About the SVTR model using fp16

[AshleyD707]

What is the feature?

After adopting half-precision computations for the model, the predictions exhibited substantial discrepancies when compared to those derived from single-precision (fp32) calculations.

Any other context?

No response

[AshleyD707]

And is the model compatible with AMP training? I tried it train with about 1W data. However, the loss fails to decrease, and the model's predictions consistently yield None outputs.