diffusers/docs/source/en/quantization/bitsandbytes.md

<!--Copyright 2025 The HuggingFace Team. All rights reserved.

Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at

http://www.apache.org/licenses/LICENSE-2.0

Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.

-->

# bitsandbytes

[bitsandbytes](https://huggingface.co/docs/bitsandbytes/index) is the easiest option for quantizing a model to 8 and 4-bit. 8-bit quantization multiplies outliers in fp16 with non-outliers in int8, converts the non-outlier values back to fp16, and then adds them together to return the weights in fp16. This reduces the degradative effect outlier values have on a model's performance.

4-bit quantization compresses a model even further, and it is commonly used with [QLoRA](https://hf.co/papers/2305.14314) to finetune quantized LLMs.

This guide demonstrates how quantization can enable running
[FLUX.1-dev](https://huggingface.co/black-forest-labs/FLUX.1-dev)
on less than 16GB of VRAM and even on a free Google
Colab instance.

![comparison image](https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/quant-bnb/comparison.png)

To use bitsandbytes, make sure you have the following libraries installed:

```bash
pip install diffusers transformers accelerate bitsandbytes -U
```

Now you can quantize a model by passing a [`BitsAndBytesConfig`] to [`~ModelMixin.from_pretrained`]. This works for any model in any modality, as long as it supports loading with [Accelerate](https://hf.co/docs/accelerate/index) and contains `torch.nn.Linear` layers.

<hfoptions id="bnb">
<hfoption id="8-bit">

Quantizing a model in 8-bit halves the memory-usage:

bitsandbytes is supported in both Transformers and Diffusers, so you can quantize both the
[`FluxTransformer2DModel`] and [`~transformers.T5EncoderModel`].

For Ada and higher-series GPUs. we recommend changing `torch_dtype` to `torch.bfloat16`.

> [!TIP]
> The [`CLIPTextModel`] and [`AutoencoderKL`] aren't quantized because they're already small in size and because [`AutoencoderKL`] only has a few `torch.nn.Linear` layers.

```py
from diffusers import BitsAndBytesConfig as DiffusersBitsAndBytesConfig
from transformers import BitsAndBytesConfig as TransformersBitsAndBytesConfig
import torch
from diffusers import AutoModel
from transformers import T5EncoderModel

quant_config = TransformersBitsAndBytesConfig(load_in_8bit=True,)

text_encoder_2_8bit = T5EncoderModel.from_pretrained(
    "black-forest-labs/FLUX.1-dev",
    subfolder="text_encoder_2",
    quantization_config=quant_config,
    torch_dtype=torch.float16,
)

quant_config = DiffusersBitsAndBytesConfig(load_in_8bit=True,)

transformer_8bit = AutoModel.from_pretrained(
    "black-forest-labs/FLUX.1-dev",
    subfolder="transformer",
    quantization_config=quant_config,
    torch_dtype=torch.float16,
)
```

By default, all the other modules such as `torch.nn.LayerNorm` are converted to `torch.float16`. You can change the data type of these modules with the `torch_dtype` parameter.

```diff
transformer_8bit = AutoModel.from_pretrained(
    "black-forest-labs/FLUX.1-dev",
    subfolder="transformer",
    quantization_config=quant_config,
+   torch_dtype=torch.float32,
)
```

Let's generate an image using our quantized models.

Setting `device_map="auto"` automatically fills all available space on the GPU(s) first, then the
CPU, and finally, the hard drive (the absolute slowest option) if there is still not enough memory.

```py
from diffusers import FluxPipeline

pipe = FluxPipeline.from_pretrained(
    "black-forest-labs/FLUX.1-dev",
    transformer=transformer_8bit,
    text_encoder_2=text_encoder_2_8bit,
    torch_dtype=torch.float16,
    device_map="auto",
)

pipe_kwargs = {
    "prompt": "A cat holding a sign that says hello world",
    "height": 1024,
    "width": 1024,
    "guidance_scale": 3.5,
    "num_inference_steps": 50,
    "max_sequence_length": 512,
}

image = pipe(**pipe_kwargs, generator=torch.manual_seed(0),).images[0]
```

<div class="flex justify-center">
   <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/quant-bnb/8bit.png"/>
</div>

When there is enough memory, you can also directly move the pipeline to the GPU with `.to("cuda")` and apply [`~DiffusionPipeline.enable_model_cpu_offload`] to optimize GPU memory usage.

