text
stringlengths 2
11.8k
|
|---|
Let's check out 3 important ones:
Device
If you use device=n, the pipeline automatically puts the model on the specified device.
This will work regardless of whether you are using PyTorch or Tensorflow.
py
transcriber = pipeline(model="openai/whisper-large-v2", device=0)
If the model is too large for a single GPU and you are using PyTorch, you can set device_map="auto" to automatically
determine how to load and store the model weights. Using the device_map argument requires the 🤗 Accelerate
package: |
pip install --upgrade accelerate
The following code automatically loads and stores model weights across devices:
py
transcriber = pipeline(model="openai/whisper-large-v2", device_map="auto")
Note that if device_map="auto" is passed, there is no need to add the argument device=device when instantiating your pipeline as you may encounter some unexpected behavior!
Batch size
By default, pipelines will not batch inference for reasons explained in detail here. The reason is that batching is not necessarily faster, and can actually be quite slower in some cases.
But if it works in your use case, you can use:
py
transcriber = pipeline(model="openai/whisper-large-v2", device=0, batch_size=2)
audio_filenames = [f"https://huggingface.co/datasets/Narsil/asr_dummy/resolve/main/{i}.flac" for i in range(1, 5)]
texts = transcriber(audio_filenames)
This runs the pipeline on the 4 provided audio files, but it will pass them in batches of 2
to the model (which is on a GPU, where batching is more likely to help) without requiring any further code from you.
The output should always match what you would have received without batching. It is only meant as a way to help you get more speed out of a pipeline.
Pipelines can also alleviate some of the complexities of batching because, for some pipelines, a single item (like a long audio file) needs to be chunked into multiple parts to be processed by a model. The pipeline performs this chunk batching for you.
Task specific parameters
All tasks provide task specific parameters which allow for additional flexibility and options to help you get your job done.
For instance, the [transformers.AutomaticSpeechRecognitionPipeline.__call__] method has a return_timestamps parameter which sounds promising for subtitling videos: |
transcriber = pipeline(model="openai/whisper-large-v2", return_timestamps=True)
transcriber("https://huggingface.co/datasets/Narsil/asr_dummy/resolve/main/mlk.flac")
{'text': ' I have a dream that one day this nation will rise up and live out the true meaning of its creed.', 'chunks': [{'timestamp': (0.0, 11.88), 'text': ' I have a dream that one day this nation will rise up and live out the true meaning of its'}, {'timestamp': (11.88, 12.38), 'text': ' creed.'}]} |
As you can see, the model inferred the text and also outputted when the various sentences were pronounced.
There are many parameters available for each task, so check out each task's API reference to see what you can tinker with!
For instance, the [~transformers.AutomaticSpeechRecognitionPipeline] has a chunk_length_s parameter which is helpful
for working on really long audio files (for example, subtitling entire movies or hour-long videos) that a model typically
cannot handle on its own:
thon |
transcriber = pipeline(model="openai/whisper-large-v2", chunk_length_s=30, return_timestamps=True)
transcriber("https://huggingface.co/datasets/sanchit-gandhi/librispeech_long/resolve/main/audio.wav")
{'text': " Chapter 16. I might have told you of the beginning of this liaison in a few lines, but I wanted you to see every step by which we came. I, too, agree to whatever Marguerite wished, Marguerite to be unable to live apart from me. It was the day after the evening |
If you can't find a parameter that would really help you out, feel free to request it!
Using pipelines on a dataset
The pipeline can also run inference on a large dataset. The easiest way we recommend doing this is by using an iterator:
def data():
for i in range(1000):
yield f"My example {i}"
pipe = pipeline(model="openai-community/gpt2", device=0)
generated_characters = 0
for out in pipe(data()):
generated_characters += len(out[0]["generated_text"]) |
The iterator data() yields each result, and the pipeline automatically
recognizes the input is iterable and will start fetching the data while
it continues to process it on the GPU (this uses DataLoader under the hood).
This is important because you don't have to allocate memory for the whole dataset
and you can feed the GPU as fast as possible.
Since batching could speed things up, it may be useful to try tuning the batch_size parameter here.
The simplest way to iterate over a dataset is to just load one from 🤗 Datasets: |
KeyDataset is a util that will just output the item we're interested in.
from transformers.pipelines.pt_utils import KeyDataset
from datasets import load_dataset
pipe = pipeline(model="hf-internal-testing/tiny-random-wav2vec2", device=0)
dataset = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation[:10]")
for out in pipe(KeyDataset(dataset, "audio")):
print(out)
Using pipelines for a webserver |
Using pipelines for a webserver
Creating an inference engine is a complex topic which deserves it's own
page.
Link
Vision pipeline
Using a [pipeline] for vision tasks is practically identical.
Specify your task and pass your image to the classifier. The image can be a link, a local path or a base64-encoded image. For example, what species of cat is shown below? |
from transformers import pipeline
vision_classifier = pipeline(model="google/vit-base-patch16-224")
preds = vision_classifier(
images="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/pipeline-cat-chonk.jpeg"
)
preds = [{"score": round(pred["score"], 4), "label": pred["label"]} for pred in preds]
preds
[{'score': 0.4335, 'label': 'lynx, catamount'}, {'score': 0.0348, 'label': 'cougar, puma, catamount, mountain lion, painter, panther, Felis concolor'}, {'score': 0.0324, 'label': 'snow leopard, ounce, Panthera uncia'}, {'score': 0.0239, 'label': 'Egyptian cat'}, {'score': 0.0229, 'label': 'tiger cat'}] |
Text pipeline
Using a [pipeline] for NLP tasks is practically identical. |
from transformers import pipeline
This model is a zero-shot-classification model.
It will classify text, except you are free to choose any label you might imagine
classifier = pipeline(model="facebook/bart-large-mnli")
classifier(
"I have a problem with my iphone that needs to be resolved asap!!",
candidate_labels=["urgent", "not urgent", "phone", "tablet", "computer"],
)
{'sequence': 'I have a problem with my iphone that needs to be resolved asap!!', 'labels': ['urgent', 'phone', 'computer', 'not urgent', 'tablet'], 'scores': [0.504, 0.479, 0.013, 0.003, 0.002]} |
Multimodal pipeline
The [pipeline] supports more than one modality. For example, a visual question answering (VQA) task combines text and image. Feel free to use any image link you like and a question you want to ask about the image. The image can be a URL or a local path to the image.
For example, if you use this invoice image: |
from transformers import pipeline
vqa = pipeline(model="impira/layoutlm-document-qa")
vqa(
image="https://huggingface.co/spaces/impira/docquery/resolve/2359223c1837a7587402bda0f2643382a6eefeab/invoice.png",
question="What is the invoice number?",
)
[{'score': 0.42515, 'answer': 'us-001', 'start': 16, 'end': 16}]
To run the example above you need to have pytesseract installed in addition to 🤗 Transformers:
sudo apt install -y tesseract-ocr
pip install pytesseract |
To run the example above you need to have pytesseract installed in addition to 🤗 Transformers:
sudo apt install -y tesseract-ocr
pip install pytesseract
Using pipeline on large models with 🤗 accelerate:
You can easily run pipeline on large models using 🤗 accelerate! First make sure you have installed accelerate with pip install accelerate.
