Generally you dont have to change much besides the Presets and GPU Layers. Read the --help for more info about each settings.
• Obtain and load a GGUF model. See here
• By default, you can connect to http://localhost:5001
• You can also run it using the command line. For info, please check koboldcpp.exe --help

Linux Usage (Precompiled Binary, Recommended)

On modern Linux systems, you should download the koboldcpp-linux-x64-cuda1150 prebuilt PyInstaller binary on the releases page. Simply download and run the binary (You may have to chmod +x it first).

Alternatively, you can also install koboldcpp to the current directory by running the following terminal command:

curl -fLo koboldcpp https://github.com/LostRuins/koboldcpp/releases/latest/download/koboldcpp-linux-x64-cuda1150 && chmod +x koboldcpp

After running this command you can launch Koboldcpp from the current directory using ./koboldcpp in the terminal (for CLI usage, run with --help). Finally, obtain and load a GGUF model. See here

MacOS (Precompiled Binary)

• PyInstaller binaries for Modern ARM64 MacOS (M1, M2, M3) are now available! Simply download the MacOS binary
• In a MacOS terminal window, set the file to executable chmod +x koboldcpp-mac-arm64 and run it with ./koboldcpp-mac-arm64.
• In newer MacOS you may also have to whitelist it in security settings if it's blocked. Here's a video guide.
• Alternatively, or for older x86 MacOS computers, you can clone the repo and compile from source code, see Compiling for MacOS below.
• Finally, obtain and load a GGUF model. See here

Run on Colab

• KoboldCpp now has an official Colab GPU Notebook! This is an easy way to get started without installing anything in a minute or two. Try it here!.
• Note that KoboldCpp is not responsible for your usage of this Colab Notebook, you should ensure that your own usage complies with Google Colab's terms of use.

Run on RunPod

• KoboldCpp can now be used on RunPod cloud GPUs! This is an easy way to get started without installing anything in a minute or two, and is very scalable, capable of running 70B+ models at afforable cost. Try our RunPod image here!.

Run on Novita AI

KoboldCpp can now also be run on Novita AI, a newer alternative GPU cloud provider which has a quick launch KoboldCpp template for as well. Check it out here!

Docker

• The official docker can be found at https://hub.docker.com/r/koboldai/koboldcpp
• If you're building your own docker, remember to set CUDA_DOCKER_ARCH or enable LLAMA_PORTABLE

Obtaining a GGUF model

• KoboldCpp uses GGUF models. They are not included with KoboldCpp, but you can download GGUF files from other places such as TheBloke's Huggingface. Search for "GGUF" on huggingface.co for plenty of compatible models in the .gguf format.
• For beginners, we recommend the models Airoboros Mistral or Tiefighter 13B (larger model).
• Alternatively, you can download the tools to convert models to the GGUF format yourself here. Run convert-hf-to-gguf.py to convert them, then quantize_gguf.exe to quantize the result.
• Other models for Whisper (speech recognition), Image Generation or Image Recognition can be found on the Wiki

Improving Performance

• GPU Acceleration: If you're on Windows with an Nvidia GPU you can get CUDA support out of the box using the --usecublas flag (Nvidia Only), or --usevulkan (Any GPU), make sure you select the correct .exe with CUDA support.
• GPU Layer Offloading: Add --gpulayers to offload model layers to the GPU. The more layers you offload to VRAM, the faster generation speed will become. Experiment to determine number of layers to offload, and reduce by a few if you run out of memory.
• Increasing Context Size: Use --contextsize (number) to increase context size, allowing the model to read more text. Note that you may also need to increase the max context in the KoboldAI Lite UI as well (click and edit the number text field).
• Old CPU Compatibility: If you are having crashes or issues, you can try running in a non-avx2 compatibility mode by adding the --noavx2 flag. You can also try turning off mmap with --nommap or reducing your --blasbatchssize (set -1 to avoid batching)

For more information, be sure to run the program with the --help flag, or check the wiki.

Compiling KoboldCpp From Source Code

Compiling on Linux (Using koboldcpp.sh automated compiler script)

when you can't use the precompiled binary directly, we provide an automated build script which uses conda to obtain all dependencies, and generates (from source) a ready-to-use a pyinstaller binary for linux users.

• Clone the repo with git clone https://github.com/LostRuins/koboldcpp.git
• Simply execute the build script with ./koboldcpp.sh dist and run the generated binary. (Not recommended for systems that already have an existing installation of conda. Dependencies: curl, bzip2)

./koboldcpp.sh # This launches the GUI for easy configuration and launching (X11 required).
./koboldcpp.sh --help # List all available terminal commands for using Koboldcpp, you can use koboldcpp.sh the same way as our python script and binaries.
./koboldcpp.sh rebuild # Automatically generates a new conda runtime and compiles a fresh copy of the libraries. Do this after updating Koboldcpp to keep everything functional.
./koboldcpp.sh dist # Generate your own precompiled binary (Due to the nature of Linux compiling these will only work on distributions equal or newer than your own.)

