TensorRT Engine
AI & Machine LearningENGINE can denote a TensorRT optimized inference engine, but the extension alone does not identify the format. TensorRT version, GPU, compatibility modes, input constraints and plugins influence runtime compatibility; establish build conditions.
- What it is used for
- Persisting optimized inference engines or producer-specific data
- What the file contains
- Producer-dependent
All sections, questions and sources
Content reviewed： · Editor：Kai
Opening steps
Choose your environment or purpose to see its steps.
Menu labels and support vary by app, OS and version. Preserve the original before making changes.
Troubleshooting by symptom
- Deserialization fails: review version/build compatibility.
- GPU differs: check default target and any configured compatibility mode.
- Input mismatch: review I/O types/shapes and optimization profiles.
- Deserialization fails on another GPU: inspect hardware compatibility mode and build conditions.
- Shape fails: inspect build-time minimum/optimal/maximum shapes and selected profile.
- Missing plugin: compare required build/runtime plugin libraries.
ENGINE → rebuild from the source model for the target environment
- Obtain the original model, build settings and required plugins.
- Build for the target environment and compare predictions/performance.
- Compare output shapes, values and timing with the same input before/after export; retain definitions and export settings.
Engines do not guarantee reconstruction of the original ONNX; renaming to PLAN does not change compatibility.
Compatibility of GPU artifacts
Engines depend on build conditions; version/hardware compatibility modes still require checking their supported scope.
Deserialization loads executable artifacts; avoid deserializing unknown files merely to inspect them.
Source for these steps：NVIDIA · Engine Compatibility / NVIDIA · Build your first engine / NVIDIA · TensorRT C++ API documentation
Record input profiles and precision
Dynamic input profiles constrain execution; arbitrary shapes may not be accepted.
Compare accuracy and performance on the same data after precision changes; retain source, build logs and reference outputs.
Source for these steps：NVIDIA · Engine Compatibility / NVIDIA · Build your first engine / NVIDIA · TensorRT C++ API documentation
Related-format differences
| Format | Main role | What to check |
|---|---|---|
| ENGINE / PLAN | TensorRT runtime artifact | Check build requirements |
| ONNX | Inference graph interchange | May serve as rebuild input |
| PT / KERAS | Source-framework serialization | Retain definitions/preprocessing |
Handling and security
TensorRT engines may contain executable code and plugin dependencies. Do not deserialize/run untrusted engines or enable host-code permissions merely to make an unknown plan load.
Questions and answers
Will it run on every GPU?
It depends on build targets and compatibility modes.
Is ENGINE always TensorRT?
Identify producer and content type.
Can it be converted back to ONNX?
Do not assume universal reverse conversion; obtain the original model.
Do ENGINE and PLAN suffixes require conversion?
Both are used as serialized-artifact names; inspect contents and producer workflow.
Do compatibility flags make it universal?
Modes have limits; check platform, version, hardware and plugin requirements.
Format facts: name, steward and MIME type
- Formal name
- TensorRT serialized engine (common use)
- Steward / specification
- NVIDIA TensorRT / producer-specific uses
- MIME type
- Depends on the producer or server; an extension alone does not establish a MIME type.
Primary and official sources
- Official documentationEngine Compatibility
Documents TensorRT/GPU compatibility and trusted engine handling.
NVIDIA · Link checked： - Official documentationBuild your first engine
Build and validate an engine from ONNX.
NVIDIA · Link checked： - Official documentationTensorRT C++ API documentation
Documents deserializing only trusted engines.
NVIDIA · Link checked：