First load your model using device_map="auto"! We will use facebook/opt-1.3b for our example. |
pip install accelerate
import torch
from transformers import pipeline
pipe = pipeline(model="facebook/opt-1.3b", torch_dtype=torch.bfloat16, device_map="auto")
output = pipe("This is a cool example!", do_sample=True, top_p=0.95)
You can also pass 8-bit loaded models if you install bitsandbytes and add the argument load_in_8bit=True |
You can also pass 8-bit loaded models if you install bitsandbytes and add the argument load_in_8bit=True
pip install accelerate bitsandbytes
import torch
from transformers import pipeline
pipe = pipeline(model="facebook/opt-1.3b", device_map="auto", model_kwargs={"load_in_8bit": True})
output = pipe("This is a cool example!", do_sample=True, top_p=0.95)
Note that you can replace the checkpoint with any of the Hugging Face model that supports large model loading such as BLOOM! |
Fully Sharded Data Parallel
Fully Sharded Data Parallel (FSDP) is a data parallel method that shards a model's parameters, gradients and optimizer states across the number of available GPUs (also called workers or rank). Unlike DistributedDataParallel (DDP), FSDP reduces memory-usage because a model is replicated on each GPU. This improves GPU memory-efficiency and allows you to train much larger models on fewer GPUs. FSDP is integrated with the Accelerate, a library for easily managing training in distributed environments, which means it is available for use from the [Trainer] class.
Before you start, make sure Accelerate is installed and at least PyTorch 2.1.0 or newer. |
pip install accelerate
FSDP configuration
To start, run the accelerate config command to create a configuration file for your training environment. Accelerate uses this configuration file to automatically setup the correct training environment based on your selected training options in accelerate config. |
accelerate config
When you run accelerate config, you'll be prompted with a series of options to configure your training environment. This section covers some of the most important FSDP options. To learn more about the other available FSDP options, take a look at the fsdp_config parameters.
Sharding strategy
FSDP offers a number of sharding strategies to select from: |
FULL_SHARD - shards model parameters, gradients and optimizer states across workers; select 1 for this option
SHARD_GRAD_OP- shard gradients and optimizer states across workers; select 2 for this option
NO_SHARD - don't shard anything (this is equivalent to DDP); select 3 for this option
HYBRID_SHARD - shard model parameters, gradients and optimizer states within each worker where each worker also has a full copy; select 4 for this option
HYBRID_SHARD_ZERO2 - shard gradients and optimizer states within each worker where each worker also has a full copy; select 5 for this option |
This is enabled by the fsdp_sharding_strategy flag.
CPU offload
You could also offload parameters and gradients when they are not in use to the CPU to save even more GPU memory and help you fit large models where even FSDP may not be sufficient. This is enabled by setting fsdp_offload_params: true when running accelerate config.
Wrapping policy
FSDP is applied by wrapping each layer in the network. The wrapping is usually applied in a nested way where the full weights are discarded after each forward pass to save memory for use in the next layer. The auto wrapping policy is the simplest way to implement this and you don't need to change any code. You should select fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP to wrap a Transformer layer and fsdp_transformer_layer_cls_to_wrap to specify which layer to wrap (for example BertLayer).
Otherwise, you can choose a size-based wrapping policy where FSDP is applied to a layer if it exceeds a certain number of parameters. This is enabled by setting fsdp_wrap_policy: SIZE_BASED_WRAP and min_num_param to the desired size threshold.
Checkpointing
Intermediate checkpoints should be saved with fsdp_state_dict_type: SHARDED_STATE_DICT because saving the full state dict with CPU offloading on rank 0 takes a lot of time and often results in NCCL Timeout errors due to indefinite hanging during broadcasting. You can resume training with the sharded state dicts with the [~accelerate.Accelerator.load_state]` method. |
directory containing checkpoints
accelerator.load_state("ckpt")
However, when training ends, you want to save the full state dict because sharded state dict is only compatible with FSDP.
if trainer.is_fsdp_enabled:
trainer.accelerator.state.fsdp_plugin.set_state_dict_type("FULL_STATE_DICT")
trainer.save_model(script_args.output_dir) |
TPU
PyTorch XLA supports FSDP training for TPUs and it can be enabled by modifying the FSDP configuration file generated by accelerate config. In addition to the sharding strategies and wrapping options specified above, you can add the parameters shown below to the file.
yaml
xla: True # must be set to True to enable PyTorch/XLA
xla_fsdp_settings: # XLA-specific FSDP parameters
xla_fsdp_grad_ckpt: True # use gradient checkpointing
The xla_fsdp_settings allow you to configure additional XLA-specific parameters for FSDP.
Launch training
An example FSDP configuration file may look like:
yaml
compute_environment: LOCAL_MACHINE
debug: false
distributed_type: FSDP
downcast_bf16: 'no'
fsdp_config:
fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP
fsdp_backward_prefetch_policy: BACKWARD_PRE
fsdp_cpu_ram_efficient_loading: true
fsdp_forward_prefetch: false
fsdp_offload_params: true
fsdp_sharding_strategy: 1
fsdp_state_dict_type: SHARDED_STATE_DICT
fsdp_sync_module_states: true
fsdp_transformer_layer_cls_to_wrap: BertLayer
fsdp_use_orig_params: true
machine_rank: 0
main_training_function: main
mixed_precision: bf16
num_machines: 1
num_processes: 2
rdzv_backend: static
same_network: true
tpu_env: []
tpu_use_cluster: false
tpu_use_sudo: false
use_cpu: false
To launch training, run the accelerate launch command and it'll automatically use the configuration file you previously created with accelerate config. |
accelerate launch my-trainer-script.py
accelerate launch --fsdp="full shard" --fsdp_config="path/to/fsdp_config/ my-trainer-script.py
Next steps
FSDP can be a powerful tool for training really large models and you have access to more than one GPU or TPU. By sharding the model parameters, optimizer and gradient states, and even offloading them to the CPU when they're inactive, FSDP can reduce the high cost of large-scale training. If you're interested in learning more, the following may be helpful: |
Follow along with the more in-depth Accelerate guide for FSDP.
Read the Introducing PyTorch Fully Sharded Data Parallel (FSDP) API blog post.
Read the Scaling PyTorch models on Cloud TPUs with FSDP blog post. |
The Transformer model family
Since its introduction in 2017, the original Transformer model has inspired many new and exciting models that extend beyond natural language processing (NLP) tasks. There are models for predicting the folded structure of proteins, training a cheetah to run, and time series forecasting. With so many Transformer variants available, it can be easy to miss the bigger picture. What all these models have in common is they're based on the original Transformer architecture. Some models only use the encoder or decoder, while others use both. This provides a useful taxonomy to categorize and examine the high-level differences within models in the Transformer family, and it'll help you understand Transformers you haven't encountered before.
If you aren't familiar with the original Transformer model or need a refresher, check out the How do Transformers work chapter from the Hugging Face course. |
Computer vision |
Convolutional network
For a long time, convolutional networks (CNNs) were the dominant paradigm for computer vision tasks until the Vision Transformer demonstrated its scalability and efficiency. Even then, some of a CNN's best qualities, like translation invariance, are so powerful (especially for certain tasks) that some Transformers incorporate convolutions in their architecture. ConvNeXt flipped this exchange around and incorporated design choices from Transformers to modernize a CNN. For example, ConvNeXt uses non-overlapping sliding windows to patchify an image and a larger kernel to increase its global receptive field. ConvNeXt also makes several layer design choices to be more memory-efficient and improve performance, so it competes favorably with Transformers!
Encoder[[cv-encoder]]
The Vision Transformer (ViT) opened the door to computer vision tasks without convolutions. ViT uses a standard Transformer encoder, but its main breakthrough was how it treated an image. It splits an image into fixed-size patches and uses them to create an embedding, just like how a sentence is split into tokens. ViT capitalized on the Transformers' efficient architecture to demonstrate competitive results with the CNNs at the time while requiring fewer resources to train. ViT was soon followed by other vision models that could also handle dense vision tasks like segmentation as well as detection.
