User Guide

Conceptual guides and how-tos for Torch-TensorRT.

• Torch-TensorRT Explained
  • Dynamo Frontend
  • Legacy Frontends
  • torch_tensorrt.compile
• Compilation
  • TensorRT Backend for torch.compile
  • Example: Torch Compile Advanced Usage
  • Compiling Exported Programs with Torch-TensorRT
  • CompilationSettings Reference
  • Dynamic shapes with Torch-TensorRT
  • Example: Compiling Models with Dynamic Input Shapes
  • Example: Sharing Dynamic Dimensions Across Inputs
  • Handling Unsupported Operators
• Precision & Quantization
  • Compile Mixed Precision models with Torch-TensorRT
  • An example of using Torch-TensorRT Autocast
  • Quantization (INT8 / FP8 / FP4)
  • Deploy Quantized Models using Torch-TensorRT
  • Example: ViT FP8 Quantization
  • Example: TorchAO FP8 WOQ Linear
  • Example: TorchAO Static FP8 Linear
  • Example: FLUX.1-dev FP8 WOQ
  • Example: TorchAO INT4 WOQ Linear
  • Example: FLUX.1-dev INT4 WOQ
  • Example: Qwen3-8B INT4 WOQ
  • Example: TorchAO NVFP4 WOQ Linear
  • Example: FLUX.1-dev NVFP4 WOQ
  • Example: TorchAO MXFP4 Linear
  • Example: FLUX.1-dev MXFP4
• Runtime & Serialization
  • Deploying Torch-TensorRT Programs
  • Runtime API
  • Runtime Settings (TensorRT-RTX)
  • DLA
  • Saving models compiled with Torch-TensorRT
  • Extracting a Raw TensorRT Engine
  • AOTInductor Deployment
  • MutableTorchTensorRTModule
  • Example: Saving and Loading Models with Dynamic Shapes
  • Example: Saving Models with Dynamic Shapes - Both Methods
• Performance Tuning Guide
  • Common Benchmarking Issues
  • Using the Right Precision
  • Tuning opt_shape
  • Optimization Level
  • TRT Coverage and Graph Breaks
  • CUDA Graphs
  • Engine Caching
  • Memory and Throughput Tradeoffs
  • Profiling with Nsight
  • Benchmarking Checklist