import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from safetensors.torch import load_file
from accelerate import init_empty_weights, load_checkpoint_and_dispatch

# Specify the model name and safetensors file path
MODEL_NAME = "mistral-8x7B"
SAFETENSORS_PATH = "path_to_your_model.safetensors"

# Load the tokenizer
tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)

# Initialize an empty model (no weights loaded yet)
with init_empty_weights():
    model = AutoModelForCausalLM.from_pretrained(MODEL_NAME)

# Load the model weights from the safetensors file
model_weights = load_file(SAFETENSORS_PATH)

# Use Hugging Face's `accelerate` to load the model efficiently
# This allows for sharding and offloading to CPU/disk if needed
model = load_checkpoint_and_dispatch(
    model,
    SAFETENSORS_PATH,
    device_map="auto",  # Automatically handles GPU/CPU offloading
    no_split_module_classes=["MistralLayer"],  # Specify layers not to split
    dtype=torch.float16,  # Use mixed precision for memory efficiency
)

# Move the model to the appropriate device
device = "cuda" if torch.cuda.is_available() else "cpu"
model.to(device)

# Example usage
input_text = "Hello, how are you?"
inputs = tokenizer(input_text, return_tensors="pt").to(device)

# Generate output with efficient memory usage
with torch.no_grad():
    outputs = model.generate(
        inputs["input_ids"],
        max_length=50,
        num_return_sequences=1,
        temperature=0.7,
        top_k=50,
        top_p=0.95,
    )

# Decode and print the output
generated_text = tokenizer.decode(outputs[0], skip_special_tokens=True)
print("Generated Text:", generated_text)