text
stringlengths 2
11.8k
|
|---|
tf_model = TFDistilBertModel.from_pretrained("distilbert/distilbert-base-uncased")
When you load pretrained weights, the default model configuration is automatically loaded if the model is provided by 🤗 Transformers. However, you can still replace - some or all of - the default model configuration attributes with your own if you'd like:
tf_model = TFDistilBertModel.from_pretrained("distilbert/distilbert-base-uncased", config=my_config) |
tf_model = TFDistilBertModel.from_pretrained("distilbert/distilbert-base-uncased", config=my_config)
Model heads
At this point, you have a base DistilBERT model which outputs the hidden states. The hidden states are passed as inputs to a model head to produce the final output. 🤗 Transformers provides a different model head for each task as long as a model supports the task (i.e., you can't use DistilBERT for a sequence-to-sequence task like translation). |
For example, [DistilBertForSequenceClassification] is a base DistilBERT model with a sequence classification head. The sequence classification head is a linear layer on top of the pooled outputs.
from transformers import DistilBertForSequenceClassification
model = DistilBertForSequenceClassification.from_pretrained("distilbert/distilbert-base-uncased") |
from transformers import DistilBertForSequenceClassification
model = DistilBertForSequenceClassification.from_pretrained("distilbert/distilbert-base-uncased")
Easily reuse this checkpoint for another task by switching to a different model head. For a question answering task, you would use the [DistilBertForQuestionAnswering] model head. The question answering head is similar to the sequence classification head except it is a linear layer on top of the hidden states output. |
from transformers import DistilBertForQuestionAnswering
model = DistilBertForQuestionAnswering.from_pretrained("distilbert/distilbert-base-uncased")
``
</pt>
<tf>
For example, [TFDistilBertForSequenceClassification`] is a base DistilBERT model with a sequence classification head. The sequence classification head is a linear layer on top of the pooled outputs. |
from transformers import TFDistilBertForSequenceClassification
tf_model = TFDistilBertForSequenceClassification.from_pretrained("distilbert/distilbert-base-uncased")
Easily reuse this checkpoint for another task by switching to a different model head. For a question answering task, you would use the [TFDistilBertForQuestionAnswering] model head. The question answering head is similar to the sequence classification head except it is a linear layer on top of the hidden states output. |
from transformers import TFDistilBertForQuestionAnswering
tf_model = TFDistilBertForQuestionAnswering.from_pretrained("distilbert/distilbert-base-uncased")
Tokenizer
The last base class you need before using a model for textual data is a tokenizer to convert raw text to tensors. There are two types of tokenizers you can use with 🤗 Transformers: |
[PreTrainedTokenizer]: a Python implementation of a tokenizer.
[PreTrainedTokenizerFast]: a tokenizer from our Rust-based 🤗 Tokenizer library. This tokenizer type is significantly faster - especially during batch tokenization - due to its Rust implementation. The fast tokenizer also offers additional methods like offset mapping which maps tokens to their original words or characters.
Both tokenizers support common methods such as encoding and decoding, adding new tokens, and managing special tokens. |
Both tokenizers support common methods such as encoding and decoding, adding new tokens, and managing special tokens.
Not every model supports a fast tokenizer. Take a look at this table to check if a model has fast tokenizer support.
If you trained your own tokenizer, you can create one from your vocabulary file:
from transformers import DistilBertTokenizer
my_tokenizer = DistilBertTokenizer(vocab_file="my_vocab_file.txt", do_lower_case=False, padding_side="left") |
It is important to remember the vocabulary from a custom tokenizer will be different from the vocabulary generated by a pretrained model's tokenizer. You need to use a pretrained model's vocabulary if you are using a pretrained model, otherwise the inputs won't make sense. Create a tokenizer with a pretrained model's vocabulary with the [DistilBertTokenizer] class:
from transformers import DistilBertTokenizer
slow_tokenizer = DistilBertTokenizer.from_pretrained("distilbert/distilbert-base-uncased") |
from transformers import DistilBertTokenizer
slow_tokenizer = DistilBertTokenizer.from_pretrained("distilbert/distilbert-base-uncased")
Create a fast tokenizer with the [DistilBertTokenizerFast] class:
from transformers import DistilBertTokenizerFast
fast_tokenizer = DistilBertTokenizerFast.from_pretrained("distilbert/distilbert-base-uncased")
By default, [AutoTokenizer] will try to load a fast tokenizer. You can disable this behavior by setting use_fast=False in from_pretrained. |
By default, [AutoTokenizer] will try to load a fast tokenizer. You can disable this behavior by setting use_fast=False in from_pretrained.
Image processor
An image processor processes vision inputs. It inherits from the base [~image_processing_utils.ImageProcessingMixin] class.
To use, create an image processor associated with the model you're using. For example, create a default [ViTImageProcessor] if you are using ViT for image classification: |
from transformers import ViTImageProcessor
vit_extractor = ViTImageProcessor()
print(vit_extractor)
ViTImageProcessor {
"do_normalize": true,
"do_resize": true,
"image_processor_type": "ViTImageProcessor",
"image_mean": [
0.5,
0.5,
0.5
],
"image_std": [
0.5,
0.5,
0.5
],
"resample": 2,
"size": 224
}
If you aren't looking for any customization, just use the from_pretrained method to load a model's default image processor parameters. |
If you aren't looking for any customization, just use the from_pretrained method to load a model's default image processor parameters.
Modify any of the [ViTImageProcessor] parameters to create your custom image processor: |
from transformers import ViTImageProcessor
my_vit_extractor = ViTImageProcessor(resample="PIL.Image.BOX", do_normalize=False, image_mean=[0.3, 0.3, 0.3])
print(my_vit_extractor)
ViTImageProcessor {
"do_normalize": false,
"do_resize": true,
"image_processor_type": "ViTImageProcessor",
"image_mean": [
0.3,
0.3,
0.3
],
"image_std": [
0.5,
0.5,
0.5
],
"resample": "PIL.Image.BOX",
"size": 224
}
Backbone |
Computer vision models consist of a backbone, neck, and head. The backbone extracts features from an input image, the neck combines and enhances the extracted features, and the head is used for the main task (e.g., object detection). Start by initializing a backbone in the model config and specify whether you want to load pretrained weights or load randomly initialized weights. Then you can pass the model config to the model head.
