[BUG] LR scheduler double-counted when resuming from checkpoint

When resuming training from a checkpoint, the LR scheduler's num_steps is incremented twice, causing the scheduler position to be doubled.

Root Cause

In slime/backends/megatron_utils/model.py, function initialize_model_and_optimizer() (line 786):

iteration, _ = load_checkpoint(model, optimizer, opt_param_scheduler, ...)

opt_param_scheduler.step(increment=iteration * args.global_batch_size)  # ← BUG

Megatron's load_checkpoint() already calls opt_param_scheduler.load_state_dict(), which internally calls self.step(increment=num_steps) with the checkpoint's saved num_steps. Line 786 then adds iteration * global_batch_size again.

Result: scheduler.num_steps is doubled on every resume.

Fix

Remove line 786:

# opt_param_scheduler.step(increment=iteration * args.global_batch_size)

Start in slime/backends/megatron_utils/model.py at initialize_model_and_optimizer() around line 786. Read how load_checkpoint() restores opt_param_scheduler state and compare it with the following scheduler step. Done means resuming from a checkpoint advances num_steps once rather than twice.