Urmish Thakker

BigScience Workshop, Teven Le Scao, Angela Fan et al. · allen-ai, berkeley

Large language models (LLMs) have been shown to be able to perform new tasks based on a few demonstrations or natural language instructions. While these capabilities have led to widespread adoption, most LLMs are developed by resource-rich organizations and are frequently kept from the public. As a step towards democratizing this powerful technology, we present BLOOM, a 176B-parameter open-access language model designed and built thanks to a collaboration of hundreds of researchers. BLOOM is a decoder-only Transformer language model that was trained on the ROOTS corpus, a dataset comprising hundreds of sources in 46 natural and 13 programming languages (59 in total). We find that BLOOM achieves competitive performance on a wide variety of benchmarks, with stronger results after undergoing multitask prompted finetuning. To facilitate future research and applications using LLMs, we publicly release our models and code under the Responsible AI License.

Zoltan Csaki, Pian Pawakapan, Urmish Thakker et al.

Recent large language models (LLM) exhibit sub-optimal performance on low-resource languages, as the training data of these models is usually dominated by English and other high-resource languages. Furthermore, it is challenging to train models for low-resource languages, especially from scratch, due to a lack of high quality training data. Adapting pretrained LLMs reduces the need for data in the new language while also providing cross lingual transfer capabilities. However, naively adapting to new languages leads to catastrophic forgetting and poor tokenizer efficiency. In this work, we study how to efficiently adapt any existing pretrained LLM to a new language without running into these issues. In particular, we improve the encoding efficiency of the tokenizer by adding new tokens from the target language and study the data mixing recipe to mitigate forgetting. Our experiments on adapting an English LLM to Hungarian and Thai show that our recipe can reach better performance than open source models on the target language, with minimal regressions on English.

Ravi Raju, Swayambhoo Jain, Bo Li et al.

Large Language Models (LLMs) have revolutionized the landscape of machine learning, yet current benchmarks often fall short in capturing the diverse behavior of these models in real-world applications. A benchmark's usefulness is determined by its ability to clearly differentiate between models of varying capabilities (separability) and closely align with human preferences. Existing frameworks like Alpaca-Eval 2.0 LC \cite{dubois2024lengthcontrolledalpacaevalsimpleway} and Arena-Hard v0.1 \cite{li2024crowdsourced} are limited by their focus on general-purpose queries and lack of diversity across domains such as law, medicine, and multilingual contexts. In this paper, we address these limitations by introducing a novel data pipeline that curates diverse, domain-specific evaluation sets tailored for LLM-as-a-Judge frameworks. Our approach leverages a combination of manual curation, semi-supervised learning to generate clusters, and stratified sampling to ensure balanced representation across a wide range of domains and languages. The resulting evaluation set, which includes 1573 samples across 14 categories, demonstrates high separability (84\%) across ten top-ranked models, and agreement (84\%) with Chatbot Arena and (0.915) Spearman correlation. The agreement values are 9\% better than Arena Hard and 20\% better than AlpacaEval 2.0 LC, while the Spearman coefficient is 0.7 more than the next best benchmark, showcasing a significant improvement in the usefulness of the benchmark. We further provide an open-source evaluation tool that enables fine-grained analysis of model performance across user-defined categories, offering valuable insights for practitioners. This work contributes to the ongoing effort to enhance the transparency, diversity, and effectiveness of LLM evaluation methodologies.

Xueliang Zhao, Lin Zheng, Haige Bo et al.

Formal theorem proving, a field at the intersection of mathematics and computer science, has seen renewed interest with advancements in large language models (LLMs). This paper introduces SubgoalXL, a novel approach that synergizes subgoal-based proofs with expert learning to enhance LLMs' capabilities in formal theorem proving within the Isabelle environment. SubgoalXL addresses two critical challenges: the scarcity of specialized mathematics and theorem-proving data, and the need for improved multi-step reasoning abilities in LLMs. By optimizing data efficiency and employing subgoal-level supervision, SubgoalXL extracts richer information from limited human-generated proofs. The framework integrates subgoal-oriented proof strategies with an expert learning system, iteratively refining formal statement, proof, and subgoal generators. Leveraging the Isabelle environment's advantages in subgoal-based proofs, SubgoalXL achieves a new state-of-the-art performance of 56.1\% in Isabelle on the standard miniF2F dataset, marking an absolute improvement of 4.9\%. Notably, SubgoalXL successfully solves 41 AMC12, 9 AIME, and 3 IMO problems from miniF2F. These results underscore the effectiveness of maximizing limited data utility and employing targeted guidance for complex reasoning in formal theorem proving, contributing to the ongoing advancement of AI reasoning capabilities. The implementation is available at \url{https://github.com/zhaoxlpku/SubgoalXL}.

