from transformers import AutoModelForCausalLM
import torch
import torch.utils.dlpack

# Load the original model
model_name = "./mixed_llm"
model = AutoModelForCausalLM.from_pretrained(model_name, trust_remote_code=True)

# Convert the model to a different precision
model = model.bfloat16()

# Save the model as a safetensor
model.save_pretrained(f"./mixed_llm_half", safetensors=True)