eedd21b9e7
* Fixed typo: insted to instead * Fixed typo: relase to release * Fixed typo: nighlty to nightly * Fixed typos: versatible, benchamarks, becnhmark to versatile, benchmark, benchmarks * Fixed typo in comment: quantizd to quantized * Fixed typo: architecutre to architecture * Fixed typo: contibution to contribution * Fixed typo: Presequities to Prerequisites * Fixed typo: faste to faster * Fixed typo: extendeding to extending * Fixed typo: segmetantion_maps to segmentation_maps * Fixed typo: Alternativelly to Alternatively * Fixed incorrectly defined variable: output to output_disabled * Fixed typo in library name: tranformers.onnx to transformers.onnx * Fixed missing import: import tensorflow as tf * Fixed incorrectly defined variable: token_tensor to tokens_tensor * Fixed missing import: import torch * Fixed incorrectly defined variable and typo: uromaize to uromanize * Fixed incorrectly defined variable and typo: uromaize to uromanize * Fixed typo in function args: numpy.ndarry to numpy.ndarray * Fixed Inconsistent Library Name: Torchscript to TorchScript * Fixed Inconsistent Class Name: OneformerProcessor to OneFormerProcessor * Fixed Inconsistent Class Named Typo: TFLNetForMultipleChoice to TFXLNetForMultipleChoice * Fixed Inconsistent Library Name Typo: Pytorch to PyTorch * Fixed Inconsistent Function Name Typo: captureWarning to captureWarnings * Fixed Inconsistent Library Name Typo: Pytorch to PyTorch * Fixed Inconsistent Class Name Typo: TrainingArgument to TrainingArguments * Fixed Inconsistent Model Name Typo: Swin2R to Swin2SR * Fixed Inconsistent Model Name Typo: EART to BERT * Fixed Inconsistent Library Name Typo: TensorFLow to TensorFlow * Fixed Broken Link for Speech Emotion Classification with Wav2Vec2 * Fixed minor missing word Typo * Fixed minor missing word Typo * Fixed minor missing word Typo * Fixed minor missing word Typo * Fixed minor missing word Typo * Fixed minor missing word Typo * Fixed minor missing word Typo * Fixed minor missing word Typo * Fixed Punctuation: Two commas * Fixed Punctuation: No Space between XLM-R and is * Fixed Punctuation: No Space between [~accelerate.Accelerator.backward] and method * Added backticks to display model.fit() in codeblock * Added backticks to display openai-community/gpt2 in codeblock * Fixed Minor Typo: will to with * Fixed Minor Typo: is to are * Fixed Minor Typo: in to on * Fixed Minor Typo: inhibits to exhibits * Fixed Minor Typo: they need to it needs * Fixed Minor Typo: cast the load the checkpoints To load the checkpoints * Fixed Inconsistent Class Name Typo: TFCamembertForCasualLM to TFCamembertForCausalLM * Fixed typo in attribute name: outputs.last_hidden_states to outputs.last_hidden_state * Added missing verbosity level: fatal * Fixed Minor Typo: take To takes * Fixed Minor Typo: heuristic To heuristics * Fixed Minor Typo: setting To settings * Fixed Minor Typo: Content To Contents * Fixed Minor Typo: millions To million * Fixed Minor Typo: difference To differences * Fixed Minor Typo: while extract To which extracts * Fixed Minor Typo: Hereby To Here * Fixed Minor Typo: addition To additional * Fixed Minor Typo: supports To supported * Fixed Minor Typo: so that benchmark results TO as a consequence, benchmark * Fixed Minor Typo: a To an * Fixed Minor Typo: a To an * Fixed Minor Typo: Chain-of-though To Chain-of-thought
3.9 KiB
Performance and Scalability
Training large transformer models and deploying them to production present various challenges.
During training, the model may require more GPU memory than available or exhibit slow training speed. In the deployment
phase, the model can struggle to handle the required throughput in a production environment.
This documentation aims to assist you in overcoming these challenges and finding the optimal settings for your use-case. The guides are divided into training and inference sections, as each comes with different challenges and solutions. Within each section you'll find separate guides for different hardware configurations, such as single GPU vs. multi-GPU for training or CPU vs. GPU for inference.
Use this document as your starting point to navigate further to the methods that match your scenario.
Training
Training large transformer models efficiently requires an accelerator such as a GPU or TPU. The most common case is where you have a single GPU. The methods that you can apply to improve training efficiency on a single GPU extend to other setups such as multiple GPU. However, there are also techniques that are specific to multi-GPU or CPU training. We cover them in separate sections.
- Methods and tools for efficient training on a single GPU: start here to learn common approaches that can help optimize GPU memory utilization, speed up the training, or both.
- Multi-GPU training section: explore this section to learn about further optimization methods that apply to a multi-GPU settings, such as data, tensor, and pipeline parallelism.
- CPU training section: learn about mixed precision training on CPU.
- Efficient Training on Multiple CPUs: learn about distributed CPU training.
- Training on TPU with TensorFlow: if you are new to TPUs, refer to this section for an opinionated introduction to training on TPUs and using XLA.
- Custom hardware for training: find tips and tricks when building your own deep learning rig.
- Hyperparameter Search using Trainer API
Inference
Efficient inference with large models in a production environment can be as challenging as training them. In the following sections we go through the steps to run inference on CPU and single/multi-GPU setups.
- Inference on a single CPU
- Inference on a single GPU
- Multi-GPU inference
- XLA Integration for TensorFlow Models
Training and inference
Here you'll find techniques, tips and tricks that apply whether you are training a model, or running inference with it.
Contribute
This document is far from being complete and a lot more needs to be added, so if you have additions or corrections to make please don't hesitate to open a PR or if you aren't sure start an Issue and we can discuss the details there.
When making contributions that A is better than B, please try to include a reproducible benchmark and/or a link to the source of that information (unless it comes directly from you).