Amir Yazdanbakhsh

Aman Madaan, Niket Tandon, Prakhar Gupta et al. · allen-ai, cmu

Like humans, large language models (LLMs) do not always generate the best output on their first try. Motivated by how humans refine their written text, we introduce Self-Refine, an approach for improving initial outputs from LLMs through iterative feedback and refinement. The main idea is to generate an initial output using an LLMs; then, the same LLMs provides feedback for its output and uses it to refine itself, iteratively. Self-Refine does not require any supervised training data, additional training, or reinforcement learning, and instead uses a single LLM as the generator, refiner, and feedback provider. We evaluate Self-Refine across 7 diverse tasks, ranging from dialog response generation to mathematical reasoning, using state-of-the-art (GPT-3.5, ChatGPT, and GPT-4) LLMs. Across all evaluated tasks, outputs generated with Self-Refine are preferred by humans and automatic metrics over those generated with the same LLM using conventional one-step generation, improving by ~20% absolute on average in task performance. Our work demonstrates that even state-of-the-art LLMs like GPT-4 can be further improved at test time using our simple, standalone approach.

Alexander Shypula, Aman Madaan, Yimeng Zeng et al. · cmu

With the decline of Moore's law, optimizing program performance has become a major focus of software research. However, high-level optimizations such as API and algorithm changes remain elusive due to the difficulty of understanding the semantics of code. Simultaneously, pretrained large language models (LLMs) have demonstrated strong capabilities at solving a wide range of programming tasks. To that end, we introduce a framework for adapting LLMs to high-level program optimization. First, we curate a dataset of performance-improving edits made by human programmers of over 77,000 competitive C++ programming submission pairs, accompanied by extensive unit tests. A major challenge is the significant variability of measuring performance on commodity hardware, which can lead to spurious "improvements." To isolate and reliably evaluate the impact of program optimizations, we design an environment based on the gem5 full system simulator, the de facto simulator used in academia and industry. Next, we propose a broad range of adaptation strategies for code optimization; for prompting, these include retrieval-based few-shot prompting and chain-of-thought, and for finetuning, these include performance-conditioned generation and synthetic data augmentation based on self-play. A combination of these techniques achieves a mean speedup of 6.86 with eight generations, higher than average optimizations from individual programmers (3.66). Using our model's fastest generations, we set a new upper limit on the fastest speedup possible for our dataset at 9.64 compared to using the fastest human submissions available (9.56).

Joo Hyung Lee, Wonpyo Park, Nicole Mitchell et al. · mila

This paper introduces JaxPruner, an open-source JAX-based pruning and sparse training library for machine learning research. JaxPruner aims to accelerate research on sparse neural networks by providing concise implementations of popular pruning and sparse training algorithms with minimal memory and latency overhead. Algorithms implemented in JaxPruner use a common API and work seamlessly with the popular optimization library Optax, which, in turn, enables easy integration with existing JAX based libraries. We demonstrate this ease of integration by providing examples in four different codebases: Scenic, t5x, Dopamine and FedJAX and provide baseline experiments on popular benchmarks.

Aman Madaan, Amir Yazdanbakhsh · cmu

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

Chaitanya Mamatha Ananda, Rajiv Gupta, Mircea Trofin et al.

Post-link optimizers (PLOs) such as Propeller and BOLT have demonstrated that precise, profile-guided code layout can extract significant performance gains from heavily optimized binaries. However, these systems are currently restricted to intraprocedural techniques, leaving the global potential of interprocedural layout largely untapped. Interprocedural code layout is historically difficult due to a combinatorially intractable search space and complex call-return semantics that are challenging to model. Consequently, the performance potential of fine-grained interprocedural layout remains unproven in practice. AI-PROPELLER uses Magellan, an agentic workflow that evolves the compiler heuristic in Propeller into a fine-grained interprocedural optimizer and fine-tunes the resulting policy hyperparameters. To ensure high-fidelity, we move away from approximate static cost models and the agentic workflow generates multiple layout variants that are executed on actual hardware to measure real performance counters, providing a precise reward signal for the evolutionary loop. AI-PROPELLER has been evaluated on several benchmarks including large warehouse-scale applications and experiments show performance improvements of 0.23% to 1.6% optimized with state-of-the-art FDO and PLO which is significant for real-world binaries. This is the first time ever that large warehouse-scale applications in industrial settings have been optimized with fine-grained interprocedural code layout.

Zheng Li, Soroush Ghodrati, Amir Yazdanbakhsh et al.

Self-attention is a key enabler of state-of-art accuracy for various transformer-based Natural Language Processing models. This attention mechanism calculates a correlation score for each word with respect to the other words in a sentence. Commonly, only a small subset of words highly correlates with the word under attention, which is only determined at runtime. As such, a significant amount of computation is inconsequential due to low attention scores and can potentially be pruned. The main challenge is finding the threshold for the scores below which subsequent computation will be inconsequential. Although such a threshold is discrete, this paper formulates its search through a soft differentiable regularizer integrated into the loss function of the training. This formulation piggy backs on the back-propagation training to analytically co-optimize the threshold and the weights simultaneously, striking a formally optimal balance between accuracy and computation pruning. To best utilize this mathematical innovation, we devise a bit-serial architecture, dubbed LeOPArd, for transformer language models with bit-level early termination microarchitectural mechanism. We evaluate our design across 43 back-end tasks for MemN2N, BERT, ALBERT, GPT-2, and Vision transformer models. Post-layout results show that, on average, LeOPArd yields 1.9x and 3.9x speedup and energy reduction, respectively, while keeping the average accuracy virtually intact (<0.2% degradation)

Sheng-Chun Kao, Amir Yazdanbakhsh, Suvinay Subramanian et al.

Sparsity has become one of the promising methods to compress and accelerate Deep Neural Networks (DNNs). Among different categories of sparsity, structured sparsity has gained more attention due to its efficient execution on modern accelerators. Particularly, N:M sparsity is attractive because there are already hardware accelerator architectures that can leverage certain forms of N:M structured sparsity to yield higher compute-efficiency. In this work, we focus on N:M sparsity and extensively study and evaluate various training recipes for N:M sparsity in terms of the trade-off between model accuracy and compute cost (FLOPs). Building upon this study, we propose two new decay-based pruning methods, namely "pruning mask decay" and "sparse structure decay". Our evaluations indicate that these proposed methods consistently deliver state-of-the-art (SOTA) model accuracy, comparable to unstructured sparsity, on a Transformer-based model for a translation task. The increase in the accuracy of the sparse model using the new training recipes comes at the cost of marginal increase in the total training compute (FLOPs).

