Showing 1–11 of 11 results for author: Damani, S

Improving Efficiency of GPU Kernel Optimization Agents using a Domain-Specific Language and Speed-of-Light Guidance

Authors: Siva Kumar Sastry Hari, Vignesh Balaji, Sana Damani, Qijing Huang, Christos Kozyrakis

Abstract: Optimizing GPU kernels with LLM agents is an iterative process over a large design space. Every candidate must be generated, compiled, validated, and profiled, so fewer trials will save both runtime and cost. We make two key observations. First, the abstraction level that agents operate at is important. If it is too low, the LLM wastes reasoning on low-impact details. If it is too high, it may mis… ▽ More Optimizing GPU kernels with LLM agents is an iterative process over a large design space. Every candidate must be generated, compiled, validated, and profiled, so fewer trials will save both runtime and cost. We make two key observations. First, the abstraction level that agents operate at is important. If it is too low, the LLM wastes reasoning on low-impact details. If it is too high, it may miss important optimization choices. Second, agents cannot easily tell when they reach the point of diminishing returns, wasting resources as they continue searching. These observations motivate two design principles to improve efficiency: (1) a compact domain-specific language (DSL) that can be learned in context and lets the model reason at a higher level while preserving important optimization levers, and (2) Speed-of-Light (SOL) guidance that uses first-principles performance bounds to steer and budget search. We implement these principles in $μ$CUTLASS, a DSL with a compiler for CUTLASS-backed GPU kernels that covers kernel configuration, epilogue fusion, and multi-stage pipelines. We use SOL guidance to estimate headroom and guide optimization trials, deprioritize problems that are near SOL, and flag kernels that game the benchmark. On 59 KernelBench problems with the same iteration budgets, switching from generating low-level code to DSL code using GPT-5-mini turns a 0.40x geomean regression into a 1.27x speedup over PyTorch. Adding SOL-guided steering raises this to 1.56x. Across model tiers, $μ$CUTLASS + SOL-guidance lets weaker models outperform stronger baseline agents at lower token cost. SOL-guided budgeting saves 19-43% of tokens while retaining at least 95% of geomean speedup, with the best policy reaching a 1.68x efficiency gain. Lastly, SOL analysis helps detect benchmark-gaming cases, where kernels may appear fast while failing to perform the intended computation. △ Less

Submitted 30 March, 2026; originally announced March 2026.

SOL-ExecBench: Speed-of-Light Benchmarking for Real-World GPU Kernels Against Hardware Limits

Authors: Edward Lin, Sahil Modi, Siva Kumar Sastry Hari, Qijing Huang, Zhifan Ye, Nestor Qin, Fengzhe Zhou, Yuan Zhang, Jingquan Wang, Sana Damani, Dheeraj Peri, Ouye Xie, Aditya Kane, Moshe Maor, Michael Behar, Triston Cao, Rishabh Mehta, Vartika Singh, Vikram Sharma Mailthody, Terry Chen, Zihao Ye, Hanfeng Chen, Tianqi Chen, Vinod Grover, Wei Chen , et al. (8 additional authors not shown)

Abstract: As agentic AI systems become increasingly capable of generating and optimizing GPU kernels, progress is constrained by benchmarks that reward speedup over software baselines rather than proximity to hardware-efficient execution. We present SOL-ExecBench, a benchmark of 235 CUDA kernel optimization problems extracted from 124 production and emerging AI models spanning language, diffusion, vision, a… ▽ More As agentic AI systems become increasingly capable of generating and optimizing GPU kernels, progress is constrained by benchmarks that reward speedup over software baselines rather than proximity to hardware-efficient execution. We present SOL-ExecBench, a benchmark of 235 CUDA kernel optimization problems extracted from 124 production and emerging AI models spanning language, diffusion, vision, audio, video, and hybrid architectures, targeting NVIDIA Blackwell GPUs. The benchmark covers forward and backward workloads across BF16, FP8, and NVFP4, including kernels whose best performance is expected to rely on Blackwell-specific capabilities. Unlike prior benchmarks that evaluate kernels primarily relative to software implementations, SOL-ExecBench measures performance against analytically derived Speed-of-Light (SOL) bounds computed by SOLAR, our pipeline for deriving hardware-grounded SOL bounds, yielding a fixed target for hardware-efficient optimization. We report a SOL Score that quantifies how much of the gap between a release-defined scoring baseline and the hardware SOL bound a candidate kernel closes. To support robust evaluation of agentic optimizers, we additionally provide a sandboxed harness with GPU clock locking, L2 cache clearing, isolated subprocess execution, and static analysis based checks against common reward-hacking strategies. SOL-ExecBench reframes GPU kernel benchmarking from beating a mutable software baseline to closing the remaining gap to hardware Speed-of-Light. △ Less

Submitted 19 March, 2026; originally announced March 2026.

