---
license: mit
pipeline_tag: video-text-to-text
library_name: transformers
---

# M4-LongVA-7B-Qwen2

[Project Page](https://omnimmi.github.io/)

This is the model described in the paper [OmniMMI: A Comprehensive Multi-modal Interaction Benchmark in Streaming Video Contexts](https://huggingface.co/papers/2503.22952).

The abstract of the paper is the following:

> The rapid advancement of multi-modal language models (MLLMs) like GPT-4o has propelled the development of Omni language models, designed to process and proactively respond to continuous streams of multi-modal data. Despite their potential, evaluating their real-world interactive capabilities in streaming video contexts remains a formidable challenge. In this work, we introduce OmniMMI, a comprehensive multi-modal interaction benchmark tailored for OmniLLMs in streaming video contexts. OmniMMI encompasses over 1,121 videos and 2,290 questions, addressing two critical yet underexplored challenges in existing video benchmarks: streaming video understanding and proactive reasoning, across six distinct subtasks. Moreover, we propose a novel framework, Multi-modal Multiplexing Modeling (M4), designed to enable an inference-efficient streaming model that can see, listen while generating.

![images](./assets/framework.png)

Enhancing Interactive Capabilities in MLLM

M4-7B is an extension of [LongVA-7B](https://github.com/EvolvingLMMs-Lab/LongVA), further trained using the [M4-IT](https://huggingface.co/datasets/ColorfulAI/M4-IT) dataset, which comprises 9,963 visual instruction tuning instances. This training was conducted without any special modifications to the existing training pipeline.

## Usage

*Please refer to [M4](https://github.com/patrick-tssn/M4) to install relvevant packages*

```python
import os
from PIL import Image
import numpy as np
import torchaudio
import torch
from decord import VideoReader, cpu
import whisper
# fix seed
torch.manual_seed(0)

from intersuit.model.builder import load_pretrained_model
from intersuit.mm_utils import tokenizer_image_speech_tokens, process_images
from intersuit.constants import IMAGE_TOKEN_INDEX, SPEECH_TOKEN_INDEX


import warnings
warnings.filterwarnings("ignore")

model_path = "checkpoints/M4-LongVA-7B-Qwen2"
video_path = "local_demo/assets/water.mp4"
max_frames_num = 16 # you can change this to several thousands so long you GPU memory can handle it :)
gen_kwargs = {"do_sample": True, "temperature": 0.5, "top_p": None, "num_beams": 1, "use_cache": True, "max_new_tokens": 1024}
tokenizer, model, image_processor, _ = load_pretrained_model(model_path, None, "llava_qwen", device_map="cuda:0", attn_implementation="eager")

# original query
query = "Give a detailed caption of the video as if I am blind."
prompt = f"<|im_start|>system
You are a helpful assistant.<|im_end|>
<|im_start|>user
<image>{query}
<|im_end|>
<|im_start|>assistant
"
input_ids = tokenizer_image_speech_tokens(prompt, tokenizer, IMAGE_TOKEN_INDEX, SPEECH_TOKEN_INDEX, return_tensors="pt").unsqueeze(0).to(model.device)
pad_token_ids = (tokenizer.pad_token_id if tokenizer.pad_token_id is not None else tokenizer.eos_token_id)
attention_masks = input_ids.ne(pad_token_ids).to(input_ids.device)

# new query
new_query = "How many people in the video?"
new_query = "Okay, I see."
new_query = "Sorry to interrupt."
new_query_pos = 10 # which token encounter the new query
new_prompt = f"<|im_start|>system
You are a helpful assistant.<|im_end|>
<|im_start|>user
{new_query}
<|im_end|>
<|im_start|>assistant
"
new_input_ids = tokenizer_image_speech_tokens(new_prompt, tokenizer, IMAGE_TOKEN_INDEX, SPEECH_TOKEN_INDEX, return_tensors="pt").unsqueeze(0).to(model.device)

#video input
vr = VideoReader(video_path, ctx=cpu(0))
total_frame_num = len(vr)
uniform_sampled_frames = np.linspace(0, total_frame_num - 1, max_frames_num, dtype=int)
frame_idx = uniform_sampled_frames.tolist()
frames = vr.get_batch(frame_idx).asnumpy()
video_tensor = image_processor.preprocess(frames, return_tensors="pt")["pixel_values"].to(model.device, dtype=torch.bfloat16)


with torch.inference_mode():
    output_ids = model.generate_parallel(input_ids, 
                                attention_mask=attention_masks,
                                images=[video_tensor], 
                                modalities=["video"], 
                                new_query=new_input_ids,
                                new_query_pos=new_query_pos,
                                query_str=query,
                                new_query_str=new_query,
                                tokenizer=tokenizer,
                                **gen_kwargs)
outputs = tokenizer.batch_decode(output_ids, skip_special_tokens=True)[0].strip()

```

For more information about the interaction inference pipeline, please visit the [M4 GitHub repository](https://github.com/patrick-tssn/M4).