EMS: Adaptive Evict-then-Merge Strategy for Head-wise KV Cache Compression Based on Global-Local Importance

As large language models (LLMs) continue to advance, the demand for higher quality and faster processing of long contexts across various applications is growing. KV cache is widely adopted as it stores previously generated key and value tokens, effectively reducing redundant computations during inference. However, as memory overhead becomes a significant concern, efficient compression of KV cache has gained increasing attention. Most existing methods perform compression from two perspectives: identifying important tokens and designing compression strategies. However, these approaches often produce biased distributions of important tokens due to the influence of accumulated attention scores or positional encoding. Furthermore, they overlook the sparsity and redundancy across different heads, which leads to difficulties in preserving the most effective information at the head level. To this end, we propose EMS to overcome these limitations, while achieving better KV cache compression under extreme compression ratios. Specifically, we introduce a Global-Local score that combines accumulated attention scores from both global and local KV tokens to better identify the token importance. For the compression strategy, we design an adaptive and unified Evict-then-Merge framework that accounts for the sparsity and redundancy of KV tokens across different heads. Additionally, we implement the head-wise parallel compression through a zero-class mechanism to enhance efficiency. Extensive experiments demonstrate our SOTA performance even under extreme compression ratios. EMS consistently achieves the lowest perplexity, improves scores by over 1.28 points across four LLMs on LongBench under a 256 cache budget, and preserves 95% retrieval accuracy with a cache budget less than 2% of the context length in the Needle-in-a-Haystack task.

GAQAT: gradient-adaptive quantization-aware training for domain generalization

Research on loss surface geometry, such as Sharpness-Aware Minimization (SAM), shows that flatter minima improve generalization. Recent studies further reveal that flatter minima can also reduce the domain generalization (DG) gap. However, existing flatness-based DG techniques predominantly operate within a full-precision training process, which is impractical for deployment on resource-constrained edge devices that typically rely on lower bit-width representations (e.g., 4 bits, 3 bits). Consequently, low-precision quantization-aware training is critical for optimizing these techniques in real-world applications. In this paper, we observe a significant degradation in performance when applying state-of-the-art DG-SAM methods to quantized models, suggesting that current approaches fail to preserve generalizability during the low-precision training process. To address this limitation, we propose a novel Gradient-Adaptive Quantization-Aware Training (GAQAT) framework for DG. Our approach begins by identifying the scale-gradient conflict problem in low-precision quantization, where the task loss and smoothness loss induce conflicting gradients for the scaling factors of quantizers, with certain layers exhibiting opposing gradient directions. This conflict renders the optimization of quantized weights highly unstable. To mitigate this, we further introduce a mechanism to quantify gradient inconsistencies and selectively freeze the gradients of scaling factors, thereby stabilizing the training process and enhancing out-of-domain generalization. Extensive experiments validate the effectiveness of the proposed GAQAT framework. On PACS, our 3-bit and 4-bit models outperform direct DG-QAT integration by up to 4.5%. On DomainNet, the 4-bit model achieves near-lossless performance compared to full precision, with improvements of 1.39% (4-bit) and 1.06% (3-bit) over the SOTA QAT baseline.

Mixed Degradation Image Restoration via Local Dynamic Optimization and Conditional Embedding

Multiple-in-one image restoration (IR) has made significant progress, aiming to handle all types of single degraded image restoration with a single model. However, in real-world scenarios, images often suffer from combinations of multiple degradation factors. Existing multiple-in-one IR models encounter challenges related to degradation diversity and prompt singularity when addressing this issue. In this paper, we propose a novel multiple-in-one IR model that can effectively restore images with both single and mixed degradations. To address degradation diversity, we design a Local Dynamic Optimization (LDO) module which dynamically processes degraded areas of varying types and granularities. To tackle the prompt singularity issue, we develop an efficient Conditional Feature Embedding (CFE) module that guides the decoder in leveraging degradation-type-related features, significantly improving the model's performance in mixed degradation restoration scenarios. To validate the effectiveness of our model, we introduce a new dataset containing both single and mixed degradation elements. Experimental results demonstrate that our proposed model achieves state-of-the-art (SOTA) performance not only on mixed degradation tasks but also on classic single-task restoration benchmarks.

