Large Multimodal Models (LMMs) extend Large Language Models (LLMs) by handling diverse inputs such as images, audio, and video, but at the cost of adding a multimodal encoding stage that increases both computational and memory overhead. This step negatively affects key Service Level Objectives (SLOs), such as time to first token (TTFT) and time per output token (TPOT). We introduce Encode-Prefill-Decode (EPD) Disaggregation, a novel framework that separates the encoding, prefill, and decode stages onto dedicated resources. Unlike current systems, which bundle encoding and prefill together, our approach decouples these steps, unlocking new opportunities and optimizations. These include a mechanism to cache multimedia tokens for efficient transfer, a novel way to parallelize the encoding load within a request, a module for optimal resource allocation for disaggregated serving, and a novel role-switching method to handle changing workload characteristics. Experimental evaluations with popular LMMs show substantial gains in memory efficiency (up to 15$\times$ lower peak memory utilization), batch sizes (up to 22$\times$ larger), 10$\times$ more images per request, and 2.2$\times$ larger KV caches. Furthermore, it leads to significant improvements in SLO attainment (up to 90–100% improvement) and TTFT (up to 71% reduction), compared to systems that do not disaggregate. The code is available at https://github.com/vbdi/epdserve.

Huawei Cloud users leverage LoRA (Low-Rank Adaptation) as an efficient and scalable method to fine-tune and customize large language models (LLMs) for application-specific needs. However, tasks that require complex reasoning or deep contextual understanding are often hindered by biases or interference from the base model when using typical decoding methods like greedy or beam search. These biases can lead to generic or task-agnostic responses from the base model instead of leveraging the LoRA-specific adaptations. In this paper, we introduce Contrastive LoRA Decoding (CoLD), a novel decoding framework designed to maximize the use of task-specific knowledge in LoRA-adapted models, resulting in better downstream performance. CoLD uses contrastive decoding by scoring candidate tokens based on the divergence between the probability distributions of a LoRA-adapted expert model and the corresponding base model. This approach prioritizes tokens that better align with the LoRA's learned representations, enhancing performance for specialized tasks. While effective, a naive implementation of CoLD is computationally expensive because each decoding step requires evaluating multiple token candidates across both models. To address this, we developed an optimized kernel for Huawei's Ascend NPU. CoLD achieves up to a 5.54% increase in task accuracy while reducing end-to-end latency by 28% compared to greedy decoding. This work provides practical and efficient decoding strategies for fine-tuned LLMs in resource-constrained environments and has broad implications for applied data science in both cloud and on-premises settings.

Math Word Problem (MWP) is a crucial NLP task aimed at providing solutions for given mathematical descriptions. A notable sub-category of MWP is the Linear Programming Word Problem (LPWP), which holds significant relevance in real-world decision-making and operations research. While the recent rise of generative large language models (LLMs) has brought more advanced solutions to LPWPs, existing evaluation methodologies for this task still diverge from human judgment and face challenges in recognizing mathematically equivalent answers. In this paper, we introduce a novel evaluation metric rooted in graph edit distance, featuring benefits such as permutation invariance and more accurate program equivalence identification. Human evaluations empirically validate the superior efficacy of our proposed metric when particularly assessing LLM-based solutions for LPWP.

The multi-head self-attention mechanism of the transformer model has been thoroughly investigated recently.In one vein of study, researchers are interested in understanding why and how transformers work.In another vein, researchers propose new attention augmentation methods to make transformers more accurate, efficient and interpretable.In this paper, we combine these two lines of research in a human-in-the-loop pipeline to first discover important task-specific attention patterns.Then those patterns are injected, not only to smaller models, but also to the original model.The benefits of our pipeline and discovered patterns are demonstrated in two case studies with extractive summarization and topic segmentation.After discovering interpretable patterns in BERT-based models fine-tuned for the two downstream tasks, experiments indicate that when we inject the patterns into attention heads, the models show considerable improvements in accuracy and efficiency.

Dialogue topic segmentation is critical in several dialogue modeling problems.However, popular unsupervised approaches only exploit surface features in assessing topical coherence among utterances.In this work, we address this limitation by leveraging supervisory signals from the utterance-pair coherence scoring task.First, we present a simple yet effective strategy to generate a training corpus for utterance-pair coherence scoring.Then, we train a BERT-based neural utterance-pair coherence model with the obtained training corpus.Finally, such model is used to measure the topical relevance between utterances, acting as the basis of the segmentation inference 1 .Experiments on three public datasets in English and Chinese demonstrate that our proposal outperforms the state-of-the-art baselines.

Existing research on fairness evaluation of document classification models mainly uses synthetic monolingual data without ground truth for author demographic attributes. In this work, we assemble and publish a multilingual Twitter corpus for the task of hate speech detection with inferred four author demographic factors: age, country, gender and race/ethnicity. The corpus covers five languages: English, Italian, Polish, Portuguese and Spanish. We evaluate the inferred demographic labels with a crowdsourcing platform, Figure Eight. To examine factors that can cause biases, we take an empirical analysis of demographic predictability on the English corpus. We measure the performance of four popular document classifiers and evaluate the fairness and bias of the baseline classifiers on the author-level demographic attributes.

Linzi Xing, Michael J. Paul, Giuseppe Carenini. Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP). 2019.

Bayesian inference methods for probabilistic topic models can quantify uncertainty in the parameters, which has primarily been used to increase the robustness of parameter estimates. In this work, we explore other rich information that can be obtained by analyzing the posterior distributions in topic models. Experimenting with latent Dirichlet allocation on two datasets, we propose ideas incorporating information about the posterior distributions at the topic level and at the word level. At the topic level, we propose a metric called topic stability that measures the variability of the topic parameters under the posterior. We show that this metric is correlated with human judgments of topic quality as well as with the consistency of topics appearing across multiple models. At the word level, we experiment with different methods for adjusting individual word probabilities within topics based on their uncertainty. Humans prefer words ranked by our adjusted estimates nearly twice as often when compared to the traditional approach. Finally, we describe how the ideas presented in this work could potentially applied to other predictive or exploratory models in future work.