Jonathan Ragan‐Kelley

Ahmed H. Mahmoud, Rahul Goel, Jonathan Ragan-Kelley et al.

We present a GPU-based system for automatic differentiation (AD) of functions defined on triangle meshes, designed to exploit the locality and sparsity in mesh-based computation. Our system evaluates derivatives using per-element forward-mode AD, confining all computation to registers and shared memory and assembling global gradients, sparse Jacobians, and sparse Hessians directly on the GPU. By avoiding global computation graphs, intermediate buffers, and device-host synchronization, our approach minimizes memory traffic and enables efficient differentiation under both static and dynamically changing sparsity. Our programming model lets users express energy terms over mesh neighborhoods, while our system automatically manages parallel execution, derivative propagation, sparse assembly, and matrix-free operations such as Hessian-vector products. Our system supports both scalar- and vector-valued objectives, dynamic interaction-driven sparsity updates, and seamless integration with external GPU sparse linear solvers. We evaluate our system on applications including elastic and cloth simulation, surface parameterization, mesh smoothing, frame field design, ARAP deformation, and spherical manifold optimization. Across these tasks, our system consistently outperforms state-of-the-art differentiation frameworks, including PyTorch, JAX, Warp, DrJIT, EnzymeAD, and Thallo. We demonstrate speedups across a range of solver types, from Newton and Gauss-Newton for nonlinear least squares to L-BFGS and gradient descent, and across different derivative usage modes, including Hessian-vector products as well as full sparse Hessian and Jacobian construction. Our system is available as open source at https://github.com/owensgroup/RXMesh.

Kartik Chandra, Tzu-Mao Li, Joshua Tenenbaum et al.

We design new visual illusions by finding "adversarial examples" for principled models of human perception -- specifically, for probabilistic models, which treat vision as Bayesian inference. To perform this search efficiently, we design a differentiable probabilistic programming language, whose API exposes MCMC inference as a first-class differentiable function. We demonstrate our method by automatically creating illusions for three features of human vision: color constancy, size constancy, and face perception.

William Brandon, Aniruddha Nrusimha, Kevin Qian et al.

To help address the growing demand for ever-longer sequence lengths in transformer models, Liu et al. recently proposed Ring Attention, an exact attention algorithm capable of overcoming per-device memory bottle- necks by distributing self-attention across multiple devices. In this paper, we study the performance characteristics of Ring Attention in the important special case of causal transformer models, and identify a key workload imbal- ance due to triangular structure of causal attention computations. We propose a simple extension to Ring Attention, which we call Striped Attention to fix this imbalance. Instead of devices having contiguous subsequences, each device has a subset of tokens distributed uniformly throughout the sequence, which we demonstrate leads to more even workloads. In experiments running Striped Attention on A100 GPUs and TPUv4s, we are able to achieve up to 1.45x end-to-end throughput improvements over the original Ring Attention algorithm on causal transformer training at a sequence length of 256k. Furthermore, on 16 TPUv4 chips, we were able to achieve 1.65x speedups at sequence lengths of 786k. We release the code for our experiments as open source

Tian Jin, Nolan Clement, Xin Dong et al.

How does scaling the number of parameters in large language models (LLMs) affect their core capabilities? We study two natural scaling techniques -- weight pruning and simply training a smaller or larger model, which we refer to as dense scaling -- and their effects on two core capabilities of LLMs: (a) recalling facts presented during pre-training and (b) processing information presented in-context during inference. By curating a suite of tasks that help disentangle these two capabilities, we find a striking difference in how these two abilities evolve due to scaling. Reducing the model size by more than 30\% (via either scaling approach) significantly decreases the ability to recall facts seen in pre-training. Yet, a 60--70\% reduction largely preserves the various ways the model can process in-context information, ranging from retrieving answers from a long context to learning parameterized functions from in-context exemplars. The fact that both dense scaling and weight pruning exhibit this behavior suggests that scaling model size has an inherently disparate effect on fact recall and in-context learning.

Matthew Caren, Kartik Chandra, Joshua B. Tenenbaum et al.

We present a method for automatically producing human-like vocal imitations of sounds: the equivalent of "sketching," but for auditory rather than visual representation. Starting with a simulated model of the human vocal tract, we first try generating vocal imitations by tuning the model's control parameters to make the synthesized vocalization match the target sound in terms of perceptually-salient auditory features. Then, to better match human intuitions, we apply a cognitive theory of communication to take into account how human speakers reason strategically about their listeners. Finally, we show through several experiments and user studies that when we add this type of communicative reasoning to our method, it aligns with human intuitions better than matching auditory features alone does. This observation has broad implications for the study of depiction in computer graphics.

