app.py

import os
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM
from huggingface_hub import login
import time
import torch.quantization

# Directly assign your Hugging Face token here
hf_token = "your_hugging_face_api_token"

# Log in to Hugging Face
login(token=hf_token)

# Load the Mixtral-8x7B-Instruct model and tokenizer with authorization header
model_name = 'mistralai/Mistral-7B-Instruct-v0.3'
headers = {"Authorization": f"Bearer {hf_token}"}

# Ensure sentencepiece is installed
try:
    import sentencepiece
except ImportError:
    raise ImportError("The sentencepiece library is required for this tokenizer. Please install it with `pip install sentencepiece`.")

# Start time to measure execution time
start_time = time.time()

# Load tokenizer and model
tokenizer = AutoTokenizer.from_pretrained(model_name, use_auth_token=hf_token)
model = AutoModelForCausalLM.from_pretrained(model_name, use_auth_token=hf_token)

# Quantize the model
quantized_model = torch.quantization.quantize_dynamic(model, {torch.nn.Linear}, dtype=torch.qint8)

# Check if a GPU is available and if not, fall back to CPU
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
quantized_model.to(device)

# Measure time for loading tokenizer, model, and quantization
loading_time = time.time() - start_time
print(f"Time taken to load tokenizer, model, and quantize: {loading_time:.2f} seconds")

# Example text input
text_input = "How did Tesla perform in Q1 2024?"

# Start time for inference
inference_start_time = time.time()

# Tokenize the input text
inputs = tokenizer(text_input, return_tensors="pt").to(device)

# Measure time for tokenization
tokenization_time = time.time() - inference_start_time

# Generate a response
outputs = quantized_model.generate(**inputs, max_length=150, do_sample=False)

# Measure time for inference
inference_time = time.time() - inference_start_time
print(f"Time taken for inference with quantized model: {inference_time:.2f} seconds")

# Decode the generated tokens to a readable string
response = tokenizer.decode(outputs[0], skip_special_tokens=True)

# Print the response
print(f"Generated response: {response}")

# Total execution time
total_time = time.time() - start_time
print(f"Total execution time with quantized model: {total_time:.2f} seconds")