import torchfrom peft import PeftModelfrom transformers import AutoConfig, AutoTokenizer, Qwen3_5ForCausalLM BASE = "Qwen/Qwen3.5-4B"REVISION = "851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a"ADAPTER = "anthonyyazdaniml/umls-agent-qwen3.5-4b-thinking" tokenizer = AutoTokenizer.from_pretrained(ADAPTER, padding_side="left")config = AutoConfig.from_pretrained(BASE, revision=REVISION).text_configbase = Qwen3_5ForCausalLM.from_pretrained( BASE, revision=REVISION, config=config, dtype=torch.float32, device_map="cpu", attn_implementation="sdpa", key_mapping={r"^model\.language_model\.": "model."},)base.config.eos_token_id = tokenizer.eos_token_idbase.generation_config.eos_token_id = tokenizer.eos_token_idmodel = PeftModel.from_pretrained(base, ADAPTER, is_trainable=False)model = model.merge_and_unload(safe_merge=True).to(torch.bfloat16)model.save_pretrained("merged", max_shard_size="2GB")tokenizer.save_pretrained("merged")