KernelBlaster: Continual Cross-Task CUDA Optimization via Memory-Augmented In-Context Reinforcement Learning

Authors: Kris Shengjun Dong, Sahil Modi, Dima Nikiforov, Sana Damani, Edward Lin, Siva Kumar Sastry Hari, Christos Kozyrakis

Abstract: Optimizing CUDA code across multiple generations of GPU architectures is challenging, as achieving peak performance requires an extensive exploration of an increasingly complex, hardware-specific optimization space. Traditional compilers are constrained by fixed heuristics, whereas finetuning Large Language Models (LLMs) can be expensive. However, agentic workflows for CUDA code optimization have… ▽ More Optimizing CUDA code across multiple generations of GPU architectures is challenging, as achieving peak performance requires an extensive exploration of an increasingly complex, hardware-specific optimization space. Traditional compilers are constrained by fixed heuristics, whereas finetuning Large Language Models (LLMs) can be expensive. However, agentic workflows for CUDA code optimization have limited ability to aggregate knowledge from prior exploration, leading to biased sampling and suboptimal solutions. We propose KernelBlaster, a Memory-Augmented In-context Reinforcement Learning (MAIC-RL) framework designed to improve CUDA optimization search capabilities of LLM-based GPU coding agents. KernelBlaster enables agents to learn from experience and make systematically informed decisions on future tasks by accumulating knowledge into a retrievable Persistent CUDA Knowledge Base. We propose a novel profile-guided, textual-gradient-based agentic flow for CUDA generation and optimization to achieve high performance across generations of GPU architectures. KernelBlaster guides LLM agents to systematically explore high-potential optimization strategies beyond naive rewrites. Compared to the PyTorch baseline, our method achieves geometric mean speedups of 1.43x, 2.50x, and 1.50x on KernelBench Levels 1, 2, and 3, respectively. We release KernelBlaster as an open-source agentic framework, accompanied by a test harness, verification components, and a reproducible evaluation pipeline. △ Less

Submitted 15 February, 2026; originally announced February 2026.

Comments: 15 pages, 33 pages with appendix

ProofWright: Towards Agentic Formal Verification of CUDA

Authors: Bodhisatwa Chatterjee, Drew Zagieboylo, Sana Damani, Siva Hari, Christos Kozyrakis

Abstract: Large Language Models (LLMs) are increasingly used to automatically generate optimized CUDA kernels, substantially improving developer productivity. However, despite rapid generation, these kernels often contain subtle correctness bugs and lack formal safety guarantees. Runtime testing is inherently unreliable - limited input coverage and reward hacking can mask incorrect behavior - while manual f… ▽ More Large Language Models (LLMs) are increasingly used to automatically generate optimized CUDA kernels, substantially improving developer productivity. However, despite rapid generation, these kernels often contain subtle correctness bugs and lack formal safety guarantees. Runtime testing is inherently unreliable - limited input coverage and reward hacking can mask incorrect behavior - while manual formal verification is reliable but cannot scale to match LLM output rates, creating a critical validation bottleneck. We present ProofWright, an agentic verification framework that bridges this gap by integrating automated formal verification with LLM-based code generation. ProofWright provides end-to-end guarantees of memory safety, thread safety, and semantic correctness for LLM-generated CUDA kernels. On KernelBench L1, ProofWright verifies safety properties for 74% of generated kernels, uncovers subtle correctness errors missed by conventional testing, and establishes semantic equivalence for a class of element-wise kernels. With a modest overhead of 3 minutes per kernel, ProofWright demonstrates that scalable, automated formal verification of LLM-generated GPU code is feasible - offering a path toward trustworthy high-performance code generation without sacrificing developer productivity. △ Less

Submitted 18 March, 2026; v1 submitted 15 November, 2025; originally announced November 2025.

GPUArmor: A Hardware-Software Co-design for Efficient and Scalable Memory Safety on GPUs

Authors: Mohamed Tarek Ibn Ziad, Sana Damani, Mark Stephenson, Stephen W. Keckler, Aamer Jaleel

