Bottom Line: It is 100% safe to clear NVIDIA Triton Inference Server temporary caches. Triton automatically loads and compiles model engines on service startup.
Why Does Triton Server Cache Exist?
NVIDIA Triton Inference Server deploys production AI models (TensorRT, ONNX Runtime, PyTorch). It caches compiled TensorRT engine plan files (.plan, .engine) in /var/tmp/triton/cache.
- Primary Purpose: GPU engine compilation caching for ultra-low latency model inference.
- Storage Growth: Re-compiling models across GPU architectures generates 5 GB to 40 GB+ of engine cache files.
What Happens If You Delete Triton Cache?
- System Safety: ✅ 100% Safe to Delete. Original model repository files (
model.onnx,config.pbtxt) remain untouched. - Re-compilation: Triton will re-compile TensorRT engines when the model repository loads.
- Reclaimed Storage: Frees 5 GB to 40 GB+ of server SSD storage.
How to Clear Triton Server Cache
# Delete Triton temporary engine cache
sudo rm -rf /var/tmp/triton/cache/*Frequently Asked Questions (FAQ)
Why is Triton startup slow after clearing cache?
TensorRT optimizes neural network layers specifically for your GPU during initial startup, which takes 2–5 minutes per model before caching.
Discussion
Loading authentication...