arXiv subjects

Jack Zhang

Publications and source records attributed to Jack Zhang.

10 recordsLinked to original sources

arXiv2026

Dion3: Full-Stack Orthogonal Updates

The Muon optimizer incurs a significant overhead cost due to its cubic-time Newton-Schulz orthogonalization step. When weights are sharded, communication overhead compounds this computational cost, eroding the benefits of Muon in many settings. We present Dion3, a revision of Muon that targets this overhead at every level of the stack. Our Gram Newton-Schulz algorithm reduces the FLOP cost of orthogonalization, our CuteDSL kernels accelerate it by exploiting symmetry, and our megabatching strategy reduces communication overhead. Moreover, we propose a simple change to the update rule that cuts costs even further: selecting only a fraction of the momentum matrix's rows to orthogonalize at each step. This update rule improves on Dion (another "compressed" version of Muon), in both speed and performance. Overall, Dion3 matches or improves on the loss achieved by Muon but reduces optimizer step time by up to 6x. Dion3 is available via the dion package (https://github.com/microsoft/dion) as a drop-in replacement for Muon.

cs.LG↗

arXiv2026

CODA: Rewriting Transformer Blocks as GEMM-Epilogue Programs

Transformer training systems are built around dense linear algebra, yet a nontrivial fraction of end-to-end time is spent on surrounding memory-bound operators. Normalization, activations, residual updates, reductions, and related computations repeatedly move large intermediate tensors through global memory while performing little arithmetic, making data movement an increasingly important bottleneck in otherwise highly optimized training stacks. We introduce CODA, a GPU kernel abstraction that expresses these computations as GEMM-plus-epilogue programs. CODA is based on the observation that many Transformer operators exposed as separate framework kernels can be algebraically reparameterized to execute while a GEMM output tile remains on chip, before it is written to memory. The abstraction fixes the GEMM mainloop and exposes a small set of composable epilogue primitives for scaling, reductions, pairwise transformations, and accumulation. This constrained interface preserves the performance structure of expert-written GEMMs while remaining expressive enough to cover nearly all non-attention computation in the forward and backward pass of a standard Transformer block. Across representative Transformer workloads, both human- and LLM-authored CODA kernels achieve high performance, suggesting that GEMM-plus-epilogue programming offers a practical path toward combining framework-level productivity with hardware-level efficiency.

cs.LG↗

arXiv2026

ExecuTorch -- A Unified PyTorch Solution to Run AI Models On-Device

Local execution of AI on edge devices is important for low latency and offline operation. However, deploying models on diverse hardware remains fragmented, often requiring model conversion or complete reimplementation outside the PyTorch ecosystem where the model was originally authored. We introduce ExecuTorch, a unified PyTorch-native deployment framework for edge AI. ExecuTorch enables seamless deployment of machine learning models across heterogeneous compute environments. It scales from embedded microcontrollers to complex system-on-chips (SoCs) with dedicated accelerators, powering devices ranging from wearables and smartphones to large compute clusters. ExecuTorch preserves PyTorch semantics while allowing customization, support for optimizations like quantization, and pluggable execution "backends". These features together enable fast experimentation, allowing researchers to validate deployment behavior entirely within PyTorch, bridging the gap between research and production.

cs.LG↗

arXiv2024

The Llama 3 Herd of Models

Modern artificial intelligence (AI) systems are powered by foundation models. This paper presents a new set of foundation models, called Llama 3. It is a herd of language models that natively support multilinguality, coding, reasoning, and tool usage. Our largest model is a dense Transformer with 405B parameters and a context window of up to 128K tokens. This paper presents an extensive empirical evaluation of Llama 3. We find that Llama 3 delivers comparable quality to leading language models such as GPT-4 on a plethora of tasks. We publicly release Llama 3, including pre-trained and post-trained versions of the 405B parameter language model and our Llama Guard 3 model for input and output safety. The paper also presents the results of experiments in which we integrate image, video, and speech capabilities into Llama 3 via a compositional approach. We observe this approach performs competitively with the state-of-the-art on image, video, and speech recognition tasks. The resulting models are not yet being broadly released as they are still under development.

cs.AI↗

arXiv2024

Dynamic Embeddings with Task-Oriented prompting

This paper introduces Dynamic Embeddings with Task-Oriented prompting (DETOT), a novel approach aimed at improving the adaptability and efficiency of machine learning models by implementing a flexible embedding layer. Unlike traditional static embeddings [14], DETOT dynamically adjusts embeddings based on task-specific requirements and performance feedback, optimizing input data representation for individual tasks [4]. This method enhances both accuracy and computational performance by tailoring the representation layer to meet the unique needs of each task. The structure of DETOT is detailed, highlighting its task-specific adaptation, continuous feedback loop, and mechanisms for preventing overfitting. Empirical evaluations demonstrate its superiority over existing methods.