For example, to load a ResNet backbone into a MaskFormer model with an instance segmentation head: |
Set use_pretrained_backbone=True to load pretrained ResNet weights for the backbone.
from transformers import MaskFormerConfig, MaskFormerForInstanceSegmentation, ResNetConfig
config = MaskFormerConfig(backbone="microsoft/resnet50", use_pretrained_backbone=True) # backbone and neck config
model = MaskFormerForInstanceSegmentation(config) # head
You could also load the backbone config separately and then pass it to the model config. |
You could also load the backbone config separately and then pass it to the model config.
from transformers import MaskFormerConfig, MaskFormerForInstanceSegmentation, ResNetConfig
backbone_config = ResNetConfig.from_pretrained("microsoft/resnet-50")
config = MaskFormerConfig(backbone_config=backbone_config)
model = MaskFormerForInstanceSegmentation(config)
Set use_pretrained_backbone=False to randomly initialize a ResNet backbone. |
Set use_pretrained_backbone=False to randomly initialize a ResNet backbone.
from transformers import MaskFormerConfig, MaskFormerForInstanceSegmentation, ResNetConfig
config = MaskFormerConfig(backbone="microsoft/resnet50", use_pretrained_backbone=False) # backbone and neck config
model = MaskFormerForInstanceSegmentation(config) # head
You could also load the backbone config separately and then pass it to the model config. |
You could also load the backbone config separately and then pass it to the model config.
from transformers import MaskFormerConfig, MaskFormerForInstanceSegmentation, ResNetConfig
backbone_config = ResNetConfig()
config = MaskFormerConfig(backbone_config=backbone_config)
model = MaskFormerForInstanceSegmentation(config) |
timm models are loaded with [TimmBackbone] and [TimmBackboneConfig].
thon
from transformers import TimmBackboneConfig, TimmBackbone
backbone_config = TimmBackboneConfig("resnet50")
model = TimmBackbone(config=backbone_config) |
Feature extractor
A feature extractor processes audio inputs. It inherits from the base [~feature_extraction_utils.FeatureExtractionMixin] class, and may also inherit from the [SequenceFeatureExtractor] class for processing audio inputs.
To use, create a feature extractor associated with the model you're using. For example, create a default [Wav2Vec2FeatureExtractor] if you are using Wav2Vec2 for audio classification: |
from transformers import Wav2Vec2FeatureExtractor
w2v2_extractor = Wav2Vec2FeatureExtractor()
print(w2v2_extractor)
Wav2Vec2FeatureExtractor {
"do_normalize": true,
"feature_extractor_type": "Wav2Vec2FeatureExtractor",
"feature_size": 1,
"padding_side": "right",
"padding_value": 0.0,
"return_attention_mask": false,
"sampling_rate": 16000
}
If you aren't looking for any customization, just use the from_pretrained method to load a model's default feature extractor parameters. |
If you aren't looking for any customization, just use the from_pretrained method to load a model's default feature extractor parameters.
Modify any of the [Wav2Vec2FeatureExtractor] parameters to create your custom feature extractor: |
Modify any of the [Wav2Vec2FeatureExtractor] parameters to create your custom feature extractor:
from transformers import Wav2Vec2FeatureExtractor
w2v2_extractor = Wav2Vec2FeatureExtractor(sampling_rate=8000, do_normalize=False)
print(w2v2_extractor)
Wav2Vec2FeatureExtractor {
"do_normalize": false,
"feature_extractor_type": "Wav2Vec2FeatureExtractor",
"feature_size": 1,
"padding_side": "right",
"padding_value": 0.0,
"return_attention_mask": false,
"sampling_rate": 8000
} |
Processor
For models that support multimodal tasks, 🤗 Transformers offers a processor class that conveniently wraps processing classes such as a feature extractor and a tokenizer into a single object. For example, let's use the [Wav2Vec2Processor] for an automatic speech recognition task (ASR). ASR transcribes audio to text, so you will need a feature extractor and a tokenizer.
Create a feature extractor to handle the audio inputs: |
from transformers import Wav2Vec2FeatureExtractor
feature_extractor = Wav2Vec2FeatureExtractor(padding_value=1.0, do_normalize=True)
Create a tokenizer to handle the text inputs:
from transformers import Wav2Vec2CTCTokenizer
tokenizer = Wav2Vec2CTCTokenizer(vocab_file="my_vocab_file.txt")
Combine the feature extractor and tokenizer in [Wav2Vec2Processor]:
from transformers import Wav2Vec2Processor
processor = Wav2Vec2Processor(feature_extractor=feature_extractor, tokenizer=tokenizer) |
With two basic classes - configuration and model - and an additional preprocessing class (tokenizer, image processor, feature extractor, or processor), you can create any of the models supported by 🤗 Transformers. Each of these base classes are configurable, allowing you to use the specific attributes you want. You can easily setup a model for training or modify an existing pretrained model to fine-tune. |
Methods and tools for efficient training on a single GPU
This guide demonstrates practical techniques that you can use to increase the efficiency of your model's training by
optimizing memory utilization, speeding up the training, or both. If you'd like to understand how GPU is utilized during
training, please refer to the Model training anatomy conceptual guide first. This guide
focuses on practical techniques. |
If you have access to a machine with multiple GPUs, these approaches are still valid, plus you can leverage additional methods outlined in the multi-GPU section.
When training large models, there are two aspects that should be considered at the same time:
Data throughput/training time
Model performance |
Maximizing the throughput (samples/second) leads to lower training cost. This is generally achieved by utilizing the GPU
as much as possible and thus filling GPU memory to its limit. If the desired batch size exceeds the limits of the GPU memory,
the memory optimization techniques, such as gradient accumulation, can help.
However, if the preferred batch size fits into memory, there's no reason to apply memory-optimizing techniques because they can
slow down the training. Just because one can use a large batch size, does not necessarily mean they should. As part of
hyperparameter tuning, you should determine which batch size yields the best results and then optimize resources accordingly.
The methods and tools covered in this guide can be classified based on the effect they have on the training process:
| Method/tool | Improves training speed | Optimizes memory utilization |
|:-----------------------------------------------------------|:------------------------|:-----------------------------|
| Batch size choice | Yes | Yes |
| Gradient accumulation | No | Yes |
| Gradient checkpointing | No | Yes |
| Mixed precision training | Yes | (No) |
| Optimizer choice | Yes | Yes |
| Data preloading | Yes | No |
| DeepSpeed Zero | No | Yes |
| torch.compile | Yes | No |
| Parameter-Efficient Fine Tuning (PEFT) | No | Yes | |
Note: when using mixed precision with a small model and a large batch size, there will be some memory savings but with a
large model and a small batch size, the memory use will be larger. |
You can combine the above methods to get a cumulative effect. These techniques are available to you whether you are
training your model with [Trainer] or writing a pure PyTorch loop, in which case you can configure these optimizations
with 🤗 Accelerate.
If these methods do not result in sufficient gains, you can explore the following options:
* Look into building your own custom Docker container with efficient softare prebuilds
* Consider a model that uses Mixture of Experts (MoE)
* Convert your model to BetterTransformer to leverage PyTorch native attention
Finally, if all of the above is still not enough, even after switching to a server-grade GPU like A100, consider moving
to a multi-GPU setup. All these approaches are still valid in a multi-GPU setup, plus you can leverage additional parallelism
techniques outlined in the multi-GPU section.
Batch size choice
To achieve optimal performance, start by identifying the appropriate batch size. It is recommended to use batch sizes and
input/output neuron counts that are of size 2^N. Often it's a multiple of 8, but it can be
higher depending on the hardware being used and the model's dtype.
For reference, check out NVIDIA's recommendation for input/output neuron counts and
batch size for
fully connected layers (which are involved in GEMMs (General Matrix Multiplications)).
Tensor Core Requirements
define the multiplier based on the dtype and the hardware. For instance, for fp16 data type a multiple of 8 is recommended, unless
it's an A100 GPU, in which case use multiples of 64.
For parameters that are small, consider also Dimension Quantization Effects.
This is where tiling happens and the right multiplier can have a significant speedup.
