arXiv subjects

Krishna Teja Chitty-Venkata

Publications and source records attributed to Krishna Teja Chitty-Venkata.

12 recordsLinked to original sources

arXiv2026

H-Spec: Parallel Speculative Decoding Without a Drafter-Side KV Cache

Speculative decoding losslessly accelerates large language model inference by having a lightweight draft model predict future tokens for verification by the target model. Recent block diffusion drafters further reduce drafting latency by predicting multiple tokens in parallel. However, existing block drafters project target hidden states at every input position into a separate drafter-side KV cache, incurring per-request memory and KV-write overhead that grow with concurrency; directly reusing target KVs in place removes this cache but fails to sustain draft quality throughout the block. We propose a hybrid target-context injection method that complements direct target KV reuse with target hidden states only at the last input position, requiring no separate drafter-side KV cache. Building on this design, we propose H-Spec, a hybrid Mamba-attention parallel drafter that consumes the two target-context sources through complementary modules. Mamba modules are initialized with projected last-token target hidden states, while attention modules reuse target KVs in place. Despite its recurrent formulation, Mamba's parallel scan allows H-Spec to preserve block-parallel drafting. Across three target models and diverse tasks, H-Spec improves over the best baseline by 5.0--13.3% in mean accepted length and 5.3--12.6% in batch-size-1 inter-token latency speedup. Under concurrent serving, H-Spec consistently achieves higher throughput while maintaining lower KV cache utilization than baselines across evaluated concurrency levels.

cs.LG↗

arXiv2026

PreLoRA: Hybrid Pre-training of Vision Transformers with Full Training and Low-Rank Adapters

Training large models ranging from millions to billions of parameters is highly resource-intensive, requiring significant time, compute, and memory. It is observed that most of the learning (higher change in weights) takes place in the earlier stage of the training loop. As training progresses, these changes stabilize, suggesting that the resulting updates may be amenable to approximation using low intrinsic-rank matrices. Therefore, we propose an approach to identify such states of partial convergence and dynamically switch from full parameter training to Low Rank Adaptation (LoRA) on the ViT-Large model. We introduce a flexible approach that leverages user-defined hyperparameters to determine the switching point and assign a rank specific to each module layer based on its level of convergence. Experimental results show that this approach preserves model accuracy while reducing the number of trainable parameters to 10% of its original size, resulting in a 3x improvement in throughput, and a 1.5x reduction in average training time per epoch while also reducing GPU memory consumption by 20%.

cs.LG↗

arXiv2026

Swimba: Switch Mamba Model Scales State Space Models

Mixture-of-experts (MoE) is a common approach for increasing parameter capacity, but applying MoE to state space model (SSM) token mixers can multiply the cost of the recurrent state update. We study how to introduce expert specialization into selective SSMs while preserving computational efficiency. We show that MoE--SSM can refer to two designs: (1) MoE over separated SSMs, which maintains multiple state trajectories and thus scales compute with the number of experts; and (2) MoE-parameterized SSM, which mixes experts in parameter space, maintains a single state trajectory, and evaluates the recurrence once. Our method, Switch Mamba (Swimba), follows the second design by routing over expert-produced SSM streams. Theoretically, we establish well-definedness and stability for MoE-parameterized SSMs and characterize the relationship between the two designs. Empirically, we evaluate Swimba on standard benchmark tasks and measure real-time throughput and latency. Under matched FLOPs, Swimba achieves slightly better average performance than the baseline, with a small slowdown in real-time latency and throughput. Overall, these results suggest that parameter-space MoE can increase SSM capacity while keeping the dominant recurrence cost fixed.

cs.LG↗

arXiv2025

ImageNet-Think-250K: A Large-Scale Synthetic Dataset for Multimodal Reasoning for Vision Language Models

We develop ImageNet-Think, a multimodal reasoning dataset designed to aid the development of Vision Language Models (VLMs) with explicit reasoning capabilities. Our dataset is built on 250,000 images from ImageNet21k dataset, providing structured thinking tokens and corresponding answers. Our synthetic dataset is generated by two state-of-the-art VLMs: GLM-4.1V-9B-Thinking and Kimi-VL-A3B-Thinking-2506. Each image is accompanied by two pairs of thinking-answer sequences, creating a resource for training and evaluating multimodal reasoning models. We capture the step-by-step reasoning process of VLMs and the final descriptive answers. Our goal with this dataset is to enable the development of more robust VLMs while contributing to the broader understanding of multimodal reasoning mechanisms. The dataset and evaluation benchmarks will be publicly available to aid research in reasoning/thinking multimodal VLMs.

cs.CV↗

arXiv2025

PagedEviction: Structured Block-wise KV Cache Pruning for Efficient Large Language Model Inference

