map_location hasn't been implemented in load_from_checkpoint

[haossr]

Describe the bug
Current load_from_checkpoint function does not support to load a gpu checkpoint on cpu devices, i.e. map the location.

To Reproduce

# on a cpu only device:
Model.load_from_checkpoint("GPU_CHECKPOINT_FILE")

Expected behavior
AssertionError: Torch not compiled with CUDA enabled

Desktop (please complete the following information):

• OS: macOS Catalina 10.15.1
• Browser: chrome
• Version: 0.5.3.1

[williamFalcon]

weird, i thought we did. mind submitting a PR?

[haossr]

Sure will do.

[gaoalexander]

This appears to still be unresolved, currently using Lightning 2.3.0.