Searcharxiv⌕ Search

arXiv subjects

Nandita Vijaykumar

Publications and source records attributed to Nandita Vijaykumar.

At least 19 recordsLinked to original sources

CertiSum: Bitwise-Identical MXFP4 Training Across Platforms and Parallelism Layouts

Training with MXFP4, a 4-bit microscaling format, is non-deterministic: identical inputs can produce different results across parallelism layouts and GPU vendors. These differences complicate debugging and training audits and can cause training and inference engines to disagree. We present CertiSum to make MXFP4 training bitwise invariant across runs, parallelism layouts, and GPU vendors. Its key insight is that native FP32 accumulation of MXFP4 products is often already exact for every summation order. CertiSum collects statistics during quantization to certify native matrix multiplication results, recomputes uncertified outputs exactly, and uses exact kernels directly where certification rarely succeeds. CertiSum keeps quantization blocks unchanged across layouts and merges partial sums without rounding, including in attention and weight-gradient computation. In 200-step Qwen3-8B-Base experiments, CertiSum preserves training states bit for bit across AMD MI350X and NVIDIA GB200 and across changes in data, tensor, and context parallelism on MI350X. Across three 8B--14B models on FineWeb-Edu using MI350X, it adds 11.8--60.5% to training-step latency relative to its non-exact MXFP4 counterpart at the larger evaluated input shapes, with speedups of up to $1.24\times$ over BF16 and $1.70\times$ over deterministic BF16.

cs.LG↗

FOCUS: From Privileged States to RGB-D with Controlled Modality Switching and Representation Alignment

Vision-based reinforcement learning for robotic manipulation is sample-inefficient because RGB-D observations are high-dimensional and noisy. Privileged state information available in simulation can accelerate training, but its absence at test time creates a train-test modality gap. We propose FOCUS, a single-stage PPO framework that trains the critic on privileged state while automatically regulating whether the actor collects rollouts from RGB-D or privileged-state latents. Regulation is driven by the KL divergence between the action distributions induced by the two modalities, while representation alignment encourages consistent action selection across them. Together, these mechanisms limit RGB-D rollouts when the actor's action distributions from RGB-D and privileged-state latents disagree. As they align, RGB-D exposure increases, shifting on-policy training toward the RGB-D inputs used at test time. Across five manipulation tasks, FOCUS raises average test success from 0.71 to 0.93 relative to the strongest RGB-D-at-test baseline on each task. When accounting for each method's complete training pipeline, budget-normalized training-success AUC increases from 0.47 to 0.65. On Pick-and-Place, test success rises from 0.47 to 0.86, while AUC increases from 0.12 to 0.61, a 5.0x improvement in learning efficiency over the fixed interaction budget.

cs.RO↗

Activation Sparsity with Weight Approximation for Faster LLM Decoding on Offloaded Weights

Deploying LLMs on consumer-grade GPUs with insufficient memory to hold their weights can result in prohibitively slow inference, because decoding repeatedly transfers offloaded weights from system RAM or flash storage into GPU at much lower bandwidth than local GPU-memory access. Activation sparsity reduces these transfers by skipping weights associated with zero or near-zero activations. However, as more activation contributions are omitted, model quality eventually degrades rapidly, indicating that weights associated with small-magnitude activations collectively influence model quality sharply. In this work, we improve the trade-off between model quality and decoding performance when exploiting activation sparsity. Our key idea is to replace the binary choice of whether or not to read a weight with three options: fully retain it, approximate it using a compressed weight representation, or omit it entirely. SpAx skips weights associated with activations closest to zero, reads approximate weights for smaller-magnitude activations, and reads original weights for the largest-magnitude activations. Smaller-magnitude activations attenuate the errors introduced by approximate weights, while compressed weight representations require fewer bytes to be transferred. With weights offloaded to CPU memory, SpAx speeds up decoding by 3.86X on average (up to 5.57X) with 16-bit weights and 2.06X (up to 2.74X) with 4-bit weights, at a WikiText-2 perplexity increase of at most 10%. With weights offloaded to flash storage, the speedups are 3.31X on average (up to 4.81X) and 1.54X (up to 2.03X).