Gradient Accumulation
The gradient accumulation method aims to calculate gradients in smaller increments instead of computing them for the
entire batch at once. This approach involves iteratively calculating gradients in smaller batches by performing forward
and backward passes through the model and accumulating the gradients during the process. Once a sufficient number of
gradients have been accumulated, the model's optimization step is executed. By employing gradient accumulation, it
becomes possible to increase the effective batch size beyond the limitations imposed by the GPU's memory capacity.
However, it is important to note that the additional forward and backward passes introduced by gradient accumulation can
slow down the training process.
You can enable gradient accumulation by adding the gradient_accumulation_steps argument to [TrainingArguments]:
py
training_args = TrainingArguments(per_device_train_batch_size=1, gradient_accumulation_steps=4, **default_args)
In the above example, your effective batch size becomes 4.
Alternatively, use 🤗 Accelerate to gain full control over the training loop. Find the 🤗 Accelerate example
further down in this guide.
While it is advised to max out GPU usage as much as possible, a high number of gradient accumulation steps can
result in a more pronounced training slowdown. Consider the following example. Let's say, the per_device_train_batch_size=4
without gradient accumulation hits the GPU's limit. If you would like to train with batches of size 64, do not set the
per_device_train_batch_size to 1 and gradient_accumulation_steps to 64. Instead, keep per_device_train_batch_size=4
and set gradient_accumulation_steps=16. This results in the same effective batch size while making better use of
the available GPU resources.
For additional information, please refer to batch size and gradient accumulation benchmarks for RTX-3090
and A100.
Gradient Checkpointing
Some large models may still face memory issues even when the batch size is set to 1 and gradient accumulation is used.
This is because there are other components that also require memory storage.
Saving all activations from the forward pass in order to compute the gradients during the backward pass can result in
significant memory overhead. The alternative approach of discarding the activations and recalculating them when needed
during the backward pass, would introduce a considerable computational overhead and slow down the training process.
Gradient checkpointing offers a compromise between these two approaches and saves strategically selected activations
throughout the computational graph so only a fraction of the activations need to be re-computed for the gradients. For
an in-depth explanation of gradient checkpointing, refer to this great article.
To enable gradient checkpointing in the [Trainer], pass the corresponding a flag to [TrainingArguments]:
py
training_args = TrainingArguments(
per_device_train_batch_size=1, gradient_accumulation_steps=4, gradient_checkpointing=True, **default_args
)
Alternatively, use 🤗 Accelerate - find the 🤗 Accelerate example further in this guide. |
While gradient checkpointing may improve memory efficiency, it slows training by approximately 20%. |
Mixed precision training
Mixed precision training is a technique that aims to optimize the computational efficiency of training models by
utilizing lower-precision numerical formats for certain variables. Traditionally, most models use 32-bit floating point
precision (fp32 or float32) to represent and process variables. However, not all variables require this high precision
level to achieve accurate results. By reducing the precision of certain variables to lower numerical formats like 16-bit
floating point (fp16 or float16), we can speed up the computations. Because in this approach some computations are performed
in half-precision, while some are still in full precision, the approach is called mixed precision training.
Most commonly mixed precision training is achieved by using fp16 (float16) data types, however, some GPU architectures
(such as the Ampere architecture) offer bf16 and tf32 (CUDA internal data type) data types. Check
out the NVIDIA Blog to learn more about
the differences between these data types.
fp16
The main advantage of mixed precision training comes from saving the activations in half precision (fp16).
Although the gradients are also computed in half precision they are converted back to full precision for the optimization
step so no memory is saved here.
While mixed precision training results in faster computations, it can also lead to more GPU memory being utilized, especially for small batch sizes.
This is because the model is now present on the GPU in both 16-bit and 32-bit precision (1.5x the original model on the GPU).
To enable mixed precision training, set the fp16 flag to True:
py
training_args = TrainingArguments(per_device_train_batch_size=4, fp16=True, **default_args)
If you prefer to use 🤗 Accelerate, find the 🤗 Accelerate example further in this guide.
BF16
If you have access to an Ampere or newer hardware you can use bf16 for mixed precision training and evaluation. While
bf16 has a worse precision than fp16, it has a much bigger dynamic range. In fp16 the biggest number you can have
is 65535 and any number above that will result in an overflow. A bf16 number can be as large as 3.39e+38 (!) which
is about the same as fp32 - because both have 8-bits used for the numerical range.
You can enable BF16 in the 🤗 Trainer with:
python
training_args = TrainingArguments(bf16=True, **default_args)
TF32
The Ampere hardware uses a magical data type called tf32. It has the same numerical range as fp32 (8-bits), but instead
of 23 bits precision it has only 10 bits (same as fp16) and uses only 19 bits in total. It's "magical" in the sense that
you can use the normal fp32 training and/or inference code and by enabling tf32 support you can get up to 3x throughput
improvement. All you need to do is to add the following to your code:
python
import torch
torch.backends.cuda.matmul.allow_tf32 = True
torch.backends.cudnn.allow_tf32 = True
CUDA will automatically switch to using tf32 instead of fp32 where possible, assuming that the used GPU is from the Ampere series.
According to NVIDIA research, the
majority of machine learning training workloads show the same perplexity and convergence with tf32 training as with fp32.
If you're already using fp16 or bf16 mixed precision it may help with the throughput as well.
You can enable this mode in the 🤗 Trainer:
python
TrainingArguments(tf32=True, **default_args) |
tf32 can't be accessed directly via tensor.to(dtype=torch.tf32) because it is an internal CUDA data type. You need torch>=1.7 to use tf32 data types. |
For additional information on tf32 vs other precisions, please refer to the following benchmarks:
RTX-3090 and
A100.
Flash Attention 2
You can speedup the training throughput by using Flash Attention 2 integration in transformers. Check out the appropriate section in the single GPU section to learn more about how to load a model with Flash Attention 2 modules.
Optimizer choice
The most common optimizer used to train transformer models is Adam or AdamW (Adam with weight decay). Adam achieves
good convergence by storing the rolling average of the previous gradients; however, it adds an additional memory
footprint of the order of the number of model parameters. To remedy this, you can use an alternative optimizer.
For example if you have NVIDIA/apex installed for NVIDIA GPUs, or ROCmSoftwarePlatform/apex for AMD GPUs, adamw_apex_fused will give you the
fastest training experience among all supported AdamW optimizers.
[Trainer] integrates a variety of optimizers that can be used out of box: adamw_hf, adamw_torch, adamw_torch_fused,
adamw_apex_fused, adamw_anyprecision, adafactor, or adamw_bnb_8bit. More optimizers can be plugged in via a third-party implementation.
Let's take a closer look at two alternatives to AdamW optimizer:
1. adafactor which is available in [Trainer]
2. adamw_bnb_8bit is also available in Trainer, but a third-party integration is provided below for demonstration.
For comparison, for a 3B-parameter model, like “google-t5/t5-3b”:
* A standard AdamW optimizer will need 24GB of GPU memory because it uses 8 bytes for each parameter (83 => 24GB)
* Adafactor optimizer will need more than 12GB. It uses slightly more than 4 bytes for each parameter, so 43 and then some extra.
* 8bit BNB quantized optimizer will use only (2*3) 6GB if all optimizer states are quantized.