Amir Yazdanbakhsh, Ashkan Moradifirouzabadi, Zheng Li et al.

As its core computation, a self-attention mechanism gauges pairwise correlations across the entire input sequence. Despite favorable performance, calculating pairwise correlations is prohibitively costly. While recent work has shown the benefits of runtime pruning of elements with low attention scores, the quadratic complexity of self-attention mechanisms and their on-chip memory capacity demands are overlooked. This work addresses these constraints by architecting an accelerator, called SPRINT, which leverages the inherent parallelism of ReRAM crossbar arrays to compute attention scores in an approximate manner. Our design prunes the low attention scores using a lightweight analog thresholding circuitry within ReRAM, enabling SPRINT to fetch only a small subset of relevant data to on-chip memory. To mitigate potential negative repercussions for model accuracy, SPRINT re-computes the attention scores for the few fetched data in digital. The combined in-memory pruning and on-chip recompute of the relevant attention scores enables SPRINT to transform quadratic complexity to a merely linear one. In addition, we identify and leverage a dynamic spatial locality between the adjacent attention operations even after pruning, which eliminates costly yet redundant data fetches. We evaluate our proposed technique on a wide range of state-of-the-art transformer models. On average, SPRINT yields 7.5x speedup and 19.6x energy reduction when total 16KB on-chip memory is used, while virtually on par with iso-accuracy of the baseline models (on average 0.36% degradation).

Yucheng Lu, Shivani Agrawal, Suvinay Subramanian et al.

Recent innovations on hardware (e.g. Nvidia A100) have motivated learning N:M structured sparsity masks from scratch for fast model inference. However, state-of-the-art learning recipes in this regime (e.g. SR-STE) are proposed for non-adaptive optimizers like momentum SGD, while incurring non-trivial accuracy drop for Adam-trained models like attention-based LLMs. In this paper, we first demonstrate such gap origins from poorly estimated second moment (i.e. variance) in Adam states given by the masked weights. We conjecture that learning N:M masks with Adam should take the critical regime of variance estimation into account. In light of this, we propose STEP, an Adam-aware recipe that learns N:M masks with two phases: first, STEP calculates a reliable variance estimate (precondition phase) and subsequently, the variance remains fixed and is used as a precondition to learn N:M masks (mask-learning phase). STEP automatically identifies the switching point of two phases by dynamically sampling variance changes over the training trajectory and testing the sample concentration. Empirically, we evaluate STEP and other baselines such as ASP and SR-STE on multiple tasks including CIFAR classification, machine translation and LLM fine-tuning (BERT-Base, GPT-2). We show STEP mitigates the accuracy drop of baseline recipes and is robust to aggressive structured sparsity ratios.

Ondrej Sykora, Phitchaya Mangpo Phothilimthana, Charith Mendis et al.

Analytical hardware performance models yield swift estimation of desired hardware performance metrics. However, developing these analytical models for modern processors with sophisticated microarchitectures is an extremely laborious task and requires a firm understanding of target microarchitecture's internal structure. In this paper, we introduce GRANITE, a new machine learning model that estimates the throughput of basic blocks across different microarchitectures. GRANITE uses a graph representation of basic blocks that captures both structural and data dependencies between instructions. This representation is processed using a graph neural network that takes advantage of the relational information captured in the graph and learns a rich neural representation of the basic block that allows more precise throughput estimation. Our results establish a new state-of-the-art for basic block performance estimation with an average test error of 6.9% across a wide range of basic blocks and microarchitectures for the x86-64 target. Compared to recent work, this reduced the error by 1.7% while improving training and inference throughput by approximately 3.0x. In addition, we propose the use of multi-task learning with independent multi-layer feed forward decoder networks. Our results show that this technique further improves precision of all learned models while significantly reducing per-microarchitecture training costs. We perform an extensive set of ablation studies and comparisons with prior work, concluding a set of methods to achieve high accuracy for basic block performance estimation.

Hanjiang Wu, Abhimanyu Rajeshkumar Bambhaniya, Sarbartha Banerjee et al.

Modern large language model (LLM) inference has progressively disaggregated to keep pace with growing model sizes and tight TTFT and TPOT service-level objectives: from chunked-prefill aggregation, to prefill-decode (P/D) disaggregation, and most recently to operator-level Attention-FFN Disaggregation (AFD). This trend is especially important for mixture-of-experts (MoE) models, where memory-bound attention, compute-intensive expert FFNs, and MoE dispatch/combine communication create distinct resource demands. AFD further exposes this heterogeneity by placing attention and MoE-FFN execution on separate GPU groups. Each level of disaggregation deepens the scheduling design space across workload characteristics, resource allocation, and interconnect topology, raising the central question: when does each level actually pay off? We systematically characterize this trade-off for MoE inference across realistic workloads spanning input/output sequence lengths, prefix-KV reuse, and per-user latency constraints. Using chunked-prefill and P/D disaggregation as baselines, we study the benefits and limits of AFD at scale through a framework that fuses on-device kernel measurements with high-fidelity network simulation. Under strict TTFT/TPOT SLOs, AFD sustains around 4k tokens/s of system throughput on DeepSeek-V3.2 across chat, coding, and agentic-coding workloads, where non-AFD deployments are infeasible. We distill concrete takeaways for jointly optimizing throughput and interactivity, including how to partition attention and FFN across GPUs as a function of workload and model architecture, providing design principles for current rack- and cluster-scale deployments as well as future disaggregated AI infrastructure.

Liu Yang, Zeyu Nie, Andrew Liu et al.

