vlm.py

"""
File: vlm.py
Description: Vision language model utility functions.
Author: Didier Guillevic
Date: 2025-05-08
"""

from transformers import AutoProcessor
from transformers import Mistral3ForConditionalGeneration
from transformers import TextIteratorStreamer
from threading import Thread
import re
import time
import torch
import base64
import spaces

import logging
logger = logging.getLogger(__name__)
logging.basicConfig(level=logging.INFO)

#
# Load the model: OPEA/Mistral-Small-3.1-24B-Instruct-2503-int4-AutoRound-awq-sym
#
model_id = "OPEA/Mistral-Small-3.1-24B-Instruct-2503-int4-AutoRound-awq-sym"
device = 'cuda' if torch.cuda.is_available() else 'cpu'
processor = AutoProcessor.from_pretrained(model_id)
model = Mistral3ForConditionalGeneration.from_pretrained(
    model_id, 
    #_attn_implementation="flash_attention_2",
    torch_dtype=torch.float16
).eval().to(device)

#
# Encode images as base64
#
def encode_image(image_path):
    """Encode the image to base64."""
    try:
        with open(image_path, "rb") as image_file:
            return base64.b64encode(image_file.read()).decode('utf-8')
    except FileNotFoundError:
        print(f"Error: The file {image_path} was not found.")
        return None
    except Exception as e:  # Added general exception handling
        print(f"Error: {e}")
        return None


#
# Build messages
#
def normalize_message_content(msg: dict) -> dict:
    content = msg.get("content")

    # Case 1: Already in expected format
    if isinstance(content, list) and all(isinstance(item, dict) for item in content):
        return {"role": msg["role"], "content": content}

    # Case 2: String (assume text)
    if isinstance(content, str):
        return {"role": msg["role"], "content": [{"type": "text", "text": content}]}

    # Case 3: Tuple with image path(s)
    if isinstance(content, tuple):
        return {
            "role": msg["role"],
            "content": [
                {"type": "image", "image": encode_image(path)}  # your `encode_image()` function
                for path in content if isinstance(path, str)
            ]
        }

    logger.warning(f"Unexpected content format in message: {msg}")
    return {"role": msg["role"], "content": [{"type": "text", "text": str(content)}]}


def build_messages(message: dict, history: list[dict]):
    """Build messages given message & history from a **multimodal** chat interface.
    Args:
        message: dictionary with keys: 'text', 'files'
        history: list of dictionaries
    
    Returns:
        list of messages (to be sent to the model)
    """
    logger.info(f"{message=}")
    logger.info(f"{history=}")
    
    # Get the user's text and list of images
    user_text = message.get("text", "")
    user_images = message.get("files", [])  # List of images
    
    # Build the user message's content from the provided message
    user_content = []
    if user_text:
        user_content.append({"type": "text", "text": user_text})
    for image in user_images:
        user_content.append(
            {
                "type": "image",
                "image": f"data:image/jpeg;base64,{encode_image(image)}"
            }
        )

    # Normalize existing history content
    messages = [normalize_message_content(msg) for msg in history]
    
    # Append new user message
    messages.append({'role': 'user', 'content': user_content})
    logger.info(f"{messages=}")

    return messages


#
# stream response
#
@spaces.GPU
@torch.inference_mode()
def stream_response(
        messages: list[dict],
        max_new_tokens: int=1_024,
        temperature: float=0.15
    ):
    """Stream the model's response to the chat interface.
    
    Args:
        messages: list of messages to send to the model
    """
    # Generate model's response
    inputs = processor.apply_chat_template(
        messages,
        add_generation_prompt=True,
        tokenize=True,
        return_dict=True,
        return_tensors="pt",
    ).to(model.device, dtype=torch.float16)
    
    # Generate
    streamer = TextIteratorStreamer(
        processor, skip_prompt=True, skip_special_tokens=True)
    generation_args = dict(
        inputs,
        streamer=streamer,
        max_new_tokens=max_new_tokens,
        temperature=temperature,
        top_p=0.9,
        do_sample=True
    )

    thread = Thread(target=model.generate, kwargs=generation_args)
    thread.start()

    partial_message = ""
    for new_text in streamer:
        partial_message += new_text
        yield partial_message