Making DeepSpeed ZeRO run efficiently on more-affordable hardware

Amazon researchers optimize the distributed-training tool to run efficiently on the Elastic Fabric Adapter network interface.

Most modern natural-language-processing applications are built on top of pretrained language models, which encode the probabilities of word sequences for entire languages. Over time, these models have trended larger and larger, into the regime of billions or even trillions of parameters.

Training these models within a reasonable amount of time requires very large computing clusters, whose large communication volume can block computation, resulting in low or inefficient GPU utilization. Communication between the GPUs needs to be carefully managed to avoid becoming a performance bottleneck.

Related content
How SageMaker’s data-parallel and model-parallel engines make training neural networks easier, faster, and cheaper.

Microsoft’s DeepSpeed distributed-training library introduced one such management technique, called the Zero Redundancy Optimizer (ZeRO). ZeRO works by partitioning the state of a machine learning model across distributed workers and fetching the necessary model state from other workers as needed during training. ZeRO has several “stages,” each of which allows the training of larger models through reducing memory requirements, typically at the cost of additional communication volume.

While Microsoft researchers were able to achieve ideal scaling performance with this technique, they reported experiments only on a specialized hypercluster that uses expensive high-speed InfiniBand networking (specifically, an Nvidia DGX system).

To reduce costs for customers in need of high-performance computing, Amazon Web Services (AWS) uses an Elastic Fabric Adapter (EFA) network instead of InfiniBand. The EFA available on instances of AWS’s p4d.24xlarge computational infrastructure has less communication bandwidth than InfiniBand on the Nvidia DGX hypercluster, so we would expect some performance dropoff for bandwidth-intensive tasks. When we tried to reproduce Microsoft’s results, however, we found that the relative dropoff in ZeRO’s third stage was twice that of the dropoff in the second stage.

We profiled the training process to look for bottlenecks and observed that in ZeRO Stage 3, communication dominated training time. We have made a series of optimizations to ZeRO Stage 3 in order to close the performance gap relative to results obtained on InfiniBand-equipped DGX clusters. Below is a table showing the overall performance improvement conferred by our optimizations, measured when training a RoBERTa language model on AWS p4d.24xlarge instances.

Model

Number of GPUs

TFLOPS/GPU

RoBERTa-10B

64

Optimized: 123 teraflops
Unoptimized: 73 teraflops

RoBERTa-50B

64

Optimized: 154 teraflops
Unoptimized: 89 teraflops

In January, we merged our optimizations into the DeepSpeed code repository for public use.

Optimizations

Related content
Earlier this year, we reported a speech recognition system trained on a million hours of data, a feat possible through semi-supervised learning, in which training data is annotated by machines rather than by people. These sorts of massive machine learning projects are becoming more common, and they require distributing the training process across multiple processors. Otherwise, training becomes too time consuming.

Our optimizations can roughly be categorized as (1) improving overlap between communication and computation, (2) improving bandwidth utilization, and (3) improving memory efficiency.

Synchronization/Parallelism

Finer-grained synchronization between communication and computation streams

In lower-bandwidth or large clusters where communication times dominate, it is critical to mask communication costs by overlapping computation with communication. Through profiling, we found that this overlapping was limited by ZeRO’s overly coarse synchronization.

This resulted in a suboptimal level of overlapping for two distributed-computing operations: allgather, which aggregates data (in this case, model parameters) from all workers across the network, and reduce-scatter, which reduces data (in this case, summing gradients) across workers. These two operations were causing poor GPU utilization because communication was constantly blocking computation operations. In response, we made significant changes to the parameter gathering and gradient reduce-scatter paths to reduce or remove synchronization while maintaining correctness.

After these changes, we were able to achieve much better overlapping and thus much fewer and smaller computation bubbles.

Precomputation/caching of Python fetching and partitioning decisions 

Related content
“Anytime query” approach adapts to the available resources.

During training, many complex decisions need to be made, relating to which parameters should be fetched, which parameters will be used next, which parameters may be reused soon and should be kept, and which can be released. These operations were slow enough to frequently prevent the Python process from keeping GPUs fed with work, creating large computation bubbles.

