fix: disable torch.compile tracing at the LightningModule.log boundary

Move the torch.compiler.disable from the generic is_param_in_hook_signature
helper to LightningModule.log, and revert the helper to pure Python.

Only self.log is reachable from a compiled training_step; its signature
introspection (inspect.getfullargspec on a bound method) fails to trace on
PyTorch 2.12/2.13. Disabling at the log boundary matches the existing idiom
(toggle_optimizer, _ResultCollection.log) and keeps the introspection helper
torch-free, since its other callers run outside the compiled graph.

Author: bhimrazy

src/lightning/pytorch/core/module.py

@@ -383,6 +383,7 @@ def forward(self, x):
             else:
                 print(*args, **kwargs)
 
+    @torch.compiler.disable
     def log(
         self,
         name: str,
@@ -408,6 +409,13 @@ def log(
 
         The default behavior per hook is documented here: :ref:`extensions/logging:Automatic Logging`.
 
+        .. note::
+            This method is decorated with :func:`torch.compiler.disable` so that it runs as regular Python when
+            called from a :func:`torch.compile`-wrapped ``training_step``. Logging is bookkeeping that does not
+            belong in the compiled graph, and tracing it fails on newer PyTorch versions (the signature
+            introspection it performs is not supported by Dynamo). Disabling the compiler keeps behavior identical
+            for eager users.
+
         Args:
             name: key to log. Must be identical across all processes if using DDP or any other distributed strategy.
             value: value to log. Can be a ``float``, ``Tensor``, or a ``Metric``.

src/lightning/pytorch/utilities/signature_utils.py

@@ -14,10 +14,7 @@
 import inspect
 from typing import Callable, Optional
 
-import torch
 
-
-@torch.compiler.disable
 def is_param_in_hook_signature(
     hook_fx: Callable, param: str, explicit: bool = False, min_args: Optional[int] = None
 ) -> bool: