Knowledge distillation methods have recently shown to be a promising direction to speedup the synthesis of large-scale diffusion models by requiring only a few inference steps. While several powerful distillation methods were recently proposed, the overall quality of student samples is typically lower compared to the teacher ones, which hinders their practical usage. In this work, we investigate the relative quality of samples produced by the teacher text-to-image diffusion model and its distilled student version. As our main empirical finding, we discover that a noticeable portion of student samples exhibit superior fidelity compared to the teacher ones, despite the ``approximate'' nature of the student. Based on this finding, we propose an adaptive collaboration between student and teacher diffusion models for effective text-to-image synthesis. Specifically, the distilled model produces the initial sample, and then an oracle decides whether it needs further improvements with a slow teacher model. Extensive experiments demonstrate that the designed pipeline surpasses state-of-the-art text-to-image alternatives for various inference budgets in terms of human preference. Furthermore, the proposed approach can be naturally used in popular applications such as text-guided image editing and controllable generation.
Large language models (LLMs) are useful in many NLP tasks and become more capable with size, with the best open-source models having over 50 billion parameters. However, using these 50B+ models requires high-end hardware, making them inaccessible to most researchers. In this work, we investigate methods for cost-efficient inference and fine-tuning of LLMs, comparing local and distributed strategies. We observe that a large enough model (50B+) can run efficiently even on geodistributed devices in a consumer-grade network. This could allow running LLM efficiently by pooling together idle compute resources of multiple research groups and volunteers. We address two open problems: (1) how to perform inference and fine-tuning reliably if any device can disconnect abruptly and (2) how to partition LLMs between devices with uneven hardware, joining and leaving at will. In order to do that, we develop special fault-tolerant inference algorithms and load-balancing protocols that automatically assign devices to maximize the total system throughput. We showcase these algorithms in Petals - a decentralized system that runs Llama 2 (70B) and BLOOM (176B) over the Internet up to 10x faster than offloading for interactive generation. We evaluate the performance of our system in simulated conditions and a real-world setup spanning two continents.
The statistical distribution of content uploaded and searched on media sharing sites changes over time due to seasonal, sociological and technical factors. We investigate the impact of this "content drift" for large-scale similarity search tools, based on nearest neighbor search in embedding space. Unless a costly index reconstruction is performed frequently, content drift degrades the search accuracy and efficiency. The degradation is especially severe since, in general, both the query and database distributions change. We introduce and analyze real-world image and video datasets for which temporal information is available over a long time period. Based on the learnings, we devise DeDrift, a method that updates embedding quantizers to continuously adapt large-scale indexing structures on-the-fly. DeDrift almost eliminates the accuracy degradation due to the query and database content drift while being up to 100x faster than a full index reconstruction.
Recent advances in diffusion models enable many powerful instruments for image editing. One of these instruments is text-driven image manipulations: editing semantic attributes of an image according to the provided text description. % Popular text-conditional diffusion models offer various high-quality image manipulation methods for a broad range of text prompts. Existing diffusion-based methods already achieve high-quality image manipulations for a broad range of text prompts. However, in practice, these methods require high computation costs even with a high-end GPU. This greatly limits potential real-world applications of diffusion-based image editing, especially when running on user devices. In this paper, we address efficiency of the recent text-driven editing methods based on unconditional diffusion models and develop a novel algorithm that learns image manipulations 4.5-10 times faster and applies them 8 times faster. We carefully evaluate the visual quality and expressiveness of our approach on multiple datasets using human annotators. Our experiments demonstrate that our algorithm achieves the quality of much more expensive methods. Finally, we show that our approach can adapt the pretrained model to the user-specified image and text description on the fly just for 4 seconds. In this setting, we notice that more compact unconditional diffusion models can be considered as a rational alternative to the popular text-conditional counterparts.
