Explicitly pass expert_tensor_parallel_size to initialize_model_parallel

[nathan-az]

Due to the default handling in MCore, if this isn't passed, expert_tensor_parallel_size is forced to tensor_model_parallel_size. With this PR, we first check whether the parameter is passed, and if not, fall back to the default behaviour with None.

Not extensively tested, but came across the issue and the fix seems straightforward.

(Hopefully) fixes #546

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[ananthsub]

/ok to test 91fba7d

[ananthsub]

Thanks for your contribution @nathan-az - could you share the full stack trace of the error you ran into?

[nathan-az]

@ananthsub created an issue with a repro and stacktrace. Let me know if I've overlooked anything. I'm working out the kinks of using NeMO-RL and reporting/fixing issues along the way, but I'm not at all experienced in the megatron/nemo stack.

[nathan-az]

Also CC @yaoyu-33 since (AFAICT) some EP cases are currently broken, and this fixes them - I'm not sure how often the NeMO-RL submodule branch yuya/nemo-rl-use-chunkpatch is sync'd with main.

[yfw]

Thank you @nathan-az for your contribution! I've verified this fixes an issue I was seeing with expert_tensor_parallel_size not being set correctly when testing NeMo RL with deepseek models. I've included this change in this PR for NeMo RL: NVIDIA-NeMo/RL#1059 to make sure it gets included into NeMo RL.