Adafactor
Adafactor doesn't store rolling averages for each element in weight matrices. Instead, it keeps aggregated information
(sums of rolling averages row- and column-wise), significantly reducing its footprint. However, compared to Adam,
Adafactor may have slower convergence in certain cases.
You can switch to Adafactor by setting optim="adafactor" in [TrainingArguments]:
py
training_args = TrainingArguments(per_device_train_batch_size=4, optim="adafactor", **default_args)
Combined with other approaches (gradient accumulation, gradient checkpointing, and mixed precision training)
you can notice up to 3x improvement while maintaining the throughput! However, as mentioned before, the convergence of
Adafactor can be worse than Adam.
8-bit Adam
Instead of aggregating optimizer states like Adafactor, 8-bit Adam keeps the full state and quantizes it. Quantization
means that it stores the state with lower precision and dequantizes it only for the optimization. This is similar to the
idea behind mixed precision training.
To use adamw_bnb_8bit, you simply need to set optim="adamw_bnb_8bit" in [TrainingArguments]:
py
training_args = TrainingArguments(per_device_train_batch_size=4, optim="adamw_bnb_8bit", **default_args)
However, we can also use a third-party implementation of the 8-bit optimizer for demonstration purposes to see how that can be integrated.
First, follow the installation guide in the GitHub repo to install the bitsandbytes library
that implements the 8-bit Adam optimizer.
Next you need to initialize the optimizer. This involves two steps:
* First, group the model's parameters into two groups - one where weight decay should be applied, and the other one where it should not. Usually, biases and layer norm parameters are not weight decayed.
* Then do some argument housekeeping to use the same parameters as the previously used AdamW optimizer. |
import bitsandbytes as bnb
from torch import nn
from transformers.trainer_pt_utils import get_parameter_names
training_args = TrainingArguments(per_device_train_batch_size=4, **default_args)
decay_parameters = get_parameter_names(model, [nn.LayerNorm])
decay_parameters = [name for name in decay_parameters if "bias" not in name]
optimizer_grouped_parameters = [
{
"params": [p for n, p in model.named_parameters() if n in decay_parameters],
"weight_decay": training_args.weight_decay,
},
{
"params": [p for n, p in model.named_parameters() if n not in decay_parameters],
"weight_decay": 0.0,
},
]
optimizer_kwargs = {
"betas": (training_args.adam_beta1, training_args.adam_beta2),
"eps": training_args.adam_epsilon,
}
optimizer_kwargs["lr"] = training_args.learning_rate
adam_bnb_optim = bnb.optim.Adam8bit(
optimizer_grouped_parameters,
betas=(training_args.adam_beta1, training_args.adam_beta2),
eps=training_args.adam_epsilon,
lr=training_args.learning_rate,
) |
Finally, pass the custom optimizer as an argument to the Trainer:
py
trainer = Trainer(model=model, args=training_args, train_dataset=ds, optimizers=(adam_bnb_optim, None))
Combined with other approaches (gradient accumulation, gradient checkpointing, and mixed precision training),
you can expect to get about a 3x memory improvement and even slightly higher throughput as using Adafactor.
multi_tensor
pytorch-nightly introduced torch.optim._multi_tensor which should significantly speed up the optimizers for situations
with lots of small feature tensors. It should eventually become the default, but if you want to experiment with it sooner, take a look at this GitHub issue.
Data preloading
One of the important requirements to reach great training speed is the ability to feed the GPU at the maximum speed it
can handle. By default, everything happens in the main process, and it might not be able to read the data from disk fast
enough, and thus create a bottleneck, leading to GPU under-utilization. Configure the following arguments to reduce the bottleneck: |
DataLoader(pin_memory=True, ) - ensures the data gets preloaded into the pinned memory on CPU and typically leads to much faster transfers from CPU to GPU memory.
DataLoader(num_workers=4, ) - spawn several workers to preload data faster. During training, watch the GPU utilization stats; if it's far from 100%, experiment with increasing the number of workers. Of course, the problem could be elsewhere, so many workers won't necessarily lead to better performance. |
When using [Trainer], the corresponding [TrainingArguments] are: dataloader_pin_memory (True by default), and dataloader_num_workers (defaults to 0).
DeepSpeed ZeRO
DeepSpeed is an open-source deep learning optimization library that is integrated with 🤗 Transformers and 🤗 Accelerate.
It provides a wide range of features and optimizations designed to improve the efficiency and scalability of large-scale
deep learning training.
If your model fits onto a single GPU and you have enough space to fit a small batch size, you don't need to use DeepSpeed
as it'll only slow things down. However, if the model doesn't fit onto a single GPU or you can't fit a small batch, you can
leverage DeepSpeed ZeRO + CPU Offload, or NVMe Offload for much larger models. In this case, you need to separately
install the library, then follow one of the guides to create a configuration file
and launch DeepSpeed: |
For an in-depth guide on DeepSpeed integration with [Trainer], review the corresponding documentation, specifically the
section for a single GPU. Some adjustments are required to use DeepSpeed in a notebook; please take a look at the corresponding guide.
If you prefer to use 🤗 Accelerate, refer to 🤗 Accelerate DeepSpeed guide. |
Using torch.compile
PyTorch 2.0 introduced a new compile function that doesn't require any modification to existing PyTorch code but can
optimize your code by adding a single line of code: model = torch.compile(model).
If using [Trainer], you only need to pass the torch_compile option in the [TrainingArguments]:
python
training_args = TrainingArguments(torch_compile=True, **default_args)
torch.compile uses Python's frame evaluation API to automatically create a graph from existing PyTorch programs. After
capturing the graph, different backends can be deployed to lower the graph to an optimized engine.
You can find more details and benchmarks in PyTorch documentation.
torch.compile has a growing list of backends, which can be found in by calling torchdynamo.list_backends(), each of which with its optional dependencies.
Choose which backend to use by specifying it via torch_compile_backend in the [TrainingArguments]. Some of the most commonly used backends are:
Debugging backends:
* dynamo.optimize("eager") - Uses PyTorch to run the extracted GraphModule. This is quite useful in debugging TorchDynamo issues.
* dynamo.optimize("aot_eager") - Uses AotAutograd with no compiler, i.e, just using PyTorch eager for the AotAutograd's extracted forward and backward graphs. This is useful for debugging, and unlikely to give speedups.
Training & inference backends:
* dynamo.optimize("inductor") - Uses TorchInductor backend with AotAutograd and cudagraphs by leveraging codegened Triton kernels Read more
* dynamo.optimize("nvfuser") - nvFuser with TorchScript. Read more
* dynamo.optimize("aot_nvfuser") - nvFuser with AotAutograd. Read more
* dynamo.optimize("aot_cudagraphs") - cudagraphs with AotAutograd. Read more
Inference-only backends:
* dynamo.optimize("ofi") - Uses Torchscript optimize_for_inference. Read more
* dynamo.optimize("fx2trt") - Uses NVIDIA TensorRT for inference optimizations. Read more
* dynamo.optimize("onnxrt") - Uses ONNXRT for inference on CPU/GPU. Read more
* dynamo.optimize("ipex") - Uses IPEX for inference on CPU. Read more
For an example of using torch.compile with 🤗 Transformers, check out this blog post on fine-tuning a BERT model for Text Classification using the newest PyTorch 2.0 features
Using 🤗 PEFT
Parameter-Efficient Fine Tuning (PEFT) methods freeze the pretrained model parameters during fine-tuning and add a small number of trainable parameters (the adapters) on top of it.