Compiling on Linux (Manual Method)

• To compile your binaries from source, clone the repo with git clone https://github.com/LostRuins/koboldcpp.git
• A makefile is provided, simply run make.
• Optional Vulkan: Link your own install of Vulkan SDK manually with make LLAMA_VULKAN=1
• Optional CLBlast: Link your own install of CLBlast manually with make LLAMA_CLBLAST=1
• Note: for these you will need to obtain and link OpenCL and CLBlast libraries.
  • For Arch Linux: Install cblas and clblast.
  • For Debian: Install libclblast-dev.
• You can attempt a CuBLAS build with LLAMA_CUBLAS=1, (or LLAMA_HIPBLAS=1 for AMD). You will need CUDA Toolkit installed. Some have also reported success with the CMake file, though that is more for windows.
• For a full featured build (all backends), do make LLAMA_CLBLAST=1 LLAMA_CUBLAS=1 LLAMA_VULKAN=1. (Note that LLAMA_CUBLAS=1 will not work on windows, you need visual studio)
• To make your build sharable and capable of working on other devices, you must use LLAMA_PORTABLE=1
• After all binaries are built, you can run the python script with the command koboldcpp.py [ggml_model.gguf] [port]

Compiling on Windows

• You're encouraged to use the .exe released, but if you want to compile your binaries from source at Windows, the easiest way is:
  • Get the latest release of w64devkit (https://github.com/skeeto/w64devkit). Be sure to use the "vanilla one", not i686 or other different stuff. If you try they will conflit with the precompiled libs!
  • Clone the repo with git clone https://github.com/LostRuins/koboldcpp.git
  • Make sure you are using the w64devkit integrated terminal, then run make at the KoboldCpp source folder. This will create the .dll files for a pure CPU native build.
  • For a full featured build (all backends), do make LLAMA_CLBLAST=1 LLAMA_VULKAN=1. (Note that LLAMA_CUBLAS=1 will not work on windows, you need visual studio)
  • To make your build sharable and capable of working on other devices, you must use LLAMA_PORTABLE=1
  • If you want to generate the .exe file, make sure you have the python module PyInstaller installed with pip (pip install PyInstaller). Then run the script make_pyinstaller.bat
  • The koboldcpp.exe file will be at your dist folder.
• Building with CUDA: Visual Studio, CMake and CUDA Toolkit is required. Clone the repo, then open the CMake file and compile it in Visual Studio. Copy the koboldcpp_cublas.dll generated into the same directory as the koboldcpp.py file. If you are bundling executables, you may need to include CUDA dynamic libraries (such as cublasLt64_11.dll and cublas64_11.dll) in order for the executable to work correctly on a different PC.
• Replacing Libraries (Not Recommended): If you wish to use your own version of the additional Windows libraries (OpenCL, CLBlast, Vulkan), you can do it with:
  • OpenCL - tested with https://github.com/KhronosGroup/OpenCL-SDK . If you wish to compile it, follow the repository instructions. You will need vcpkg.
  • CLBlast - tested with https://github.com/CNugteren/CLBlast . If you wish to compile it you will need to reference the OpenCL files. It will only generate the ".lib" file if you compile using MSVC.
  • Move the respectives .lib files to the /lib folder of your project, overwriting the older files.
  • Also, replace the existing versions of the corresponding .dll files located in the project directory root (e.g. clblast.dll).
  • Make the KoboldCpp project using the instructions above.

Compiling on MacOS

• You can compile your binaries from source. You can clone the repo with git clone https://github.com/LostRuins/koboldcpp.git
• A makefile is provided, simply run make.
• If you want Metal GPU support, instead run make LLAMA_METAL=1, note that MacOS metal libraries need to be installed.
• To make your build sharable and capable of working on other devices, you must use LLAMA_PORTABLE=1
• After all binaries are built, you can run the python script with the command koboldcpp.py --model [ggml_model.gguf] (and add --gpulayers (number of layer) if you wish to offload layers to GPU).

