- June 18, 2024Research
SpecExec: Massively Parallel Speculative Decoding for Interactive LLM Inference on Consumer Devices
- July 11, 2023Research
Petals: decentralized inference and finetuning of large language models
- September 30, 2021Research
A few attention heads for reasoning in multiple languages
Natural language processing
Language is one of the key forms of communication. We study methods of language representation and understanding to simplify human-computer interactions.
Posts
Publications
Cache Me If You Must: Adaptive Key-Value Quantization for Large Language Models
ICML, 2025Efficient real-world deployments of large language models (LLMs) rely on Key-Value (KV) caching for processing and generating long outputs, reducing the need for repetitive computation. For large contexts, Key-Value caches can take up tens of gigabytes of device memory, as they store vector representations for each token and layer. Recent work has shown that the cached vectors can be compressed through quantization, pruning or merging, but these techniques often compromise quality towards higher compression rates. In this work, we aim to improve Key & Value compression by exploiting two observations: 1) the inherent dependencies between keys and values across different layers, and 2) the existence of high-compression methods for internal network states (e.g. attention Keys & Values). We propose AQUA-KV, an adaptive quantization for Key-Value caches that relies on compact adapters to exploit existing dependencies between Keys and Values, and aims to “optimally” compress the information that cannot be predicted. AQUA-KV significantly improves compression rates, while maintaining high accuracy on state-of-the-art LLM families. On Llama 3.2 LLMs, we achieve near-lossless inference at 2-2.5 bits per value with under 1 relative error in perplexity and LongBench scores. AQUA-KV is one-shot, simple, and efficient: it can be calibrated on a single GPU within 1-6 hours, even for 70B models.
EvoPress: Accurate Dynamic Model Compression via Evolutionary Search
ICML, 2025The high computational costs of large language models (LLMs) have led to a flurry of research on LLM compression, via methods such as quantization, sparsification, or structured pruning. A new frontier in this area is given by dynamic, non-uniform compression methods, which adjust the compression levels (e.g., sparsity) per-block or even per-layer in order to minimize accuracy loss, while guaranteeing a global compression threshold. Yet, current methods rely on estimating the “importance” of a given layer, implicitly assuming that layers contribute independently to the overall compression error. We begin from the motivating observation that this independence assumption does not generally hold for LLM compression: pruning a model further may even significantly recover performance. To address this, we propose EvoPress, a novel evolutionary framework for dynamic LLM compression. By formulating dynamic compression as a general optimization problem, EvoPress identifies optimal compression profiles in a highly efficient manner, and generalizes across diverse models and compression techniques. Via EvoPress, we achieve state-of-the-art performance for dynamic compression of Llama, Mistral, and Phi models, setting new benchmarks for structural pruning (block/layer dropping), unstructured sparsity, and quantization with dynamic bitwidths.
FRUGAL: Memory-Efficient Optimization by Reducing State Overhead for Scalable Training
ICML, 2025With the increase in the number of parameters in large language models, the training process increasingly demands larger volumes of GPU memory. A significant portion of this memory is typically consumed by the optimizer state. To overcome this challenge, recent approaches such as low-rank adaptation (LoRA), low-rank gradient projection (GaLore), and blockwise optimization (BAdam) have been proposed. However, in all these algorithms, the effective rank of the weight updates remains low-rank, which can lead to a substantial loss of information from the gradient. This loss can be critically important, especially during the pre-training stage. In this paper, we introduce FRUGAL (Full-Rank Updates with GrAdient spLitting), a new memory-efficient optimization framework. FRUGAL leverages gradient splitting to perform low-dimensional updates using advanced algorithms (such as Adam), while updates along the remaining directions are executed via state-free methods like SGD or signSGD. Our framework can be integrated with various low-rank update selection techniques, including GaLore and BAdam. We provide theoretical convergence guarantees for our framework when using SGDM for low-dimensional updates and SGD for state-free updates. Additionally, our method consistently outperforms concurrent approaches, achieving state-of-the-art results in pre-training and fine-tuning tasks while balancing memory efficiency and performance metrics.
Datasets
Shifts Dataset
The Shifts Dataset contains curated and labeled examples of real, 'in-the-wild' distributional shifts across three large-scale tasks. Specifically, it contains tabular weather prediction, machine translation, and vehicle motion prediction tasks' data used in Shifts Challenge 2021. Dataset shift is ubiquitous in all of these tasks and modalities.
Text-to-Image dataset for billion-scale similarity search
Yandex Text-to-Image (T2I) dataset is collected to foster the research in billion-scale nearest neighbor search (NNS) when query distribution differs from the indexing one. In particular, this dataset addresses the cross-domain setting: a user specifies a textual query and requests the search engine to retrieve the most relevant images to the query. Notably, current large-scale indexing methods perform poorly in this setting. Therefore, novel highly-performant indexing solutions robust to out-of-domain queries are in high demand.
The dataset represents a snapshot of the Yandex visual search engine and contains 1 billion 200-dimensional image embeddings for indexing. The image embeddings are produced by the Se-ResNext-101 model. The embeddings for textual queries are extracted by a variant of the DSSM model.
Read more about the data format and how to download the dataset in the related post.