NVIDIA DALI

Back to troubleshooting start

Note

Nvidia DALI info and considerations:

NVIDIA DALI installation

You have to check NVIDIA DALI Path, it is needed.

NVIDIA DALI Path

installing DALI with TF2 CUDA11

#for DALI
pip install nvidia-dali-cuda110
#for TF2 plugin
pip install nvidia-dali-tf-plugin-cuda110

installing DALI with TF2 CUDA12

#for DALI
pip install nvidia-dali-cuda120
#for TF2 plugin
pip install nvidia-dali-tf-plugin-cuda120

I used CUDA12, and had a lot of problems to get DALI working, I downloaded oauthlib 3.2.2 then (TF2.16.2+Cuda/DALI 1.47+Plugin), any newer TF2 version is not supported(Bug that DALI has, they are aware of it on their Github), anything older than 2.16 has compatibility issues, so it took time to find the correct one, also TF2 pip, not Conda TF-GPU. This is the situation in the time of writing (April 2025).

Note

DALI may not be always faster than base dataloaders in TensorFlow or PyTorch, here is an example where PyTorch dataloader was faster.

In some cases, DALI may not be able to hide the latency of access to the files on the network drive (as it uses only one thread to perform the read operation while the PyTorch dataloader may read as many files in parallel as it has threads). On top of that it mmaps files to avoid unnecessary memcopies. You can set dont_use_mmap=True and see if that helps.

NVIDIA DALI Path