Zachary Ankner, Rishab Parthasarathy, Aniruddha Nrusimha et al.

To combat the memory bandwidth-bound nature of autoregressive LLM inference, previous research has proposed the speculative decoding frame-work. To perform speculative decoding, a small draft model proposes candidate continuations of the input sequence that are then verified in parallel by the base model. One way to specify the draft model, as used in the recent Medusa decoding framework, is as a collection of lightweight heads, called draft heads, that operate on the base model's hidden states. To date, all existing draft heads have been sequentially independent, meaning that they speculate tokens in the candidate continuation independently of any preceding tokens in the candidate continuation. In this work, we propose Hydra heads: a sequentially-dependent drop-in replacement for standard draft heads that significantly improves the accuracy of draft head speculation. We further explore the design space of Hydra head training objectives and architectures, and propose a carefully tuned Hydra head recipe, which we call Hydra++, that improves decoding throughput by up to 1.31x and 2.70x compared to Medusa decoding and autoregressive de-coding respectively. Overall, Hydra heads are a simple and well-motivated intervention on standard draft heads that significantly improve the end-to-end speed of draft head-based speculative decoding. We make our code publicly available at https://github.com/zankner/Hydra.

Han Guo, William Brandon, Radostin Cholakov et al.

The deployment of large language models (LLMs) is often constrained by memory bandwidth, where the primary bottleneck is the cost of transferring model parameters from the GPU's global memory to its registers. When coupled with custom kernels that fuse the dequantization and matmul operations, weight-only quantization can thus enable faster inference by reducing the amount of memory movement. However, developing high-performance kernels for weight-quantized LLMs presents substantial challenges, especially when the weights are compressed to non-evenly-divisible bit widths (e.g., 3 bits) with non-uniform, lookup table (LUT) quantization. This paper describes FLUTE, a flexible lookup table engine for LUT-quantized LLMs, which uses offline restructuring of the quantized weight matrix to minimize bit manipulations associated with unpacking, and vectorization and duplication of the lookup table to mitigate shared memory bandwidth constraints. At batch sizes < 32 and quantization group size of 128 (typical in LLM inference), the FLUTE kernel can be 2-4x faster than existing GEMM kernels. As an application of FLUTE, we explore a simple extension to lookup table-based NormalFloat quantization and apply it to quantize LLaMA3 to various configurations, obtaining competitive quantization performance against strong baselines while obtaining an end-to-end throughput increase of 1.5 to 2 times.

Hasan Genc, Seah Kim, Alon Amid et al.

DNN accelerators are often developed and evaluated in isolation without considering the cross-stack, system-level effects in real-world environments. This makes it difficult to appreciate the impact of System-on-Chip (SoC) resource contention, OS overheads, and programming-stack inefficiencies on overall performance/energy-efficiency. To address this challenge, we present Gemmini, an open-source*, full-stack DNN accelerator generator. Gemmini generates a wide design-space of efficient ASIC accelerators from a flexible architectural template, together with flexible programming stacks and full SoCs with shared resources that capture system-level effects. Gemmini-generated accelerators have also been fabricated, delivering up to three orders-of-magnitude speedups over high-performance CPUs on various DNN benchmarks. * https://github.com/ucb-bar/gemmini

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.

Kartik Chandra, Katherine M. Collins, Will Crichton et al.

Often, a good explanation for a program's unexpected behavior is a bug in the programmer's code. But sometimes, an even better explanation is a bug in the programmer's mental model of the language or API they are using. Instead of merely debugging our current code ("giving the programmer a fish"), what if our tools could directly debug our mental models ("teaching the programmer to fish")? In this paper, we apply recent ideas from computational cognitive science to offer a principled framework for doing exactly that. Given a "why?" question about a program, we automatically infer potential misconceptions about the language/API that might cause the user to be surprised by the program's behavior -- and then analyze those misconceptions to provide explanations of the program's behavior. Our key idea is to formally represent misconceptions as counterfactual (erroneous) semantics for the language/API, which can be inferred and debugged using program synthesis techniques. We demonstrate our framework, WatChat, by building systems for explanation in two domains: JavaScript type coercion, and the Git version control system. We evaluate WatChatJS and WatChatGit by comparing their outputs to experimentally-collected human-written explanations in these two domains: we show that WatChat's explanations exhibit key features of human-written explanation, unlike those of a state-of-the-art language model.