We optimized this by precomputing or caching as many decisions as possible, speeding up their computation to the point that it has become a nonfactor for training throughput.

Communication/bandwidth use

Batching allgather/reduce-scatter calls

We found that batching the collective operations — allgather and reduce-scatter — uses bandwidth more efficiently and amortizes the fixed costs of running the computational kernels that execute the operations. To implement collective batching, we flatten tensor data into a single, contiguous buffer to be sent in a single transaction. Each collective requires a special interleaving scheme to ensure that each worker receives the correct data.

Allgather-interleaving.jpeg
Allgather interleaving scheme.
Reduce-scatter interleaving.jpeg
Reduce-scatter interleaving scheme.

Memory

Our implementation of ZeRO, like the Microsoft implementation, uses the Compute Unified Device Architecture (CUDA), Nvidia’s parallel-computing platform. CUDA memory allocations are both synchronous and slow (ignoring the stream-ordered alternatives cudaMallocAsync and cudaMemcpyAsync, which are not yet used in PyTorch), so PyTorch uses a caching allocator to avoid the large costs of constantly reallocating memory. If there are no cached or free blocks for an allocation request, the allocator will flush its cache. This is disastrous for a few reasons:

  • Before the flush can begin, several cudaEventSynchronize calls are necessary to allow computation on held memory to complete. This and the subsequent cudaFree calls can take multiple seconds.
  • Different workers are not guaranteed to flush their caches simultaneously. This means that for any collective, if even a single worker is currently flushing its cache, the other N-1 workers sit blocked waiting for that worker to join. As cluster size increases, so does the probability that at least one worker is flushing its cache for any given collective.
  • After the cache flush, subsequent allocations require cudaMalloc calls, which as mentioned earlier are both synchronous and slow.

For these reasons, memory efficiency is critical for performance.

Memory-efficient batched PyTorch collectives

Although our use of batched collectives significantly reduced kernel launch overhead and improved bandwidth utilization, it also increased memory consumption because of its flattening of batched tensors into an additional buffer.

Related content
New method enables two- to 14-fold speedups over best-performing predecessors.

To avoid redundant flatten operations in PyTorch collectives, we used the *_base variants of the collective operations, which accept pre-flattened tensors, avoiding the need to internally allocate additional flattened buffers. In future work, we plan to use group-based batching operations from the Nvidia Collective Communications Library (NCCL) to eliminate all flattening operations.

More aggressive initialization-time defragmentation of parameter partitions

Even with more than 10GB of free GPU memory, we continued to see evidence of allocator cache flushes, suggesting memory fragmentation. In order to reduce this, we made initialization-time defragmentation changes to move all persisted tensors into a single contiguous buffer.

Miscellaneous

In addition to the optimizations described above, we also

  • optimized gradient normalization by reducing host-device data movement and synchronization and pulling math operations out of a for-loop into a single kernel launch with parallelized computation; and
  • removed tensor operations (.norm()) that were being added to debug messages via string formatting. (These were causing copies from host to device, which meant data movement and host-device synchronization.)

By making DeepSpeed ZeRO Stage 3 performant on widely available public cloud offerings, we hope to further democratize the training of large language models.

