✨[Feature] Route FP16 Layer Norm properly
Open
@narendasan is already working on this.
Since Jan 30, 2026.
feature request
story: Quantization & Precision
- Dominant language
- Python
- Stars
- 3k
- Forks
- 410
- Avg merge
- 3d 18h
- Merged PRs (30d)
- 78
Description
Is your feature request related to a problem? Please describe.
We now see these warnings:
WARNING:torch_tensorrt [TensorRT Conversion Context]:Detected layernorm nodes in FP16.
WARNING:torch_tensorrt [TensorRT Conversion Context]:Running layernorm after self-attention with FP16 Reduce or Pow may cause overflow. Forcing Reduce or Pow Layers in FP32 precision, or exporting the model to use INormalizationLayer (available with ONNX opset >= 17) can help preserving accuracy.
Describe the solution you'd like
Add a lowering pass to detect this case and properly route the subgraph to the right converters
Describe alternatives you've considered
Additional context
Contributor guide
First steps
- Read the whole issue, then the project's contributing guide.
- Comment on the issue to say you are picking it up — it saves two people doing the same work.
- Fork the repository and make your change on a branch.
- Open a pull request that references the issue number.
Assessment
This issue has not been assessed yet.