BioMiner-Instruct

import torch
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer, Qwen3VLForConditionalGeneration
import os
# 1. set path
base_model_path = "./qwen3-vl-32b-instruct"  # path of base model
lora_adapter_path = "./biominer-instruct-lora-adapter" # lora adapter path
save_path = "./biominer-instruct" # save path

print(f"Loading base model from: {base_model_path}")

# 2. load model and tokenizer
base_model = Qwen3VLForConditionalGeneration.from_pretrained(
    base_model_path,
    torch_dtype=torch.bfloat16,
    device_map="cpu"
)
tokenizer = AutoTokenizer.from_pretrained(base_model_path)

print(f"Loading LoRA adapter from: {lora_adapter_path}")

# 3. load PETF model 
model_with_lora = PeftModel.from_pretrained(base_model, lora_adapter_path, device_map="cpu")

print("Merging LoRA weights into the base model...")

# 4. key step：merge lora weight and remove lora layer
merged_model = model_with_lora.merge_and_unload()

print(f"Saving merged model to: {save_path}")

# 5. save model and tokenizer
merged_model.save_pretrained(save_path)
tokenizer.save_pretrained(save_path)

print("Done! The merged model has been saved.")