Fix push_to_hub: use upload_folder instead (create_model_card kwarg compat)

Deploy BF-Router trainer (Qwen3-4B QLoRA + Gradio)

fix: remove redundant evaluate() call — use epoch eval_loss from callback

fix: tokenizer -> processing_class for trl 0.29 compat

fix: rename max_seq_length to max_length for trl >= 0.16 compat

Deploy BF-Router trainer (Qwen3-4B QLoRA + Gradio)