arXiv subjects

Harry Dong

Publications and source records attributed to Harry Dong.

14 recordsLinked to original sources

arXiv2026

Almost Free State Prediction Separation

State--prediction separation (SPS) relieves a language model's hidden state of two competing burdens---summarizing the context and predicting the next token---by splitting the forward pass into a state stream and a prediction stream. The separation works, but it is expensive: the prediction stream is a second pass over the whole backbone, costing $\sim$1.9$\times$ the pretraining FLOPs, and even more in terms of wall-clock time when using a flexible attention mask. This paper makes state--prediction separation almost free. We take the separation to its limit with a free pause token: a prediction stream that writes no keys or values at all and so rides the sequence's existing positions. It improves next-token prediction of a standard Transformer by 2-3 centinats in practice on a 1B parameter model, and because it adds no position it costs nothing at inference---no added context length, no KV cache, no decode steps, and essentially no latency, with the growth in inference flops typically irrelevant as it is not the active bottleneck on throughput. The cost is therefore entirely in training where we use four mechanisms to drive it down: a two-pass split that keeps FlashAttention kernels viable, the $w{=}0$ prediction window, a shared gated FFN that evaluates one FFN per position rather than one per stream, and phasing the separation onto the tail of the run. Together these bring the overhead versus an optimized pretraining pipeline to $1.33\times$ wall-clock while recovering ~94% of the gain compared to SPS, and to as low as $1.09\times$ along a graceful quality/compute tradeoff. Furthermore, the FFN optimization reduces the raw flops required at inference time. The result is an isoflop, isoparameter, and isotoken improvement over standard next token trained transformers.

cs.LG↗

arXiv2026

Full-bandwidth transformer

Autoregressive transformers compute along two axes: horizontally across generated tokens, and vertically through model depth. Dense attention gives each token broad horizontal access to the past, but the vertical feedback channel between decoding steps remains narrow: only the sampled token returns to the bottom of the stack, while the top-layer hidden state is discarded. We introduce the \emph{full-bandwidth transformer}, which widens this channel with \emph{latent feedback}: at each decoding step, the previous top-layer hidden state is fused with the sampled token embedding through a gated linear unit and fed back as the next input. Latent feedback lets non-verbalized computation re-enter the stack with a renewed depth budget, while preserving the standard transformer architecture, KV cache, and language-modeling objective. To train full-bandwidth transformers without losing parallel teacher forcing, we use a scheduled multi-pass objective that introduces latent feedback late in pretraining and mixes a small fraction of deeper feedback passes for stability. We train 1B-parameter full-bandwidth transformers up to 400B tokens and find that latent feedback improves validation loss, 5-shot language-model evaluation, math and coding generation, and instruction-tuned performance. With negligible per-token decoding overhead, full-bandwidth transformers match or approach standard transformers trained with roughly $1.5\times$ more tokens, and manage to produce shorter reasoning traces at equal or better accuracy.

cs.AI↗

arXiv2026

Multimodal Diffusion to Mutually Enhance Polarized Light and Low Resolution EBSD Data

In spite of the utility of 3-D electron back-scattered diffraction (EBSD) microscopy, the data collection process can be time-consuming with serial-sectioning. Hence, it is natural to look at other modalities, such as polarized light (PL) data, to accelerate EBSD data collection, supplemented with shared information. Complementarily, features in chaotic PL data could even be enriched with a handful of EBSD measurements. To inherently learn the complex dynamics between EBSD and PL to solve these inverse problems, we use an unconditional multimodal diffusion model, motivated by progress in diffusion models for inverse problems. Although trained solely on synthetic data once, our model has strong generalizable capabilities on real data which can be low-resolution, noisy, corrupted, and misregistered. With inference-time scaling, we show gains in performance on a variety of objectives including grain boundary prediction, super-resolution, and denoising. With our model, we demonstrate that there is little difference from full resolution performance with only 25% (1/4 the resolution) of EBSD data and corrupted PL data.

