installation.mdx

# Installation Guide

Welcome to the installation guide for the `bitsandbytes` library! This document provides step-by-step instructions to install `bitsandbytes` across various platforms and hardware configurations. The library primarily supports CUDA-based GPUs, but the team is actively working on enabling support for additional backends like AMD ROCm, Intel, and Apple Silicon.

> [!TIP]
> For a high-level overview of backend support and compatibility, see the [Multi-backend Support](#multi-backend) section.

## Table of Contents

- [CUDA](#cuda)
  - [Installation via PyPI](#cuda-pip)
  - [Compile from Source](#cuda-compile)
- [Multi-backend Support (Alpha Release)](#multi-backend)
  - [Supported Backends](#multi-backend-supported-backends)
  - [Pre-requisites](#multi-backend-pre-requisites)
  - [Installation](#multi-backend-pip)
  - [Compile from Source](#multi-backend-compile)
- [PyTorch CUDA Versions](#pytorch-cuda-versions)

## CUDA[[cuda]]

`bitsandbytes` is currently only supported on CUDA GPUs for CUDA versions **11.0 - 12.8**. However, there's an ongoing multi-backend effort under development, which is currently in alpha. If you're interested in providing feedback or testing, check out [the multi-backend section below](#multi-backend).

### Supported CUDA Configurations[[cuda-pip]]

The latest version of the distributed `bitsandbytes` package is built with the following configurations:

| **OS**      | **CUDA Toolkit** | **Host Compiler**         |
|-------------|------------------|----------------------|
| **Linux**   | 11.8 - 12.3      | GCC 11.4             |
|             | 12.4 - 12.8      | GCC 13.2             |
| **Windows** | 11.8 - 12.8      | MSVC 19.42+ (VS2022) |

For CUDA systems, ensure your hardware meets the following requirements:

| **Feature**                     | **Minimum Hardware Requirement**                              |
|---------------------------------|---------------------------------------------------------------|
| LLM.int8()                      | NVIDIA Turing (RTX 20 series, T4) or newer GPUs               |
| 8-bit optimizers/quantization   | NVIDIA Maxwell (GTX 900 series, TITAN X, M40) or newer GPUs * |
| NF4/FP4 quantization            | NVIDIA Maxwell (GTX 900 series, TITAN X, M40) or newer GPUs * |

> [!WARNING]
> `bitsandbytes >= 0.45.0` no longer supports Kepler GPUs.
>
> Support for Maxwell GPUs is deprecated and will be removed in a future release. For the best results, a Turing generation device or newer is recommended.

```bash
pip install bitsandbytes
```

### `pip install` pre-built wheel from latest `main` commit

If you would like to use new feature even before they are officially released and help us test them, feel free to install the wheel directly from our CI (*the wheel links will remain stable!*):

<hfoptions id="OS">
<hfoption id="Linux">

```
# Note, if you don't want to reinstall BNBs dependencies, append the `--no-deps` flag!
pip install --force-reinstall 'https://github.com/bitsandbytes-foundation/bitsandbytes/releases/download/continuous-release_main/bitsandbytes-0.44.2.dev0-py3-none-manylinux_2_24_x86_64.whl'
```

</hfoption>
<hfoption id="Windows">

```
# Note, if you don't want to reinstall BNBs dependencies, append the `--no-deps` flag!
pip install --force-reinstall 'https://github.com/bitsandbytes-foundation/bitsandbytes/releases/download/continuous-release_multi-backend-refactor/bitsandbytes-0.44.1.dev0-py3-none-macosx_13_1_arm64.whl'
```
</hfoption>
</hfoptions>

### Compile from source[[cuda-compile]]

> [!TIP]
> Don't hesitate to compile from source! The process is pretty straight forward and resilient. This might be needed for older CUDA versions or other less common configurations, which we don't support out of the box due to package size.

For Linux and Windows systems, compiling from source allows you to customize the build configurations. See below for detailed platform-specific instructions (see the `CMakeLists.txt` if you want to check the specifics and explore some additional options):

<hfoptions id="source">
<hfoption id="Linux">

To compile from source, you need CMake >= **3.22.1** and Python >= **3.9** installed. Make sure you have a compiler installed to compile C++ (`gcc`, `make`, headers, etc.).