Venkat Srinivasan, Darshan Gandhi, Urmish Thakker et al.

Large foundation language models have shown their versatility in being able to be adapted to perform a wide variety of downstream tasks, such as text generation, sentiment analysis, semantic search etc. However, training such large foundational models is a non-trivial exercise that requires a significant amount of compute power and expertise from machine learning and systems experts. As models get larger, these demands are only increasing. Sparsity is a promising technique to relieve the compute requirements for training. However, sparsity introduces new challenges in training the sparse model to the same quality as the dense counterparts. Furthermore, sparsity drops the operation intensity and introduces irregular memory access patterns that makes it challenging to efficiently utilize compute resources. This paper demonstrates an end-to-end training flow on a large language model - 13 billion GPT - using sparsity and dataflow. The dataflow execution model and architecture enables efficient on-chip irregular memory accesses as well as native kernel fusion and pipelined parallelism that helps recover device utilization. We show that we can successfully train GPT 13B to the same quality as the dense GPT 13B model, while achieving an end-end speedup of 4.5x over dense A100 baseline.

Shubhangi Upasani, Chen Wu, Jay Rainton et al.

Test-time adaptation enables large language models (LLMs) to modify their behavior at inference without updating model parameters. A common approach is many-shot prompting, where large numbers of in-context learning (ICL) examples are injected as an input-space test-time update. Although performance can improve as more demonstrations are added, the reliability and limits of this update mechanism remain poorly understood, particularly for open-source models. We present an empirical study of many-shot prompting across tasks and model backbones, analyzing how performance varies with update magnitude, example ordering, and selection policy. We further study Dynamic and Reinforced ICL as alternative test-time update strategies that control which information is injected and how it constrains model behavior. We find that many-shot prompting is effective for structured tasks where demonstrations provide high information gain, but is highly sensitive to selection strategy and often shows limited benefits for open-ended generation tasks. Overall, we characterize the practical limits of prompt-based test-time adaptation and outline when input-space updates are beneficial versus harmful.

Ravi Raju, Mengmeng Ji, Shubhangi Upasani et al.

Rapidly increasing context lengths have led to the assumption that large language models (LLMs) can directly reason over entire codebases. Concurrently, recent advances in LLMs have enabled strong performance on software engineering benchmarks, particularly when paired with agentic workflows. In this work, we systematically evaluate whether current LLMs can reliably perform long-context code debugging and patch generation. Using SWE-bench Verified as a controlled experimental setting, we first evaluate state-of-the-art models within an agentic harness (mini-SWE-agent), where performance improves substantially: GPT-5-nano achieves up to a 31\% resolve rate on 100 samples, and open-source models such as Deepseek-R1-0528 obtain competitive results. However, token-level analysis shows that successful agentic trajectories typically remain under 20k tokens, and that longer accumulated contexts correlate with lower success rates, indicating that agentic success primarily arises from task decomposition into short-context steps rather than effective long-context reasoning. To directly test long-context capability, we construct a data pipeline where we artificially inflate the context length of the input by placing the relevant files into the context (ensuring perfect retrieval recall); we then study single-shot patch generation under genuinely long contexts (64k-128k tokens). Despite this setup, performance degrades sharply: Qwen3-Coder-30B-A3B achieves only a 7\% resolve rate at 64k context, while GPT-5-nano solves none of the tasks. Qualitative analysis reveals systematic failure modes, including hallucinated diffs, incorrect file targets, and malformed patch headers. Overall, our findings highlight a significant gap between nominal context length and usable context capacity in current LLMs, and suggest that existing agentic coding benchmarks do not meaningfully evaluate long-context reasoning.

Jonathan Li, Nasim Farahini, Evgenii Iuliugin et al.