eess.IV↗

arXiv2026

STEM: Scaling Transformers with Embedding Modules

Fine-grained sparsity promises higher parametric capacity without proportional per-token compute, but often suffers from training instability, load balancing, and communication overhead. We introduce STEM (Scaling Transformers with Embedding Modules), a static, token-indexed approach that replaces the FFN up-projection with a layer-local embedding lookup while keeping the gate and down-projection dense. This removes runtime routing, enables CPU offload with asynchronous prefetch, and decouples capacity from both per-token FLOPs and cross-device communication. Empirically, STEM trains stably despite extreme sparsity. It improves downstream performance over dense baselines while reducing per-token FLOPs and parameter accesses (eliminating roughly one-third of FFN parameters). STEM learns embedding spaces with large angular spread which enhances its knowledge storage capacity. More interestingly, this enhanced knowledge capacity comes with better interpretability. The token-indexed nature of STEM embeddings allows simple ways to perform knowledge editing and knowledge injection in an interpretable manner without any intervention in the input text or additional computation. In addition, STEM strengthens long-context performance: as sequence length grows, more distinct parameters are activated, yielding practical test-time capacity scaling. Across 350M and 1B model scales, STEM delivers up to ~3--4% accuracy improvements overall, with notable gains on knowledge and reasoning-heavy benchmarks (ARC-Challenge, OpenBookQA, GSM8K, MMLU). Overall, STEM is an effective way of scaling parametric memory while providing better interpretability, better training stability and improved efficiency.

cs.LG↗

arXiv2025

Generalized Parallel Scaling with Interdependent Generations

Parallel LLM inference scaling involves sampling a set of $N>1$ responses for a single input prompt. However, these $N$ parallel responses tend to be generated independently from each other, partitioning compute resources and leaving potentially useful information in one generation untapped by others. This is in contrast to response length scaling where past computation is used in all future steps. For higher quality responses and response sets, we propose Bridge to generate interdependent responses in parallel by rethinking batched LLM hidden states as holistic tensors rather than independent slices. With only a small amount (2.8%-5.1%) of new parameters, Bridge improves the relative mean accuracy gains from reinforcement learning with verifiable rewards by up to 39% and boosts consistency of correct responses. Trained once, Bridge scales to any generation width, all with greater performance than independent generations, unlocking a more general mode of parallel scaling that effectively leverages information between sequences, compatible with any post-generation aggregation technique.

cs.AI↗

arXiv2025

Scalable LLM Reasoning Acceleration with Low-rank Distillation

Due to long generations, large language model (LLM) math reasoning demands significant computational resources and time. While many existing efficient inference methods have been developed with excellent performance preservation on language tasks, they often severely degrade math performance. In this paper, we propose Caprese, a resource-efficient distillation method to recover lost capabilities from deploying efficient inference methods, focused primarily in feedforward blocks. With original weights unperturbed, roughly 1% of additional parameters, and only 20K synthetic training samples, we are able to recover much if not all of the reasoning capabilities lost from efficient inference for thinking LLMs and without harm to language tasks for instruct LLMs. Moreover, Caprese slashes the number of active parameters (~2B cut for Gemma 2 9B and Llama 3.1 8B) and integrates cleanly into existing model layers to reduce latency (>16% time-to-next-token reduction) while encouraging response brevity (up to 8.5% fewer tokens).

cs.CL↗

arXiv2024

Towards Low-bit Communication for Tensor Parallel LLM Inference

Tensor parallelism provides an effective way to increase server large language model (LLM) inference efficiency despite adding an additional communication cost. However, as server LLMs continue to scale in size, they will need to be distributed across more devices, magnifying the communication cost. One way to approach this problem is with quantization, but current methods for LLMs tend to avoid quantizing the features that tensor parallelism needs to communicate. Taking advantage of consistent outliers in communicated features, we introduce a quantization method that reduces communicated values on average from 16 bits to 4.2 bits while preserving nearly all of the original performance. For instance, our method maintains around 98.0% and 99.5% of Gemma 2 27B's and Llama 2 13B's original performance, respectively, averaged across all tasks we evaluated on.