Abstract: Memory safety errors continue to pose a significant threat to current computing systems, and graphics processing units (GPUs) are no exception. A prominent class of memory safety algorithms is allocation-based solutions. The key idea is to maintain each allocation's metadata (base address and size) in a disjoint table and retrieve it at runtime to verify memory accesses. While several previous sol… ▽ More Memory safety errors continue to pose a significant threat to current computing systems, and graphics processing units (GPUs) are no exception. A prominent class of memory safety algorithms is allocation-based solutions. The key idea is to maintain each allocation's metadata (base address and size) in a disjoint table and retrieve it at runtime to verify memory accesses. While several previous solutions have adopted allocation-based algorithms (e.g., cuCatch and GPUShield), they typically suffer from high memory overheads or scalability problems. In this work, we examine the key characteristics of real-world GPU workloads and observe several differences between GPU and CPU applications regarding memory access patterns, memory footprint, number of live allocations, and active allocation working set. Our observations motivate GPUArmor, a hardware-software co-design framework for memory safety on GPUs. We show that a simple compiler analysis combined with lightweight hardware support using a small Memory Lookaside Buffer (MLB) can help prevent spatial and temporal memory violations on modern GPU workloads with 2.3% average run time overheads. More importantly, GPUArmor achieves speed-of-light performance with negligible storage requirements. This result benefits both base and bounds solutions and memory tagging techniques, which we showcase with GPUArmor-HWOnly, a variation of GPUArmor that does not require recompilation, and achieves 2.2% slowdowns while significantly reducing storage overheads beyond traditional memory tagging approaches. △ Less

Submitted 25 February, 2025; v1 submitted 24 February, 2025; originally announced February 2025.

Comments: arXiv version of submission

AutoTask: Task Aware Multi-Faceted Single Model for Multi-Task Ads Relevance

Authors: Shouchang Guo, Sonam Damani, Keng-hao Chang

Abstract: Ads relevance models are crucial in determining the relevance between user search queries and ad offers, often framed as a classification problem. The complexity of modeling increases significantly with multiple ad types and varying scenarios that exhibit both similarities and differences. In this work, we introduce a novel multi-faceted attention model that performs task aware feature combination… ▽ More Ads relevance models are crucial in determining the relevance between user search queries and ad offers, often framed as a classification problem. The complexity of modeling increases significantly with multiple ad types and varying scenarios that exhibit both similarities and differences. In this work, we introduce a novel multi-faceted attention model that performs task aware feature combination and cross task interaction modeling. Our technique formulates the feature combination problem as "language" modeling with auto-regressive attentions across both feature and task dimensions. Specifically, we introduce a new dimension of task ID encoding for task representations, thereby enabling precise relevance modeling across diverse ad scenarios with substantial improvement in generality capability for unseen tasks. We demonstrate that our model not only effectively handles the increased computational and maintenance demands as scenarios proliferate, but also outperforms generalized DNN models and even task-specific models across a spectrum of ad applications using a single unified model. △ Less

Submitted 9 July, 2024; originally announced July 2024.

LoPT: Low-Rank Prompt Tuning for Parameter Efficient Language Models

Authors: Shouchang Guo, Sonam Damani, Keng-hao Chang

Abstract: In prompt tuning, a prefix or suffix text is added to the prompt, and the embeddings (soft prompts) or token indices (hard prompts) of the prefix/suffix are optimized to gain more control over language models for specific tasks. This approach eliminates the need for hand-crafted prompt engineering or explicit model fine-tuning. Prompt tuning is significantly more parameter-efficient than model fin… ▽ More In prompt tuning, a prefix or suffix text is added to the prompt, and the embeddings (soft prompts) or token indices (hard prompts) of the prefix/suffix are optimized to gain more control over language models for specific tasks. This approach eliminates the need for hand-crafted prompt engineering or explicit model fine-tuning. Prompt tuning is significantly more parameter-efficient than model fine-tuning, as it involves optimizing partial inputs of language models to produce desired outputs. In this work, we aim to further reduce the amount of trainable parameters required for a language model to perform well on specific tasks. We propose Low-rank Prompt Tuning (LoPT), a low-rank model for prompts that achieves efficient prompt optimization. The proposed method demonstrates similar outcomes to full parameter prompt tuning while reducing the number of trainable parameters by a factor of 5. It also provides promising results compared to the state-of-the-art methods that would require 10 to 20 times more parameters. △ Less

Submitted 27 June, 2024; originally announced June 2024.

VEGETA: Vertically-Integrated Extensions for Sparse/Dense GEMM Tile Acceleration on CPUs

Authors: Geonhwa Jeong, Sana Damani, Abhimanyu Rajeshkumar Bambhaniya, Eric Qin, Christopher J. Hughes, Sreenivas Subramoney, Hyesoon Kim, Tushar Krishna