The proliferation of 100B+ parameter Large Language Models (LLMs) with 100k+ context length support have resulted in increasing demands for on-chip memory to support large KV caches. Techniques such as StreamingLLM and SnapKV demonstrate how to control KV cache size while maintaining model accuracy. Yet, these techniques are not commonly used within industrial deployments using frameworks like vLLM or SGLang. The reason is twofold: on one hand, the static graphs and continuous batching methodology employed by these frameworks make it difficult to admit modifications to the standard multi-head attention algorithm, while on the other hand, the accuracy implications of such techniques on modern instruction-following and reasoning models are not well understood, obfuscating the need for implementing these techniques. In this paper, we explore these accuracy implications on Llama-3.1-8B-Instruct and DeepSeek-R1, and develop SnapStream, a KV cache compression method that can be deployed at scale. We demonstrate the efficacy of SnapStream in a 16-way tensor-parallel deployment of DeepSeek-671B on SambaNova SN40L accelerators running at 128k context length and up to 1832 tokens per second in a real production setting. SnapStream enables $4\times$ improved on-chip memory usage and introduces minimal accuracy degradation on LongBench-v2, AIME24 and LiveCodeBench. To the best of our knowledge, this is the first implementation of sparse KV attention techniques deployed in a production inference system with static graphs and continuous batching.

Shubhangi Upasani, Ravi Shanker Raju, Bo Li et al.

Prompt length is a major bottleneck in agentic large language model (LLM) workloads, where repeated inference steps and multi-call loops incur substantial prefill cost. Recent work on speculative prefill demonstrates that attention-based token importance estimation can enable training-free prompt compression, but this assumes the existence of a draft model that shares the same tokenizer as the target model. In practice, however, agentic pipelines frequently employ models without any smaller in-family draft model. In this work, we study cross-family speculative prefill, where a lightweight draft model from one model family is used to perform prompt compression for a target model from a different family. Using the same speculative prefill mechanism as prior work, we evaluate a range of cross-family draft-target combinations, including Qwen, LLaMA, and DeepSeek models. Across a broad diversity of tasks, we find that attention-based token importance estimation transfers reliably across different model families despite differences in model architectures and tokenizers between draft and target models. Cross-model prompt compression largely retains 90~100% of full-prompt baseline performance and, in some cases, slightly improves accuracy due to denoising effects, while delivering substantial reductions in time to first token (TTFT). These results suggest that speculative prefill depends mainly on task priors and semantic structure, thus serving as a generalizable prompt compression primitive. We discuss the implications of our findings for agentic systems, where repeated long-context inference and heterogeneous model stacks make cross-model prompt compression both necessary and practical.

Qizheng Zhang, Changran Hu, Shubhangi Upasani et al. · stanford

Large language model (LLM) applications such as agents and domain-specific reasoning increasingly rely on context adaptation -- modifying inputs with instructions, strategies, or evidence, rather than weight updates. Prior approaches improve usability but often suffer from brevity bias, which drops domain insights for concise summaries, and from context collapse, where iterative rewriting erodes details over time. Building on the adaptive memory introduced by Dynamic Cheatsheet, we introduce ACE (Agentic Context Engineering), a framework that treats contexts as evolving playbooks that accumulate, refine, and organize strategies through a modular process of generation, reflection, and curation. ACE prevents collapse with structured, incremental updates that preserve detailed knowledge and scale with long-context models. Across agent and domain-specific benchmarks, ACE optimizes contexts both offline (e.g., system prompts) and online (e.g., agent memory), consistently outperforming strong baselines: +10.6% on agents and +8.6% on finance, while significantly reducing adaptation latency and rollout cost. Notably, ACE could adapt effectively without labeled supervision and instead by leveraging natural execution feedback. On the AppWorld leaderboard, ACE matches the top-ranked production-level agent on the overall average and surpasses it on the harder test-challenge split, despite using a smaller open-source model. These results show that comprehensive, evolving contexts enable scalable, efficient, and self-improving LLM systems with low overhead.

Stephen H. Bach, Victor Sanh, Zheng-Xin Yong et al.