Acknowledgments: Zhen Zhang, Stephen Rawls, Yida Wang

Related content

IN, HR, Gurugram
Our customers have immense faith in our ability to deliver packages timely and as expected. A well planned network seamlessly scales to handle millions of package movements a day. It has monitoring mechanisms that detect failures before they even happen (such as predicting network congestion, operations breakdown), and perform proactive corrective actions. When failures do happen, it has inbuilt redundancies to mitigate impact (such as determine other routes or service providers that can handle the extra load), and avoids relying on single points of failure (service provider, node, or arc). Finally, it is cost optimal, so that customers can be passed the benefit from an efficiently set up network. Amazon Shipping is hiring Applied Scientists to help improve our ability to plan and execute package movements. As an Applied Scientist in Amazon Shipping, you will work on multiple challenging machine learning problems spread across a wide spectrum of business problems. You will build ML models to help our transportation cost auditing platforms effectively audit off-manifest (discrepancies between planned and actual shipping cost). You will build models to improve the quality of financial and planning data by accurately predicting ship cost at a package level. Your models will help forecast the packages required to be pick from shipper warehouses to reduce First Mile shipping cost. Using signals from within the transportation network (such as network load, and velocity of movements derived from package scan events) and outside (such as weather signals), you will build models that predict delivery delay for every package. These models will help improve buyer experience by triggering early corrective actions, and generating proactive customer notifications. Your role will require you to demonstrate Think Big and Invent and Simplify, by refining and translating Transportation domain-related business problems into one or more Machine Learning problems. You will use techniques from a wide array of machine learning paradigms, such as supervised, unsupervised, semi-supervised and reinforcement learning. Your model choices will include, but not be limited to, linear/logistic models, tree based models, deep learning models, ensemble models, and Q-learning models. You will use techniques such as LIME and SHAP to make your models interpretable for your customers. You will employ a family of reusable modelling solutions to ensure that your ML solution scales across multiple regions (such as North America, Europe, Asia) and package movement types (such as small parcel movements and truck movements). You will partner with Applied Scientists and Research Scientists from other teams in US and India working on related business domains. Your models are expected to be of production quality, and will be directly used in production services. You will work as part of a diverse data science and engineering team comprising of other Applied Scientists, Software Development Engineers and Business Intelligence Engineers. You will participate in the Amazon ML community by authoring scientific papers and submitting them to Machine Learning conferences. You will mentor Applied Scientists and Software Development Engineers having a strong interest in ML. You will also be called upon to provide ML consultation outside your team for other problem statements. If you are excited by this charter, come join us!
US, WA, Seattle
Do you want to re-invent how millions of people consume video content on their TVs, Tablets and Alexa? We are building a free to watch streaming service called Fire TV Channels (https://techcrunch.com/2023/08/21/amazon-launches-fire-tv-channels-app-400-fast-channels/). Our goal is to provide customers with a delightful and personalized experience for consuming content across News, Sports, Cooking, Gaming, Entertainment, Lifestyle and more. You will work closely with engineering and product stakeholders to realize our ambitious product vision. You will get to work with Generative AI and other state of the art technologies to help build personalization and recommendation solutions from the ground up. You will be in the driver's seat to present customers with content they will love. Using Amazon’s large-scale computing resources, you will ask research questions about customer behavior, build state-of-the-art models to generate recommendations and run these models to enhance the customer experience. You will participate in the Amazon ML community and mentor Applied Scientists and Software Engineers with a strong interest in and knowledge of ML. Your work will directly benefit customers and you will measure the impact using scientific tools.
US, MA, Boston
The Artificial General Intelligence (AGI) team is looking for a passionate, talented, and inventive Applied Scientist with a strong deep learning background, to build industry-leading technology with Large Language Models (LLMs) and multi-modal systems. You will support projects that work on technologies including multi-modal model alignment, moderation systems and evaluation. Key job responsibilities As an Applied Scientist with the AGI team, you will support the development of novel algorithms and modeling techniques, to advance the state of the art with LLMs. Your work will directly impact our customers in the form of products and services that make use of speech and language technology. You will leverage Amazon’s heterogeneous data sources and large-scale computing resources to accelerate advances in generative artificial intelligence (GenAI). You are also expected to publish in top tier conferences. About the team The AGI team has a mission to push the envelope in LLMs and multimodal systems. Specifically, we focus on model alignment with an aim to maintain safety while not denting utility, in order to provide the best-possible experience for our customers.
