ExLlamaV3

ExLlamaV3 is an inference library for running local LLMs on modern consumer GPUs. Headline features:

• New EXL3 quantization format based on QTIP
• Flexible tensor-parallel and expert-parallel inference for consumer hardware setups
• OpenAI-compatible server provided via TabbyAPI
• Continuous, dynamic batching
• HF Transformers plugin (see here)
• HF model support (see supported architectures)
• Speculative decoding
• 2-8 bit cache quantization
• Multimodal support
• LoRA support

The official and recommended backend server for ExLlamaV3 is TabbyAPI, which provides an OpenAI-compatible API for local or remote inference, with extended features like HF model downloading, embedding model support and support for HF Jinja2 chat templates.

Architecture support

• AFM (ArceeForCausalLM)
• AfMoE (AfmoeForCausalLM)
• Apertus (ApertursForCausalLM)
• Command-R etc. (CohereForCausalLM)
• Command-A, Command-R7B, Command-R+ etc. (Cohere2ForCausalLM)
• DeciLM, Nemotron (DeciLMForCausalLM)
• Deepseek V3 (DeepseekV3ForCausalLM)
• Deepseek V4 (DeepseekV4ForCausalLM)
• dots.llm1 (Dots1ForCausalLM) (n_group>1 currently not supported)
• ERNIE 4.5 (Ernie4_5_ForCausalLM, Ernie4_5_MoeForCausalLM)
• EXAONE 4.0 (Exaone4ForCausalLM)
• Gemma 2 (Gemma2ForCausalLM)
• Gemma 3 (Gemma3ForCausalLM, Gemma3ForConditionalGeneration) - multimodal
• Gemma 4 (Gemma4ForConditionalGeneration, Gemma4UnifiedForConditionalGeneration) - multimodal (E2B/E4B currently not supported)
• GLM 4, GLM 4.5, GLM 4.5-Air, GLM 4.6 (Glm4ForCausalLM, Glm4MoeForCausalLM)
• GLM 4.1V, GLM 4.5V (Glm4vForConditionalGeneration, Glm4vMoeForConditionalGeneration) - multimodal
• GLM 5.2 (GlmMoeDsaForCausalLM)
• GLM 5.3-Flash (Glm5NextForConditionalGeneration) - multimodal
• GPT-OSS (GptOssForCausalLM)
• HyperCLOVAX (HyperCLOVAXForCausalLM, HCXVisionV2ForCausalLM) - multimodal
• Hy3 (HYV3ForCausalLM)
• IQuest-Coder (IQuestCoderForCausalLM)
• Laguna 2.1 (LagunaForCausalLM)
• LFM 2.5 (Lfm2MoeForCausalLM)
• Llama, Llama 2, Llama 3, Llama 3.1-Nemotron etc. (LlamaForCausalLM)
• MiMo-RL (MiMoForCausalLM)
• MiniMax-M2 (MiniMaxM2ForCausalLM)
• Mistral, Ministral 3, Devstral 2, Mistral-4 etc. (MistralForCausalLM, Mistral3ForConditionalGeneration) - multimodal
• Mixtral (MixtralForCausalLM)
• NemotronH, Nemotron-3 (NemotronHForCausalLM)
• Olmo 3.1 (Olmo3ForCausalLM)
• Olmo-Hybrid (OlmoHybridForCausalLM)
• Phi3, Phi4 (Phi3ForCausalLM)
• Qwen 2, Qwen 2.5, Qwen 2.5 VL (Qwen2ForCausalLM, Qwen2_5_VLForConditionalGeneration) - multimodal
• Qwen 3 (Qwen3ForCausalLM, Qwen3MoeForCausalLM)
• Qwen 3-Next (Qwen3NextForCausalLM)
• Qwen 3-VL (Qwen3VLForConditionalGeneration) - multimodal
• Qwen 3-VL MoE (Qwen3VLMoeForConditionalGeneration) - multimodal
• Qwen 3.5 (Qwen3_5ForConditionalGeneration) - multimodal
• Qwen 3.5 MoE (Qwen3_5MoeForConditionalGeneration) - multimodal
• Qwen 3.8-Flash-Next (Qwen4ExpForConditionalGeneration) - multimodal
• Seed-OSS (SeedOssForCausalLM)
• SmolLM (SmolLM3ForCausalLM)
• SolarOpen (SolarOpenForCausalLM)
• Step 3.5 Flash (Step3p5ForCausalLM)
• Step 3.7 Flash (Step3p7ForConditionalGeneration) - multimodal

Always adding more, stay tuned.

