Skip to main content

PyTorch 2.0 introduces accelerated Transformer API to democratize ML

The PyTorch team has officially released PyTorch 2.0, which was first previewed back in December 2022 at the PyTorch Conference. 

PyTorch is a Linux Foundation machine learning framework that was originally developed by Meta. 

This release includes a high-performance implementation of the Transformer API. It supports more use cases now, such as models using Cross-Attention, Transformer Decoders, and for training models. The goal of releasing this new API is to make training and deployment of Transformer models more cost effective and affordable, the team explained. 

PyTorch 2.0 also introduces torch.compile as the main API for wrapping models and returning a compiled model. This is a completely additive feature, helping to maintain backwards compatibility.  

Torch.compile is built on four other new technologies: 

  1. TorchDynamo, which uses Python Frame Evaluation Hooks to safely capture PyTorch programs
  2. AOTAutogram, which can be used to generate ahead-of-time backward traces
  3. PrimTorch, which condenses over 2,000 PyTorch operators down into a set of 250 that can be targeted to build a complete PyTorch backend, significantly reducing the barrier to entry
  4. TorchInductor, which is a deep learning compiler that makes use of OpenAI Triton.

“We have achieved major speedups for training transformer models and in particular large language models with Accelerated PyTorch 2 Transformers using a combination of custom kernels and torch.compile(),” the PyTorch team wrote in a blog post. 

This release also adds support for 60 new operators to the Metal Performance Shaders (MPS) backend, which provides GPU accelerated training on macOS platforms. This brings the total coverage to 300 operators to-date. 

AWS customers will see improved performance on AWS Graviton compared to previous releases. These improvements focus on GEMM kernels, bfloat16 support, primitive caching, and the memory allocator. 

This release also includes several beta updates to PyTorch domain libraries and other libraries like TorchAudio, TorchVision, and TorchText. 

There are also several features in the prototype stage across many features, including TensorParallel, DTensor, 2D parallel, TorchDynamo, AOTAutograd, PrimTorch and TorchInductor.

The post PyTorch 2.0 introduces accelerated Transformer API to democratize ML appeared first on SD Times.



from SD Times https://ift.tt/0YoK9hI
Labels:

Comments

Post a Comment

Popular posts from this blog

Difference between Web Designer and Web Developer Neeraj Mishra The Crazy Programmer

Have you ever wondered about the distinctions between web developers’ and web designers’ duties and obligations? You’re not alone! Many people have trouble distinguishing between these two. Although they collaborate to publish new websites on the internet, web developers and web designers play very different roles. To put these job possibilities into perspective, consider the construction of a house. To create a vision for the house, including the visual components, the space planning and layout, the materials, and the overall appearance and sense of the space, you need an architect. That said, to translate an idea into a building, you need construction professionals to take those architectural drawings and put them into practice. Image Source In a similar vein, web development and design work together to create websites. Let’s examine the major responsibilities and distinctions between web developers and web designers. Let’s get going, shall we? What Does a Web Designer Do?
Post a Comment
Read more

A guide to data integration tools

CData Software is a leader in data access and connectivity solutions. It specializes in the development of data drivers and data access technologies for real-time access to online or on-premise applications, databases and web APIs. The company is focused on bringing data connectivity capabilities natively into tools organizations already use. It also features ETL/ELT solutions, enterprise connectors, and data visualization. Matillion ’s data transformation software empowers customers to extract data from a wide number of sources, load it into their chosen cloud data warehouse (CDW) and transform that data from its siloed source state, into analytics-ready insights – prepared for advanced analytics, machine learning, and artificial intelligence use cases. Only Matillion is purpose-built for Snowflake, Amazon Redshift, Google BigQuery, and Microsoft Azure, enabling businesses to achieve new levels of simplicity, speed, scale, and savings. Trusted by companies of all sizes to meet
Post a Comment
Read more

2022: The year of hybrid work

Remote work was once considered a luxury to many, but in 2020, it became a necessity for a large portion of the workforce, as the scary and unknown COVID-19 virus sickened and even took the lives of so many people around the world.  Some workers were able to thrive in a remote setting, while others felt isolated and struggled to keep up a balance between their work and home lives. Last year saw the availability of life-saving vaccines, so companies were able to start having the conversation about what to do next. Should they keep everyone remote? Should they go back to working in the office full time? Or should they do something in between? Enter hybrid work, which offers a mix of the two. A Fall 2021 study conducted by Google revealed that over 75% of survey respondents expect hybrid work to become a standard practice within their organization within the next three years.  Thus, two years after the world abruptly shifted to widespread adoption of remote work, we are declaring 20
Post a Comment
Read more