Direct_Transformers/models/qwen.py

from transformers import Qwen2VLForConditionalGeneration, AutoTokenizer, AutoProcessor, Qwen2_5_VLForConditionalGeneration
from qwen_vl_utils import process_vision_info
from transformers import TextIteratorStreamer
from threading import Thread
import logging
import torch
import time
import pynvml

class Qwen2VL:
    def __init__(self, model_id):
        self.model_id = model_id
        if "2.5" in model_id:
            self.model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
                model_id, torch_dtype="float16", device_map="auto"
            )
        else:   
            self.model = Qwen2VLForConditionalGeneration.from_pretrained(
                model_id, torch_dtype="float16", device_map="auto"
            )
        self.processor = AutoProcessor.from_pretrained(model_id)

        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 hasattr(self, 'handle') and self.handle:
            try:
                pynvml.nvmlShutdown()
            except:
                pass

    def generate(self, video, prompt):

        start_time = time.time()

        # Preparation for inference
        video_paths = [f"file://{path}" for path in video]
        messages = [
            {
                "role": "user",
                "content": [
                    {
                        "type": "video",
                        "video": video_paths,
                        "resized_height": 280,
                        "resized_width": 420,
                    },
                    {"type": "text", "text": prompt},
                ],
            }
        ]
        text = self.processor.apply_chat_template(
            messages, tokenize=False, add_generation_prompt=True
        )

        image_inputs, video_inputs = process_vision_info(messages)
        inputs = self.processor(
            text=[text],
            images=image_inputs,
            videos=video_inputs,
            padding=True,
            return_tensors="pt",
        )
        inputs = inputs.to("cuda")
        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=256
        )

        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,
        }