Abhinav Bhatele

Daniel Nichols, Aniruddha Marathe, Harshitha Menon et al.

Parallel programs in high performance computing (HPC) continue to grow in complexity and scale in the exascale era. The diversity in hardware and parallel programming models make developing, optimizing, and maintaining parallel software even more burdensome for developers. One way to alleviate some of these burdens is with automated development and analysis tools. Such tools can perform complex and/or remedial tasks for developers that increase their productivity and decrease the chance for error. Until recently, such tools for code development and performance analysis have been limited in the complexity of tasks they can perform, especially for parallel programs. However, with recent advancements in language modeling, and the availability of large amounts of open-source code related data, these tools have started to utilize predictive language models to automate more complex tasks. In this paper, we show how large language models (LLMs) can be applied to tasks specific to high performance and scientific codes. We introduce a new dataset of HPC and scientific codes and use it to fine-tune several pre-trained models. We compare several pre-trained LLMs on HPC-related tasks and introduce a new model, HPC-Coder, fine-tuned on parallel codes. In our experiments, we show that this model can auto-complete HPC functions where generic models cannot, decorate for loops with OpenMP pragmas, and model performance changes in scientific application repositories as well as programming competition solutions.

Vivan Madan, Prajwal Singhania, Abhinav Bhatele et al.

Mixture-of-Experts (MoE) models have gained popularity as a means of scaling the capacity of large language models (LLMs) while maintaining sparse activations and reduced per-token compute. However, in memory-constrained inference settings, expert weights must be offloaded to CPU, creating a performance bottleneck from CPU-GPU transfers during decoding. We propose an expert prefetching scheme that leverages currently computed internal model representations to speculate future experts, enabling memory transfers to overlap with computation. Across multiple MoE architectures, we demonstrate that future experts can be reliably predicted by these internal representations. We also demonstrate that executing speculated experts generally maintains downstream task accuracy, thus preserving more effective compute-memory overlap by eliminating the need to re-fetch true router-selected experts. Integrated into an optimized inference engine, our approach achieves up to 14\% reduction in time per output token (TPOT) over on-demand loading of experts from CPU memory. For MoEs where speculative execution alone yields suboptimal accuracy, we further examine lightweight estimators that improve expert prediction hit rates, thereby reducing performance degradation. Our code is released in open-source at https://github.com/axonn-ai/yalis/tree/offload_prefetch.

Joshua H. Davis, Klaudiusz Rydzy, Srinivasan Ramesh et al.

Performance profiles of GPU kernels generated by tools such as Nsight Compute are rich in detail but are often challenging to interpret. To achieve the best performance possible on a given GPU architecture, kernel developers need to spend significant time analyzing and comparing profiles in the tool's graphical interface to identify and understand kernel performance bottlenecks. Large Language Models (LLMs) have shown promise in understanding complex data and generating natural language explanations. In this paper, we propose the Kernel Execution Explanation Toolkit (KEET), an LLM-based agentic framework for interpreting Nsight Compute profiles to generate useful and data-grounded natural language explanations of performance issues in GPU kernels, and suggestions for optimizations. We evaluate \toolname using several CUDA kernels of varying complexity on NVIDIA H100 GPUs. We find that the generated explanations, when provided as context, improve the quality of LLM code optimization and multiple-choice question answering in downstream tasks. We further demonstrate that the tool can be used to interpret performance data from large sets of profiles to improve the quality of optimization suggestions.

Siddharth Singh, Olatunji Ruwase, Ammar Ahmad Awan et al.

Mixture-of-Experts (MoE) is a neural network architecture that adds sparsely activated expert blocks to a base model, increasing the number of parameters without impacting computational costs. However, current distributed deep learning frameworks are limited in their ability to train high-quality MoE models with large base models. In this work, we present DeepSpeed-TED, a novel, three-dimensional, hybrid parallel algorithm that combines data, tensor, and expert parallelism to enable the training of MoE models with 4 to 8x larger base models than the current state-of-the-art. We also describe memory optimizations in the optimizer step, and communication optimizations that eliminate unnecessary data movement. We implement our approach in DeepSpeed and achieve speedups of 26% over a baseline (i.e. without our communication optimizations) when training a 40 billion parameter MoE model (6.7 billion base model with 16 experts) on 128 V100 GPUs.

