Hossein Rajabzadeh

Hossein Rajabzadeh, Aref Jafari, Aman Sharma et al.

Large Language Models (LLMs), with their increasing depth and number of parameters, have demonstrated outstanding performance across a variety of natural language processing tasks. However, this growth in scale leads to increased computational demands, particularly during inference and fine-tuning. To address these challenges, we introduce EchoAtt, a novel framework aimed at optimizing transformer-based models by analyzing and leveraging the similarity of attention patterns across layers. Our analysis reveals that many inner layers in LLMs, especially larger ones, exhibit highly similar attention matrices. By exploiting this similarity, EchoAtt enables the sharing of attention matrices in less critical layers, significantly reducing computational requirements without compromising performance. We incorporate this approach within a knowledge distillation setup, where a pre-trained teacher model guides the training of a smaller student model. The student model selectively shares attention matrices in layers with high similarity while inheriting key parameters from the teacher. Our best results with TinyLLaMA-1.1B demonstrate that EchoAtt improves inference speed by 15\%, training speed by 25\%, and reduces the number of parameters by approximately 4\%, all while improving zero-shot performance. These findings highlight the potential of attention matrix sharing to enhance the efficiency of LLMs, making them more practical for real-time and resource-limited applications.

Hossein Rajabzadeh, Suyuchen Wang, Hyock Ju Kwon et al.

We employ a tool-interacting divide-and-conquer strategy enabling large language models (LLMs) to answer complex multimodal multi-hop questions. In particular, we harness the power of large language models to divide a given multimodal multi-hop question into unimodal single-hop sub-questions to be answered by the appropriate tool from a predefined set of tools. After all corresponding tools provide the LLM with their answers, the LLM generates the next relevant unimodal single-hop question. To increase the reasoning ability of LLMs, we prompt chatGPT to generate a tool-interacting divide-and-conquer dataset. This dataset is then used to efficiently finetune the corresponding LLM. To assess the effectiveness of this approach, we conduct an evaluation on two recently introduced complex question-answering datasets. The experimental analysis demonstrate substantial improvements over existing state-of-the-art solutions, indicating the efficacy and generality of our strategy

Hossein Rajabzadeh, Maryam Dialameh, Chul B. Park et al.

Autoregressive large language models (LLMs) are bottlenecked by sequential decoding, where each new token typically requires executing all transformer layers. Existing dynamic-depth and layer-skipping methods reduce this cost, but often rely on auxiliary routing mechanisms or incur accuracy degradation when bypassed layers are left uncompensated. We present \textbf{LoRA-Drop}, a plug-and-play inference framework that accelerates decoding by applying a \emph{temporal compute schedule} to a fixed subset of intermediate layers: on most decoding steps, selected layers reuse the previous-token hidden state and apply a low-rank LoRA correction, while periodic \emph{refresh} steps execute the full model to prevent drift. LoRA-Drop requires no routing network, is compatible with standard KV caching, and can reduce KV-cache footprint by skipping KV updates in droppable layers during LoRA steps and refreshing periodically. Across \textbf{LLaMA2-7B}, \textbf{LLaMA3-8B}, \textbf{Qwen2.5-7B}, and \textbf{Qwen2.5-14B}, LoRA-Drop achieves up to \textbf{2.6$\times$ faster decoding} and \textbf{45--55\% KV-cache reduction} while staying within \textbf{0.5 percentage points (pp)} of baseline accuracy. Evaluations on reasoning (GSM8K, MATH, BBH), code generation (HumanEval, MBPP), and long-context/multilingual benchmarks (LongBench, XNLI, XCOPA) identify a consistent \emph{safe zone} of scheduling configurations that preserves quality while delivering substantial efficiency gains, providing a simple path toward adaptive-capacity inference in LLMs. Codes are available at https://github.com/hosseinbv/LoRA-Drop.git.

Maryam Dialameh, Hossein Rajabzadeh, Jung Suk Sim et al.

Papillary thyroid microcarcinoma (PTMC) is increasingly managed with radio-frequency ablation (RFA), yet accurate lesion segmentation in ultrasound videos remains difficult due to low contrast, probe-induced motion, and heat-related artifacts. The recent Segment Anything Model 2 (SAM-2) generalizes well to static images, but its frame-independent design yields unstable predictions and temporal drift in interventional ultrasound. We introduce \textbf{EMA-SAM}, a lightweight extension of SAM-2 that incorporates a confidence-weighted exponential moving average pointer into the memory bank, providing a stable latent prototype of the tumour across frames. This design preserves temporal coherence through probe pressure and bubble occlusion while rapidly adapting once clear evidence reappears. On our curated PTMC-RFA dataset (124 minutes, 13 patients), EMA-SAM improves \emph{maxDice} from 0.82 (SAM-2) to 0.86 and \emph{maxIoU} from 0.72 to 0.76, while reducing false positives by 29\%. On external benchmarks, including VTUS and colonoscopy video polyp datasets, EMA-SAM achieves consistent gains of 2--5 Dice points over SAM-2. Importantly, the EMA pointer adds \textless0.1\% FLOPs, preserving real-time throughput of $\sim$30\,FPS on a single A100 GPU. These results establish EMA-SAM as a robust and efficient framework for stable tumour tracking, bridging the gap between foundation models and the stringent demands of interventional ultrasound. Codes are available here \hyperref[code {https://github.com/mdialameh/EMA-SAM}.