US, MA, Boston
The Artificial General Intelligence (AGI) team is looking for a passionate, talented, and inventive Senior Applied Scientist with a strong deep learning background, to build industry-leading technology with Large Language Models (LLMs) and multimodal systems. Key job responsibilities As a Senior Applied Scientist with the AGI team, you will work with talented peers to lead the development of novel algorithms and modeling techniques, to advance the state of the art with LLMs. Your work will directly impact our customers in the form of products and services that make use of speech and language technology. You will leverage Amazon’s heterogeneous data sources and large-scale computing resources to accelerate advances in generative artificial intelligence (GenAI). About the team The AGI team has a mission to push the envelope in LLMs and multimodal systems, in order to provide the best-possible experience for our customers.
IN, KA, Bengaluru
The Amazon Alexa AI team in India is seeking a talented, self-driven Applied Scientist to work on prototyping, optimizing, and deploying ML algorithms within the realm of Generative AI. Key responsibilities include: - Research, experiment and build Proof Of Concepts advancing the state of the art in AI & ML for GenAI. - Collaborate with cross-functional teams to architect and execute technically rigorous AI projects. - Thrive in dynamic environments, adapting quickly to evolving technical requirements and deadlines. - Engage in effective technical communication (written & spoken) with coordination across teams. - Conduct thorough documentation of algorithms, methodologies, and findings for transparency and reproducibility. - Publish research papers in internal and external venues of repute - Support on-call activities for critical issues Basic Qualifications: - Master’s or PhD in computer science, statistics or a related field - 2-7 years experience in deep learning, machine learning, and data science. - Proficiency in coding and software development, with a strong focus on machine learning frameworks. - Experience in Python, or another language; command line usage; familiarity with Linux and AWS ecosystems. - Understanding of relevant statistical measures such as confidence intervals, significance of error measurements, development and evaluation data sets, etc. - Excellent communication skills (written & spoken) and ability to collaborate effectively in a distributed, cross-functional team setting. - Papers published in AI/ML venues of repute Preferred Qualifications: - Track record of diving into data to discover hidden patterns and conducting error/deviation analysis - Ability to develop experimental and analytic plans for data modeling processes, use of strong baselines, ability to accurately determine cause and effect relations - The motivation to achieve results in a fast-paced environment. - Exceptional level of organization and strong attention to detail - Comfortable working in a fast paced, highly collaborative, dynamic work environment
IN, KA, Bengaluru
Amazon is investing heavily in building a world class advertising business and we are responsible for defining and delivering a collection of self-service performance advertising products that drive discovery and sales. Our products are strategically important to our Retail and Marketplace businesses driving long term growth. We deliver billions of ad impressions and millions of clicks daily and are breaking fresh ground to create world-class products. We are highly motivated, collaborative and fun-loving with an entrepreneurial spirit and bias for action. With a broad mandate to experiment and innovate, we are growing at an unprecedented rate with a seemingly endless range of new opportunities. The ATT team, based in Bangalore, is responsible for ensuring that ads are relevant and is of good quality, leading to higher conversion for the sellers and providing a great experience for the customers. We deal with one of the world’s largest product catalog, handle billions of requests a day with plans to grow it by order of magnitude and use automated systems to validate tens of millions of offers submitted by thousands of merchants in multiple countries and languages. In this role, you will build and develop ML models to address content understanding problems in Ads. These models will rely on a variety of visual and textual features requiring expertise in both domains. These models need to scale to multiple languages and countries. You will collaborate with engineers and other scientists to build, train and deploy these models. As part of these activities, you will develop production level code that enables moderation of millions of ads submitted each day.
US, WA, Seattle
