Expose hardcoded Megatron infrastructure parameters to user config

[nic-nvidia]

Summary

NeMo RL hardcodes several Megatron infrastructure parameters (I/O, timeouts, diagnostics) that the underlying framework already supports as configurable. In order to maximize performance on varied storage backends (e.g. NFS), we'd like these exposed in user config.

Parameters to expose

All in nemo_rl/models/megatron/setup.py:

Parameter	Line	Current	Issue
async_save	904	False	Checkpoints block training; async plumbing exists but is disabled
init_process_group timeout	188	600s (default)	Too short for 120B weight conversion; env vars don't override it
check_for_nan_in_grad	608	True	Unnecessary overhead in production
logging_level	600	0	Not overridable
fully_parallel_save/load, load_rng	904	Hardcoded	No user override for debugging or reproducibility

Proposed config

checkpointing:
  async_save: true
  fully_parallel_save: true
  fully_parallel_load: true
  load_rng: false

megatron_cfg:
  distributed_timeout_minutes: 120
  check_for_nan_in_grad: false
  logging_level: 0

Environment

• NeMo RL v0.5.0rc0, Megatron-LM (3rdparty/Megatron-LM, branch yifu/superv3)
• 6x DGX B200, Kubernetes 1.31, KubeRay 1.6.0, NFS storage
• Model: Nemotron 3 Super 120B, GRPO with DAPO-Math-17k

Workarounds

• async_save: patch line 904 (async_save=True)
• Timeout: patch line 188 (timeout=timedelta(minutes=120))
• NaN check: patch line 608 (check_for_nan_in_grad=False)

[zeel2104]

Is it open for contribution?