cs.LG↗

Squeeze3D: Extreme Neural Compression with Latent Space Bridging

We propose Squeeze3D, a novel framework that leverages implicit prior knowledge learnt by existing pre-trained encoders and decoders to compress 3D data at extremely high compression ratios. Our approach bridges the latent spaces between a pre-trained encoder and a pretrained decoder model through trainable mapping networks. Any 3D asset represented as a mesh, point cloud, or radiance field is first encoded by the pre-trained encoder and then transformed (i.e. compressed) into a highly compact latent code by a mapping network. This latent code can effectively be used as an extremely compressed representation of the mesh, point cloud, or radiance field. A mapping network transforms the compressed latent code into the latent space of a powerful generative model; the decoder of this generative model then recreates the original 3D asset (i.e. decompression). Squeeze3D is trained entirely on generated synthetic data and does not require any 3D datasets. The Squeeze3D architecture can be flexibly used with existing pre-trained 3D encoders and existing generative models. It can flexibly support different formats, including meshes, point clouds, and radiance fields. Our experiments demonstrate that Squeeze3D achieves compression ratios of up to 2187$\times$ for textured meshes, 58.5$\times$ for point clouds, and more than 650$\times$ for radiance fields while maintaining visual quality comparable to many existing methods. Squeeze3D only incurs a small compression and decompression latency since it does not involve training object-specific networks to compress an object.

cs.GR↗

KerColle: Unlocking Fine-Grained GPU Concurrency in Vision-Language-Action Models

Vision-Language-Action (VLA) models have emerged as foundational models for next-generation robotics. High VLA inference throughput is critical for meeting the control-rate requirements of robots. VLA models comprise two phases, a vision-language model (VLM) and an action head, that can be decoupled and executed asynchronously and concurrently across independent robot requests. Through a detailed characterization of four state-of-the-art VLAs, we observe that GPUs are severely underutilized in VLA inference as the Cooperative Thread Array (CTA) scheduler of GPU is unable to fully overlap the two independent phases of VLA execution. We identify that this inefficiency is caused by head-of-line blocking in the hardware thread-block dispatcher. We demonstrate that prior scheduling frameworks do not address the challenges posed by VLA concurrency. First, the independent phases across different robot requests each comprise numerous kernels, and at any given time, there are different combinations of kernels that are executed in parallel. This makes static or ahead-of-time scheduling policies largely ineffective. Second, many of the action-head operators are short-running kernels and there are numerous such kernels. This leaves no headroom for online profiling or preemption-based mechanisms. To address these challenges, we present KerColle, a lightweight GPU scheduling framework that leverages online Streaming Multiprocessor (SM) utilization and individual kernel resource requirements to intelligently and dynamically co-schedule kernels to efficiently overlap the two phases of execution by (1) mitigating head-of-line blocking, and (2) co-scheduling kernels with complementary resource requirements. We demonstrate in simulation, across two GPU architectures, for 4 state-of-the-art VLA models, that KerColle delivers throughput gains of up to $28\%$.

cs.AR↗

HIP: Hessian Interatomic Potentials without derivatives

Molecular Hessians, the second derivatives of the potential energy, are fundamental to many workflows in computational chemistry. Usually, accurate Hessians are computationally expensive to calculate and scale poorly with system size, whether computed using quantum chemistry methods or machine-learning interatomic potentials (MLIPs). In this work, we introduce Hessian interatomic potentials (HIPs), a deep learning model that directly predicts Hessians without relying on automatic differentiation or finite differences. To do so, we construct SE(3)-equivariant, symmetric Hessians from irreducible representation (irrep) features up to degree $l$=2, computed by a graph neural network. HIP Hessians are one to two orders of magnitude faster, more accurate, more memory efficient, easier to train, and exhibit more favourable scaling with system size. We validate our predictions across a wide range of downstream tasks, demonstrating consistently superior performance in transition state search, geometry optimization, zero-point energy corrections, and vibrational analysis. We open-source the HIP code and model weights.

