Prajwal Singhania

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.

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.

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.

Shalmoli Ghosh, Prajwal Singhania, Siddharth Singh et al.

Online forums and social media platforms are increasingly being used to discuss topics of varying polarities where different people take different stances. Several methodologies for automatic stance detection from text have been proposed in literature. To our knowledge, there has not been any systematic investigation towards their reproducibility, and their comparative performances. In this work, we explore the reproducibility of several existing stance detection models, including both neural models and classical classifier-based models. Through experiments on two datasets -- (i)~the popular SemEval microblog dataset, and (ii)~a set of health-related online news articles -- we also perform a detailed comparative analysis of various methods and explore their shortcomings. Implementations of all algorithms discussed in this paper are available at https://github.com/prajwal1210/Stance-Detection-in-Web-and-Social-Media.

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.

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.

Tianjian Huang, Prajwal Singhania, Maziar Sanjabi et al.

Quantization of the parameters of machine learning models, such as deep neural networks, requires solving constrained optimization problems, where the constraint set is formed by the Cartesian product of many simple discrete sets. For such optimization problems, we study the performance of the Alternating Direction Method of Multipliers for Quantization ($\texttt{ADMM-Q}$) algorithm, which is a variant of the widely-used ADMM method applied to our discrete optimization problem. We establish the convergence of the iterates of $\texttt{ADMM-Q}$ to certain $\textit{stationary points}$. To the best of our knowledge, this is the first analysis of an ADMM-type method for problems with discrete variables/constraints. Based on our theoretical insights, we develop a few variants of $\texttt{ADMM-Q}$ that can handle inexact update rules, and have improved performance via the use of "soft projection" and "injecting randomness to the algorithm". We empirically evaluate the efficacy of our proposed approaches.