Stateful optimizers maintain gradient statistics over time, for example, the exponentially smoothed sum (SGD with momentum) or squared sum (Adam) of past gradient values. This state can be used to accelerate optimization compared to plain stochastic gradient descent, but uses memory that might otherwise be allocated to model parameters. As a result, this limits the maximum size of models that can be trained in practice. Now take a look at the biggest models that can be trained with 8-bit optimizers.
<figcaption class="text-center">Depending on your GPU size, you can train a much larger model with a 8-bit optimizer.</figcaption>
</figure>
</div>
bitsandbytes optimizers use 8-bit statistics, while maintaining the performance levels of using 32-bit optimizer states.
To overcome the resulting computational, quantization and stability challenges, 8-bit optimizers have three components:
1. Block-wise quantization: divides input tensors into smaller blocks that are independently quantized, isolating outliers and distributing the error more equally over all bits. Each block is processed in parallel across cores, yielding faster optimization and high precision quantization.
2. Dynamic quantization: quantizes both small and large values with high precision.
3. Stable embedding layer: improves stability during optimization for models with word embeddings.
With these components, performing an optimizer update with 8-bit states is straightforward. The 8-bit optimizer states are dequantized to 32-bit before you perform the update, and then the states are quantized back to 8-bit for storage.
The 8-bit to 32-bit conversion happens element-by-element in registers, meaning no slow copies to GPU memory or additional temporary memory are needed to perform quantization and dequantization. For GPUs, this makes 8-bit optimizers much faster than regular 32-bit optimizers.
<figcaption class="text-center">A comparison of memory and time saved using 8-bit and 32-bit optimizers.</figcaption>
</figure>
</div>
## Stable embedding layer
The stable embedding layer improves the training stability of the standard word embedding layer for NLP tasks. It addresses the challenge of non-uniform input distributions and mitigates extreme gradient variations. This means the stable embedding layer can support more aggressive quantization strategies without compromising training stability, and it can help achieve stable training outcomes, which is particularly important for models dealing with diverse and complex language data.
There are three features of the stable embedding layer:
- Initialization: utilizes Xavier uniform initialization to maintain consistent variance, reducing the likelihood of large gradients.
- Normalization: incorporates layer normalization before adding positional embeddings, aiding in output stability.
- Optimizer states: employs 32-bit optimizer states exclusively for this layer to enhance stability, while the rest of the model may use standard 16-bit precision.
## Paged optimizers
Paged optimizers are built on top of the [unified memory](https://developer.nvidia.com/blog/unified-memory-cuda-beginners/) feature of CUDA. Unified memory provides a single memory space the GPU and CPU can easily access. While this feature is not supported by PyTorch, it has been added to bitsandbytes.
Paged optimizers works like regular CPU paging, which means that it *only becomes active if you run out of GPU memory*. When that happens, memory is transferred page-by-page from GPU to CPU. The memory is mapped, meaning that pages are pre-allocated on the CPU but they are not updated automatically. Pages are only updated if the memory is accessed or a swapping operation is launched.
The unified memory feature is less efficient than regular asynchronous memory transfers, and you usually won't be able to get full PCIe memory bandwidth utilization. If you do a manual prefetch, transfer speeds can be high but still only about half or worse than the full PCIe memory bandwidth (tested on 16x lanes PCIe 3.0).
This means performance depends highly on the particular use-case. For example, if you evict 1 GB of memory per forward-backward-optimizer loop, then you can expect about 50% of the PCIe bandwidth as time in the best case. So, 1 GB for PCIe 3.0 with 16x lanes would run at 16 GB/s, which is `1/(16*0.5) = 1/8 = 125ms` of overhead per optimizer step. Other overhead can be estimated for the particular use-case given a PCIe interface, lanes, and the memory evicted in each iteration.
Compared to CPU offloading, a paged optimizer has zero overhead if all the memory fits onto the device and only some overhead if some of memory needs to be evicted. For offloading, you usually offload fixed parts of the model and need to off and onload all this memory with each iteration through the model (sometimes twice for both forward and backward pass).
The `bitsandbytes` library is a lightweight Python wrapper around CUDA custom functions, in particular 8-bit optimizers, matrix multiplication (LLM.int8()), and 8 + 4-bit quantization functions.
bitsandbytes enables accessible large language models via k-bit quantization for PyTorch. bitsandbytes provides three main features for dramatically reducing memory consumption for inference and training:
The library includes quantization primitives for 8-bit & 4-bit operations, through `bitsandbytes.nn.Linear8bitLt` and `bitsandbytes.nn.Linear4bit` and 8bit optimizers through `bitsandbytes.optim` module.