cs.LG↗

RevaMp3D: Architecting the Processor Core and Cache Hierarchy for Systems with Monolithically-Integrated Logic and Memory

Recent nano-technological advances enable the Monolithic 3D (M3D) integration of multiple memory and logic layers in a single chip, allowing for fine-grained connections between layers and significantly alleviating main memory bottlenecks. We show for a variety of workloads, on a state-of-the-art M3D-based system, that the performance and energy bottlenecks shift from main memory to the processor core and cache hierarchy. Therefore, there is a need to revisit current designs that have been conventionally tailored to tackle the memory bottleneck. Based on the insights from our design space exploration, we propose RevaMp3D, introducing five key changes. First, we propose removing the shared last-level cache, as this delivers speedups comparable to or exceeding those from increasing its size or reducing its latency across all workloads. Second, since improving L1 cache latency has a large impact on performance, we reduce L1 latency by leveraging an M3D layout to shorten its wires. Third, we repurpose the area from the removed cache to widen and scale up pipeline structures, accommodating more in-flight requests that are efficiently served by M3D memory. To avoid latency penalties from these larger structures, we leverage M3D layouts. Fourth, to facilitate high thread-level parallelism, we propose a new fine-grained synchronization technique, using M3D's dense inter-layer connectivity. Fifth, we leverage the M3D main memory to mitigate the core bottlenecks. We propose a processor frontend design that memoizes the repetitive fetched, decoded, and reordered instructions, stores them in main memory, and turns off the relevant parts of the core when possible. RevaMp3D provides 1.2x-2.9x speedup and 1.2x-1.4x energy reduction compared to a state-of-the-art M3D system. We also analyze RevaMp3D's design decisions across various memory latencies to facilitate latency-aware design decisions.

cs.AR↗

zkComposer: Decomposing Proof Construction to Scale zkML

Zero-knowledge machine learning (zkML) enables a server to perform verifiable inference while keeping model parameters private from the client. However, existing zkML systems incur prohibitive proof-generation costs. We observe that proof generation exhibits limited parallelism; that is, prover time does not decrease significantly as the number of threads increases. This limitation is because existing systems rely on monolithic proof computation, constructing a single proof for the entire machine learning model. We introduce zkComposer, a modular proof-construction framework that unlocks an additional dimension of parallelism, in addition to the parallelism in existing proof kernels. zkComposer decomposes the zkML proof of correct inference into independent sub-proofs, each covering a subset of the computation for inference e.g., each independent sub-proof can cover a subset of contiguous layers in the ML model. Adjacent sub-proofs are cryptographically linked through shared commitments to the activations from the boundary layer. zkComposer provides the same guarantees as the monolithic proof without requiring additional linking proofs or changes to the underlying cryptographic primitives. We implement zkComposer and evaluate it on three CNNs and GPT-2. We show that, on CNN workloads, zkComposer reduces prover time and response time by up to 3.25x relative to zkCNN [1]. On GPT-2, zkComposer reduces these times by up to 4.83x relative to zkGPT [2], when partitioning along the model layers. When partitioning across both model layers and input sequences in GPT-2, we show that zkComposer reduces prover time and response time by up to 6.84x relative to zkGPT [2].