As a result the memory associated to the optimizer states and gradients are greatly reduced.
For example with a vanilla AdamW, the memory requirement for the optimizer state would be:
* fp32 copy of parameters: 4 bytes/param
* Momentum: 4 bytes/param
* Variance: 4 bytes/param
Suppose a model with 7B parameters and 200 millions parameters injected with Low Rank Adapters.
The memory requirement for the optimizer state of the plain model would be 12 * 7 = 84 GB (assuming 7B trainable parameters).
Adding Lora increases slightly the memory associated to the model weights and substantially decreases memory requirement for the optimizer state to 12 * 0.2 = 2.4GB.
Read more about PEFT and its detailed usage in the PEFT documentation or PEFT repository.
Using 🤗 Accelerate
With 🤗 Accelerate you can use the above methods while gaining full
control over the training loop and can essentially write the loop in pure PyTorch with some minor modifications.
Suppose you have combined the methods in the [TrainingArguments] like so:
py
training_args = TrainingArguments(
per_device_train_batch_size=1,
gradient_accumulation_steps=4,
gradient_checkpointing=True,
fp16=True,
**default_args,
)
The full example training loop with 🤗 Accelerate is only a handful of lines of code long: |
from accelerate import Accelerator
from torch.utils.data.dataloader import DataLoader
dataloader = DataLoader(ds, batch_size=training_args.per_device_train_batch_size)
if training_args.gradient_checkpointing:
model.gradient_checkpointing_enable()
accelerator = Accelerator(fp16=training_args.fp16)
model, optimizer, dataloader = accelerator.prepare(model, adam_bnb_optim, dataloader)
model.train()
for step, batch in enumerate(dataloader, start=1):
loss = model(**batch).loss
loss = loss / training_args.gradient_accumulation_steps
accelerator.backward(loss)
if step % training_args.gradient_accumulation_steps == 0:
optimizer.step()
optimizer.zero_grad() |
First we wrap the dataset in a DataLoader.
Then we can enable gradient checkpointing by calling the model's [~PreTrainedModel.gradient_checkpointing_enable] method.
When we initialize the Accelerator
we can specify if we want to use mixed precision training and it will take care of it for us in the [prepare] call.
During the prepare
call the dataloader will also be distributed across workers should we use multiple GPUs. We use the same 8-bit optimizer from the earlier example.
Finally, we can add the main training loop. Note that the backward call is handled by 🤗 Accelerate. We can also see
how gradient accumulation works: we normalize the loss, so we get the average at the end of accumulation and once we have
enough steps we run the optimization.
Implementing these optimization techniques with 🤗 Accelerate only takes a handful of lines of code and comes with the
benefit of more flexibility in the training loop. For a full documentation of all features have a look at the
Accelerate documentation.
Efficient Software Prebuilds
PyTorch's pip and conda builds come prebuilt with the cuda toolkit
which is enough to run PyTorch, but it is insufficient if you need to build cuda extensions.
At times, additional efforts may be required to pre-build some components. For instance, if you're using libraries like apex that
don't come pre-compiled. In other situations figuring out how to install the right cuda toolkit system-wide can be complicated.
To address these scenarios PyTorch and NVIDIA released a new version of NGC docker container which already comes with
everything prebuilt. You just need to install your programs on it, and it will run out of the box.
This approach is also useful if you want to tweak the pytorch source and/or make a new customized build.
To find the docker image version you want start with PyTorch release notes,
choose one of the latest monthly releases. Go into the release's notes for the desired release, check that the environment's
components are matching your needs (including NVIDIA Driver requirements!) and then at the very top of that document go
to the corresponding NGC page. If for some reason you get lost, here is the index of all PyTorch NGC images.
Next follow the instructions to download and deploy the docker image.
Mixture of Experts
Some recent papers reported a 4-5x training speedup and a faster inference by integrating
Mixture of Experts (MoE) into the Transformer models.
Since it has been discovered that more parameters lead to better performance, this technique allows to increase the
number of parameters by an order of magnitude without increasing training costs.
In this approach every other FFN layer is replaced with a MoE Layer which consists of many experts, with a gated function
that trains each expert in a balanced way depending on the input token's position in a sequence. |
(source: GLAM)
You can find exhaustive details and comparison tables in the papers listed at the end of this section.
The main drawback of this approach is that it requires staggering amounts of GPU memory - almost an order of magnitude
larger than its dense equivalent. Various distillation and approaches are proposed to how to overcome the much higher memory requirements.
There is direct trade-off though, you can use just a few experts with a 2-3x smaller base model instead of dozens or
hundreds experts leading to a 5x smaller model and thus increase the training speed moderately while increasing the
memory requirements moderately as well.
Most related papers and implementations are built around Tensorflow/TPUs: |
GShard: Scaling Giant Models with Conditional Computation and Automatic Sharding
Switch Transformers: Scaling to Trillion Parameter Models with Simple and Efficient Sparsity
GLaM: Generalist Language Model (GLaM) |
And for Pytorch DeepSpeed has built one as well: DeepSpeed-MoE: Advancing Mixture-of-Experts Inference and Training to Power Next-Generation AI Scale, Mixture of Experts - blog posts: 1, 2 and specific deployment with large transformer-based natural language generation models: blog post, Megatron-Deepspeed branch.
Using PyTorch native attention and Flash Attention
PyTorch's torch.nn.functional.scaled_dot_product_attention (SDPA) can also call FlashAttention and memory-efficient attention kernels under the hood. SDPA support is currently being added natively in Transformers and is used by default for torch>=2.1.1 when an implementation is available. Please refer to PyTorch scaled dot product attention for a list of supported models and more details.
Check out this blogpost to learn more about acceleration and memory-savings with SDPA. |
Installation
Install 🤗 Transformers for whichever deep learning library you're working with, setup your cache, and optionally configure 🤗 Transformers to run offline.
🤗 Transformers is tested on Python 3.6+, PyTorch 1.1.0+, TensorFlow 2.0+, and Flax. Follow the installation instructions below for the deep learning library you are using:
PyTorch installation instructions.
TensorFlow 2.0 installation instructions.
Flax installation instructions. |
PyTorch installation instructions.
TensorFlow 2.0 installation instructions.
Flax installation instructions.
Install with pip
You should install 🤗 Transformers in a virtual environment. If you're unfamiliar with Python virtual environments, take a look at this guide. A virtual environment makes it easier to manage different projects, and avoid compatibility issues between dependencies.
