How to Read Data into GPU Memory in Python with NVIDIA GPUDirect Storage
GPU applications are processing ever-growing datasets with increasingly powerful accelerators. According to the 2025 Stanford AI Index, compute for notable AI...
First seen by Oriel on .
AI summary
The tutorial shows how to read files into GPU memory with NVIDIA KvikIO and verify whether cuFile used the accelerated GPUDirect Storage path or a CPU-staged compatibility path.
- KvikIO provides Python and C++ interfaces over cuFile, but KvikIO and cuFile make separate decisions about whether and how to use GDS.
- The example reads a 1 GiB file into a CuPy GPU buffer; level-three cuFile statistics showed 256 nvfs reads and zero posix operations with the tutorial configuration.
- KvikIO defaults to 4 MiB tasks, splitting the read into 256 cuFile requests; KvikIO task settings and cuFile execution settings independently affect request decomposition and scheduling.
- Setting KVIKIO_COMPAT_MODE=on bypasses cuFile, while CUFILE_FORCE_COMPAT_MODE=true makes cuFile use its CPU-staged compatibility path.
- The tutorial was tested on Linux with an NVIDIA H100, CUDA 13.3, cuFile 1.18, local ext4 NVMe storage, and the nvidia-fs kernel module.
#gpudirect storage #kvikio #cuda #gpu io
Generated by gpt-5.6-sol. Check the original for details.
Aliases
| Source | External ID | URL |
|---|---|---|
| NVIDIA Developer Blog | https://developer.nvidia.com/blog/?p=123962 | https://developer.nvidia.com/blog/how-to-read-data-into-gpu-memory-in-python-with-nvidia-gpudirect-storage/ Read discussion |