cs.AI↗

arXiv2024

ShadowKV: KV Cache in Shadows for High-Throughput Long-Context LLM Inference

With the widespread deployment of long-context large language models (LLMs), there has been a growing demand for efficient support of high-throughput inference. However, as the key-value (KV) cache expands with the sequence length, the increasing memory footprint and the need to access it for each token generation both result in low throughput when serving long-context LLMs. While various dynamic sparse attention methods have been proposed to speed up inference while maintaining generation quality, they either fail to sufficiently reduce GPU memory consumption or introduce significant decoding latency by offloading the KV cache to the CPU. We present ShadowKV, a high-throughput long-context LLM inference system that stores the low-rank key cache and offloads the value cache to reduce the memory footprint for larger batch sizes and longer sequences. To minimize decoding latency, ShadowKV employs an accurate KV selection strategy that reconstructs minimal sparse KV pairs on-the-fly. By evaluating ShadowKV on a broad range of benchmarks, including RULER, LongBench, and Needle In A Haystack, and models like Llama-3.1-8B, Llama-3-8B-1M, GLM-4-9B-1M, Yi-9B-200K, Phi-3-Mini-128K, and Qwen2-7B-128K, we demonstrate that it can support up to 6$\times$ larger batch sizes and boost throughput by up to 3.04$\times$ on an A100 GPU without sacrificing accuracy, even surpassing the performance achievable with infinite batch size under the assumption of infinite GPU memory. The code is available at https://github.com/bytedance/ShadowKV.

cs.LG↗

arXiv2024

Leveraging Multimodal Diffusion Models to Accelerate Imaging with Side Information

Diffusion models have found phenomenal success as expressive priors for solving inverse problems, but their extension beyond natural images to more structured scientific domains remains limited. Motivated by applications in materials science, we aim to reduce the number of measurements required from an expensive imaging modality of interest, by leveraging side information from an auxiliary modality that is much cheaper to obtain. To deal with the non-differentiable and black-box nature of the forward model, we propose a framework to train a multimodal diffusion model over the joint modalities, turning inverse problems with black-box forward models into simple linear inpainting problems. Numerically, we demonstrate the feasibility of training diffusion models over materials imagery data, and show that our approach achieves superior image reconstruction by leveraging the available side information, requiring significantly less amount of data from the expensive microscopy modality.

cs.CV↗

arXiv2024

Prompt-prompted Adaptive Structured Pruning for Efficient LLM Generation

With the development of transformer-based large language models (LLMs), they have been applied to many fields due to their remarkable utility, but this comes at a considerable computational cost at deployment. Fortunately, some methods such as pruning or constructing a mixture of experts (MoE) aim at exploiting sparsity in transformer feedforward (FF) blocks to gain boosts in speed and reduction in memory requirements. However, these techniques can be very costly and inflexible in practice, as they often require training or are restricted to specific types of architectures. To address this, we introduce GRIFFIN, a novel training-free and calibration-free method that selects unique FF experts at the sequence level for efficient generation across a plethora of LLMs with different non-ReLU activation functions. This is possible due to a critical observation that many trained LLMs naturally produce highly structured FF activation patterns within a sequence, which we call flocking. Despite our method's simplicity, we show with 50% of the FF parameters, GRIFFIN maintains the original model's performance with little to no degradation on a variety of classification and generation tasks, all while improving latency (e.g. 1.29$\times$ and 1.25$\times$ speed-ups in Gemma 7B and Llama 2 13B, respectively, on an NVIDIA L40). Code is available at https://github.com/hdong920/GRIFFIN.

cs.LG↗

arXiv2024

Get More with LESS: Synthesizing Recurrence with KV Cache Compression for Efficient LLM Inference