One of these models is the Swin Transformer. It builds hierarchical feature maps (like a CNN 👀 and unlike ViT) from smaller-sized patches and merges them with neighboring patches in deeper layers. Attention is only computed within a local window, and the window is shifted between attention layers to create connections to help the model learn better. Since the Swin Transformer can produce hierarchical feature maps, it is a good candidate for dense prediction tasks like segmentation and detection. The SegFormer also uses a Transformer encoder to build hierarchical feature maps, but it adds a simple multilayer perceptron (MLP) decoder on top to combine all the feature maps and make a prediction.
Other vision models, like BeIT and ViTMAE, drew inspiration from BERT's pretraining objective. BeIT is pretrained by masked image modeling (MIM); the image patches are randomly masked, and the image is also tokenized into visual tokens. BeIT is trained to predict the visual tokens corresponding to the masked patches. ViTMAE has a similar pretraining objective, except it must predict the pixels instead of visual tokens. What's unusual is 75% of the image patches are masked! The decoder reconstructs the pixels from the masked tokens and encoded patches. After pretraining, the decoder is thrown away, and the encoder is ready to be used in downstream tasks.
Decoder[[cv-decoder]]
Decoder-only vision models are rare because most vision models rely on an encoder to learn an image representation. But for use cases like image generation, the decoder is a natural fit, as we've seen from text generation models like GPT-2. ImageGPT uses the same architecture as GPT-2, but instead of predicting the next token in a sequence, it predicts the next pixel in an image. In addition to image generation, ImageGPT could also be finetuned for image classification.
Encoder-decoder[[cv-encoder-decoder]]
Vision models commonly use an encoder (also known as a backbone) to extract important image features before passing them to a Transformer decoder. DETR has a pretrained backbone, but it also uses the complete Transformer encoder-decoder architecture for object detection. The encoder learns image representations and combines them with object queries (each object query is a learned embedding that focuses on a region or object in an image) in the decoder. DETR predicts the bounding box coordinates and class label for each object query.
Natural language processing |
Encoder[[nlp-encoder]]
BERT is an encoder-only Transformer that randomly masks certain tokens in the input to avoid seeing other tokens, which would allow it to "cheat". The pretraining objective is to predict the masked token based on the context. This allows BERT to fully use the left and right contexts to help it learn a deeper and richer representation of the inputs. However, there was still room for improvement in BERT's pretraining strategy. RoBERTa improved upon this by introducing a new pretraining recipe that includes training for longer and on larger batches, randomly masking tokens at each epoch instead of just once during preprocessing, and removing the next-sentence prediction objective.
The dominant strategy to improve performance is to increase the model size. But training large models is computationally expensive. One way to reduce computational costs is using a smaller model like DistilBERT. DistilBERT uses knowledge distillation - a compression technique - to create a smaller version of BERT while keeping nearly all of its language understanding capabilities.
However, most Transformer models continued to trend towards more parameters, leading to new models focused on improving training efficiency. ALBERT reduces memory consumption by lowering the number of parameters in two ways: separating the larger vocabulary embedding into two smaller matrices and allowing layers to share parameters. DeBERTa added a disentangled attention mechanism where the word and its position are separately encoded in two vectors. The attention is computed from these separate vectors instead of a single vector containing the word and position embeddings. Longformer also focused on making attention more efficient, especially for processing documents with longer sequence lengths. It uses a combination of local windowed attention (attention only calculated from fixed window size around each token) and global attention (only for specific task tokens like [CLS] for classification) to create a sparse attention matrix instead of a full attention matrix.
Decoder[[nlp-decoder]]
GPT-2 is a decoder-only Transformer that predicts the next word in the sequence. It masks tokens to the right so the model can't "cheat" by looking ahead. By pretraining on a massive body of text, GPT-2 became really good at generating text, even if the text is only sometimes accurate or true. But GPT-2 lacked the bidirectional context from BERT's pretraining, which made it unsuitable for certain tasks. XLNET combines the best of both BERT and GPT-2's pretraining objectives by using a permutation language modeling objective (PLM) that allows it to learn bidirectionally.
After GPT-2, language models grew even bigger and are now known as large language models (LLMs). LLMs demonstrate few- or even zero-shot learning if pretrained on a large enough dataset. GPT-J is an LLM with 6B parameters and trained on 400B tokens. GPT-J was followed by OPT, a family of decoder-only models, the largest of which is 175B and trained on 180B tokens. BLOOM was released around the same time, and the largest model in the family has 176B parameters and is trained on 366B tokens in 46 languages and 13 programming languages.
Encoder-decoder[[nlp-encoder-decoder]]
BART keeps the original Transformer architecture, but it modifies the pretraining objective with text infilling corruption, where some text spans are replaced with a single mask token. The decoder predicts the uncorrupted tokens (future tokens are masked) and uses the encoder's hidden states to help it. Pegasus is similar to BART, but Pegasus masks entire sentences instead of text spans. In addition to masked language modeling, Pegasus is pretrained by gap sentence generation (GSG). The GSG objective masks whole sentences important to a document, replacing them with a mask token. The decoder must generate the output from the remaining sentences. T5 is a more unique model that casts all NLP tasks into a text-to-text problem using specific prefixes. For example, the prefix Summarize: indicates a summarization task. T5 is pretrained by supervised (GLUE and SuperGLUE) training and self-supervised training (randomly sample and drop out 15% of tokens).
Audio |
Encoder[[audio-encoder]]
Wav2Vec2 uses a Transformer encoder to learn speech representations directly from raw audio waveforms. It is pretrained with a contrastive task to determine the true speech representation from a set of false ones. HuBERT is similar to Wav2Vec2 but has a different training process. Target labels are created by a clustering step in which segments of similar audio are assigned to a cluster which becomes a hidden unit. The hidden unit is mapped to an embedding to make a prediction.
Encoder-decoder[[audio-encoder-decoder]]
Speech2Text is a speech model designed for automatic speech recognition (ASR) and speech translation. The model accepts log mel-filter bank features extracted from the audio waveform and pretrained autoregressively to generate a transcript or translation. Whisper is also an ASR model, but unlike many other speech models, it is pretrained on a massive amount of ✨ labeled ✨ audio transcription data for zero-shot performance. A large chunk of the dataset also contains non-English languages, meaning Whisper can also be used for low-resource languages. Structurally, Whisper is similar to Speech2Text. The audio signal is converted to a log-mel spectrogram encoded by the encoder. The decoder generates the transcript autoregressively from the encoder's hidden states and the previous tokens.
Multimodal |
Encoder[[mm-encoder]]
VisualBERT is a multimodal model for vision-language tasks released shortly after BERT. It combines BERT and a pretrained object detection system to extract image features into visual embeddings, passed alongside text embeddings to BERT. VisualBERT predicts the masked text based on the unmasked text and the visual embeddings, and it also has to predict whether the text is aligned with the image. When ViT was released, ViLT adopted ViT in its architecture because it was easier to get the image embeddings this way. The image embeddings are jointly processed with the text embeddings. From there, ViLT is pretrained by image text matching, masked language modeling, and whole word masking.
CLIP takes a different approach and makes a pair prediction of (image, text) . An image encoder (ViT) and a text encoder (Transformer) are jointly trained on a 400 million (image, text) pair dataset to maximize the similarity between the image and text embeddings of the (image, text) pairs. After pretraining, you can use natural language to instruct CLIP to predict the text given an image or vice versa. OWL-ViT builds on top of CLIP by using it as its backbone for zero-shot object detection. After pretraining, an object detection head is added to make a set prediction over the (class, bounding box) pairs.
