Visual Auto-Regressive modeling (VAR) has shown promise in bridging the speed and quality gap between autoregressive image models and diffusion models. VAR reformulates autoregressive modeling by decomposing an image into successive resolution scales. During inference, an image is generated by predicting all the tokens in the next (higher-resolution) scale, conditioned on all tokens in all previous (lower-resolution) scales. However, this formulation suffers from reduced image quality due to the parallel generation of all tokens in a resolution scale; has sequence lengths scaling superlinearly in image resolution; and requires retraining to change the sampling schedule.We introduce Hierarchical Masked AutoRegressive modeling (HMAR), a new image generation algorithm that alleviates these issues using next-scale prediction and masked prediction to generate high-quality images with fast sampling. HMAR reformulates next-scale prediction as a Markovian process, wherein the prediction of each resolution scale is conditioned only on tokens in its immediate predecessor instead of the tokens in all predecessor resolutions. When predicting a resolution scale, HMAR uses a controllable multi-step masked generation procedure to generate a subset of the tokens in each step. On ImageNet 256 × 256 and 512 × 512 benchmarks, HMAR models match or outperform parameter-matched VAR, diffusion, and autoregressive baselines. We develop efficient IO-aware block-sparse attention kernels that allow HMAR to achieve faster training and inference times over VAR by over 2.5× and 1.75× respectively, as well as over 3× lower inference memory foot-print. Finally, HMAR yields additional flexibility over VAR; its sampling schedule can be changed without further training, and it can be applied to image editing tasks in a zero-shot manner.

Designing model architectures requires decisions such as selecting operators (e.g., attention, convolution) and configurations (e.g., depth, width). However, evaluating the impact of these decisions on model quality requires costly pretraining, limiting architectural investigation. Inspired by how new software is built on existing code, we ask: can new architecture designs be studied using pretrained models? To this end, we present *grafting*, a simple approach for editing pretrained diffusion transformers (DiTs) to materialize new architectures under small compute budgets. Informed by our analysis of activation behavior and attention locality, we construct a testbed based on the DiT-XL/2 design to study the impact of grafting on model quality. Using this testbed, we develop a family of hybrid designs via grafting: replacing softmax attention with gated convolution, local attention, and linear attention, and replacing MLPs with variable expansion ratio and convolutional variants. Notably, many hybrid designs achieve good quality (FID: 2.38–2.64 vs. 2.27 for DiT-XL/2) using $<2$% pretraining compute. We then graft a text-to-image model (PixArt-$\Sigma$), achieving a 1.43$\times$ speedup with less than a 2% drop in GenEval score. Finally, we present a case study that restructures DiT-XL/2 by converting every pair of sequential transformer blocks into parallel blocks via grafting. This reduces model depth by 2$\times$ and yields better quality (FID: 2.77) than other models of comparable depth. Together, we show that new diffusion model designs can be explored by grafting pretrained DiTs, with edits ranging from operator replacement to architecture restructuring. Code and grafted models: https://grafting.stanford.edu.

Convolution models with long filters have demonstrated state-of-the-art reasoning abilities in many long-sequence tasks but lag behind the most optimized Transformers in wall-clock time. A major bottleneck is the Fast Fourier Transform (FFT)---which allows long convolutions to run in $O(N\log N)$ time in sequence length $N$ but has poor hardware utilization. In this paper, we study how to optimize the FFT convolution. We find two key bottlenecks: the FFT does not effectively use specialized matrix multiply units, and it incurs expensive I/O between layers of the memory hierarchy. In response, we propose FlashFFTConv. FlashFFTConv uses a matrix decomposition that computes the FFT using matrix multiply units and enables kernel fusion for long sequences, reducing I/O. We also present two sparse convolution algorithms---1) partial convolutions and 2) frequency-sparse convolutions---which can be implemented simply by skipping blocks in the matrix decomposition, enabling further opportunities for memory and compute savings. FlashFFTConv speeds up exact FFT convolutions by up to 8.7$\times$ over PyTorch and achieves up to 4.4$\times$ speedup end-to-end. Given the same compute budget, FlashFFTConv allows Hyena-GPT-s to achieve 2.3 points better perplexity and M2-BERT-base to achieve 3.3 points higher GLUE score---matching models with twice the parameter count. FlashFFTConv also achieves 96.1% accuracy on Path-512, a high-resolution vision task where no model had previously achieved better than 50%. Furthermore, partial convolutions enable longer-sequence models---yielding the first DNA model that can process the longest human genes (2.3M base pairs)---and frequency-sparse convolutions speed up pretrained models while maintaining or improving model quality.