The Search Supply & Experiences team, within Sponsored Products, is seeking an Applied Scientist to solve challenging problems in natural language understanding, personalization, and other areas using the latest techniques in machine learning. In our team, you will have the opportunity to create new ads experiences that elevate the shopping experience for our hundreds of millions customers worldwide. As an Applied Scientist, you will partner with other talented scientists and engineers to design, train, test, and deploy machine learning models. You will be responsible for translating business and engineering requirements into deliverables, and performing detailed experiment analysis to determine how shoppers and advertisers are responding to your changes. We are looking for candidates who thrive in an exciting, fast-paced environment and who have a strong personal interest in learning, researching, and creating new technologies with high customer impact. Key job responsibilities As an Applied Scientist on the Search Supply & Experiences team you will: - Perform hands-on analysis and modeling of enormous datasets to develop insights that increase traffic monetization and merchandise sales, without compromising the shopper experience. - Drive end-to-end machine learning projects that have a high degree of ambiguity, scale, and complexity. - Build machine learning models, perform proof-of-concept, experiment, optimize, and deploy your models into production; work closely with software engineers to assist in productionizing your ML models. - Design and run experiments, gather data, and perform statistical analysis. - Establish scalable, efficient, automated processes for large-scale data analysis, machine-learning model development, model validation and serving. - Stay up to date on the latest advances in machine learning. About the team We are a customer-obsessed team of engineers, technologists, product leaders, and scientists. We are focused on continuous exploration of contexts and creatives where advertising delivers value to shoppers and advertisers. We specifically work on new ads experiences globally with the goal of helping shoppers make the most informed purchase decision. We obsess about our customers and we are continuously innovating on their behalf to enrich their shopping experience on Amazon
US, WA, Seattle
Have you ever wondered how Amazon launches and maintains a consistent customer experience across hundreds of countries and languages it serves its customers? Are you passionate about data and mathematics, and hope to impact the experience of millions of customers? Are you obsessed with designing simple algorithmic solutions to very challenging problems? If so, we look forward to hearing from you! At Amazon, we strive to be Earth's most customer-centric company, where both internal and external customers can find and discover anything they want in their own language of preference. Our Translations Services (TS) team plays a pivotal role in expanding the reach of our marketplace worldwide and enables thousands of developers and other stakeholders (Product Managers, Program Managers, Linguists) in developing locale specific solutions. Amazon Translations Services (TS) is seeking an Applied Scientist to be based in our Seattle office. As a key member of the Science and Engineering team of TS, this person will be responsible for designing algorithmic solutions based on data and mathematics for translating billions of words annually across 130+ and expanding set of locales. The successful applicant will ensure that there is minimal human touch involved in any language translation and accurate translated text is available to our worldwide customers in a streamlined and optimized manner. With access to vast amounts of data, cutting-edge technology, and a diverse community of talented individuals, you will have the opportunity to make a meaningful impact on the way customers and stakeholders engage with Amazon and our platform worldwide. Together, we will drive innovation, solve complex problems, and shape the future of e-commerce. Key job responsibilities * Apply your expertise in LLM models to design, develop, and implement scalable machine learning solutions that address complex language translation-related challenges in the eCommerce space. * Collaborate with cross-functional teams, including software engineers, data scientists, and product managers, to define project requirements, establish success metrics, and deliver high-quality solutions. * Conduct thorough data analysis to gain insights, identify patterns, and drive actionable recommendations that enhance seller performance and customer experiences across various international marketplaces. * Continuously explore and evaluate state-of-the-art modeling techniques and methodologies to improve the accuracy and efficiency of language translation-related systems. * Communicate complex technical concepts effectively to both technical and non-technical stakeholders, providing clear explanations and guidance on proposed solutions and their potential impact. About the team We are a start-up mindset team. As the long-term technical strategy is still taking shape, there is a lot of opportunity for this fresh Science team to innovate by leveraging Gen AI technoligies to build scalable solutions from scratch. Our Vision: Language will not stand in the way of anyone on earth using Amazon products and services. Our Mission: We are the enablers and guardians of translation for Amazon's customers. We do this by offering hands-off-the-wheel service to all Amazon teams, optimizing translation quality and speed at the lowest cost possible.
US, WA, Seattle
