Blake A. Hechtman

Felix Chern, Blake Hechtman, Andy Davis et al.

This paper presents a novel nearest neighbor search algorithm achieving TPU (Google Tensor Processing Unit) peak performance, outperforming state-of-the-art GPU algorithms with similar level of recall. The design of the proposed algorithm is motivated by an accurate accelerator performance model that takes into account both the memory and instruction bottlenecks. Our algorithm comes with an analytical guarantee of recall in expectation and does not require maintaining sophisticated index data structure or tuning, making it suitable for applications with frequent updates. Our work is available in the open-source package of Jax and Tensorflow on TPU.

Jevin Jiang, Ying Chen, Blake A. Hechtman et al.

Large Language Model (LLM) deployment is increasingly shifting to cost-efficient accelerators like Google's Tensor Processing Units (TPUs), prioritizing both performance and total cost of ownership (TCO). However, existing LLM inference kernels and serving systems remain largely GPU-centric, and there is no well-established approach for efficiently mapping LLM workloads onto TPU architectures--particularly under the dynamic and ragged execution patterns common in modern serving. In this paper, we present Ragged Paged Attention (RPA), a high-performance and flexible attention kernel for TPUs, implemented using Pallas and Mosaic. RPA addresses these challenges through three key techniques: (1) fine-grained tiling to enable efficient dynamic slicing over ragged memory, (2) a custom software pipeline that fuses KV cache updates with attention computation, and (3) a distribution-aware compilation strategy that generates specialized kernels for decode, prefill, and mixed workloads. Evaluated on Llama 3 8B on TPU7x, RPA achieves up to 86% memory bandwidth utilization (MBU) in decode and 73% model FLOPs utilization (MFU) in prefill. Integrated as the primary TPU backend in vLLM and SGLang, RPA provides a production-grade foundation for efficient TPU inference and offers practical insights into kernel design.

Noam Shazeer, Youlong Cheng, Niki Parmar et al.

Batch-splitting (data-parallelism) is the dominant distributed Deep Neural Network (DNN) training strategy, due to its universal applicability and its amenability to Single-Program-Multiple-Data (SPMD) programming. However, batch-splitting suffers from problems including the inability to train very large models (due to memory constraints), high latency, and inefficiency at small batch sizes. All of these can be solved by more general distribution strategies (model-parallelism). Unfortunately, efficient model-parallel algorithms tend to be complicated to discover, describe, and to implement, particularly on large clusters. We introduce Mesh-TensorFlow, a language for specifying a general class of distributed tensor computations. Where data-parallelism can be viewed as splitting tensors and operations along the "batch" dimension, in Mesh-TensorFlow, the user can specify any tensor-dimensions to be split across any dimensions of a multi-dimensional mesh of processors. A Mesh-TensorFlow graph compiles into a SPMD program consisting of parallel operations coupled with collective communication primitives such as Allreduce. We use Mesh-TensorFlow to implement an efficient data-parallel, model-parallel version of the Transformer sequence-to-sequence model. Using TPU meshes of up to 512 cores, we train Transformer models with up to 5 billion parameters, surpassing state of the art results on WMT'14 English-to-French translation task and the one-billion-word language modeling benchmark. Mesh-Tensorflow is available at https://github.com/tensorflow/mesh .

Aditya Agrawal, Matthew Hedlund, Blake Hechtman

eXmY is a novel data type for quantization of ML models. It supports both arbitrary bit widths and arbitrary integer and floating point formats. For example, it seamlessly supports 3, 5, 6, 7, 9 bit formats. For a specific bit width, say 7, it defines all possible formats e.g. e0m6, e1m5, e2m4, e3m3, e4m2, e5m1 and e6m0. For non-power of two bit widths e.g. 5, 6, 7, we created a novel encoding and decoding scheme which achieves perfect compression, byte addressability and is amenable to sharding and vector processing. We implemented libraries for emulation, encoding and decoding tensors and checkpoints in C++, TensorFlow, JAX and PAX. For optimal performance, the codecs use SIMD instructions on CPUs and vector instructions on TPUs and GPUs. eXmY is also a technique and exploits the statistical distribution of exponents in tensors. It can be used to quantize weights, static and dynamic activations, gradients, master weights and optimizer state. It can reduce memory (CPU DRAM and accelerator HBM), network and disk storage and transfers. It can increase multi tenancy and accelerate compute. eXmY has been deployed in production for almost 2 years.

Aidan Clark, Diego de las Casas, Aurelia Guy et al.

The performance of a language model has been shown to be effectively modeled as a power-law in its parameter count. Here we study the scaling behaviors of Routing Networks: architectures that conditionally use only a subset of their parameters while processing an input. For these models, parameter count and computational requirement form two independent axes along which an increase leads to better performance. In this work we derive and justify scaling laws defined on these two variables which generalize those known for standard language models and describe the performance of a wide range of routing architectures trained via three different techniques. Afterwards we provide two applications of these laws: first deriving an Effective Parameter Count along which all models scale at the same rate, and then using the scaling coefficients to give a quantitative comparison of the three routing techniques considered. Our analysis derives from an extensive evaluation of Routing Networks across five orders of magnitude of size, including models with hundreds of experts and hundreds of billions of parameters.