Retrieval pipelines are an integral component of many machine learning systems. However, they perform poorly in domains where documents are long (e.g., 10K tokens or more) and where identifying the relevant document requires synthesizing information across the entire text. Developing long-context retrieval encoders suitable for these domains raises three challenges: (1) how to evaluate long-context retrieval performance, (2) how to pretrain a base language model to represent both short contexts (corresponding to queries) and long contexts (corresponding to documents), and (3) how to finetune this model for retrieval under the batch size limitations imposed by GPU memory constraints. To address these challenges, we first introduce LoCoV1, a 12 task benchmark constructed to measure long-context retrieval where chunking is not possible or not effective. We next present the M2-BERT retrieval encoder, an 80M parameter state-space encoder model built from the Monarch Mixer architecture, capable of scaling to documents up to 32K tokens long. We describe a pretraining data mixture which allows this encoder to process both short and long context sequences, and a finetuning approach that adapts this base model to retrieval with only single-sample batches. Finally, we validate the M2-BERT retrieval encoder on LoCoV1, finding that it outperforms competitive Transformer-based models by at least 22.2 points, despite containing 90× fewer parameters.

Large language models are increasingly becoming a cornerstone technology in artificial intelligence, the sciences, and society as a whole, yet the optimal strategies for dataset composition and filtering remain largely elusive. Many of the top-performing models lack transparency in their dataset curation and model development processes, posing an obstacle to the development of fully open language models. In this paper, we identify three core data-related challenges that must be addressed to advance open-source language models. These include (1) transparency in model development, including the data curation process, (2) access to large quantities of high-quality data, and (3) availability of artifacts and metadata for dataset curation and analysis. To address these challenges, we release RedPajama-V1, an open reproduction of the LLaMA training dataset. In addition, we release RedPajama-V2, a massive web-only dataset consisting of raw, unfiltered text data together with quality signals and metadata.Together, the RedPajama datasets comprise over 100 trillion tokens spanning multiple domains and with their quality signals facilitate the filtering of data, aiming to inspire the development of numerous new datasets. To date, these datasets have already been used in the training of strong language models used in production, such as Snowflake Arctic, Salesforce's XGen and AI2's OLMo. To provide insight into the quality of RedPajama, we present a series of analyses and ablation studies with decoder-only language models with up to 1.6B parameters. Our findings demonstrate how quality signals for web data can be effectively leveraged to curate high-quality subsets of the dataset, underscoring the potential of RedPajama to advance the development of transparent and high-performing language models at scale.

State space models (SSMs) have high performance on long sequence modeling but require sophisticated initialization techniques and specialized implementations for high quality and runtime performance. We study whether a simple alternative can match SSMs in performance and efficiency: directly learning long convolutions over the sequence. We find that a key requirement to achieving high performance is keeping the convolution kernels smooth. We find that simple interventions-such as squashing the kernel weights-result in smooth kernels and recover SSM performance on a range of tasks including the long range arena, image classification, language modeling, and brain data modeling. Next, we develop FlashButterfly, an IO-aware algorithm to improve the runtime performance of long convolutions. FlashButterfly appeals to classic Butterfly decompositions of the convolution to reduce GPU memory IO and increase FLOP utilization. FlashButterfly speeds up convolutions by 2.2$\times$, and allows us to train on Path256, a challenging task with sequence length 64K, where we set state-of-the-art by 29.1 points while training 7.2$\times$ faster than prior work. Lastly, we introduce an extension to FlashButterfly that learns the coefficients of the Butterfly decomposition, increasing expressivity without increasing runtime. Using this extension, we outperform a Transformer on WikiText103 by 0.2 PPL with 30% fewer parameters.

Recent advances in deep learning have relied heavily on the use of large Transformers due to their ability to learn at scale. However, the core building block of Transformers, the attention operator, exhibits quadratic cost in sequence length, limiting the amount of context accessible. Existing subquadratic methods based on low-rank and sparse approximations need to be combined with dense attention layers to match Transformers at scale, indicating a gap in capability. In this work, we propose Hyena, a subquadratic drop-in replacement for attention constructed by interleaving implicitly parametrized long convolutions and data-controlled gating. In challenging reasoning tasks on sequences of thousands to hundreds of thousands of tokens, Hyena improves accuracy by more than 50 points over operators relying on state-space models, transfer functions, and other implicit and explicit methods, matching attention-based models. We set a new state-of-the-art for dense-attention-free architectures on language modeling in standard datasets WikiText103 and The Pile, reaching Transformer quality with a 20% reduction in training compute required at sequence length 2k. Hyena operators are 2x faster than highly optimized attention at sequence length 8k, with speedups of 100x at 64k.