Start by creating a virtual environment in your project directory: |
python -m venv .env
Activate the virtual environment. On Linux and MacOs:
source .env/bin/activate
Activate Virtual environment on Windows
.env/Scripts/activate
Now you're ready to install 🤗 Transformers with the following command:
pip install transformers
For CPU-support only, you can conveniently install 🤗 Transformers and a deep learning library in one line. For example, install 🤗 Transformers and PyTorch with:
pip install 'transformers[torch]'
🤗 Transformers and TensorFlow 2.0: |
pip install 'transformers[torch]'
🤗 Transformers and TensorFlow 2.0:
pip install 'transformers[tf-cpu]'
M1 / ARM Users
You will need to install the following before installing TensorFLow 2.0
brew install cmake
brew install pkg-config
🤗 Transformers and Flax:
pip install 'transformers[flax]'
Finally, check if 🤗 Transformers has been properly installed by running the following command. It will download a pretrained model: |
🤗 Transformers and Flax:
pip install 'transformers[flax]'
Finally, check if 🤗 Transformers has been properly installed by running the following command. It will download a pretrained model:
python -c "from transformers import pipeline; print(pipeline('sentiment-analysis')('we love you'))"
Then print out the label and score:
[{'label': 'POSITIVE', 'score': 0.9998704791069031}]
Install from source
Install 🤗 Transformers from source with the following command: |
pip install git+https://github.com/huggingface/transformers
This command installs the bleeding edge main version rather than the latest stable version. The main version is useful for staying up-to-date with the latest developments. For instance, if a bug has been fixed since the last official release but a new release hasn't been rolled out yet. However, this means the main version may not always be stable. We strive to keep the main version operational, and most issues are usually resolved within a few hours or a day. If you run into a problem, please open an Issue so we can fix it even sooner!
Check if 🤗 Transformers has been properly installed by running the following command: |
python -c "from transformers import pipeline; print(pipeline('sentiment-analysis')('I love you'))"
Editable install
You will need an editable install if you'd like to:
Use the main version of the source code.
Contribute to 🤗 Transformers and need to test changes in the code.
Clone the repository and install 🤗 Transformers with the following commands: |
git clone https://github.com/huggingface/transformers.git
cd transformers
pip install -e .
These commands will link the folder you cloned the repository to and your Python library paths. Python will now look inside the folder you cloned to in addition to the normal library paths. For example, if your Python packages are typically installed in ~/anaconda3/envs/main/lib/python3.7/site-packages/, Python will also search the folder you cloned to: ~/transformers/. |
You must keep the transformers folder if you want to keep using the library.
Now you can easily update your clone to the latest version of 🤗 Transformers with the following command:
cd ~/transformers/
git pull
Your Python environment will find the main version of 🤗 Transformers on the next run.
Install with conda
Install from the conda channel conda-forge: |
conda install conda-forge::transformers
Cache setup
Pretrained models are downloaded and locally cached at: ~/.cache/huggingface/hub. This is the default directory given by the shell environment variable TRANSFORMERS_CACHE. On Windows, the default directory is given by C:\Users\username\.cache\huggingface\hub. You can change the shell environment variables shown below - in order of priority - to specify a different cache directory: |
Shell environment variable (default): HUGGINGFACE_HUB_CACHE or TRANSFORMERS_CACHE.
Shell environment variable: HF_HOME.
Shell environment variable: XDG_CACHE_HOME + /huggingface.
🤗 Transformers will use the shell environment variables PYTORCH_TRANSFORMERS_CACHE or PYTORCH_PRETRAINED_BERT_CACHE if you are coming from an earlier iteration of this library and have set those environment variables, unless you specify the shell environment variable TRANSFORMERS_CACHE. |
Offline mode
Run 🤗 Transformers in a firewalled or offline environment with locally cached files by setting the environment variable TRANSFORMERS_OFFLINE=1.
Add 🤗 Datasets to your offline training workflow with the environment variable HF_DATASETS_OFFLINE=1. |
HF_DATASETS_OFFLINE=1 TRANSFORMERS_OFFLINE=1 \
python examples/pytorch/translation/run_translation.py --model_name_or_path google-t5/t5-small --dataset_name wmt16 --dataset_config ro-en
This script should run without hanging or waiting to timeout because it won't attempt to download the model from the Hub.
You can also bypass loading a model from the Hub from each [~PreTrainedModel.from_pretrained] call with the [local_files_only] parameter. When set to True, only local files are loaded: |
from transformers import T5Model
model = T5Model.from_pretrained("./path/to/local/directory", local_files_only=True)
Fetch models and tokenizers to use offline
Another option for using 🤗 Transformers offline is to download the files ahead of time, and then point to their local path when you need to use them offline. There are three ways to do this:
Download a file through the user interface on the Model Hub by clicking on the ↓ icon. |
Download a file through the user interface on the Model Hub by clicking on the ↓ icon.
Use the [PreTrainedModel.from_pretrained] and [PreTrainedModel.save_pretrained] workflow:
Download your files ahead of time with [PreTrainedModel.from_pretrained]:
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
tokenizer = AutoTokenizer.from_pretrained("bigscience/T0_3B")
model = AutoModelForSeq2SeqLM.from_pretrained("bigscience/T0_3B") |
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
tokenizer = AutoTokenizer.from_pretrained("bigscience/T0_3B")
model = AutoModelForSeq2SeqLM.from_pretrained("bigscience/T0_3B")
Save your files to a specified directory with [PreTrainedModel.save_pretrained]:
tokenizer.save_pretrained("./your/path/bigscience_t0")
model.save_pretrained("./your/path/bigscience_t0")
Now when you're offline, reload your files with [PreTrainedModel.from_pretrained] from the specified directory: |
Now when you're offline, reload your files with [PreTrainedModel.from_pretrained] from the specified directory:
tokenizer = AutoTokenizer.from_pretrained("./your/path/bigscience_t0")
model = AutoModel.from_pretrained("./your/path/bigscience_t0")
Programmatically download files with the huggingface_hub library:
Install the huggingface_hub library in your virtual environment:
python -m pip install huggingface_hub |
Programmatically download files with the huggingface_hub library:
Install the huggingface_hub library in your virtual environment:
python -m pip install huggingface_hub
Use the hf_hub_download function to download a file to a specific path. For example, the following command downloads the config.json file from the T0 model to your desired path:
from huggingface_hub import hf_hub_download
hf_hub_download(repo_id="bigscience/T0_3B", filename="config.json", cache_dir="./your/path/bigscience_t0") |
from huggingface_hub import hf_hub_download
hf_hub_download(repo_id="bigscience/T0_3B", filename="config.json", cache_dir="./your/path/bigscience_t0")
Once your file is downloaded and locally cached, specify it's local path to load and use it:
from transformers import AutoConfig
config = AutoConfig.from_pretrained("./your/path/bigscience_t0/config.json")
See the How to download files from the Hub section for more details on downloading files stored on the Hub. |
Quantization
Quantization techniques focus on representing data with less information while also trying to not lose too much accuracy. This often means converting a data type to represent the same information with fewer bits. For example, if your model weights are stored as 32-bit floating points and they're quantized to 16-bit floating points, this halves the model size which makes it easier to store and reduces memory-usage. Lower precision can also speedup inference because it takes less time to perform calculations with fewer bits.
Transformers supports several quantization schemes to help you run inference with large language models (LLMs) and finetune adapters on quantized models. This guide will show you how to use Activation-aware Weight Quantization (AWQ), AutoGPTQ, and bitsandbytes. |
Interested in adding a new quantization method to Transformers? Read the HfQuantizer guide to learn how! |
AQLM
Try AQLM on Google Colab!