KV caching significantly improves the efficiency of Large Language Model (LLM) inference by storing attention states from previously processed tokens, enabling faster generation of subsequent tokens. However, as sequence length increases, the KV cache quickly becomes a major memory bottleneck. To address this, we propose PagedEviction, a novel fine-grained, structured KV cache pruning strategy that enhances the memory efficiency of vLLM's PagedAttention. Unlike existing approaches that rely on attention-based token importance or evict tokens across different vLLM pages, PagedEviction introduces an efficient block-wise eviction algorithm tailored for paged memory layouts. Our method integrates seamlessly with PagedAttention without requiring any modifications to its CUDA attention kernels. We evaluate PagedEviction across Llama-3.1-8B-Instruct, Llama-3.2-1B-Instruct, and Llama-3.2-3B-Instruct models on the LongBench benchmark suite, demonstrating improved memory usage with better accuracy than baselines on long context tasks.

cs.LG↗

arXiv2025

MoPEQ: Mixture of Mixed Precision Quantized Experts

Large Language and Vision Models using a Mixture-of-Experts (MoE) architecture pose significant challenges for deployment due to their computational and memory demands. Mixed Precision Quantization assigns different precisions to different layers of an LLM/VLM based on layer sensitivity and importance within the model. In this work, we propose a Post Training Quantization algorithm, MoPEQ, that assigns optimal bit width to each expert. Our method balances accuracy and model size by analyzing each expert's sensitivity using Hessian trace approximation instead of relying on the activation frequency of the expert. This per-expert granularity approach clusters similar experts to maintain model performance while reducing memory requirements. The experimental results on VLMEvalKit benchmark datasets using State-of-the-art VLMs Deepseek-VL2 -tiny, -small, -base, and MolmoE models demonstrate that our mixed precision quantized MoEs achieve competitive accuracy with substantial improvements in memory footprint compared to uniform-precision baseline methods. We perform a comprehensive study to analyze the impact of expert activation frequency and sensitivity using Hessian trace approximation at both layer-wise and model-wide expert precision allocation of 2, 3, and 4 bits to provide a thorough understanding of mixed precision quantization of VLM-MoEs.

cs.LG↗

arXiv2025

LExI: Layer-Adaptive Active Experts for Efficient MoE Model Inference

Mixture-of-Experts (MoE) models scale efficiently by activating only a subset of experts per token, offering a computationally sparse alternative to dense architectures. While prior post-training optimizations, such as inter- and intra-expert pruning, reduce memory usage they provide limited gains in inference-time compute efficiency. Moreover, existing MoE architectures typically activate a fixed number of experts uniformly across all layers, resulting in redundant computation and suboptimal performance. In this work, we first demonstrate that MoE pruning strategies improve only the memory footprint but do not significantly improve inference performance on GPU using optimized frameworks such as vLLM. To address this, we introduce LExI, a data-free optimization technique that determines the optimal number of active experts per layer in a pretrained MoE model. LExI leverages only the model weights to estimate the relative importance of each layer and adaptively assigns the number of active experts accordingly per layer. Experiments on state-of-the-art language and vision MoE benchmarks demonstrate that LExI significantly outperforms traditional MoE pruning approaches in terms of inference efficiency with negligible accuracy loss. For example, using LExI, Qwen1.5-MoE achieves the same throughput on Nvidia H100 GPU with 10% better accuracy than traditional expert pruning.

cs.LG↗

arXiv2025

MoE-Inference-Bench: Performance Evaluation of Mixture of Expert Large Language and Vision Models

Mixture of Experts (MoE) models have enabled the scaling of Large Language Models (LLMs) and Vision Language Models (VLMs) by achieving massive parameter counts while maintaining computational efficiency. However, MoEs introduce several inference-time challenges, including load imbalance across experts and the additional routing computational overhead. To address these challenges and fully harness the benefits of MoE, a systematic evaluation of hardware acceleration techniques is essential. We present MoE-Inference-Bench, a comprehensive study to evaluate MoE performance across diverse scenarios. We analyze the impact of batch size, sequence length, and critical MoE hyperparameters such as FFN dimensions and number of experts on throughput. We evaluate several optimization techniques on Nvidia H100 GPUs, including pruning, Fused MoE operations, speculative decoding, quantization, and various parallelization strategies. Our evaluation includes MoEs from the Mixtral, DeepSeek, OLMoE and Qwen families. The results reveal performance differences across configurations and provide insights for the efficient deployment of MoEs.

cs.LG↗

arXiv2025

LangVision-LoRA-NAS: Neural Architecture Search for Variable LoRA Rank in Vision Language Models