For example, to install a compiler and CMake on Ubuntu:

```bash
apt-get install -y build-essential cmake
```

You should also install CUDA Toolkit by following the [NVIDIA CUDA Installation Guide for Linux](https://docs.nvidia.com/cuda/cuda-installation-guide-linux/index.html) guide from NVIDIA. The current minimum supported CUDA Toolkit version is **11.8**.

```bash
git clone https://github.com/bitsandbytes-foundation/bitsandbytes.git && cd bitsandbytes/
cmake -DCOMPUTE_BACKEND=cuda -S .
make
pip install -e .   # `-e` for "editable" install, when developing BNB (otherwise leave that out)
```

> [!TIP]
> If you have multiple versions of CUDA installed or installed it in a non-standard location, please refer to CMake CUDA documentation for how to configure the CUDA compiler.

</hfoption>
<hfoption id="Windows">

Windows systems require Visual Studio with C++ support as well as an installation of the CUDA SDK.

To compile from source, you need CMake >= **3.22.1** and Python >= **3.9** installed. You should also install CUDA Toolkit by following the [CUDA Installation Guide for Windows](https://docs.nvidia.com/cuda/cuda-installation-guide-microsoft-windows/index.html) guide from NVIDIA. The current minimum supported CUDA Toolkit version is **11.8**.

```bash
git clone https://github.com/bitsandbytes-foundation/bitsandbytes.git && cd bitsandbytes/
cmake -DCOMPUTE_BACKEND=cuda -S .
cmake --build . --config Release
pip install -e .   # `-e` for "editable" install, when developing BNB (otherwise leave that out)
```

Big thanks to [wkpark](https://github.com/wkpark), [Jamezo97](https://github.com/Jamezo97), [rickardp](https://github.com/rickardp), [akx](https://github.com/akx) for their amazing contributions to make bitsandbytes compatible with Windows.

</hfoption>
</hfoptions>

### PyTorch CUDA versions[[pytorch-cuda-versions]]

Some bitsandbytes features may need a newer CUDA version than the one currently supported by PyTorch binaries from Conda and pip. In this case, you should follow these instructions to load a precompiled bitsandbytes binary.

1. Determine the path of the CUDA version you want to use. Common paths include:

* `/usr/local/cuda`
* `/usr/local/cuda-XX.X` where `XX.X` is the CUDA version number

Then locally install the CUDA version you need with this script from bitsandbytes:

```bash
wget https://raw.githubusercontent.com/bitsandbytes-foundation/bitsandbytes/main/install_cuda.sh
# Syntax cuda_install CUDA_VERSION INSTALL_PREFIX EXPORT_TO_BASH
#   CUDA_VERSION in {118, 120, 121, 122, 123, 124, 125, 126, 128}
#   EXPORT_TO_BASH in {0, 1} with 0=False and 1=True

# For example, the following installs CUDA 12.6 to ~/local/cuda-12.6 and exports the path to your .bashrc

bash install_cuda.sh 126 ~/local 1
```

2. Set the environment variables `BNB_CUDA_VERSION` and `LD_LIBRARY_PATH` by manually overriding the CUDA version installed by PyTorch.

> [!TIP]
> It is recommended to add the following lines to the `.bashrc` file to make them permanent.

```bash
export BNB_CUDA_VERSION=<VERSION>
export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:<PATH>
```

For example, to use a local install path:

```bash
export BNB_CUDA_VERSION=126
export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/home/YOUR_USERNAME/local/cuda-12.6
```

3. Now when you launch bitsandbytes with these environment variables, the PyTorch CUDA version is overridden by the new CUDA version (in this example, version 12.6) and a different bitsandbytes library is loaded.

## Multi-backend Support (Alpha Release)[[multi-backend]]

> [!TIP]
> This functionality is currently in preview and not yet production-ready. We very much welcome community feedback, contributions and leadership on topics like Apple Silicon as well as other less common accellerators! For more information, see [this guide on multi-backend support](./non_cuda_backends).