Denoising diffusion probabilistic models are currently becoming the leading paradigm of generative modeling for many important data modalities. Being the most prevalent in the computer vision community, diffusion models have also recently gained some attention in other domains, including speech, NLP, and graph-like data. In this work, we investigate if the framework of diffusion models can be advantageous for general tabular problems, where datapoints are typically represented by vectors of heterogeneous features. The inherent heterogeneity of tabular data makes it quite challenging for accurate modeling, since the individual features can be of completely different nature, i.e., some of them can be continuous and some of them can be discrete. To address such data types, we introduce TabDDPM -- a diffusion model that can be universally applied to any tabular dataset and handles any type of feature. We extensively evaluate TabDDPM on a wide set of benchmarks and demonstrate its superiority over existing GAN/VAE alternatives, which is consistent with the advantage of diffusion models in other fields. Additionally, we show that TabDDPM is eligible for privacy-oriented setups, where the original datapoints cannot be publicly shared.
Many NLP tasks benefit from using large language models (LLMs) that often have more than 100 billion parameters. With the release of BLOOM-176B and OPT-175B, everyone can download pretrained models of this scale. Still, using these models requires high-end hardware unavailable to many researchers. In some cases, LLMs can be used more affordably via RAM offloading or hosted APIs. However, these techniques have innate limitations: offloading is too slow for interactive inference, while APIs are not flexible enough for research. In this work, we propose Petals $-$ a system for inference and fine-tuning of large models collaboratively by joining the resources of multiple parties trusted to process client's data. We demonstrate that this strategy significantly outperforms offloading for very large models, running inference of BLOOM-176B on consumer GPUs with $\approx$ 1 step per second. Unlike most inference APIs, Petals also natively exposes the hidden states of served models, allowing its users to train and share custom model extensions based on efficient fine-tuning methods.
Despite the broad range of algorithms for Approximate Nearest Neighbor Search, most empirical evaluations of algorithms have focused on smaller datasets, typically of 1 million points~\citep{Benchmark}. However, deploying recent advances in embedding based techniques for search, recommendation and ranking at scale require ANNS indices at billion, trillion or larger scale. Barring a few recent papers, there is limited consensus on which algorithms are effective at this scale vis-\`a-vis their hardware cost. This competition compares ANNS algorithms at billion-scale by hardware cost, accuracy and performance. We set up an open source evaluation framework and leaderboards for both standardized and specialized hardware. The competition involves three tracks. The standard hardware track T1 evaluates algorithms on an Azure VM with limited DRAM, often the bottleneck in serving billion-scale indices, where the embedding data can be hundreds of GigaBytes in size. It uses FAISS~\citep{Faiss17} as the baseline. The standard hardware track T2 additional allows inexpensive SSDs in addition to the limited DRAM and uses DiskANN~\citep{DiskANN19} as the baseline. The specialized hardware track T3 allows any hardware configuration, and again uses FAISS as the baseline. We compiled six diverse billion-scale datasets, four newly released for this competition, that span a variety of modalities, data types, dimensions, deep learning models, distance functions and sources. The outcome of the competition was ranked leaderboards of algorithms in each track based on recall at a query throughput threshold. Additionally, for track T3, separate leaderboards were created based on recall as well as cost-normalized and power-normalized query throughput.
Denoising diffusion probabilistic models have recently received much research attention since they outperform alternative approaches, such as GANs, and currently provide state-of-the-art generative performance. The superior performance of diffusion models has made them an appealing tool in several applications, including inpainting, super-resolution, and semantic editing. In this paper, we demonstrate that diffusion models can also serve as an instrument for semantic segmentation, especially in the setup when labeled data is scarce. In particular, for several pretrained diffusion models, we investigate the intermediate activations from the networks that perform the Markov step of the reverse diffusion process. We show that these activations effectively capture the semantic information from an input image and appear to be excellent pixel-level representations for the segmentation problem. Based on these observations, we describe a simple segmentation method, which can work even if only a few training images are provided. Our approach significantly outperforms the existing alternatives on several datasets for the same amount of human supervision.