fix(train): raise clear error when computed train_iters is 0

[Xuan-1998]

What this fixes

Launching training with a rollout configuration whose total sample count is smaller than the global batch size, for example:

--num-rollout 2 --rollout-batch-size 8 --n-samples-per-prompt 8 --global-batch-size 256

computes train_iters = 2 * 8 * 8 // 256 = 0 in get_optimizer_param_scheduler (slime/backends/megatron_utils/model.py). The zero propagates into Megatron's OptimizerParamScheduler, which fails during actor bring-up with a bare AssertionError at megatron/core/optimizer_param_scheduler.py (assert self.lr_decay_steps > 0, no message). Nothing in the traceback points at the slime config knobs that caused it.

What changes

A ValueError is raised right after train_iters is computed, at the point where slime owns the math. The message names the four inputs with their values, shows the product versus global_batch_size, and says how to fix the config:

ValueError: computed train_iters is 0: num_rollout (2) * rollout_batch_size (8) * n_samples_per_prompt (8) = 128 < global_batch_size (256). The optimizer LR schedule needs at least one training step; increase num_rollout, rollout_batch_size, or n_samples_per_prompt, or decrease global_batch_size.

No behavior change for valid configs.

Prior art

• fix(train): skip optimizer and scheduler for eval-only #2296 (which replaced fix(train): support eval-only mode (--num-rollout 0) #2109) handles the eval-only case: num_rollout == 0 returns from initialize_model_and_optimizer before get_optimizer_param_scheduler is called, so that path is unaffected by this change.
• [Bug] train.py num_rollout==0 error #1785 asked for a clear error on num_rollout == 0 and was addressed by fix(train): skip optimizer and scheduler for eval-only #2296.
• The remaining case, num_rollout > 0 but num_rollout * rollout_batch_size * n_samples_per_prompt < global_batch_size, was still unguarded; the validation in slime/utils/arguments.py only checks the batch math when --num-steps-per-rollout is set.

Test

tests/test_train_iters_validation.py is a CPU unit test following the megatron-stubbing pattern of tests/test_model_provider_freeze.py: it loads model.py with stubbed megatron.* modules and calls get_optimizer_param_scheduler directly. One case asserts the ValueError names all four inputs; one case asserts a valid config still computes train_iters and constructs the scheduler with the same step counts as before.

tests/test_train_iters_validation.py::test_zero_train_iters_raises_value_error_naming_inputs PASSED
tests/test_train_iters_validation.py::test_valid_config_builds_scheduler PASSED