Maryam Dialameh, Hossein Rajabzadeh, Weiwei Zhang et al.

We present Bayesian Mixture of Experts (Bayesian-MoE), a post-hoc uncertainty estimation framework for fine-tuned large language models (LLMs) based on Mixture-of-Experts architectures. Our method applies a structured Laplace approximation to the second linear layer of each expert, enabling calibrated uncertainty estimation without modifying the original training procedure or introducing new parameters. Unlike prior approaches, which apply Bayesian inference to added adapter modules, Bayesian-MoE directly targets the expert pathways already present in MoE models, leveraging their modular design for tractable block-wise posterior estimation. We use Kronecker-factored low-rank approximations to model curvature and derive scalable estimates of predictive uncertainty and marginal likelihood. Experiments on common-sense reasoning benchmarks with Qwen1.5-MoE and DeepSeek-MoE demonstrate that Bayesian-MoE improves both expected calibration error (ECE) and negative log-likelihood (NLL) over baselines, confirming its effectiveness for reliable downstream decision-making.

Hossein Rajabzadeh, Mojtaba Valipour, Tianshu Zhu et al.

Finetuning large language models requires huge GPU memory, restricting the choice to acquire Larger models. While the quantized version of the Low-Rank Adaptation technique, named QLoRA, significantly alleviates this issue, finding the efficient LoRA rank is still challenging. Moreover, QLoRA is trained on a pre-defined rank and, therefore, cannot be reconfigured for its lower ranks without requiring further fine-tuning steps. This paper proposes QDyLoRA -Quantized Dynamic Low-Rank Adaptation-, as an efficient quantization approach for dynamic low-rank adaptation. Motivated by Dynamic LoRA, QDyLoRA is able to efficiently finetune LLMs on a set of pre-defined LoRA ranks. QDyLoRA enables fine-tuning Falcon-40b for ranks 1 to 64 on a single 32 GB V100-GPU through one round of fine-tuning. Experimental results show that QDyLoRA is competitive to QLoRA and outperforms when employing its optimal rank.

Maryam Dialameh, Hossein Rajabzadeh, Moslem Sadeghi-Goughari et al.

Precise segmentation of papillary thyroid microcarcinoma (PTMC) during ultrasound-guided radiofrequency ablation (RFA) is critical for effective treatment but remains challenging due to acoustic artifacts, small lesion size, and anatomical variability. In this study, we propose DualSwinUnet++, a dual-decoder transformer-based architecture designed to enhance PTMC segmentation by incorporating thyroid gland context. DualSwinUnet++ employs independent linear projection heads for each decoder and a residual information flow mechanism that passes intermediate features from the first (thyroid) decoder to the second (PTMC) decoder via concatenation and transformation. These design choices allow the model to condition tumor prediction explicitly on gland morphology without shared gradient interference. Trained on a clinical ultrasound dataset with 691 annotated RFA images and evaluated against state-of-the-art models, DualSwinUnet++ achieves superior Dice and Jaccard scores while maintaining sub-200ms inference latency. The results demonstrate the model's suitability for near real-time surgical assistance and its effectiveness in improving segmentation accuracy in challenging PTMC cases.

Hossam Amer, Maryam Dialameh, Hossein Rajabzadeh et al.

Scaling training compute, measured in FLOPs, has long been shown to improve the accuracy of large language models, yet training remains resource-intensive. Prior work shows that increasing test-time compute (TTC)-for example through iterative sampling-can allow smaller models to rival or surpass much larger ones at lower overall cost. We introduce TTC-aware training, where an intermediate checkpoint and a corresponding TTC configuration can together match or exceed the accuracy of a fully trained model while requiring substantially fewer training FLOPs. Building on this insight, we propose an early stopping algorithm that jointly selects a checkpoint and TTC configuration to minimize training compute without sacrificing accuracy. To make this practical, we develop an efficient TTC evaluation method that avoids exhaustive search, and we formalize a break-even bound that identifies when increased inference compute compensates for reduced training compute. Experiments demonstrate up to 92\% reductions in training FLOPs while maintaining and sometimes remarkably improving accuracy. These results highlight a new perspective for balancing training and inference compute in model development, enabling faster deployment cycles and more frequent model refreshes. Codes will be publicly released.