The transition from sequential to parallel computing is essential for modern high-performance applications but is hindered by the steep learning curve of concurrent programming. This challenge is magnified for irregular data structures (such as sparse graphs, unbalanced trees, and non-uniform meshes) where static scheduling fails and data dependencies are unpredictable. Current Large Language Models (LLMs) often fail catastrophically on these tasks, generating code plagued by subtle race conditions, deadlocks, and sub-optimal scaling. We bridge this gap with ParEVO, a framework designed to synthesize high-performance parallel algorithms for irregular data. Our contributions include: (1) The Parlay-Instruct Corpus, a curated dataset of 13,820 tasks synthesized via a "Critic-Refine" pipeline that explicitly filters for empirically performant algorithms that effectively utilize Work-Span parallel primitives; (2) specialized DeepSeek, Qwen, and Gemini models fine-tuned to align probabilistic generation with the rigorous semantics of the ParlayLib library; and (3) an Evolutionary Coding Agent (ECA) that improves the "last mile" of correctness by iteratively repairing code using feedback from compilers, dynamic race detectors, and performance profilers. On the ParEval benchmark, ParEVO achieves an average 106x speedup (with a maximum of 1103x) across the suite, and a robust 13.6x speedup specifically on complex irregular graph problems, outperforming state-of-the-art commercial models. Furthermore, our evolutionary approach matches state-of-the-art expert human baselines, achieving up to a 4.1x speedup on specific highly-irregular kernels. Source code and datasets are available at https://github.com/WildAlg/ParEVO.

Jeffrey Jian Ma, Milad Hashemi, Amir Yazdanbakhsh et al.

Optimizing the performance of large-scale software repositories demands expertise in code reasoning and software engineering (SWE) to reduce runtime while preserving program correctness. However, most benchmarks emphasize what to fix rather than how to fix code. We introduce SWE-fficiency, a benchmark for evaluating repository-level performance optimization on real workloads. Our suite contains 498 tasks across nine widely used data-science, machine-learning, and HPC repositories (e.g., numpy, pandas, scipy): given a complete codebase and a slow workload, an agent must investigate code semantics, localize bottlenecks and relevant tests, and produce a patch that matches or exceeds expert speedup while passing the same unit tests. To enable this how-to-fix evaluation, our automated pipeline scrapes GitHub pull requests for performance-improving edits, combining keyword filtering, static analysis, coverage tooling, and execution validation to both confirm expert speedup baselines and identify relevant repository unit tests. Empirical evaluation of state-of-the-art agents reveals significant underperformance. On average, agents achieve less than 0.15x the expert speedup: agents struggle in localizing optimization opportunities, reasoning about execution across functions, and maintaining correctness in proposed edits. We release the benchmark and accompanying data pipeline to facilitate research on automated performance engineering and long-horizon software reasoning.

Abhimanyu Rajeshkumar Bambhaniya, Amir Yazdanbakhsh, Suvinay Subramanian et al.

N:M Structured sparsity has garnered significant interest as a result of relatively modest overhead and improved efficiency. Additionally, this form of sparsity holds considerable appeal for reducing the memory footprint owing to their modest representation overhead. There have been efforts to develop training recipes for N:M structured sparsity, they primarily focus on low-sparsity regions ($\sim$50\%). Nonetheless, performance of models trained using these approaches tends to decline when confronted with high-sparsity regions ($>$80\%). In this work, we study the effectiveness of existing sparse training recipes at \textit{high-sparsity regions} and argue that these methods fail to sustain the model quality on par with low-sparsity regions. We demonstrate that the significant factor contributing to this disparity is the presence of elevated levels of induced noise in the gradient magnitudes. To mitigate this undesirable effect, we employ decay mechanisms to progressively restrict the flow of gradients towards pruned elements. Our approach improves the model quality by up to 2$\%$ and 5$\%$ in vision and language models at high sparsity regime, respectively. We also evaluate the trade-off between model accuracy and training compute cost in terms of FLOPs. At iso-training FLOPs, our method yields better performance compared to conventional sparse training recipes, exhibiting an accuracy improvement of up to 2$\%$. The source code is available at https://github.com/abhibambhaniya/progressive_gradient_flow_nm_sparsity.

Gheorghe Comanici, Eric Bieber, Mike Schaekermann et al. · amazon-science, baidu

In this report, we introduce the Gemini 2.X model family: Gemini 2.5 Pro and Gemini 2.5 Flash, as well as our earlier Gemini 2.0 Flash and Flash-Lite models. Gemini 2.5 Pro is our most capable model yet, achieving SoTA performance on frontier coding and reasoning benchmarks. In addition to its incredible coding and reasoning skills, Gemini 2.5 Pro is a thinking model that excels at multimodal understanding and it is now able to process up to 3 hours of video content. Its unique combination of long context, multimodal and reasoning capabilities can be combined to unlock new agentic workflows. Gemini 2.5 Flash provides excellent reasoning abilities at a fraction of the compute and latency requirements and Gemini 2.0 Flash and Flash-Lite provide high performance at low latency and cost. Taken together, the Gemini 2.X model generation spans the full Pareto frontier of model capability vs cost, allowing users to explore the boundaries of what is possible with complex agentic problem solving.

Shvetank Prakash, Andrew Cheng, Jason Yik et al.

We introduce QuArch, a dataset of 1500 human-validated question-answer pairs designed to evaluate and enhance language models' understanding of computer architecture. The dataset covers areas including processor design, memory systems, and performance optimization. Our analysis highlights a significant performance gap: the best closed-source model achieves 84% accuracy, while the top small open-source model reaches 72%. We observe notable struggles in memory systems, interconnection networks, and benchmarking. Fine-tuning with QuArch improves small model accuracy by up to 8%, establishing a foundation for advancing AI-driven computer architecture research. The dataset and leaderboard are at https://harvard-edge.github.io/QuArch/.

Arash Nasr-Esfahany, Mohammad Alizadeh, Victor Lee et al.

Cycle-level simulators such as gem5 are widely used in microarchitecture design, but they are prohibitively slow for large-scale design space explorations. We present Concorde, a new methodology for learning fast and accurate performance models of microarchitectures. Unlike existing simulators and learning approaches that emulate each instruction, Concorde predicts the behavior of a program based on compact performance distributions that capture the impact of different microarchitectural components. It derives these performance distributions using simple analytical models that estimate bounds on performance induced by each microarchitectural component, providing a simple yet rich representation of a program's performance characteristics across a large space of microarchitectural parameters. Experiments show that Concorde is more than five orders of magnitude faster than a reference cycle-level simulator, with about 2% average Cycles-Per-Instruction (CPI) prediction error across a range of SPEC, open-source, and proprietary benchmarks. This enables rapid design-space exploration and performance sensitivity analyses that are currently infeasible, e.g., in about an hour, we conducted a first-of-its-kind fine-grained performance attribution to different microarchitectural components across a diverse set of programs, requiring nearly 150 million CPI evaluations.