Encoder-decoder[[mm-encoder-decoder]]
Optical character recognition (OCR) is a long-standing text recognition task that typically involves several components to understand the image and generate the text. TrOCR simplifies the process using an end-to-end Transformer. The encoder is a ViT-style model for image understanding and processes the image as fixed-size patches. The decoder accepts the encoder's hidden states and autoregressively generates text. Donut is a more general visual document understanding model that doesn't rely on OCR-based approaches. It uses a Swin Transformer as the encoder and multilingual BART as the decoder. Donut is pretrained to read text by predicting the next word based on the image and text annotations. The decoder generates a token sequence given a prompt. The prompt is represented by a special token for each downstream task. For example, document parsing has a special parsing token that is combined with the encoder hidden states to parse the document into a structured output format (JSON).
Reinforcement learning |
Decoder[[rl-decoder]]
The Decision and Trajectory Transformer casts the state, action, and reward as a sequence modeling problem. The Decision Transformer generates a series of actions that lead to a future desired return based on returns-to-go, past states, and actions. For the last K timesteps, each of the three modalities are converted into token embeddings and processed by a GPT-like model to predict a future action token. Trajectory Transformer also tokenizes the states, actions, and rewards and processes them with a GPT architecture. Unlike the Decision Transformer, which is focused on reward conditioning, the Trajectory Transformer generates future actions with beam search. |
Custom hardware for training
The hardware you use to run model training and inference can have a big effect on performance. For a deep dive into GPUs make sure to check out Tim Dettmer's excellent blog post.
Let's have a look at some practical advice for GPU setups.
GPU
When you train bigger models you have essentially three options:
bigger GPUs
more GPUs
more CPU and NVMe (offloaded to by DeepSpeed-Infinity) |
Let's start at the case where you have a single GPU.
Power and Cooling
If you bought an expensive high end GPU make sure you give it the correct power and sufficient cooling.
Power:
Some high end consumer GPU cards have 2 and sometimes 3 PCI-E 8-Pin power sockets. Make sure you have as many independent 12V PCI-E 8-Pin cables plugged into the card as there are sockets. Do not use the 2 splits at one end of the same cable (also known as pigtail cable). That is if you have 2 sockets on the GPU, you want 2 PCI-E 8-Pin cables going from your PSU to the card and not one that has 2 PCI-E 8-Pin connectors at the end! You won't get the full performance out of your card otherwise.
Each PCI-E 8-Pin power cable needs to be plugged into a 12V rail on the PSU side and can supply up to 150W of power.
Some other cards may use a PCI-E 12-Pin connectors, and these can deliver up to 500-600W of power.
Low end cards may use 6-Pin connectors, which supply up to 75W of power.
Additionally you want the high-end PSU that has stable voltage. Some lower quality ones may not give the card the stable voltage it needs to function at its peak.
And of course the PSU needs to have enough unused Watts to power the card.
Cooling:
When a GPU gets overheated it will start throttling down and will not deliver full performance and it can even shutdown if it gets too hot.
It's hard to tell the exact best temperature to strive for when a GPU is heavily loaded, but probably anything under +80C is good, but lower is better - perhaps 70-75C is an excellent range to be in. The throttling down is likely to start at around 84-90C. But other than throttling performance a prolonged very high temperature is likely to reduce the lifespan of a GPU.
Next let's have a look at one of the most important aspects when having multiple GPUs: connectivity.
Multi-GPU Connectivity
If you use multiple GPUs the way cards are inter-connected can have a huge impact on the total training time. If the GPUs are on the same physical node, you can run: |
nvidia-smi topo -m
and it will tell you how the GPUs are inter-connected. On a machine with dual-GPU and which are connected with NVLink, you will most likely see something like:
GPU0 GPU1 CPU Affinity NUMA Affinity
GPU0 X NV2 0-23 N/A
GPU1 NV2 X 0-23 N/A
on a different machine w/o NVLink we may see:
GPU0 GPU1 CPU Affinity NUMA Affinity
GPU0 X PHB 0-11 N/A
GPU1 PHB X 0-11 N/A
The report includes this legend:
X = Self
SYS = Connection traversing PCIe as well as the SMP interconnect between NUMA nodes (e.g., QPI/UPI)
NODE = Connection traversing PCIe as well as the interconnect between PCIe Host Bridges within a NUMA node
PHB = Connection traversing PCIe as well as a PCIe Host Bridge (typically the CPU)
PXB = Connection traversing multiple PCIe bridges (without traversing the PCIe Host Bridge)
PIX = Connection traversing at most a single PCIe bridge
NV# = Connection traversing a bonded set of # NVLinks
So the first report NV2 tells us the GPUs are interconnected with 2 NVLinks, and the second report PHB we have a typical consumer-level PCIe+Bridge setup.
Check what type of connectivity you have on your setup. Some of these will make the communication between cards faster (e.g. NVLink), others slower (e.g. PHB).
Depending on the type of scalability solution used, the connectivity speed could have a major or a minor impact. If the GPUs need to sync rarely, as in DDP, the impact of a slower connection will be less significant. If the GPUs need to send messages to each other often, as in ZeRO-DP, then faster connectivity becomes super important to achieve faster training.
NVlink
NVLink is a wire-based serial multi-lane near-range communications link developed by Nvidia.
Each new generation provides a faster bandwidth, e.g. here is a quote from Nvidia Ampere GA102 GPU Architecture: |
Third-Generation NVLink®
GA102 GPUs utilize NVIDIA’s third-generation NVLink interface, which includes four x4 links,
with each link providing 14.0625 GB/sec bandwidth in each direction between two GPUs. Four
links provide 56.25 GB/sec bandwidth in each direction, and 112.5 GB/sec total bandwidth
between two GPUs. Two RTX 3090 GPUs can be connected together for SLI using NVLink.
(Note that 3-Way and 4-Way SLI configurations are not supported.) |
So the higher X you get in the report of NVX in the output of nvidia-smi topo -m the better. The generation will depend on your GPU architecture.
Let's compare the execution of a openai-community/gpt2 language model training over a small sample of wikitext.
The results are:
| NVlink | Time |
| ----- | ---: |
| Y | 101s |
| N | 131s |
You can see that NVLink completes the training ~23% faster. In the second benchmark we use NCCL_P2P_DISABLE=1 to tell the GPUs not to use NVLink.
Here is the full benchmark code and outputs:
```bash
DDP w/ NVLink
rm -r /tmp/test-clm; CUDA_VISIBLE_DEVICES=0,1 torchrun \
--nproc_per_node 2 examples/pytorch/language-modeling/run_clm.py --model_name_or_path openai-community/gpt2 \
--dataset_name wikitext --dataset_config_name wikitext-2-raw-v1 --do_train \
--output_dir /tmp/test-clm --per_device_train_batch_size 4 --max_steps 200
{'train_runtime': 101.9003, 'train_samples_per_second': 1.963, 'epoch': 0.69}
DDP w/o NVLink
rm -r /tmp/test-clm; CUDA_VISIBLE_DEVICES=0,1 NCCL_P2P_DISABLE=1 torchrun \
--nproc_per_node 2 examples/pytorch/language-modeling/run_clm.py --model_name_or_path openai-community/gpt2 \
--dataset_name wikitext --dataset_config_name wikitext-2-raw-v1 --do_train
--output_dir /tmp/test-clm --per_device_train_batch_size 4 --max_steps 200
{'train_runtime': 131.4367, 'train_samples_per_second': 1.522, 'epoch': 0.69} |
Hardware: 2x TITAN RTX 24GB each + NVlink with 2 NVLinks (NV2 in nvidia-smi topo -m)
Software: pytorch-1.8-to-be + cuda-11.0 / transformers==4.3.0.dev0 |
Attention mechanisms
Most transformer models use full attention in the sense that the attention matrix is square. It can be a big
computational bottleneck when you have long texts. Longformer and reformer are models that try to be more efficient and
use a sparse version of the attention matrix to speed up training.