Maryam Dialameh, Rezaul Karim, Hossein Rajabzadeh et al.

This paper introduces ECHO-LLaMA, an efficient LLaMA architecture designed to improve both the training speed and inference throughput of LLaMA architectures while maintaining its learning capacity. ECHO-LLaMA transforms LLaMA models into shared KV caching across certain layers, significantly reducing KV computational complexity while maintaining or improving language performance. Experimental results demonstrate that ECHO-LLaMA achieves up to 77\% higher token-per-second throughput during training, up to 16\% higher Model FLOPs Utilization (MFU), and up to 14\% lower loss when trained on an equal number of tokens. Furthermore, on the 1.1B model, ECHO-LLaMA delivers approximately 7\% higher test-time throughput compared to the baseline. By introducing a computationally efficient adaptation mechanism, ECHO-LLaMA offers a scalable and cost-effective solution for pretraining and finetuning large language models, enabling faster and more resource-efficient training without compromising performance.

Benyamin Jamialahmadi, Parsa Kavehzadeh, Mehdi Rezagholizadeh et al.

Deploying large language models (LLMs) in real-world applications is often hindered by strict computational and latency constraints. While dynamic inference offers the flexibility to adjust model behavior based on varying resource budgets, existing methods are frequently limited by hardware inefficiencies or performance degradation. In this paper, we introduce Balcony, a simple yet highly effective framework for depth-based dynamic inference. By freezing the pretrained LLM and inserting additional transformer layers at selected exit points, Balcony maintains the full model's performance while enabling real-time adaptation to different computational budgets. These additional layers are trained using a straightforward self-distillation loss, aligning the sub-model outputs with those of the full model. This approach requires significantly fewer training tokens and tunable parameters, drastically reducing computational costs compared to prior methods. When applied to the LLaMA3-8B model, using only 0.2% of the original pretraining data, Balcony achieves minimal performance degradation while enabling significant speedups. Remarkably, we show that Balcony outperforms state-of-the-art methods such as Flextron and Layerskip as well as other leading compression techniques on multiple models and at various scales, across a variety of benchmarks.

Mojtaba Valipour, Mehdi Rezagholizadeh, Hossein Rajabzadeh et al.

Deep neural networks (DNNs) must cater to a variety of users with different performance needs and budgets, leading to the costly practice of training, storing, and maintaining numerous user/task-specific models. There are solutions in the literature to deal with single dynamic or many-in-one models instead of many individual networks; however, they suffer from significant drops in performance, lack of generalization across different model architectures or different dimensions (e.g. depth, width, attention blocks), heavy model search requirements during training, and training a limited number of sub-models. To address these limitations, we propose SortedNet, a generalized and scalable training solution to harness the inherent modularity of DNNs. Thanks to a generalized nested architecture (which we refer as \textit{sorted} architecture in this paper) with shared parameters and its novel update scheme combining random sub-model sampling and a new gradient accumulation mechanism, SortedNet enables the training of sub-models simultaneously along with the training of the main model (without any significant extra training or inference overhead), simplifies dynamic model selection, customizes deployment during inference, and reduces the model storage requirement significantly. The versatility and scalability of SortedNet are validated through various architectures and tasks, including LLaMA, BERT, RoBERTa (NLP tasks), ResNet and MobileNet (image classification) demonstrating its superiority over existing dynamic training methods. For example, we introduce a novel adaptive self-speculative approach based on sorted-training to accelerate large language models decoding. Moreover, SortedNet is able to train 160 sub-models at once, achieving at least 96\% of the original model's performance.

Hossein Rajabzadeh, Mansoor Zolghadri Jahromi, Mohammad Sadegh Zare et al.

Distance metric learning is a successful way to enhance the performance of the nearest neighbor classifier. In most cases, however, the distribution of data does not obey a regular form and may change in different parts of the feature space. Regarding that, this paper proposes a novel local distance metric learning method, namely Local Mahalanobis Distance Learning (LMDL), in order to enhance the performance of the nearest neighbor classifier. LMDL considers the neighborhood influence and learns multiple distance metrics for a reduced set of input samples. The reduced set is called as prototypes which try to preserve local discriminative information as much as possible. The proposed LMDL can be kernelized very easily, which is significantly desirable in the case of highly nonlinear data. The quality as well as the efficiency of the proposed method assesses through a set of different experiments on various datasets and the obtained results show that LDML as well as the kernelized version is superior to the other related state-of-the-art methods.