Shruthi Gorantala, Jianming Tong, Asra Ali et al.

The deployment of Fully Homomorphic Encryption (FHE) at scale is hindered due to its heavy computational overhead. While specialized hardware accelerators like Google Tensor Processing Units (TPUs) can help, mapping complex cryptographic kernels onto such architectures remains a challenge. Efficient execution requires co-optimization between the systolic array-based Matrix Multiplication Unit (MXU) and Vector Processing Units (VPUs), as well as the orchestration of data movement across the vector register files. Existing compiler stacks often abstract low-level hardware utilization, requiring developers to adopt a manual trial-and-error process that often results in fragmented execution and underutilized resources. To accelerate this development process, we use AlphaEvolve to automate the exploration of hardware-aware cryptographic-kernel optimizations. We frame optimization as an evolutionary search problem, utilizing the closed-loop system provided by AlphaEvolve, that leverages LLM-driven code generation. We use real-world feedback from hardware execution and rigorous correctness testing to guide the evolution process. We evaluate AlphaEvolve optimization on primitives for both the TFHE (Jaxite) and CKKS (CROSS) FHE schemes on Google Cloud TPUv5e, a contemporary TPU architecture. Within 24 hours of automated exploration, AlphaEvolve discovered implementation-level optimizations that improve TFHE bootstrap latency by 2.5x and CKKS rotation and multiplication latency by 1.31x and 1.18x, respectively, relative to human-engineered state of the art. These results demonstrate that AlphaEvolve can be used to enable researchers to navigate the optimization trade-offs between cryptography, compilers, and hardware accelerators.

Haoran You, Yichao Fu, Zheng Wang et al.

Autoregressive Large Language Models (LLMs) have achieved impressive performance in language tasks but face two significant bottlenecks: (1) quadratic complexity in the attention module as the number of tokens increases, and (2) limited efficiency due to the sequential processing nature of autoregressive LLMs during generation. While linear attention and speculative decoding offer potential solutions, their applicability and synergistic potential for enhancing autoregressive LLMs remain uncertain. We conduct the first comprehensive study on the efficacy of existing linear attention methods for autoregressive LLMs, integrating them with speculative decoding. We introduce an augmentation technique for linear attention that ensures compatibility with speculative decoding, enabling more efficient training and serving of LLMs. Extensive experiments and ablation studies involving seven existing linear attention models and five encoder/decoder-based LLMs consistently validate the effectiveness of our augmented linearized LLMs. Notably, our approach achieves up to a 6.67 reduction in perplexity on the LLaMA model and up to a 2$\times$ speedup during generation compared to prior linear attention methods. Codes and models are available at https://github.com/GATECH-EIC/Linearized-LLM.

Haoran You, Yipin Guo, Yichao Fu et al.

Large language models (LLMs) have shown impressive performance on language tasks but face challenges when deployed on resource-constrained devices due to their extensive parameters and reliance on dense multiplications, resulting in high memory demands and latency bottlenecks. Shift-and-add reparameterization offers a promising solution by replacing costly multiplications with hardware-friendly primitives in both the attention and multi-layer perceptron (MLP) layers of an LLM. However, current reparameterization techniques require training from scratch or full parameter fine-tuning to restore accuracy, which is resource-intensive for LLMs. To address this, we propose accelerating pretrained LLMs through post-training shift-and-add reparameterization, creating efficient multiplication-free models, dubbed ShiftAddLLM. Specifically, we quantize each weight matrix into binary matrices paired with group-wise scaling factors. The associated multiplications are reparameterized into (1) shifts between activations and scaling factors and (2) queries and adds according to the binary matrices. To reduce accuracy loss, we present a multi-objective optimization method to minimize both weight and output activation reparameterization errors. Additionally, based on varying sensitivity across layers to reparameterization, we develop an automated bit allocation strategy to further reduce memory usage and latency. Experiments on five LLM families and eight tasks consistently validate the effectiveness of ShiftAddLLM, achieving average perplexity improvements of 5.6 and 22.7 points at comparable or lower latency compared to the most competitive quantized LLMs at 3 and 2 bits, respectively, and more than 80% memory and energy reductions over the original LLMs. Codes and models are available at https://github.com/GATECH-EIC/ShiftAddLLM.

Wenqi Jiang, Suvinay Subramanian, Cat Graves et al.

Retrieval-augmented generation (RAG), which combines large language models (LLMs) with retrievals from external knowledge databases, is emerging as a popular approach for reliable LLM serving. However, efficient RAG serving remains an open challenge due to the rapid emergence of many RAG variants and the substantial differences in workload characteristics across them. In this paper, we make three fundamental contributions to advancing RAG serving. First, we introduce RAGSchema, a structured abstraction that captures the wide range of RAG algorithms, serving as a foundation for performance optimization. Second, we analyze several representative RAG workloads with distinct RAGSchema, revealing significant performance variability across these workloads. Third, to address this variability and meet diverse performance requirements, we propose RAGO (Retrieval-Augmented Generation Optimizer), a system optimization framework for efficient RAG serving. Our evaluation shows that RAGO achieves up to a 2x increase in QPS per chip and a 55% reduction in time-to-first-token latency compared to RAG systems built on LLM-system extensions.

Ali TehraniJamsaz, Arijit Bhattacharjee, Le Chen et al.

