TRTEXEC with CLIP

The trtexec tool is a command-line wrapper included as part of the TensorRT samples. Refer to the official TensorRT documentation for instructions on installing or building trtexec on your target platform (including x86 with discrete GPU and Jetson devices).

This section describes how to generate a TensorRT engine using trtexec, which allows you to deploy TAO-trained models on TensorRT, Triton, and Deepstream.

To generate an ONNX file for CLIP, refer to Exporting the Model.

Sample Command: Combined Engine

A combined CLIP engine contains both the vision and text encoders in a single TensorRT engine. Use this format when you run vision and text encoding together at inference time.

Warning

attention_mask is currently accepted as an explicit graph input for backward compatibility only. This input is deprecated and scheduled for removal. Remove it from your shape profiles and inference code to avoid a future breaking change.

trtexec \
  --onnx=/path/to/clip_model.onnx \
  --minShapes=image:1x3x256x256,input_ids:1x77,attention_mask:1x77 \
  --optShapes=image:8x3x256x256,input_ids:8x77,attention_mask:8x77 \
  --maxShapes=image:16x3x256x256,input_ids:16x77,attention_mask:16x77 \
  --fp16 \
  --saveEngine=/path/to/clip_model.engine

Sample Command: Separate Vision Engine

Export the vision encoder as a standalone engine when you want to pre-compute image embeddings independently of text encoding.

trtexec \
  --onnx=/path/to/clip_model_vision.onnx \
  --minShapes=image:1x3x256x256 \
  --optShapes=image:8x3x256x256 \
  --maxShapes=image:16x3x256x256 \
  --fp16 \
  --saveEngine=/path/to/clip_model_vision.engine

Sample Command: Separate Text Engine

Export the text encoder as a standalone engine when you want to pre-compute text embeddings independently — for example, to index a fixed set of captions or class names offline.

trtexec \
  --onnx=/path/to/clip_model_text.onnx \
  --minShapes=input_ids:1x77,attention_mask:1x77 \
  --optShapes=input_ids:8x77,attention_mask:8x77 \
  --maxShapes=input_ids:16x77,attention_mask:16x77 \
  --fp16 \
  --saveEngine=/path/to/clip_model_text.engine