Additive Quantization of Language Models (AQLM) is a Large Language Models compression method. It quantizes multiple weights together and take advantage of interdependencies between them. AQLM represents groups of 8-16 weights as a sum of multiple vector codes.
Inference support for AQLM is realised in the aqlm library. Make sure to install it to run the models (note aqlm works only with python>=3.10): |
pip install aqlm[gpu,cpu]
The library provides efficient kernels for both GPU and CPU inference and training.
The instructions on how to quantize models yourself, as well as all the relevant code can be found in the corresponding GitHub repository.
PEFT
Starting with version aqlm 1.0.2, AQLM supports Parameter-Efficient Fine-Tuning in a form of LoRA integrated into the PEFT library.
AQLM configurations
AQLM quantization setpus vary mainly on the number of codebooks used as well as codebook sizes in bits. The most popular setups, as well as inference kernels they support are:
| Kernel | Number of codebooks | Codebook size, bits | Notation | Accuracy | Speedup | Fast GPU inference | Fast CPU inference |
|---|---------------------|---------------------|----------|-------------|-------------|--------------------|--------------------|
| Triton | K | N | KxN | - | Up to ~0.7x | ✅ | ❌ |
| CUDA | 1 | 16 | 1x16 | Best | Up to ~1.3x | ✅ | ❌ |
| CUDA | 2 | 8 | 2x8 | OK | Up to ~3.0x | ✅ | ❌ |
| Numba | K | 8 | Kx8 | Good | Up to ~4.0x | ❌ | ✅ |
AWQ |
Try AWQ quantization with this notebook! |
Activation-aware Weight Quantization (AWQ) doesn't quantize all the weights in a model, and instead, it preserves a small percentage of weights that are important for LLM performance. This significantly reduces quantization loss such that you can run models in 4-bit precision without experiencing any performance degradation.
There are several libraries for quantizing models with the AWQ algorithm, such as llm-awq, autoawq or optimum-intel. Transformers supports loading models quantized with the llm-awq and autoawq libraries. This guide will show you how to load models quantized with autoawq, but the process is similar for llm-awq quantized models.
Make sure you have autoawq installed: |
pip install autoawq
AWQ-quantized models can be identified by checking the quantization_config attribute in the model's config.json file:
json
{
"_name_or_path": "/workspace/process/huggingfaceh4_zephyr-7b-alpha/source",
"architectures": [
"MistralForCausalLM"
],
"quantization_config": {
"quant_method": "awq",
"zero_point": true,
"group_size": 128,
"bits": 4,
"version": "gemm"
}
}
A quantized model is loaded with the [~PreTrainedModel.from_pretrained] method. If you loaded your model on the CPU, make sure to move it to a GPU device first. Use the device_map parameter to specify where to place the model: |
from transformers import AutoModelForCausalLM, AutoTokenizer
model_id = "TheBloke/zephyr-7B-alpha-AWQ"
model = AutoModelForCausalLM.from_pretrained(model_id, device_map="cuda:0")
Loading an AWQ-quantized model automatically sets other weights to fp16 by default for performance reasons. If you want to load these other weights in a different format, use the torch_dtype parameter: |
from transformers import AutoModelForCausalLM, AutoTokenizer
model_id = "TheBloke/zephyr-7B-alpha-AWQ"
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype=torch.float32)
AWQ quantization can also be combined with FlashAttention-2 to further accelerate inference:
from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained("TheBloke/zephyr-7B-alpha-AWQ", attn_implementation="flash_attention_2", device_map="cuda:0") |
Fused modules
Fused modules offers improved accuracy and performance and it is supported out-of-the-box for AWQ modules for Llama and Mistral architectures, but you can also fuse AWQ modules for unsupported architectures.
Fused modules cannot be combined with other optimization techniques such as FlashAttention-2. |
To enable fused modules for supported architectures, create an [AwqConfig] and set the parameters fuse_max_seq_len and do_fuse=True. The fuse_max_seq_len parameter is the total sequence length and it should include the context length and the expected generation length. You can set it to a larger value to be safe.
For example, to fuse the AWQ modules of the TheBloke/Mistral-7B-OpenOrca-AWQ model.
thon
import torch
from transformers import AwqConfig, AutoModelForCausalLM
model_id = "TheBloke/Mistral-7B-OpenOrca-AWQ"
quantization_config = AwqConfig(
bits=4,
fuse_max_seq_len=512,
do_fuse=True,
)
model = AutoModelForCausalLM.from_pretrained(model_id, quantization_config=quantization_config).to(0) |
For architectures that don't support fused modules yet, you need to create a custom fusing mapping to define which modules need to be fused with the modules_to_fuse parameter. For example, to fuse the AWQ modules of the TheBloke/Yi-34B-AWQ model.
thon
import torch
from transformers import AwqConfig, AutoModelForCausalLM
model_id = "TheBloke/Yi-34B-AWQ"
quantization_config = AwqConfig(
bits=4,
fuse_max_seq_len=512,
modules_to_fuse={
"attention": ["q_proj", "k_proj", "v_proj", "o_proj"],
"layernorm": ["ln1", "ln2", "norm"],
"mlp": ["gate_proj", "up_proj", "down_proj"],
"use_alibi": False,
"num_attention_heads": 56,
"num_key_value_heads": 8,
"hidden_size": 7168
}
)
model = AutoModelForCausalLM.from_pretrained(model_id, quantization_config=quantization_config).to(0) |
The parameter modules_to_fuse should include: |
"attention": The names of the attention layers to fuse in the following order: query, key, value and output projection layer. If you don't want to fuse these layers, pass an empty list.
"layernorm": The names of all the LayerNorm layers you want to replace with a custom fused LayerNorm. If you don't want to fuse these layers, pass an empty list.
"mlp": The names of the MLP layers you want to fuse into a single MLP layer in the order: (gate (dense, layer, post-attention) / up / down layers).
"use_alibi": If your model uses ALiBi positional embedding.
"num_attention_heads": The number of attention heads.
"num_key_value_heads": The number of key value heads that should be used to implement Grouped Query Attention (GQA). If num_key_value_heads=num_attention_heads, the model will use Multi Head Attention (MHA), if num_key_value_heads=1 the model will use Multi Query Attention (MQA), otherwise GQA is used.
"hidden_size": The dimension of the hidden representations. |
Exllama-v2 support
Recent versions of autoawq supports exllama-v2 kernels for faster prefill and decoding. To get started, first install the latest version of autoawq by running: |
pip install git+https://github.com/casper-hansen/AutoAWQ.git
Get started by passing an AwqConfig() with version="exllama".
thon
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, AwqConfig
quantization_config = AwqConfig(version="exllama")
model = AutoModelForCausalLM.from_pretrained(
"TheBloke/Mistral-7B-Instruct-v0.1-AWQ",
quantization_config=quantization_config,
device_map="auto",
)
input_ids = torch.randint(0, 100, (1, 128), dtype=torch.long, device="cuda")
output = model(input_ids)
print(output.logits)
tokenizer = AutoTokenizer.from_pretrained("TheBloke/Mistral-7B-Instruct-v0.1-AWQ")
input_ids = tokenizer.encode("How to make a cake", return_tensors="pt").to(model.device)
output = model.generate(input_ids, do_sample=True, max_length=50, pad_token_id=50256)
print(tokenizer.decode(output[0], skip_special_tokens=True)) |
Note this feature is supported on AMD GPUs.
