trt-engine-rebuild

name	trt-engine-rebuild
description	Use when building, rebuilding, or debugging TensorRT engines on this Jetson — shape profile errors ("does not satisfy any optimization profiles"), new ONNX exports, FP16 engines, or engine/version mismatch on deserialize.

TensorRT Engine Rebuild (Jetson Orin, TRT 10.3.0.30, SM87)

Hard rules

• Engines are tied to this exact machine (TRT 10.3.0.30 + sm87). Never copy .engine files between hosts. A Deserialize error usually means version/arch mismatch — rebuild from ONNX.
• trtexec lives at /usr/src/tensorrt/bin/trtexec (not on PATH).
• Set MAXN_SUPER before building or timing: sudo nvpmodel -m 2 && sudo jetson_clocks.

Shape profile errors

Error: Set dimension [1,3,H,W] for tensor X does not satisfy any optimization profiles. Valid range: [min]..[max] → the input exceeds the engine's max profile. Either resize the input upstream (VIC resize in RtspSource::connect_resized) or rebuild with a bigger --maxShapes.

Current xfeat profile covers 1080p: min 1x3x240x320, opt 1x3x640x640, max 1x3x1088x1920. Inputs must be multiples of 32 (use pad32).

Rebuild commands

# XFeat backbone (input tensor name: "image")
/usr/src/tensorrt/bin/trtexec \
    --onnx=models/xfeat/xfeat_backbone.onnx \
    --saveEngine=models/xfeat/xfeat_backbone_fp16.engine \
    --fp16 \
    --minShapes=image:1x3x240x320 \
    --optShapes=image:1x3x640x640 \
    --maxShapes=image:1x3x1088x1920 \
    --memPoolSize=workspace:2048

# Timing sanity check on an existing engine
/usr/src/tensorrt/bin/trtexec --loadEngine=<engine> --shapes=image:1x3x640x640

Inspecting an engine

/usr/src/tensorrt/bin/trtexec --loadEngine=<engine_path> --verbose 2>&1 | \
    grep -iA2 "input\|output\|profile" | head -30

Shows I/O tensor names, dtypes, and profile shapes — check this FIRST when inference fails with "unknown tensor" or shape errors. (Programmatic access: Engine exposes the same via the named-tensor API in crates/vrt.)

Performance expectations (MAXN_SUPER)

• XFeat backbone @ 640×640 opt shape: ~3ms GPU
• Larger inputs scale ~linearly with pixel area; running far from opt shape is slower than the area ratio suggests — prefer VIC-resizing input toward opt.
• Build time: ~5 min. Keep workspace at 2048 MB (unified memory — don't go higher).