Recent advancements in Large Language Models (LLMs) have renewed interest in automatic programming language translation. Encoder-decoder transformer models, in particular, have shown promise in translating between different programming languages. However, translating between a language and its high-performance computing (HPC) extensions remains underexplored due to challenges such as complex parallel semantics. In this paper, we introduce CodeRosetta, an encoder-decoder transformer model designed specifically for translating between programming languages and their HPC extensions. CodeRosetta is evaluated on C++ to CUDA and Fortran to C++ translation tasks. It uses a customized learning framework with tailored pretraining and training objectives to effectively capture both code semantics and parallel structural nuances, enabling bidirectional translation. Our results show that CodeRosetta outperforms state-of-the-art baselines in C++ to CUDA translation by 2.9 BLEU and 1.72 CodeBLEU points while improving compilation accuracy by 6.05%. Compared to general closed-source LLMs, our method improves C++ to CUDA translation by 22.08 BLEU and 14.39 CodeBLEU, with 2.75% higher compilation accuracy. Finally, CodeRosetta exhibits proficiency in Fortran to parallel C++ translation, marking it, to our knowledge, as the first encoder-decoder model for this complex task, improving CodeBLEU by at least 4.63 points compared to closed-source and open-code LLMs.

Tian Jin, Ellie Y. Cheng, Zack Ankner et al.

Decoding with autoregressive large language models (LLMs) traditionally occurs sequentially, generating one token after another. An emerging line of work explored parallel decoding by identifying and simultaneously generating semantically independent chunks of LLM responses. However, these techniques rely on hand-crafted heuristics tied to syntactic structures like lists and paragraphs, making them rigid and imprecise. We present PASTA, a learning-based system that teaches LLMs to identify semantic independence and express parallel decoding opportunities in their own responses. At its core are PASTA-LANG and its interpreter: PASTA-LANG is an annotation language that enables LLMs to express semantic independence in their own responses; the language interpreter acts on these annotations to orchestrate parallel decoding on-the-fly at inference time. Through a two-stage finetuning process, we train LLMs to generate PASTA-LANG annotations that optimize both response quality and decoding speed. Evaluation on AlpacaEval, an instruction following benchmark, shows that our approach Pareto-dominates existing methods in terms of decoding speed and response quality; our results demonstrate geometric mean speedups ranging from 1.21x to 1.93x with corresponding quality changes of +2.2% to -7.1%, measured by length-controlled win rates against sequential decoding baseline.

Yoonsung Kim, Changhun Oh, Jinwoo Hwang et al.

Deep neural network (DNN) video analytics is crucial for autonomous systems such as self-driving vehicles, unmanned aerial vehicles (UAVs), and security robots. However, real-world deployment faces challenges due to their limited computational resources and battery power. To tackle these challenges, continuous learning exploits a lightweight "student" model at deployment (inference), leverages a larger "teacher" model for labeling sampled data (labeling), and continuously retrains the student model to adapt to changing scenarios (retraining). This paper highlights the limitations in state-of-the-art continuous learning systems: (1) they focus on computations for retraining, while overlooking the compute needs for inference and labeling, (2) they rely on power-hungry GPUs, unsuitable for battery-operated autonomous systems, and (3) they are located on a remote centralized server, intended for multi-tenant scenarios, again unsuitable for autonomous systems due to privacy, network availability, and latency concerns. We propose a hardware-algorithm co-designed solution for continuous learning, DaCapo, that enables autonomous systems to perform concurrent executions of inference, labeling, and training in a performant and energy-efficient manner. DaCapo comprises (1) a spatially-partitionable and precision-flexible accelerator enabling parallel execution of kernels on sub-accelerators at their respective precisions, and (2) a spatiotemporal resource allocation algorithm that strategically navigates the resource-accuracy tradeoff space, facilitating optimal decisions for resource allocation to achieve maximal accuracy. Our evaluation shows that DaCapo achieves 6.5% and 5.5% higher accuracy than a state-of-the-art GPU-based continuous learning systems, Ekya and EOMU, respectively, while consuming 254x less power.

Santosh Pandey, Amir Yazdanbakhsh, Hang Liu

Microarchitecture simulators are indispensable tools for microarchitecture designers to validate, estimate, and optimize new hardware that meets specific design requirements. While the quest for a fast, accurate and detailed microarchitecture simulation has been ongoing for decades, existing simulators excel and fall short at different aspects: (i) Although execution-driven simulation is accurate and detailed, it is extremely slow and requires expert-level experience to design. (ii) Trace-driven simulation reuses the execution traces in pursuit of fast simulation but faces accuracy concerns and fails to achieve significant speedup. (iii) Emerging deep learning (DL)-based simulations are remarkably fast and have acceptable accuracy but fail to provide adequate low-level microarchitectural performance metrics crucial for microarchitectural bottleneck analysis. Additionally, they introduce substantial overheads from trace regeneration and model re-training when simulating a new microarchitecture. Re-thinking the advantages and limitations of the aforementioned simulation paradigms, this paper introduces TAO that redesigns the DL-based simulation with three primary contributions: First, we propose a new training dataset design such that the subsequent simulation only needs functional trace as inputs, which can be rapidly generated and reused across microarchitectures. Second, we redesign the input features and the DL model using self-attention to support predicting various performance metrics. Third, we propose techniques to train a microarchitecture agnostic embedding layer that enables fast transfer learning between different microarchitectural configurations and reduces the re-training overhead of conventional DL-based simulators. Our extensive evaluation shows TAO can reduce the overall training and simulation time by 18.06x over the state-of-the-art DL-based endeavors.

Mohammad Mozaffari, Amir Yazdanbakhsh, Maryam Mehri Dehnavi

Conventional model compression techniques for LLMs address high memory consumption and slow inference challenges but typically require computationally expensive retraining to preserve accuracy. In contrast, one-shot compression methods eliminate retraining cost, but struggle to achieve accuracy comparable to dense models. This paper presents SLIM, a new one-shot compression framework that holistically integrates hardware-friendly quantization, sparsity, and low-rank approximation into a unified process. First, we formulate the quantization process using a probabilistic approach (SLIM-Quant) that enables us to apply uniform quantization. Then, we use an existing one-shot pruning method to apply semi-structured sparsity on top of the quantized weights. Finally, to compensate for the introduced aggregated quantization and sparsity error, we use a novel saliency function with unique invertible and additive features that enables us to mathematically compute the value of low-rank adapters. SLIM improves model accuracy by up to 5.66% (LLaMA-2-7B) for 2:4 sparsity with 4-bit weight quantization, outperforming prior methods. Models compressed with SLIM achieve up to 4.3x and 3.8x on Nvidia RTX3060 and A100 GPUs, respectively. Additionally, they achieve up to 0.23x end-to-end memory reduction in comparison to their dense counterparts. We also propose an optional PEFT recipe that further improves accuracy by up to 1.66% (LLaMA-2-13B) compared to SLIM without fine-tuning.

