Daniel Alexander

Shiqi Huang, Yipei Wang, Natasha Thorley et al.

Foundation models leverage large-scale pretraining to capture extensive knowledge, demonstrating generalization in a wide range of language tasks. By comparison, vision foundation models (VFMs) often exhibit uneven improvements across downstream tasks, despite substantial computational investment. We postulate that this limitation arises from a mismatch between pretraining objectives and the demands of downstream vision-and-imaging tasks. Pretraining strategies like masked image reconstruction or contrastive learning shape representations for tasks such as recovery of generic visual patterns or global semantic structures, which may not align with the task-specific requirements of downstream applications including segmentation, classification, or image synthesis. To investigate this in a concrete real-world clinical area, we assess two VFMs, a reconstruction-focused MAE-based model (ProFound) and a contrastive-learning-based model (ProViCNet), on five prostate multiparametric MR imaging tasks, examining how such task alignment influences transfer performance, i.e., from pretraining to fine-tuning. Our findings indicate that better alignment between pretraining and downstream tasks, measured by simple divergence metrics such as maximum-mean-discrepancy (MMD) between the same features before and after fine-tuning, correlates with greater performance improvements and faster convergence, emphasizing the importance of designing and analyzing pretraining objectives with downstream applicability in mind.

Xuting Liu, Daniel Alexander, Siva Kesava Reddy Kakarla et al.

Early-Exit (EE) is a Large Language Model (LLM) architecture that accelerates inference by allowing easier tokens to be generated using only a subset of the model's layers. However, traditional batching frameworks are ill-suited for EE LLMs, as not all requests in a batch may be ready to exit at the same time. Existing solutions either force a uniform decision on the batch, which overlooks EE opportunities, or degrade output quality by forcing premature exits. We propose Dynamic Rebatching, a solution where we dynamically reorganize the batch at each early-exit point. Requests that meet the exit criteria are immediately processed, while those that continue are held in a buffer, re-grouped into a new batch, and forwarded to deeper layers. We introduce DREX, an early-exit inference system that implements Dynamic Rebatching with two key optimizations: 1) a copy-free rebatching buffer that avoids physical data movement, and 2) an EE and SLA-aware scheduler that analytically predicts whether a given rebatching operation will be profitable. DREX also efficiently handles the missing KV cache from skipped layers using memory-efficient state-copying. Our evaluation shows that DREX improves throughput by 2-12% compared to baseline approaches while maintaining output quality. Crucially, DREX completely eliminates involuntary exits, providing a key guarantee for preserving the output quality intended by the EE model.