Machine learning models are increasingly being scaled in both sequence length and model dimension to reach longer contexts and better performance. However, existing architectures such as Transformers scale quadratically along both these axes. We ask: are there performant architectures that can scale sub-quadratically along sequence length and model dimension? We introduce Monarch Mixer (M2), a new architecture that uses the same sub-quadratic primitive along both sequence length and model dimension: Monarch matrices, a simple class of expressive structured matrices that captures many linear transforms, achieves high hardware efficiency on GPUs, and scales sub-quadratically. As a proof of concept, we explore the performance of M2 in three domains: non-causal BERT-style language modeling, ViT-style image classification, and causal GPT-style language modeling. For non-causal BERT-style modeling, M2 matches BERT-base and BERT-large in downstream GLUE quality with up to 27% fewer parameters, and achieves up to 9.1$\times$ higher throughput at sequence length 4K. On ImageNet, M2 outperforms ViT-b by 1% in accuracy, with only half the parameters. Causal GPT-style models introduce a technical challenge: enforcing causality via masking introduces a quadratic bottleneck. To alleviate this bottleneck, we develop a novel theoretical view of Monarch matrices based on multivariate polynomial evaluation and interpolation, which lets us parameterize M2 to be causal while remaining sub-quadratic. Using this parameterization, M2 matches GPT-style Transformers at 360M parameters in pretraining perplexity on The PILE—showing for the first time that it may be possible to match Transformer quality without attention or MLPs.

Recent advances in attention-free sequence models rely on convolutions as alternatives to the attention operator at the core of Transformers. In particular, long convolution sequence models have achieved state-of-the-art performance in many domains, but incur a significant cost during auto-regressive inference workloads -- naively requiring a full pass (or caching of activations) over the input sequence for each generated token -- similarly to attention-based models. In this paper, we seek to enable $\mathcal O(1)$ compute and memory cost per token in any pre-trained long convolution architecture to reduce memory footprint and increase throughput during generation. Concretely, our methods consist in extracting low-dimensional linear state-space models from each convolution layer, building upon rational interpolation and model-order reduction techniques. We further introduce architectural improvements to convolution-based layers such as Hyena: by weight-tying the filters across channels into heads, we achieve higher pre-training quality and reduce the number of filters to be distilled. The resulting model achieves 10x higher throughput than Transformers and 1.5x higher than Hyena at 1.3B parameters, without any loss in quality after distillation.

An ideal learned representation should display transferability and robustness. Supervised contrastive learning (SupCon) is a promising method for training accurate models, but produces representations that do not capture these properties due to class collapse—when all points in a class map to the same representation. Recent work suggests that "spreading out" these representations improves them, but the precise mechanism is poorly understood. We argue that creating spread alone is insufficient for better representations, since spread is invariant to permutations within classes. Instead, both the correct degree of spread and a mechanism for breaking this invariance are necessary. We first prove that adding a weighted class-conditional InfoNCE loss to SupCon controls the degree of spread. Next, we study three mechanisms to break permutation invariance: using a constrained encoder, adding a class-conditional autoencoder, and using data augmentation. We show that the latter two encourage clustering of latent subclasses under more realistic conditions than the former. Using these insights, we show that adding a properly-weighted class-conditional InfoNCE loss and a class-conditional autoencoder to SupCon achieves 11.1 points of lift on coarse-to-fine transfer across 5 standard datasets and 4.7 points on worst-group robustness on 3 datasets, setting state-of-the-art on CelebA by 11.5 points.

Transformers are slow and memory-hungry on long sequences, since the time and memory complexity of self-attention are quadratic in sequence length. Approximate attention methods have attempted to address this problem by trading off model quality to reduce the compute complexity, but often do not achieve wall-clock speedup. We argue that a missing principle is making attention algorithms IO-aware---accounting for reads and writes between levels of GPU memory. We propose FlashAttention, an IO-aware exact attention algorithm that uses tiling to reduce the number of memory reads/writes between GPU high bandwidth memory (HBM) and GPU on-chip SRAM. We analyze the IO complexity of FlashAttention, showing that it requires fewer HBM accesses than standard attention, and is optimal for a range of SRAM sizes. We also extend FlashAttention, yielding an approximate attention algorithm that is faster than any existing approximate attention method. FlashAttention, 3x speedup on GPT-2 (seq. length 1K), and 2.4x speedup on long-range arena (seq. length 1K-4K). FlashAttention, yielding higher quality models (0.7 better perplexity on GPT-2 and 6.4 points of lift on long-document classification) and entirely new capabilities: the first Transformers to achieve better-than-chance performance on the Path-X challenge (seq. length 16K, 61.4% accuracy) and Path-256 (seq. length 64K, 63.1% accuracy).