Amazon.com strives to be Earth's most customer-centric company where customers can shop in our stores to find and discover anything they want to buy. We hire the world's brightest minds, offering them a fast paced, technologically sophisticated and friendly work environment. Economists at Amazon partner closely with senior management, business stakeholders, scientist and engineers, and economist leadership to solve key business problems ranging from Amazon Web Services, Kindle, Prime, inventory planning, international retail, third party merchants, search, pricing, labor and employment planning, effective benefits (health, retirement, etc.) and beyond. Amazon Economists build econometric models using our world class data systems and apply approaches from a variety of skillsets – applied macro/time series, applied micro, econometric theory, empirical IO, empirical health, labor, public economics and related fields are all highly valued skillsets at Amazon. You will work in a fast moving environment to solve business problems as a member of either a cross-functional team embedded within a business unit or a central science and economics organization. You will be expected to develop techniques that apply econometrics to large data sets, address quantitative problems, and contribute to the design of automated systems around the company. About the team The International Seller Services (ISS) Economics team is a dynamic group at the forefront of shaping Amazon's global seller ecosystem. As part of ISS, we drive innovation and growth through sophisticated economic analysis and data-driven insights. Our mission is critical: we're transforming how Amazon empowers millions of international sellers to succeed in the digital marketplace. Our team stands at the intersection of innovative technology and practical business solutions. We're leading Amazon's transformation in seller services through work with Large Language Models (LLMs) and generative AI, while tackling fundamental questions about seller growth, marketplace dynamics, and operational efficiency. What sets us apart is our unique blend of rigorous economic methodology and practical business impact. We're not just analyzing data – we're building the frameworks and measurement systems that will define the future of Amazon's seller services. Whether we're optimizing the seller journey, evaluating new technologies, or designing innovative service models, our team transforms complex economic challenges into actionable insights that drive real-world results. Join us in shaping how millions of businesses worldwide succeed on Amazon's marketplace, while working on problems that combine economic theory, advanced analytics, and innovative technology.
GB, London
Are you looking to work at the forefront of Machine Learning and AI? Would you be excited to apply cutting edge Generative AI algorithms to solve real world problems with significant impact? The AWS Industries Team at AWS helps AWS customers implement Generative AI solutions and realize transformational business opportunities for AWS customers in the most strategic industry verticals. This is a team of data scientists, engineers, and architects working step-by-step with customers to build bespoke solutions that harness the power of generative AI. The team helps customers imagine and scope the use cases that will create the greatest value for their businesses, select and train and fine tune the right models, define paths to navigate technical or business challenges, develop proof-of-concepts, and build applications to launch these solutions at scale. The AWS Industries team provides guidance and implements best practices for applying generative AI responsibly and cost efficiently. You will work directly with customers and innovate in a fast-paced organization that contributes to game-changing projects and technologies. You will design and run experiments, research new algorithms, and find new ways of optimizing risk, profitability, and customer experience. In this Data Scientist role you will be capable of using GenAI and other techniques to design, evangelize, and implement and scale cutting-edge solutions for never-before-solved problems. Key job responsibilities - Collaborate with AI/ML scientists, engineers, and architects to research, design, develop, and evaluate cutting-edge generative AI algorithms and build ML systems to address real-world challenges - Interact with customers directly to understand the business problem, help and aid them in implementation of generative AI solutions, deliver briefing and deep dive sessions to customers and guide customer on adoption patterns and paths to production - Create and deliver best practice recommendations, tutorials, blog posts, publications, sample code, and presentations adapted to technical, business, and executive stakeholder - Provide customer and market feedback to Product and Engineering teams to help define product direction About the team Diverse Experiences Amazon values diverse experiences. Even if you do not meet all of the preferred qualifications and skills listed in the job description, we encourage candidates to apply. If your career is just starting, hasn’t followed a traditional path, or includes alternative experiences, don’t let it stop you from applying. Why AWS Amazon Web Services (AWS) is the world’s most comprehensive and broadly adopted cloud platform. We pioneered cloud computing and never stopped innovating — that’s why customers from the most successful startups to Global 500 companies trust our robust suite of products and services to power their businesses. Work/Life Balance We value work-life harmony. Achieving success at work should never come at the expense of sacrifices at home, which is why we strive for flexibility as part of our working culture. When we feel supported in the workplace and at home, there’s nothing we can’t achieve in the cloud. Inclusive Team Culture Here at AWS, it’s in our nature to learn and be curious. Our employee-led affinity groups foster a culture of inclusion that empower us to be proud of our differences. Ongoing events and learning experiences, including our Conversations on Race and Ethnicity (CORE) and AmazeCon (gender diversity) conferences, inspire us to never stop embracing our uniqueness. Mentorship and Career Growth We’re continuously raising our performance bar as we strive to become Earth’s Best Employer. That’s why you’ll find endless knowledge-sharing, mentorship and other career-advancing resources here to help you develop into a better-rounded professional.