What's missing?

Currently on the to-do list:

• ROCm support

As for what is implemented, expect that some things may be a little broken at first. Please be patient, raise issues and/or contribute. 👉👈

How to?

TabbyAPI has a startup script that manages and installs prerequisites if you want to get started quickly with inference in an OAI-compatible client.

Otherwise, start by making sure you have the appropriate version of PyTorch installed (CUDA 12.4 or later) since the Torch dependency is not automatically handled by pip. Then pick a method below:

Method 1: Installing from prebuilt wheel (recommended if you're unsure)

Pick a wheel from the releases page, then e.g.:

pip install https://github.com/turboderp-org/exllamav3/releases/download/v0.0.6/exllamav3-0.0.6+cu128.torch2.8.0-cp313-cp313-linux_x86_64.whl

Method 2: Installing from PyPi:

pip install exllamav3

Note that the PyPi package does not contain a prebuilt extension and requires the CUDA toolkit and build prerequisites (i.e. VS Build Tools on Windows, gcc on Linux, python-dev headers etc.).

Method 3: Building from source

exllamav3 declares a minimum torch version (>= 2.6.0) and CUDA version (>= 12.4), but beyond that the user is free to select a version of torch that is compatible with their environment.

torch can be installed in three ways (from least to most effort):

1. with uv, setting only --extra cuXXX installs torch automatically with the specified CUDA version, torch version is selected by uv from compatible versions in the specific index associated with the chosen CUDA version (options 1 and 2)
2. with uv, creating a thin project that depends on exllamav3[cuXXX] and pins a specific torch version — like (1) but torch is pinned in the thin project's pyproject.toml, see pinning a specific PyTorch version (optional) for details
3. Manually with uv pip or pip (options 3 and 4)

The flavor extras (--extra) are cu124, cu126, cu128, cu129, cu130, and cu132 — pick the one matching your installed CUDA build. Both uv sync and pip install . build the package in an isolated environment where your torch is not visible, so they install the extension sources and compile them at first import (JIT, a few minutes once per torch version). For a precompiled install run pip install --no-build-isolation . in an environment that already has torch, or use the release wheels. Selecting a flavor installs the matching CUDA build of torch.

Option 1 — Working in the cloned repo directly (uv sync):

git clone https://github.com/turboderp-org/exllamav3
cd exllamav3
# (Optional) switch to dev branch for latest in-progress features
git checkout dev

uv venv
uv sync --extra cu130
# add --extra examples and/or --extra eval for those extra dependencies

Option 2 — Using exllamav3 as a dependency from another project (uv add):

# `uv add` works inside an existing project (a directory with a pyproject.toml).
# `uv init` creates one if you're starting a new project, if integrating into
# an existing project skip `uv init`.
uv init my-project
cd my-project

# local checkout
uv add 'path/to/exllamav3[cu130]'               # non-editable
uv add 'path/to/exllamav3[cu130]' --editable    # editable

# straight from GitHub
uv add 'git+https://github.com/turboderp-org/exllamav3.git[cu130]'                 # default branch
uv add 'git+https://github.com/turboderp-org/exllamav3.git[cu130]' --branch dev    # specific branch

Option 3 — Bring your own torch and let uv pick the backend automatically:

uv venv            # or: uv venv --python-preference only-managed
source .venv/bin/activate
uv pip install torch --torch-backend=auto
uv pip install .

--torch-backend=auto inspects your system and installs the matching PyTorch CUDA build; see Automatic backend selection.

Option 4 — With pip:

On Windows, you should also make sure you have the triton-windows package installed. ExLlamaV3 may work without it, but many things will work suboptimally.

# install a CUDA-enabled torch first so it matches your setup, e.g.:
pip install torch --index-url https://download.pytorch.org/whl/cu128
pip install .