Siddharth Singh, Abhinav Bhatele

Parallel training of neural networks at scale is challenging due to significant overheads arising from communication. Recently, deep learning researchers have developed a variety of pruning algorithms that are capable of pruning (i.e. setting to zero) 80-90% of the parameters in a neural network to yield sparse subnetworks that equal the accuracy of the unpruned parent network. In this work, we propose a novel approach that exploits these sparse subnetworks to optimize the memory utilization and communication in two popular algorithms for parallel deep learning namely -- data and inter-layer parallelism. We integrate our approach into AxoNN, a highly scalable framework for parallel deep learning that relies on data and inter-layer parallelism, and demonstrate the reduction in communication time and memory utilization. On 512 NVIDIA V100 GPUs, our optimizations reduce the memory consumption of a 2.7 billion parameter model by 74%, and the total communication time by 40%, thus providing an overall speedup of 34% over AxoNN, 32% over DeepSpeed-3D and 46% over Sputnik, a sparse matrix computation baseline.

Cunyang Wei, Siddharth Singh, Aishwarya Sarkar et al.

Graph neural networks (GNNs) are widely used for learning on graph datasets derived from various real-world scenarios. Learning from extremely large graphs requires distributed training, and mini-batching with sampling is a popular approach for parallelizing GNN training. Existing distributed mini-batch approaches have significant performance bottlenecks due to expensive sampling methods and limited scaling when using data parallelism. In this work, we present ScaleGNN, a 4D parallel framework for scalable mini-batch GNN training that combines communication-free distributed sampling, 3D parallel matrix multiplication (PMM), and data parallelism. ScaleGNN introduces a uniform vertex sampling algorithm, enabling each process (GPU device) to construct its local mini-batch, i.e., subgraph partitions without any inter-process communication. 3D PMM enables scaling mini-batch training to much larger GPU counts than vanilla data parallelism with significantly lower communication overheads. We also present additional optimizations to overlap sampling with training, reduce communication overhead by sending data in lower precision, kernel fusion, and communication-computation overlap. We evaluate ScaleGNN on five graph datasets and demonstrate strong scaling up to 2048 GPUs on Perlmutter, 2048 GCDs on Frontier, and 1024 GPUs on Tuolumne. On Perlmutter, ScaleGNN achieves 3.5x end-to-end training speedup over the SOTA baseline on ogbn-products.

Siddharth Singh, Zachary Sating, Abhinav Bhatele

Despite their better convergence properties compared to first-order optimizers, second-order optimizers for deep learning have been less popular due to their significant computational costs. The primary efficiency bottleneck in such optimizers is matrix inverse calculations in the preconditioning step, which are expensive to compute on GPUs. In this paper, we introduce Jorge, a second-order optimizer that promises the best of both worlds -- rapid convergence benefits of second-order methods, and high computational efficiency typical of first-order methods. We address the primary computational bottleneck of computing matrix inverses by completely eliminating them using an approximation of the preconditioner computation. This makes Jorge extremely efficient on GPUs in terms of wall-clock time. Further, we describe an approach to determine Jorge's hyperparameters directly from a well-tuned SGD baseline, thereby significantly minimizing tuning efforts. Our empirical evaluations demonstrate the distinct advantages of using Jorge, outperforming state-of-the-art optimizers such as SGD, AdamW, and Shampoo across multiple deep learning models, both in terms of sample efficiency and wall-clock time.

Daniel Nichols, Prajwal Singhania, Charles Jekel et al.

Language models (LMs) are becoming increasingly dependent on external tools. LM-based agentic frameworks frequently interact with their environment via such tools to search files, run code, call APIs, etc. Further, modern reasoning-based LMs use tools such as web search and Python code execution to enhance their reasoning capabilities. While tools greatly improve the capabilities of LMs, they also introduce performance bottlenecks during the inference process. In this paper, we introduce novel systems optimizations to address such performance bottlenecks by speculating tool calls and forcing sequences to remain resident in the inference engine to minimize overheads. Our optimizations lead to throughput improvements of several hundred tokens per second when hosting inference for LM agents. We provide a theoretical analysis of our algorithms to provide insights into speculation configurations that will yield the best performance. Further, we recommend a new "tool cache" API endpoint to enable LM providers to easily adopt these optimizations.

