[DDP][PT2D] Fix some tracing bugs of DDP (#124421)

1. We need to clear the cache of get_legacy_mod_inlinelist to ensure the newly added rule will be captured.
2. Don't add the hook if the parameter does not require gradient.

Differential Revision: [D56315534](https://our.internmc.facebook.com/intern/diff/D56315534/)

Pull Request resolved: https://github.com/pytorch/pytorch/pull/124421
Approved by: https://github.com/yf225

torch/nn/parallel/distributed.py

@@ -894,6 +894,7 @@ class DistributedDataParallel(Module, Joinable):
             torch._dynamo.trace_rules.LEGACY_MOD_INLINELIST.add(
                 "torch.nn.parallel.distributed"
             )
+            torch._dynamo.trace_rules.get_legacy_mod_inlinelist.cache_clear()
         self._force_to_disable_cpp_reducer = (
             optimize_ddp == "python_reducer_without_compiled_forward"
         )
@@ -926,6 +927,8 @@ class DistributedDataParallel(Module, Joinable):
                 param.grad.copy_(gradient)
 
         for index, param in enumerate(self._module_parameters):
+            if not param.requires_grad:
+                continue
             self._accum_grad_hooks.append(
                 param.register_post_accumulate_grad_hook(
                     functools.partial(