Samir Khaki, Xiuyu Li, Junxian Guo et al.

Fine-tuning LLMs is both computationally and memory-intensive. While parameter-efficient fine-tuning methods, such as QLoRA and DoRA, reduce the number of trainable parameters and lower memory usage, they do not decrease computational cost. In some cases, they may even slow down fine-tuning. In this paper, we introduce SparseLoRA, a method that accelerates LLM fine-tuning through contextual sparsity. We propose a lightweight, training-free SVD sparsity estimator that dynamically selects a sparse subset of weights for loss and gradient computation. Also, we systematically analyze and address sensitivity across layers, tokens, and training steps. Our experimental results show that SparseLoRA reduces computational cost by up to 2.2 times and a measured speedup of up to 1.6 times while maintaining accuracy across various downstream tasks, including commonsense and arithmetic reasoning, code generation, and instruction following.

Tian Jin, Ahmed Imtiaz Humayun, Utku Evci et al.

Pruning eliminates unnecessary parameters in neural networks; it offers a promising solution to the growing computational demands of large language models (LLMs). While many focus on post-training pruning, sparse pre-training--which combines pruning and pre-training into a single phase--provides a simpler alternative. In this work, we present the first systematic exploration of optimal sparse pre-training configurations for LLMs through an examination of 80 unique pruning schedules across different sparsity levels and training durations. We find that initiating pruning at 25% of total training compute and concluding at 75% achieves near-optimal final evaluation loss. These findings provide valuable insights for efficient and effective sparse pre-training of LLMs. Furthermore, we propose a new scaling law that modifies the Chinchilla scaling law to use the average parameter count over pre-training. Through empirical and theoretical validation, we demonstrate that this modified scaling law accurately models evaluation loss for both sparsely and densely pre-trained LLMs, unifying scaling laws across pre-training paradigms. Our findings indicate that while sparse pre-training achieves the same final model quality as dense pre-training for equivalent compute budgets, it provides substantial benefits through reduced model size, enabling significant potential computational savings during inference.

Amir Yazdanbakhsh

For decades, Moore's Law has served as a steadfast pillar in computer architecture and system design, promoting a clear abstraction between hardware and software. This traditional Moore's computing paradigm has deepened the rift between the two, enabling software developers to achieve near-exponential performance gains often without needing to delve deeply into hardware-specific optimizations. Yet today, Moore's Law -- with its once relentless performance gains now diminished to incremental improvements -- faces inevitable physical barriers. This stagnation necessitates a reevaluation of the conventional system design philosophy. The traditional decoupled system design philosophy, which maintains strict abstractions between hardware and software, is increasingly obsolete. The once-clear boundary between software and hardware is rapidly dissolving, replaced by co-design. It is imperative for the computing community to intensify its commitment to hardware-software co-design, elevating system abstractions to first-class citizens and reimagining design principles to satisfy the insatiable appetite of modern computing. Hardware-software co-design is not a recent innovation. To illustrate its historical evolution, I classify its development into five relatively distinct ``epochs''. This post also highlights the growing influence of the architecture community in interdisciplinary teams -- particularly alongside ML researchers -- and explores why current co-design paradigms are struggling in today's computing landscape. Additionally, I will examine the concept of the ``hardware lottery'' and explore directions to mitigate its constraining influence on the next era of computing innovation.

Mohammad Mozaffari, Samuel Kushnir, Maryam Mehri Dehnavi et al.

Post-training model pruning is a promising solution, yet it faces a trade-off: simple heuristics that zero weights are fast but degrade accuracy, while principled joint optimization methods recover accuracy but are computationally infeasible at modern scale. One-shot methods such as SparseGPT offer a practical trade-off in optimality by applying efficient, approximate heuristic weight updates. To close this gap, we introduce OPTIMA, a practical one-shot post-training pruning method that balances accuracy and scalability. OPTIMA casts layer-wise weight reconstruction after mask selection as independent, row-wise Quadratic Programs (QPs) that share a common layer Hessian. Solving these QPs yields the per-row globally optimal update with respect to the reconstruction objective given the estimated Hessian. The shared-Hessian structure makes the problem highly amenable to batching on accelerators. We implement an accelerator-friendly QP solver that accumulates one Hessian per layer and solves many small QPs in parallel, enabling one-shot post-training pruning at scale on a single accelerator without fine-tuning. OPTIMA integrates with existing mask selectors and consistently improves zero-shot performance across multiple LLM families and sparsity regimes, yielding up to 3.97% absolute accuracy improvement. On an NVIDIA H100, OPTIMA prunes a 8B-parameter transformer end-to-end in 40 hours with 60GB peak memory. Together, these results set a new state-of-the-art accuracy-efficiency trade-offs for one-shot post-training pruning.

Hongzheng Chen, Alexander Novikov, Ngân Vũ et al.

Modern compilers rely on hand-crafted heuristics to guide optimization passes. These human-designed rules often struggle to adapt to the complexity of modern software and hardware and lead to high maintenance burden. To address this challenge, we present Magellan, an agentic framework that evolves the compiler pass itself by synthesizing executable C++ decision logic. Magellan couples an LLM coding agent with evolutionary search and autotuning in a closed loop of generation, evaluation on user-provided macro-benchmarks, and refinement, producing compact heuristics that integrate directly into existing compilers. Across several production optimization tasks, Magellan discovers policies that match or surpass expert baselines. In LLVM function inlining, Magellan synthesizes new heuristics that outperform decades of manual engineering for both binary-size reduction and end-to-end performance. In register allocation, it learns a concise priority rule for live-range processing that matches intricate human-designed policies on a large-scale workload. We also report preliminary results on XLA problems, demonstrating portability beyond LLVM with reduced engineering effort.

Shvetank Prakash, Andrew Cheng, Arya Tschand et al.