Utkarsh Singhal, Brian Cheung, Kartik Chandra et al.

How much can you say about the gradient of a neural network without computing a loss or knowing the label? This may sound like a strange question: surely the answer is "very little." However, in this paper, we show that gradients are more structured than previously thought. Gradients lie in a predictable low-dimensional subspace which depends on the network architecture and incoming features. Exploiting this structure can significantly improve gradient-free optimization schemes based on directional derivatives, which have struggled to scale beyond small networks trained on toy datasets. We study how to narrow the gap in optimization performance between methods that calculate exact gradients and those that use directional derivatives. Furthermore, we highlight new challenges in overcoming the large gap between optimizing with exact gradients and guessing the gradients.

Aniruddha Nrusimha, William Brandon, Mayank Mishra et al.

The size and compute characteristics of modern large language models have led to an increased interest in developing specialized kernels tailored for training and inference. Existing kernels primarily optimize for compute utilization, targeting the large-batch training and inference settings. However, low-batch inference, where memory bandwidth and kernel launch overheads contribute are significant factors, remains important for many applications of interest such as in edge deployment and latency-sensitive applications. This paper describes FlashFormer, a proof-of-concept kernel for accelerating single-batch inference for transformer-based large language models. Across various model sizes and quantizations settings, we observe nontrivial speedups compared to existing state-of-the-art inference kernels.

Kartik Chandra, Tony Chen, Tzu-Mao Li et al.

A single panel of a comic book can say a lot: it can depict not only where the characters currently are, but also their motions, their motivations, their emotions, and what they might do next. More generally, humans routinely infer complex sequences of past and future events from a *static snapshot* of a *dynamic scene*, even in situations they have never seen before. In this paper, we model how humans make such rapid and flexible inferences. Building on a long line of work in cognitive science, we offer a Monte Carlo algorithm whose inferences correlate well with human intuitions in a wide variety of domains, while only using a small, cognitively-plausible number of samples. Our key technical insight is a surprising connection between our inference problem and Monte Carlo path tracing, which allows us to apply decades of ideas from the computer graphics community to this seemingly-unrelated theory of mind task.

Kartik Chandra, Tzu-Mao Li, Josh Tenenbaum et al.

Great storytellers know how to take us on a journey. They direct characters to act -- not necessarily in the most rational way -- but rather in a way that leads to interesting situations, and ultimately creates an impactful experience for audience members looking on. If audience experience is what matters most, then can we help artists and animators *directly* craft such experiences, independent of the concrete character actions needed to evoke those experiences? In this paper, we offer a novel computational framework for such tools. Our key idea is to optimize animations with respect to *simulated* audience members' experiences. To simulate the audience, we borrow an established principle from cognitive science: that human social intuition can be modeled as "inverse planning," the task of inferring an agent's (hidden) goals from its (observed) actions. Building on this model, we treat storytelling as "*inverse* inverse planning," the task of choosing actions to manipulate an inverse planner's inferences. Our framework is grounded in literary theory, naturally capturing many storytelling elements from first principles. We give a series of examples to demonstrate this, with supporting evidence from human subject studies.

Yuka Ikarashi, Jonathan Ragan-Kelley, Tsukasa Fukusato et al.

Writing high-performance image processing code is challenging and labor-intensive. The Halide programming language simplifies this task by decoupling high-level algorithms from "schedules" which optimize their implementation. However, even with this abstraction, it is still challenging for Halide programmers to understand complicated scheduling strategies and productively write valid, optimized schedules. To address this, we propose a programming support method called "guided optimization." Guided optimization provides programmers a set of valid optimization options and interactive feedback about their current choices, which enables them to comprehend and efficiently optimize image processing code without the time-consuming trial-and-error process of traditional text editors. We implemented a proof-of-concept system, Roly-poly, which integrates guided optimization, program visualization, and schedule cost estimation to support the comprehension and development of efficient Halide image processing code. We conducted a user study with novice Halide programmers and confirmed that Roly-poly and its guided optimization was informative, increased productivity, and resulted in higher-performing schedules in less time.

Vaishaal Shankar, Alex Fang, Wenshuo Guo et al.

