Fix division by zero in LR scheduler when max_steps equals warmup_steps

[Br1an67]

Fixes #1393

Summary

This PR fixes a division by zero error that occurs in the learning rate scheduler when max_steps equals lr_warmup_steps. The issue was caused by the CosineAnnealingLR scheduler receiving a T_max value of 0, which results in a ZeroDivisionError during training.

Changes

• Added validation in the get_lr_scheduler function across all finetune modules (adapter.py, adapter_v2.py, full.py, lora.py, lora_legacy.py)
• The validation ensures that max_steps > warmup_steps before creating the scheduler
• If validation fails, a clear error message is raised indicating the problematic values

Testing

The fix was verified by:

1. Ensuring all modified Python files parse correctly
2. The validation will catch the problematic case early with a clear error message instead of failing during training with a cryptic division by zero error

Files Changed

 litgpt/finetune/adapter.py     | 2 ++
 litgpt/finetune/adapter_v2.py  | 2 ++
 litgpt/finetune/full.py        | 2 ++
 litgpt/finetune/lora.py        | 2 ++
 litgpt/finetune/lora_legacy.py | 2 ++
 5 files changed, 10 insertions(+)