Describe the bug
When we run gpt-oss-20b PEFT training on a single GPU, the script fails because n_routed_experts is not recognized.

[rank0]:   File "/opt/Automodel/nemo_automodel/components/moe/layers.py", line 493, in forward
[rank0]:     assert self.n_routed_experts % self.ep_size == 0, (
[rank0]:            ^^^^^^^^^^^^^^^^^^^^^
[rank0]:   File "/usr/local/lib/python3.12/dist-packages/torch/nn/modules/module.py", line 1964, in __getattr__
[rank0]:     raise AttributeError(
[rank0]: AttributeError: 'GroupedExpertsDeepEP' object has no attribute 'n_routed_experts'

When we disable DeepEP, the script runs as expected.

The difference between GroupedExperts and GroupedExpertsDeepEP is that n_routed_experts is only set in the DeepEP version by the method init_token_dispatcher, which appears to only be called in multi-GPU settings. In vanilla GroupedExperts it's set in the __init__ method.

Steps/Code to reproduce bug

Run an MOE config with DeepEP enabled on a single GPU.

Expected behavior

Script runs.