gradient value should be unscaled when precision=16

[AndersJing]

🐛 Bug

On 1.4.x, when track_grad_norm=2 and precision=16, the logged gradient value is the scaled value.

To Reproduce

https://colab.research.google.com/drive/1kV3wP8HmprPVflKXfY5eMdqIOar8reKk?usp=sharing

Expected behavior

The gradient value should be unscaled when preciosn=16

Environment

• CUDA:
  • GPU:
    • Tesla K80
  • available: True
  • version: 10.2
• Packages:
  • numpy: 1.19.5
  • pyTorch_debug: False
  • pyTorch_version: 1.9.0+cu102
  • pytorch-lightning: 1.4.4
  • tqdm: 4.62.0
• System:
  • OS: Linux
  • architecture:
    • 64bit
  • processor: x86_64
  • python: 3.7.11
  • version: Sat Jun 5 09:50:34 PDT 2021

Additional context

On v1.3.8, the behavior is normal, see https://colab.research.google.com/drive/1AXMhr1gPll5ZPDTphQqAOJlI_9iEMISU?usp=sharing

@carmocca

[cowwoc]

I believe this is fixed in release 1.4.9: https://github.com/PyTorchLightning/pytorch-lightning/releases/tag/1.4.9

Can this issue be closed?

[AndersJing]

I run the code again on 1.4.9, the issue is fixed.