TensorRT 10.3 FP16 inference produces incorrect segmentation results for DINOv3 ViT models while TensorRT 8.6.1 works correctly

Description

I trained a semantic segmentation model based on DINOv3 ViT backbone using the Lightly_train framework. The model uses pretrained weights from ViT-B16 and ViT-L16 and is fine-tuned for a segmentation task.
However, I encountered inconsistent inference results when converting the model to TensorRT with different TensorRT versions.

Case 1: ViT-B16 backbone

When converting the model to FP16 TensorRT engines:

• TensorRT 8.6.1

  • Engine builds successfully
  • FP16 inference results are correct

• TensorRT 10.3

  • Engine also builds successfully
  • But inference results are incorrect
  • The predicted segmentation map contains almost a single class value across the entire image

I verified that:

• The ONNX model produces correct results
• The inference code is identical between TensorRT versions
• CNN-based segmentation models run correctly on TensorRT 10.3 using the same pipeline

Therefore, the issue seems specific to Transformer-based models (ViT / DINOv3).
After inspecting the TensorRT engine behavior, I suspect there may be numerical instability or FP16 overflow in TensorRT 10.3, possibly related to attention or LayerNorm operations.

Case 2: ViT-L16 backbone

For a larger model using ViT-L16 weights, the situation is worse.

When converting to FP16 TensorRT engines:

• TensorRT 8.6.1
• TensorRT 10.3
• TensorRT 10.15

All versions build successfully, but inference results are incorrect.

From preliminary analysis, this may be caused by precision overflow or numerical instability in FP16, since ViT-L16 has a deeper Transformer architecture.

Observations
During conversion, I also noticed a difference in attention handling:

1. In TensorRT 8.6.1, multi-head attention seems to be fused into optimized kernels
2. In TensorRT 10.3, attention operations appear to be fully decomposed into MatMul / Transpose / Softmax layers
   This difference might be related to the observed inference errors.

Questions

1、Why does TensorRT 8.6.1 produce correct results while TensorRT 10.3 produces incorrect results for the same ViT-B16 model?

2、Could this be related to:

• numerical instability in FP16
• attention decomposition in TensorRT 10.x
• LayerNorm precision issues?

3、For deeper Transformer models such as ViT-L16, what is the recommended way to build stable FP16 TensorRT engines?

For example:

• Should certain layers (e.g., LayerNorm or Softmax) be forced to FP32?
• Are there recommended TensorRT build flags or plugins for Transformer models?

Any guidance would be greatly appreciated.

Start by comparing the ONNX outputs with FP16 TensorRT inference for the ViT-B16 and ViT-L16 cases across TensorRT 8.6.1, 10.3, and 10.15. Inspect the reported difference between fused multi-head attention and decomposed MatMul/Transpose/Softmax operations, along with possible LayerNorm or FP16 overflow. Done means identifying the numerical cause or documenting a stable build configuration.