M.Tech Computer Science, IIIT Lucknow · Research Intern, PACE Lab at IIT Madras

I work on GPU performance and efficient LLM inference. My projects span CUDA and Triton kernels, profiling, and experiments with prefill/decode scheduling on a single GPU. I like understanding where time is spent before changing the code.

• LLM inference on NVIDIA L40S: Measuring prefill and decode separately, then studying how SM allocation with CUDA Green Contexts affects their overlap and end-to-end throughput. I compare against a serial baseline and use repeat runs and Nsight traces to investigate gaps between expected and measured performance.
• GPU performance at IIT Madras: Studying clocks, power, and temperature under sustained compute and memory workloads.

• NVIDIA cuDF — merged PR #21068: Helped expose the native split_part operation through Cython/Python bindings. Separately, I built a lazy string-extraction prototype that measured 137 ms → 0.85 ms on a 5-million-row Tesla T4 benchmark. The prototype's speedup is not a performance claim for the merged PR.
• Triton extensions — merged PR #29: Worked on masked-load behavior in the triton-loop-split MLIR pass and added FileCheck coverage.
• PyTorch — Green Context issue #195378: Reported an L40S SM/TPC allocation observation and requested better visibility into the actual resources assigned to green contexts.

• FP8 MoE kernels on B200: Triton quantization, GEMM, and scatter kernels for a FlashInfer-Bench contest submission. The repository reports up to 4.68× versus its BF16 PyTorch reference on a Grok-style configuration and 832 TFLOP/s on a large-batch configuration. The DeepSeek-style configuration was slower than the reference; the repository includes that result too.
• H100 GEMM optimization: CUDA and Triton matrix multiplication experiments with tile tuning and profiling.
• Learning MLIR: Notes and small examples on IR, dialects, and compiler passes.

Python · C++/CUDA · Triton · PyTorch · MLIR · Nsight Systems · Nsight Compute · NVML · Linux

LinkedIn · Portfolio