LSH attention
Reformer uses LSH attention. In the softmax(QK^t), only the biggest elements (in the softmax
dimension) of the matrix QK^t are going to give useful contributions. So for each query q in Q, we can consider only
the keys k in K that are close to q. A hash function is used to determine if q and k are close. The attention mask is
modified to mask the current token (except at the first position), because it will give a query and a key equal (so
very similar to each other). Since the hash can be a bit random, several hash functions are used in practice
(determined by a n_rounds parameter) and then are averaged together.
Local attention
Longformer uses local attention: often, the local context (e.g., what are the two tokens to the
left and right?) is enough to take action for a given token. Also, by stacking attention layers that have a small
window, the last layer will have a receptive field of more than just the tokens in the window, allowing them to build a
representation of the whole sentence.
Some preselected input tokens are also given global attention: for those few tokens, the attention matrix can access
all tokens and this process is symmetric: all other tokens have access to those specific tokens (on top of the ones in
their local window). This is shown in Figure 2d of the paper, see below for a sample attention mask: |
Using those attention matrices with less parameters then allows the model to have inputs having a bigger sequence
length.
Other tricks
Axial positional encodings
Reformer uses axial positional encodings: in traditional transformer models, the positional encoding
E is a matrix of size \(l\) by \(d\), \(l\) being the sequence length and \(d\) the dimension of the
hidden state. If you have very long texts, this matrix can be huge and take way too much space on the GPU. To alleviate
that, axial positional encodings consist of factorizing that big matrix E in two smaller matrices E1 and E2, with
dimensions \(l_{1} \times d_{1}\) and \(l_{2} \times d_{2}\), such that \(l_{1} \times l_{2} = l\) and
\(d_{1} + d_{2} = d\) (with the product for the lengths, this ends up being way smaller). The embedding for time
step \(j\) in E is obtained by concatenating the embeddings for timestep \(j \% l1\) in E1 and \(j // l1\)
in E2. |
Use tokenizers from 🤗 Tokenizers
The [PreTrainedTokenizerFast] depends on the 🤗 Tokenizers library. The tokenizers obtained from the 🤗 Tokenizers library can be
loaded very simply into 🤗 Transformers.
Before getting in the specifics, let's first start by creating a dummy tokenizer in a few lines:
thon |
from tokenizers import Tokenizer
from tokenizers.models import BPE
from tokenizers.trainers import BpeTrainer
from tokenizers.pre_tokenizers import Whitespace
tokenizer = Tokenizer(BPE(unk_token="[UNK]"))
trainer = BpeTrainer(special_tokens=["[UNK]", "[CLS]", "[SEP]", "[PAD]", "[MASK]"])
tokenizer.pre_tokenizer = Whitespace()
files = []
tokenizer.train(files, trainer) |
We now have a tokenizer trained on the files we defined. We can either continue using it in that runtime, or save it to
a JSON file for future re-use.
Loading directly from the tokenizer object
Let's see how to leverage this tokenizer object in the 🤗 Transformers library. The
[PreTrainedTokenizerFast] class allows for easy instantiation, by accepting the instantiated
tokenizer object as an argument:
thon |
from transformers import PreTrainedTokenizerFast
fast_tokenizer = PreTrainedTokenizerFast(tokenizer_object=tokenizer)
This object can now be used with all the methods shared by the 🤗 Transformers tokenizers! Head to the tokenizer
page for more information.
Loading from a JSON file
In order to load a tokenizer from a JSON file, let's first start by saving our tokenizer:
thon
tokenizer.save("tokenizer.json") |
tokenizer.save("tokenizer.json")
The path to which we saved this file can be passed to the [PreTrainedTokenizerFast] initialization
method using the tokenizer_file parameter:
thon
from transformers import PreTrainedTokenizerFast
fast_tokenizer = PreTrainedTokenizerFast(tokenizer_file="tokenizer.json")
This object can now be used with all the methods shared by the 🤗 Transformers tokenizers! Head to the tokenizer
page for more information. |
Instantiating a big model
When you want to use a very big pretrained model, one challenge is to minimize the use of the RAM. The usual workflow
from PyTorch is:
Create your model with random weights.
Load your pretrained weights.
Put those pretrained weights in your random model. |
Create your model with random weights.
Load your pretrained weights.
Put those pretrained weights in your random model.
Step 1 and 2 both require a full version of the model in memory, which is not a problem in most cases, but if your model starts weighing several GigaBytes, those two copies can make you get out of RAM. Even worse, if you are using torch.distributed to launch a distributed training, each process will load the pretrained model and store these two copies in RAM. |
Note that the randomly created model is initialized with "empty" tensors, which take the space in memory without filling it (thus the random values are whatever was in this chunk of memory at a given time). The random initialization following the appropriate distribution for the kind of model/parameters instantiated (like a normal distribution for instance) is only performed after step 3 on the non-initialized weights, to be as fast as possible! |
In this guide, we explore the solutions Transformers offer to deal with this issue. Note that this is an area of active development, so the APIs explained here may change slightly in the future.
Sharded checkpoints
Since version 4.18.0, model checkpoints that end up taking more than 10GB of space are automatically sharded in smaller pieces. In terms of having one single checkpoint when you do model.save_pretrained(save_dir), you will end up with several partial checkpoints (each of which being of size < 10GB) and an index that maps parameter names to the files they are stored in.
You can control the maximum size before sharding with the max_shard_size parameter, so for the sake of an example, we'll use a normal-size models with a small shard size: let's take a traditional BERT model. |
from transformers import AutoModel
model = AutoModel.from_pretrained("google-bert/bert-base-cased")
If you save it using [~PreTrainedModel.save_pretrained], you will get a new folder with two files: the config of the model and its weights:
import os
import tempfile
with tempfile.TemporaryDirectory() as tmp_dir:
model.save_pretrained(tmp_dir)
print(sorted(os.listdir(tmp_dir)))
['config.json', 'pytorch_model.bin']
Now let's use a maximum shard size of 200MB: |
Now let's use a maximum shard size of 200MB:
with tempfile.TemporaryDirectory() as tmp_dir:
model.save_pretrained(tmp_dir, max_shard_size="200MB")
print(sorted(os.listdir(tmp_dir)))
['config.json', 'pytorch_model-00001-of-00003.bin', 'pytorch_model-00002-of-00003.bin', 'pytorch_model-00003-of-00003.bin', 'pytorch_model.bin.index.json'] |
On top of the configuration of the model, we see three different weights files, and an index.json file which is our index. A checkpoint like this can be fully reloaded using the [~PreTrainedModel.from_pretrained] method:
with tempfile.TemporaryDirectory() as tmp_dir:
model.save_pretrained(tmp_dir, max_shard_size="200MB")
new_model = AutoModel.from_pretrained(tmp_dir) |
The main advantage of doing this for big models is that during step 2 of the workflow shown above, each shard of the checkpoint is loaded after the previous one, capping the memory usage in RAM to the model size plus the size of the biggest shard.