cs.CR↗

DataGuard: Guaranteeing Private Training in Systolic-array Based Accelerators

Differential privacy (DP) and federated learning (FL) have emerged as important privacy-preserving approaches when using sensitive data to train machine learning (ML) models. FL ensures that raw sensitive data does not leave the users' devices by training the model locally on the device. DP ensures that the model does not leak any information about an individual by clipping and adding noise to the gradients before updating the model. It provides formalism to constrain privacy loss during training to a privacy budget determined a priori by the owner of sensitive data. However, real-life deployments of FL algorithms typically assume that a third-party FL application can be trusted to correctly implement DP algorithms. Thus, the third-party application is given full access to sensitive data. In this work, we propose DataGuard, a hardware-based mechanism that guarantees that the only data that can leave the device is the result of computation that meets DP requirements. DataGuard can thus be used to ensure that the privacy budget defined by the data owner is not exceeded during FL training without the need to trust a third-party application. We evaluate DataGuard in simulations of four accelerators for various ML models and demonstrate only small area overheads of less than 0.01\% and performance slowdowns of less than 0.3\%.

cs.AR↗

QPILOTS: Efficient Test-Time Q-Steering for Flow Policies

Flow-matching and diffusion policies are expressive action generators, but optimizing them with temporal-difference reinforcement learning (RL) remains difficult. Effective policy extraction requires exploiting the critic's action gradient, yet directly backpropagating this signal through a multi-step denoising process can be numerically unstable. Existing methods work around this either by discarding gradient information, distilling the policy into a simpler one-step actor, or repeatedly fine-tuning the denoising policy as the critic improves. We propose QPILOTS, a method that leaves the original policy unmodified and steers the denoising process at inference time. At each denoising step, instead of evaluating the critic on the noisy intermediate action where critic predictions are unreliable, we first project that intermediate state to an estimate of the final clean action and compute the critic gradient there. We introduce two variants: QPILOTS-U uses a fast single-point approximation, while QPILOTS-M draws differentiable posterior samples via a learned auxiliary network. On a standard offline-to-online RL benchmark, QPILOTS achieves the best aggregate performance, reaching an average success rate of 90% across 50 tasks. We also apply QPILOTS to steer a large, frozen, pretrained Vision-Language Action (VLA) foundation model, outperforming or matching prior inference-time approaches across six manipulation tasks in simulation.

cs.LG↗

XPR: An Extensible Cross-Platform Point-Based Differentiable Renderer

Point-based differentiable rendering underpins modern 3D reconstruction, novel-view synthesis, and learning-based graphics pipelines, but developing new rendering methods often requires extensive low-level implementation, hardware-specific kernels, and manually written backward passes. This limits rapid prototyping, reproducibility, exploration, and deployment, especially across diverse hardware platforms. This paper presents XPR, an extensible cross-platform framework for point-based differentiable rendering. XPR introduces a high-level programming interface that separates method-specific logic from the shared rendering pipeline, allowing users to implement new methods in a few lines of code. Its pipeline decomposes rendering into modular, statically shaped parallel operations that can be lowered by a cross-platform compiler to GPUs, TPUs, CPUs, and other ML accelerators. We demonstrate implementations of 3DGS, 3DGUT, and LinPrim, with only a few 100s lines of Python code, each of which can be compiled to a range of hardware platforms with the XLA compiler. These results show that XPR enables fast experimentation and portable execution for emerging point-based differentiable rendering systems.

cs.GR↗

FG-Attn: Leveraging Fine-Grained Sparse Attention in Video Diffusion Models

Using diffusion transformers for media generation may require evaluating attention over extremely long sequences, with attention layers accounting for the majority of generation latency. Exploiting sparsity in attention maps offers a promising opportunity to reduce this cost. In this work, we show that attention maps in diffusion transformers exhibit significant fine-grained sparsity in video generation models. Existing sparse attention methods, however, are too coarse-grained, leaving a large fraction of redundant computation unaddressed, or incur high overheads at finer granularity. We propose FG-Attn, a novel, low-overhead fine-grained sparse attention mechanism that skips score computations at the granularity of a MxN tile, where N>=1 and M>=16, and where each block is the result of query-key dot products between M queries and N keys. FG-Attn addresses the key challenge of hardware underutilization in sparse attention kernels on GPUs, without incurring the overheads of irregular memory access and redundant operations. FG-Attn can fully supersede existing sparse attention methods and extend block sparse attention methods to finer granularities on modern GPUs. At 70% sparsity, FG-Attn is up to 2.45X faster than the state-of-art FlashInfer, and reduces attention kernel time by 14.7% on average. FG-Attn speeds up end-to-end video generation times by up to 1.40X (1.18X on average) over Flash Attention 3.