PromptSource is a system for creating, sharing, and using natural language prompts. Prompts are functions that map an example from a dataset to a natural language input and target output. Using prompts to train and query language models is an emerging area in NLP that requires new tools that let users develop and refine these prompts collaboratively. PromptSource addresses the emergent challenges in this new setting with (1) a templating language for defining data-linked prompts, (2) an interface that lets users quickly iterate on prompt development by observing outputs of their prompts on many examples, and (3) a community-driven set of guidelines for contributing new prompts to a common pool. Over 2,000 prompts for roughly 170 datasets are already available in PromptSource. PromptSource is available at https://github.com/bigscience-workshop/promptsource.

Victor Sanh, Albert Webson, Colin Raffel et al.

Large language models have recently been shown to attain reasonable zero-shot generalization on a diverse set of tasks (Brown et al., 2020). It has been hypothesized that this is a consequence of implicit multitask learning in language models' pretraining (Radford et al., 2019). Can zero-shot generalization instead be directly induced by explicit multitask learning? To test this question at scale, we develop a system for easily mapping any natural language tasks into a human-readable prompted form. We convert a large set of supervised datasets, each with multiple prompts with diverse wording. These prompted datasets allow for benchmarking the ability of a model to perform completely held-out tasks. We fine-tune a pretrained encoder-decoder model (Raffel et al., 2020; Lester et al., 2021) on this multitask mixture covering a wide variety of tasks. The model attains strong zero-shot performance on several standard datasets, often outperforming models up to 16x its size. Further, our approach attains strong performance on a subset of tasks from the BIG-bench benchmark, outperforming models up to 6x its size. All trained models are available at https://github.com/bigscience-workshop/t-zero and all prompts are available at https://github.com/bigscience-workshop/promptsource.

Colby Banbury, Chuteng Zhou, Igor Fedorov et al.

Executing machine learning workloads locally on resource constrained microcontrollers (MCUs) promises to drastically expand the application space of IoT. However, so-called TinyML presents severe technical challenges, as deep neural network inference demands a large compute and memory budget. To address this challenge, neural architecture search (NAS) promises to help design accurate ML models that meet the tight MCU memory, latency and energy constraints. A key component of NAS algorithms is their latency/energy model, i.e., the mapping from a given neural network architecture to its inference latency/energy on an MCU. In this paper, we observe an intriguing property of NAS search spaces for MCU model design: on average, model latency varies linearly with model operation (op) count under a uniform prior over models in the search space. Exploiting this insight, we employ differentiable NAS (DNAS) to search for models with low memory usage and low op count, where op count is treated as a viable proxy to latency. Experimental results validate our methodology, yielding our MicroNet models, which we deploy on MCUs using Tensorflow Lite Micro, a standard open-source NN inference runtime widely used in the TinyML community. MicroNets demonstrate state-of-the-art results for all three TinyMLperf industry-standard benchmark tasks: visual wake words, audio keyword spotting, and anomaly detection. Models and training scripts can be found at github.com/ARM-software/ML-zoo.

Raghu Prabhakar, Ram Sivaramakrishnan, Darshan Gandhi et al.

Monolithic large language models (LLMs) like GPT-4 have paved the way for modern generative AI applications. Training, serving, and maintaining monolithic LLMs at scale, however, remains prohibitively expensive and challenging. The disproportionate increase in compute-to-memory ratio of modern AI accelerators have created a memory wall, necessitating new methods to deploy AI. Composition of Experts (CoE) is an alternative modular approach that lowers the cost and complexity of training and serving. However, this approach presents two key challenges when using conventional hardware: (1) without fused operations, smaller models have lower operational intensity, which makes high utilization more challenging to achieve; and (2) hosting a large number of models can be either prohibitively expensive or slow when dynamically switching between them. In this paper, we describe how combining CoE, streaming dataflow, and a three-tier memory system scales the AI memory wall. We describe Samba-CoE, a CoE system with 150 experts and a trillion total parameters. We deploy Samba-CoE on the SambaNova SN40L Reconfigurable Dataflow Unit (RDU) - a commercial dataflow accelerator architecture that has been co-designed for enterprise inference and training applications. The chip introduces a new three-tier memory system with on-chip distributed SRAM, on-package HBM, and off-package DDR DRAM. A dedicated inter-RDU network enables scaling up and out over multiple sockets. We demonstrate speedups ranging from 2$\times$ to 13$\times$ on various benchmarks running on eight RDU sockets compared with an unfused baseline. We show that for CoE inference deployments, the 8-socket RDU Node reduces machine footprint by up to 19$\times$, speeds up model switching time by 15$\times$ to 31$\times$, and achieves an overall speedup of 3.7$\times$ over a DGX H100 and 6.6$\times$ over a DGX A100.