Behind the scenes, the index file is used to determine which keys are in the checkpoint, and where the corresponding weights are stored. We can load that index like any json and get a dictionary: |
import json
with tempfile.TemporaryDirectory() as tmp_dir:
model.save_pretrained(tmp_dir, max_shard_size="200MB")
with open(os.path.join(tmp_dir, "pytorch_model.bin.index.json"), "r") as f:
index = json.load(f)
print(index.keys())
dict_keys(['metadata', 'weight_map'])
The metadata just consists of the total size of the model for now. We plan to add other information in the future:
index["metadata"]
{'total_size': 433245184} |
The metadata just consists of the total size of the model for now. We plan to add other information in the future:
index["metadata"]
{'total_size': 433245184}
The weights map is the main part of this index, which maps each parameter name (as usually found in a PyTorch model state_dict) to the file it's stored in:
index["weight_map"]
{'embeddings.LayerNorm.bias': 'pytorch_model-00001-of-00003.bin',
'embeddings.LayerNorm.weight': 'pytorch_model-00001-of-00003.bin', |
index["weight_map"]
{'embeddings.LayerNorm.bias': 'pytorch_model-00001-of-00003.bin',
'embeddings.LayerNorm.weight': 'pytorch_model-00001-of-00003.bin',
If you want to directly load such a sharded checkpoint inside a model without using [~PreTrainedModel.from_pretrained] (like you would do model.load_state_dict() for a full checkpoint) you should use [~modeling_utils.load_sharded_checkpoint]: |
from transformers.modeling_utils import load_sharded_checkpoint
with tempfile.TemporaryDirectory() as tmp_dir:
model.save_pretrained(tmp_dir, max_shard_size="200MB")
load_sharded_checkpoint(model, tmp_dir) |
Low memory loading
Sharded checkpoints reduce the memory usage during step 2 of the workflow mentioned above, but in order to use that model in a low memory setting, we recommend leveraging our tools based on the Accelerate library.
Please read the following guide for more information: Large model loading using Accelerate |
Efficient Training on CPU
This guide focuses on training large models efficiently on CPU.
Mixed precision with IPEX
Mixed precision uses single (fp32) and half-precision (bf16/fp16) data types in a model to accelerate training or inference while still preserving much of the single-precision accuracy. Modern CPUs such as 3rd and 4th Gen Intel® Xeon® Scalable processors natively support bf16, so you should get more performance out of the box by enabling mixed precision training with bf16.
To further maximize training performance, you can use Intel® Extension for PyTorch (IPEX), which is a library built on PyTorch and adds additional CPU instruction level architecture (ISA) level support such as Intel® Advanced Vector Extensions 512 Vector Neural Network Instructions (Intel® AVX512-VNNI), and Intel® Advanced Matrix Extensions (Intel® AMX) for an extra performance boost on Intel CPUs. However, CPUs with only AVX2 (e.g., AMD or older Intel CPUs) are not guaranteed to have better performance under IPEX.
Auto Mixed Precision (AMP) for CPU backends has been enabled since PyTorch 1.10. AMP support for bf16 on CPUs and bf16 operator optimization is also supported in IPEX and partially upstreamed to the main PyTorch branch. You can get better performance and user experience with IPEX AMP.
Check more detailed information for Auto Mixed Precision.
IPEX installation:
IPEX release is following PyTorch, to install via pip:
| PyTorch Version | IPEX version |
| :---------------: | :----------: |
| 2.1.x | 2.1.100+cpu |
| 2.0.x | 2.0.100+cpu |
| 1.13 | 1.13.0+cpu |
| 1.12 | 1.12.300+cpu |
Please run pip list | grep torch to get your pytorch_version, so you can get the IPEX version_name. |
pip install intel_extension_for_pytorch==<version_name> -f https://developer.intel.com/ipex-whl-stable-cpu
You can check the latest versions in ipex-whl-stable-cpu if needed.
Check more approaches for IPEX installation.
Usage in Trainer
To enable auto mixed precision with IPEX in Trainer, users should add use_ipex, bf16 and no_cuda in training command arguments.
Take an example of the use cases on Transformers question-answering
Training with IPEX using BF16 auto mixed precision on CPU: |
python run_qa.py \
--model_name_or_path google-bert/bert-base-uncased \
--dataset_name squad \
--do_train \
--do_eval \
--per_device_train_batch_size 12 \
--learning_rate 3e-5 \
--num_train_epochs 2 \
--max_seq_length 384 \
--doc_stride 128 \
--output_dir /tmp/debug_squad/ \
--use_ipex \
--bf16 \
--use_cpu
If you want to enable use_ipex and bf16 in your script, add these parameters to TrainingArguments like this:
diff
training_args = TrainingArguments(
output_dir=args.output_path,
+ bf16=True,
+ use_ipex=True,
+ use_cpu=True,
**kwargs
)
Practice example
Blog: Accelerating PyTorch Transformers with Intel Sapphire Rapids |
Model training anatomy
To understand performance optimization techniques that one can apply to improve efficiency of model training
speed and memory utilization, it's helpful to get familiar with how GPU is utilized during training, and how compute
intensity varies depending on an operation performed.
Let's start by exploring a motivating example of GPU utilization and the training run of a model. For the demonstration,
we'll need to install a few libraries: |
pip install transformers datasets accelerate nvidia-ml-py3
The nvidia-ml-py3 library allows us to monitor the memory usage of the models from within Python. You might be familiar
with the nvidia-smi command in the terminal - this library allows to access the same information in Python directly.
Then, we create some dummy data: random token IDs between 100 and 30000 and binary labels for a classifier.
In total, we get 512 sequences each with length 512 and store them in a [~datasets.Dataset] with PyTorch format. |
import numpy as np
from datasets import Dataset
seq_len, dataset_size = 512, 512
dummy_data = {
"input_ids": np.random.randint(100, 30000, (dataset_size, seq_len)),
"labels": np.random.randint(0, 1, (dataset_size)),
}
ds = Dataset.from_dict(dummy_data)
ds.set_format("pt")
To print summary statistics for the GPU utilization and the training run with the [Trainer] we define two helper functions: |
from pynvml import *
def print_gpu_utilization():
nvmlInit()
handle = nvmlDeviceGetHandleByIndex(0)
info = nvmlDeviceGetMemoryInfo(handle)
print(f"GPU memory occupied: {info.used//1024**2} MB.")
def print_summary(result):
print(f"Time: {result.metrics['train_runtime']:.2f}")
print(f"Samples/second: {result.metrics['train_samples_per_second']:.2f}")
print_gpu_utilization()
Let's verify that we start with a free GPU memory: |
Let's verify that we start with a free GPU memory:
print_gpu_utilization()
GPU memory occupied: 0 MB. |
print_gpu_utilization()
GPU memory occupied: 0 MB.
That looks good: the GPU memory is not occupied as we would expect before we load any models. If that's not the case on
your machine make sure to stop all processes that are using GPU memory. However, not all free GPU memory can be used by
the user. When a model is loaded to the GPU the kernels are also loaded, which can take up 1-2GB of memory. To see how
much it is we load a tiny tensor into the GPU which triggers the kernels to be loaded as well. |
import torch
torch.ones((1, 1)).to("cuda")
print_gpu_utilization()
GPU memory occupied: 1343 MB.
We see that the kernels alone take up 1.3GB of GPU memory. Now let's see how much space the model uses.
