Souvik Kundu
8 indexed papers
Publications per year
Top categories
Frequent co-authors
Research Timeline
The paper systematically analyzes the benefits and limits of Attention-FFN Disaggregation (AFD) for Mixture-of-Experts (MoE) LLM serving, demonstrating that AFD is crucial for achieving high throughput under strict latency constraints.
The paper systematically analyzes the benefits and limits of Attention-FFN Disaggregation (AFD) for Mixture-of-Experts (MoE) LLM serving, demonstrating that AFD is crucial for achieving high throughput under strict latency constraints.
SPARQLe is a hardware-software co-design framework that exploits the inherent sub-precision sparsity of LLM activations to reduce memory traffic and enable efficient computation on lower-bit datapaths, significantly accelerating inference.
SPARQLe is a hardware-software co-design framework that exploits the inherent sub-precision sparsity of LLM activations to reduce memory traffic and enable efficient computation on lower-bit datapaths, significantly accelerating inference.
MOSAIC is a novel scheduling framework that significantly accelerates Mixture-of-Agents (MoA) workloads by jointly optimizing expert placement and utilizing confidence-aware adaptive aggregation.
MOSAIC is a novel scheduling framework that significantly accelerates Mixture-of-Agents (MoA) workloads by jointly optimizing expert placement and utilizing confidence-aware adaptive aggregation.
The paper presents Dyserve, a workflow-aware serving layer for agentic AI applications that compiles per-node model and verifier choices into an integer linear program, allowing for efficient model selection and verification under serving-time conditions.
The paper introduces TRACE-Router, a task-level routing framework for large language models that aligns routing with the unit of supervision using delayed task feedback.
Papers
TRACE-ROUTER: Task-Consistent and Adaptive Online Routing for Agentic AI
Ritik Raj, Souvik Kundu, Sarbartha Banerjee, Dheemanth Joshi +2 more
The paper introduces TRACE-Router, a task-level routing framework for large language models that aligns routing with the unit of supervision using delayed task feedback.