cs.CL↗

arXiv2024

Case Study: Testing Model Capabilities in Some Reasoning Tasks

Large Language Models (LLMs) excel in generating personalized content and facilitating interactive dialogues, showcasing their remarkable aptitude for a myriad of applications. However, their capabilities in reasoning and providing explainable outputs, especially within the context of reasoning abilities, remain areas for improvement. In this study, we delve into the reasoning abilities of LLMs, highlighting the current challenges and limitations that hinder their effectiveness in complex reasoning scenarios.

cs.CL↗

arXiv2024

Improving Agent Interactions in Virtual Environments with Language Models

Enhancing AI systems with efficient communication skills for effective human assistance necessitates proactive initiatives from the system side to discern specific circumstances and interact aptly. This research focuses on a collective building assignment in the Minecraft dataset, employing language modeling to enhance task understanding through state-of-the-art methods. These models focus on grounding multi-modal understanding and task-oriented dialogue comprehension tasks, providing insights into their interpretative and responsive capabilities. Our experimental results showcase a substantial improvement over existing methods, indicating a promising direction for future research in this domain.

cs.CL↗

arXiv2022

The Renovated Thacher Observatory and First Science Results

Located on the campus of the Thacher School in Southern California, the Thacher Observatory has a legacy of astronomy research and education that dates back to the late 1950's. In 2016, the observatory was fully renovated with upgrades including a new 0.7-m telescope, a research grade camera, and a slit dome with full automation capabilities. The low-elevation site is bordered by the Los Padres National Forest and therefore affords dark to very dark skies allowing for accurate and precise photometric observations. We present a characterization of the site including sky brightness, weather, and seeing, and we demonstrate the on-sky performance of the facility. Our primary research programs are based around our multi-band photometric capabilities and include photometric monitoring of variable sources, a nearby supernova search and followup program, a quick response transient followup effort, and exoplanet and eclipsing binary light curves. Select results from these programs are included in this work which highlight the broad range of science available to an automated observatory with a moderately sized telescope.

astro-ph.IM↗

arXiv2021

Physics Driven Domain Specific Transporter Framework with Attention Mechanism for Ultrasound Imaging

Most applications of deep learning techniques in medical imaging are supervised and require a large number of labeled data which is expensive and requires many hours of careful annotation by experts. In this paper, we propose an unsupervised, physics driven domain specific transporter framework with an attention mechanism to identify relevant key points with applications in ultrasound imaging. The proposed framework identifies key points that provide a concise geometric representation highlighting regions with high structural variation in ultrasound videos. We incorporate physics driven domain specific information as a feature probability map and use the radon transform to highlight features in specific orientations. The proposed framework has been trained on130 Lung ultrasound (LUS) videos and 113 Wrist ultrasound (WUS) videos and validated on 100 Lung ultrasound (LUS) videos and 58 Wrist ultrasound (WUS) videos acquired from multiple centers across the globe. Images from both datasets were independently assessed by experts to identify clinically relevant features such as A-lines, B-lines and pleura from LUS and radial metaphysis, radial epiphysis and carpal bones from WUS videos. The key points detected from both datasets showed high sensitivity (LUS = 99\% , WUS = 74\%) in detecting the image landmarks identified by experts. Also, on employing for classification of the given lung image into normal and abnormal classes, the proposed approach, even with no prior training, achieved an average accuracy of 97\% and an average F1-score of 95\% respectively on the task of co-classification with 3 fold cross-validation. With the purely unsupervised nature of the proposed approach, we expect the key point detection approach to increase the applicability of ultrasound in various examination performed in emergency and point of care.

eess.IV↗

arXiv2021

Domain Specific Transporter Framework to Detect Fractures in Ultrasound

Ultrasound examination for detecting fractures is ideally suited for Emergency Departments (ED) as it is relatively fast, safe (from ionizing radiation), has dynamic imaging capability and is easily portable. High interobserver variability in manual assessment of ultrasound scans has piqued research interest in automatic assessment techniques using Deep Learning (DL). Most DL techniques are supervised and are trained on large numbers of labeled data which is expensive and requires many hours of careful annotation by experts. In this paper, we propose an unsupervised, domain specific transporter framework to identify relevant keypoints from wrist ultrasound scans. Our framework provides a concise geometric representation highlighting regions with high structural variation in a 3D ultrasound (3DUS) sequence. We also incorporate domain specific information represented by instantaneous local phase (LP) which detects bone features from 3DUS. We validate the technique on 3DUS videos obtained from 30 subjects. Each ultrasound scan was independently assessed by three readers to identify fractures along with the corresponding x-ray. Saliency of keypoints detected in the image\ are compared against manual assessment based on distance from relevant features.The transporter neural network was able to accurately detect 180 out of 250 bone regions sampled from wrist ultrasound videos. We expect this technique to increase the applicability of ultrasound in fracture detection.

eess.IV↗