Saeed Maleki

Zhiqi Lin, Youshan Miao, Guanbin Xu et al. · microsoft-research

Increasingly complex and diverse deep neural network (DNN) models necessitate distributing the execution across multiple devices for training and inference tasks, and also require carefully planned schedules for performance. However, existing practices often rely on predefined schedules that may not fully exploit the benefits of emerging diverse model-aware operator placement strategies. Handcrafting high-efficiency schedules can be challenging due to the large and varying schedule space. This paper presents Tessel, an automated system that searches for efficient schedules for distributed DNN training and inference for diverse operator placement strategies. To reduce search costs, Tessel leverages the insight that the most efficient schedules often exhibit repetitive pattern (repetend) across different data inputs. This leads to a two-phase approach: repetend construction and schedule completion. By exploring schedules for various operator placement strategies, Tessel significantly improves both training and inference performance. Experiments with representative DNN models demonstrate that Tessel achieves up to 5.5x training performance speedup and up to 38% inference latency reduction.

Saeed Maleki

AI models are increasing in size and recent advancement in the community has shown that unlike HPC applications where double precision datatype are required, lower-precision datatypes such as fp8 or int4 are sufficient to bring the same model quality both for training and inference. Following these trends, GPU vendors such as NVIDIA and AMD have added hardware support for fp16, fp8 and int8 GeMM operations with an exceptional performance via Tensor Cores. However, this paper proposes a new algorithm called msGeMM which shows that AI models with low-precision datatypes can run with ~2.5x fewer multiplication and add instructions. Efficient implementation of this algorithm requires special CUDA cores with the ability to add elements from a small look-up table at the rate of Tensor Cores.

Saeed Maleki, Madan Musuvathi, Todd Mytkowicz et al.

Stochastic gradient descent (SGD) is an inherently sequential training algorithm--computing the gradient at batch $i$ depends on the model parameters learned from batch $i-1$. Prior approaches that break this dependence do not honor them (e.g., sum the gradients for each batch, which is not what sequential SGD would do) and thus potentially suffer from poor convergence. This paper introduces a novel method to combine gradients called Adasum (for adaptive sum) that converges faster than prior work. Adasum is easy to implement, almost as efficient as simply summing gradients, and is integrated into the open-source toolkit Horovod. This paper first provides a formal justification for Adasum and then empirically demonstrates Adasum is more accurate than prior gradient accumulation methods. It then introduces a series of case-studies to show Adasum works with multiple frameworks, (TensorFlow and PyTorch), scales multiple optimizers (Momentum-SGD, Adam, and LAMB) to larger batch-sizes while still giving good downstream accuracy. Finally, it proves that Adasum converges. To summarize, Adasum scales Momentum-SGD on the MLPerf Resnet50 benchmark to 64K examples before communication (no MLPerf v0.5 entry converged with more than 16K), the Adam optimizer to 64K examples before communication on BERT-LARGE (prior work showed Adam stopped scaling at 16K), and the LAMB optimizer to 128K before communication on BERT-LARGE (prior work used 64K), all while maintaining downstream accuracy metrics. Finally, if a user does not need to scale, we show LAMB with Adasum on BERT-LARGE converges in 30% fewer steps than the baseline.

Saeed Maleki, Yang Cheng, John Crassidis et al.

This work provides a theoretical framework for the pose estimation problem using total least squares for vector observations from landmark features. First, the optimization framework is formulated with observation vectors extracted from point cloud features. Then, error-covariance expressions are derived. The attitude and position solutions obtained via the derived optimization framework are proven to reach the bounds defined by the Cramér-Rao lower bound under the small-angle approximation of attitude errors. The measurement data for the simulation of this problem is provided through a series of vector observation scans, and a fully populated observation noise-covariance matrix is assumed as the weight in the cost function to cover the most general case of the sensor uncertainty. Here, previous derivations are expanded for the pose estimation problem to include more generic correlations in the errors than previous cases involving an isotropic noise assumption. The proposed solution is simulated in a Monte-Carlo framework to validate the error-covariance analysis.