Zoltan Csaki, Bo Li, Jonathan Li et al.

Despite the widespread availability of LLMs, there remains a substantial gap in their capabilities and availability across diverse languages. One approach to address these issues has been to take an existing pre-trained LLM and continue to train it on new languages. While prior works have experimented with language adaptation, many questions around best practices and methodology have not been covered. In this paper, we present a comprehensive investigation into the adaptation of LLMs to new languages. Our study covers the key components in this process, including vocabulary extension, direct preference optimization and the data scarcity problem for human alignment in low-resource languages. We scale these experiments across 9 languages and 2 parameter scales (7B and 70B). We compare our models against Llama 2, Aya-101, XGLM, BLOOM and existing language experts, outperforming all prior published baselines. Additionally, all evaluation code and checkpoints are made public to facilitate future research.

Hadi Amini, Md Jueal Mia, Yasaman Saadati et al.

Language models (LMs) are machine learning models designed to predict linguistic patterns by estimating the probability of word sequences based on large-scale datasets, such as text. LMs have a wide range of applications in natural language processing (NLP) tasks, including autocomplete and machine translation. Although larger datasets typically enhance LM performance, scalability remains a challenge due to constraints in computational power and resources. Distributed computing strategies offer essential solutions for improving scalability and managing the growing computational demand. Further, the use of sensitive datasets in training and deployment raises significant privacy concerns. Recent research has focused on developing decentralized techniques to enable distributed training and inference while utilizing diverse computational resources and enabling edge AI. This paper presents a survey on distributed solutions for various LMs, including large language models (LLMs), vision language models (VLMs), multimodal LLMs (MLLMs), and small language models (SLMs). While LLMs focus on processing and generating text, MLLMs are designed to handle multiple modalities of data (e.g., text, images, and audio) and to integrate them for broader applications. To this end, this paper reviews key advancements across the MLLM pipeline, including distributed training, inference, fine-tuning, and deployment, while also identifying the contributions, limitations, and future areas of improvement. Further, it categorizes the literature based on six primary focus areas of decentralization. Our analysis describes gaps in current methodologies for enabling distributed solutions for LMs and outline future research directions, emphasizing the need for novel solutions to enhance the robustness and applicability of distributed LMs.

Guangtao Wang, Shubhangi Upasani, Chen Wu et al.

Efficient long-context inference is critical as large language models (LLMs) adopt context windows of ranging from 128K to 1M tokens. However, the growing key-value (KV) cache and the high computational complexity of attention create significant bottlenecks in memory usage and latency. In this paper, we find that attention in diverse long-context tasks exhibits sparsity, and LLMs implicitly "know" which tokens can be dropped or evicted at the head level after the pre-filling stage. Based on this insight, we propose Self-Attention Guided Eviction~(SAGE-KV), a simple and effective KV eviction cache method for long-context inference. After prefilling, our method performs a one-time top-k selection at both the token and head levels to compress the KV cache, enabling efficient inference with the reduced cache. Evaluations on LongBench and three long-context LLMs (Llama3.1-8B-Instruct-128k, Llama3-8B-Prolong-512k-Instruct, and Qwen2.5-7B-Instruct-128k) show that SAGE-KV maintains accuracy comparable to full attention while significantly improving efficiency. Specifically, SAGE-KV achieves 4x higher memory efficiency with improved accuracy over the static KV cache selection method StreamLLM, and 2x higher memory efficiency with better accuracy than the dynamic KV cache selection method Quest.

Fenglu Hong, Ravi Raju, Jonathan Lingjie Li et al.

