Forcing layernorm layers to run in FP32 precision

[de1star]

Hi, when I build tensorRT engine, there was a warning:
[W] Running layernorm after self-attention in FP16 may cause overflow. Forcing layernorm layers to run in FP32 precision can help with preserving accuracy.
But I did not found an approach to force layernorm run in fp32 precision, could you help me with that? Thanks a lot!

[rajeevsrao]

@de1star is this an ONNX model you are trying to run? If so can you try exporting to opset 17 (which added LayerNormalization operator) and running with TRT 8.6? Precision requirements for LayerNormalization operator is handled automatically by TensorRT optimizer in 8.6.

[de1star]

Thank you, I will have try!

[de1star]

Hi, I tried to set opset_version as 17, but an error was raised:
ValueError: Unsupported ONNX opset version: 17.
It seems like torch does not support opset_version=17. Any suggestions?

[rajeevsrao]

@de1star you will need to use torch v1.13.0 or newer version.
https://github.com/pytorch/pytorch/blob/v1.13.0-rc1/torch/onnx/symbolic_opset17.py

[ttyio]

Closing since no activity for more than 3 weeks, pls reopen if you still have question. thanks!

[monsterlyg]

I've exported model to opset 17 onnx. The warning still exists. @rajeevsrao

[jinluyang]

Sorry, my bad. I found TensorRT8.6 to be working fine. I got the following error because I had previously a TensorRT8.4 installed, I missed the libnvonnxparser.so while removing TRT8.4. now with TRT8.6 and ONNX opset17, everything works fine, thank you.

Also the way to manually set the layernorm layer to fp32 through tensorrt python api, can be figured out by this link #1196 (comment)

[1193700079]

How can TensorRT be implemented in C++?

[w1005444804]

@rajeevsrao How to Forcing layernorm layers to run in FP32 precision with c++?? I have set "config->setFlag(BuilderFlag::kFP16);"

[focusunsink]

still no solution

[lantudou]

trtexec --onnx=sim_cnn.onnx --saveEngine=model.trt --fp16 --verbose
you will find something like that in output:

[11/23/2024-18:16:34] [W] [TRT] Detected layernorm nodes in FP16.
[11/23/2024-18:16:34] [V] [TRT] /downsample_layers.0/downsample_layers.0.1/ReduceMean_1,/downsample_layers.0/downsample_layers.0.1/ReduceMean,/downsample_layers.0/downsample_layers.0.1/Pow,/downsample_layers.1/downsample_layers.1.0/ReduceMean,/downsample_layers.1/downsample_layers.1.0/Pow,/downsample_layers.1/downsample_layers.1.0/ReduceMean_1,/downsample_layers.2/downsample_layers.2.0/ReduceMean,/downsample_layers.2/downsample_layers.2.0/Pow,/downsample_layers.2/downsample_layers.2.0/ReduceMean_1,/downsample_layers.3/downsample_layers.3.0/ReduceMean,/downsample_layers.3/downsample_layers.3.0/Pow,/downsample_layers.3/downsample_layers.3.0/ReduceMean_1
[11/23/2024-18:16:34] [W] [TRT] 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.

Copy this layer name,and set them to fp32 percision like this command:

trtexec --onnx=sim_cnn.onnx --saveEngine=model.trt --fp16 --precisionConstraints=obey --layerPrecisions=/downsample_layers.0/downsample_layers.0.1/ReduceMean_1:fp32,/downsample_layers.0/downsample_layers.0.1/ReduceMean:fp32,/downsample_layers.0/downsample_layers.0.1/Pow:fp32,/downsample_layers.1/downsample_layers.1.0/ReduceMean:fp32,/downsample_layers.1/downsample_layers.1.0/Pow:fp32,/downsample_layers.1/downsample_layers.1.0/ReduceMean_1:fp32,/downsample_layers.2/downsample_layers.2.0/ReduceMean:fp32,/downsample_layers.2/downsample_layers.2.0/Pow:fp32,/downsample_layers.2/downsample_layers.2.0/ReduceMean_1:fp32,/downsample_layers.3/downsample_layers.3.0/ReduceMean:fp32,/downsample_layers.3/downsample_layers.3.0/Pow:fp32,/downsample_layers.3/downsample_layers.3.0/ReduceMean_1:fp32