Gurbinder Gill, Roshan Dathathri, Saeed Maleki et al.

Many applications today, such as NLP, network analysis, and code analysis, rely on semantically embedding objects into low-dimensional fixed-length vectors. Such embeddings naturally provide a way to perform useful downstream tasks, such as identifying relations among objects or predicting objects for a given context, etc. Unfortunately, the training necessary for accurate embeddings is usually computationally intensive and requires processing large amounts of data. Furthermore, distributing this training is challenging. Most embedding training uses stochastic gradient descent (SGD), an "inherently" sequential algorithm. Prior approaches to parallelizing SGD do not honor these dependencies and thus potentially suffer poor convergence. This paper presents a distributed training framework for a class of applications that use Skip-gram-like models to generate embeddings. We call this class Any2Vec and it includes Word2Vec, DeepWalk, and Node2Vec among others. We first formulate Any2Vec training algorithm as a graph application and leverage the state-of-the-art distributed graph analytics framework, D-Galois. We adapt D-Galois to support dynamic graph generation and repartitioning, and incorporate novel communication optimizations. Finally, we introduce a novel way to combine gradients during distributed training to prevent accuracy loss. We show that our framework, called GraphAny2Vec, matches on a cluster of 32 hosts the accuracy of the state-of-the-art shared-memory implementations of Word2Vec and Vertex2Vec on 1 host, and gives a geo-mean speedup of 12x and 5x respectively. Furthermore, GraphAny2Vec is on average 2x faster than the state-of-the-art distributed Word2Vec implementation, DMTK, on 32 hosts. We also show the superiority of our Gradient Combiner independent of GraphAny2Vec by incorporating it in DMTK, which raises its accuracy by > 30%.

Roshan Dathathri, Olli Saarikivi, Hao Chen et al.

Fully Homomorphic Encryption (FHE) refers to a set of encryption schemes that allow computations to be applied directly on encrypted data without requiring a secret key. This enables novel application scenarios where a client can safely offload storage and computation to a third-party cloud provider without having to trust the software and the hardware vendors with the decryption keys. Recent advances in both FHE schemes and implementations have moved such applications from theoretical possibilities into the realm of practicalities. This paper proposes a compact and well-reasoned interface called the Homomorphic Instruction Set Architecture (HISA) for developing FHE applications. Just as the hardware ISA interface enabled hardware advances to proceed independent of software advances in the compiler and language runtimes, HISA decouples compiler optimizations and runtimes for supporting FHE applications from advancements in the underlying FHE schemes. This paper demonstrates the capabilities of HISA by building an end-to-end software stack for evaluating neural network models on encrypted data. Our stack includes an end-to-end compiler, runtime, and a set of optimizations. Our approach shows generated code, on a set of popular neural network architectures, is faster than hand-optimized implementations.

Saeed Maleki, Madanlal Musuvathi, Todd Mytkowicz

Stochastic gradient descent (SGD) is a well known method for regression and classification tasks. However, it is an inherently sequential algorithm at each step, the processing of the current example depends on the parameters learned from the previous examples. Prior approaches to parallelizing linear learners using SGD, such as HOGWILD! and ALLREDUCE, do not honor these dependencies across threads and thus can potentially suffer poor convergence rates and/or poor scalability. This paper proposes SYMSGD, a parallel SGD algorithm that, to a first-order approximation, retains the sequential semantics of SGD. Each thread learns a local model in addition to a model combiner, which allows local models to be combined to produce the same result as what a sequential SGD would have produced. This paper evaluates SYMSGD's accuracy and performance on 6 datasets on a shared-memory machine shows upto 11x speedup over our heavily optimized sequential baseline on 16 cores and 2.2x, on average, faster than HOGWILD!.