**Link to give us feedback** (bugs, install issues, perf results, requests, etc.)**:**

<hfoptions id="platform">
<hfoption id="ROCm">

[**Multi-backend refactor: Alpha release (AMD ROCm ONLY)**](https://github.com/bitsandbytes-foundation/bitsandbytes/discussions/1339)

</hfoption>
<hfoption id="Intel CPU+GPU">

[**Multi-backend refactor: Alpha release (INTEL ONLY)**](https://github.com/bitsandbytes-foundation/bitsandbytes/discussions/1338)

</hfoption>
<hfoption id="Apple Silicon / Metal (MPS)">

[**Github Discussion space on coordinating the kickoff of MPS backend development**](https://github.com/bitsandbytes-foundation/bitsandbytes/discussions/1340)

</hfoption>
</hfoptions>

### Supported Backends[[multi-backend-supported-backends]]

| **Backend** | **Supported Versions** | **Python versions** | **Architecture Support** | **Status** |
|-------------|------------------------|---------------------------|-------------------------|------------|
| **AMD ROCm** | 6.1+                   | 3.10+                     | minimum CDNA - `gfx90a`, RDNA - `gfx1100` | Alpha      |
| **Apple Silicon (MPS)** | WIP                        | 3.10+                     | M1/M2 chips                    | Planned    |
| **Intel CPU** | v2.4.0+ (`ipex`)         | 3.10+                     | Intel CPU | Alpha |
| **Intel GPU** | v2.4.0+ (`ipex`)         | 3.10+                     | Intel GPU | Experimental |
| **Ascend NPU** | 2.1.0+ (`torch_npu`)         | 3.10+                     | Ascend NPU | Experimental |

For each supported backend, follow the respective instructions below:

### Pre-requisites[[multi-backend-pre-requisites]]

To use bitsandbytes non-CUDA backends, be sure to install:

```
pip install "transformers>=4.45.1"
```

<hfoptions id="backend">
<hfoption id="AMD ROCm">

> [!WARNING]
> Pre-compiled binaries are only built for ROCm versions `6.1.2`/`6.2.4`/`6.3.2` and `gfx90a`, `gfx942`, `gfx1100` GPU architectures. [Find the pip install instructions here](#multi-backend-pip).
>
> Other supported versions that don't come with pre-compiled binaries [can be compiled for with these instructions](#multi-backend-compile).
>
> **Windows is not supported for the ROCm backend**; also not WSL2 to our knowledge.

> [!TIP]
> If you would like to install ROCm and PyTorch on bare metal, skip the Docker steps and refer to ROCm's official guides at [ROCm installation overview](https://rocm.docs.amd.com/projects/install-on-linux/en/latest/tutorial/install-overview.html#rocm-install-overview) and [Installing PyTorch for ROCm](https://rocm.docs.amd.com/projects/install-on-linux/en/latest/how-to/3rd-party/pytorch-install.html#using-wheels-package) (Step 3 of wheels build for quick installation). Special note: please make sure to get the respective ROCm-specific PyTorch wheel for the installed ROCm version, e.g. `https://download.pytorch.org/whl/nightly/rocm6.2/`!

```bash
# Create a docker container with latest ROCm image, which includes ROCm libraries
docker pull rocm/dev-ubuntu-22.04:6.1.2-complete
docker run -it --device=/dev/kfd --device=/dev/dri --group-add video rocm/dev-ubuntu-22.04:6.1.2-complete
apt-get update && apt-get install -y git && cd home

# Install pytorch compatible with above ROCm version
pip install torch --index-url https://download.pytorch.org/whl/rocm6.1/
```

</hfoption>
<hfoption id="Intel CPU + GPU">

Compatible hardware and functioning `import intel_extension_for_pytorch as ipex` capable environment with Python `3.10` as the minimum requirement.

Please refer to [the official Intel installations instructions](https://intel.github.io/intel-extension-for-pytorch/index.html#installation?platform=cpu&version=v2.4.0%2bcpu&os=linux%2fwsl2) for guidance on how to pip install the necessary `intel_extension_for_pytorch` dependency.

</hfoption>
<hfoption id="Apple Silicon (MPS)">

> [!TIP]
> Apple Silicon support is still a WIP. Please visit and write us in [this Github Discussion space on coordinating the kickoff of MPS backend development](https://github.com/bitsandbytes-foundation/bitsandbytes/discussions/1340) and coordinate a community-led effort to implement this backend.

</hfoption>
</hfoptions>

### Installation

You can install the pre-built wheels for each backend, or compile from source for custom configurations.

