Fix tokenizer auto-detection in SFT and DPO training to support GPT-2 vocab size

Fix LoRA unwrapping in SFT final save so output model can be served directly

upgrading the model

Initial commit