/exllamav2

A fast inference library for running LLMs locally on modern consumer-class GPUs

Primary LanguagePythonMIT LicenseMIT

ExLlamaV2

This is a very initial release of ExLlamaV2, an inference library for running local LLMs on modern consumer GPUs.

It still needs a lot of testing and tuning, and a few key features are not yet implemented. Don't be surprised if things are a bit broken to start with, as almost all of this code is completely new and only tested on a few setups so far.

Overview of differences compared to V1

  • Faster, better kernels
  • Cleaner and more versatile codebase
  • Support for a new quant format (see below)

Performance

Some quick tests to compare performance with V1. There may be more performance optimizations in the future, and speeds will vary across GPUs, with slow CPUs still being a potential bottleneck:

Model Mode Size grpsz act V1: 3090Ti V1: 4090 V2: 3090Ti V2: 4090
Llama GPTQ 7B 128 no 143 t/s 173 t/s 175 t/s 195 t/s
Llama GPTQ 13B 128 no 84 t/s 102 t/s 105 t/s 110 t/s
Llama GPTQ 33B 128 yes 37 t/s 45 t/s 45 t/s 48 t/s
OpenLlama GPTQ 3B 128 yes 194 t/s 226 t/s 295 t/s 321 t/s
CodeLlama EXL2 4.0 bpw 34B - - - - 42 t/s 48 t/s
Llama2 EXL2 3.0 bpw 7B - - - - 195 t/s 224 t/s
Llama2 EXL2 4.0 bpw 7B - - - - 164 t/s 197 t/s
Llama2 EXL2 5.0 bpw 7B - - - - 144 t/s 160 t/s
Llama2 EXL2 2.5 bpw 70B - - - - 30 t/s 35 t/s
TinyLlama EXL2 3.0 bpw 1.1B - - - - 536 t/s 635 t/s
TinyLlama EXL2 4.0 bpw 1.1B - - - - 509 t/s 590 t/s

How to

Clone the repository and install dependencies:

git clone https://github.com/turboderp/exllamav2
cd exllamav2
pip install -r requirements.txt

python test_inference.py -m <path_to_model> -p "Once upon a time,"

A simple console chatbot is included. Run it with:

python examples/chat.py -m <path_to_model> -mode llama

For a chat with colored code, run:

python examples/chatcode.py -m <path_to_model> -mode llama

The -mode argument chooses the prompt format to use. llama is for the Llama(2)-chat finetunes, while codellama probably works better for CodeLlama-instruct. raw will produce a simple chatlog-style chat that works with base models and various other finetunes. You can also provide a custom system prompt with -sp.

Installation

Method 1: Install from source

To install the current dev version, clone the repo and run the setup script:

git clone https://github.com/turboderp/exllamav2
cd exllamav2
python setup.py install --user

By default this will also compile and install the Torch C++ extension (exllamav2_ext) that the library relies on. You can skip this step by setting the EXLLAMA_NOCOMPILE environment variable:

EXLLAMA_NOCOMPILE= python setup.py install --user

This will install the "JIT version" of the package, i.e. it will install the Python components without building the C++ extension in the process. Instead, the extension will be built the first time the library is used, then cached in ~/.cache/torch_extensions for subsequent use.

Method 2: Install from release (with prebuilt extension)

Releases are available here, with prebuilt wheels that contain the extension binaries. Make sure to grab the right version, matching your platform, Python version (cp) and CUDA version. Download an appropriate wheel, then run:

pip install exllamav2-0.0.4+cu118-cp310-cp310-linux_x86_64.whl

The py3-none-any.whl version is the JIT version which will build the extension on first launch. The .tar.gz file can also be installed this way, and it will build the extension while installing.

Method 3: Install from PyPI

A PyPI package is available as well. It can be installed with:

pip install exllamav2

The version available through PyPI is the JIT version (see above). Still working on a solution for distributing prebuilt wheels via PyPI.

EXL2 quantization

ExLlamaV2 supports the same 4-bit GPTQ models as V1, but also a new "EXL2" format. EXL2 is based on the same optimization method as GPTQ and supports 2, 3, 4, 5, 6 and 8-bit quantization. The format allows for mixing quantization levels within a model to achieve any average bitrate between 2 and 8 bits per weight.

Moreover, it's possible to apply multiple quantization levels to each linear layer, producing something akin to sparse quantization wherein more important weights (columns) are quantized with more bits. The same remapping trick that lets ExLlama work efficiently with act-order models allows this mixing of formats to happen with little to no impact on performance.

Parameter selection is done automatically by quantizing each matrix multiple times, measuring the quantization error (with respect to the chosen calibration data) for each of a number of possible settings, per layer. Finally, a combination is chosen that minimizes the maximum quantization error over the entire model while meeting a target average bitrate.

In my tests, this scheme allows Llama2 70B to run on a single 24 GB GPU with a 2048-token context, producing coherent and mostly stable output with 2.55 bits per weight. 13B models run at 2.65 bits within 8 GB of VRAM, although currently none of them uses GQA which effectively limits the context size to 2048. In either case it's unlikely that the model will fit alongside a desktop environment. For now.

chat_screenshot chat_screenshot

Conversion

A script is provided to quantize models. Converting large models can be somewhat slow, so be warned. The conversion script and its options are explained in detail here

HuggingFace repos

I've uploaded a few EXL2-quantized models to HuggingFace to play around with, here.

Note that these were produced over a period of time with different calibration data, so they're not useful as a way to measure quantization loss. Thorough perplexity and accuracy tests are coming, once I've had time to convert models for that purpose.

More to come

There are still things that need to be ported over from V1, and other planned features. Among them:

  • Example web UI
  • Web server
  • More samplers

Recent updates

2023-09-27: Prebuilt wheels are now available, credit to @jllllll. They're on the releases page here. A solution to installing prebuilt wheels straight from PyPI is still pending. Updated installation instructions above.

2023-10-03: Added support for extended vocabularies and alternative BOS/EOS/UNK tokens and the ability to encode/decode sequences with special tokens. Added Orca template to the chatbot example.

2023-10-07: (Multi) LoRA support as well as some experimental optimizations.

2023-10-13: Merged speculative sampling into streaming generator. Now supports streaming and stop conditions. Chat example updated to take draft model.

2023-10-15: Got the 8-bit cache mode to a fairly working state. Added the -c8 option to the chatbot. Big VRAM savings for CodeLlama-13B, at least.

2023-10-22: Added auto GPU split option. -gs auto will load the model while allocating the cache and running a forward pass to precisely measure VRAM usage, then automatically use all available VRAM starting from the first CUDA device.