Aniruddha Tupe

Intern @Tenstorrent

Santa Clara, CA, US

Explored tt-metal and ttnn - tt-metal is Tenstorrent’s low-level AI hardware SDK offering near-metal control over custom kernels, while ttnn is the operator library built atop it for model support - and focused on optimizing LLM execution strategies using these frameworks.• Implemented DP = 8 for SentenceBERT and ViT, enabling robust data-parallel execution on the T3K hardware.• Integrated the multi-scale deformable attention module of the UniAD (autonomous driving) model within the ttnn framework, enhancing its functionality in the accelerated environment.• Worked on tt-transformers, enabling seamless out-of-the-box execution of LLMs with architectures like Llama, Qwen, or DeepSeek - bypassing the traditional model bring-up cycle.• Benchmarked finetuned LLMs via tt-inference-server - which orchestrates LLM containers using vLLM under the hood - together with lm-eval-harness to quantify performance and reliability.• Streamlined tool integration by integrating tt-studio with tt-inference-server, delegating Docker container creation to tt-inference-server for a more modular and automated deployment flow.

Savitribai Phule Pune University

Bachelor's degree, Information Technology

Rutgers University–New Brunswick

Master of Science - MS, Computer Science