[docs] refine optimizers, integrations, etc (#1125)

* optim, integration

* toctree

* feedback

Author: stevhliu

docs/source/_toctree.yml

@@ -22,12 +22,12 @@
     title: FAQs
 - title: Explanation
   sections:
-  - local: resources
+  - local: explanations/optimizers
+    title: 8-bit optimizers
+  - local: explanations/resources
     title: Papers, resources & how to cite
 - title: API reference
   sections:
-  - local: reference/quantization
-    title: Quantization
   - title: Optimizers
     sections:
     - local: reference/optim/optim_overview

docs/source/explanations/optimizers.mdx

@@ -0,0 +1,51 @@
+# 8-bit optimizers
+
+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.
+
+<div class="flex justify-center">
+    <figure>
+        <img class="rounded-xl" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/bitsandbytes/optimizer_largest_model.png"/>
+        <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.
+
+<div class="flex justify-center">
+    <figure>
+        <img class="rounded-xl" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/bitsandbytes/optimizer_comparison.png"/>
+        <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).

docs/source/resources.mdx renamed to docs/source/explanations/resources.mdx

@@ -1,19 +1,13 @@
-# `bitsandbytes`
+# bitsandbytes
 
-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.
-
-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.
-
-## API documentation
-
-- [Quantization](quantization)
-- [Integrations](integrations)
-- [Optimizers](optimizers)
+* 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.
+* 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.
 
 # 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.

docs/source/index.mdx

@@ -21,7 +21,7 @@ To install from PyPI.
 pip install bitsandbytes
 ```
 
-## Alternative: Compiling from source
+## Compile from source
 
 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:
 

docs/source/installation.mdx

@@ -1,31 +1,89 @@
-# Transformers
+# Integrations
 
-With Transformers it's very easy to load any model in 4 or 8-bit, quantizing them on the fly with `bitsandbytes` primitives.
+bitsandbytes is widely integrated with many of the libraries in the Hugging Face and wider PyTorch ecosystem. This guide provides a brief overview of the integrations and how to use bitsandbytes with them. For more details, you should refer to the linked documentation for each library.
 
-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).
+> [!TIP]
+> Learn more in the bitsandbytes Transformers integration [guide](https://huggingface.co/docs/transformers/quantization#bitsandbytes).
+
+With Transformers, 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]
-> **Beware: bf16 is the optimal compute data type!**
->
-> 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:
+> 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`]!
 
 ```py
-import torch
-from transformers import BitsAndBytesConfig
+from transformers import AutoModelForCausalLM, BitsAndBytesConfig
 
 quantization_config = BitsAndBytesConfig(load_in_4bit=True, bnb_4bit_compute_dtype=torch.bfloat16)
+model_4bit = AutoModelForCausalLM.from_pretrained(
+    "bigscience/bloom-1b7",
+    device_map=device_map,
+    quantization_config=quantization_config,
+)
+```
+
+### 8-bit optimizers
+
+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's prepare the 4-bit model from the section above for training.
+
+Call the [`~peft.prepare_model_for_kbit_training`] method to prepare the model for training. This only works for Transformers models!
+
+```py
+from peft import prepare_model_for_kbit_training
+
+model_4bit = prepare_model_for_kbit_training(model_4bit)
 ```
 
-# PEFT
-With `PEFT`, you can use QLoRA out of the box with `LoraConfig` and a 4-bit base model.
+Setup a [`~peft.LoraConfig`] to use QLoRA:
+
+```py
+from peft import LoraConfig
+
+config = LoraConfig(
+    r=16,
+    lora_alpha=8,
+    target_modules="all-linear",
+    lora_dropout=0.05
+    bias="none",
+    task_type="CAUSAL_LM"
+)
+```
 
-Please review the [bitsandbytes section in the PEFT docs](https://huggingface.co/docs/peft/developer_guides/quantization#quantize-a-model).
+Now call the [`~peft.get_peft_model`] function on your model and config to create a trainable [`PeftModel`].
+
+```py
+from peft import get_peft_model
+
+model = get_peft_model(model_4bit, config)
+```
 
-# Accelerate
+## Accelerate
 
-Bitsandbytes is also easily usable from within Accelerate, where you can quantize any PyTorch model simply by passing a quantization config; e.g:
+> [!TIP]
+> Learn more in the bitsandbytes Accelerate integration [guide](https://huggingface.co/docs/accelerate/usage_guides/quantization).
+
+bitsandbytes is also easily usable from Accelerate and you can quantize any PyTorch model by passing a [`~accelerate.utils.BnbQuantizationConfig`] with your desired settings, and then calling the [`~accelerate.utils.load_and_quantize_model`] function to quantize it.
 
 ```py
 from accelerate import init_empty_weights
@@ -55,37 +113,25 @@ quantized_model = load_and_quantize_model(
 )
 ```
 
-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).
-
-
-
-# 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
+## PyTorch Lightning and Lightning Fabric
 
-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.
+bitsandbytes is available from:
 
-Please review the [bitsandbytes section in the Lit-GPT quantization docs](https://github.com/Lightning-AI/lit-gpt/blob/main/tutorials/quantize.md).
+- [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.
+- [Lightning Fabric](https://lightning.ai/docs/fabric/stable/), a fast and lightweight way to scale PyTorch models without boilerplate.
 
+Learn more in the bitsandbytes PyTorch Lightning integration [guide](https://lightning.ai/docs/pytorch/stable/common/precision_intermediate.html#quantization-via-bitsandbytes).
 
 
-# Trainer for the optimizers
+## Lit-GPT
 
-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`).
+bitsandbytes is integrated with [Lit-GPT](https://github.com/Lightning-AI/lit-gpt), a hackable implementation of state-of-the-art open-source large language models. Lit-GPT is based on Lightning Fabric, and it can be used for quantization during training, finetuning, and inference.
 
-See the [official API docs for reference](https://huggingface.co/docs/transformers/main_classes/trainer#transformers.Trainer).
+Learn more in the bitsandbytes Lit-GPT integration [guide](https://github.com/Lightning-AI/lit-gpt/blob/main/tutorials/quantize.md).
 
-Here we point out to relevant doc sections in transformers / peft / Trainer + very briefly explain how these are integrated:
-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`):
+## Blog posts
 
-# Blog posts
+To learn in more detail about some of bitsandbytes integrations, take a look at the following blog posts:
 
-- [Making LLMs even more accessible with `bitsandbytes`, 4-bit quantization and QLoRA](https://huggingface.co/blog/4bit-transformers-bitsandbytes)
-- [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)
+- [Making LLMs even more accessible with bitsandbytes, 4-bit quantization and QLoRA](https://huggingface.co/blog/4bit-transformers-bitsandbytes)
+- [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)