Speculative decoding is an effective method for accelerating inference of large language models (LLMs) by employing a small draft model to predict the output of a target model. However, when adapting speculative decoding to domain-specific target models, the acceptance rate of the generic draft model drops significantly due to domain shift. In this work, we systematically investigate knowledge distillation techniques for training domain draft models to improve their speculation accuracy. We compare white-box and black-box distillation approaches and explore their effectiveness in various data accessibility scenarios, including historical user queries, curated domain data, and synthetically generated alignment data. Our experiments across Function Calling, Biology, and Chinese domains show that offline distillation consistently outperforms online distillation by 11% to 25%, white-box distillation surpasses black-box distillation by 2% to 10%, and data scaling trends hold across domains. Additionally, we find that synthetic data can effectively align draft models and achieve 80% to 93% of the performance of training on historical user queries. These findings provide practical guidelines for training domain-specific draft models to improve speculative decoding efficiency.

Jonathan Li, Zoltan Csaki, Nidhi Hiremath et al.

Modern VLMs have achieved near-saturation accuracy in English document visual question-answering (VQA). However, this task remains challenging in lower resource languages due to a dearth of suitable training and evaluation data. In this paper we present scalable methods for curating such datasets by focusing on Hungarian, approximately the 17th highest resource language on the internet. Specifically, we present HuDocVQA and HuDocVQA-manual, document VQA datasets that modern VLMs significantly underperform on compared to English DocVQA. HuDocVQA-manual is a small manually curated dataset based on Hungarian documents from Common Crawl, while HuDocVQA is a larger synthetically generated VQA data set from the same source. We apply multiple rounds of quality filtering and deduplication to HuDocVQA in order to match human-level quality in this dataset. We also present HuCCPDF, a dataset of 117k pages from Hungarian Common Crawl PDFs along with their transcriptions, which can be used for training a model for Hungarian OCR. To validate the quality of our datasets, we show how finetuning on a mixture of these datasets can improve accuracy on HuDocVQA for Llama 3.2 11B Instruct by +7.2%. Our datasets and code will be released to the public to foster further research in multilingual DocVQA.

Colby Banbury, Vijay Janapa Reddi, Peter Torelli et al.

Advancements in ultra-low-power tiny machine learning (TinyML) systems promise to unlock an entirely new class of smart applications. However, continued progress is limited by the lack of a widely accepted and easily reproducible benchmark for these systems. To meet this need, we present MLPerf Tiny, the first industry-standard benchmark suite for ultra-low-power tiny machine learning systems. The benchmark suite is the collaborative effort of more than 50 organizations from industry and academia and reflects the needs of the community. MLPerf Tiny measures the accuracy, latency, and energy of machine learning inference to properly evaluate the tradeoffs between systems. Additionally, MLPerf Tiny implements a modular design that enables benchmark submitters to show the benefits of their product, regardless of where it falls on the ML deployment stack, in a fair and reproducible manner. The suite features four benchmarks: keyword spotting, visual wake words, image classification, and anomaly detection.

Urmish Thakker, Paul N. Whatmough, Zhigang Liu et al.

Structured matrices, such as those derived from Kronecker products (KP), are effective at compressing neural networks, but can lead to unacceptable accuracy loss when applied to large models. In this paper, we propose the notion of doping -- addition of an extremely sparse matrix to a structured matrix. Doping facilitates additional degrees of freedom for a small number of parameters, allowing them to independently diverge from the fixed structure. To train LSTMs with doped structured matrices, we introduce the additional parameter matrix while slowly annealing its sparsity level. However, we find that performance degrades as we slowly sparsify the doping matrix, due to co-matrix adaptation (CMA) between the structured and the sparse matrices. We address this over dependence on the sparse matrix using a co-matrix dropout regularization (CMR) scheme. We provide empirical evidence to show that doping, CMA and CMR are concepts generally applicable to multiple structured matrices (Kronecker Product, LMF, Hybrid Matrix Decomposition). Additionally, results with doped kronecker product matrices demonstrate state-of-the-art accuracy at large compression factors (10 - 25x) across 4 natural language processing applications with minor loss in accuracy. Doped KP compression technique outperforms previous state-of-the art compression results by achieving 1.3 - 2.4x higher compression factor at a similar accuracy, while also beating strong alternatives like pruning and low-rank methods by a large margin (8% or more). Additionally, we show that doped KP can be deployed on commodity hardware using the current software stack and achieve 2.5 - 5.5x inference run-time speed-up over baseline.

Urmish Thakker, Jesse Beu, Dibakar Gope et al.