Abstract: Deep Learning (DL) acceleration support in CPUs has recently gained a lot of traction, with several companies (Arm, Intel, IBM) announcing products with specialized matrix engines accessible via GEMM instructions. CPUs are pervasive and need to handle diverse requirements across DL workloads running in edge/HPC/cloud platforms. Therefore, as DL workloads embrace sparsity to reduce the computations… ▽ More Deep Learning (DL) acceleration support in CPUs has recently gained a lot of traction, with several companies (Arm, Intel, IBM) announcing products with specialized matrix engines accessible via GEMM instructions. CPUs are pervasive and need to handle diverse requirements across DL workloads running in edge/HPC/cloud platforms. Therefore, as DL workloads embrace sparsity to reduce the computations and memory size of models, it is also imperative for CPUs to add support for sparsity to avoid under-utilization of the dense matrix engine and inefficient usage of the caches and registers. This work presents VEGETA, a set of ISA and microarchitecture extensions over dense matrix engines to support flexible structured sparsity for CPUs, enabling programmable support for diverse DL models with varying degrees of sparsity. Compared to the state-of-the-art (SOTA) dense matrix engine in CPUs, a VEGETA engine provides 1.09x, 2.20x, 3.74x, and 3.28x speed-ups when running 4:4 (dense), 2:4, 1:4, and unstructured (95%) sparse DNN layers. △ Less

Submitted 23 February, 2023; v1 submitted 16 February, 2023; originally announced February 2023.

Comments: This paper is accepted to HPCA 2023

Using AI for Economic Upliftment of Handicraft Industry

Authors: Nitya Raviprakash, Sonam Damani, Ankush Chatterjee, Meghana Joshi, Puneet Agrawal

Abstract: The handicraft industry is a strong pillar of Indian economy which provides large-scale employment opportunities to artisans in rural and underprivileged communities. However, in this era of globalization, diverse modern designs have rendered traditional designs old and monotonous, causing an alarming decline of handicraft sales. For this age-old industry to survive the global competition, it is i… ▽ More The handicraft industry is a strong pillar of Indian economy which provides large-scale employment opportunities to artisans in rural and underprivileged communities. However, in this era of globalization, diverse modern designs have rendered traditional designs old and monotonous, causing an alarming decline of handicraft sales. For this age-old industry to survive the global competition, it is imperative to integrate contemporary designs with Indian handicrafts. In this paper, we use novel AI techniques to generate contemporary designs for two popular Indian handicrafts - Ikat and Block Print. These techniques were successfully employed by communities across India to manufacture and sell products with greater appeal and revenue. The designs are evaluated to be significantly more likeable and marketable than the current designs used by artisans. △ Less

Submitted 31 May, 2019; originally announced July 2019.

Using AI to Design Stone Jewelry

Authors: Khyatti Gupta, Sonam Damani, Kedhar Nath Narahari

Abstract: Jewelry has been an integral part of human culture since ages. One of the most popular styles of jewelry is created by putting together precious and semi-precious stones in diverse patterns. While technology is finding its way in the production process of such jewelry, designing it remains a time-consuming and involved task. In this paper, we propose a unique approach using optimization methods co… ▽ More Jewelry has been an integral part of human culture since ages. One of the most popular styles of jewelry is created by putting together precious and semi-precious stones in diverse patterns. While technology is finding its way in the production process of such jewelry, designing it remains a time-consuming and involved task. In this paper, we propose a unique approach using optimization methods coupled with machine learning techniques to generate novel stone jewelry designs at scale. Our evaluation shows that designs generated by our approach are highly likeable and visually appealing. △ Less

Submitted 21 November, 2018; originally announced November 2018.

Ruuh: A Deep Learning Based Conversational Social Agent

Authors: Sonam Damani, Nitya Raviprakash, Umang Gupta, Ankush Chatterjee, Meghana Joshi, Khyatti Gupta, Kedhar Nath Narahari, Puneet Agrawal, Manoj Kumar Chinnakotla, Sneha Magapu, Abhishek Mathur

Abstract: Dialogue systems and conversational agents are becoming increasingly popular in the modern society but building an agent capable of holding intelligent conversation with its users is a challenging problem for artificial intelligence. In this demo, we demonstrate a deep learning based conversational social agent called "Ruuh" (facebook.com/Ruuh) designed by a team at Microsoft India to converse on… ▽ More Dialogue systems and conversational agents are becoming increasingly popular in the modern society but building an agent capable of holding intelligent conversation with its users is a challenging problem for artificial intelligence. In this demo, we demonstrate a deep learning based conversational social agent called "Ruuh" (facebook.com/Ruuh) designed by a team at Microsoft India to converse on a wide range of topics. Ruuh needs to think beyond the utilitarian notion of merely generating "relevant" responses and meet a wider range of user social needs, like expressing happiness when user's favorite team wins, sharing a cute comment on showing the pictures of the user's pet and so on. The agent also needs to detect and respond to abusive language, sensitive topics and trolling behavior of the users. Many of these problems pose significant research challenges which will be demonstrated in our demo. Our agent has interacted with over 2 million real world users till date which has generated over 150 million user conversations. △ Less

Submitted 22 October, 2018; originally announced October 2018.

Comments: 2 pages, 1 figure