Update custom model files, README, and requirements

asr_modeling.py

@@ -200,17 +200,16 @@ class ASRModel(PreTrainedModel, GenerationMixin):
                     **cache_kwargs,
                 )
                 if adapter_config_file is not None:
-                    # Load saved adapter weights
-                    from pathlib import Path
-
+                    # Load saved adapter weights using the original repo_id/path
+                    # PEFT handles Hub downloads and caching internally
                     from peft import PeftModel
 
-                    adapter_dir = str(Path(adapter_config_file).parent)
                     # language_model is bare (not PEFT-wrapped) since we skipped _setup_lora
                     model.language_model = PeftModel.from_pretrained(
                         model.language_model,
-                        adapter_dir,
+                        pretrained_model_name_or_path,  # Use original repo_id, not cache path
                         is_trainable=True,
+                        **cache_kwargs,
                     )
                 else:
                     # No saved adapters - initialize fresh LoRA for training