Once a model is quantized, you can push the model to the Hub with the [`~ModelMixin.push_to_hub`] method. The quantization `config.json` file is pushed first, followed by the quantized model weights. You can also save the serialized 8-bit models locally with [`~ModelMixin.save_pretrained`].

</hfoption>
<hfoption id="4-bit">

Quantizing a model in 4-bit reduces your memory-usage by 4x:

bitsandbytes is supported in both Transformers and Diffusers, so you can can quantize both the
[`FluxTransformer2DModel`] and [`~transformers.T5EncoderModel`].

For Ada and higher-series GPUs. we recommend changing `torch_dtype` to `torch.bfloat16`.

> [!TIP]
> The [`CLIPTextModel`] and [`AutoencoderKL`] aren't quantized because they're already small in size and because [`AutoencoderKL`] only has a few `torch.nn.Linear` layers.

```py
from diffusers import BitsAndBytesConfig as DiffusersBitsAndBytesConfig
from transformers import BitsAndBytesConfig as TransformersBitsAndBytesConfig
import torch
from diffusers import AutoModel
from transformers import T5EncoderModel

quant_config = TransformersBitsAndBytesConfig(load_in_4bit=True,)

text_encoder_2_4bit = T5EncoderModel.from_pretrained(
    "black-forest-labs/FLUX.1-dev",
    subfolder="text_encoder_2",
    quantization_config=quant_config,
    torch_dtype=torch.float16,
)

quant_config = DiffusersBitsAndBytesConfig(load_in_4bit=True,)

transformer_4bit = AutoModel.from_pretrained(
    "black-forest-labs/FLUX.1-dev",
    subfolder="transformer",
    quantization_config=quant_config,
    torch_dtype=torch.float16,
)
```

By default, all the other modules such as `torch.nn.LayerNorm` are converted to `torch.float16`. You can change the data type of these modules with the `torch_dtype` parameter.

```diff
transformer_4bit = AutoModel.from_pretrained(
    "black-forest-labs/FLUX.1-dev",
    subfolder="transformer",
    quantization_config=quant_config,
+   torch_dtype=torch.float32,
)
```

Let's generate an image using our quantized models.

Setting `device_map="auto"` automatically fills all available space on the GPU(s) first, then the CPU, and finally, the hard drive (the absolute slowest option) if there is still not enough memory.

```py
from diffusers import FluxPipeline

pipe = FluxPipeline.from_pretrained(
    "black-forest-labs/FLUX.1-dev",
    transformer=transformer_4bit,
    text_encoder_2=text_encoder_2_4bit,
    torch_dtype=torch.float16,
    device_map="auto",
)

pipe_kwargs = {
    "prompt": "A cat holding a sign that says hello world",
    "height": 1024,
    "width": 1024,
    "guidance_scale": 3.5,
    "num_inference_steps": 50,
    "max_sequence_length": 512,
}

image = pipe(**pipe_kwargs, generator=torch.manual_seed(0),).images[0]
```

<div class="flex justify-center">
   <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/quant-bnb/4bit.png"/>
</div>

When there is enough memory, you can also directly move the pipeline to the GPU with `.to("cuda")` and apply [`~DiffusionPipeline.enable_model_cpu_offload`] to optimize GPU memory usage.

Once a model is quantized, you can push the model to the Hub with the [`~ModelMixin.push_to_hub`] method. The quantization `config.json` file is pushed first, followed by the quantized model weights. You can also save the serialized 4-bit models locally with [`~ModelMixin.save_pretrained`].

</hfoption>
</hfoptions>

> [!WARNING]
> Training with 8-bit and 4-bit weights are only supported for training *extra* parameters.

Check your memory footprint with the `get_memory_footprint` method:

```py
print(model.get_memory_footprint())
```

Note that this only tells you the memory footprint of the model params and does _not_ estimate the inference memory requirements.