#### Pre-built Wheel Installation (recommended)[[multi-backend-pip]]

<hfoptions id="platform">
<hfoption id="Linux">

```
# Note, if you don't want to reinstall BNBs dependencies, append the `--no-deps` flag!
pip install --force-reinstall 'https://github.com/bitsandbytes-foundation/bitsandbytes/releases/download/continuous-release_multi-backend-refactor/bitsandbytes-0.44.1.dev0-py3-none-manylinux_2_24_x86_64.whl'
```

</hfoption>
<hfoption id="Windows">

```
# Note, if you don't want to reinstall BNBs dependencies, append the `--no-deps` flag!
pip install --force-reinstall 'https://github.com/bitsandbytes-foundation/bitsandbytes/releases/download/continuous-release_multi-backend-refactor/bitsandbytes-0.44.1.dev0-py3-none-win_amd64.whl'
```

</hfoption>
<hfoption id="Ascend NPU">

Compatible hardware and functioning `import torch_npu` capable environment with Python `3.10` as the minimum requirement.

Please refer to [the official Ascend installations instructions](https://www.hiascend.com/document/detail/zh/Pytorch/60RC3/configandinstg/instg/insg_0001.html) for guidance on how to pip install the necessary `torch_npu` dependency.

</hfoption>
<hfoption id="Mac">

> [!WARNING]
> bitsandbytes does not yet support Apple Silicon / Metal with a dedicated backend. However, the build infrastructure is in place and the below pip install will eventually provide Apple Silicon support as it becomes available on the `multi-backend-refactor` branch based on community contributions.

```
# Note, if you don't want to reinstall BNBs dependencies, append the `--no-deps` flag!
pip install --force-reinstall 'https://github.com/bitsandbytes-foundation/bitsandbytes/releases/download/continuous-release_multi-backend-refactor/bitsandbytes-0.44.1.dev0-py3-none-macosx_13_1_arm64.whl'
```

</hfoption>
</hfoptions>

#### Compile from Source[[multi-backend-compile]]

<hfoptions id="backend">
<hfoption id="AMD ROCm">

#### AMD GPU

bitsandbytes is fully supported from ROCm 6.1 onwards (currently in alpha release).

```bash
# Install bitsandbytes from source
# Clone bitsandbytes repo, ROCm backend is currently enabled on multi-backend-refactor branch
git clone -b multi-backend-refactor https://github.com/bitsandbytes-foundation/bitsandbytes.git && cd bitsandbytes/

# Compile & install
apt-get install -y build-essential cmake  # install build tools dependencies, unless present
cmake -DCOMPUTE_BACKEND=hip -S .  # Use -DBNB_ROCM_ARCH="gfx90a;gfx942" to target specific gpu arch
make
pip install -e .   # `-e` for "editable" install, when developing BNB (otherwise leave that out)
```

</hfoption>
<hfoption id="Intel CPU + GPU">

#### Intel CPU

> [!TIP]
> Intel CPU backend only supports building from source; for now, please follow the instructions below.

Similar to the CUDA case, you can compile bitsandbytes from source for Linux and Windows systems.

The below commands are for Linux. For installing on Windows, please adapt the below commands according to the same pattern as described [the section above on compiling from source under the Windows tab](#cuda-compile).

```
git clone --depth 1 -b multi-backend-refactor https://github.com/bitsandbytes-foundation/bitsandbytes.git && cd bitsandbytes/
pip install intel_extension_for_pytorch
cmake -DCOMPUTE_BACKEND=cpu -S .
make
pip install -e .   # `-e` for "editable" install, when developing BNB (otherwise leave that out)
```

</hfoption>
<hfoption id="Ascend NPU">

#### Ascend NPU

> [!TIP]
> Ascend NPU backend only supports building from source; for now, please follow the instructions below.


```
# Install bitsandbytes from source
# Clone bitsandbytes repo, Ascend NPU backend is currently enabled on multi-backend-refactor branch
git clone -b multi-backend-refactor https://github.com/bitsandbytes-foundation/bitsandbytes.git && cd bitsandbytes/

# Compile & install
apt-get install -y build-essential cmake  # install build tools dependencies, unless present
cmake -DCOMPUTE_BACKEND=npu -S .
make
pip install -e .   # `-e` for "editable" install, when developing BNB (otherwise leave that out)
```


</hfoption>
<hfoption id="Apple Silicon (MPS)">

#### Apple Silicon

WIP

</hfoption>
</hfoptions>