🐛 Bug

When the persistent cache is enabled and XLA_IR_DEBUG and XLA_HLO_DEBUG env var changed, PyTorch/XLA will not recompile a graph with the HLO metadata

To Reproduce

run the same pytorch/xla problem with persistent cache but flip the XLA_IR_DEBUG and XLA_HLO_DEBUG.

Expected behavior

We recompile

Environment

• Reproducible on XLA backend: all
• torch_xla version: nightly

cc @jonb377