The field of computer architecture, which bridges high-level software abstractions and low-level hardware implementations, remains absent from current large language model (LLM) evaluations. To this end, we present QuArch (pronounced 'quark'), the first benchmark designed to facilitate the development and evaluation of LLM knowledge and reasoning capabilities specifically in computer architecture. QuArch provides a comprehensive collection of 2,671 expert-validated question-answer (QA) pairs covering various aspects of computer architecture, including processor design, memory systems, and interconnection networks. Our evaluation reveals that while frontier models possess domain-specific knowledge, they struggle with skills that require higher-order thinking in computer architecture. Frontier model accuracies vary widely (from 34% to 72%) on these advanced questions, highlighting persistent gaps in architectural reasoning across analysis, design, and implementation QAs. By holistically assessing fundamental skills, QuArch provides a foundation for building and measuring LLM capabilities that can accelerate innovation in computing systems. With over 140 contributors from 40 institutions, this benchmark represents a community effort to set the standard for architectural reasoning in LLM evaluation.

Abhimanyu Rajeshkumar Bambhaniya, Hanjiang Wu, Suvinay Subramanian et al.

The rapid evolution of Large Language Models (LLMs) has driven the need for increasingly sophisticated inference pipelines and hardware platforms. Modern LLM serving extends beyond traditional prefill-decode workflows, incorporating multi-stage processes such as Retrieval Augmented Generation (RAG), key-value (KV) cache retrieval, dynamic model routing, and multi step reasoning. These stages exhibit diverse computational demands, requiring distributed systems that integrate GPUs, ASICs, CPUs, and memory-centric architectures. However, existing simulators lack the fidelity to model these heterogeneous, multi-engine workflows, limiting their ability to inform architectural decisions. To address this gap, we introduce HERMES, a Heterogeneous Multi-stage LLM inference Execution Simulator. HERMES models diverse request stages; including RAG, KV retrieval, reasoning, prefill, and decode across complex hardware hierarchies. HERMES supports heterogeneous clients executing multiple models concurrently unlike prior frameworks while incorporating advanced batching strategies and multi-level memory hierarchies. By integrating real hardware traces with analytical modeling, HERMES captures critical trade-offs such as memory bandwidth contention, inter-cluster communication latency, and batching efficiency in hybrid CPU-accelerator deployments. Through case studies, we explore the impact of reasoning stages on end-to-end latency, optimal batching strategies for hybrid pipelines, and the architectural implications of remote KV cache retrieval. HERMES empowers system designers to navigate the evolving landscape of LLM inference, providing actionable insights into optimizing hardware-software co-design for next-generation AI workloads.

Aviral Kumar, Amir Yazdanbakhsh, Milad Hashemi et al.

Industry has gradually moved towards application-specific hardware accelerators in order to attain higher efficiency. While such a paradigm shift is already starting to show promising results, designers need to spend considerable manual effort and perform a large number of time-consuming simulations to find accelerators that can accelerate multiple target applications while obeying design constraints. Moreover, such a "simulation-driven" approach must be re-run from scratch every time the set of target applications or design constraints change. An alternative paradigm is to use a "data-driven", offline approach that utilizes logged simulation data, to architect hardware accelerators, without needing any form of simulations. Such an approach not only alleviates the need to run time-consuming simulation, but also enables data reuse and applies even when set of target applications changes. In this paper, we develop such a data-driven offline optimization method for designing hardware accelerators, dubbed PRIME, that enjoys all of these properties. Our approach learns a conservative, robust estimate of the desired cost function, utilizes infeasible points, and optimizes the design against this estimate without any additional simulator queries during optimization. PRIME architects accelerators -- tailored towards both single and multiple applications -- improving performance upon state-of-the-art simulation-driven methods by about 1.54x and 1.20x, while considerably reducing the required total simulation time by 93% and 99%, respectively. In addition, PRIME also architects effective accelerators for unseen applications in a zero-shot setting, outperforming simulation-based methods by 1.26x.

Sheng-Chun Kao, Suvinay Subramanian, Gaurav Agrawal et al.

Attention mechanisms, primarily designed to capture pairwise correlations between words, have become the backbone of machine learning, expanding beyond natural language processing into other domains. This growth in adaptation comes at the cost of prohibitively large memory requirements and computational complexity, especially at higher number of input elements. This limitation is due to inherently limited data reuse opportunities and quadratic growth in memory footprints, leading to severe memory-boundedness and limited scalability of input elements. This work addresses these challenges by devising a tailored dataflow optimization, called FLAT, for attention mechanisms without altering their functionality. This dataflow processes costly attention operations through a unique fusion mechanism, transforming the memory footprint quadratic growth to merely a linear one. To realize the full potential of this bespoke mechanism, we propose a tiling approach to enhance the data reuse across attention operations. Our method both mitigates the off-chip bandwidth bottleneck as well as reduces the on-chip memory requirement. FLAT delivers 1.94x (1.76x) speedup and 49% and (42%) of energy savings compared to the state-of-the-art Edge (Cloud) accelerators with no customized dataflow optimization. When on-chip resources are scarce (20 KB-200 KB), FLAT yields, on average, 1.5x end-to-end latency reduction across a diverse range of conventional attention-based models with input sequence lengths ranging from 512-token to 64K-token. Our evaluations demonstrate that state-of-the-art DNN dataflow applied to attention operations reach the efficiency limit for inputs above 512 elements. In contrast, FLAT unblocks transformer models for inputs with up to 64K elements

Kiran Seshadri, Berkin Akin, James Laudon et al.

Edge TPUs are a domain of accelerators for low-power, edge devices and are widely used in various Google products such as Coral and Pixel devices. In this paper, we first discuss the major microarchitectural details of Edge TPUs. Then, we extensively evaluate three classes of Edge TPUs, covering different computing ecosystems, that are either currently deployed in Google products or are the product pipeline, across 423K unique convolutional neural networks. Building upon this extensive study, we discuss critical and interpretable microarchitectural insights about the studied classes of Edge TPUs. Mainly, we discuss how Edge TPU accelerators perform across convolutional neural networks with different structures. Finally, we present our ongoing efforts in developing high-accuracy learned machine learning models to estimate the major performance metrics of accelerators such as latency and energy consumption. These learned models enable significantly faster (in the order of milliseconds) evaluations of accelerators as an alternative to time-consuming cycle-accurate simulators and establish an exciting opportunity for rapid hard-ware/software co-design.