Prajwal Singhania, Siddharth Singh, Lannie Dalton Hough et al.

As large language models (LLMs) continue to grow in size, distributed inference has become increasingly important. Model-parallel strategies must now efficiently scale not only across multiple GPUs but also across multiple nodes. In this work, we present a detailed performance study of multi-node distributed inference using LLMs on GPU-based supercomputers. We conduct experiments with several state-of-the-art inference engines alongside YALIS, a research-oriented prototype engine designed for controlled experimentation. We analyze the strong-scaling behavior of different model-parallel schemes and identify key bottlenecks. Since all-reduce operations are a common performance bottleneck, we develop NVRAR, a hierarchical all-reduce algorithm based on recursive doubling with NVSHMEM. NVRAR achieves up to 1.9x-3.6x lower latency than NCCL for message sizes between 128 KB and 2 MB on HPE Slingshot and InfiniBand interconnects. Integrated into YALIS, NVRAR achieves up to a 1.72x reduction in end-to-end batch latency for the Llama 3.1 405B model in multi-node decode-heavy workloads using tensor parallelism.

Sean McLeish, John Kirchenbauer, David Yu Miller et al.

Scaling laws are typically fit using a family of models with a narrow range of frozen hyperparameter choices. In this work we study scaling laws using multiple architectural shapes and hyperparameter choices, highlighting their impact on resulting prescriptions. As a primary artifact of our research, we release the Gemstones: an open-source scaling law dataset, consisting of over 4000 checkpoints from transformers with up to 2 billion parameters and diverse architectural shapes; including ablations over learning rate and cooldown. Our checkpoints enable more complex studies of scaling, such as analyzing the relationship between width and depth. By examining our model suite, we find that the prescriptions of scaling laws can be highly sensitive to the experimental design process and the specific model checkpoints used during fitting.

Siddharth Singh, Prajwal Singhania, Aditya Ranjan et al.

Training and fine-tuning large language models (LLMs) with hundreds of billions to trillions of parameters requires tens of thousands of GPUs, and a highly scalable software stack. In this work, we present a novel four-dimensional hybrid parallel algorithm implemented in a highly scalable, portable, open-source framework called AxoNN. We describe several performance optimizations in AxoNN to improve matrix multiply kernel performance, overlap non-blocking collectives with computation, and performance modeling to choose performance optimal configurations. These have resulted in unprecedented scaling and peak flop/s (bf16) for training of GPT-style transformer models on Perlmutter (620.1 Petaflop/s), Frontier (1.381 Exaflop/s) and Alps (1.423 Exaflop/s). While the abilities of LLMs improve with the number of trainable parameters, so do privacy and copyright risks caused by memorization of training data, which can cause disclosure of sensitive or private information at inference time. We highlight this side effect of scale through experiments that explore "catastrophic memorization", where models are sufficiently large to memorize training data in a single pass, and present an approach to prevent it. As part of this study, we demonstrate fine-tuning of a 405-billion parameter LLM using AxoNN on Frontier.

Aman Chaturvedi, Daniel Nichols, Siddharth Singh et al.

Large Language Model (LLM) based coding tools have been tremendously successful as software development assistants, yet they are often designed for general purpose programming tasks and perform poorly for more specialized domains such as high performance computing. Creating specialized models and tools for these domains is crucial towards gaining the benefits of LLMs in areas such as HPC. While previous work has explored HPC-specific models, LLMs still struggle to generate parallel code and it is not at all clear what hurdles are still holding back these LLMs and what must be done to overcome them. In this work, we conduct an in-depth study along the many axes of fine-tuning a specialized HPC LLM in order to better understand the challenges. Based on our findings we fine-tune and evaluate a specialized HPC LLM that is shown to be the best performing open-source code LLM for parallel code generation to date.

Jonas Geiping, Sean McLeish, Neel Jain et al.