We investigate the connections between neural networks and simple building blocks in kernel space. In particular, using well established feature space tools such as direct sum, averaging, and moment lifting, we present an algebra for creating "compositional" kernels from bags of features. We show that these operations correspond to many of the building blocks of "neural tangent kernels (NTK)". Experimentally, we show that there is a correlation in test error between neural network architectures and the associated kernels. We construct a simple neural network architecture using only 3x3 convolutions, 2x2 average pooling, ReLU, and optimized with SGD and MSE loss that achieves 96% accuracy on CIFAR10, and whose corresponding compositional kernel achieves 90% accuracy. We also use our constructions to investigate the relative performance of neural networks, NTKs, and compositional kernels in the small dataset regime. In particular, we find that compositional kernels outperform NTKs and neural networks outperform both kernel methods.

Yuanming Hu, Luke Anderson, Tzu-Mao Li et al.

We present DiffTaichi, a new differentiable programming language tailored for building high-performance differentiable physical simulators. Based on an imperative programming language, DiffTaichi generates gradients of simulation steps using source code transformations that preserve arithmetic intensity and parallelism. A light-weight tape is used to record the whole simulation program structure and replay the gradient kernels in a reversed order, for end-to-end backpropagation. We demonstrate the performance and productivity of our language in gradient-based learning and optimization tasks on 10 different physical simulators. For example, a differentiable elastic object simulator written in our language is 4.2x shorter than the hand-engineered CUDA version yet runs as fast, and is 188x faster than the TensorFlow implementation. Using our differentiable programs, neural network controllers are typically optimized within only tens of iterations.

Jing Pu, Steven Bell, Xuan Yang et al.

Specialized image processing accelerators are necessary to deliver the performance and energy efficiency required by important applications in computer vision, computational photography, and augmented reality. But creating, "programming,"and integrating this hardware into a hardware/software system is difficult. We address this problem by extending the image processing language, Halide, so users can specify which portions of their applications should become hardware accelerators, and then we provide a compiler that uses this code to automatically create the accelerator along with the "glue" code needed for the user's application to access this hardware. Starting with Halide not only provides a very high-level functional description of the hardware, but also allows our compiler to generate the complete software program including the sequential part of the workload, which accesses the hardware for acceleration. Our system also provides high-level semantics to explore different mappings of applications to a heterogeneous system, with the added flexibility of being able to map at various throughput rates. We demonstrate our approach by mapping applications to a Xilinx Zynq system. Using its FPGA with two low-power ARM cores, our design achieves up to 6x higher performance and 8x lower energy compared to the quad-core ARM CPU on an NVIDIA Tegra K1, and 3.5x higher performance with 12x lower energy compared to the K1's 192-core GPU.

Xuan Yang, Jing Pu, Blaine Burton Rister et al.

Convolutional Neural Networks (CNNs) are the state of the art solution for many computer vision problems, and many researchers have explored optimized implementations. Most implementations heuristically block the computation to deal with the large data sizes and high data reuse of CNNs. This paper explores how to block CNN computations for memory locality by creating an analytical model for CNN-like loop nests. Using this model we automatically derive optimized blockings for common networks that improve the energy efficiency of custom hardware implementations by up to an order of magnitude. Compared to traditional CNN CPU implementations based on highly-tuned, hand-optimized BLAS libraries,our x86 programs implementing the optimal blocking reduce the number of memory accesses by up to 90%.

Zachary DeVito, Michael Mara, Michael Zollhöfer et al.

Many graphics and vision problems can be expressed as non-linear least squares optimizations of objective functions over visual data, such as images and meshes. The mathematical descriptions of these functions are extremely concise, but their implementation in real code is tedious, especially when optimized for real-time performance on modern GPUs in interactive applications. In this work, we propose a new language, Opt (available under http://optlang.org), for writing these objective functions over image- or graph-structured unknowns concisely and at a high level. Our compiler automatically transforms these specifications into state-of-the-art GPU solvers based on Gauss-Newton or Levenberg-Marquardt methods. Opt can generate different variations of the solver, so users can easily explore tradeoffs in numerical precision, matrix-free methods, and solver approaches. In our results, we implement a variety of real-world graphics and vision applications. Their energy functions are expressible in tens of lines of code, and produce highly-optimized GPU solver implementations. These solver have performance competitive with the best published hand-tuned, application-specific GPU solvers, and orders of magnitude beyond a general-purpose auto-generated solver.