* 8-bit optimizers uses block-wise quantization to maintain 32-bit performance at a small fraction of the memory cost.
* LLM.Int() or 8-bit quantization enables large language model inference with only half the required memory and without any performance degradation. This method is based on vector-wise quantization to quantize most features to 8-bits and separately treating outliers with 16-bit matrix multiplication.
There are ongoing efforts to support further hardware backends, i.e. Intel CPU + GPU, AMD GPU, Apple Silicon. Windows support is on its way as well.
* QLoRA or 4-bit quantization enables large language model training with several memory-saving techniques that don't compromise performance. This method quantizes a model to 4-bits and inserts a small set of trainable low-rank adaptation (LoRA) weights to allow training.
## API documentation
- [Quantization](quantization)
- [Integrations](integrations)
- [Optimizers](optimizers)
# License
# License
The majority of bitsandbytes is licensed under MIT, however portions of the project are available under separate license terms, as the parts adapted from Pytorch are licensed under the BSD license.
bitsandbytes is MIT licensed.
We thank Fabio Cannizzo for his work on [FastBinarySearch](https://github.com/fabiocannizzo/FastBinarySearch) which we use for CPU quantization.
We thank Fabio Cannizzo for his work on [FastBinarySearch](https://github.com/fabiocannizzo/FastBinarySearch) which we use for CPU quantization.
To compile from source, you need CMake >= **3.22.1** and Python >= **3.8** 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:
To compile from source, you need CMake >= **3.22.1** and Python >= **3.8** 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:
Please review the [`bitsandbytes` section in the Transformers docs](https://huggingface.co/docs/transformers/main/en/quantization#bitsandbytes).
##Transformers
Details about the BitsAndBytesConfig can be found [here](https://huggingface.co/docs/transformers/v4.37.2/en/main_classes/quantization#transformers.BitsAndBytesConfig).
WithTransformers,it's very easy to load any model in 4 or 8-bit and quantize them on the fly. To configure the quantization parameters, specify them in the [`~transformers.BitsAndBytesConfig`] class.
For example, to load and quantize a model to 4-bits and use the bfloat16 data type for compute:
> [!WARNING]
> [!WARNING]
> **Beware: bf16 is the optimal compute data type!**
> bfloat16 is the optimal compute data type if your hardware supports it. The default is float32 for backward compatibility and numerical stability, but it can often lead to numerical instabilities. bfloat16 provides the best of both worlds, numerical stability equivalent to float32, but combined with the memory footprint and significant computation speedup of a 16-bit data type. Make sure to check if your hardware supports bfloat16 and if it does, configure it using the `bnb_4bit_compute_dtype` parameter in [`~transformers.BitsAndBytesConfig`]!
>
> If your hardware supports it, `bf16` is the optimal compute dtype. The default is `float32` for backward compatibility and numerical stability. `float16` often leads to numerical instabilities, but `bfloat16` provides the benefits of both worlds: numerical stability equivalent to float32, but combined with the memory footprint and significant computation speedup of a 16-bit data type. Therefore, be sure to check if your hardware supports `bf16` and configure it using the `bnb_4bit_compute_dtype` parameter in BitsAndBytesConfig:
```py
```py
import torch
from transformers import AutoModelForCausalLM, BitsAndBytesConfig
You can use any of the 8-bit or paged optimizers with Transformers by passing them to the [`~transformers.Trainer`] class on initialization. All bitsandbytes optimizers are supported by passing the correct string in the [`~transformers.TrainingArguments`] `optim` parameter. For example, to load a [`~bitsandbytes.optim.PagedAdamW32bit`] optimizer:
```py
from transformers import TrainingArguments, Trainer
training_args = TrainingArguments(
...,
optim="paged_adamw_32bit",
)
trainer = Trainer(model, training_args, ...)
trainer.train()
```
## PEFT
> [!TIP]
> Learn more in the bitsandbytes PEFT integration [guide](https://huggingface.co/docs/peft/developer_guides/quantization#quantization).
PEFT builds on the bitsandbytes Transformers integration, and extends it for training with a few more steps. Let'spreparethe4-bitmodelfromthesectionabovefortraining.
For further details, e.g. model saving, cpu-offloading andfine-tuning, please review the [`bitsandbytes` section in the Accelerate docs](https://huggingface.co/docs/accelerate/en/usage_guides/quantization).
##PyTorchLightningandLightningFabric
# PyTorch Lightning and Lightning Fabric
Bitsandbytes is available from within both
- [PyTorch Lightning](https://lightning.ai/docs/pytorch/stable/), a deep learning framework for professional AI researchers and machine learning engineers who need maximal flexibility without sacrificing performance at scale;
- and [Lightning Fabric](https://lightning.ai/docs/fabric/stable/), a fast and lightweight way to scale PyTorch models without boilerplate).