Jack W. Rae, Sebastian Borgeaud, Trevor Cai et al.

Language modelling provides a step towards intelligent communication systems by harnessing large repositories of written human knowledge to better predict and understand the world. In this paper, we present an analysis of Transformer-based language model performance across a wide range of model scales -- from models with tens of millions of parameters up to a 280 billion parameter model called Gopher. These models are evaluated on 152 diverse tasks, achieving state-of-the-art performance across the majority. Gains from scale are largest in areas such as reading comprehension, fact-checking, and the identification of toxic language, but logical and mathematical reasoning see less benefit. We provide a holistic analysis of the training dataset and model's behaviour, covering the intersection of model scale with bias and toxicity. Finally we discuss the application of language models to AI safety and the mitigation of downstream harms.

Yuanzhong Xu, HyoukJoong Lee, Dehao Chen et al.

We present GSPMD, an automatic, compiler-based parallelization system for common machine learning computations. It allows users to write programs in the same way as for a single device, then give hints through a few annotations on how to distribute tensors, based on which GSPMD will parallelize the computation. Its representation of partitioning is simple yet general, allowing it to express different or mixed paradigms of parallelism on a wide variety of models. GSPMD infers the partitioning for every operator based on limited user annotations, making it convenient to scale existing single-device programs. It solves several technical challenges for production usage, allowing GSPMD to achieve 50% to 62% compute utilization on up to 2048 Cloud TPUv3 cores for models with up to one trillion parameters.

Ashish Vaswani, Prajit Ramachandran, Aravind Srinivas et al.

Self-attention has the promise of improving computer vision systems due to parameter-independent scaling of receptive fields and content-dependent interactions, in contrast to parameter-dependent scaling and content-independent interactions of convolutions. Self-attention models have recently been shown to have encouraging improvements on accuracy-parameter trade-offs compared to baseline convolutional models such as ResNet-50. In this work, we aim to develop self-attention models that can outperform not just the canonical baseline models, but even the high-performing convolutional models. We propose two extensions to self-attention that, in conjunction with a more efficient implementation of self-attention, improve the speed, memory usage, and accuracy of these models. We leverage these improvements to develop a new self-attention model family, HaloNets, which reach state-of-the-art accuracies on the parameter-limited setting of the ImageNet classification benchmark. In preliminary transfer learning experiments, we find that HaloNet models outperform much larger models and have better inference performance. On harder tasks such as object detection and instance segmentation, our simple local self-attention and convolutional hybrids show improvements over very strong baselines. These results mark another step in demonstrating the efficacy of self-attention models on settings traditionally dominated by convolutional models.

Sameer Kumar, James Bradbury, Cliff Young et al.

Recent results in language understanding using neural networks have required training hardware of unprecedentedscale, with thousands of chips cooperating on a single training run. This paper presents techniques to scaleML models on the Google TPU Multipod, a mesh with 4096 TPU-v3 chips. We discuss model parallelism toovercome scaling limitations from the fixed batch size in data parallelism, communication/collective optimizations,distributed evaluation of training metrics, and host input processing scaling optimizations. These techniques aredemonstrated in both the TensorFlow and JAX programming frameworks. We also present performance resultsfrom the recent Google submission to the MLPerf-v0.7 benchmark contest, achieving record training times from16 to 28 seconds in four MLPerf models on the Google TPU-v3 Multipod machine.

Yuanzhong Xu, HyoukJoong Lee, Dehao Chen et al.

In data-parallel synchronous training of deep neural networks, different devices (replicas) run the same program with different partitions of the training batch, but weight update computation is repeated on all replicas, because the weights do not have a batch dimension to partition. This can be a bottleneck for performance and scalability in typical language models with large weights, and models with small per-replica batch size which is typical in large-scale training. This paper presents an approach to automatically shard the weight update computation across replicas with efficient communication primitives and data formatting, using static analysis and transformations on the training computation graph. We show this technique achieves substantial speedups on typical image and language models on Cloud TPUs, requiring no change to model code. This technique helps close the gap between traditionally expensive (ADAM) and cheap (SGD) optimizers, as they will only take a small part of training step time and have similar peak memory usage. It helped us to achieve state-of-the-art training performance in Google's MLPerf 0.6 submission.

Sameer Kumar, Victor Bitorff, Dehao Chen et al.

The recent submission of Google TPU-v3 Pods to the industry wide MLPerf v0.6 training benchmark demonstrates the scalability of a suite of industry relevant ML models. MLPerf defines a suite of models, datasets and rules to follow when benchmarking to ensure results are comparable across hardware, frameworks and companies. Using this suite of models, we discuss the optimizations and techniques including choice of optimizer, spatial partitioning and weight update sharding necessary to scale to 1024 TPU chips. Furthermore, we identify properties of models that make scaling them challenging, such as limited data parallelism and unscaled weights. These optimizations contribute to record performance in transformer, Resnet-50 and SSD in the Google MLPerf-0.6 submission.