We study a novel language model architecture that is capable of scaling test-time computation by implicitly reasoning in latent space. Our model works by iterating a recurrent block, thereby unrolling to arbitrary depth at test-time. This stands in contrast to mainstream reasoning models that scale up compute by producing more tokens. Unlike approaches based on chain-of-thought, our approach does not require any specialized training data, can work with small context windows, and can capture types of reasoning that are not easily represented in words. We scale a proof-of-concept model to 3.5 billion parameters and 800 billion tokens. We show that the resulting model can improve its performance on reasoning benchmarks, sometimes dramatically, up to a computation load equivalent to 50 billion parameters.

Kaiyu Yue, Vasu Singla, Menglin Jia et al.

Vision language models (VLMs) typically pair a modestly sized vision encoder with a large language model (LLM), e.g., Llama-70B, making the decoder the primary computational burden during training. To reduce costs, a potential promising strategy is to first train the vision encoder using a small language model before transferring it to the large one. We construct small "surrogate models" that share the same embedding space and representation language as the large target LLM by directly inheriting its shallow layers. Vision encoders trained on the surrogate can then be directly transferred to the larger model, a process we call zero-shot grafting -- when plugged directly into the full-size target LLM, the grafted pair surpasses the encoder-surrogate pair and, on some benchmarks, even performs on par with full decoder training with the target LLM. Furthermore, our surrogate training approach reduces overall VLM training costs by ~45% when using Llama-70B as the decoder. The code is at https://github.com/facebookresearch/zero.

Vasu Singla, Kaiyu Yue, Sukriti Paul et al.

Training large vision-language models requires extensive, high-quality image-text pairs. Existing web-scraped datasets, however, are noisy and lack detailed image descriptions. To bridge this gap, we introduce PixelProse, a comprehensive dataset of over 16M (million) synthetically generated captions, leveraging cutting-edge vision-language models for detailed and accurate descriptions. To ensure data integrity, we rigorously analyze our dataset for problematic content, including child sexual abuse material (CSAM), personally identifiable information (PII), and toxicity. We also provide valuable metadata such as watermark presence and aesthetic scores, aiding in further dataset filtering. We hope PixelProse will be a valuable resource for future vision-language research. PixelProse is available at https://huggingface.co/datasets/tomg-group-umd/pixelprose

Daniel Nichols, Joshua H. Davis, Zhaojun Xie et al.

Large language models are increasingly becoming a popular tool for software development. Their ability to model and generate source code has been demonstrated in a variety of contexts, including code completion, summarization, translation, and lookup. However, they often struggle to generate code for complex programs. In this paper, we study the capabilities of state-of-the-art language models to generate parallel code. In order to evaluate language models, we create a benchmark, ParEval, consisting of prompts that represent 420 different coding tasks related to scientific and parallel computing. We use ParEval to evaluate the effectiveness of several state-of-the-art open- and closed-source language models on these tasks. We introduce novel metrics for evaluating the performance of generated code, and use them to explore how well each large language model performs for 12 different computational problem types and six different parallel programming models.

Daniel Nichols, Pranav Polasam, Harshitha Menon et al.

Optimizing scientific software is a difficult task because codebases are often large and complex, and performance can depend upon several factors including the algorithm, its implementation, and hardware among others. Causes of poor performance can originate from disparate sources and be difficult to diagnose. Recent years have seen a multitude of work that use large language models (LLMs) to assist in software development tasks. However, these tools are trained to model the distribution of code as text, and are not specifically designed to understand performance aspects of code. In this work, we introduce a reinforcement learning based methodology to align the outputs of code LLMs with performance. This allows us to build upon the current code modeling capabilities of LLMs and extend them to generate better performing code. We demonstrate that our fine-tuned model improves the expected speedup of generated code over base models for a set of benchmark tasks from 0.9 to 1.6 for serial code and 1.9 to 4.5 for OpenMP code.

Keita Teranishi, Harshitha Menon, William F. Godoy et al.

We discuss the challenges and propose research directions for using AI to revolutionize the development of high-performance computing (HPC) software. AI technologies, in particular large language models, have transformed every aspect of software development. For its part, HPC software is recognized as a highly specialized scientific field of its own. We discuss the challenges associated with leveraging state-of-the-art AI technologies to develop such a unique and niche class of software and outline our research directions in the two US Department of Energy--funded projects for advancing HPC Software via AI: Ellora and Durban.