cs.CV↗

CRAFT: Fine-Grained Cost-Aware Expert Replication For Efficient Mixture-of-Experts Serving

Mixture-of-Experts (MoE) has recently emerged as the mainstream architecture for efficiently scaling large language models while maintaining near-constant computational cost. Expert parallelism distributes parameters by partitioning experts across devices, but this introduces token-level load imbalance during inference. Expert replication is a widely adopted load-balancing technique in serving frameworks that alleviates load imbalance in large-scale deployments by replicating experts with high loads. In this work, we demonstrate that existing replication schemes often over-replicate, with many replicas providing marginal improvement. Replicas consume substantial GPU memory, which may lead to resource contention and throughput degradation. We present CRAFT, an efficient expert replication framework that maximizes load balance under a given memory budget by performing fine-grained, per-layer replication based on the estimated replication benefit. CRAFT can be seamlessly integrated into existing serving frameworks without any additional training or model changes. Our evaluation shows that CRAFT increases end-to-end serving throughput by $1.14\times$ on average (up to $1.2\times$) over existing replication techniques in large-scale deployments with models ranging from hundreds of billions to a trillion parameters.

cs.DC↗

SpecBench: Evaluating Specification-Level Reasoning for Software Engineering LLM Agents

Software engineering (SWE) agents are transitioning from code generation to full software development lifecycle automation. A critical phase in this lifecycle is specification design: transforming initial proposals into carefully considered requirements through expert review. Existing benchmarks such as SWE-Bench are implementation-focused by measuring the agent's ability to generate code given fixed, precise design requirements. This formulation assumes specifications are correct and complete. In real-world complex and critical software systems, initial specifications are often incomplete and flawed, requiring extensive expert reviews and revisions before being accepted for implementation. To fill this gap, we introduce SpecBench to evaluate specification-level reasoning: the ability to generate complete, unambiguous, consistent, and correct system specifications. SpecBench tasks are derived from the Request for Comments (RFC) process used by mature open-source projects. For each task, an agent is given an initial design proposal, the project codebase, and all past project RFC discussions. The agent is tasked with identifying specification deficiencies: omissions, ambiguities, inconsistencies, or incorrect assumptions in the initial proposal. We evaluate predictions against critiques raised by expert maintainers during historical RFC reviews. SpecBench contains tasks from 5 diverse repositories: Kubernetes, React, Rust, TVM, and vLLM. We evaluate state-of-the-art SWE agents on SpecBench, analyzing their capacity to reason about system design without execution feedback. The best performing agent, GPT-5.4, achieves 44.4% accuracy.

cs.MA↗

LuxRemix: Lighting Decomposition and Remixing for Indoor Scenes

We present a novel approach for interactive light editing in indoor scenes from a single multi-view scene capture. Our method leverages a generative image-based light decomposition model that factorizes complex indoor scene illumination into its constituent light sources. This factorization enables independent manipulation of individual light sources, specifically allowing control over their state (on/off), chromaticity, and intensity. We further introduce multi-view lighting harmonization to ensure consistent propagation of the lighting decomposition across all scene views. This is integrated into a relightable 3D Gaussian splatting representation, providing real-time interactive control over the individual light sources. Our results demonstrate highly photorealistic lighting decomposition and relighting outcomes across diverse indoor scenes. We evaluate our method on both synthetic and real-world datasets and provide a quantitative and qualitative comparison to state-of-the-art techniques. For video results and interactive demos, see https://luxremix.github.io.