Load Model
First, we load the google-bert/bert-large-uncased model. We load the model weights directly to the GPU so that we can check
how much space just the weights use. |
from transformers import AutoModelForSequenceClassification
model = AutoModelForSequenceClassification.from_pretrained("google-bert/bert-large-uncased").to("cuda")
print_gpu_utilization()
GPU memory occupied: 2631 MB. |
We can see that the model weights alone take up 1.3 GB of GPU memory. The exact number depends on the specific
GPU you are using. Note that on newer GPUs a model can sometimes take up more space since the weights are loaded in an
optimized fashion that speeds up the usage of the model. Now we can also quickly check if we get the same result
as with nvidia-smi CLI: |
nvidia-smi
```bash
Tue Jan 11 08:58:05 2022
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 460.91.03 Driver Version: 460.91.03 CUDA Version: 11.2 |
|-------------------------------+----------------------+----------------------+
| GPU Name Persistence-M| Bus-Id Disp.A | Volatile Uncorr. ECC |
| Fan Temp Perf Pwr:Usage/Cap| Memory-Usage | GPU-Util Compute M. |
| | | MIG M. |
|===============================+======================+======================|
| 0 Tesla V100-SXM2 On | 00000000:00:04.0 Off | 0 |
| N/A 37C P0 39W / 300W | 2631MiB / 16160MiB | 0% Default |
| | | N/A |
+-------------------------------+----------------------+----------------------+
+-----------------------------------------------------------------------------+
| Processes: |
| GPU GI CI PID Type Process name GPU Memory |
| ID ID Usage |
|=============================================================================|
| 0 N/A N/A 3721 C nvs/codeparrot/bin/python 2629MiB |
+-----------------------------------------------------------------------------+ |
We get the same number as before and you can also see that we are using a V100 GPU with 16GB of memory. So now we can
start training the model and see how the GPU memory consumption changes. First, we set up a few standard training
arguments:
py
default_args = {
"output_dir": "tmp",
"evaluation_strategy": "steps",
"num_train_epochs": 1,
"log_level": "error",
"report_to": "none",
} |
If you plan to run multiple experiments, in order to properly clear the memory between experiments, restart the Python
kernel between experiments.
Memory utilization at vanilla training
Let's use the [Trainer] and train the model without using any GPU performance optimization techniques and a batch size of 4: |
Memory utilization at vanilla training
Let's use the [Trainer] and train the model without using any GPU performance optimization techniques and a batch size of 4:
from transformers import TrainingArguments, Trainer, logging
logging.set_verbosity_error()
training_args = TrainingArguments(per_device_train_batch_size=4, **default_args)
trainer = Trainer(model=model, args=training_args, train_dataset=ds)
result = trainer.train()
print_summary(result) |
Time: 57.82
Samples/second: 8.86
GPU memory occupied: 14949 MB.
We see that already a relatively small batch size almost fills up our GPU's entire memory. However, a larger batch size
can often result in faster model convergence or better end performance. So ideally we want to tune the batch size to our
model's needs and not to the GPU limitations. What's interesting is that we use much more memory than the size of the model.
To understand a bit better why this is the case let's have a look at a model's operations and memory needs.
Anatomy of Model's Operations
Transformers architecture includes 3 main groups of operations grouped below by compute-intensity. |
Tensor Contractions
Linear layers and components of Multi-Head Attention all do batched matrix-matrix multiplications. These operations are the most compute-intensive part of training a transformer.
Statistical Normalizations
Softmax and layer normalization are less compute-intensive than tensor contractions, and involve one or more reduction operations, the result of which is then applied via a map. |
Element-wise Operators
These are the remaining operators: biases, dropout, activations, and residual connections. These are the least compute-intensive operations. |
This knowledge can be helpful to know when analyzing performance bottlenecks.
This summary is derived from Data Movement Is All You Need: A Case Study on Optimizing Transformers 2020
Anatomy of Model's Memory
We've seen that training the model uses much more memory than just putting the model on the GPU. This is because there
are many components during training that use GPU memory. The components on GPU memory are the following: |
model weights
optimizer states
gradients
forward activations saved for gradient computation
temporary buffers
functionality-specific memory
A typical model trained in mixed precision with AdamW requires 18 bytes per model parameter plus activation memory. For
inference there are no optimizer states and gradients, so we can subtract those. And thus we end up with 6 bytes per
model parameter for mixed precision inference, plus activation memory.
Let's look at the details.
Model Weights: |
4 bytes * number of parameters for fp32 training
6 bytes * number of parameters for mixed precision training (maintains a model in fp32 and one in fp16 in memory)
Optimizer States:
8 bytes * number of parameters for normal AdamW (maintains 2 states)
2 bytes * number of parameters for 8-bit AdamW optimizers like bitsandbytes
4 bytes * number of parameters for optimizers like SGD with momentum (maintains only 1 state)
Gradients |
Gradients
4 bytes * number of parameters for either fp32 or mixed precision training (gradients are always kept in fp32)
Forward Activations
size depends on many factors, the key ones being sequence length, hidden size and batch size. |
There are the input and output that are being passed and returned by the forward and the backward functions and the
forward activations saved for gradient computation.
Temporary Memory
Additionally, there are all kinds of temporary variables which get released once the calculation is done, but in the
moment these could require additional memory and could push to OOM. Therefore, when coding it's crucial to think
strategically about such temporary variables and sometimes to explicitly free those as soon as they are no longer needed.
Functionality-specific memory
Then, your software could have special memory needs. For example, when generating text using beam search, the software
needs to maintain multiple copies of inputs and outputs.
forward vs backward Execution Speed
For convolutions and linear layers there are 2x flops in the backward compared to the forward, which generally translates
into ~2x slower (sometimes more, because sizes in the backward tend to be more awkward). Activations are usually
bandwidth-limited, and it’s typical for an activation to have to read more data in the backward than in the forward
(e.g. activation forward reads once, writes once, activation backward reads twice, gradOutput and output of the forward,
and writes once, gradInput).
As you can see, there are potentially a few places where we could save GPU memory or speed up operations.
Now that you understand what affects GPU utilization and computation speed, refer to
the Methods and tools for efficient training on a single GPU documentation page to learn about
performance optimization techniques. |
Perplexity of fixed-length models
[[open-in-colab]]
Perplexity (PPL) is one of the most common metrics for evaluating language models. Before diving in, we should note
that the metric applies specifically to classical language models (sometimes called autoregressive or causal language
models) and is not well defined for masked language models like BERT (see summary of the models).
Perplexity is defined as the exponentiated average negative log-likelihood of a sequence. If we have a tokenized
sequence \(X = (x_0, x_1, \dots, x_t)\), then the perplexity of \(X\) is,
$$\text{PPL}(X) = \exp \left{ {-\frac{1}{t}\sum_i^t \log p_\theta (x_i|x_{<i}) } \right}$$
where \(\log p_\theta (x_i|x_{<i})\) is the log-likelihood of the ith token conditioned on the preceding tokens \(x_{<i}\) according to our model. Intuitively, it can be thought of as an evaluation of the model's ability to predict uniformly among the set of specified tokens in a corpus. Importantly, this means that the tokenization procedure has a direct impact on a model's perplexity which should always be taken into consideration when comparing different models.
This is also equivalent to the exponentiation of the cross-entropy between the data and model predictions. For more
intuition about perplexity and its relationship to Bits Per Character (BPC) and data compression, check out this
fantastic blog post on The Gradient.
Calculating PPL with fixed-length models
If we weren't limited by a model's context size, we would evaluate the model's perplexity by autoregressively
factorizing a sequence and conditioning on the entire preceding subsequence at each step, as shown below. |
When working with approximate models, however, we typically have a constraint on the number of tokens the model can
process. The largest version of GPT-2, for example, has a fixed length of 1024 tokens, so we
cannot calculate \(p_\theta(x_t|x_{<t})\) directly when \(t\) is greater than 1024.