Quantized models can be loaded from the [`~ModelMixin.from_pretrained`] method without needing to specify the `quantization_config` parameters:

```py
from diffusers import AutoModel, BitsAndBytesConfig

quantization_config = BitsAndBytesConfig(load_in_4bit=True)

model_4bit = AutoModel.from_pretrained(
    "hf-internal-testing/flux.1-dev-nf4-pkg", subfolder="transformer"
)
```

## 8-bit (LLM.int8() algorithm)

> [!TIP]
> Learn more about the details of 8-bit quantization in this [blog post](https://huggingface.co/blog/hf-bitsandbytes-integration)!

This section explores some of the specific features of 8-bit models, such as outlier thresholds and skipping module conversion.

### Outlier threshold

An "outlier" is a hidden state value greater than a certain threshold, and these values are computed in fp16. While the values are usually normally distributed ([-3.5, 3.5]), this distribution can be very different for large models ([-60, 6] or [6, 60]). 8-bit quantization works well for values ~5, but beyond that, there is a significant performance penalty. A good default threshold value is 6, but a lower threshold may be needed for more unstable models (small models or finetuning).

To find the best threshold for your model, we recommend experimenting with the `llm_int8_threshold` parameter in [`BitsAndBytesConfig`]:

```py
from diffusers import AutoModel, BitsAndBytesConfig

quantization_config = BitsAndBytesConfig(
    load_in_8bit=True, llm_int8_threshold=10,
)

model_8bit = AutoModel.from_pretrained(
    "black-forest-labs/FLUX.1-dev",
    subfolder="transformer",
    quantization_config=quantization_config,
)
```

### Skip module conversion

For some models, you don't need to quantize every module to 8-bit which can actually cause instability. For example, for diffusion models like [Stable Diffusion 3](../api/pipelines/stable_diffusion/stable_diffusion_3), the `proj_out` module can be skipped using the `llm_int8_skip_modules` parameter in [`BitsAndBytesConfig`]:

```py
from diffusers import SD3Transformer2DModel, BitsAndBytesConfig

quantization_config = BitsAndBytesConfig(
    load_in_8bit=True, llm_int8_skip_modules=["proj_out"],
)

model_8bit = SD3Transformer2DModel.from_pretrained(
    "stabilityai/stable-diffusion-3-medium-diffusers",
    subfolder="transformer",
    quantization_config=quantization_config,
)
```


## 4-bit (QLoRA algorithm)

> [!TIP]
> Learn more about its details in this [blog post](https://huggingface.co/blog/4bit-transformers-bitsandbytes).

This section explores some of the specific features of 4-bit models, such as changing the compute data type, using the Normal Float 4 (NF4) data type, and using nested quantization.


### Compute data type

To speedup computation, you can change the data type from float32 (the default value) to bf16 using the `bnb_4bit_compute_dtype` parameter in [`BitsAndBytesConfig`]:

```py
import torch
from diffusers import BitsAndBytesConfig

quantization_config = BitsAndBytesConfig(load_in_4bit=True, bnb_4bit_compute_dtype=torch.bfloat16)
```

### Normal Float 4 (NF4)

NF4 is a 4-bit data type from the [QLoRA](https://hf.co/papers/2305.14314) paper, adapted for weights initialized from a normal distribution. You should use NF4 for training 4-bit base models. This can be configured with the `bnb_4bit_quant_type` parameter in the [`BitsAndBytesConfig`]:

```py
from diffusers import BitsAndBytesConfig as DiffusersBitsAndBytesConfig
from transformers import BitsAndBytesConfig as TransformersBitsAndBytesConfig

from diffusers import AutoModel
from transformers import T5EncoderModel

quant_config = TransformersBitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
)

text_encoder_2_4bit = T5EncoderModel.from_pretrained(
    "black-forest-labs/FLUX.1-dev",
    subfolder="text_encoder_2",
    quantization_config=quant_config,
    torch_dtype=torch.float16,
)

quant_config = DiffusersBitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
)

transformer_4bit = AutoModel.from_pretrained(
    "black-forest-labs/FLUX.1-dev",
    subfolder="transformer",
    quantization_config=quant_config,
    torch_dtype=torch.float16,
)
```

For inference, the `bnb_4bit_quant_type` does not have a huge impact on performance. However, to remain consistent with the model weights, you should use the `bnb_4bit_compute_dtype` and `torch_dtype` values.

