Varun Sundar
Chennai, Tamil Nadu, India
Summary
🤩
Rockstar🎓
Top SchoolVarun Sundar is an experienced software engineer based in Chennai with 11 years of hands-on expertise across C, C++ and Java and a strong interest in both machine learning research and teaching. He contributes to high-performance ML infrastructure—most notably optimizing int8 and fp8 Cutlass kernels and dynamic FP8 quantization in the popular vllm inference engine—demonstrating a focus on hardware-accelerated model serving and MLOps. Comfortable bridging research and production, he works at the intersection of NLP, information retrieval, program synthesis and computer vision. Varun pairs solid systems-level programming skills with a researcher’s curiosity, regularly tuning low-level kernels for SM80/SM90 architectures to squeeze out performance. He aspires to either deepen his impact as a top-tier software engineer or translate his knowledge into effective teaching for the next generation of developers. Based on his contributions, he brings pragmatic optimization experience that’s often invisible but critical to scalable ML deployments.
11 years of coding experience
Saveetha Engineering College
Github Skills (8)
cuda10
quantization10
pytorch10
mlops10
inference9
triton9
deepspeed8
transformer8
Programming languages (1)
Github contributions (5)
vllm-project/vllm
May 2024 - Apr 2025
A high-throughput and memory-efficient inference and serving engine for LLMs
Role in this project:
MLOps Engineer
MLOps EngineerContributions:186 reviews, 54 PRs, 302 comments in 11 months
Contributions summary:Varun's commits focus on updating and configuring Cutlass kernels, specifically for int8 and fp8 matrix multiplication operations. They are involved in optimizing these kernels for different architectures, including SM80 and SM90, indicating a focus on performance and hardware acceleration. The user's work involves modifying configurations and dispatch logic within the `csrc/quantization` directory, aligning with a role in managing and deploying optimized kernels. The contributions also involve the integration of dynamic-per-token FP8 quantization kernels.
amdcudagpthpuinference
neuralmagic/vllm
Aug 2024 - Apr 2025
A high-throughput and memory-efficient inference and serving engine for LLMs
Contributions:12 reviews, 3 PRs, 300 pushes in 7 months
Find and Hire Top DevelopersWe’ve analyzed the programming source code of over 60 million software developers on GitHub and scored them by 50,000 skills. Sign-up on Prog,AI to search for software developers.