Compiling on Android (Termux Installation)

• Install and run Termux from F-Droid
• Enter the command termux-change-repo and choose Mirror by BFSU
• Install dependencies with pkg install wget git python (plus any other missing packages)
• Install dependencies apt install openssl (if needed)
• Clone the repo git clone https://github.com/LostRuins/koboldcpp.git
• Navigate to the koboldcpp folder cd koboldcpp
• Build the project make
• To make your build sharable and capable of working on other devices, you must use LLAMA_PORTABLE=1, this disables usage of ARM instrinsics.
• Grab a small GGUF model, such as wget https://huggingface.co/concedo/KobbleTinyV2-1.1B-GGUF/resolve/main/KobbleTiny-Q4_K.gguf
• Start the python server python koboldcpp.py --model KobbleTiny-Q4_K.gguf
• Connect to http://localhost:5001 on your mobile browser
• If you encounter any errors, make sure your packages are up-to-date with pkg up
• GPU acceleration for Termux may be possible but I have not explored it. If you find a good cross-device solution, do share or PR it.

AMD Users

• For most users, you can get very decent speeds by selecting the Vulkan option instead, which supports both Nvidia and AMD GPUs.
• Alternatively, you can try the ROCM fork at https://github.com/YellowRoseCx/koboldcpp-rocm

Third Party Resources

• These unofficial resources have been contributed by the community, and may be outdated or unmaintained. No official support will be provided for them!
  • Arch Linux Packages: CUBLAS, and HIPBLAS.
  • Unofficial Dockers: korewaChino and noneabove1182
  • Nix & NixOS: KoboldCpp is available on Nixpkgs and can be installed by adding just koboldcpp to your environment.systemPackages (or it can also be placed in home.packages).
    • Example Nix Setup and further information
    • If you face any issues with running KoboldCpp on Nix, please open an issue here.
• GPTLocalhost - KoboldCpp is supported by GPTLocalhost, a local Word Add-in for you to use KoboldCpp in Microsoft Word. A local alternative to "Copilot in Word."

Questions and Help Wiki

• First, please check out The KoboldCpp FAQ and Knowledgebase which may already have answers to your questions! Also please search through past issues and discussions.
• If you cannot find an answer, open an issue on this github, or find us on the KoboldAI Discord.

KoboldCpp and KoboldAI API Documentation

• Documentation for KoboldAI and KoboldCpp endpoints can be found here

KoboldCpp Public Demo

• A public KoboldCpp demo can be found at our Huggingface Space. Please do not abuse it.

Considerations

• For Windows: No installation, single file executable, (It Just Works)
• Since v1.15, requires CLBlast if enabled, the prebuilt windows binaries are included in this repo. If not found, it will fall back to a mode without CLBlast.
• Since v1.33, you can set the context size to be above what the model supports officially. It does increases perplexity but should still work well below 4096 even on untuned models. (For GPT-NeoX, GPT-J, and Llama models) Customize this with --ropeconfig.
• Since v1.42, supports GGUF models for LLAMA and Falcon
• Since v1.55, lcuda paths on Linux are hardcoded and may require manual changes to the makefile if you do not use koboldcpp.sh for the compilation.
• Since v1.60, provides native image generation with StableDiffusion.cpp, you can load any SD1.5 or SDXL .safetensors model and it will provide an A1111 compatible API to use.
• I try to keep backwards compatibility with ALL past llama.cpp models. But you are also encouraged to reconvert/update your models if possible for best results.
• Since v1.75, openblas has been deprecated and removed in favor of the native CPU implementation.

License

• The original GGML library and llama.cpp by ggerganov are licensed under the MIT License
• However, KoboldAI Lite is licensed under the AGPL v3.0 License
• KoboldCpp code and other files are also under the AGPL v3.0 License unless otherwise stated

Notes

• If you wish, after building the koboldcpp libraries with make, you can rebuild the exe yourself with pyinstaller by using make_pyinstaller.bat
• API documentation available at /api (e.g. http://localhost:5001/api) and https://lite.koboldai.net/koboldcpp_api. An OpenAI compatible API is also provided at /v1 route (e.g. http://localhost:5001/v1).
• All up-to-date GGUF models are supported, and KoboldCpp also includes backward compatibility for older versions/legacy GGML .bin models, though some newer features might be unavailable.
• An incomplete list of models and architectures is listed, but there are many hundreds of other GGUF models. In general, if it's GGUF, it should work.
  • Llama / Llama2 / Llama3 / Alpaca / GPT4All / Vicuna / Koala / Pygmalion / Metharme / WizardLM
  • Mistral / Mixtral / Miqu
  • Qwen / Qwen2 / Yi
  • Gemma / Gemma2
  • GPT-2 / Cerebras
  • Phi-2 / Phi-3
  • GPT-NeoX / Pythia / StableLM / Dolly / RedPajama
  • GPT-J / RWKV4 / MPT / Falcon / Starcoder / Deepseek and many more
  • Stable Diffusion 1.5 and SDXL safetensor models
  • LLaVA based Vision models and multimodal projectors (mmproj)
  • Whisper models for Speech-To-Text