Siddharth Singh, Mahua Singh, Abhinav Bhatele

We evaluate the current state of collective communication on GPU-based supercomputers for large language model (LLM) training at scale. Existing libraries such as RCCL and Cray-MPICH exhibit critical limitations on systems such as Frontier -- Cray-MPICH underutilizes network and compute resources, while RCCL suffers from severe scalability issues. To address these challenges, we introduce PCCL, a communication library with highly optimized implementations of all-gather and reduce-scatter operations tailored for distributed deep learning workloads. PCCL is designed to maximally utilize all available network and compute resources and to scale efficiently to thousands of GPUs. It achieves substantial performance improvements, delivering 6-33x speedups over RCCL and 28-70x over Cray-MPICH for all-gather on 2048 GCDs of Frontier. These gains translate directly to end-to-end performance: in large-scale GPT-3-style training, PCCL provides up to 60% and 40% speedups over RCCL for 7B and 13B parameter models, respectively.

Daniel Nichols, Konstantinos Parasyris, Harshitha Menon et al.

Code LLMs have become extremely popular recently for modeling source code across a variety of tasks, such as generation, translation, and summarization. However, transformer-based models are limited in their capabilities to reason through structured, analytical properties of code, such as control and data flow. Previous work has explored the modeling of these properties with structured data and graph neural networks. However, these approaches lack the generative capabilities and scale of modern LLMs. In this work, we introduce a novel approach to combine the strengths of modeling both code as text and more structured forms.

Nirav Koley, Prajwal Singhania, Abhinav Bhatele

Efficient inference on GPUs using large language models remains challenging due to memory bandwidth limitations, particularly during data transfers between High Bandwidth Memory (HBM) and SRAM in attention computations. Approximate attention methods address this issue by reducing computational and memory overhead but often rely on expensive top-$k$ operations, which perform poorly on GPUs. We propose SiftAttention, a novel approximate attention method that replaces the top-$k$ step with a computationally efficient element-wise filtering operation based on a threshold value. Our intuition for doing this is based on our empirical observation that the $τ$-th quantile of attention scores follows a predictable power-law over sequential generation steps. Exploiting this insight, our approach dynamically estimates a threshold value per prompt at each generation step. Only attention scores above this threshold and their corresponding value vectors are loaded/used to compute the attention output, reducing data movement between HBM and SRAM. Our evaluation demonstrates that SiftAttention preserves model quality better than existing approximate attention methods while reducing memory bandwidth usage when loading value vectors.

Aditya K. Ranjan, Siddharth Singh, Cunyang Wei et al.

Graph neural networks (GNNs) leverage the connectivity and structure of real-world graphs to learn intricate properties and relationships between nodes. Many real-world graphs exceed the memory capacity of a GPU due to their sheer size, and training GNNs on such graphs requires techniques such as mini-batch sampling to scale. The alternative approach of distributed full-graph training suffers from high communication overheads and load imbalance due to the irregular structure of graphs. We propose a three-dimensional (3D) parallel approach for full-graph training that tackles these issues and scales to billion-edge graphs. In addition, we introduce optimizations such as a double permutation scheme for load balancing, and a performance model to predict the optimal 3D configuration of our parallel implementation -- Plexus. We evaluate Plexus on six different graph datasets and show scaling results on up to 2048 GPUs of Perlmutter, and 1024 GPUs of Frontier. Plexus achieves unprecedented speedups of 2.3-12.5x over prior state of the art, and a reduction in time-to-solution by 5.2-8.7x on Perlmutter and 7.0-54.2x on Frontier.

Abhimanyu Hans, Yuxin Wen, Neel Jain et al.

Large language models can memorize and repeat their training data, causing privacy and copyright risks. To mitigate memorization, we introduce a subtle modification to the next-token training objective that we call the goldfish loss. During training, randomly sampled subsets of tokens are excluded from the loss computation. These dropped tokens are not memorized by the model, which prevents verbatim reproduction of a complete chain of tokens from the training set. We run extensive experiments training billion-scale Llama-2 models, both pre-trained and trained from scratch, and demonstrate significant reductions in extractable memorization with little to no impact on downstream benchmarks.

Prajwal Singhania, Siddharth Singh, Shwai He et al.