Pinning a specific PyTorch version (optional)

The flavor extra picks the index, but by default torch resolves to the latest version on that index that satisfies >=2.6.0. To pin a specific torch version while developing on exllamav3, create a "thin" project that consumes your local checkout as an editable install and declares the exact torch version itself. This keeps the pin out of the exllamav3 pyproject, so you can change the torch version freely without touching the repo.

my-exllamav3-dev/          # thin project (uv init)
├── pyproject.toml
└── src/                  # package sources (auto-generated)

In pyproject.toml:

[project]
name = "my-exllamav3-dev"
version = "0.1.0"
description = "Dev environment for exllamav3"
requires-python = ">=3.10.11"
dependencies = [
    "exllamav3[cu130]",   # select correct CUDA version
    "torch==2.13.0",      # pin the exact torch version you need
]

[tool.uv.sources]
exllamav3 = { path = "../exllamav3", editable = true }

Adjust ../exllamav3 to point at your local checkout, then a plain uv sync sets up an environment with the correct PyTorch index (routed via the cuXXX extra), the pinned version of torch from that index (as long as it exists), and an editable install of exllamav3 so code changes apply immediately. Switch CUDA flavors by changing the extra (exllamav3[cu124], exllamav3[cu128], …) and/or the torch pin in the thin project.

Or, if you're installing torch manually with uv pip install torch (e.g. as in Option 3 above), specify the version directly, e.g. uv pip install "torch==2.11.0" --torch-backend=auto.

---

After installing with one of the options above, you should be able to run the conversion, eval and example scripts from the main repo directory, e.g., uv run python convert.pt -i ... or, for manual installations once the venv is active, python convert.py -i ...

Relevant env variables for building:

• MAX_JOBS: by default ninja may launch too many processes and run out of system memory for compilation. Set this to a reasonable value like 4 in that case.
• EXLLAMA_NOCOMPILE: set to install the library without compiling the C++/CUDA extension. Torch will build/load it at runtime instead.

Conversion

To convert a model to EXL3 format, use:

# Convert model
python convert.py -i <input_dir> -o <output_dir> -w <working_dir> -b <bitrate>

# Resume an interrupted quant job
python convert.py -w <working_dir> -r

# More options
python convert.py -h

The working directory is temporary storage for state checkpoints and for storing quantized tensors until the converted model can be compiled. It should have enough free space to store an entire copy of the output model. Note that while EXL2 conversion by default resumes an interrupted job when pointed to an existing folder, EXL3 needs you to explicitly resume with the -r/--resume argument.

See here for more information.

Examples

A number of example scripts are provided to showcase the features of the backend and generator. Some of them have hardcoded model paths and should be edited before you run them, but there is a simple CLI chatbot that you can start with:

python examples/chat.py -m <input_dir> -mode <prompt_mode> 

# E.g.:
python examples/chat.py -m /mnt/models/llama3.1-8b-instruct-exl3 -mode llama3

# Wealth of options
python examples/chat.py -h

EXL3 quantization

Despite their amazing achievements, most SOTA quantization techniques remain cumbersome or even prohibitively expensive to use. For instance, AQLM quantization of a 70B model takes around 720 GPU-hours on an A100 server, costing $850 US at the time of writing. ExLlamaV3 aims to address this with the EXL3 format, which is a streamlined variant of QTIP from Cornell RelaxML. The conversion process is designed to be simple and efficient and requires only an input model (in HF format) and a target bitrate. By computing Hessians on the fly and thanks to a fused Viterbi kernel, the quantizer can convert a model in a single step, taking a couple of minutes for smaller models, up to a few hours for larger ones (70B+) (on a single RTX 4090 or equivalent GPU.)

The Marlin-inspired GEMM kernel achieves roughly memory-bound latency under optimal conditions (4bpw, RTX 4090), though it still needs some work to achieve the same efficiency on Ampere GPUs and to remain memory-bound at lower bitrates.

Since converted models largely retain the original file structure (unlike EXL2 which renames some tensors in its quest to turn every model into a Llama variant), it will be possible to extend EXL3 support to other frameworks like HF Transformers and vLLM.

There are some benchmark results here, and a full writeup on the format is coming soon.

Fun fact: Llama-3.1-70B-EXL3 is coherent at 1.6 bpw. With the output layer quantized to 3 bpw and a 4096-token cache, inference is possible in under 16 GB of VRAM.

Community

You are always welcome to join the ExLlama discord server ←🎮

🤗 HuggingFace repos

A selection of EXL3-quantized models is available here. Also shout out the following lovely people:

• ArtusDev
• MikeRoz
• MetaphoricalCode
• Ready.Art
• isogen

Acknowledgements

This project owes its existence to a wonderful community of FOSS developers and some very generous supporters (🐈❤️!) The following projects in particular deserve a special mention:

• TabbyAPI
• PyTorch
• FlashAttention
• QTIP
• Transformers
• Marlin
• Flash Linear Attention