cs.CV↗

Sparse Code Uplifting for Efficient 3D Language Gaussian Splatting

3D Language Gaussian Splatting (3DLGS) augments 3D Gaussian Splatting with language-aligned visual features for open-vocabulary 3D scene understanding. A core challenge is efficiently associating high-dimensional vision-language embeddings with millions of 3D Gaussians while preserving efficient feature rendering for text-based querying. Existing methods either store dense features directly on Gaussians, causing high storage costs and slow rendering, or learn compact representations through expensive per-scene optimization with repeated feature rasterization. No existing method simultaneously achieves fast 3D semantic reconstruction, efficient storage, and fast rendering. We propose SCOUP (Sparse COde UPlifting), which addresses all three by decoupling language representation learning from 3D Gaussian optimization. Rather than working directly in 3D, we learn sparse codebook-based representations entirely using features associated with 2D image regions, associating each region with a sparse set of codebook coefficients. We then uplift these coefficients to 3D Gaussians with our weighted sparse aggregation using Gaussian-to-pixel associations, where each Gaussian accumulates coefficients over codebook atoms across views. Top-$K$ filtering then extracts the most dominant multi-view coefficients per Gaussian, enabling efficient storage and fast rendering. Our method achieves up to $400\times$ training speedup while being $3\times$ more memory efficient during training compared to the state-of-the-art in rendering speed. Across multiple benchmarks, SCOUP matches or outperforms existing methods in open-vocabulary querying accuracy.

cs.CV↗

DPQuant: Efficient and Differentially-Private Model Training via Dynamic Quantization Scheduling

Differentially-Private SGD (DP-SGD) and its adaptive variant DP-Adam are powerful techniques to protect user privacy when using sensitive data to train neural networks. During training, converting model weights and activations into low-precision formats, i.e., quantization, can drastically reduce training times, energy consumption, and cost, and is thus a widely used technique. In this work, we demonstrate for the first time that quantization causes significantly higher accuracy degradation in DP training compared to regular SGD. We observe that this is caused by noise injection, which amplifies quantization variance, leading to disproportionately large accuracy degradation. To address this challenge, we present DPQuant, a dynamic quantization framework that adaptively selects a changing subset of layers to quantize at each epoch. Our method combines two key ideas that effectively reduce quantization variance: (i) probabilistic sampling that rotates which layers are quantized every epoch, and (ii) loss-aware layer prioritization, which uses a differentially private loss sensitivity estimator to identify layers that can be quantized with minimal impact on model quality. This estimator consumes a negligible fraction of the overall privacy budget, preserving DP guarantees. Empirical evaluations on ResNet18, ResNet50, and DenseNet121 across a range of datasets demonstrate that DPQuant consistently outperforms static quantization baselines, achieving near Pareto-optimal accuracy-compute trade-offs and up to $2.21\times$ theoretical throughput improvements on low-precision hardware, with less than 2% drop in validation accuracy. We further show that our framework extends to DP-Adam with similar gains.

cs.LG↗

AccelOpt: A Self-Improving LLM Agentic System for AI Accelerator Kernel Optimization

We present AccelOpt, a self-improving large language model (LLM) agentic system that autonomously optimizes kernels for emerging AI acclerators, eliminating the need for expert-provided hardware-specific optimization knowledge. AccelOpt explores the kernel optimization space through iterative generation, informed by an optimization memory that curates experiences and insights from previously encountered slow-fast kernel pairs. We build NKIBench, a new benchmark suite of AWS Trainium accelerator kernels with varying complexity extracted from real-world LLM workloads to evaluate the effectiveness of AccelOpt. Our evaluation confirms that AccelOpt's capability improves over time, boosting the average percentage of peak throughput from $49\%$ to $61\%$ on Trainium 1 and from $45\%$ to $59\%$ on Trainium 2 for NKIBench kernels. Moreover, AccelOpt is highly cost-effective: using open-source models, it matches the kernel improvements of Claude Sonnet 4 while being $26\times$ cheaper. The code is open-sourced at https://github.com/zhang677/AccelOpt.

cs.LG↗