Direct_Transformers/models/lfm.py

from transformers import AutoProcessor, AutoModelForImageTextToText, TextIteratorStreamer
from transformers.image_utils import load_image
from threading import Thread
import logging
import torch
import time
import pynvml

class LFM2:
    def __init__(self, model_id):
        self.model_id = model_id
        self.model = AutoModelForImageTextToText.from_pretrained(
            model_id,
            device_map="auto",
            torch_dtype=torch.bfloat16,
            trust_remote_code=True
        )
        self.processor = AutoProcessor.from_pretrained(model_id, trust_remote_code=True)
        
        self.handle = None
        if torch.cuda.is_available():
            try:
                pynvml.nvmlInit()
                device_id = next(self.model.parameters()).device.index
                self.handle = pynvml.nvmlDeviceGetHandleByIndex(device_id)
            except Exception as e:
                logging.error(f"Failed to initialize NVML: {e}")

    def __del__(self):
        if self.handle:
            try:
                pynvml.nvmlShutdown()
            except:
                pass

    def generate(self, video, prompt):
        start_time = time.time()
        conversation = [
            {
                "role": "user",
                "content": [
                    {"type": "text", "text": prompt},
                ],
            },
        ]
        # Assume video is a list of image paths
        images = [load_image(image_path) for image_path in video]
        for image in images:
            conversation[0]["content"].append({"type": "image", "image": image})

        # Generate Answer
        inputs = self.processor.apply_chat_template(
            conversation,
            add_generation_prompt=True,
            return_tensors="pt",
            return_dict=True,
            tokenize=True,
        ).to(self.model.device)
        
        logging.info(f"Prompt token length: {len(inputs.input_ids[0])}")
        streamer = TextIteratorStreamer(self.processor, skip_prompt=True, skip_special_tokens=True)

        generation_kwargs = dict(
            **inputs,
            streamer=streamer,
            max_new_tokens=512
        )

        thread = Thread(target=self.model.generate, kwargs=generation_kwargs)
        thread.start()
        
        full_response = ""
        print("Response: ", end="")
        first_token_time = None
        for new_text in streamer:
            if first_token_time is None:
                first_token_time = time.time()
            full_response += new_text
            print(new_text, end="", flush=True)
        print()
        thread.join()

        end_time = time.time()
        
        if first_token_time is not None:
            generation_time = end_time - first_token_time
        else:
            generation_time = 0
        
        num_generated_tokens = len(self.processor.tokenizer(full_response).input_ids)
        tokens_per_second = num_generated_tokens / generation_time if generation_time > 0 else 0

        peak_memory_mb = 0
        if self.handle:
            mem_info = pynvml.nvmlDeviceGetMemoryInfo(self.handle)
            peak_memory_mb = mem_info.used / (1024 * 1024)

        return {
            "response": full_response,
            "tokens_per_second": tokens_per_second,
            "peak_gpu_memory_mb": peak_memory_mb,
            "num_generated_tokens": num_generated_tokens,
        }