Mohammad Sharifkhani

Sajjad Kachuee, Mohammad Sharifkhani

Mixture-of-Experts (MoE) embedding models combine expert outputs using weighted linear summation, implicitly assuming a linear subspace structure in the embedding space. This assumption is shown to be inconsistent with the geometry of expert representations. Geometric analysis of a modern MoE embedding model reveals that expert outputs lie on a shared hyperspherical manifold characterized by tightly concentrated norms and substantial angular separation. Under this geometry, linear aggregation induces inward collapse toward the manifold interior, distorting vector magnitude and direction and reducing embedding comparability. To address this inconsistency, Spherical Barycentric Aggregation (SBA) is introduced as a geometry-preserving aggregation operator that separates radial and angular components to maintain hyperspherical structure while remaining fully compatible with existing routing mechanisms. Experiments on selected tasks from the Massive Text Embedding Benchmark (MTEB), including semantic similarity, clustering, and duplicate question detection, demonstrate consistent performance improvements with identical training cost and full stability. Additional geometric analyses confirm that SBA prevents aggregation-induced collapse and preserves hyperspherical consistency, highlighting the importance of geometry-aware aggregation in MoE embedding architectures.

Sajjad Kachuee, Mohammad Sharifkhani

Using Large Language Models (LLMs) in real-world applications presents significant challenges, particularly in balancing computational efficiency with model performance. Optimizing acceleration after fine-tuning and during inference is critical for building efficient architectures. This paper introduces Zero-Shot Adjustable Acceleration, a novel training and inference method that dynamically adjusts hardware utilization during inference without requiring additional fine-tuning. The proposed approach is applied to recent LLMs and evaluated across multiple classification and text generation tasks. Experimental results demonstrate that the method supports a wide range of zero-shot acceleration and achieves up to 11x speedup compared to the baseline.

Sajjad Kachuee, Mohammad Sharifkhani

Adjusting the latency, power, and accuracy of natural language understanding models is a desirable objective of an efficient architecture. This paper proposes an efficient Transformer architecture that adjusts the inference computational cost adaptively with a desired inference latency speedup. In fine-tuning phase, the proposed method detects less important hidden sequence elements (word-vectors) and eliminates them in each encoder layer using a proposed Attention Context Contribution (ACC) metric. After the fine-tuning phase, with the novel offline-tuning property, the inference latency of the model can be adjusted in a wide range of inference speedup selections without any further training. Extensive experiments reveal that most word-vectors in higher Transformer layers contribute less to subsequent layers, allowing their removal to improve inference latency. Experimental results on various language understanding, text generation, and instruction tuning tasks and benchmarks demonstrate the approach's effectiveness across diverse datasets, with minimal impact on the input's global context. The technique improves Time-to-First-Token (TTFT) of Llama3 by up to 2.9x, with minor performance drop. The suggested approach posits that in Large Language Models (LLMs), although the complete network is necessary for training, it can be truncated during the fine-tuning phase.