### Nested quantization

Nested quantization is a technique that can save additional memory at no additional performance cost. This feature performs a second quantization of the already quantized weights to save an additional 0.4 bits/parameter. 

```py
from diffusers import BitsAndBytesConfig as DiffusersBitsAndBytesConfig
from transformers import BitsAndBytesConfig as TransformersBitsAndBytesConfig

from diffusers import AutoModel
from transformers import T5EncoderModel

quant_config = TransformersBitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_use_double_quant=True,
)

text_encoder_2_4bit = T5EncoderModel.from_pretrained(
    "black-forest-labs/FLUX.1-dev",
    subfolder="text_encoder_2",
    quantization_config=quant_config,
    torch_dtype=torch.float16,
)

quant_config = DiffusersBitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_use_double_quant=True,
)

transformer_4bit = AutoModel.from_pretrained(
    "black-forest-labs/FLUX.1-dev",
    subfolder="transformer",
    quantization_config=quant_config,
    torch_dtype=torch.float16,
)
```

## Dequantizing `bitsandbytes` models

Once quantized, you can dequantize a model to its original precision, but this might result in a small loss of quality. Make sure you have enough GPU RAM to fit the dequantized model. 

```python
from diffusers import BitsAndBytesConfig as DiffusersBitsAndBytesConfig
from transformers import BitsAndBytesConfig as TransformersBitsAndBytesConfig

from diffusers import AutoModel
from transformers import T5EncoderModel

quant_config = TransformersBitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_use_double_quant=True,
)

text_encoder_2_4bit = T5EncoderModel.from_pretrained(
    "black-forest-labs/FLUX.1-dev",
    subfolder="text_encoder_2",
    quantization_config=quant_config,
    torch_dtype=torch.float16,
)

quant_config = DiffusersBitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_use_double_quant=True,
)

transformer_4bit = AutoModel.from_pretrained(
    "black-forest-labs/FLUX.1-dev",
    subfolder="transformer",
    quantization_config=quant_config,
    torch_dtype=torch.float16,
)

text_encoder_2_4bit.dequantize()
transformer_4bit.dequantize()
```

## torch.compile

Speed up inference with `torch.compile`. Make sure you have the latest `bitsandbytes` installed and we also recommend installing [PyTorch nightly](https://pytorch.org/get-started/locally/).

<hfoptions id="bnb">
<hfoption id="8-bit">
```py
torch._dynamo.config.capture_dynamic_output_shape_ops = True

quant_config = DiffusersBitsAndBytesConfig(load_in_8bit=True)
transformer_4bit = AutoModel.from_pretrained(
    "black-forest-labs/FLUX.1-dev",
    subfolder="transformer",
    quantization_config=quant_config,
    torch_dtype=torch.float16,
)
transformer_4bit.compile(fullgraph=True)
```

</hfoption>
<hfoption id="4-bit">

```py
quant_config = DiffusersBitsAndBytesConfig(load_in_4bit=True)
transformer_4bit = AutoModel.from_pretrained(
    "black-forest-labs/FLUX.1-dev",
    subfolder="transformer",
    quantization_config=quant_config,
    torch_dtype=torch.float16,
)
transformer_4bit.compile(fullgraph=True)
```
</hfoption>
</hfoptions>

On an RTX 4090 with compilation, 4-bit Flux generation completed in 25.809 seconds versus 32.570 seconds without.

Check out the [benchmarking script](https://gist.github.com/sayakpaul/0db9d8eeeb3d2a0e5ed7cf0d9ca19b7d) for more details.

## Resources

* [End-to-end notebook showing Flux.1 Dev inference in a free-tier Colab](https://gist.github.com/sayakpaul/c76bd845b48759e11687ac550b99d8b4)
* [Training](https://github.com/huggingface/diffusers/blob/8c661ea586bf11cb2440da740dd3c4cf84679b85/examples/dreambooth/README_hidream.md#using-quantization)