Sequence model based NLP applications can be large. Yet, many applications that benefit from them run on small devices with very limited compute and storage capabilities, while still having run-time constraints. As a result, there is a need for a compression technique that can achieve significant compression without negatively impacting inference run-time and task accuracy. This paper proposes a new compression technique called Hybrid Matrix Factorization that achieves this dual objective. HMF improves low-rank matrix factorization (LMF) techniques by doubling the rank of the matrix using an intelligent hybrid-structure leading to better accuracy than LMF. Further, by preserving dense matrices, it leads to faster inference run-time than pruning or structure matrix based compression technique. We evaluate the impact of this technique on 5 NLP benchmarks across multiple tasks (Translation, Intent Detection, Language Modeling) and show that for similar accuracy values and compression factors, HMF can achieve more than 2.32x faster inference run-time than pruning and 16.77% better accuracy than LMF.

Colby R. Banbury, Vijay Janapa Reddi, Max Lam et al.

Recent advancements in ultra-low-power machine learning (TinyML) hardware promises to unlock an entirely new class of smart applications. However, continued progress is limited by the lack of a widely accepted benchmark for these systems. Benchmarking allows us to measure and thereby systematically compare, evaluate, and improve the performance of systems and is therefore fundamental to a field reaching maturity. In this position paper, we present the current landscape of TinyML and discuss the challenges and direction towards developing a fair and useful hardware benchmark for TinyML workloads. Furthermore, we present our four benchmarks and discuss our selection methodology. Our viewpoints reflect the collective thoughts of the TinyMLPerf working group that is comprised of over 30 organizations.

Ahmed Imteaj, Urmish Thakker, Shiqiang Wang et al.

Nowadays, devices are equipped with advanced sensors with higher processing/computing capabilities. Further, widespread Internet availability enables communication among sensing devices. As a result, vast amounts of data are generated on edge devices to drive Internet-of-Things (IoT), crowdsourcing, and other emerging technologies. The collected extensive data can be pre-processed, scaled, classified, and finally, used for predicting future events using machine learning (ML) methods. In traditional ML approaches, data is sent to and processed in a central server, which encounters communication overhead, processing delay, privacy leakage, and security issues. To overcome these challenges, each client can be trained locally based on its available data and by learning from the global model. This decentralized learning structure is referred to as Federated Learning (FL). However, in large-scale networks, there may be clients with varying computational resource capabilities. This may lead to implementation and scalability challenges for FL techniques. In this paper, we first introduce some recently implemented real-life applications of FL. We then emphasize on the core challenges of implementing the FL algorithms from the perspective of resource limitations (e.g., memory, bandwidth, and energy budget) of client clients. We finally discuss open issues associated with FL and highlight future directions in the FL area concerning resource-constrained devices.

Urmish Thakker, Paul N. Whatmough, Zhi-Gang Liu et al.

Kronecker Products (KP) have been used to compress IoT RNN Applications by 15-38x compression factors, achieving better results than traditional compression methods. However when KP is applied to large Natural Language Processing tasks, it leads to significant accuracy loss (approx 26%). This paper proposes a way to recover accuracy otherwise lost when applying KP to large NLP tasks, by allowing additional degrees of freedom in the KP matrix. More formally, we propose doping, a process of adding an extremely sparse overlay matrix on top of the pre-defined KP structure. We call this compression method doped kronecker product compression. To train these models, we present a new solution to the phenomenon of co-matrix adaption (CMA), which uses a new regularization scheme called co matrix dropout regularization (CMR). We present experimental results that demonstrate compression of a large language model with LSTM layers of size 25 MB by 25x with 1.4% loss in perplexity score. At 25x compression, an equivalent pruned network leads to 7.9% loss in perplexity score, while HMD and LMF lead to 15% and 27% loss in perplexity score respectively.

Dibakar Gope, Jesse Beu, Urmish Thakker et al.