CATP-LLM: Empowering Large Language Models for Cost-Aware Tool Planning

Utilizing large language models (LLMs) for tool planning has emerged as a promising avenue for developing general AI systems, where LLMs automatically schedule external tools (e.g. vision models) to tackle complex tasks based on task descriptions. To push this paradigm toward practical applications, it is crucial for LLMs to consider tool execution costs (e.g. execution time) for tool planning. Unfortunately, prior studies overlook the tool execution costs, leading to the generation of expensive plans of which the costs outweigh task performance. To fill this gap, we propose the Cost-Aware Tool Planning with LLMs (CATP-LLM) framework, which for the first time provides a coherent design to empower LLMs for cost-aware tool planning. Specifically, CATP-LLM incorporates a tool planning language to enhance the LLM to generate non-sequential plans of multiple branches for efficient concurrent tool execution and cost reduction. Moreover, it further designs a cost-aware offline reinforcement learning algorithm to fine-tune the LLM to optimize the performance-cost trade-off in tool planning. In lack of public cost-related datasets, we further present OpenCATP, the first platform for cost-aware planning evaluation. Experiments on OpenCATP show that CATP-LLM outperforms GPT-4 even when using Llama2-7B as its backbone, with the average improvement of 28.2%-30.2% higher plan performance and 24.7%-45.8% lower costs even on the challenging planning tasks. The codes of CATP-LLM and OpenCATP will be publicly available.

RTF-Q: Unsupervised domain adaptation based retraining-free quantization network

Performing unsupervised domain adaptation on resource-constrained edge devices is a significant task. Although existing research allows edge devices to use subnets with different computational budgets for inference, they often require expensive pre-training and do not consider the issues of parameter precision redundancy in the model, which is not conducive to the deployment of the model on edge devices. In this paper, we introduce a ReTraining-Free Quantized (RTF-Q) network based on unsupervised domain adaptation, featuring quantized subnets of varying computational costs that can operate on devices with dynamically changing computation budgets. Our network has three switchable dimensions: width (number of channels), input resolution, and quantization bit-width. Specifically, we choose subnet dimensions that have minimal impact on network performance and then directly load the official weight files without requiring expensive and time-consuming pre-training on Imagenet-1K. To further reduce the network's computational load and memory usage, we use quantization-aware training, reducing the BitOPs of full-precision networks by at least 1/16. We propose a training method called SandwichQ for multiple quantization bit widths, which can efficiently train multiple quantization subnets. By training in multiple quantization bit-width spaces simultaneously and using the proposed SandwichQ rule, we achieve better network performance compared to using a single quantization bit-width alone. Experimental results show that our method achieves classification accuracy comparable to SOTA methods on various UDA tasks, significantly reducing network size and computational overhead. Code will be available at https://github.com/dunanyang/RTF-Q.

PRANCE: Joint Token-Optimization and Structural Channel-Pruning for Adaptive ViT Inference

We introduce PRANCE, a Vision Transformer compression framework that jointly optimizes the activated channels and reduces tokens, based on the characteristics of inputs. Specifically, PRANCE~ leverages adaptive token optimization strategies for a certain computational budget, aiming to accelerate ViTs' inference from a unified data and architectural perspective. However, the joint framework poses challenges to both architectural and decision-making aspects. Firstly, while ViTs inherently support variable-token inference, they do not facilitate dynamic computations for variable channels. To overcome this limitation, we propose a meta-network using weight-sharing techniques to support arbitrary channels of the Multi-head Self-Attention and Multi-layer Perceptron layers, serving as a foundational model for architectural decision-making. Second, simultaneously optimizing the structure of the meta-network and input data constitutes a combinatorial optimization problem with an extremely large decision space, reaching up to around $10^{14}$, making supervised learning infeasible. To this end, we design a lightweight selector employing Proximal Policy Optimization for efficient decision-making. Furthermore, we introduce a novel "Result-to-Go" training mechanism that models ViTs' inference process as a Markov decision process, significantly reducing action space and mitigating delayed-reward issues during training. Extensive experiments demonstrate the effectiveness of PRANCE~ in reducing FLOPs by approximately 50\%, retaining only about 10\% of tokens while achieving lossless Top-1 accuracy. Additionally, our framework is shown to be compatible with various token optimization techniques such as pruning, merging, and sequential pruning-merging strategies. The code is available at \href{https://github.com/ChildTang/PRANCE}{https://github.com/ChildTang/PRANCE}.

