Projects

GPU kernel and LLM-inference projects I've built or am actively working on. Code on GitHub.

• 

  FlashAttention from scratch

  Active

  A fused, IO-aware attention kernel in CUDA — tiling Q·Kᵀ through shared memory so the full score matrix never touches HBM.

  • CUDA
  • Attention
  • Tensor Cores

  GitHub
• 

  Tiled GEMM kernel

  Active

  A from-scratch HGEMM/SGEMM approaching cuBLAS via shared-memory tiling, register blocking, and vectorized loads — with a roofline write-up.

  • CUDA
  • GEMM
  • Performance

  GitHub
• 

  llm-infer-bench

  Active

  A reproducible LLM inference benchmark harness: tokens/s, TTFT, and roofline utilization across batch sizes, sequence lengths, and quantization.

  • Inference
  • Benchmarking
  • Python

  GitHub
• 

  Nsight case studies

  Past

  Annotated Nsight Compute walkthroughs — turning profiler counters (SOL, stalls, the in-tool roofline) into measured kernel speedups.

  • Profiling
  • Nsight
  • CUDA

  GitHub

More on GitHub: @ai-hpc