AutoGPTQ
Try GPTQ quantization with PEFT in this notebook and learn more about it's details in this blog post! |
The AutoGPTQ library implements the GPTQ algorithm, a post-training quantization technique where each row of the weight matrix is quantized independently to find a version of the weights that minimizes the error. These weights are quantized to int4, but they're restored to fp16 on the fly during inference. This can save your memory-usage by 4x because the int4 weights are dequantized in a fused kernel rather than a GPU's global memory, and you can also expect a speedup in inference because using a lower bitwidth takes less time to communicate.
Before you begin, make sure the following libraries are installed: |
pip install auto-gptq
pip install git+https://github.com/huggingface/optimum.git
pip install git+https://github.com/huggingface/transformers.git
pip install --upgrade accelerate
To quantize a model (currently only supported for text models), you need to create a [GPTQConfig] class and set the number of bits to quantize to, a dataset to calibrate the weights for quantization, and a tokenizer to prepare the dataset. |
from transformers import AutoModelForCausalLM, AutoTokenizer, GPTQConfig
model_id = "facebook/opt-125m"
tokenizer = AutoTokenizer.from_pretrained(model_id)
gptq_config = GPTQConfig(bits=4, dataset="c4", tokenizer=tokenizer) |
You could also pass your own dataset as a list of strings, but it is highly recommended to use the same dataset from the GPTQ paper.
py
dataset = ["auto-gptq is an easy-to-use model quantization library with user-friendly apis, based on GPTQ algorithm."]
gptq_config = GPTQConfig(bits=4, dataset=dataset, tokenizer=tokenizer)
Load a model to quantize and pass the gptq_config to the [~AutoModelForCausalLM.from_pretrained] method. Set device_map="auto" to automatically offload the model to a CPU to help fit the model in memory, and allow the model modules to be moved between the CPU and GPU for quantization.
py
quantized_model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto", quantization_config=gptq_config)
If you're running out of memory because a dataset is too large, disk offloading is not supported. If this is the case, try passing the max_memory parameter to allocate the amount of memory to use on your device (GPU and CPU):
py
quantized_model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto", max_memory={0: "30GiB", 1: "46GiB", "cpu": "30GiB"}, quantization_config=gptq_config) |
Depending on your hardware, it can take some time to quantize a model from scratch. It can take ~5 minutes to quantize the facebook/opt-350m model on a free-tier Google Colab GPU, but it'll take ~4 hours to quantize a 175B parameter model on a NVIDIA A100. Before you quantize a model, it is a good idea to check the Hub if a GPTQ-quantized version of the model already exists. |
Once your model is quantized, you can push the model and tokenizer to the Hub where it can be easily shared and accessed. Use the [~PreTrainedModel.push_to_hub] method to save the [GPTQConfig]:
py
quantized_model.push_to_hub("opt-125m-gptq")
tokenizer.push_to_hub("opt-125m-gptq")
You could also save your quantized model locally with the [~PreTrainedModel.save_pretrained] method. If the model was quantized with the device_map parameter, make sure to move the entire model to a GPU or CPU before saving it. For example, to save the model on a CPU: |
quantized_model.save_pretrained("opt-125m-gptq")
tokenizer.save_pretrained("opt-125m-gptq")
if quantized with device_map set
quantized_model.to("cpu")
quantized_model.save_pretrained("opt-125m-gptq")
Reload a quantized model with the [~PreTrainedModel.from_pretrained] method, and set device_map="auto" to automatically distribute the model on all available GPUs to load the model faster without using more memory than needed. |
from transformers import AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained("{your_username}/opt-125m-gptq", device_map="auto")
ExLlama
ExLlama is a Python/C++/CUDA implementation of the Llama model that is designed for faster inference with 4-bit GPTQ weights (check out these benchmarks). The ExLlama kernel is activated by default when you create a [GPTQConfig] object. To boost inference speed even further, use the ExLlamaV2 kernels by configuring the exllama_config parameter: |
import torch
from transformers import AutoModelForCausalLM, GPTQConfig
gptq_config = GPTQConfig(bits=4, exllama_config={"version":2})
model = AutoModelForCausalLM.from_pretrained("{your_username}/opt-125m-gptq", device_map="auto", quantization_config=gptq_config)
Only 4-bit models are supported, and we recommend deactivating the ExLlama kernels if you're finetuning a quantized model with PEFT. |
The ExLlama kernels are only supported when the entire model is on the GPU. If you're doing inference on a CPU with AutoGPTQ (version > 0.4.2), then you'll need to disable the ExLlama kernel. This overwrites the attributes related to the ExLlama kernels in the quantization config of the config.json file.
py
import torch
from transformers import AutoModelForCausalLM, GPTQConfig
gptq_config = GPTQConfig(bits=4, use_exllama=False)
model = AutoModelForCausalLM.from_pretrained("{your_username}/opt-125m-gptq", device_map="cpu", quantization_config=gptq_config)
bitsandbytes
bitsandbytes 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 to finetune quantized LLMs.
To use bitsandbytes, make sure you have the following libraries installed: |
pip install transformers accelerate bitsandbytes>0.37.0
pip install bitsandbytes>=0.39.0
pip install --upgrade accelerate
pip install --upgrade transformers
Now you can quantize a model with the load_in_8bit or load_in_4bit parameters in the [~PreTrainedModel.from_pretrained] method. This works for any model in any modality, as long as it supports loading with Accelerate and contains torch.nn.Linear layers. |
Quantizing a model in 8-bit halves the memory-usage, and for large models, set device_map="auto" to efficiently use the GPUs available:
from transformers import AutoModelForCausalLM
model_8bit = AutoModelForCausalLM.from_pretrained("bigscience/bloom-1b7", device_map="auto", load_in_8bit=True)
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 if you want: |
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 if you want:
import torch
from transformers import AutoModelForCausalLM
model_8bit = AutoModelForCausalLM.from_pretrained("facebook/opt-350m", load_in_8bit=True, torch_dtype=torch.float32)
model_8bit.model.decoder.layers[-1].final_layer_norm.weight.dtype |
Once a model is quantized to 8-bit, you can't push the quantized weights to the Hub unless you're using the latest version of Transformers and bitsandbytes. If you have the latest versions, then you can push the 8-bit model to the Hub with the [~PreTrainedModel.push_to_hub] method. The quantization config.json file is pushed first, followed by the quantized model weights. |
from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained("bigscience/bloom-560m", device_map="auto", load_in_8bit=True)
tokenizer = AutoTokenizer.from_pretrained("bigscience/bloom-560m")
model.push_to_hub("bloom-560m-8bit")
Quantizing a model in 4-bit reduces your memory-usage by 4x, and for large models, set device_map="auto" to efficiently use the GPUs available: |
Quantizing a model in 4-bit reduces your memory-usage by 4x, and for large models, set device_map="auto" to efficiently use the GPUs available:
from transformers import AutoModelForCausalLM
model_4bit = AutoModelForCausalLM.from_pretrained("bigscience/bloom-1b7", device_map="auto", load_in_4bit=True)
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 if you want: |
Subsets and Splits