Q-DiT: Accurate Post-Training Quantization for Diffusion Transformers

Recent advancements in diffusion models, particularly the trend of architectural transformation from UNet-based Diffusion to Diffusion Transformer (DiT), have significantly improved the quality and scalability of image synthesis. Despite the incredible generative quality, the large computational requirements of these large-scale models significantly hinder the deployments in real-world scenarios. Post-training Quantization (PTQ) offers a promising solution by compressing model sizes and speeding up inference for the pretrained models while eliminating model retraining. However, we have observed the existing PTQ frameworks exclusively designed for both ViT and conventional Diffusion models fall into biased quantization and result in remarkable performance degradation. In this paper, we find that the DiTs typically exhibit considerable variance in terms of both weight and activation, which easily runs out of the limited numerical representations. To address this issue, we devise Q-DiT, which seamlessly integrates three techniques: fine-grained quantization to manage substantial variance across input channels of weights and activations, an automatic search strategy to optimize the quantization granularity and mitigate redundancies, and dynamic activation quantization to capture the activation changes across timesteps. Extensive experiments on the ImageNet dataset demonstrate the effectiveness of the proposed Q-DiT. Specifically, when quantizing DiT-XL/2 to W8A8 on ImageNet 256x256, Q-DiT achieves a remarkable reduction in FID by 1.26 compared to the baseline. Under a W4A8 setting, it maintains high fidelity in image generation, showcasing only a marginal increase in FID and setting a new benchmark for efficient, high-quality quantization in diffusion transformers. Code is available at \href{https://github.com/Juanerx/Q-DiT}{https://github.com/Juanerx/Q-DiT}.

Towards Lightweight Graph Neural Network Search with Curriculum Graph Sparsification

Graph Neural Architecture Search (GNAS) has achieved superior performance on various graph-structured tasks. However, existing GNAS studies overlook the applications of GNAS in resource-constraint scenarios. This paper proposes to design a joint graph data and architecture mechanism, which identifies important sub-architectures via the valuable graph data. To search for optimal lightweight Graph Neural Networks (GNNs), we propose a Lightweight Graph Neural Architecture Search with Graph SparsIfication and Network Pruning (GASSIP) method. In particular, GASSIP comprises an operation-pruned architecture search module to enable efficient lightweight GNN search. Meanwhile, we design a novel curriculum graph data sparsification module with an architecture-aware edge-removing difficulty measurement to help select optimal sub-architectures. With the aid of two differentiable masks, we iteratively optimize these two modules to efficiently search for the optimal lightweight architecture. Extensive experiments on five benchmarks demonstrate the effectiveness of GASSIP. Particularly, our method achieves on-par or even higher node classification performance with half or fewer model parameters of searched GNNs and a sparser graph.

One QuantLLM for ALL: Fine-tuning Quantized LLMs Once for Efficient Deployments

Large Language Models (LLMs) have advanced rapidly but face significant memory demands. While quantization has shown promise for LLMs, current methods typically require lengthy training to alleviate the performance degradation from quantization loss. However, deploying LLMs across diverse scenarios with different resource constraints, e.g., servers and personal computers, requires repeated training per application, which amplifies the lengthy training problem. Given that, it is advantageous to train a once-for-all (OFA) supernet capable of yielding diverse optimal subnets for downstream applications through one-shot training. Nonetheless, the scale of current language models impedes efficiency and amplifies interference from weight sharing between subnets. We make an initial attempt to extend the once-for-all framework to large language models. Specifically, we decouple shared weights to eliminate the interference and incorporate Low-Rank adapters for training efficiency. Furthermore, we observe the imbalance allocation of training resources from the traditional uniform sampling. A non-parametric scheduler is introduced to adjust the sampling rate for each quantization configuration, achieving a more balanced allocation among subnets with varying demands. We validate the approach on LLaMA2 families, and downstream evaluation confirms our ability to maintain high performance while significantly reducing deployment time faced with multiple scenarios.

LLM-Enhanced Causal Discovery in Temporal Domain from Interventional Data

In the field of Artificial Intelligence for Information Technology Operations, causal discovery is pivotal for operation and maintenance of graph construction, facilitating downstream industrial tasks such as root cause analysis. Temporal causal discovery, as an emerging method, aims to identify temporal causal relationships between variables directly from observations by utilizing interventional data. However, existing methods mainly focus on synthetic datasets with heavy reliance on intervention targets and ignore the textual information hidden in real-world systems, failing to conduct causal discovery for real industrial scenarios. To tackle this problem, in this paper we propose to investigate temporal causal discovery in industrial scenarios, which faces two critical challenges: 1) how to discover causal relationships without the interventional targets that are costly to obtain in practice, and 2) how to discover causal relations via leveraging the textual information in systems which can be complex yet abundant in industrial contexts. To address these challenges, we propose the RealTCD framework, which is able to leverage domain knowledge to discover temporal causal relationships without interventional targets. Specifically, we first develop a score-based temporal causal discovery method capable of discovering causal relations for root cause analysis without relying on interventional targets through strategic masking and regularization. Furthermore, by employing Large Language Models (LLMs) to handle texts and integrate domain knowledge, we introduce LLM-guided meta-initialization to extract the meta-knowledge from textual information hidden in systems to boost the quality of discovery. We conduct extensive experiments on simulation and real-world datasets to show the superiority of our proposed RealTCD framework over existing baselines in discovering temporal causal structures.