Sujan Kumar Gonugondla

This paper proposes a effective way to implement speculative sampling under batched settings.

This paper proposes a novel approach for token alignment in subword completion tasks, leveraging character-level matching techniques to improve accuracy and efficiency.

This paper introduces Bifurcated Attention for Single-Context Large-Batch Sampling and generalized multi-group attention mechanisms to improve the IO efficiency of attention computations in deep learning models.

This work presents a comprehensive evaluation of code generation models across multiple programming languages, providing insights into their performance and generalizability.

This research explores the application of quantization techniques to large code generation models, aiming to reduce their environmental impact while maintaining performance.

IMPQ is a novel technique for reducing the complexity of neural networks by assigning granular precision to weights and activations, leading to more efficient neural-nets.

This work investigates the fundamental limits on energy-delay-accuracy tradeoffs in in-memory architectures for inference applications, providing insights into for system design using analog computing architectures.

Explore my complete list of publications on Google Scholar to learn more about my research contributions.