Many computational factors limit broader deployment of large language models. In this paper, we focus on a memory bottleneck imposed by the key-value (KV) cache, a computational shortcut that requires storing previous KV pairs during decoding. While existing KV cache methods approach this problem by pruning or evicting large swaths of relatively less important KV pairs to dramatically reduce the memory footprint of the cache, they can have limited success in tasks that require recollecting a majority of previous tokens. To alleviate this issue, we propose LESS, a simple integration of a (nearly free) constant sized cache with eviction-based cache methods, such that all tokens can be queried at later decoding steps. Its ability to retain information throughout time shows merit on a variety of tasks where we demonstrate LESS can help reduce the performance gap from caching everything, sometimes even matching it, all while being efficient. Relevant code can be found at https://github.com/hdong920/LESS.

cs.LG↗

arXiv2023

A Lightweight Transformer for Faster and Robust EBSD Data Collection

Three dimensional electron back-scattered diffraction (EBSD) microscopy is a critical tool in many applications in materials science, yet its data quality can fluctuate greatly during the arduous collection process, particularly via serial-sectioning. Fortunately, 3D EBSD data is inherently sequential, opening up the opportunity to use transformers, state-of-the-art deep learning architectures that have made breakthroughs in a plethora of domains, for data processing and recovery. To be more robust to errors and accelerate this 3D EBSD data collection, we introduce a two step method that recovers missing slices in an 3D EBSD volume, using an efficient transformer model and a projection algorithm to process the transformer's outputs. Overcoming the computational and practical hurdles of deep learning with scarce high dimensional data, we train this model using only synthetic 3D EBSD data with self-supervision and obtain superior recovery accuracy on real 3D EBSD data, compared to existing methods.

cs.CV↗

arXiv2022

Deep Unfolded Tensor Robust PCA with Self-supervised Learning

Tensor robust principal component analysis (RPCA), which seeks to separate a low-rank tensor from its sparse corruptions, has been crucial in data science and machine learning where tensor structures are becoming more prevalent. While powerful, existing tensor RPCA algorithms can be difficult to use in practice, as their performance can be sensitive to the choice of additional hyperparameters, which are not straightforward to tune. In this paper, we describe a fast and simple self-supervised model for tensor RPCA using deep unfolding by only learning four hyperparameters. Despite its simplicity, our model expunges the need for ground truth labels while maintaining competitive or even greater performance compared to supervised deep unfolding. Furthermore, our model is capable of operating in extreme data-starved scenarios. We demonstrate these claims on a mix of synthetic data and real-world tasks, comparing performance against previously studied supervised deep unfolding methods and Bayesian optimization baselines.

stat.ML↗

arXiv2022

Fast and Provable Tensor Robust Principal Component Analysis via Scaled Gradient Descent

An increasing number of data science and machine learning problems rely on computation with tensors, which better capture the multi-way relationships and interactions of data than matrices. When tapping into this critical advantage, a key challenge is to develop computationally efficient and provably correct algorithms for extracting useful information from tensor data that are simultaneously robust to corruptions and ill-conditioning. This paper tackles tensor robust principal component analysis (RPCA), which aims to recover a low-rank tensor from its observations contaminated by sparse corruptions, under the Tucker decomposition. To minimize the computation and memory footprints, we propose to directly recover the low-dimensional tensor factors -- starting from a tailored spectral initialization -- via scaled gradient descent (ScaledGD), coupled with an iteration-varying thresholding operation to adaptively remove the impact of corruptions. Theoretically, we establish that the proposed algorithm converges linearly to the true low-rank tensor at a constant rate that is independent with its condition number, as long as the level of corruptions is not too large. Empirically, we demonstrate that the proposed algorithm achieves better and more scalable performance than state-of-the-art matrix and tensor RPCA algorithms through synthetic experiments and real-world applications.

stat.ML↗