Foundation models offer an exciting new paradigm for constructing models with out-of-the-box embeddings and a few labeled examples. However, it is not clear how to best apply foundation models without labeled data. A potential approach is to fuse foundation models with weak supervision frameworks, which use weak label sources—pre-trained models, heuristics, crowd-workers—to construct pseudolabels. The challenge is building a combination that best exploits the signal available in both foundation models and weak sources. We propose LIGER, a combination that uses foundation model embeddings to improve two crucial elements of existing weak supervision techniques. First, we produce finer estimates of weak source quality by partitioning the embedding space and learning per-part source accuracies. Second, we improve source coverage by extending source votes in embedding space. Despite the black-box nature of foundation models, we prove results characterizing how our approach improves performance and show that lift scales with the smoothness of label distributions in embedding space. On six benchmark NLP and video tasks, LIGER outperforms vanilla weak supervision by 14.1 points, weakly-supervised kNN and adapters by 11.8 points, and kNN and adapters supervised by traditional hand labels by 7.2 points.

Cable (TV) news reaches millions of US households each day. News stakeholders such as communications researchers, journalists, and media monitoring organizations are interested in the visual content of cable news, especially who is on-screen. Manual analysis, however, is labor intensive and limits the size of prior studies. We conduct a large-scale, quantitative analysis of the faces in a decade of cable news video from the top three US cable news networks (CNN, FOX, and MSNBC), totaling 244,038 hours between January 2010 and July 2019. Our work uses technologies such as automatic face and gender recognition to measure the "screen time" of faces and to enable visual analysis and exploration at scale. Our analysis method gives insight into a broad set of socially relevant topics. For instance, male-presenting faces receive much more screen time than female-presenting faces (2.4x in 2010, 1.9x in 2019). To make our dataset and annotations accessible, we release a public interface at https://tvnews.stanford.edu that allows the general public to write queries and to perform their own analyses.

Weak supervision is a popular method for building machine learning models without relying on ground truth annotations. Instead, it generates probabilistic training labels by estimating the accuracies of multiple noisy labeling sources (e.g., heuristics, crowd workers). Existing approaches use latent variable estimation to model the noisy sources, but these methods can be computationally expensive, scaling superlinearly in the data. In this work, we show that, for a class of latent variable models highly applicable to weak supervision, we can find a closed-form solution to model parameters, obviating the need for iterative solutions like stochastic gradient descent (SGD). We use this insight to build FlyingSquid, a weak supervision framework that runs orders of magnitude faster than previous weak supervision approaches and requires fewer assumptions. In particular, we prove bounds on generalization error without assuming that the latent variable model can exactly parameterize the underlying data distribution. Empirically, we validate FlyingSquid on benchmark weak supervision datasets and find that it achieves the same or higher quality compared to previous approaches without the need to tune an SGD procedure, recovers model parameters 170 times faster on average, and enables new video analysis and online learning applications.

Since manually labeling training data is slow and expensive, recent industrial and scientific research efforts have turned to weaker or noisier forms of supervision sources. However, existing weak supervision approaches fail to model multi-resolution sources for sequential data, like video, that can assign labels to individual elements or collections of elements in a sequence. A key challenge in weak supervision is estimating the unknown accuracies and correlations of these sources without using labeled data. Multi-resolution sources exacerbate this challenge due to complex correlations and sample complexity that scales in the length of the sequence. We propose Dugong, the first framework to model multi-resolution weak supervision sources with complex correlations to assign probabilistic labels to training data. Theoretically, we prove that Dugong, under mild conditions, can uniquely recover the unobserved accuracy and correlation parameters and use parameter sharing to improve sample complexity. Our method assigns clinician-validated labels to population-scale biomedical video repositories, helping outperform traditional supervision by 36.8 F1 points and addressing a key use case where machine learning has been severely limited by the lack of expert labeled data. On average, Dugong improves over traditional supervision by 16.0 F1 points and existing weak supervision approaches by 24.2 F1 points across several video and sensor classification tasks.