Instead, the sequence is typically broken into subsequences equal to the model's maximum input size. If a model's max
input size is \(k\), we then approximate the likelihood of a token \(x_t\) by conditioning only on the
\(k-1\) tokens that precede it rather than the entire context. When evaluating the model's perplexity of a
sequence, a tempting but suboptimal approach is to break the sequence into disjoint chunks and add up the decomposed
log-likelihoods of each segment independently. |
This is quick to compute since the perplexity of each segment can be computed in one forward pass, but serves as a poor
approximation of the fully-factorized perplexity and will typically yield a higher (worse) PPL because the model will
have less context at most of the prediction steps.
Instead, the PPL of fixed-length models should be evaluated with a sliding-window strategy. This involves repeatedly
sliding the context window so that the model has more context when making each prediction. |
This is a closer approximation to the true decomposition of the sequence probability and will typically yield a more
favorable score. The downside is that it requires a separate forward pass for each token in the corpus. A good
practical compromise is to employ a strided sliding window, moving the context by larger strides rather than sliding by
1 token a time. This allows computation to proceed much faster while still giving the model a large context to make
predictions at each step.
Example: Calculating perplexity with GPT-2 in 🤗 Transformers
Let's demonstrate this process with GPT-2.
thon
from transformers import GPT2LMHeadModel, GPT2TokenizerFast
device = "cuda"
model_id = "openai-community/gpt2-large"
model = GPT2LMHeadModel.from_pretrained(model_id).to(device)
tokenizer = GPT2TokenizerFast.from_pretrained(model_id) |
We'll load in the WikiText-2 dataset and evaluate the perplexity using a few different sliding-window strategies. Since
this dataset is small and we're just doing one forward pass over the set, we can just load and encode the entire
dataset in memory.
thon
from datasets import load_dataset
test = load_dataset("wikitext", "wikitext-2-raw-v1", split="test")
encodings = tokenizer("\n\n".join(test["text"]), return_tensors="pt") |
With 🤗 Transformers, we can simply pass the input_ids as the labels to our model, and the average negative
log-likelihood for each token is returned as the loss. With our sliding window approach, however, there is overlap in
the tokens we pass to the model at each iteration. We don't want the log-likelihood for the tokens we're just treating
as context to be included in our loss, so we can set these targets to -100 so that they are ignored. The following
is an example of how we could do this with a stride of 512. This means that the model will have at least 512 tokens
for context when calculating the conditional likelihood of any one token (provided there are 512 preceding tokens
available to condition on).
thon
import torch
from tqdm import tqdm
max_length = model.config.n_positions
stride = 512
seq_len = encodings.input_ids.size(1)
nlls = []
prev_end_loc = 0
for begin_loc in tqdm(range(0, seq_len, stride)):
end_loc = min(begin_loc + max_length, seq_len)
trg_len = end_loc - prev_end_loc # may be different from stride on last loop
input_ids = encodings.input_ids[:, begin_loc:end_loc].to(device)
target_ids = input_ids.clone()
target_ids[:, :-trg_len] = -100
with torch.no_grad():
outputs = model(input_ids, labels=target_ids) |
# loss is calculated using CrossEntropyLoss which averages over valid labels
# N.B. the model only calculates loss over trg_len - 1 labels, because it internally shifts the labels
# to the left by 1.
neg_log_likelihood = outputs.loss
nlls.append(neg_log_likelihood)
prev_end_loc = end_loc
if end_loc == seq_len:
break
ppl = torch.exp(torch.stack(nlls).mean()) |
Running this with the stride length equal to the max input length is equivalent to the suboptimal, non-sliding-window
strategy we discussed above. The smaller the stride, the more context the model will have in making each prediction,
and the better the reported perplexity will typically be.
When we run the above with stride = 1024, i.e. no overlap, the resulting PPL is 19.44, which is about the same
as the 19.93 reported in the GPT-2 paper. By using stride = 512 and thereby employing our striding window
strategy, this jumps down to 16.45. This is not only a more favorable score, but is calculated in a way that is
closer to the true autoregressive decomposition of a sequence likelihood. |
Philosophy
🤗 Transformers is an opinionated library built for:
machine learning researchers and educators seeking to use, study or extend large-scale Transformers models.
hands-on practitioners who want to fine-tune those models or serve them in production, or both.
engineers who just want to download a pretrained model and use it to solve a given machine learning task.
The library was designed with two strong goals in mind:
Be as easy and fast to use as possible: |
The library was designed with two strong goals in mind:
Be as easy and fast to use as possible:
We strongly limited the number of user-facing abstractions to learn, in fact, there are almost no abstractions,
just three standard classes required to use each model: configuration,
models, and a preprocessing class (tokenizer for NLP, image processor for vision, feature extractor for audio, and processor for multimodal inputs). |
All of these classes can be initialized in a simple and unified way from pretrained instances by using a common
from_pretrained() method which downloads (if needed), caches and
loads the related class instance and associated data (configurations' hyperparameters, tokenizers' vocabulary,
and models' weights) from a pretrained checkpoint provided on Hugging Face Hub or your own saved checkpoint.
On top of those three base classes, the library provides two APIs: [pipeline] for quickly
using a model for inference on a given task and [Trainer] to quickly train or fine-tune a PyTorch model (all TensorFlow models are compatible with Keras.fit). |
As a consequence, this library is NOT a modular toolbox of building blocks for neural nets. If you want to
extend or build upon the library, just use regular Python, PyTorch, TensorFlow, Keras modules and inherit from the base
classes of the library to reuse functionalities like model loading and saving. If you'd like to learn more about our coding philosophy for models, check out our Repeat Yourself blog post. |
Provide state-of-the-art models with performances as close as possible to the original models:
We provide at least one example for each architecture which reproduces a result provided by the official authors
of said architecture.
The code is usually as close to the original code base as possible which means some PyTorch code may be not as
pytorchic as it could be as a result of being converted TensorFlow code and vice versa.
A few other goals: |
A few other goals:
Expose the models' internals as consistently as possible:
We give access, using a single API, to the full hidden-states and attention weights.
The preprocessing classes and base model APIs are standardized to easily switch between models.
Incorporate a subjective selection of promising tools for fine-tuning and investigating these models:
A simple and consistent way to add new tokens to the vocabulary and embeddings for fine-tuning.
Simple ways to mask and prune Transformer heads. |
A simple and consistent way to add new tokens to the vocabulary and embeddings for fine-tuning.
Simple ways to mask and prune Transformer heads.
Easily switch between PyTorch, TensorFlow 2.0 and Flax, allowing training with one framework and inference with another.
Main concepts
The library is built around three types of classes for each model: |
Model classes can be PyTorch models (torch.nn.Module), Keras models (tf.keras.Model) or JAX/Flax models (flax.linen.Module) that work with the pretrained weights provided in the library.
Configuration classes store the hyperparameters required to build a model (such as the number of layers and hidden size). You don't always need to instantiate these yourself. In particular, if you are using a pretrained model without any modification, creating the model will automatically take care of instantiating the configuration (which is part of the model).
Preprocessing classes convert the raw data into a format accepted by the model. A tokenizer stores the vocabulary for each model and provide methods for encoding and decoding strings in a list of token embedding indices to be fed to a model. Image processors preprocess vision inputs, feature extractors preprocess audio inputs, and a processor handles multimodal inputs. |
All these classes can be instantiated from pretrained instances, saved locally, and shared on the Hub with three methods: |
Subsets and Splits