ls -1 AI Infrastructure & GPU Study
Hands-on experiments exploring CUDA streams, PyTorch and vLLM inference, and numerical errors in CUDA graphs. Includes reproducible examples and profiling notes from NVIDIA GPU hardware.
- CUDA
- PyTorch
- vLLM
- GPU Profiling
Selected studies
2 studiesFocused experiments, reproducible examples, and visual explanations.
cuda-streams/ EXPERIMENT
CUDA Streams
From CUDA microbenchmarks to real inference: study concurrency, hidden serialization, and the streams used by PyTorch and vLLM.
- CUDA
- PyTorch
- vLLM
- Nsight Systems
cuda-graph-errors/ VISUAL REPORT
CUDA Graph Numerical Errors
Explore what CUDA graph replay freezes, why numerical errors occur, and how to fix them through code comparisons, diagrams, and measured results.
- CUDA Graphs
- PyTorch
- Numerical Correctness