Inference on large language models (LLMs) can be expensive in terms of the compute and memory costs involved, especially when long sequence lengths are used. In particular, the self-attention mechanism used in LLM inference contributes significantly to these costs, which has sparked an interest in approximating the self-attention computation to reduce such costs. In this work, we propose to approximate self-attention by focusing on the dimensionality of key vectors computed in the attention block. Our analysis reveals that key vectors lie in a significantly lower-dimensional space, consistently across several datasets and models. Exploiting this observation, we propose Loki, a novel sparse attention method that ranks and selects tokens in the KV-cache based on attention scores computed in low-dimensional space. Our evaluations show that Loki is able to speed up the attention computation due to reduced data movement (load/store) and compute costs while maintaining the efficacy of the models better than other popular approximation methods.

Siddharth Singh, Prajwal Singhania, Aditya K. Ranjan et al.

Heavy communication, in particular, collective operations, can become a critical performance bottleneck in scaling the training of billion-parameter neural networks to large-scale parallel systems. This paper introduces a four-dimensional (4D) approach to optimize communication in parallel training. This 4D approach is a hybrid of 3D tensor and data parallelism, and is implemented in the AxoNN framework. In addition, we employ two key strategies to further minimize communication overheads. First, we aggressively overlap expensive collective operations (reduce-scatter, all-gather, and all-reduce) with computation. Second, we develop an analytical model to identify high-performing configurations within the large search space defined by our 4D algorithm. This model empowers practitioners by simplifying the tuning process for their specific training workloads. When training an 80-billion parameter GPT on 1024 GPUs of Perlmutter, AxoNN surpasses Megatron-LM, a state-of-the-art framework, by a significant 26%. Additionally, it achieves a significantly high 57% of the theoretical peak FLOP/s or 182 PFLOP/s in total.

Daniel Nichols, Siddharth Singh, Shu-Huai Lin et al.

The field of deep learning has witnessed a remarkable shift towards extremely compute- and memory-intensive neural networks. These newer larger models have enabled researchers to advance state-of-the-art tools across a variety of fields. This phenomenon has spurred the development of algorithms for distributed training of neural networks over a larger number of hardware accelerators. In this paper, we discuss and compare current state-of-the-art frameworks for large scale distributed deep learning. First, we survey current practices in distributed learning and identify the different types of parallelism used. Then, we present empirical results comparing their performance on large image and language training tasks. Additionally, we address their statistical efficiency and memory consumption behavior. Based on our results, we discuss algorithmic and implementation portions of each framework which hinder performance.

Siddharth Singh, Abhinav Bhatele

In the last few years, the memory requirements to train state-of-the-art neural networks have far exceeded the DRAM capacities of modern hardware accelerators. This has necessitated the development of efficient algorithms to train these neural networks in parallel on large-scale GPU-based clusters. Since computation is relatively inexpensive on modern GPUs, designing and implementing extremely efficient communication in these parallel training algorithms is critical for extracting the maximum performance. This paper presents AxoNN, a parallel deep learning framework that exploits asynchrony and message-driven execution to schedule neural network operations on each GPU, thereby reducing GPU idle time and maximizing hardware efficiency. By using the CPU memory as a scratch space for offloading data periodically during training, AxoNN is able to reduce GPU memory consumption by four times. This allows us to increase the number of parameters per GPU by four times, thus reducing the amount of communication and increasing performance by over 13%. When tested against large transformer models with 12-100 billion parameters on 48-384 NVIDIA Tesla V100 GPUs, AxoNN achieves a per-GPU throughput of 49.4-54.78% of theoretical peak and reduces the training time by 22-37 days (15-25% speedup) as compared to the state-of-the-art.

Ian J. Costello, Abhinav Bhatele

In recent years, several HPC facilities have started continuous monitoring of their systems and jobs to collect performance-related data for understanding performance and operational efficiency. Such data can be used to optimize the performance of individual jobs and the overall system by creating data-driven models that can predict the performance of jobs waiting in the scheduler queue. In this paper, we model the performance of representative control jobs using longitudinal system-wide monitoring data and machine learning to explore the causes of performance variability. We analyze these prediction models in great detail to identify the features that are dominant predictors of performance. We demonstrate that such models can be application-agnostic and can be used for predicting performance of applications that are not included in training.