What changed

get_optimizer_param_scheduler now computes the estimated training-iteration count once and clamps the scheduler-visible train_iters to at least 1 before deriving Megatron LR/WD schedule steps.

A CPU regression test (tests/test_eval_only_optimizer_scheduler.py) stubs Megatron's OptimizerParamScheduler, preserves its lr_decay_steps > 0 assertion, and is registered in the cpu-unittest matrix.

Why

train.py has an eval-only path for --num-rollout 0 with --eval-interval, but model and optimizer setup run before that branch. With num_rollout == 0, the old estimate produced train_iters == 0, then lr_decay_steps == 0, so Megatron aborted before eval could start.

The clamp only gives the scheduler a valid nonzero size for zero-estimated runs. It does not add training iterations; the training loop is still controlled by args.num_rollout. For normal configs that already estimate at least one optimizer step, the value is unchanged.

Validation

tests/test_eval_only_optimizer_scheduler.py covers both the eval-only startup case (num_rollout=0 no longer trips Megatron's scheduler assertion and sets train_iters == 1) and a normal training config where train_iters remains 16.

Fixes #1785