Yanqi Zhou, Xuanyi Dong, Berkin Akin et al.

Neural architectures and hardware accelerators have been two driving forces for the progress in deep learning. Previous works typically attempt to optimize hardware given a fixed model architecture or model architecture given fixed hardware. And the dominant hardware architecture explored in this prior work is FPGAs. In our work, we target the optimization of hardware and software configurations on an industry-standard edge accelerator. We systematically study the importance and strategies of co-designing neural architectures and hardware accelerators. We make three observations: 1) the software search space has to be customized to fully leverage the targeted hardware architecture, 2) the search for the model architecture and hardware architecture should be done jointly to achieve the best of both worlds, and 3) different use cases lead to very different search outcomes. Our experiments show that the joint search method consistently outperforms previous platform-aware neural architecture search, manually crafted models, and the state-of-the-art EfficientNet on all latency targets by around 1% on ImageNet top-1 accuracy. Our method can reduce energy consumption of an edge accelerator by up to 2x under the same accuracy constraint, when co-adapting the model architecture and hardware accelerator configurations.

Amir Yazdanbakhsh, Christof Angermueller, Berkin Akin et al.

The looming end of Moore's Law and ascending use of deep learning drives the design of custom accelerators that are optimized for specific neural architectures. Architecture exploration for such accelerators forms a challenging constrained optimization problem over a complex, high-dimensional, and structured input space with a costly to evaluate objective function. Existing approaches for accelerator design are sample-inefficient and do not transfer knowledge between related optimizations tasks with different design constraints, such as area and/or latency budget, or neural architecture configurations. In this work, we propose a transferable architecture exploration framework, dubbed Apollo, that leverages recent advances in black-box function optimization for sample-efficient accelerator design. We use this framework to optimize accelerator configurations of a diverse set of neural architectures with alternative design constraints. We show that our framework finds high reward design configurations (up to 24.6% speedup) more sample-efficiently than a baseline black-box optimization approach. We further show that by transferring knowledge between target architectures with different design constraints, Apollo is able to find optimal configurations faster and often with better objective value (up to 25% improvements). This encouraging outcome portrays a promising path forward to facilitate generating higher quality accelerators.

Byung Hoon Ahn, Prannoy Pilligundla, Amir Yazdanbakhsh et al.

Achieving faster execution with shorter compilation time can foster further diversity and innovation in neural networks. However, the current paradigm of executing neural networks either relies on hand-optimized libraries, traditional compilation heuristics, or very recently genetic algorithms and other stochastic methods. These methods suffer from frequent costly hardware measurements rendering them not only too time consuming but also suboptimal. As such, we devise a solution that can learn to quickly adapt to a previously unseen design space for code optimization, both accelerating the search and improving the output performance. This solution dubbed Chameleon leverages reinforcement learning whose solution takes fewer steps to converge, and develops an adaptive sampling algorithm that not only focuses on the costly samples (real hardware measurements) on representative points but also uses a domain-knowledge inspired logic to improve the samples itself. Experimentation with real hardware shows that Chameleon provides 4.45x speed up in optimization time over AutoTVM, while also improving inference time of the modern deep networks by 5.6%.

Ahmed T. Elthakeb, Prannoy Pilligundla, FatemehSadat Mireshghallah et al.

Deep Neural Networks (DNNs) typically require massive amount of computation resource in inference tasks for computer vision applications. Quantization can significantly reduce DNN computation and storage by decreasing the bitwidth of network encodings. Recent research affirms that carefully selecting the quantization levels for each layer can preserve the accuracy while pushing the bitwidth below eight bits. However, without arduous manual effort, this deep quantization can lead to significant accuracy loss, leaving it in a position of questionable utility. As such, deep quantization opens a large hyper-parameter space (bitwidth of the layers), the exploration of which is a major challenge. We propose a systematic approach to tackle this problem, by automating the process of discovering the quantization levels through an end-to-end deep reinforcement learning framework (ReLeQ). We adapt policy optimization methods to the problem of quantization, and focus on finding the best design decisions in choosing the state and action spaces, network architecture and training framework, as well as the tuning of various hyperparamters. We show how ReLeQ can balance speed and quality, and provide an asymmetric general solution for quantization of a large variety of deep networks (AlexNet, CIFAR-10, LeNet, MobileNet-V1, ResNet-20, SVHN, and VGG-11) that virtually preserves the accuracy (=< 0.3% loss) while minimizing the computation and storage cost. With these DNNs, ReLeQ enables conventional hardware to achieve 2.2x speedup over 8-bit execution. Similarly, a custom DNN accelerator achieves 2.0x speedup and energy reduction compared to 8-bit runs. These encouraging results mark ReLeQ as the initial step towards automating the deep quantization of neural networks.

Amir Yazdanbakhsh, Hajar Falahati, Philip J. Wolfe et al.

Generative Adversarial Networks (GANs) are one of the most recent deep learning models that generate synthetic data from limited genuine datasets. GANs are on the frontier as further extension of deep learning into many domains (e.g., medicine, robotics, content synthesis) requires massive sets of labeled data that is generally either unavailable or prohibitively costly to collect. Although GANs are gaining prominence in various fields, there are no accelerators for these new models. In fact, GANs leverage a new operator, called transposed convolution, that exposes unique challenges for hardware acceleration. This operator first inserts zeros within the multidimensional input, then convolves a kernel over this expanded array to add information to the embedded zeros. Even though there is a convolution stage in this operator, the inserted zeros lead to underutilization of the compute resources when a conventional convolution accelerator is employed. We propose the GANAX architecture to alleviate the sources of inefficiency associated with the acceleration of GANs using conventional convolution accelerators, making the first GAN accelerator design possible. We propose a reorganization of the output computations to allocate compute rows with similar patterns of zeros to adjacent processing engines, which also avoids inconsequential multiply-adds on the zeros. This compulsory adjacency reclaims data reuse across these neighboring processing engines, which had otherwise diminished due to the inserted zeros. The reordering breaks the full SIMD execution model, which is prominent in convolution accelerators. Therefore, we propose a unified MIMD-SIMD design for GANAX that leverages repeated patterns in the computation to create distinct microprograms that execute concurrently in SIMD mode.