MobileNets family of computer vision neural networks have fueled tremendous progress in the design and organization of resource-efficient architectures in recent years. New applications with stringent real-time requirements on highly constrained devices require further compression of MobileNets-like already compute-efficient networks. Model quantization is a widely used technique to compress and accelerate neural network inference and prior works have quantized MobileNets to 4-6 bits albeit with a modest to significant drop in accuracy. While quantization to sub-byte values (i.e. precision less than or equal to 8 bits) has been valuable, even further quantization of MobileNets to binary or ternary values is necessary to realize significant energy savings and possibly runtime speedups on specialized hardware, such as ASICs and FPGAs. Under the key observation that convolutional filters at each layer of a deep neural network may respond differently to ternary quantization, we propose a novel quantization method that generates per-layer hybrid filter banks consisting of full-precision and ternary weight filters for MobileNets. The layer-wise hybrid filter banks essentially combine the strengths of full-precision and ternary weight filters to derive a compact, energy-efficient architecture for MobileNets. Using this proposed quantization method, we quantized a substantial portion of weight filters of MobileNets to ternary values resulting in 27.98% savings in energy, and a 51.07% reduction in the model size, while achieving comparable accuracy and no degradation in throughput on specialized hardware in comparison to the baseline full-precision MobileNets.

Urmish Thakker, Igor Fedorov, Jesse Beu et al.

Recurrent Neural Networks (RNN) can be difficult to deploy on resource constrained devices due to their size. As a result, there is a need for compression techniques that can significantly compress RNNs without negatively impacting task accuracy. This paper introduces a method to compress RNNs for resource constrained environments using Kronecker product (KP). KPs can compress RNN layers by 16-38x with minimal accuracy loss. We show that KP can beat the task accuracy achieved by other state-of-the-art compression techniques (pruning and low-rank matrix factorization) across 4 benchmarks spanning 3 different applications, while simultaneously improving inference run-time.

Newsha Ardalani, Urmish Thakker, Aws Albarghouthi et al.

Porting code from CPU to GPU is costly and time-consuming; Unless much time is invested in development and optimization, it is not obvious, a priori, how much speed-up is achievable or how much room is left for improvement. Knowing the potential speed-up a priori can be very useful: It can save hundreds of engineering hours, help programmers with prioritization and algorithm selection. We aim to address this problem using machine learning in a supervised setting, using solely the single-threaded source code of the program, without having to run or profile the code. We propose a static analysis-based cross-architecture performance prediction framework (Static XAPP) which relies solely on program properties collected using static analysis of the CPU source code and predicts whether the potential speed-up is above or below a given threshold. We offer preliminary results that show we can achieve 94% accuracy in binary classification, in average, across different thresholds

Urmish Thakker, Jesse Beu, Dibakar Gope et al.

Recurrent neural networks can be large and compute-intensive, yet many applications that benefit from RNNs run on small devices with very limited compute and storage capabilities while still having run-time constraints. As a result, there is a need for compression techniques that can achieve significant compression without negatively impacting inference run-time and task accuracy. This paper explores a new compressed RNN cell implementation called Hybrid Matrix Decomposition (HMD) that achieves this dual objective. This scheme divides the weight matrix into two parts - an unconstrained upper half and a lower half composed of rank-1 blocks. This results in output features where the upper sub-vector has "richer" features while the lower-sub vector has "constrained features". HMD can compress RNNs by a factor of 2-4x while having a faster run-time than pruning (Zhu &Gupta, 2017) and retaining more model accuracy than matrix factorization (Grachev et al., 2017). We evaluate this technique on 5 benchmarks spanning 3 different applications, illustrating its generality in the domain of edge computing.

Urmish Thakker, Jesse Beu, Dibakar Gope et al.

Recurrent Neural Networks (RNN) can be difficult to deploy on resource constrained devices due to their size.As a result, there is a need for compression techniques that can significantly compress RNNs without negatively impacting task accuracy. This paper introduces a method to compress RNNs for resource constrained environments using Kronecker product (KP). KPs can compress RNN layers by 15-38x with minimal accuracy loss. By quantizing the resulting models to 8-bits, we further push the compression factor to 50x. We show that KP can beat the task accuracy achieved by other state-of-the-art compression techniques across 5 benchmarks spanning 3 different applications, while simultaneously improving inference run-time. We show that the KP compression mechanism does introduce an accuracy loss, which can be mitigated by a proposed hybrid KP (HKP) approach. Our HKP algorithm provides fine-grained control over the compression ratio, enabling us to regain accuracy lost during compression by adding a small number of model parameters.