Please review the [bitsandbytes section in the PyTorch Lightning docs](https://lightning.ai/docs/pytorch/stable/common/precision_intermediate.html#quantization-via-bitsandbytes).
# Lit-GPT
Bitsandbytes is integrated into [Lit-GPT](https://github.com/Lightning-AI/lit-gpt), a hackable implementation of state-of-the-art open-source large language models, based on Lightning Fabric, where it can be used for quantization during training, finetuning, and inference.
bitsandbytesisavailablefrom:
Please review the [bitsandbytes section in the Lit-GPT quantization docs](https://github.com/Lightning-AI/lit-gpt/blob/main/tutorials/quantize.md).
You can use any of the 8-bit and/or paged optimizers by simple passing them to the `transformers.Trainer` class on initialization.All bnb optimizers are supported by passing the correct string in `TrainingArguments`'s `optim` attribute - e.g. (`paged_adamw_32bit`).
Here we point out to relevant doc sections in transformers / peft / Trainer + very briefly explain how these are integrated:
##Blogposts
e.g. for transformers state that you can load any model in 8-bit / 4-bit precision, for PEFT, you can use QLoRA out of the box with `LoraConfig` + 4-bit base model, for Trainer: all bnb optimizers are supported by passing the correct string in `TrainingArguments`'s `optim` attribute - e.g. (`paged_adamw_32bit`):
- [A Gentle Introduction to 8-bit Matrix Multiplication for transformers at scale using Hugging Face Transformers, Accelerate and `bitsandbytes`](https://huggingface.co/blog/hf-bitsandbytes-integration)
>8-bitoptimizersreducememoryusageandaccelerateoptimizationonawiderangeoftasks.However,since8-bitoptimizersonlyreducememoryproportionaltothenumberofparameters,modelsthatuselargeamountsofactivationmemory,suchasconvolutionalnetworks,don't really benefit from 8-bit optimizers. 8-bit optimizers are most beneficial for training or finetuning models with many parameters on highly memory-constrained GPUs.
Generally,our8-bitoptimizershavethreecomponents:
8-bit optimizers are a drop-in replacement for regular optimizers which means they also accept the same arguments as a regular optimizer. For NLP models, it is recommended to use the [`~nn.StableEmbedding`] class to improve stability and results.
By default, all parameter tensors with less than 4096 elements are kept at 32-bits even if you initialize those parameters with 8-bit optimizers. This is done because small tensors do not save much memory and often contain highly variable parameters (biases) or parameters that require high precision (batch norm, layer norm).
You can change this value with the `min_8bit_size` parameter. For example, if you want to optimize parameters to 8-bits only if the minimum size is 16384 values (it is recommended to use multiples of 4096):
Other parameters you can configure include the learning rate (`lr`), the decay rates (`betas`), the number of bits of the optimizer state (`optim_bits`), and percentile clipping (`percentile_clipping`) which can increase stability. For example, to initialize a 32-bit [`~bitsandbytes.optim.Adam`] optimizer with 5th percentile clipping:
To optimize some unstable parameters with 32-bit Adam and others with 8-bit Adam, use the [`~bitsandbytes.optim.GlobalOptimManager`] class to override the specific hyperparameters for a particular layer. You'llneedto:
Youcanalsooverridemultiplelayersatoncebypassingthemasalistandthenewhyperparametersasadictionary.Forexample,let's override the `model.special.weight` and `model.also_special.weight` layers to use sparse optimization and a lower learning and decay rate.
For a specific layer, we recommend overriding locally in each module. Pass the module, the parameter, and its attribute name to the [`~bitsandbytes.optim.GlobalOptimManager`]:
```py
```py
class MyModule(torch.nn.Module):
class MyModule(torch.nn.Module):
def __init__(d_in, d_out):
def __init__(d_in, d_out):
...
@@ -178,13 +89,6 @@ class MyModule(torch.nn.Module):
...
@@ -178,13 +89,6 @@ class MyModule(torch.nn.Module):
```
```
##APIDocs[[optim_api_docs]]
## Next steps
...underconstruction...
Herewe'll provide further auto-generated API docs soon. Please feel free to contribute doc-strings for the respective optimizers, as `bitsandbytes` is a community effort.
### StableEmbedding[[stable-emb-api]]
[[autodoc]] bitsandbytes.nn.StableEmbedding
For more conceptual details and explanation about 8-bit optimizers, take a look at the [8-bit optimizers](./explanations/optimizers) guide.