Vision Language Models (VLMs) integrate visual and text modalities to enable multimodal understanding and generation. These models typically combine a Vision Transformer (ViT) as an image encoder and a Large Language Model (LLM) for text generation. LoRA (Low-Rank Adaptation) is an efficient fine-tuning method to adapt pre-trained models to new tasks by introducing low-rank updates to their weights. While LoRA has emerged as a powerful technique for fine-tuning large models by introducing low-rank updates, current implementations assume a fixed rank, potentially limiting flexibility and efficiency across diverse tasks. This paper introduces \textit{LangVision-LoRA-NAS}, a novel framework that integrates Neural Architecture Search (NAS) with LoRA to optimize VLMs for variable-rank adaptation. Our approach leverages NAS to dynamically search for the optimal LoRA rank configuration tailored to specific multimodal tasks, balancing performance and computational efficiency. Through extensive experiments using the LLaMA-3.2-11B model on several datasets, LangVision-LoRA-NAS demonstrates notable improvement in model performance while reducing fine-tuning costs. Our Base and searched fine-tuned models on LLaMA-3.2-11B-Vision-Instruct can be found \href{https://huggingface.co/collections/krishnateja95/llama-32-11b-vision-instruct-langvision-lora-nas-6786cac480357a6a6fcc59ee}{\textcolor{blue}{here}} and the code for LangVision-LoRA-NAS can be found \href{https://github.com/krishnateja95/LangVision-NAS}{\textcolor{blue}{here}}.

cs.CV↗

arXiv2025

BaKlaVa -- Budgeted Allocation of KV cache for Long-context Inference

In Large Language Model (LLM) inference, Key-Value (KV) caches (KV-caches) are essential for reducing time complexity. However, they result in a linear increase in GPU memory as the context length grows. While recent work explores KV-cache eviction and compression policies to reduce memory usage, they often consider uniform KV-caches across all attention heads, leading to suboptimal performance. We introduce BaKlaVa, a method to allocate optimal memory for individual KV-caches across the model by estimating the importance of each KV-cache. Our empirical analysis demonstrates that not all KV-caches are equally critical for LLM performance. Using a one-time profiling approach, BaKlaVa assigns optimal memory budgets to each KV-cache. We evaluated our method on LLaMA-3-8B, and Qwen2.5-7B models, achieving up to a 70\% compression ratio while keeping baseline performance and delivering up to an order-of-magnitude accuracy improvement at higher compression levels.

cs.LG↗

arXiv2024

LLM-Inference-Bench: Inference Benchmarking of Large Language Models on AI Accelerators

Large Language Models (LLMs) have propelled groundbreaking advancements across several domains and are commonly used for text generation applications. However, the computational demands of these complex models pose significant challenges, requiring efficient hardware acceleration. Benchmarking the performance of LLMs across diverse hardware platforms is crucial to understanding their scalability and throughput characteristics. We introduce LLM-Inference-Bench, a comprehensive benchmarking suite to evaluate the hardware inference performance of LLMs. We thoroughly analyze diverse hardware platforms, including GPUs from Nvidia and AMD and specialized AI accelerators, Intel Habana and SambaNova. Our evaluation includes several LLM inference frameworks and models from LLaMA, Mistral, and Qwen families with 7B and 70B parameters. Our benchmarking results reveal the strengths and limitations of various models, hardware platforms, and inference frameworks. We provide an interactive dashboard to help identify configurations for optimal performance for a given hardware platform.

cs.LG↗

arXiv2023

A Survey of Techniques for Optimizing Transformer Inference

Recent years have seen a phenomenal rise in performance and applications of transformer neural networks. The family of transformer networks, including Bidirectional Encoder Representations from Transformer (BERT), Generative Pretrained Transformer (GPT) and Vision Transformer (ViT), have shown their effectiveness across Natural Language Processing (NLP) and Computer Vision (CV) domains. Transformer-based networks such as ChatGPT have impacted the lives of common men. However, the quest for high predictive performance has led to an exponential increase in transformers' memory and compute footprint. Researchers have proposed techniques to optimize transformer inference at all levels of abstraction. This paper presents a comprehensive survey of techniques for optimizing the inference phase of transformer networks. We survey techniques such as knowledge distillation, pruning, quantization, neural architecture search and lightweight network design at the algorithmic level. We further review hardware-level optimization techniques and the design of novel hardware accelerators for transformers. We summarize the quantitative results on the number of parameters/FLOPs and accuracy of several models/techniques to showcase the tradeoff exercised by them. We also outline future directions in this rapidly evolving field of research. We believe that this survey will educate both novice and seasoned researchers and also spark a plethora of research efforts in this field.

cs.LG↗